diff --git a/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16.json b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16.json new file mode 100644 index 0000000000..daa633d535 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T12:28:29.944656+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256","mode":"low-latency","nodes":2,"phase":"decode","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256],"ladder_dropped":[512],"ladder_cap":256},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":61.65099889039993,"p90":64.7369995713234,"p95":65.77800214290619,"p99":67.42099672555923},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":65.13699889183044,"p90":68.0219978094101,"p95":69.58399713039398,"p99":72.94899970293045},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":126.78799778223038,"p90":132.7589973807335,"p95":135.36199927330017,"p99":140.36999642848969},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":105.51799833774567,"p90":108.60300064086914,"p95":110.68499833345413,"p99":114.85199630260468},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":119.29800361394882,"p90":121.1019977927208,"p95":122.34299629926682,"p99":124.30600076913834},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":51.75799876451492,"p90":54.040998220443726,"p95":54.482001811265945,"p99":55.76400086283684},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":48.99299889802933,"p90":51.316998898983,"p95":53.15899848937988,"p99":55.48299849033356},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.248995661735535,"p95":6.248995661735535,"p99":6.248995661735535},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.769999861717224,"p90":11.69700175523758,"p95":14.301002025604248,"p99":17.305999994277954},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.0410025715827942,"p90":1.1620000004768372,"p95":1.1620000004768372,"p99":1.1620000004768372},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":51.316998898983,"p90":52.91900038719177,"p95":53.60100045800209,"p99":54.84199896454811},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":50.71600154042244,"p90":51.596999168395996,"p95":52.156999707221985,"p99":53.599998354911804},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":100.6309986114502,"p90":104.75599765777588,"p95":105.1969975233078,"p99":106.31900280714035},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.348000943660736,"p90":20.872004330158234,"p95":21.83300256729126,"p99":23.876003921031952},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":9,"mean":5.125,"min":2,"total":82},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":134117.9191210641,"p90":132120.03345630792,"p95":130779.86058852056,"p99":128714.62279375613},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":62.9739984869957,"p90":65.89800119400024,"p95":66.85999780893326,"p99":68.82300227880478},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":67.30099767446518,"p90":70.7859992980957,"p95":73.02899658679962,"p99":76.19400322437286},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":130.27499616146088,"p90":136.68400049209595,"p95":139.88899439573288,"p99":145.01700550317764},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":109.40399765968323,"p90":112.44799941778183,"p95":114.21100050210953,"p99":118.0569976568222},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":121.06099724769592,"p90":123.58500063419342,"p95":124.70699846744537,"p99":126.62899494171143},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":54.84199896454811,"p90":57.04500153660774,"p95":57.60600045323372,"p99":58.36699903011322},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":49.19400066137314,"p90":51.47700011730194,"p95":52.39899829030037,"p99":56.52499943971634},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129004061222076,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.930999457836151,"p90":11.859007179737091,"p95":14.341004192829132,"p99":18.307000398635864},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.921003520488739,"p90":1.5220046043395996,"p95":1.5220046043395996,"p99":1.5220046043395996},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":55.4019995033741,"p90":56.405000388622284,"p95":56.76500126719475,"p99":57.60600045323372},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":50.19500106573105,"p90":51.51699855923653,"p95":52.11799964308739,"p99":53.55999991297722},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":103.31500321626663,"p90":104.95699942111969,"p95":106.31900280714035,"p99":108.64199697971344},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.746999859809875,"p90":19.548997282981873,"p95":20.270995795726776,"p99":22.673003375530243},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":15,"mean":10.375,"min":7,"total":166},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":264329.5588795345,"p90":258931.09872385487,"p95":256601.47700815336,"p99":252706.7360420093},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":65.05700200796127,"p90":67.82200187444687,"p95":68.90299916267395,"p99":71.10600173473358},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":70.34499943256378,"p90":73.02899658679962,"p95":74.63099807500839,"p99":78.6380022764206},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":135.40200144052505,"p90":140.8509984612465,"p95":143.53399723768234,"p99":149.74400401115417},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":111.20600253343582,"p90":113.65000158548355,"p95":114.85199630260468,"p99":117.29499697685242},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":126.10900402069092,"p90":127.9119998216629,"p95":129.1929930448532,"p99":131.39599561691284},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":56.44400045275688,"p90":58.04700031876564,"p95":58.36800113320351,"p99":59.087999165058136},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":51.596999168395996,"p90":53.95999923348427,"p95":55.121999233961105,"p99":58.006998151540756},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129004061222076,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1689987778663635,"p90":9.81399416923523,"p95":11.01599633693695,"p99":14.22099769115448},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.8800029754638672,"p90":0.6819963455200195,"p95":0.6819963455200195,"p99":0.6819963455200195},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":56.644998490810394,"p90":57.76499956846237,"p95":58.04700031876564,"p99":58.687999844551086},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":55.362001061439514,"p90":56.56500160694122,"p95":57.00499936938286,"p99":58.12700092792511},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":107.76100307703018,"p90":109.40299928188324,"p95":110.04400253295898,"p99":111.52700334787369},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.387995660305023,"p90":20.59100568294525,"p95":21.272003650665283,"p99":23.235000669956207},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":26,"mean":20.5625,"min":15,"total":329},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":507497.4661563374,"p90":500343.98718830047,"p95":495382.90345034807,"p99":487077.2484314746},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":68.34200024604797,"p90":71.0659995675087,"p95":72.14800268411636,"p99":74.31100308895111},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":72.14800268411636,"p90":74.47099685668945,"p95":75.95399767160416,"p99":78.27699929475784},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":140.49000293016434,"p90":145.53699642419815,"p95":148.10200035572052,"p99":152.58800238370895},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":114.2909973859787,"p90":116.69400334358215,"p95":119.098000228405,"p99":121.22099846601486},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":124.78700280189514,"p90":127.03000009059906,"p95":128.151997923851,"p99":130.11400401592255},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":54.520998150110245,"p90":57.486001402139664,"p95":58.166999369859695,"p99":59.04800072312355},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":52.3190014064312,"p90":56.08399957418442,"p95":56.96500092744827,"p99":59.32899937033653},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.0889944434165955,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.727998912334442,"p90":13.90000432729721,"p95":16.18500053882599,"p99":18.30700784921646},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.5209967494010925,"p90":0.7219985127449036,"p95":0.7219985127449036,"p99":0.7219985127449036},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":56.84499815106392,"p90":58.12700092792511,"p95":58.44699963927269,"p99":59.28900092840195},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":49.59399998188019,"p90":51.83799937367439,"p95":52.43900045752525,"p99":53.881000727415085},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":109.96399819850922,"p90":110.96599698066711,"p95":111.32600158452988,"p99":112.08800226449966},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.862999320030212,"p90":17.305999994277954,"p95":18.306002020835876,"p99":20.309992134571075},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":49,"mean":40.5625,"min":36,"total":649},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":1025747.8513463908,"p90":1007635.9907794154,"p95":998813.924665136,"p99":983752.6787995561},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":79.67899739742279,"p90":82.28199928998947,"p95":83.36500078439713,"p99":85.80800145864487},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":77.79700309038162,"p90":80.08000254631042,"p95":81.28199726343155,"p99":83.24500173330307},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":157.4760004878044,"p90":162.3620018362999,"p95":164.64699804782867,"p99":169.05300319194794},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":123.58500063419342,"p90":125.02799928188324,"p95":125.9080022573471,"p99":129.63399291038513},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":132.87800550460815,"p90":135.72299480438232,"p95":136.96500658988953,"p99":139.0880048274994},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":58.566998690366745,"p90":63.4550005197525,"p95":64.17600065469742,"p99":65.05700200796127},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":54.802000522613525,"p90":59.56900119781494,"p95":60.371000319719315,"p99":61.531998217105865},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.129004061222076,"p95":6.129004061222076,"p99":6.129004061222076},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.805997014045715,"p90":8.77300649881363,"p95":9.773999452590942,"p99":12.538999319076538},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.5209967494010925,"p90":0.6819963455200195,"p95":0.6819963455200195,"p99":0.6819963455200195},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":58.848001062870026,"p90":60.451000928878784,"p95":60.97099930047989,"p99":63.21399658918381},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":51.95600166916847,"p90":53.1190000474453,"p95":53.961001336574554,"p99":56.16400018334389},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":121.8620017170906,"p90":122.82299995422363,"p95":123.10399860143661,"p99":123.6639991402626},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.976999998092651,"p90":13.90199363231659,"p95":15.222005546092987,"p99":17.347000539302826},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":89,"mean":82.125,"min":72,"total":1314},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":1926579.188390377,"p90":1886194.748126307,"p95":1869090.5536662634,"p99":1840561.3073355819},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":92.45800226926804,"p90":95.34300118684769,"p95":96.46400064229965,"p99":98.42699766159058},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":88.93299847841263,"p90":91.21699631214142,"p95":92.49799698591232,"p99":94.70099955797195},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":181.39100074768066,"p90":186.5599974989891,"p95":188.96199762821198,"p99":193.12799721956253},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":147.70099520683289,"p90":149.02299642562866,"p95":149.384006857872,"p99":151.2259989976883},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":153.30900251865387,"p90":155.2720069885254,"p95":156.31400048732758,"p99":158.0359935760498},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":68.86199861764908,"p90":75.8339986205101,"p95":76.91500335931778,"p99":77.83599942922592},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":64.05600160360336,"p90":70.74499875307083,"p95":71.90799713134766,"p99":72.8290006518364},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.168991327285767,"p95":6.168991327285767,"p99":6.168991327285767},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.727005958557129,"p90":8.612990379333496,"p95":10.25499403476715,"p99":14.261007308959961},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.6409883499145508,"p90":-0.5210041999816895,"p95":-0.5210041999816895,"p99":-0.5210041999816895},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":66.2589967250824,"p90":68.5029998421669,"p95":69.58399713039398,"p99":71.5470016002655},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":62.613002955913544,"p90":63.93600255250931,"p95":64.77700173854828,"p99":66.89999997615814},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":144.29600536823273,"p90":145.53800225257874,"p95":145.93900740146637,"p99":147.09900319576263},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.05299186706543,"p90":11.017009615898132,"p95":11.857002973556519,"p99":13.380005955696106},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.004048583097755909,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":178,"mean":164.0,"min":148,"total":2624},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":3339660.3695057137,"p90":3297439.183856475,"p95":3275458.362039093,"p99":3239768.285783683},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":122.50299751758575,"p90":125.58799982070923,"p95":126.79000198841095,"p99":128.9130002260208},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":119.098000228405,"p90":121.46099656820297,"p95":122.50299751758575,"p99":124.94699656963348},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":241.60099774599075,"p90":247.0489963889122,"p95":249.2929995059967,"p99":253.8599967956543},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":211.5969955921173,"p90":212.9189968109131,"p95":213.4779989719391,"p99":214.60099518299103},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":215.20200371742249,"p90":217.72600710391998,"p95":218.72800588607788,"p99":220.210000872612},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":101.51100158691406,"p90":108.08099806308746,"p95":108.56200009584427,"p99":109.52399671077728},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":96.90500050783157,"p90":103.6359965801239,"p95":104.27600145339966,"p99":105.39700090885162},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.130009889602661,"p90":6.209000945091248,"p95":6.209000945091248,"p99":6.209000945091248},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":9.893998503684998,"p95":10.815009474754333,"p99":13.420000672340393},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.8410066366195679,"p90":0.881001353263855,"p95":0.881001353263855,"p99":0.881001353263855},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":100.14999657869339,"p90":102.75399684906006,"p95":103.83599996566772,"p99":105.87800294160843},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":94.5810005068779,"p90":95.90300172567368,"p95":96.54399752616882,"p99":98.46699982881546},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":206.6279947757721,"p90":207.99100399017334,"p95":208.3909958600998,"p99":209.15299654006958},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.652999997138977,"p90":11.256009340286255,"p95":12.178003787994385,"p99":14.142006635665894},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.006103515625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":363,"mean":328.5,"min":292,"total":5256},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":4758320.007766258,"p90":4703158.862924666,"p95":4681613.567735534,"p99":4650106.6978895655},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":172.89799451828003,"p90":184.23500657081604,"p95":187.19999492168427,"p99":191.00500643253326},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":167.17100143432617,"p90":172.61800169944763,"p95":173.61900210380554,"p99":176.02400481700897},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":340.0689959526062,"p90":356.8530082702637,"p95":360.8189970254898,"p99":367.02901124954224},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":303.1730055809021,"p90":309.9839985370636,"p95":311.4660084247589,"p99":313.7100040912628},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":314.3500089645386,"p90":318.5169994831085,"p95":319.47800517082214,"p99":321.9200074672699},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":141.93199574947357,"p90":154.5110046863556,"p95":155.87300062179565,"p99":157.7560007572174},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":130.63499331474304,"p90":143.77500116825104,"p95":147.10000157356262,"p99":151.62600576877594},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1699748039245605,"p90":6.3700079917907715,"p95":6.3700079917907715,"p99":6.3700079917907715},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.582008123397827,"p90":21.792978048324585,"p95":24.07398819923401,"p99":27.723997831344604},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.882016658782959,"p90":3.3240020275115967,"p95":3.3240020275115967,"p99":3.3240020275115967},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":145.77800035476685,"p90":153.9900004863739,"p95":155.2720069885254,"p99":156.83400630950928},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":140.93099534511566,"p90":145.29700577259064,"p95":145.9379941225052,"p99":147.37999439239502},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":296.0430085659027,"p90":304.81499433517456,"p95":306.2179982662201,"p99":307.81999230384827},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":15.262991189956665,"p90":23.313015699386597,"p95":24.917006492614746,"p99":27.3989737033844},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.006103515625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":692,"mean":660.8125,"min":603,"total":10573},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":6515030.830589327,"p90":6429798.105983379,"p95":6410456.954321322,"p99":6361828.87827568},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":254.61998581886292,"p90":266.0380005836487,"p95":269.68300342559814,"p99":275.9320139884949},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":263.27401399612427,"p90":273.68998527526855,"p95":276.9339978694916,"p99":279.29800748825073},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":517.8939998149872,"p90":539.7279858589172,"p95":546.6170012950897,"p99":555.2300214767456},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":501.9490122795105,"p90":515.6509876251221,"p95":517.3730254173279,"p99":521.6190218925476},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":496.5820014476776,"p90":504.11397218704224,"p95":505.7569742202759,"p99":508.84199142456055},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":202.78300344944,"p90":234.8710000514984,"p95":238.55699598789215,"p99":241.20000004768372},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":207.3490023612976,"p90":237.31400072574615,"p95":244.2059963941574,"p99":250.21401047706604},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.371008396148682,"p90":7.852017879486084,"p95":7.852017879486084,"p99":7.852017879486084},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.19002628326416,"p90":33.33202004432678,"p95":38.6199951171875,"p99":48.59501123428345},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-12.057989835739136,"p90":-9.814977645874023,"p95":-9.814977645874023,"p99":-9.814977645874023},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":230.46499490737915,"p90":237.59500682353973,"p95":238.75699937343597,"p99":240.11799693107605},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":239.51800167560577,"p90":248.5709935426712,"p95":250.3339946269989,"p99":253.45799326896667},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":480.3600013256073,"p90":487.2890114784241,"p95":490.13200402259827,"p99":493.5790002346039},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":16.663014888763428,"p90":33.37201476097107,"p95":37.41699457168579,"p99":42.98597574234009},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.006103515625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":1373,"mean":1323.8125,"min":1247,"total":21181},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":8248385.942420379,"p90":8125146.744554533,"p95":8098751.394016448,"p99":8049650.125243764},"tokens_per_rank":256}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"InterNodeV1LL","block_num":32,"rdma_block_num":16,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":0,"scale_type_size":0,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"inter-node-v1ll-vllm-cudagraph-spin-ar","rank_alignment":{"method":"in-graph-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":true,"cuda_graph_supported":true,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9943,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2102,"max":2103,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2102,"max":2104,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2074,"p50":2100,"max":2102,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2074,"p50":2104,"max":2106,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2069,"p50":2096,"max":2097,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2076,"p50":2106,"max":2108,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2080,"p50":2108,"max":2110,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2069,"p50":2097,"max":2099,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2074,"p50":2103,"max":2105,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2069,"p50":2104,"max":2106,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2082,"p50":2106,"max":2108,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2072,"p50":2101,"max":2102,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":54,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2104,"max":2106,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2071,"p50":2102,"max":2103,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2072,"p50":2103,"max":2104,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2071,"p50":2102,"max":2103,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":731.1}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-fp8.json b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-fp8.json new file mode 100644 index 0000000000..987efbeccb --- /dev/null +++ b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T12:43:45.138911+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256","mode":"low-latency","nodes":2,"phase":"decode","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256],"ladder_dropped":[512],"ladder_cap":256},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":61.57200038433075,"p90":64.17600065469742,"p95":65.25799632072449,"p99":66.97999686002731},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":68.0219978094101,"p90":71.0659995675087,"p95":72.42900133132935,"p99":75.83300024271011},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":129.59399819374084,"p90":135.24200022220612,"p95":137.68699765205383,"p99":142.81299710273743},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":108.88300091028214,"p90":111.28699779510498,"p95":112.68799751996994,"p99":115.53200334310532},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":121.78300321102142,"p90":123.94600361585617,"p95":125.06699562072754,"p99":127.07099318504333},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":51.79699882864952,"p90":53.679998964071274,"p95":54.200999438762665,"p99":55.282000452280045},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":52.79900133609772,"p90":54.721999913454056,"p95":55.52300065755844,"p99":57.406000792980194},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.249003112316132,"p90":6.249003112316132,"p95":6.249003112316132,"p99":6.249003112316132},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.849996745586395,"p90":10.734997689723969,"p95":12.258000671863556,"p99":15.142001211643219},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.8419975638389587,"p90":0.8010044693946838,"p95":0.8010044693946838,"p99":0.8010044693946838},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":51.99800059199333,"p90":53.961001336574554,"p95":54.361000657081604,"p99":55.08299916982651},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":54.680999368429184,"p90":55.68299815058708,"p95":56.24400079250336,"p99":57.9259991645813},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":105.07699847221375,"p90":106.9990023970604,"p95":107.51999914646149,"p99":109.68399792909622},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":16.94600284099579,"p90":19.46999877691269,"p95":20.429998636245728,"p99":22.59499579668045},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":587776,"scale_bytes":18368,"total_logical_bytes":606144},"roundtrip":{"activation_data_bytes":1763328,"scale_bytes":18368,"total_logical_bytes":1781696},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":587776,"scale_bytes":18368,"total_logical_bytes":606144},"roundtrip":{"activation_data_bytes":1763328,"scale_bytes":18368,"total_logical_bytes":1781696},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":9,"mean":5.125,"min":2,"total":82},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":131381.22380079384,"p90":129088.47024699998,"p95":127931.43323375954,"p99":125913.86593398603},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":63.21399658918381,"p90":65.65800309181213,"p95":66.7399987578392,"p99":68.62200051546097},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":70.22599875926971,"p90":73.8309994339943,"p95":75.63400268554688,"p99":79.96000349521637},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":133.43999534845352,"p90":139.48900252580643,"p95":142.37400144338608,"p99":148.58200401067734},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":111.80800199508667,"p90":114.81200158596039,"p95":116.37400090694427,"p99":119.65899914503098},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":122.42300063371658,"p90":125.10700523853302,"p95":126.30900740623474,"p99":128.55200469493866},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":53.961001336574554,"p90":55.883001536130905,"p95":56.36399984359741,"p99":57.326000183820724},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":52.71900072693825,"p90":54.40099909901619,"p95":55.16299977898598,"p99":58.44699963927269},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1689987778663635,"p90":6.209000945091248,"p95":6.209000945091248,"p99":6.209000945091248},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.4909979701042175,"p90":10.614998638629913,"p95":12.217998504638672,"p99":14.862000942230225},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.4409998655319214,"p90":1.2020021677017212,"p95":1.2020021677017212,"p99":1.2020021677017212},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":54.80299890041351,"p90":56.16400018334389,"p95":56.56500160694122,"p99":57.326000183820724},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":53.19900065660477,"p90":54.40099909901619,"p95":54.882001131772995,"p99":56.724999099969864},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":106.1989963054657,"p90":108.08099806308746,"p95":109.08199846744537,"p99":110.80600321292877},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":16.22600108385086,"p90":18.269002437591553,"p95":19.189998507499695,"p99":21.072007715702057},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":1189888,"scale_bytes":37184,"total_logical_bytes":1227072},"roundtrip":{"activation_data_bytes":3569664,"scale_bytes":37184,"total_logical_bytes":3606848},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":1189888,"scale_bytes":37184,"total_logical_bytes":1227072},"roundtrip":{"activation_data_bytes":3569664,"scale_bytes":37184,"total_logical_bytes":3606848},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":15,"mean":10.375,"min":7,"total":166},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":261388.78996882605,"p90":255781.040709813,"p95":253346.93587672393,"p99":248926.4953583404},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":65.53799659013748,"p90":68.10200214385986,"p95":69.06399875879288,"p99":71.50600105524063},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":73.18899780511856,"p90":76.23399794101715,"p95":78.15700024366379,"p99":81.60199970006943},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":138.72699439525604,"p90":144.336000084877,"p95":147.22099900245667,"p99":153.10800075531006},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":114.37100172042847,"p90":118.21699887514114,"p95":119.41900104284286,"p99":121.98299914598465},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":128.9529949426651,"p90":131.4370036125183,"p95":132.59899616241455,"p99":134.80199873447418},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":56.3649982213974,"p90":58.006998151540756,"p95":58.28800052404404,"p99":58.96800011396408},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":54.96300011873245,"p90":57.245999574661255,"p95":58.246999979019165,"p99":61.45099923014641},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.209000945091248,"p90":6.209999322891235,"p95":6.209999322891235,"p99":6.209999322891235},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.210001349449158,"p90":11.817000806331635,"p95":13.379000127315521,"p99":15.463002026081085},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.002999186515808,"p90":1.7629936337471008,"p95":1.7629936337471008,"p99":1.7629936337471008},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":55.60300126671791,"p90":57.04500153660774,"p95":57.565998286008835,"p99":58.48800018429756},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":58.36699903011322,"p90":60.770999640226364,"p95":61.170998960733414,"p99":62.532998621463776},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":111.44600063562393,"p90":112.92900145053864,"p95":113.5689988732338,"p99":115.73299765586853},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.627999186515808,"p90":19.86999809741974,"p95":20.791999995708466,"p99":22.754989564418793},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":2358272,"scale_bytes":73696,"total_logical_bytes":2431968},"roundtrip":{"activation_data_bytes":7074816,"scale_bytes":73696,"total_logical_bytes":7148512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":2358272,"scale_bytes":73696,"total_logical_bytes":2431968},"roundtrip":{"activation_data_bytes":7074816,"scale_bytes":73696,"total_logical_bytes":7148512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":26,"mean":20.5625,"min":15,"total":329},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":496304.8747216425,"p90":486925.28162521595,"p95":482658.2542269723,"p99":474770.4084571016},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":67.06000119447708,"p90":69.86399739980698,"p95":71.02599740028381,"p99":73.10900092124939},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":73.43000173568726,"p90":76.71499997377396,"p95":78.59800010919571,"p99":82.80400186777115},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":140.49000293016434,"p90":146.57899737358093,"p95":149.62399750947952,"p99":155.91300278902054},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":114.93299901485443,"p90":118.0569976568222,"p95":119.29800361394882,"p99":122.06300348043442},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":127.63099372386932,"p90":130.11500239372253,"p95":131.35699927806854,"p99":133.12000036239624},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":56.36399984359741,"p90":58.407001197338104,"p95":58.96800011396408,"p99":59.849999845027924},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":55.121999233961105,"p90":57.8869991004467,"p95":58.407001197338104,"p99":60.130998492240906},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.170004606246948,"p90":6.209999322891235,"p95":6.209999322891235,"p99":6.209999322891235},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.56900554895401,"p90":11.414997279644012,"p95":12.778997421264648,"p99":16.103997826576233},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.7219985127449036,"p90":0.9609982371330261,"p95":0.9609982371330261,"p99":0.9609982371330261},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":57.96699970960617,"p90":58.928001672029495,"p95":59.20900031924248,"p99":59.72899869084358},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":54.52200025320053,"p90":56.20399862527847,"p95":56.80499970912933,"p99":57.52599984407425},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":111.7670014500618,"p90":113.24899643659592,"p95":113.81100118160248,"p99":115.37200212478638},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":15.902996063232422,"p90":18.146999180316925,"p95":19.14999634027481,"p99":21.152004599571228},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":4652032,"scale_bytes":145376,"total_logical_bytes":4797408},"roundtrip":{"activation_data_bytes":13956096,"scale_bytes":145376,"total_logical_bytes":14101472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":4652032,"scale_bytes":145376,"total_logical_bytes":4797408},"roundtrip":{"activation_data_bytes":13956096,"scale_bytes":145376,"total_logical_bytes":14101472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":49,"mean":40.5625,"min":36,"total":649},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":1002891.1964513026,"p90":983745.1304245255,"p95":974443.6969745165,"p99":961538.458920839},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":72.5879967212677,"p90":75.71300119161606,"p95":76.6339972615242,"p99":77.99699902534485},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":77.23499834537506,"p90":79.96000349521637,"p95":81.48200064897537,"p99":83.96600186824799},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":149.82299506664276,"p90":155.67300468683243,"p95":158.11599791049957,"p99":161.96300089359283},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":122.18300253152847,"p90":124.30600076913834,"p95":126.74899399280548,"p99":130.35599887371063},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":137.12500035762787,"p90":139.0880048274994,"p95":140.00999927520752,"p99":142.29300618171692},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":60.57000160217285,"p90":64.29599970579147,"p95":64.65700268745422,"p99":65.37800282239914},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":56.56500160694122,"p90":59.81000140309334,"p95":60.29000133275986,"p99":60.89099869132042},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129004061222076,"p90":6.169997155666351,"p95":6.169997155666351,"p99":6.169997155666351},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.04700094461441,"p90":9.65399295091629,"p95":10.736003518104553,"p99":13.61999660730362},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.4409998655319214,"p90":0.5209967494010925,"p95":0.5209967494010925,"p99":0.5209967494010925},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":64.13599848747253,"p90":65.01699984073639,"p95":65.2180016040802,"p99":65.699003636837},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":56.1240017414093,"p90":57.44599923491478,"p95":58.327000588178635,"p99":60.690999031066895},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":120.54000049829483,"p90":121.6220036149025,"p95":121.9019964337349,"p99":122.58300185203552},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":16.58499240875244,"p90":18.589995801448822,"p95":19.54999566078186,"p99":21.63299173116684},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":9418752,"scale_bytes":294336,"total_logical_bytes":9713088},"roundtrip":{"activation_data_bytes":28256256,"scale_bytes":294336,"total_logical_bytes":28550592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":9418752,"scale_bytes":294336,"total_logical_bytes":9713088},"roundtrip":{"activation_data_bytes":28256256,"scale_bytes":294336,"total_logical_bytes":28550592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":89,"mean":82.125,"min":72,"total":1314},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":1866909.7490052218,"p90":1840561.3073355819,"p95":1828440.8351206353,"p99":1799104.5861598586},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":90.89600294828415,"p90":93.41999888420105,"p95":94.46100145578384,"p99":96.62400186061859},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":82.72399753332138,"p90":84.80700105428696,"p95":86.32899820804596,"p99":88.69300037622452},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":173.62000048160553,"p90":178.226999938488,"p95":180.7899996638298,"p99":185.3170022368431},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":141.49199426174164,"p90":143.25399696826935,"p95":144.21600103378296,"p99":147.54100143909454},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":147.74100482463837,"p90":151.58699452877045,"p95":152.86900103092194,"p99":155.35199642181396},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":72.10800051689148,"p90":76.07399672269821,"p95":76.79499685764313,"p99":77.51599699258804},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":61.813000589609146,"p90":65.57799875736237,"p95":66.05900079011917,"p99":67.0199990272522},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1690062284469604,"p90":6.209000945091248,"p95":6.209000945091248,"p99":6.209000945091248},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.609005689620972,"p90":9.77499783039093,"p95":11.016994714736938,"p99":13.98099958896637},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.7610023021697998,"p90":0.9609907865524292,"p95":0.9609907865524292,"p99":0.9609907865524292},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":70.58600336313248,"p90":72.8290006518364,"p95":73.38999956846237,"p99":75.07199794054031},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":57.08499997854233,"p90":58.687999844551086,"p95":59.44899842143059,"p99":61.33100017905235},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":138.88800144195557,"p90":140.21000266075134,"p95":140.65000414848328,"p99":141.3709968328476},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.693009614944458,"p90":12.659013271331787,"p95":13.741999864578247,"p99":16.62600040435791},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":18808832,"scale_bytes":587776,"total_logical_bytes":19396608},"roundtrip":{"activation_data_bytes":56426496,"scale_bytes":587776,"total_logical_bytes":57014272},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":18808832,"scale_bytes":587776,"total_logical_bytes":19396608},"roundtrip":{"activation_data_bytes":56426496,"scale_bytes":587776,"total_logical_bytes":57014272},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":178,"mean":164.0,"min":148,"total":2624},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":3465524.0135108056,"p90":3377598.4647734733,"p95":3349272.8842810583,"p99":3295741.360219217},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":121.1400032043457,"p90":124.22599643468857,"p95":125.30699372291565,"p99":127.11000442504883},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":104.4360026717186,"p90":106.79999738931656,"p95":108.12199860811234,"p99":110.20500212907791},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":225.5760058760643,"p90":231.02599382400513,"p95":233.42899233102798,"p99":237.31500655412674},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":195.53300738334656,"p90":196.81400060653687,"p95":197.17499613761902,"p99":198.33600521087646},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":200.29999315738678,"p90":202.78300344944,"p95":203.94399762153625,"p99":205.50799369812012},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":97.74699807167053,"p90":105.1969975233078,"p95":106.51899874210358,"p99":108.00100117921829},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":78.43700051307678,"p90":86.08900010585785,"p95":87.41000294685364,"p99":89.05299752950668},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1690062284469604,"p90":6.209000945091248,"p95":6.209000945091248,"p99":6.209000945091248},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129011511802673,"p90":9.25399363040924,"p95":10.335013270378113,"p99":14.020010828971863},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.7200092077255249,"p90":0.9209960699081421,"p95":0.9209960699081421,"p99":0.9209960699081421},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":99.90999847650528,"p90":102.87400335073471,"p95":103.87600213289261,"p99":106.07899725437164},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":79.8799991607666,"p90":81.28099888563156,"p95":81.80300146341324,"p99":84.28599685430527},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":191.52599573135376,"p90":192.96899437904358,"p95":193.449005484581,"p99":194.25100088119507},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.812010288238525,"p90":11.497989296913147,"p95":12.337997555732727,"p99":14.301985502243042},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":37675008,"scale_bytes":1177344,"total_logical_bytes":38852352},"roundtrip":{"activation_data_bytes":113025024,"scale_bytes":1177344,"total_logical_bytes":114202368},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":37675008,"scale_bytes":1177344,"total_logical_bytes":38852352},"roundtrip":{"activation_data_bytes":113025024,"scale_bytes":1177344,"total_logical_bytes":114202368},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":363,"mean":328.5,"min":292,"total":5256},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":5112331.677392453,"p90":5049732.879882679,"p95":5020986.21161806,"p99":4982774.5460071955},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":169.6929931640625,"p90":180.74999749660492,"p95":183.51399898529053,"p99":187.07899749279022},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":127.55100429058075,"p90":137.8059983253479,"p95":138.5679990053177,"p99":140.53000509738922},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":297.24399745464325,"p90":318.5559958219528,"p95":322.0819979906082,"p99":327.60900259017944},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":267.2800123691559,"p90":275.21300315856934,"p95":279.6579897403717,"p99":285.02699732780457},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":278.3370018005371,"p90":282.0630073547363,"p95":282.9039990901947,"p99":284.4260036945343},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":138.5270059108734,"p90":152.34799683094025,"p95":154.31100130081177,"p99":155.7929962873459},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":93.69999915361404,"p90":109.68399792909622,"p95":111.60700023174286,"p99":116.37400090694427},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.2499940395355225,"p90":6.328999996185303,"p95":6.328999996185303,"p99":6.328999996185303},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.497978448867798,"p90":21.311014890670776,"p95":22.956013679504395,"p99":26.360005140304565},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.083003520965576,"p90":4.527002573013306,"p95":4.527002573013306,"p99":4.527002573013306},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":146.54000103473663,"p90":153.83000671863556,"p95":154.71099317073822,"p99":156.19300305843353},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":107.76100307703018,"p90":111.84699833393097,"p95":112.44799941778183,"p99":113.88999968767166},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":260.4300081729889,"p90":269.7640061378479,"p95":271.16501331329346,"p99":272.96900749206543},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.219999313354492,"p90":21.634012460708618,"p95":23.196011781692505,"p99":25.72000026702881},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00495867757126689,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":75787264,"scale_bytes":2368352,"total_logical_bytes":78155616},"roundtrip":{"activation_data_bytes":227361792,"scale_bytes":2368352,"total_logical_bytes":229730144},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":75787264,"scale_bytes":2368352,"total_logical_bytes":78155616},"roundtrip":{"activation_data_bytes":227361792,"scale_bytes":2368352,"total_logical_bytes":229730144},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":692,"mean":660.8125,"min":603,"total":10573},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":7357986.85317321,"p90":7260789.066977275,"p95":7239204.841876633,"p99":7200466.8117458625},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":253.73899936676025,"p90":263.87399435043335,"p95":268.0410146713257,"p99":272.8089988231659},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":195.37299871444702,"p90":205.06800711154938,"p95":209.073007106781,"p99":212.95900642871857},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":449.1119980812073,"p90":468.9420014619827,"p95":477.1140217781067,"p99":485.76800525188446},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":438.9359951019287,"p90":445.06698846817017,"p95":446.42698764801025,"p99":449.4740068912506},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":432.04599618911743,"p90":441.9020116329193,"p95":443.34399700164795,"p99":445.7879960536957},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":181.87199532985687,"p90":200.05899667739868,"p95":210.79500019550323,"p99":229.98300194740295},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":135.8029991388321,"p90":140.80999791622162,"p95":150.66500008106232,"p99":168.93300414085388},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.050991058349609,"p90":7.170975208282471,"p95":7.170975208282471,"p99":7.170975208282471},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":16.907989978790283,"p90":29.082000255584717,"p95":33.329010009765625,"p99":40.82000255584717},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.4460127353668213,"p90":5.128026008605957,"p95":5.128026008605957,"p99":5.128026008605957},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":232.66899585723877,"p90":238.47800493240356,"p95":239.35799300670624,"p99":240.75999855995178},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":174.62100088596344,"p90":185.5980008840561,"p95":187.0799958705902,"p99":188.52199614048004},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":413.77899050712585,"p90":420.30999064445496,"p95":423.03401231765747,"p99":428.9189875125885},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.909990787506104,"p90":32.60999917984009,"p95":37.93799877166748,"p99":45.18899321556091},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00495867757126689,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":151825408,"scale_bytes":4744544,"total_logical_bytes":156569952},"roundtrip":{"activation_data_bytes":455476224,"scale_bytes":4744544,"total_logical_bytes":460220768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":151825408,"scale_bytes":4744544,"total_logical_bytes":156569952},"roundtrip":{"activation_data_bytes":455476224,"scale_bytes":4744544,"total_logical_bytes":460220768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":1373,"mean":1323.8125,"min":1247,"total":21181},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":9480472.070402147,"p90":9269023.204634063,"p95":9238875.518110996,"p99":9188224.080189526},"tokens_per_rank":256}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"InterNodeV1LL","block_num":32,"rdma_block_num":16,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":56,"scale_type_size":4,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"inter-node-v1ll-vllm-cudagraph-spin-ar","rank_alignment":{"method":"in-graph-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":true,"cuda_graph_supported":true,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9938,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2080,"p50":2101,"max":2103,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.2},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2080,"p50":2102,"max":2104,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.2},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2079,"p50":2100,"max":2102,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.2},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2084,"p50":2104,"max":2106,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.2},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2096,"max":2098,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.2},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2082,"p50":2105,"max":2107,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.2},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2109,"max":2110,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.2},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2076,"p50":2097,"max":2098,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.2},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2103,"max":2105,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.2},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2104,"max":2106,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.1},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2086,"p50":2107,"max":2108,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.1},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2081,"p50":2101,"max":2102,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":54,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.1},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2082,"p50":2103,"max":2105,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.1},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2081,"p50":2102,"max":2103,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.1},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2081,"p50":2103,"max":2105,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.1},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2082,"p50":2102,"max":2104,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":772.1}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-bf16.json b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-bf16.json new file mode 100644 index 0000000000..c7ac237eaf --- /dev/null +++ b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T09:25:36.775538+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":2,"phase":"decode","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":133.2789957523346,"p90":136.3230049610138,"p95":137.2860074043274,"p99":139.20800387859344},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":82.88399875164032,"p90":84.52700078487396,"p95":85.88799834251404,"p99":88.13200145959854},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":216.16299450397491,"p90":220.85000574588776,"p95":223.17400574684143,"p99":227.340005338192},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":193.9699947834015,"p90":195.09300589561462,"p95":195.41199505329132,"p99":196.492999792099},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":205.5490016937256,"p90":207.5909972190857,"p95":208.23200047016144,"p99":210.51600575447083},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":121.22199684381485,"p90":122.90400266647339,"p95":123.58500063419342,"p99":124.70699846744537},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":65.33800065517426,"p90":68.58199834823608,"p95":69.06399875879288,"p99":69.70400363206863},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129011511802673,"p90":6.170004606246948,"p95":6.170004606246948,"p99":6.170004606246948},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.365997552871704,"p90":8.492007851600647,"p95":9.73498821258545,"p99":11.618003249168396},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.5210041999816895,"p90":0.6410032510757446,"p95":0.6410032510757446,"p99":0.6410032510757446},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":120.66099792718887,"p90":122.3829984664917,"p95":123.02400171756744,"p99":124.46700036525726},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":63.17500025033951,"p90":64.33600187301636,"p95":64.57599997520447,"p99":65.37800282239914},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":187.04000115394592,"p90":188.08099627494812,"p95":188.3620023727417,"p99":189.28299844264984},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.62899959087372,"p90":20.99299430847168,"p95":22.193998098373413,"p99":24.4780033826828},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":9,"mean":5.125,"min":2,"total":82},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":77840.31967151316,"p90":77074.63336241913,"p95":76837.37352507794,"p99":76003.72210491744},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":139.00800049304962,"p90":142.13299751281738,"p95":143.25499534606934,"p99":145.53800225257874},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":84.4469964504242,"p90":86.81000024080276,"p95":88.29300105571747,"p99":91.89800173044205},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":223.45499694347382,"p90":228.94299775362015,"p95":231.5479964017868,"p99":237.43600398302078},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":201.34100317955017,"p90":203.06499302387238,"p95":203.9860039949417,"p99":206.18900656700134},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":212.67899870872498,"p90":214.68199789524078,"p95":215.60299396514893,"p99":217.84700453281403},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":126.30900740623474,"p90":128.11200320720673,"p95":128.55200469493866,"p99":129.51399385929108},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":70.3049972653389,"p90":73.35100322961807,"p95":74.06999915838242,"p99":75.3529965877533},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1690062284469604,"p90":6.209000945091248,"p95":6.209000945091248,"p99":6.209000945091248},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.330998778343201,"p90":10.776996612548828,"p95":12.097999453544617,"p99":14.380007982254028},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.9219944477081299,"p90":0.5209892988204956,"p95":0.5209892988204956,"p99":0.5209892988204956},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":125.26699900627136,"p90":126.62899494171143,"p95":127.06999480724335,"p99":128.11100482940674},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":64.69699740409851,"p90":65.85799902677536,"p95":66.37900322675705,"p99":68.70199739933014},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":189.04300034046173,"p90":190.60499966144562,"p95":191.24600291252136,"p99":194.28999722003937},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":23.67599308490753,"p90":26.0000079870224,"p95":27.081996202468872,"p99":29.325008392333984},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":15,"mean":10.375,"min":7,"total":166},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":150461.49452596248,"p90":149057.6774658822,"p95":148420.94449380712,"p99":146892.0817553857},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":141.97200536727905,"p90":144.97600495815277,"p95":145.81699669361115,"p99":148.10200035572052},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":97.22600132226944,"p90":100.31100362539291,"p95":101.23199969530106,"p99":104.03600335121155},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":239.1980066895485,"p90":245.28700858354568,"p95":247.0489963889122,"p99":252.13800370693207},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":210.55500209331512,"p90":213.359996676445,"p95":214.48099613189697,"p99":216.08300507068634},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":218.36699545383453,"p90":220.89199721813202,"p95":221.81199491024017,"p99":225.17700493335724},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":127.1899938583374,"p90":129.79400157928467,"p95":130.31500577926636,"p99":131.4769983291626},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":74.67100024223328,"p90":77.27599889039993,"p95":78.07700335979462,"p99":78.79800349473953},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.168991327285767,"p90":6.250008940696716,"p95":6.250008940696716,"p99":6.250008940696716},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.969002842903137,"p90":9.454011917114258,"p95":10.736003518104553,"p99":13.539999723434448},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.881001353263855,"p90":-0.5210041999816895,"p95":-0.5210041999816895,"p99":-0.5210041999816895},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":125.02799928188324,"p90":128.63199412822723,"p95":129.31300699710846,"p99":130.2350014448166},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":69.98399645090103,"p90":71.30700349807739,"p95":71.70700281858444,"p99":72.86900281906128},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":203.74499261379242,"p90":205.8670073747635,"p95":206.34900033473969,"p99":207.43000507354736},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.862999320030212,"p90":18.748998641967773,"p95":19.83100175857544,"p99":22.234007716178894},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":26,"mean":20.5625,"min":15,"total":329},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":293084.58389963233,"p90":289734.3534668649,"p95":288532.63785801415,"p99":284220.8511430431},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":148.30200374126434,"p90":150.90599656105042,"p95":151.82699263095856,"p99":154.14999425411224},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":101.67200118303299,"p90":103.79599779844284,"p95":104.8780009150505,"p99":107.92099684476852},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":249.97400492429733,"p90":254.70199435949326,"p95":256.70499354600906,"p99":262.07099109888077},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":224.57599639892578,"p90":228.90299558639526,"p95":229.74400222301483,"p99":231.90799355506897},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":235.95400154590607,"p90":238.27700316905975,"p95":239.03900384902954,"p99":241.08099937438965},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":137.44600117206573,"p90":138.8079971075058,"p95":139.44800198078156,"p99":140.61099290847778},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":80.04000037908554,"p90":83.12399685382843,"p95":84.08600091934204,"p99":86.20999753475189},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1690062284469604,"p90":6.248995661735535,"p95":6.248995661735535,"p99":6.248995661735535},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.247997283935547,"p90":10.496988892555237,"p95":12.137994170188904,"p99":17.504990100860596},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.0409951210021973,"p90":0.7220059633255005,"p95":0.7220059633255005,"p99":0.7220059633255005},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":138.00600171089172,"p90":139.36899602413177,"p95":139.65000212192535,"p99":140.36999642848969},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":77.51599699258804,"p90":79.31800186634064,"p95":79.83999699354172,"p99":81.12099766731262},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":217.32500195503235,"p90":219.2479968070984,"p95":219.9680060148239,"p99":221.13099694252014},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.710002303123474,"p90":21.393999457359314,"p95":22.31399714946747,"p99":24.115994572639465},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":49,"mean":40.5625,"min":36,"total":649},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":542478.6151596456,"p90":537189.9020787281,"p95":535477.4657647138,"p99":530941.8839815778},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":169.05300319194794,"p90":172.25800454616547,"p95":173.37900400161743,"p99":175.5020022392273},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":119.13900077342987,"p90":121.18099629878998,"p95":122.22299724817276,"p99":125.06699562072754},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":288.1920039653778,"p90":293.43900084495544,"p95":295.6020012497902,"p99":300.56899785995483},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":257.26500153541565,"p90":259.18900966644287,"p95":260.0710093975067,"p99":263.27499747276306},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":259.95001196861267,"p90":262.35300302505493,"p95":263.3950114250183,"p99":265.6779885292053},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":152.10700035095215,"p90":153.78999710083008,"p95":154.14999425411224,"p99":155.15199303627014},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":97.18599915504456,"p90":98.90799969434738,"p95":99.46899861097336,"p99":100.75200349092484},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.170004606246948,"p90":6.208986043930054,"p95":6.208986043930054,"p99":6.208986043930054},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.68699836730957,"p90":9.172976016998291,"p95":10.333985090255737,"p99":14.46099579334259},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.6809830665588379,"p90":0.7610023021697998,"p95":0.7610023021697998,"p99":0.7610023021697998},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":144.8580026626587,"p90":146.3789939880371,"p95":146.8999981880188,"p99":148.50200712680817},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":87.36900240182877,"p90":88.81200104951859,"p95":89.4939973950386,"p99":91.37500077486038},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":247.44999408721924,"p90":249.01199340820312,"p95":249.37300384044647,"p99":250.25400519371033},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":12.459993362426758,"p90":15.464991331100464,"p95":16.74601435661316,"p99":20.071014761924744},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":89,"mean":82.125,"min":72,"total":1314},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":984804.724805746,"p90":975784.5233261988,"p95":971924.254050941,"p99":963572.4864420169},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":181.23100697994232,"p90":183.95499885082245,"p95":185.03600358963013,"p99":187.56000697612762},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":136.40399277210236,"p90":139.5689994096756,"p95":140.53000509738922,"p99":142.81399548053741},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":317.6349997520447,"p90":323.52399826049805,"p95":325.56600868701935,"p99":330.37400245666504},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":281.46201372146606,"p90":283.34400057792664,"p95":283.9469909667969,"p99":285.7489883899689},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":285.3870093822479,"p90":287.4299883842468,"p95":288.39099407196045,"p99":290.7550036907196},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":154.7510027885437,"p90":160.67999601364136,"p95":161.88299655914307,"p99":162.88499534130096},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":110.60599982738495,"p90":117.17599630355835,"p95":118.41700226068497,"p99":120.02000212669373},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.209015846252441,"p95":6.209015846252441,"p99":6.209015846252441},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.650000810623169,"p90":10.816991329193115,"p95":11.896997690200806,"p99":13.418972492218018},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.7219910621643066,"p90":2.083003520965576,"p95":2.083003520965576,"p99":2.083003520965576},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":148.46199750900269,"p90":150.78599750995636,"p95":151.58799290657043,"p99":153.7500023841858},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":106.64000362157822,"p90":108.40299725532532,"p95":109.2429980635643,"p99":111.7670014500618},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":276.3749957084656,"p90":277.93601155281067,"p95":278.45799922943115,"p99":280.01898527145386},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.012997150421143,"p90":11.335998773574829,"p95":12.136995792388916,"p99":14.099985361099243},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.004048583097755909,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":178,"mean":164.0,"min":148,"total":2624},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":1794055.031125212,"p90":1781303.3458274363,"p95":1775367.5063523091,"p99":1760932.7217103443},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":199.69899952411652,"p90":202.90300250053406,"p95":203.90500128269196,"p99":206.3080072402954},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":157.11599588394165,"p90":159.60000455379486,"p95":160.8009934425354,"p99":163.68499398231506},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":356.81499540805817,"p90":362.5030070543289,"p95":364.70599472522736,"p99":369.9930012226105},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":319.47800517082214,"p90":321.44099473953247,"p95":322.162002325058,"p99":323.76399636268616},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":323.9240050315857,"p90":326.28801465034485,"p95":327.2489905357361,"p99":331.4549922943115},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":165.68799316883087,"p90":177.2249937057495,"p95":179.30899560451508,"p99":181.551992893219},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":122.46300280094147,"p90":134.72099602222443,"p95":137.00400292873383,"p99":140.0499939918518},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.209015846252441,"p90":6.36899471282959,"p95":6.36899471282959,"p99":6.36899471282959},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.45098090171814,"p90":9.974002838134766,"p95":11.458009481430054,"p99":15.020996332168579},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.7210075855255127,"p90":1.4030039310455322,"p95":1.4030039310455322,"p99":1.4030039310455322},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":167.29100048542023,"p90":169.81400549411774,"p95":170.73599994182587,"p99":172.6589947938919},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":126.74899399280548,"p90":128.55200469493866,"p95":129.27299737930298,"p99":131.03599846363068},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":313.7899935245514,"p90":315.39300084114075,"p95":315.87299704551697,"p99":316.7549967765808},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.213971138000488,"p90":13.218998908996582,"p95":14.380007982254028,"p99":17.865002155303955},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.006103515625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":363,"mean":328.5,"min":292,"total":5256},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":3161235.3023980125,"p90":3138331.639601699,"p95":3129115.840276909,"p99":3089408.890516126},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":236.99399828910828,"p90":242.28200316429138,"p95":243.56399476528168,"p99":246.5679943561554},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":196.13400101661682,"p90":199.77900385856628,"p95":200.7409930229187,"p99":202.70399749279022},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":433.1279993057251,"p90":442.06100702285767,"p95":444.3049877882004,"p99":449.2719918489456},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":419.7080135345459,"p90":437.09298968315125,"p95":442.02399253845215,"p99":447.0309913158417},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":402.92298793792725,"p90":406.4069986343384,"p95":407.2889983654022,"p99":409.0920090675354},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":166.37000441551208,"p90":172.85899817943573,"p95":174.02200400829315,"p99":175.86399614810944},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":123.66499751806259,"p90":126.38899683952332,"p95":127.99100577831268,"p99":132.7580064535141},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.171005010604858,"p90":7.850974798202515,"p95":7.850974798202515,"p99":7.850974798202515},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":21.272003650665283,"p90":36.41700744628906,"p95":41.141003370285034,"p99":49.914002418518066},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.6430020332336426,"p90":5.809009075164795,"p95":5.809009075164795,"p99":5.809009075164795},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":207.87200331687927,"p90":212.27799355983734,"p95":213.44000101089478,"p99":215.4030054807663},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":168.3720052242279,"p90":171.77699506282806,"p95":172.53799736499786,"p99":173.73999953269958},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":391.6279971599579,"p90":394.54901218414307,"p95":395.1930105686188,"p99":396.35300636291504},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11.215001344680786,"p90":15.861988067626953,"p95":17.023980617523193,"p99":19.306987524032593},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.006103515625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":692,"mean":660.8125,"min":603,"total":10573},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":5082857.174472028,"p90":5039283.296995268,"p95":5028370.538412193,"p99":5006208.761369142},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":311.8259906768799,"p90":331.1769962310791,"p95":340.70900082588196,"p99":351.08500719070435},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":311.706006526947,"p90":316.3129985332489,"p95":317.3559904098511,"p99":319.68000531196594},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":623.5319972038269,"p90":647.489994764328,"p95":658.064991235733,"p99":670.7650125026703},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":604.6680212020874,"p90":619.0890073776245,"p95":623.8960027694702,"p99":632.5089931488037},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":604.3050289154053,"p90":608.7520122528076,"p95":609.8350286483765,"p99":612.2390031814575},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":232.70900547504425,"p90":285.9489917755127,"p95":293.92099380493164,"p99":299.32698607444763},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":228.50200533866882,"p90":280.73999285697937,"p95":291.7959988117218,"p99":297.4439859390259},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.050991058349609,"p90":8.373022079467773,"p95":8.373022079467773,"p99":8.373022079467773},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":19.109010696411133,"p90":37.69904375076294,"p95":44.82901096343994,"p99":58.28297138214111},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":15.263020992279053,"p90":19.589006900787354,"p95":19.589006900787354,"p99":19.589006900787354},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":293.5999929904938,"p90":299.00801181793213,"p95":299.6479868888855,"p99":300.6089925765991},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":289.4330024719238,"p90":295.16100883483887,"p95":296.20298743247986,"p99":298.4060049057007},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":590.1619791984558,"p90":595.209002494812,"p95":596.2110161781311,"p99":597.8530049324036},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.943016529083252,"p90":36.73797845840454,"p95":44.310033321380615,"p99":52.08003520965576},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.006103515625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":1373,"mean":1323.8125,"min":1247,"total":21181},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":6778033.946451547,"p90":6728519.853005396,"p95":6716570.56020261,"p99":6690197.747473488},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":486.32800579071045,"p90":513.3669972419739,"p95":521.5830206871033,"p99":534.9590182304382},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":527.3889899253845,"p90":541.051983833313,"p95":547.2210049629211,"p99":554.5110106468201},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1013.716995716095,"p90":1054.4189810752869,"p95":1068.8040256500244,"p99":1089.4700288772583},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":977.1810173988342,"p90":989.799976348877,"p95":992.4830198287964,"p99":1001.30295753479},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":986.3579869270325,"p90":996.5729713439941,"p95":1000.419020652771,"p99":1005.6630373001099},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":419.06899213790894,"p90":455.24299144744873,"p95":459.52698588371277,"p99":464.4539952278137},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":461.57100796699524,"p90":504.2340159416199,"p95":513.0500197410583,"p99":521.261990070343},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.170975208282471,"p90":7.4520111083984375,"p95":7.4520111083984375,"p99":7.4520111083984375},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":22.429049015045166,"p90":52.15299129486084,"p95":61.097025871276855,"p99":83.91892910003662},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.9659738540649414,"p90":4.3669939041137695,"p95":4.3669939041137695,"p99":4.3669939041137695},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":447.952002286911,"p90":461.6129994392395,"p95":463.2929861545563,"p99":464.77699279785156},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":499.54599142074585,"p90":514.369010925293,"p95":516.772985458374,"p99":521.21901512146},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":958.5130214691162,"p90":978.547990322113,"p95":981.2300205230713,"p99":986.5170121192932},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":23.160040378570557,"p90":48.3890175819397,"p95":53.87604236602783,"p99":64.85295295715332},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"roundtrip":{"activation_data_bytes":1215807488,"scale_bytes":0,"total_logical_bytes":1215807488},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"roundtrip":{"activation_data_bytes":1215807488,"scale_bytes":0,"total_logical_bytes":1215807488},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":42404,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":2702,"mean":2650.25,"min":2578,"total":42404},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.017352504655718803,"expert_assignments_per_rank":[3108,3051,3103,2954,3171,3067,3121,2998,3045,3065,3095,3023,3150,3071,3090,3040],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,3,103,1301,3825,2960,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.17626953125,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.06268276274204254,"same_node_fraction":0.49721723794937134,"same_scaleup_domain_fraction":0.49721723794937134,"cross_node_fraction":0.5027827620506287,"cross_domain_fraction":0.5027827620506287,"gpus_per_node":8,"scale_up_domain":8,"copies":42404},"payload_copies_per_rank":[2682,2637,2676,2578,2702,2659,2687,2602,2635,2661,2672,2599,2691,2625,2662,2636],"payload_rank_cv":0.013201921246945858,"routed_copies":42404},"token_rate_at_latency_percentile":{"p50":8305301.0251601655,"p90":8220170.760754368,"p95":8188568.820547555,"p99":8145869.636407195},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"InterNodeV1","block_num":32,"rdma_block_num":16,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":0,"scale_type_size":0,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"inter-node-v1-vllm-cudagraph-spin-ar","rank_alignment":{"method":"in-graph-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":true,"cuda_graph_supported":true,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9938,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2081,"p50":2100,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2102,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2100,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2085,"p50":2105,"max":2106,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2079,"p50":2096,"max":2097,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2089,"p50":2106,"max":2107,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2109,"max":2111,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2081,"p50":2097,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2103,"max":2106,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2104,"max":2105,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2106,"max":2108,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2082,"p50":2101,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":54,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2086,"p50":2105,"max":2107,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2102,"max":2103,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2085,"p50":2103,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2084,"p50":2102,"max":2103,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":913.7}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-fp8.json b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-fp8.json new file mode 100644 index 0000000000..4a9bc611f2 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T13:02:04.052635+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":2,"phase":"decode","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":135.002002120018,"p90":137.88600265979767,"p95":138.96700739860535,"p99":141.0910040140152},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":90.45600146055222,"p90":92.2980010509491,"p95":93.62000226974487,"p99":96.50500118732452},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":225.45800358057022,"p90":230.18400371074677,"p95":232.58700966835022,"p99":237.59600520133972},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":200.10000467300415,"p90":201.62199437618256,"p95":202.0619958639145,"p99":206.3090056180954},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":210.23499965667725,"p90":212.43900060653687,"p95":213.47999572753906,"p99":216.32400155067444},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":120.5810010433197,"p90":122.54399806261063,"p95":123.34500253200531,"p99":125.74900686740875},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":72.3079964518547,"p90":75.23299753665924,"p95":75.63299685716629,"p99":76.39399915933609},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":6.1690062284469604,"p95":6.1690062284469604,"p99":6.1690062284469604},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.291002035140991,"p90":9.773999452590942,"p95":11.097997426986694,"p99":13.541996479034424},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.48001110553741455,"p90":0.7210075855255127,"p95":0.7210075855255127,"p99":0.7210075855255127},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":119.37899887561798,"p90":121.02100253105164,"p95":121.6220036149025,"p99":122.74400144815445},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":68.2620033621788,"p90":69.34399902820587,"p95":69.74399834871292,"p99":70.90599834918976},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":194.08999383449554,"p90":195.4919993877411,"p95":195.89300453662872,"p99":197.85499572753906},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":16.145989298820496,"p90":19.308999180793762,"p95":20.750999450683594,"p99":23.674994707107544},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":587776,"scale_bytes":18368,"total_logical_bytes":606144},"roundtrip":{"activation_data_bytes":1763328,"scale_bytes":18368,"total_logical_bytes":1781696},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":587776,"scale_bytes":18368,"total_logical_bytes":606144},"roundtrip":{"activation_data_bytes":1763328,"scale_bytes":18368,"total_logical_bytes":1781696},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":9,"mean":5.125,"min":2,"total":82},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":76105.31084799717,"p90":75315.73747907978,"p95":74948.47442483806,"p99":73963.12884981447},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":138.046994805336,"p90":140.89100062847137,"p95":142.05299317836761,"p99":144.25699412822723},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":92.25799888372421,"p90":94.50100362300873,"p95":95.90400010347366,"p99":99.62999820709229},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":230.3049936890602,"p90":235.3920042514801,"p95":237.95699328184128,"p99":243.88699233531952},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":210.0750058889389,"p90":212.03799545764923,"p95":212.8790020942688,"p99":216.9249951839447},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":218.40700507164001,"p90":220.4110026359558,"p95":221.29200398921967,"p99":223.7350046634674},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":125.78800320625305,"p90":127.23000347614288,"p95":127.63099372386932,"p99":128.99300456047058},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":77.47600227594376,"p90":79.96000349521637,"p95":81.12099766731262,"p99":82.24300295114517},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.168991327285767,"p90":6.209000945091248,"p95":6.209000945091248,"p99":6.209000945091248},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.9279948472976685,"p90":10.416001081466675,"p95":11.77901029586792,"p99":14.100998640060425},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.6009936332702637,"p90":0.8819997310638428,"p95":0.8819997310638428,"p99":0.8819997310638428},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":126.14800035953522,"p90":127.469003200531,"p95":127.79200077056885,"p99":128.79200279712677},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":70.98600268363953,"p90":72.22799956798553,"p95":72.94899970293045,"p99":75.63299685716629},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":196.89400494098663,"p90":198.53700697422028,"p95":199.09800589084625,"p99":201.54200494289398},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":21.591991186141968,"p90":24.03700351715088,"p95":24.99699592590332,"p99":27.761995792388916},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":1189888,"scale_bytes":37184,"total_logical_bytes":1227072},"roundtrip":{"activation_data_bytes":3569664,"scale_bytes":37184,"total_logical_bytes":3606848},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":1189888,"scale_bytes":37184,"total_logical_bytes":1227072},"roundtrip":{"activation_data_bytes":3569664,"scale_bytes":37184,"total_logical_bytes":3606848},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":15,"mean":10.375,"min":7,"total":166},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":146515.4471098747,"p90":145183.31488584145,"p95":144605.31525377164,"p99":143026.34515387088},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":142.41300523281097,"p90":145.73800563812256,"p95":146.81999385356903,"p99":149.10300076007843},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":104.39600050449371,"p90":107.2010025382042,"p95":108.16200077533722,"p99":111.56699806451797},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":246.8090057373047,"p90":252.93900817632675,"p95":254.98199462890625,"p99":260.6699988245964},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":218.12599897384644,"p90":220.77099978923798,"p95":221.17100656032562,"p99":224.17600452899933},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":225.65899789333344,"p90":228.18200290203094,"p95":229.1029989719391,"p99":231.14599287509918},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":128.1909942626953,"p90":130.11400401592255,"p95":130.39499521255493,"p99":131.11600279808044},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":83.1649973988533,"p90":85.56800335645676,"p95":85.84900200366974,"p99":86.93099766969681},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.170004606246948,"p90":6.209000945091248,"p95":6.209000945091248,"p99":6.209000945091248},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.970003247261047,"p90":11.0970139503479,"p95":12.019991874694824,"p99":14.261007308959961},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.9620040655136108,"p90":1.8430054187774658,"p95":1.8430054187774658,"p99":1.8430054187774658},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":125.30800700187683,"p90":128.9519965648651,"p95":129.7139972448349,"p99":130.83599507808685},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":76.95499807596207,"p90":78.71700078248978,"p95":79.07900214195251,"p99":80.11999726295471},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":210.67500114440918,"p90":212.99800276756287,"p95":213.6790007352829,"p99":215.00200033187866},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":15.143990516662598,"p90":18.708989024162292,"p95":19.59100365638733,"p99":21.872997283935547},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":2358272,"scale_bytes":73696,"total_logical_bytes":2431968},"roundtrip":{"activation_data_bytes":7074816,"scale_bytes":73696,"total_logical_bytes":7148512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":2358272,"scale_bytes":73696,"total_logical_bytes":2431968},"roundtrip":{"activation_data_bytes":7074816,"scale_bytes":73696,"total_logical_bytes":7148512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":26,"mean":20.5625,"min":15,"total":329},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":283613.77386888914,"p90":280477.8605939319,"p95":279350.33712866774,"p99":276881.28703396},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":149.50299263000488,"p90":152.1880030632019,"p95":153.1900018453598,"p99":155.5120050907135},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":108.924001455307,"p90":111.20600253343582,"p95":112.20899969339371,"p99":115.97400158643723},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":258.4269940853119,"p90":263.3940055966377,"p95":265.3990015387535,"p99":271.4860066771507},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":236.47400736808777,"p90":240.63900113105774,"p95":241.88299477100372,"p99":245.56699395179749},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":245.04700303077698,"p90":247.530996799469,"p95":248.21199476718903,"p99":250.25498867034912},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":137.40499317646027,"p90":138.96800577640533,"p95":139.76900279521942,"p99":141.5719985961914},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":88.37299793958664,"p90":91.49699658155441,"p95":92.97899901866913,"p99":96.10400348901749},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.209999322891235,"p90":6.328999996185303,"p95":6.328999996185303,"p99":6.328999996185303},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.369009613990784,"p90":10.655000805854797,"p95":11.897996068000793,"p99":15.141993761062622},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.7610023021697998,"p90":2.3639947175979614,"p95":2.3639947175979614,"p99":2.3639947175979614},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":138.046994805336,"p90":139.44900035858154,"p95":139.8490071296692,"p99":140.5699998140335},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":86.16899698972702,"p90":87.97100186347961,"p95":88.45199644565582,"p99":89.61399644613266},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":225.01599788665771,"p90":226.97900235652924,"p95":227.619007229805,"p99":229.42300140857697},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":20.071998238563538,"p90":22.634997963905334,"p95":23.556992411613464,"p99":24.998992681503296},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":4652032,"scale_bytes":145376,"total_logical_bytes":4797408},"roundtrip":{"activation_data_bytes":13956096,"scale_bytes":145376,"total_logical_bytes":14101472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":4652032,"scale_bytes":145376,"total_logical_bytes":4797408},"roundtrip":{"activation_data_bytes":13956096,"scale_bytes":145376,"total_logical_bytes":14101472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":49,"mean":40.5625,"min":36,"total":649},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":522348.7674481931,"p90":517106.95490672626,"p95":515688.2128926037,"p99":511478.3152978792},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":166.08799993991852,"p90":169.01299357414246,"p95":170.25500535964966,"p99":172.4579930305481},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":120.54099887609482,"p90":122.6240023970604,"p95":123.78499656915665,"p99":126.34900212287903},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":286.62899881601334,"p90":291.63699597120285,"p95":294.0400019288063,"p99":298.8069951534271},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":261.83098554611206,"p90":263.9560103416443,"p95":264.6760046482086,"p99":266.55900478363037},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":266.1589980125427,"p90":268.40299367904663,"p95":269.4439888000488,"p99":272.0080018043518},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":154.11099791526794,"p90":156.11299872398376,"p95":156.59399330615997,"p99":157.59499371051788},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":98.98900240659714,"p90":101.27100348472595,"p95":101.67299956083298,"p99":103.23499888181686},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.209015846252441,"p95":6.209015846252441,"p99":6.209015846252441},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.168991327285767,"p90":10.255008935928345,"p95":11.777013540267944,"p99":14.582008123397827},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.719994306564331,"p90":0.9220242500305176,"p95":0.9220242500305176,"p99":0.9220242500305176},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":148.74300360679626,"p90":150.9460061788559,"p95":151.70800685882568,"p99":153.5090059041977},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":93.13999861478806,"p90":94.46100145578384,"p95":95.02200037240982,"p99":96.74499928951263},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":248.93200397491455,"p90":250.5750060081482,"p95":250.97501277923584,"p99":251.8559992313385},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.307013273239136,"p90":20.390987396240234,"p95":21.55299484729767,"p99":24.556994438171387},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":9418752,"scale_bytes":294336,"total_logical_bytes":9713088},"roundtrip":{"activation_data_bytes":28256256,"scale_bytes":294336,"total_logical_bytes":28550592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":9418752,"scale_bytes":294336,"total_logical_bytes":9713088},"roundtrip":{"activation_data_bytes":28256256,"scale_bytes":294336,"total_logical_bytes":28550592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":89,"mean":82.125,"min":72,"total":1314},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":961831.0931120052,"p90":953789.6596865907,"p95":950104.6994593543,"p99":941148.7834984137},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":181.63099884986877,"p90":184.31499600410461,"p95":185.357004404068,"p99":187.27999925613403},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":137.16499507427216,"p90":139.36899602413177,"p95":140.21000266075134,"p99":143.29500496387482},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":318.79599392414093,"p90":323.6839920282364,"p95":325.56700706481934,"p99":330.57500422000885},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":279.6579897403717,"p90":281.6210091114044,"p95":282.6629877090454,"p99":284.9859893321991},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":283.30400586128235,"p90":285.50800681114197,"p95":286.7079973220825,"p99":288.4719967842102},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":150.4250019788742,"p90":156.15400671958923,"p95":158.39700400829315,"p99":161.081001162529},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":104.55600172281265,"p90":110.16499996185303,"p95":112.12799698114395,"p99":116.61499738693237},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.169021129608154,"p90":6.208986043930054,"p95":6.208986043930054,"p99":6.208986043930054},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.850004196166992,"p90":10.255992412567139,"p95":11.696010828018188,"p99":13.579994440078735},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.7600188255310059,"p90":-0.6009936332702637,"p95":-0.6009936332702637,"p99":-0.6009936332702637},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":148.6629992723465,"p90":151.22699737548828,"p95":152.10799872875214,"p99":153.58999371528625},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":104.27500307559967,"p90":106.11800104379654,"p95":106.88000172376633,"p99":108.84299874305725},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":274.49101209640503,"p90":275.97400546073914,"p95":276.49399638175964,"p99":277.536004781723},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.93297791481018,"p90":11.256009340286255,"p95":12.178003787994385,"p99":14.420002698898315},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":18808832,"scale_bytes":587776,"total_logical_bytes":19396608},"roundtrip":{"activation_data_bytes":56426496,"scale_bytes":587776,"total_logical_bytes":57014272},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":18808832,"scale_bytes":587776,"total_logical_bytes":19396608},"roundtrip":{"activation_data_bytes":56426496,"scale_bytes":587776,"total_logical_bytes":57014272},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":178,"mean":164.0,"min":148,"total":2624},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":1807245.889246963,"p90":1793294.71603463,"p95":1785789.0424480508,"p99":1774868.9845378601},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":202.5029957294464,"p90":206.22800290584564,"p95":207.50999450683594,"p99":209.75400507450104},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":148.38199317455292,"p90":150.58599412441254,"p95":151.70800685882568,"p99":153.91099452972412},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":350.88498890399933,"p90":356.8139970302582,"p95":359.2180013656616,"p99":363.66499960422516},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":309.7839951515198,"p90":312.42701411247253,"p95":313.14900517463684,"p99":314.5099878311157},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":315.15100598335266,"p90":317.51498579978943,"p95":318.51598620414734,"p99":321.00000977516174},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":161.5220010280609,"p90":166.6489988565445,"p95":168.93300414085388,"p99":172.53899574279785},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":108.28199982643127,"p90":112.96900361776352,"p95":115.17199873924255,"p99":119.69900131225586},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1699748039245605,"p90":6.2499940395355225,"p95":6.2499940395355225,"p99":6.2499940395355225},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.290976285934448,"p90":13.137996196746826,"p95":14.981985092163086,"p99":19.391000270843506},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.1619925498962402,"p90":0.840991735458374,"p95":0.840991735458374,"p99":0.840991735458374},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":164.44599628448486,"p90":167.29100048542023,"p95":168.25200617313385,"p99":170.69600522518158},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":114.17099833488464,"p90":116.09400063753128,"p95":116.77499860525131,"p99":118.73800307512283},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":301.53200030326843,"p90":303.6549985408783,"p95":304.297000169754,"p99":305.6980073451996},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.579994440078735,"p90":16.78299903869629,"p95":17.905980348587036,"p99":20.949989557266235},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":37675008,"scale_bytes":1177344,"total_logical_bytes":38852352},"roundtrip":{"activation_data_bytes":113025024,"scale_bytes":1177344,"total_logical_bytes":114202368},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":37675008,"scale_bytes":1177344,"total_logical_bytes":38852352},"roundtrip":{"activation_data_bytes":113025024,"scale_bytes":1177344,"total_logical_bytes":114202368},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":363,"mean":328.5,"min":292,"total":5256},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":3249236.018793134,"p90":3225044.6303208126,"p95":3214909.280389101,"p99":3190031.055504456},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":236.59400641918182,"p90":240.39900302886963,"p95":241.60200357437134,"p99":243.7639981508255},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":162.96400129795074,"p90":165.3680056333542,"p95":166.32899641990662,"p99":169.33399438858032},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":399.55800771713257,"p90":405.7670086622238,"p95":407.93099999427795,"p99":413.0979925394058},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":377.7239918708801,"p90":384.41699743270874,"p95":387.77899742126465,"p99":390.22400975227356},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":370.1939880847931,"p90":372.4769949913025,"p95":373.3980059623718,"p99":375.2810060977936},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":171.8980073928833,"p90":181.2320053577423,"p95":182.91500210762024,"p99":186.8000030517578},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":100.22900253534317,"p90":107.92099684476852,"p95":109.76400226354599,"p99":113.24899643659592},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.810009479522705,"p90":6.930023431777954,"p95":6.930023431777954,"p99":6.930023431777954},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.862000942230225,"p90":25.638997554779053,"p95":28.883010149002075,"p99":36.29601001739502},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.563983201980591,"p90":3.3259987831115723,"p95":3.3259987831115723,"p99":3.3259987831115723},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":208.51300656795502,"p90":211.95699274539948,"p95":212.95900642871857,"p99":215.12199938297272},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":135.48199832439423,"p90":138.00600171089172,"p95":138.64700496196747,"p99":140.1289999485016},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":359.4990074634552,"p90":361.9009852409363,"p95":362.5420033931732,"p99":363.46399784088135},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.535001754760742,"p90":13.659000396728516,"p95":14.540016651153564,"p99":16.703009605407715},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00495867757126689,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":75787264,"scale_bytes":2368352,"total_logical_bytes":78155616},"roundtrip":{"activation_data_bytes":227361792,"scale_bytes":2368352,"total_logical_bytes":229730144},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":75787264,"scale_bytes":2368352,"total_logical_bytes":78155616},"roundtrip":{"activation_data_bytes":227361792,"scale_bytes":2368352,"total_logical_bytes":229730144},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":692,"mean":660.8125,"min":603,"total":10573},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":5532234.62810775,"p90":5498326.145075944,"p95":5484764.158613053,"p99":5457243.9497945625},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":314.2710030078888,"p90":325.6070017814636,"p95":328.3720016479492,"p99":337.18401193618774},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":241.56199395656586,"p90":250.17499923706055,"p95":251.45700573921204,"p99":255.02300262451172},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":555.8329969644547,"p90":575.7820010185242,"p95":579.8290073871613,"p99":592.2070145606995},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":543.6949729919434,"p90":550.2229928970337,"p95":552.2680282592773,"p99":559.7190260887146},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":540.4099822044373,"p90":546.4199781417847,"p95":547.5010275840759,"p99":550.0649809837341},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":227.059006690979,"p90":238.4369969367981,"p95":245.44699490070343,"p99":274.49101209640503},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":160.92100739479065,"p90":168.1320071220398,"p95":170.77599465847015,"p99":208.15199613571167},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.9699883460998535,"p90":7.252037525177002,"p95":7.252037525177002,"p99":7.252037525177002},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":15.704035758972168,"p90":24.63400363922119,"p95":26.957988739013672,"p99":38.05899620056152},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.0439624786376953,"p90":2.68399715423584,"p95":2.68399715423584,"p99":2.68399715423584},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":291.1149859428406,"p90":298.8480031490326,"p95":299.6079921722412,"p99":300.6109893321991},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":222.9740023612976,"p90":232.14800655841827,"p95":233.14900696277618,"p99":235.03200709819794},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":516.8949961662292,"p90":526.5470147132874,"p95":528.4730195999146,"p99":531.1539769172668},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":21.193981170654297,"p90":36.857008934020996,"p95":39.9019718170166,"p99":46.35098576545715},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00495867757126689,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":151825408,"scale_bytes":4744544,"total_logical_bytes":156569952},"roundtrip":{"activation_data_bytes":455476224,"scale_bytes":4744544,"total_logical_bytes":460220768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":151825408,"scale_bytes":4744544,"total_logical_bytes":156569952},"roundtrip":{"activation_data_bytes":455476224,"scale_bytes":4744544,"total_logical_bytes":460220768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":1373,"mean":1323.8125,"min":1247,"total":21181},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":7579430.682038146,"p90":7496065.597618345,"p95":7481264.49748262,"p99":7446392.9564734865},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":484.16298627853394,"p90":503.23402881622314,"p95":510.1240277290344,"p99":523.622989654541},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":409.13400053977966,"p90":422.1929907798767,"p95":425.2780079841614,"p99":430.5660128593445},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":893.2969868183136,"p90":925.4270195960999,"p95":935.4020357131958,"p99":954.1890025138855},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":856.7209839820862,"p90":867.0600056648254,"p95":869.2229986190796,"p99":874.9880194664001},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":866.6989803314209,"p90":876.0290145874023,"p95":878.6370158195496,"p99":883.8409781455994},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":422.6740002632141,"p90":456.1240077018738,"p95":459.76701378822327,"p99":464.1759991645813},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":351.9279956817627,"p90":386.0990107059479,"p95":390.26400446891785,"p99":399.4390070438385},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.529033184051514,"p90":6.770014762878418,"p95":6.770014762878418,"p99":6.770014762878418},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.301976680755615,"p90":43.148040771484375,"p95":51.16003751754761,"p99":64.860999584198},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.364039421081543,"p90":4.927992820739746,"p95":4.927992820739746,"p99":4.927992820739746},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":450.59600472450256,"p90":461.8909955024719,"p95":463.4949862957001,"p99":465.13500809669495},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":383.976012468338,"p90":393.62800121307373,"p95":396.4729905128479,"p99":401.078999042511},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":848.5519886016846,"p90":859.3680262565613,"p95":861.6920113563538,"p99":866.1350011825562},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.47100257873535,"p90":41.90701246261597,"p95":48.50900173187256,"p99":57.730019092559814},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.005181347019970417,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":303951872,"scale_bytes":9498496,"total_logical_bytes":313450368},"roundtrip":{"activation_data_bytes":911855616,"scale_bytes":9498496,"total_logical_bytes":921354112},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":303951872,"scale_bytes":9498496,"total_logical_bytes":313450368},"roundtrip":{"activation_data_bytes":911855616,"scale_bytes":9498496,"total_logical_bytes":921354112},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":42404,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":2702,"mean":2650.25,"min":2578,"total":42404},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.017352504655718803,"expert_assignments_per_rank":[3108,3051,3103,2954,3171,3067,3121,2998,3045,3065,3095,3023,3150,3071,3090,3040],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,3,103,1301,3825,2960,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.17626953125,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.06268276274204254,"same_node_fraction":0.49721723794937134,"same_scaleup_domain_fraction":0.49721723794937134,"cross_node_fraction":0.5027827620506287,"cross_domain_fraction":0.5027827620506287,"gpus_per_node":8,"scale_up_domain":8,"copies":42404},"payload_copies_per_rank":[2682,2637,2676,2578,2702,2659,2687,2602,2635,2661,2672,2599,2691,2625,2662,2636],"payload_rank_cv":0.013201921246945858,"routed_copies":42404},"token_rate_at_latency_percentile":{"p50":9451955.276176078,"p90":9351288.443178244,"p95":9323531.620573604,"p99":9268635.65116404},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"InterNodeV1","block_num":32,"rdma_block_num":16,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":56,"scale_type_size":4,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"inter-node-v1-vllm-cudagraph-spin-ar","rank_alignment":{"method":"in-graph-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":true,"cuda_graph_supported":true,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9943,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2089,"p50":2101,"max":2103,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2090,"p50":2102,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2100,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2091,"p50":2104,"max":2105,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2096,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2105,"max":2107,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2095,"p50":2108,"max":2110,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2085,"p50":2097,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2103,"max":2105,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2104,"max":2106,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2093,"p50":2106,"max":2108,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2100,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2104,"max":2106,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2090,"p50":2102,"max":2103,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2090,"p50":2103,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2091,"p50":2102,"max":2103,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":958.7}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-bf16.json b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-bf16.json new file mode 100644 index 0000000000..cc3929b2a3 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T06:52:57.819795+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":1,"phase":"decode","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.201999142766,"p90":32.00700134038925,"p95":33.048998564481735,"p99":35.13199836015701},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.600999921560287,"p90":30.205000191926956,"p95":31.16600029170513,"p99":32.88799896836281},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":55.802999064326286,"p90":62.21200153231621,"p95":64.21499885618687,"p99":68.01999732851982},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":45.94799876213074,"p90":49.11300167441368,"p95":50.234999507665634,"p99":52.39799991250038},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":48.271000385284424,"p90":50.836000591516495,"p95":51.87699943780899,"p99":53.24000120162964},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":26.8389992415905,"p90":27.441000565886497,"p95":27.600999921560287,"p99":27.840999886393547},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":24.235999211668968,"p90":24.956999346613884,"p95":25.036999955773354,"p99":25.196999311447144},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129000335931778,"p90":6.169002503156662,"p95":6.169002503156662,"p99":6.169002503156662},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.402001827955246,"p90":5.528002977371216,"p95":6.609998643398285,"p99":8.572999387979507},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.20000338554382324,"p90":0.39999932050704956,"p95":0.39999932050704956,"p99":0.39999932050704956},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.28099934756756,"p90":27.520999312400818,"p95":27.600999921560287,"p99":27.842000126838684},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":23.83599989116192,"p90":24.196000769734383,"p95":24.356000125408173,"p99":25.157999247312546},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":44.505998492240906,"p90":44.94699835777283,"p95":45.187000185251236,"p99":46.10799998044968},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.727000206708908,"p90":6.289001554250717,"p95":7.212001830339432,"p99":8.813001215457916},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8,"byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"roundtrip":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"roundtrip":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":35,"assignments":48,"wire":"rank-deduplicated"},"receive":{"max":6,"mean":4.375,"min":2,"total":35},"routing":{"empty_expert_count":337,"empty_rank_count":0,"expert_assignment_rank_cv":0.3227486312389374,"expert_assignments_per_rank":[8,6,6,6,6,5,9,2],"expert_load_cv":2.708012819290161,"expert_load_max":2,"expert_load_mean":0.125,"expert_load_min":0,"fanout_histogram":[0,0,0,5,3,0,0,0],"fanout_max":5,"fanout_mean":4.375,"fanout_min":4,"hotspot_ratio":16.0,"locality":{"placement":"packed","local_rank_fraction":0.08571428805589676,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":35},"payload_copies_per_rank":[5,5,5,4,4,4,6,2],"payload_rank_cv":0.2539483904838562,"routed_copies":35},"token_rate_at_latency_percentile":{"p50":165730.97586846008,"p90":157368.7919370872,"p95":154210.9236597335,"p99":150262.95678887263},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.12200039625168,"p90":31.68800100684166,"p95":32.809000462293625,"p99":35.413000732660294},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":29.20299954712391,"p90":31.727001070976257,"p95":32.84800052642822,"p99":34.85200181603432},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":57.32499994337559,"p90":63.41500207781792,"p95":65.65700098872185,"p99":70.26500254869461},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":47.189999371767044,"p90":50.675999373197556,"p95":52.47800052165985,"p99":54.80099841952324},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":49.8340018093586,"p90":52.47800052165985,"p95":53.39900031685829,"p99":54.962001740932465},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":26.799999177455902,"p90":27.161000296473503,"p95":27.28099934756756,"p99":27.481000870466232},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":25.75800009071827,"p90":26.319000869989395,"p95":26.559999212622643,"p99":27.441000565886497},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129000335931778,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.6419999301433563,"p90":5.607999861240387,"p95":6.407998502254486,"p99":7.931999862194061},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.23999810218811035,"p90":0.31900033354759216,"p95":0.31900033354759216,"p99":0.31900033354759216},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.079999446868896,"p90":27.360999956727028,"p95":27.441000565886497,"p99":27.600999921560287},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":25.43799951672554,"p90":25.999000295996666,"p95":26.11899934709072,"p99":26.440000161528587},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":46.188998967409134,"p90":46.75000160932541,"p95":46.94899916648865,"p99":47.67100140452385},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.68499755859375,"p90":6.209000945091248,"p95":7.17100128531456,"p99":8.692998439073563},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"roundtrip":{"activation_data_bytes":2007040,"scale_bytes":0,"total_logical_bytes":2007040},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"roundtrip":{"activation_data_bytes":2007040,"scale_bytes":0,"total_logical_bytes":2007040},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":70,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":12,"mean":8.75,"min":5,"total":70},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.33592739701271057,"expert_assignments_per_rank":[18,17,13,7,13,7,13,8],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,3,6,5,2,0,0],"fanout_max":6,"fanout_mean":4.375,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.15714286267757416,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":70},"payload_copies_per_rank":[12,11,10,5,9,6,10,7],"payload_rank_cv":0.26649653911590576,"routed_copies":70},"token_rate_at_latency_percentile":{"p50":321065.9272600354,"p90":304889.66502060485,"p95":299631.07745574653,"p99":291110.21238668135},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.201999142766,"p90":32.16800093650818,"p95":33.16900134086609,"p99":35.413000732660294},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":31.445998698472977,"p90":34.331001341342926,"p95":35.41199862957001,"p99":36.89499944448471},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":59.647997841238976,"p90":66.4990022778511,"p95":68.5809999704361,"p99":72.308000177145},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":49.95400086045265,"p90":53.599998354911804,"p95":55.60300126671791,"p99":57.9259991645813},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":52.31799930334091,"p90":55.041998624801636,"p95":56.00300058722496,"p99":57.60499835014343},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":26.96000039577484,"p90":27.319999411702156,"p95":27.441000565886497,"p99":27.600999921560287},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.402000665664673,"p90":29.04300019145012,"p95":29.16399948298931,"p99":29.443999752402306},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129998713731766,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.4420002698898315,"p90":5.64899668097496,"p95":6.609000265598297,"p99":8.733000606298447},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.199001282453537,"p90":0.19999966025352478,"p95":0.19999966025352478,"p99":0.19999966025352478},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.2000003606081,"p90":27.401000261306763,"p95":27.481000870466232,"p99":27.600999921560287},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.721000835299492,"p90":28.201999142766,"p95":28.44199910759926,"p99":29.123999178409576},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":48.59200119972229,"p90":49.27299916744232,"p95":49.633998423814774,"p99":51.19600147008896},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.6859996616840363,"p90":6.129000335931778,"p95":7.09100067615509,"p99":9.054001420736313},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"roundtrip":{"activation_data_bytes":4100096,"scale_bytes":0,"total_logical_bytes":4100096},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"roundtrip":{"activation_data_bytes":4100096,"scale_bytes":0,"total_logical_bytes":4100096},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":143,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":22,"mean":17.875,"min":13,"total":143},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.20518453419208527,"expert_assignments_per_rank":[31,30,24,17,22,22,28,18],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,4,12,13,3,0,0],"fanout_max":6,"fanout_mean":4.46875,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.1258741319179535,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":143},"payload_copies_per_rank":[22,21,18,13,16,15,22,16],"payload_rank_cv":0.1800614446401596,"routed_copies":143},"token_rate_at_latency_percentile":{"p50":611644.1841451791,"p90":581374.2378457342,"p95":571397.954832078,"p99":555507.3503429815},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.522999957203865,"p90":32.3680005967617,"p95":33.48999843001366,"p99":35.81399843096733},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":31.927000731229782,"p90":34.49099883437157,"p95":35.61300039291382,"p99":37.29499876499176},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":60.45000068843365,"p90":66.85899943113327,"p95":69.10299882292747,"p99":73.10899719595909},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":50.11399835348129,"p90":54.32099848985672,"p95":56.1240017414093,"p99":58.60700085759163},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":52.99900099635124,"p90":55.52300065755844,"p95":56.64400011301041,"p99":58.00599977374077},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":27.11999975144863,"p90":27.51999907195568,"p95":27.600999921560287,"p99":27.80199982225895},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.162000700831413,"p90":28.72299961745739,"p95":28.843000531196594,"p99":29.12300080060959},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1689987778663635,"p90":6.207998842000961,"p95":6.207998842000961,"p99":6.207998842000961},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.7619989812374115,"p90":5.968000739812851,"p95":7.050998508930206,"p99":8.973002433776855},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.20099803805351257,"p90":0.28100237250328064,"p95":0.28100237250328064,"p99":0.28100237250328064},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.360999956727028,"p90":27.681000530719757,"p95":27.760999277234077,"p99":27.96199917793274},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.201999142766,"p90":28.522999957203865,"p95":28.64300087094307,"p99":28.922999277710915},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":49.47299882769585,"p90":49.91399869322777,"p95":50.11399835348129,"p99":51.316000521183014},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.4850016236305237,"p90":6.088998168706894,"p95":7.130000740289688,"p99":8.852001279592514},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"roundtrip":{"activation_data_bytes":8056832,"scale_bytes":0,"total_logical_bytes":8056832},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"roundtrip":{"activation_data_bytes":8056832,"scale_bytes":0,"total_logical_bytes":8056832},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":281,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":42,"mean":35.125,"min":26,"total":281},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.1458333283662796,"expert_assignments_per_rank":[55,56,58,37,46,43,46,43],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,1,9,24,24,6,0,0],"fanout_max":6,"fanout_mean":4.390625,"fanout_min":2,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.11743772029876709,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":281},"payload_copies_per_rank":[38,42,42,26,29,31,38,35],"payload_rank_cv":0.15879225730895996,"routed_copies":281},"token_rate_at_latency_percentile":{"p50":1207569.9314484461,"p90":1152675.454173019,"p95":1129863.707935768,"p99":1103334.1421514936},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":30.48500046133995,"p90":34.251000732183456,"p95":35.1720005273819,"p99":37.13500127196312},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":32.648999243974686,"p90":34.97200086712837,"p95":36.05400025844574,"p99":37.81599923968315},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":63.133999705314636,"p90":69.22300159931183,"p95":71.22600078582764,"p99":74.95100051164627},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":52.31799930334091,"p90":56.1630018055439,"p95":57.486001402139664,"p99":60.32999977469444},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":55.92299997806549,"p90":58.125998824834824,"p95":59.167999774217606,"p99":60.89099869132042},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":29.16399948298931,"p90":29.56400066614151,"p95":29.64399941265583,"p99":29.80400063097477},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.48299965262413,"p90":29.002999886870384,"p95":29.12300080060959,"p99":29.28300015628338},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1689987778663635,"p90":6.209999322891235,"p95":6.209999322891235,"p99":6.209999322891235},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.4409981667995453,"p90":5.728002637624741,"p95":6.650000810623169,"p99":8.611999452114105},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.199001282453537,"p90":0.2800002694129944,"p95":0.2800002694129944,"p99":0.2800002694129944},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":29.40399944782257,"p90":29.683999717235565,"p95":29.765000566840172,"p99":29.96399998664856},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.88300083577633,"p90":29.32400070130825,"p95":29.443999752402306,"p99":29.683999717235565},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":52.156001329422,"p90":52.678000181913376,"p95":52.838001400232315,"p99":53.15899848937988},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.724999725818634,"p90":6.209000945091248,"p95":7.2109997272491455,"p99":8.77300277352333},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"roundtrip":{"activation_data_bytes":16171008,"scale_bytes":0,"total_logical_bytes":16171008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"roundtrip":{"activation_data_bytes":16171008,"scale_bytes":0,"total_logical_bytes":16171008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":564,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":77,"mean":70.5,"min":62,"total":564},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.07384087145328522,"expert_assignments_per_rank":[101,108,101,91,83,97,94,93],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,2,14,52,50,10,0,0],"fanout_max":6,"fanout_mean":4.40625,"fanout_min":2,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.14716312289237976,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":564},"payload_copies_per_rank":[73,76,77,63,62,69,74,70],"payload_rank_cv":0.07472094893455505,"routed_copies":564},"token_rate_at_latency_percentile":{"p50":2288861.4711336135,"p90":2202112.6963466634,"p95":2163331.5388122327,"p99":2102116.9425858916},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":33.53000059723854,"p90":37.174999713897705,"p95":38.1769984960556,"p99":39.739999920129776},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":32.166000455617905,"p90":34.53100100159645,"p95":35.53299978375435,"p99":37.4549999833107},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":65.69600105285645,"p90":71.70600071549416,"p95":73.70999827980995,"p99":77.19499990344048},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":55.642999708652496,"p90":59.689000248909,"p95":60.65100058913231,"p99":63.21399658918381},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":58.52599814534187,"p90":61.00999936461449,"p95":62.09300085902214,"p99":63.81499767303467},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":32.127998769283295,"p90":32.52800181508064,"p95":32.6480008661747,"p99":32.92899951338768},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.562000021338463,"p90":29.083000496029854,"p95":29.203999787569046,"p99":29.44299951195717},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129000335931778,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.1240003407001495,"p90":5.688000470399857,"p95":6.250001490116119,"p99":8.77399742603302},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.19999966025352478,"p90":0.2400018274784088,"p95":0.2400018274784088,"p99":0.2400018274784088},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":32.24800154566765,"p90":32.56800025701523,"p95":32.648999243974686,"p99":32.84899890422821},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.44199910759926,"p90":28.88300083577633,"p95":29.002999886870384,"p99":29.323000460863113},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":54.72100153565407,"p90":55.32199889421463,"p95":55.56200072169304,"p99":57.36599862575531},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.7649981677532196,"p90":5.969002842903137,"p95":7.2109997272491455,"p99":9.093999862670898},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"roundtrip":{"activation_data_bytes":32628736,"scale_bytes":0,"total_logical_bytes":32628736},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"roundtrip":{"activation_data_bytes":32628736,"scale_bytes":0,"total_logical_bytes":32628736},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1138,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":149,"mean":142.25,"min":126,"total":1138},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.04658475145697594,"expert_assignments_per_rank":[189,194,196,170,193,201,198,195],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,3,26,101,106,20,0,0],"fanout_max":6,"fanout_mean":4.4453125,"fanout_min":2,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.1344463974237442,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":1138},"payload_copies_per_rank":[143,141,149,126,144,145,148,142],"payload_rank_cv":0.04686218500137329,"routed_copies":1138},"token_rate_at_latency_percentile":{"p50":4374124.459428382,"p90":4196033.480840827,"p95":4122847.9290480786,"p99":4011596.1660243706},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":43.063998222351074,"p90":46.42900079488754,"p95":47.3099984228611,"p99":48.312000930309296},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":34.57200154662132,"p90":37.255000323057175,"p95":38.33699971437454,"p99":39.85900059342384},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":77.6359997689724,"p90":83.68400111794472,"p95":85.64699813723564,"p99":88.17100152373314},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":67.94100254774094,"p90":71.0659995675087,"p95":72.54800200462341,"p99":75.51199942827225},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":70.34499943256378,"p90":72.98800349235535,"p95":73.99000227451324,"p99":75.51199942827225},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":41.62200167775154,"p90":42.18199849128723,"p95":42.30299964547157,"p99":42.86399856209755},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":31.24699927866459,"p90":31.807001680135727,"p95":32.00700134038925,"p99":32.287999987602234},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1689987778663635,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.6829967498779297,"p90":5.888000130653381,"p95":7.570996880531311,"p99":9.533993899822235},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.1600012183189392,"p90":0.39999932050704956,"p95":0.39999932050704956,"p99":0.39999932050704956},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":41.78199917078018,"p90":42.142000049352646,"p95":42.2619991004467,"p99":42.5029993057251},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":30.845999717712402,"p90":31.286999583244324,"p95":31.447000801563263,"p99":31.807001680135727},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":66.45800173282623,"p90":67.22000241279602,"p95":67.46000051498413,"p99":68.5420036315918},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.926001489162445,"p90":6.290003657341003,"p95":7.1709975600242615,"p99":9.13500040769577},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.007518797181546688,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"roundtrip":{"activation_data_bytes":64569344,"scale_bytes":0,"total_logical_bytes":64569344},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"roundtrip":{"activation_data_bytes":64569344,"scale_bytes":0,"total_logical_bytes":64569344},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2252,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":290,"mean":281.5,"min":268,"total":2252},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03184119611978531,"expert_assignments_per_rank":[396,393,399,359,380,389,380,376],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,3,57,223,191,38,0,0],"fanout_max":6,"fanout_mean":4.3984375,"fanout_min":2,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.11767318099737167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":2252},"payload_copies_per_rank":[290,286,290,268,282,282,280,274],"payload_rank_cv":0.025306938216090202,"routed_copies":2252},"token_rate_at_latency_percentile":{"p50":7278413.592011308,"p90":7014851.420804052,"p95":6919853.821606985,"p99":6780379.328802456},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":68.18100064992905,"p90":71.22600078582764,"p95":72.22700119018555,"p99":73.78900051116943},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":47.35000059008598,"p90":49.233000725507736,"p95":50.27399957180023,"p99":52.03799903392792},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":115.53100124001503,"p90":120.45900151133537,"p95":122.50100076198578,"p99":125.82699954509735},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":106.39800131320953,"p90":108.0000028014183,"p95":109.2429980635643,"p99":112.36699670553207},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":108.5209995508194,"p90":110.48399657011032,"p95":111.56599968671799,"p99":113.48900198936462},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":66.89999997615814,"p90":67.74099916219711,"p95":67.94100254774094,"p99":68.50200146436691},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":43.9860001206398,"p90":44.787000864744186,"p95":44.94699835777283,"p99":45.347001403570175},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.168000400066376,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.841999590396881,"p90":6.3299983739852905,"p95":7.610999047756195,"p99":9.893998503684998},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.160999596118927,"p90":0.2800002694129944,"p95":0.2800002694129944,"p99":0.2800002694129944},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":66.61900132894516,"p90":67.22000241279602,"p95":67.41999834775925,"p99":67.90100038051605},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":43.584998697042465,"p90":44.14600133895874,"p95":44.34499889612198,"p99":44.70599815249443},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":104.43499684333801,"p90":105.31499981880188,"p95":105.55700212717056,"p99":106.03799670934677},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.046998918056488,"p90":6.4909979701042175,"p95":7.651001214981079,"p99":8.973002433776855},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"roundtrip":{"activation_data_bytes":128651264,"scale_bytes":0,"total_logical_bytes":128651264},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"roundtrip":{"activation_data_bytes":128651264,"scale_bytes":0,"total_logical_bytes":128651264},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":4487,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":569,"mean":560.875,"min":553,"total":4487},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.01615842990577221,"expert_assignments_per_rank":[775,756,779,747,762,761,782,782],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,5,126,437,385,71,0,0],"fanout_max":6,"fanout_mean":4.3818359375,"fanout_min":2,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.12859371304512024,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":4487},"payload_copies_per_rank":[569,555,565,556,553,564,558,567],"payload_rank_cv":0.010142241604626179,"routed_copies":4487},"token_rate_at_latency_percentile":{"p50":9435961.742321312,"p90":9268310.63130664,"p95":9178423.559824992,"p99":9022900.739720682},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":87.65000104904175,"p90":90.33399820327759,"p95":91.37500077486038,"p99":92.85800158977509},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":71.26600295305252,"p90":72.98900187015533,"p95":73.95000010728836,"p99":75.75300335884094},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":158.91600400209427,"p90":163.32300007343292,"p95":165.32500088214874,"p99":168.61100494861603},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":150.94399452209473,"p90":152.14599668979645,"p95":152.62700617313385,"p99":154.22900021076202},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":151.46499872207642,"p90":153.38699519634247,"p95":154.42900359630585,"p99":156.23199939727783},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":85.487000644207,"p90":86.40799671411514,"p95":86.64800226688385,"p99":87.13000267744064},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":69.54299658536911,"p90":70.50500065088272,"p95":70.70499658584595,"p99":71.22600078582764},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129011511802673,"p90":6.1690062284469604,"p95":6.1690062284469604,"p99":6.1690062284469604},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.7230063676834106,"p90":6.290003657341003,"p95":7.85200297832489,"p99":10.736003518104553},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.5200058221817017,"p90":-0.4000067710876465,"p95":-0.4000067710876465,"p99":-0.4000067710876465},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":85.76700091362,"p90":86.64900064468384,"p95":86.88999712467194,"p99":87.37000077962875},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":67.57999956607819,"p90":68.74299794435501,"p95":69.022998213768,"p99":69.54299658536911},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":147.9800045490265,"p90":149.5019942522049,"p95":149.90299940109253,"p99":150.62299370765686},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.3250004053115845,"p90":5.968004465103149,"p95":7.330000400543213,"p99":9.534001350402832},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"roundtrip":{"activation_data_bytes":259768320,"scale_bytes":0,"total_logical_bytes":259768320},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"roundtrip":{"activation_data_bytes":259768320,"scale_bytes":0,"total_logical_bytes":259768320},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":9060,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":1149,"mean":1132.5,"min":1088,"total":9060},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.018761977553367615,"expert_assignments_per_rank":[1542,1490,1556,1492,1545,1541,1581,1541],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,11,225,859,791,162,0,0],"fanout_max":6,"fanout_mean":4.423828125,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.12792494893074036,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":9060},"payload_copies_per_rank":[1144,1088,1148,1126,1135,1149,1140,1130],"payload_rank_cv":0.016329573467373848,"routed_copies":9060},"token_rate_at_latency_percentile":{"p50":13521275.656284666,"p90":13351849.010266254,"p95":13261757.521622647,"p99":13108710.173977869},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":163.20200264453888,"p90":166.4070039987564,"p95":167.80899465084076,"p99":170.57399451732635},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":122.1420019865036,"p90":123.58400225639343,"p95":124.54500049352646,"p99":126.90800428390503},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":285.3440046310425,"p90":289.99100625514984,"p95":292.3539951443672,"p99":297.4819988012314},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":276.36998891830444,"p90":280.456006526947,"p95":282.0590138435364,"p99":286.0639989376068},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":278.5339951515198,"p90":282.5790047645569,"p95":283.94100069999695,"p99":286.7859899997711},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":162.48099505901337,"p90":166.44799709320068,"p95":168.53100061416626,"p99":173.7779974937439},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":115.73199927806854,"p90":117.93600022792816,"p95":118.3760017156601,"p99":119.45799738168716},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.168991327285767,"p90":6.287992000579834,"p95":6.287992000579834,"p99":6.287992000579834},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.884984016418457,"p90":6.889998912811279,"p95":8.451998233795166,"p99":10.015994310379028},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.5609989166259766,"p90":0.7210075855255127,"p95":0.7210075855255127,"p99":0.7210075855255127},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":158.99699926376343,"p90":161.67999804019928,"p95":162.6809984445572,"p99":165.08500277996063},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":118.49600076675415,"p90":119.53800171613693,"p95":119.81800198554993,"p99":120.33899873495102},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":274.00699257850647,"p90":277.61200070381165,"p95":279.0150046348572,"p99":281.69798851013184},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.407018423080444,"p90":7.010996341705322,"p95":8.011996746063232,"p99":9.494990110397339},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"roundtrip":{"activation_data_bytes":520626176,"scale_bytes":0,"total_logical_bytes":520626176},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"roundtrip":{"activation_data_bytes":520626176,"scale_bytes":0,"total_logical_bytes":520626176},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":18158,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":2294,"mean":2269.75,"min":2209,"total":18158},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.016587993130087852,"expert_assignments_per_rank":[3123,2960,3095,3022,3074,3092,3108,3102],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,23,454,1699,1566,354,0,0],"fanout_max":6,"fanout_mean":4.43310546875,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.1253442019224167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":18158},"payload_copies_per_rank":[2289,2209,2294,2272,2259,2292,2267,2276],"payload_rank_cv":0.01135234534740448,"routed_copies":18158},"token_rate_at_latency_percentile":{"p50":14705565.824278707,"p90":14495061.31360595,"p95":14425532.029196812,"p99":14282427.115785083},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"IntraNode","block_num":80,"rdma_block_num":0,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":0,"scale_type_size":0,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"intranode-vllm-cudagraph-spin-ar","rank_alignment":{"method":"in-graph-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":true,"cuda_graph_supported":true,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi300x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9938,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2059,"p50":2108,"max":2110,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":741.9},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2051,"p50":2099,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":741.9},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2050,"p50":2099,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":741.9},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2051,"p50":2098,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":741.9},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2067,"p50":2104,"max":2106,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":741.9},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2067,"p50":2106,"max":2107,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":741.9},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2067,"p50":2108,"max":2110,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":741.9},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2074,"p50":2111,"max":2113,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":741.9}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-fp8.json b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-fp8.json new file mode 100644 index 0000000000..43a655a70d --- /dev/null +++ b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T06:53:40.662138+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":1,"phase":"decode","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.96199917793274,"p90":31.766999512910843,"p95":32.648999243974686,"p99":34.1310016810894},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.320999652147293,"p90":30.16499988734722,"p95":31.16600029170513,"p99":33.08999910950661},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":55.28299883008003,"p90":61.931999400258064,"p95":63.81499953567982,"p99":67.22100079059601},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":45.22699862718582,"p90":48.95300045609474,"p95":50.11500045657158,"p99":52.99999937415123},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":47.99199849367142,"p90":50.755999982357025,"p95":51.596999168395996,"p99":53.35899814963341},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":26.680000126361847,"p90":27.079999446868896,"p95":27.160000056028366,"p99":27.28099934756756},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":23.91500025987625,"p90":24.356000125408173,"p95":24.475999176502228,"p99":24.636000394821167},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.127998232841492,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.8430016934871674,"p90":5.929999053478241,"p95":7.811002433300018,"p99":9.61500033736229},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.3609992563724518,"p90":-0.20100176334381104,"p95":-0.20100176334381104,"p99":-0.20100176334381104},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":26.920000091195107,"p90":27.160000056028366,"p95":27.201000601053238,"p99":27.320999652147293},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":23.715000599622726,"p90":24.075999855995178,"p95":24.235999211668968,"p99":24.557000026106834},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":44.38599944114685,"p90":44.90699991583824,"p95":45.06700113415718,"p99":45.54799944162369},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.5650022327899933,"p90":6.288997828960419,"p95":7.130999118089676,"p99":8.974000811576843},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8,"byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":250880,"scale_bytes":7840,"total_logical_bytes":258720},"roundtrip":{"activation_data_bytes":752640,"scale_bytes":7840,"total_logical_bytes":760480},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":250880,"scale_bytes":7840,"total_logical_bytes":258720},"roundtrip":{"activation_data_bytes":752640,"scale_bytes":7840,"total_logical_bytes":760480},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":35,"assignments":48,"wire":"rank-deduplicated"},"receive":{"max":6,"mean":4.375,"min":2,"total":35},"routing":{"empty_expert_count":337,"empty_rank_count":0,"expert_assignment_rank_cv":0.3227486312389374,"expert_assignments_per_rank":[8,6,6,6,6,5,9,2],"expert_load_cv":2.708012819290161,"expert_load_max":2,"expert_load_mean":0.125,"expert_load_min":0,"fanout_histogram":[0,0,0,5,3,0,0,0],"fanout_max":5,"fanout_mean":4.375,"fanout_min":4,"hotspot_ratio":16.0,"locality":{"placement":"packed","local_rank_fraction":0.08571428805589676,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":35},"payload_copies_per_rank":[5,5,5,4,4,4,6,2],"payload_rank_cv":0.2539483904838562,"routed_copies":35},"token_rate_at_latency_percentile":{"p50":166694.45430689745,"p90":157616.83353260363,"p95":155047.7765943437,"p99":149927.8524226745},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.88200043141842,"p90":31.727001070976257,"p95":32.489001750946045,"p99":34.09000113606453},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.201999142766,"p90":31.0869999229908,"p95":32.127998769283295,"p99":33.649999648332596},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":56.08399957418442,"p90":62.814000993967056,"p95":64.61700052022934,"p99":67.74000078439713},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":46.38899862766266,"p90":48.8319993019104,"p95":50.39500072598457,"p99":53.199999034404755},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":48.71299862861633,"p90":51.43599957227707,"p95":52.31799930334091,"p99":53.76100167632103},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":26.719000190496445,"p90":27.11999975144863,"p95":27.240000665187836,"p99":27.441000565886497},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":24.9170009046793,"p90":25.276999920606613,"p95":25.35800077021122,"p99":25.598999112844467},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.168000400066376,"p95":6.168000400066376,"p99":6.168000400066376},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.2819990515708923,"p90":5.247998982667923,"p95":6.009001284837723,"p99":7.812999188899994},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.12099742889404297,"p90":0.20000338554382324,"p95":0.20000338554382324,"p99":0.20000338554382324},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":26.67899988591671,"p90":26.920000091195107,"p95":27.079999446868896,"p99":27.360999956727028},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":24.67699907720089,"p90":25.117000564932823,"p95":25.237999856472015,"p99":25.51800012588501},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":45.30800133943558,"p90":45.94799876213074,"p95":46.14799842238426,"p99":46.66899889707565},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.36499884724617,"p90":6.129000335931778,"p95":6.930001080036163,"p99":8.413001894950867},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":501760,"scale_bytes":15680,"total_logical_bytes":517440},"roundtrip":{"activation_data_bytes":1505280,"scale_bytes":15680,"total_logical_bytes":1520960},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":501760,"scale_bytes":15680,"total_logical_bytes":517440},"roundtrip":{"activation_data_bytes":1505280,"scale_bytes":15680,"total_logical_bytes":1520960},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":70,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":12,"mean":8.75,"min":5,"total":70},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.33592739701271057,"expert_assignments_per_rank":[18,17,13,7,13,7,13,8],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,3,6,5,2,0,0],"fanout_max":6,"fanout_mean":4.375,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.15714286267757416,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":70},"payload_copies_per_rank":[12,11,10,5,9,6,10,7],"payload_rank_cv":0.26649653911590576,"routed_copies":70},"token_rate_at_latency_percentile":{"p50":328454.4259322365,"p90":311066.1819163648,"p95":305822.09207258956,"p99":297613.5023735464},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.12200039625168,"p90":31.766999512910843,"p95":32.648999243974686,"p99":34.651998430490494},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":31.0869999229908,"p90":33.69000181555748,"p95":34.73199903964996,"p99":36.173999309539795},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":59.20900031924248,"p90":65.45700132846832,"p95":67.38099828362465,"p99":70.82599774003029},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":49.27299916744232,"p90":52.11799964308739,"p95":53.55999991297722,"p99":57.20499902963638},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":51.876001060009,"p90":54.68200147151947,"p95":55.56299909949303,"p99":56.92499876022339},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":26.680000126361847,"p90":27.0409993827343,"p95":27.161000296473503,"p99":27.401000261306763},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":27.800999581813812,"p90":28.241999447345734,"p95":28.362000361084938,"p99":28.64300087094307},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.09000027179718,"p90":6.17000088095665,"p95":6.17000088095665,"p99":6.17000088095665},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.3229995965957642,"p90":4.967000335454941,"p95":6.769999861717224,"p99":9.093999862670898},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.39999932050704956,"p90":0.28100237250328064,"p95":0.28100237250328064,"p99":0.28100237250328064},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":26.87999978661537,"p90":27.240000665187836,"p95":27.319999411702156,"p99":27.441000565886497},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.600999921560287,"p90":28.04199978709221,"p95":28.162000700831413,"p99":28.481999412178993},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":48.392001539468765,"p90":48.87299984693527,"p95":49.0729995071888,"p99":50.43499916791916},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.4449994564056396,"p90":6.250001490116119,"p95":7.130999118089676,"p99":8.574001491069794},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":1025024,"scale_bytes":32032,"total_logical_bytes":1057056},"roundtrip":{"activation_data_bytes":3075072,"scale_bytes":32032,"total_logical_bytes":3107104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":1025024,"scale_bytes":32032,"total_logical_bytes":1057056},"roundtrip":{"activation_data_bytes":3075072,"scale_bytes":32032,"total_logical_bytes":3107104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":143,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":22,"mean":17.875,"min":13,"total":143},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.20518453419208527,"expert_assignments_per_rank":[31,30,24,17,22,22,28,18],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,4,12,13,3,0,0],"fanout_max":6,"fanout_mean":4.46875,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.1258741319179535,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":143},"payload_copies_per_rank":[22,21,18,13,16,15,22,16],"payload_rank_cv":0.1800614446401596,"routed_copies":143},"token_rate_at_latency_percentile":{"p50":616855.5660831125,"p90":585201.6959669418,"p95":575922.835675225,"p99":562143.1830818088},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.201999142766,"p90":31.766999512910843,"p95":32.76899829506874,"p99":34.331999719142914},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":31.766999512910843,"p90":34.1310016810894,"p95":35.333000123500824,"p99":37.05500066280365},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":59.96899865567684,"p90":65.89800119400024,"p95":68.10199841856956,"p99":71.38700038194656},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":49.87400025129318,"p90":52.63800173997879,"p95":54.12000045180321,"p99":57.80699849128723},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":52.75899916887283,"p90":55.121999233961105,"p95":56.08300119638443,"p99":57.805001735687256},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":26.920000091195107,"p90":27.320999652147293,"p95":27.401000261306763,"p99":27.600999921560287},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.162000700831413,"p90":28.682999312877655,"p95":28.843000531196594,"p99":29.28300015628338},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129000335931778,"p90":6.129998713731766,"p95":6.129998713731766,"p99":6.129998713731766},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.201000064611435,"p90":5.047999322414398,"p95":6.288997828960419,"p99":9.375002235174179},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.11999905109405518,"p90":0.199001282453537,"p95":0.199001282453537,"p99":0.199001282453537},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.079999446868896,"p90":27.401000261306763,"p95":27.520999312400818,"p99":27.682000771164894},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.241999447345734,"p90":28.80300022661686,"p95":28.96299958229065,"p99":29.242999851703644},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":49.313999712467194,"p90":49.95400086045265,"p95":50.11500045657158,"p99":50.95599964261055},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.323998302221298,"p90":5.888998508453369,"p95":6.72999769449234,"p99":8.574001491069794},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":2014208,"scale_bytes":62944,"total_logical_bytes":2077152},"roundtrip":{"activation_data_bytes":6042624,"scale_bytes":62944,"total_logical_bytes":6105568},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":2014208,"scale_bytes":62944,"total_logical_bytes":2077152},"roundtrip":{"activation_data_bytes":6042624,"scale_bytes":62944,"total_logical_bytes":6105568},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":281,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":42,"mean":35.125,"min":26,"total":281},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.1458333283662796,"expert_assignments_per_rank":[55,56,58,37,46,43,46,43],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,1,9,24,24,6,0,0],"fanout_max":6,"fanout_mean":4.390625,"fanout_min":2,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.11743772029876709,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":281},"payload_copies_per_rank":[38,42,42,26,29,31,38,35],"payload_rank_cv":0.15879225730895996,"routed_copies":281},"token_rate_at_latency_percentile":{"p50":1213063.1931653325,"p90":1161060.9355505577,"p95":1141165.7478153287,"p99":1107170.626732948},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":30.246000736951828,"p90":33.810000866651535,"p95":34.77200120687485,"p99":36.13400086760521},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":31.647998839616776,"p90":34.331001341342926,"p95":35.37299856543541,"p99":36.855001002550125},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":61.8939995765686,"p90":68.14100220799446,"p95":70.14499977231026,"p99":72.98900187015533},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":52.23799869418144,"p90":55.24199828505516,"p95":56.92499876022339,"p99":59.08900126814842},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":54.72100153565407,"p90":57.326000183820724,"p95":58.246999979019165,"p99":59.64900180697441},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.88300083577633,"p90":29.284000396728516,"p95":29.48400005698204,"p99":29.764000326395035},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.362000361084938,"p90":29.002999886870384,"p95":29.12300080060959,"p99":29.32400070130825},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.088998168706894,"p90":6.168000400066376,"p95":6.168000400066376,"p99":6.168000400066376},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.6029998660087585,"p90":5.767997354269028,"p95":6.849996745586395,"p99":9.97600331902504},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.12100115418434143,"p90":0.4409998655319214,"p95":0.4409998655319214,"p99":0.4409998655319214},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":29.002999886870384,"p90":29.40399944782257,"p95":29.603999108076096,"p99":29.844999313354492},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.04199978709221,"p90":28.64300087094307,"p95":28.801999986171722,"p99":29.12300080060959},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":51.155999302864075,"p90":51.79699882864952,"p95":52.03700065612793,"p99":52.55899950861931},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.4850016236305237,"p90":6.12800195813179,"p95":7.09100067615509,"p99":8.733000606298447},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":4042752,"scale_bytes":126336,"total_logical_bytes":4169088},"roundtrip":{"activation_data_bytes":12128256,"scale_bytes":126336,"total_logical_bytes":12254592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":4042752,"scale_bytes":126336,"total_logical_bytes":4169088},"roundtrip":{"activation_data_bytes":12128256,"scale_bytes":126336,"total_logical_bytes":12254592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":564,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":77,"mean":70.5,"min":62,"total":564},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.07384087145328522,"expert_assignments_per_rank":[101,108,101,91,83,97,94,93],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,2,14,52,50,10,0,0],"fanout_max":6,"fanout_mean":4.40625,"fanout_min":2,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.14716312289237976,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":564},"payload_copies_per_rank":[73,76,77,63,62,69,74,70],"payload_rank_cv":0.07472094893455505,"routed_copies":564},"token_rate_at_latency_percentile":{"p50":2339138.4734908445,"p90":2232843.728666871,"p95":2197538.0714218104,"p99":2145886.7059370256},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":33.369001001119614,"p90":36.97500005364418,"p95":37.776000797748566,"p99":39.33899849653244},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":31.566999852657318,"p90":34.21099856495857,"p95":35.1720005273819,"p99":36.93599998950958},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":64.93600085377693,"p90":71.18599861860275,"p95":72.94800132513046,"p99":76.27499848604202},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":55.121999233961105,"p90":59.56900119781494,"p95":60.451000928878784,"p99":61.81199848651886},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":57.806000113487244,"p90":60.24999916553497,"p95":61.29100173711777,"p99":62.654003500938416},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":31.727001070976257,"p90":32.16800093650818,"p95":32.3680005967617,"p99":32.52900019288063},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.12200039625168,"p90":28.762999922037125,"p95":28.96299958229065,"p99":29.363999143242836},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.167996674776077,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.2429972887039185,"p90":6.770003587007523,"p95":7.851999253034592,"p99":10.095003992319107},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.15999749302864075,"p90":0.6399974226951599,"p95":0.6399974226951599,"p99":0.6399974226951599},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":31.888000667095184,"p90":32.287999987602234,"p95":32.409001141786575,"p99":32.68799930810928},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.001999482512474,"p90":28.402000665664673,"p95":28.521999716758728,"p99":28.80300022661686},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":54.07999828457832,"p90":54.600998759269714,"p95":54.802000522613525,"p99":56.16400018334389},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.6460012197494507,"p90":6.09000027179718,"p95":7.09100067615509,"p99":8.572999387979507},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":8157184,"scale_bytes":254912,"total_logical_bytes":8412096},"roundtrip":{"activation_data_bytes":24471552,"scale_bytes":254912,"total_logical_bytes":24726464},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":8157184,"scale_bytes":254912,"total_logical_bytes":8412096},"roundtrip":{"activation_data_bytes":24471552,"scale_bytes":254912,"total_logical_bytes":24726464},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1138,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":149,"mean":142.25,"min":126,"total":1138},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.04658475145697594,"expert_assignments_per_rank":[189,194,196,170,193,201,198,195],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,3,26,101,106,20,0,0],"fanout_max":6,"fanout_mean":4.4453125,"fanout_min":2,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.1344463974237442,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":1138},"payload_copies_per_rank":[143,141,149,126,144,145,148,142],"payload_rank_cv":0.04686218500137329,"routed_copies":1138},"token_rate_at_latency_percentile":{"p50":4428606.018361584,"p90":4248962.714449972,"p95":4176795.8222971363,"p99":4085932.0345932194},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":42.70299896597862,"p90":46.06899991631508,"p95":46.950001269578934,"p99":49.27400127053261},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":33.53000059723854,"p90":36.21400147676468,"p95":37.4159999191761,"p99":39.01800140738487},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":76.23299956321716,"p90":82.28300139307976,"p95":84.36600118875504,"p99":88.29200267791748},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":66.61900132894516,"p90":69.46299970149994,"p95":70.94600051641464,"p99":73.78999888896942},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":69.18299943208694,"p90":71.78699970245361,"p95":72.94800132513046,"p99":74.39099997282028},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":41.10100120306015,"p90":41.66200011968613,"p95":41.822001338005066,"p99":42.26300120353699},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":30.285999178886414,"p90":30.805999413132668,"p95":30.966000631451607,"p99":31.2460009008646},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.129994988441467,"p95":6.129994988441467,"p99":6.129994988441467},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.6029998660087585,"p90":5.648002028465271,"p95":7.050998508930206,"p99":9.934000670909882},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.20100176334381104,"p90":0.35899877548217773,"p95":0.35899877548217773,"p99":0.35899877548217773},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":41.26099869608879,"p90":41.742000728845596,"p95":41.94200038909912,"p99":42.18300059437752},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":29.884999617934227,"p90":30.48500046133995,"p95":30.686000362038612,"p99":31.005999073386192},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":65.29700011014938,"p90":66.13899767398834,"p95":66.41899794340134,"p99":67.34099984169006},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.845997154712677,"p90":6.328999996185303,"p95":7.371000945568085,"p99":9.01300460100174},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.007518797181546688,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":16142336,"scale_bytes":504448,"total_logical_bytes":16646784},"roundtrip":{"activation_data_bytes":48427008,"scale_bytes":504448,"total_logical_bytes":48931456},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":16142336,"scale_bytes":504448,"total_logical_bytes":16646784},"roundtrip":{"activation_data_bytes":48427008,"scale_bytes":504448,"total_logical_bytes":48931456},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2252,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":290,"mean":281.5,"min":268,"total":2252},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03184119611978531,"expert_assignments_per_rank":[396,393,399,359,380,389,380,376],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,3,57,223,191,38,0,0],"fanout_max":6,"fanout_mean":4.3984375,"fanout_min":2,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.11767318099737167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":2252},"payload_copies_per_rank":[290,286,290,268,282,282,280,274],"payload_rank_cv":0.025306938216090202,"routed_copies":2252},"token_rate_at_latency_percentile":{"p50":7400662.073095018,"p90":7132210.59693487,"p95":7018698.123311254,"p99":6882552.999516956},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":67.86099821329117,"p90":70.82600146532059,"p95":71.8270018696785,"p99":74.11099970340729},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":40.25999829173088,"p90":42.42299869656563,"p95":43.584998697042465,"p99":45.34799978137016},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":108.12099650502205,"p90":113.24900016188622,"p95":115.41200056672096,"p99":119.45899948477745},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":99.3880033493042,"p90":101.10999643802643,"p95":102.1919995546341,"p99":104.71600294113159},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":101.4309972524643,"p90":103.59399765729904,"p95":104.51599955558777,"p99":106.35799914598465},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":66.21900200843811,"p90":66.94000214338303,"p95":67.29999929666519,"p99":67.781001329422},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":37.17600181698799,"p90":37.97699883580208,"p95":38.1769984960556,"p99":38.49799931049347},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129004061222076,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.883000135421753,"p90":6.530001759529114,"p95":8.212000131607056,"p99":9.814001619815826},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.24000555276870728,"p90":0.40099769830703735,"p95":0.40099769830703735,"p99":0.40099769830703735},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":66.13799929618835,"p90":66.7790025472641,"p95":67.0199990272522,"p99":67.58099794387817},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":36.6550013422966,"p90":37.29499876499176,"p95":37.53599897027016,"p99":37.97600045800209},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":97.78600186109543,"p90":98.70699793100357,"p95":98.94700348377228,"p99":99.58799928426743},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.564998507499695,"p90":6.248995661735535,"p95":7.371000945568085,"p99":8.892998099327087},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.007518797181546688,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":32162816,"scale_bytes":1005088,"total_logical_bytes":33167904},"roundtrip":{"activation_data_bytes":96488448,"scale_bytes":1005088,"total_logical_bytes":97493536},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":32162816,"scale_bytes":1005088,"total_logical_bytes":33167904},"roundtrip":{"activation_data_bytes":96488448,"scale_bytes":1005088,"total_logical_bytes":97493536},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":4487,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":569,"mean":560.875,"min":553,"total":4487},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.01615842990577221,"expert_assignments_per_rank":[775,756,779,747,762,761,782,782],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,5,126,437,385,71,0,0],"fanout_max":6,"fanout_mean":4.3818359375,"fanout_min":2,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.12859371304512024,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":4487},"payload_copies_per_rank":[569,555,565,556,553,564,558,567],"payload_rank_cv":0.010142241604626179,"routed_copies":4487},"token_rate_at_latency_percentile":{"p50":10095533.197324663,"p90":9884742.583132189,"p95":9797543.00158969,"p99":9627860.698982125},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":87.60900050401688,"p90":90.69599956274033,"p95":91.73700213432312,"p99":93.21899712085724},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":52.99900099635124,"p90":54.84199896454811,"p95":55.883001536130905,"p99":57.806000113487244},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":140.60800150036812,"p90":145.53799852728844,"p95":147.62000367045403,"p99":151.02499723434448},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":133.47899913787842,"p90":134.60099697113037,"p95":134.96099412441254,"p99":136.88400387763977},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":133.19900631904602,"p90":135.562002658844,"p95":136.40399277210236,"p99":138.28599452972412},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":86.28900349140167,"p90":87.20999956130981,"p95":87.4909982085228,"p99":87.81100064516068},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":51.51699855923653,"p90":52.43799835443497,"p95":52.63800173997879,"p99":52.99900099635124},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":6.169989705085754,"p95":6.169989705085754,"p99":6.169989705085754},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.1229982376098633,"p90":5.848005414009094,"p95":7.1300119161605835,"p99":9.85400378704071},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.3210008144378662,"p90":0.5209892988204956,"p95":0.5209892988204956,"p99":0.5209892988204956},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":85.7279971241951,"p90":86.52999997138977,"p95":86.76999807357788,"p99":87.28999644517899},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":49.35299977660179,"p90":50.39599910378456,"p95":50.675999373197556,"p99":51.11600086092949},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":129.75400686264038,"p90":131.07499480247498,"p95":131.39599561691284,"p99":131.95699453353882},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.2840073108673096,"p90":6.248995661735535,"p95":7.132008671760559,"p99":9.332999587059021},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.007518797181546688,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":64942080,"scale_bytes":2029440,"total_logical_bytes":66971520},"roundtrip":{"activation_data_bytes":194826240,"scale_bytes":2029440,"total_logical_bytes":196855680},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":64942080,"scale_bytes":2029440,"total_logical_bytes":66971520},"roundtrip":{"activation_data_bytes":194826240,"scale_bytes":2029440,"total_logical_bytes":196855680},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":9060,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":1149,"mean":1132.5,"min":1088,"total":9060},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.018761977553367615,"expert_assignments_per_rank":[1542,1490,1556,1492,1545,1541,1581,1541],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,11,225,859,791,162,0,0],"fanout_max":6,"fanout_mean":4.423828125,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.12792494893074036,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":9060},"payload_copies_per_rank":[1144,1088,1148,1126,1135,1149,1140,1130],"payload_rank_cv":0.016329573467373848,"routed_copies":9060},"token_rate_at_latency_percentile":{"p50":15375490.077564929,"p90":15107478.200613536,"p95":15014223.252406593,"p99":14809887.342277378},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":170.09399831295013,"p90":175.0209927558899,"p95":176.38300359249115,"p99":178.70700359344482},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":82.32299983501434,"p90":84.4459980726242,"p95":85.56800335645676,"p99":87.01000362634659},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":252.41699814796448,"p90":259.4669908285141,"p95":261.9510069489479,"p99":265.7170072197914},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":241.4799928665161,"p90":246.92800641059875,"p95":248.73100221157074,"p99":251.01500749588013},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":243.7629997730255,"p90":249.25200641155243,"p95":250.77399611473083,"p99":254.33900952339172},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":163.96500170230865,"p90":169.6929931640625,"p95":171.09499871730804,"p99":174.06000196933746},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":78.15700024366379,"p90":79.47800308465958,"p95":79.87900078296661,"p99":81.24099671840668},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.2499940395355225,"p95":6.2499940395355225,"p99":6.2499940395355225},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.92399525642395,"p90":7.130995392799377,"p95":8.692994713783264,"p99":11.176005005836487},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.840991735458374,"p90":0.9219944477081299,"p95":0.9219944477081299,"p99":0.9219944477081299},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":165.6070053577423,"p90":170.6549972295761,"p95":171.77599668502808,"p99":174.82100427150726},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":78.63699644804001,"p90":79.59800213575363,"p95":79.9190029501915,"p99":80.55999875068665},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":239.2359972000122,"p90":244.68499422073364,"p95":246.20699882507324,"p99":249.01199340820312},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.3669939041137695,"p90":6.650000810623169,"p95":7.851004600524902,"p99":9.57399606704712},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.007518797181546688,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":130156544,"scale_bytes":4067392,"total_logical_bytes":134223936},"roundtrip":{"activation_data_bytes":390469632,"scale_bytes":4067392,"total_logical_bytes":394537024},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":130156544,"scale_bytes":4067392,"total_logical_bytes":134223936},"roundtrip":{"activation_data_bytes":390469632,"scale_bytes":4067392,"total_logical_bytes":394537024},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":18158,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":2294,"mean":2269.75,"min":2209,"total":18158},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.016587993130087852,"expert_assignments_per_rank":[3123,2960,3095,3022,3074,3092,3108,3102],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,23,454,1699,1566,354,0,0],"fanout_max":6,"fanout_mean":4.43310546875,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.1253442019224167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":18158},"payload_copies_per_rank":[2289,2209,2294,2272,2259,2292,2267,2276],"payload_rank_cv":0.01135234534740448,"routed_copies":18158},"token_rate_at_latency_percentile":{"p50":16803206.408740863,"p90":16433167.616058785,"p95":16333431.948526481,"p99":16104489.860503638},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"IntraNode","block_num":80,"rdma_block_num":0,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":56,"scale_type_size":4,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"intranode-vllm-cudagraph-spin-ar","rank_alignment":{"method":"in-graph-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":true,"cuda_graph_supported":true,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi300x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9962,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2073,"p50":2107,"max":2109,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":779.0},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2070,"p50":2104,"max":2105,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":779.0},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2067,"p50":2102,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":779.0},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2065,"p50":2099,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":779.0},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2073,"p50":2106,"max":2107,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":779.0},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2072,"p50":2103,"max":2107,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":779.0},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2069,"p50":2103,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":779.0},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2072,"p50":2107,"max":2109,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":779.0}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16.json b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16.json new file mode 100644 index 0000000000..5a17d8c51c --- /dev/null +++ b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T09:38:03.468532+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":2,"phase":"prefill","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":885.3639960289001,"p90":903.594970703125,"p95":913.1699800491333,"p99":929.2740225791931},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1035.3529453277588,"p90":1051.0159730911255,"p95":1054.6610355377197,"p99":1061.913013458252},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1920.716941356659,"p90":1954.6109437942505,"p95":1967.831015586853,"p99":1991.187036037445},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1881.1010122299194,"p90":1905.5780172348022,"p95":1911.5469455718994,"p99":1923.5550165176392},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1883.1050395965576,"p90":1916.5949821472168,"p95":1922.2029447555542,"p99":1930.575966835022},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":847.186028957367,"p90":856.8009734153748,"p95":858.722984790802,"p99":862.3690009117126},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":982.5140237808228,"p90":1013.5200023651123,"p95":1017.5260305404663,"p99":1028.3420085906982},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129026412963867,"p90":6.5299272537231445,"p95":6.5299272537231445,"p99":6.5299272537231445},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":31.796932220458984,"p90":47.782063484191895,"p95":53.99203300476074,"p99":64.5679235458374},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.9260387420654297,"p90":11.738061904907227,"p95":11.738061904907227,"p99":11.738061904907227},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":834.3719840049744,"p90":855.5989861488342,"p95":857.6020002365112,"p99":860.8059883117676},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":983.269989490509,"p90":1013.6350393295288,"p95":1018.0820226669312,"p99":1025.6129503250122},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1834.2620134353638,"p90":1867.2709465026855,"p95":1872.0380067825317,"p99":1880.0100088119507},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":47.520995140075684,"p90":57.61599540710449,"p95":60.581088066101074,"p99":65.34910202026367},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"roundtrip":{"activation_data_bytes":2428403712,"scale_bytes":0,"total_logical_bytes":2428403712},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"roundtrip":{"activation_data_bytes":2428403712,"scale_bytes":0,"total_logical_bytes":2428403712},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":84696,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":5367,"mean":5293.5,"min":5181,"total":84696},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.012361490167677402,"expert_assignments_per_rank":[6171,6137,6243,6050,6268,6146,6227,6063,6176,6175,6160,5993,6212,6144,6110,6029],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,5,241,2597,7671,5870,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16943359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.06316709518432617,"same_node_fraction":0.49927976727485657,"same_scaleup_domain_fraction":0.49927976727485657,"cross_node_fraction":0.500720202922821,"cross_domain_fraction":0.500720202922821,"gpus_per_node":8,"scale_up_domain":8,"copies":84696},"payload_copies_per_rank":[5326,5273,5366,5236,5367,5310,5348,5225,5315,5334,5305,5181,5318,5287,5280,5225],"payload_rank_cv":0.009882226586341858,"routed_copies":84696},"token_rate_at_latency_percentile":{"p50":8700523.685874773,"p90":8548493.631995495,"p95":8523553.688595325,"p99":8486586.532442886},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1451.5700340270996,"p90":1483.497977256775,"p95":1487.3440265655518,"p99":1493.0720329284668},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1379.4289827346802,"p90":1411.877989768982,"p95":1416.0840511322021,"p99":1421.170949935913},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":2830.99901676178,"p90":2895.375967025757,"p95":2903.428077697754,"p99":2914.24298286438},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":2795.1929569244385,"p90":2802.562952041626,"p95":2806.1680793762207,"p99":2816.570043563843},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2796.4351177215576,"p90":2827.867031097412,"p95":2832.192897796631,"p99":2837.12100982666},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1327.3500204086304,"p90":1334.7220420837402,"p95":1336.6049528121948,"p99":1341.5319919586182},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1265.2109861373901,"p90":1268.615961074829,"p95":1269.8580026626587,"p99":1272.3020315170288},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.689786911010742,"p90":6.730079650878906,"p95":6.730079650878906,"p99":6.730079650878906},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.274908065795898,"p90":33.39505195617676,"p95":38.4829044342041,"p99":43.87998580932617},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.2018680572509766,"p90":2.6841163635253906,"p95":2.6841163635253906,"p99":2.6841163635253906},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1408.1920385360718,"p90":1415.0429964065552,"p95":1417.165994644165,"p99":1420.2100038528442},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1346.7329740524292,"p90":1351.8610000610352,"p95":1352.9820442199707,"p99":1356.5069437026978},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2775.7089138031006,"p90":2779.93106842041,"p95":2781.0919284820557,"p99":2783.8170528411865},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":20.735979080200195,"p90":50.98104476928711,"p95":55.74798583984375,"p99":61.03706359863281},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"roundtrip":{"activation_data_bytes":4850069504,"scale_bytes":0,"total_logical_bytes":4850069504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"roundtrip":{"activation_data_bytes":4850069504,"scale_bytes":0,"total_logical_bytes":4850069504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":169157,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":10735,"mean":10572.3125,"min":10405,"total":169157},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.008082338608801365,"expert_assignments_per_rank":[12320,12304,12326,12170,12384,12230,12308,12211,12262,12359,12237,12047,12413,12243,12482,12312],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,16,505,5263,15346,11638,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.162261962890625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.0630006417632103,"same_node_fraction":0.5010611414909363,"same_scaleup_domain_fraction":0.5010611414909363,"cross_node_fraction":0.4989388585090637,"cross_domain_fraction":0.4989388585090637,"gpus_per_node":8,"scale_up_domain":8,"copies":169157},"payload_copies_per_rank":[10630,10585,10602,10505,10644,10564,10577,10498,10504,10645,10499,10405,10642,10518,10735,10604],"payload_rank_cv":0.00741199404001236,"routed_copies":169157},"token_rate_at_latency_percentile":{"p50":11717775.889861617,"p90":11587532.100928983,"p95":11569833.405589223,"p99":11549736.471058043},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2786.684989929199,"p90":2807.516098022461,"p95":2815.368890762329,"p99":2826.9050121307373},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2562.6449584960938,"p90":2589.2040729522705,"p95":2593.890905380249,"p99":2599.579095840454},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":5349.329948425293,"p90":5396.720170974731,"p95":5409.259796142578,"p99":5426.484107971191},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":5310.925006866455,"p90":5324.718952178955,"p95":5329.56600189209,"p99":5337.657928466797},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5321.713924407959,"p90":5347.821235656738,"p95":5353.632926940918,"p99":5365.220069885254},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2686.7899894714355,"p90":2720.118999481201,"p95":2724.9529361724854,"p99":2734.5008850097656},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2509.9918842315674,"p90":2536.510944366455,"p95":2540.11607170105,"p99":2547.940969467163},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.330013275146484,"p90":6.369113922119141,"p95":6.369113922119141,"p99":6.369113922119141},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":96.07696533203125,"p90":125.29373168945312,"p95":138.55504989624023,"p99":156.005859375},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.169319152832031,"p90":8.693218231201172,"p95":8.693218231201172,"p99":8.693218231201172},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2708.4240913391113,"p90":2725.6500720977783,"p95":2729.135036468506,"p99":2736.1440658569336},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2532.6249599456787,"p90":2551.187038421631,"p95":2555.752992630005,"p99":2562.99090385437},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5254.0130615234375,"p90":5273.953914642334,"p95":5277.915954589844,"p99":5286.342144012451},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":69.23484802246094,"p90":91.28379821777344,"p95":98.1588363647461,"p99":111.7558479309082},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"roundtrip":{"activation_data_bytes":9698332672,"scale_bytes":0,"total_logical_bytes":9698332672},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"roundtrip":{"activation_data_bytes":9698332672,"scale_bytes":0,"total_logical_bytes":9698332672},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":338251,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":21435,"mean":21140.6875,"min":20921,"total":338251},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.0070151351392269135,"expert_assignments_per_rank":[24582,24524,24711,24358,24680,24435,24611,24307,24504,24638,24478,24415,25004,24532,24846,24591],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,29,1048,10572,30561,23326,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1613006591796875,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.06308037787675858,"same_node_fraction":0.5009504556655884,"same_scaleup_domain_fraction":0.5009504556655884,"cross_node_fraction":0.49904951453208923,"cross_domain_fraction":0.49904951453208923,"gpus_per_node":8,"scale_up_domain":8,"copies":338251},"payload_copies_per_rank":[21145,21167,21213,20988,21198,21055,21170,20921,21054,21231,21048,21005,21435,21070,21396,21155],"payload_rank_cv":0.00635375315323472,"routed_copies":338251},"token_rate_at_latency_percentile":{"p50":12314829.570116527,"p90":12254710.30389666,"p95":12241407.077090634,"p99":12214969.590502115},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5603.655815124512,"p90":5630.415916442871,"p95":5639.589786529541,"p99":5657.1760177612305},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4955.913066864014,"p90":4980.910778045654,"p95":4987.884044647217,"p99":4997.375965118408},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":10559.568881988525,"p90":10611.326694488525,"p95":10627.473831176758,"p99":10654.551982879639},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":10518.494606018066,"p90":10550.729751586914,"p95":10556.818962097168,"p99":10581.068992614746},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10536.469459533691,"p90":10559.303283691406,"p95":10567.556381225586,"p99":10585.143089294434},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":5384.928226470947,"p90":5461.269855499268,"p95":5474.970817565918,"p99":5497.404098510742},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":4904.369831085205,"p90":4929.687976837158,"p95":4932.692050933838,"p99":4937.725067138672},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.37054443359375,"p90":6.8511962890625,"p95":6.8511962890625,"p99":6.8511962890625},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":306.3793182373047,"p90":410.76183319091797,"p95":439.77928161621094,"p99":503.8595199584961},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-14.741897583007812,"p90":26.11827850341797,"p95":26.11827850341797,"p99":26.11827850341797},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5388.012886047363,"p90":5453.859806060791,"p95":5469.886779785156,"p99":5497.045040130615},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4927.084922790527,"p90":4933.253765106201,"p95":4937.661170959473,"p99":4951.467037200928},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10330.45482635498,"p90":10400.025367736816,"p95":10416.089057922363,"p99":10446.390151977539},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":206.01463317871094,"p90":272.4485397338867,"p95":285.1238250732422,"p99":307.7392578125},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":131072,"byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"roundtrip":{"activation_data_bytes":19408105472,"scale_bytes":0,"total_logical_bytes":19408105472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"roundtrip":{"activation_data_bytes":19408105472,"scale_bytes":0,"total_logical_bytes":19408105472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":676901,"assignments":786432,"wire":"rank-deduplicated"},"receive":{"max":42692,"mean":42306.3125,"min":41888,"total":676901},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005902660544961691,"expert_assignments_per_rank":[49152,49158,48937,48637,49532,49147,49291,48640,48855,49080,49131,49180,49736,49110,49535,49311],"expert_load_cv":0.02157296985387802,"expert_load_max":2171,"expert_load_mean":2048.0,"expert_load_min":1922,"fanout_histogram":[0,51,2031,21028,61178,46784,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.164344787597656,"fanout_min":2,"hotspot_ratio":1.06005859375,"locality":{"placement":"packed","local_rank_fraction":0.06220702826976776,"same_node_fraction":0.5000066757202148,"same_scaleup_domain_fraction":0.5000066757202148,"cross_node_fraction":0.49999335408210754,"cross_domain_fraction":0.49999335408210754,"gpus_per_node":8,"scale_up_domain":8,"copies":676901},"payload_copies_per_rank":[42365,42433,42070,41888,42530,42296,42419,41908,42030,42386,42324,42267,42692,42172,42625,42496],"payload_rank_cv":0.005482310429215431,"routed_copies":676901},"token_rate_at_latency_percentile":{"p50":12439840.546531687,"p90":12412940.179721668,"p95":12403245.866079664,"p99":12382638.467359329},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"InterNodeV1","block_num":32,"rdma_block_num":16,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":0,"scale_type_size":0,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"inter-node-v1-vllm-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9943,"throttle_fraction_max":{"power":0.0007,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2089,"p50":2093,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0003,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2096,"p50":2099,"max":2100,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2094,"p50":2097,"max":2099,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2099,"p50":2101,"max":2102,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2089,"p50":2091,"max":2093,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2096,"p50":2099,"max":2102,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0003,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2099,"p50":2103,"max":2105,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0004,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2091,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0003,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2086,"p50":2093,"max":2098,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0007,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2097,"p50":2100,"max":2102,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2100,"p50":2103,"max":2104,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2091,"p50":2095,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2099,"p50":2101,"max":2103,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2094,"p50":2098,"max":2100,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2094,"p50":2097,"max":2099,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2096,"p50":2098,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":709.0}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8.json b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8.json new file mode 100644 index 0000000000..b8193b3971 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T10:19:39.171647+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":2,"phase":"prefill","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":894.2210078239441,"p90":910.8060002326965,"p95":921.1410284042358,"p99":935.3619813919067},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":782.1729779243469,"p90":796.3939905166626,"p95":800.1999855041504,"p99":809.4130158424377},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1676.393985748291,"p90":1707.1999907493591,"p95":1721.3410139083862,"p99":1744.7749972343445},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1629.2749643325806,"p90":1649.472951889038,"p95":1656.154990196228,"p99":1663.012981414795},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1647.8309631347656,"p90":1667.178988456726,"p95":1672.4669933319092,"p99":1681.6010475158691},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":228.42100262641907,"p90":231.62800073623657,"p95":232.8290045261383,"p99":235.51300168037415},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":855.7590246200562,"p90":863.7750148773193,"p95":866.2949800491333,"p99":870.1810240745544},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":729.0099859237671,"p90":749.9639987945557,"p95":754.7280192375183,"p99":761.218011379242},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.088972091674805,"p90":6.129026412963867,"p95":6.129026412963867,"p99":6.129026412963867},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":32.65702724456787,"p90":46.228885650634766,"p95":49.76189136505127,"p99":61.77198886871338},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-4.325985908508301,"p90":10.856986045837402,"p95":10.856986045837402,"p99":10.856986045837402},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":845.2680110931396,"p90":861.407995223999,"p95":863.7310266494751,"p99":867.3359751701355},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":735.45902967453,"p90":757.5719952583313,"p95":761.8589997291565,"p99":768.468976020813},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1594.5030450820923,"p90":1613.9320135116577,"p95":1619.500994682312,"p99":1628.113031387329},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":52.36506462097168,"p90":63.98296356201172,"p95":68.02904605865479,"p99":76.12097263336182},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.009900989942252636,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":607100928,"scale_bytes":18971904,"total_logical_bytes":626072832},"roundtrip":{"activation_data_bytes":1821302784,"scale_bytes":18971904,"total_logical_bytes":1840274688},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":607100928,"scale_bytes":18971904,"total_logical_bytes":626072832},"roundtrip":{"activation_data_bytes":1821302784,"scale_bytes":18971904,"total_logical_bytes":1840274688},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":84696,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":5367,"mean":5293.5,"min":5181,"total":84696},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.012361490167677402,"expert_assignments_per_rank":[6171,6137,6243,6050,6268,6146,6227,6063,6176,6175,6160,5993,6212,6144,6110,6029],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,5,241,2597,7671,5870,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16943359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.06316709518432617,"same_node_fraction":0.49927976727485657,"same_scaleup_domain_fraction":0.49927976727485657,"cross_node_fraction":0.500720202922821,"cross_domain_fraction":0.500720202922821,"gpus_per_node":8,"scale_up_domain":8,"copies":84696},"payload_copies_per_rank":[5326,5273,5366,5236,5367,5310,5348,5225,5315,5334,5305,5181,5318,5287,5280,5225],"payload_rank_cv":0.009882226586341858,"routed_copies":84696},"token_rate_at_latency_percentile":{"p50":9942767.411550365,"p90":9827379.131718988,"p95":9796306.931809515,"p99":9743095.738554114},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1473.122000694275,"p90":1501.7249584197998,"p95":1510.9390020370483,"p99":1531.169056892395},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1110.9460592269897,"p90":1132.6580047607422,"p95":1140.0699615478516,"p99":1150.244951248169},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":2584.0680599212646,"p90":2634.382963180542,"p95":2651.0089635849,"p99":2681.414008140564},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":2520.848035812378,"p90":2545.4559326171875,"p95":2552.57511138916,"p99":2564.673900604248},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2538.9950275421143,"p90":2559.838056564331,"p95":2567.3298835754395,"p99":2581.4990997314453},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":449.23099875450134,"p90":453.95898818969727,"p95":455.68200945854187,"p99":459.646999835968},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1387.7209424972534,"p90":1433.0220222473145,"p95":1437.034010887146,"p99":1440.7600164413452},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1021.7670202255249,"p90":1066.0779476165771,"p95":1075.6880044937134,"p99":1094.601035118103},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.570100784301758,"p90":6.690025329589844,"p95":6.690025329589844,"p99":6.690025329589844},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":56.16402626037598,"p90":87.07904815673828,"p95":102.8449535369873,"p99":116.78695678710938},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-7.130861282348633,"p90":13.820886611938477,"p95":13.820886611938477,"p99":13.820886611938477},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1411.3960266113281,"p90":1428.976058959961,"p95":1430.9790134429932,"p99":1434.4240427017212},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1058.9029788970947,"p90":1085.461974143982,"p95":1092.5129652023315,"p99":1101.410984992981},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2495.8109855651855,"p90":2514.8510932922363,"p95":2520.258903503418,"p99":2528.6710262298584},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":42.01197624206543,"p90":63.36402893066406,"p95":72.85690307617188,"p99":90.80386161804199},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.011583011597394943,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":1212517376,"scale_bytes":37891168,"total_logical_bytes":1250408544},"roundtrip":{"activation_data_bytes":3637552128,"scale_bytes":37891168,"total_logical_bytes":3675443296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":1212517376,"scale_bytes":37891168,"total_logical_bytes":1250408544},"roundtrip":{"activation_data_bytes":3637552128,"scale_bytes":37891168,"total_logical_bytes":3675443296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":169157,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":10735,"mean":10572.3125,"min":10405,"total":169157},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.008082338608801365,"expert_assignments_per_rank":[12320,12304,12326,12170,12384,12230,12308,12211,12262,12359,12237,12047,12413,12243,12482,12312],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,16,505,5263,15346,11638,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.162261962890625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.0630006417632103,"same_node_fraction":0.5010611414909363,"same_scaleup_domain_fraction":0.5010611414909363,"cross_node_fraction":0.4989388585090637,"cross_domain_fraction":0.4989388585090637,"gpus_per_node":8,"scale_up_domain":8,"copies":169157},"payload_copies_per_rank":[10630,10585,10602,10505,10644,10564,10577,10498,10504,10645,10499,10405,10642,10518,10735,10604],"payload_rank_cv":0.00741199404001236,"routed_copies":169157},"token_rate_at_latency_percentile":{"p50":12905893.727456888,"p90":12800809.76840361,"p95":12763455.218448607,"p99":12693399.739480395},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2816.049098968506,"p90":2849.6599197387695,"p95":2861.9980812072754,"p99":2884.9918842315674},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1765.3980255126953,"p90":1793.1690216064453,"p95":1804.2969703674316,"p99":1829.6630382537842},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":4581.447124481201,"p90":4642.828941345215,"p95":4666.295051574707,"p99":4714.654922485352},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":4503.692150115967,"p90":4532.7348709106445,"p95":4540.447235107422,"p99":4558.5150718688965},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4543.790817260742,"p90":4573.29797744751,"p95":4581.087112426758,"p99":4597.911834716797},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":917.7709817886353,"p90":924.2209792137146,"p95":926.1090159416199,"p99":930.7109713554382},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2666.7590141296387,"p90":2716.1519527435303,"p95":2724.1640090942383,"p99":2738.8670444488525},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1657.5169563293457,"p90":1719.7309732437134,"p95":1730.3869724273682,"p99":1751.0650157928467},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.729602813720703,"p90":7.129669189453125,"p95":7.129669189453125,"p99":7.129669189453125},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":123.20518493652344,"p90":171.4777946472168,"p95":190.94610214233398,"p99":214.9810791015625},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-10.69498062133789,"p90":13.980865478515625,"p95":13.980865478515625,"p99":13.980865478515625},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2704.694986343384,"p90":2729.5329570770264,"p95":2735.6090545654297,"p99":2747.2269535064697},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1724.7380018234253,"p90":1749.8559951782227,"p95":1757.2740316390991,"p99":1771.286964416504},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4454.537868499756,"p90":4482.7399253845215,"p95":4491.192817687988,"p99":4504.171848297119},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":89.47324752807617,"p90":121.3231086730957,"p95":132.29799270629883,"p99":152.88972854614258},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":2424583168,"scale_bytes":75768224,"total_logical_bytes":2500351392},"roundtrip":{"activation_data_bytes":7273749504,"scale_bytes":75768224,"total_logical_bytes":7349517728},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":2424583168,"scale_bytes":75768224,"total_logical_bytes":2500351392},"roundtrip":{"activation_data_bytes":7273749504,"scale_bytes":75768224,"total_logical_bytes":7349517728},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":338251,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":21435,"mean":21140.6875,"min":20921,"total":338251},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.0070151351392269135,"expert_assignments_per_rank":[24582,24524,24711,24358,24680,24435,24611,24307,24504,24638,24478,24415,25004,24532,24846,24591],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,29,1048,10572,30561,23326,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1613006591796875,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.06308037787675858,"same_node_fraction":0.5009504556655884,"same_scaleup_domain_fraction":0.5009504556655884,"cross_node_fraction":0.49904951453208923,"cross_domain_fraction":0.49904951453208923,"gpus_per_node":8,"scale_up_domain":8,"copies":338251},"payload_copies_per_rank":[21145,21167,21213,20988,21198,21055,21170,20921,21054,21231,21048,21005,21435,21070,21396,21155],"payload_rank_cv":0.00635375315323472,"routed_copies":338251},"token_rate_at_latency_percentile":{"p50":14423199.18228737,"p90":14330139.939094357,"p95":14305774.675671548,"p99":14253426.850242468},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5673.800945281982,"p90":5724.717140197754,"p95":5740.500926971436,"p99":5774.230003356934},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3097.8450775146484,"p90":3135.901927947998,"p95":3148.279905319214,"p99":3183.573007583618},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":8771.64602279663,"p90":8860.619068145752,"p95":8888.78083229065,"p99":8957.803010940552},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":8655.898094177246,"p90":8693.513870239258,"p95":8702.366828918457,"p99":8725.321769714355},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8717.309951782227,"p90":8756.367683410645,"p95":8768.787384033203,"p99":8794.745445251465},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1969.5930480957031,"p90":1999.1580247879028,"p95":2005.4380893707275,"p99":2017.9760456085205},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":5411.873817443848,"p90":5488.763809204102,"p95":5509.300231933594,"p99":5524.6429443359375},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2981.6160202026367,"p90":3031.4650535583496,"p95":3045.7260608673096,"p99":3072.2720623016357},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.4105987548828125,"p90":6.649971008300781,"p95":6.649971008300781,"p99":6.649971008300781},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":320.2400207519531,"p90":472.38826751708984,"p95":498.90708923339844,"p99":559.6380233764648},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-25.87890625,"p90":22.874832153320312,"p95":22.874832153320312,"p99":22.874832153320312},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5423.332214355469,"p90":5496.922016143799,"p95":5510.396957397461,"p99":5536.500930786133},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3038.6219024658203,"p90":3073.68803024292,"p95":3080.7390213012695,"p99":3099.993944168091},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8497.30110168457,"p90":8574.776649475098,"p95":8592.122077941895,"p99":8625.17261505127},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":221.5719223022461,"p90":321.40254974365234,"p95":339.22672271728516,"p99":371.75655364990234},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01253132801502943,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":131072,"byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":4852026368,"scale_bytes":151625824,"total_logical_bytes":5003652192},"roundtrip":{"activation_data_bytes":14556079104,"scale_bytes":151625824,"total_logical_bytes":14707704928},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":4852026368,"scale_bytes":151625824,"total_logical_bytes":5003652192},"roundtrip":{"activation_data_bytes":14556079104,"scale_bytes":151625824,"total_logical_bytes":14707704928},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":676901,"assignments":786432,"wire":"rank-deduplicated"},"receive":{"max":42692,"mean":42306.3125,"min":41888,"total":676901},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005902660544961691,"expert_assignments_per_rank":[49152,49158,48937,48637,49532,49147,49291,48640,48855,49080,49131,49180,49736,49110,49535,49311],"expert_load_cv":0.02157296985387802,"expert_load_max":2171,"expert_load_mean":2048.0,"expert_load_min":1922,"fanout_histogram":[0,51,2031,21028,61178,46784,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.164344787597656,"fanout_min":2,"hotspot_ratio":1.06005859375,"locality":{"placement":"packed","local_rank_fraction":0.06220702826976776,"same_node_fraction":0.5000066757202148,"same_scaleup_domain_fraction":0.5000066757202148,"cross_node_fraction":0.49999335408210754,"cross_domain_fraction":0.49999335408210754,"gpus_per_node":8,"scale_up_domain":8,"copies":676901},"payload_copies_per_rank":[42365,42433,42070,41888,42530,42296,42419,41908,42030,42386,42324,42267,42692,42172,42625,42496],"payload_rank_cv":0.005482310429215431,"routed_copies":676901},"token_rate_at_latency_percentile":{"p50":15035831.090668373,"p90":14968763.84580357,"p95":14947562.788289826,"p99":14903444.427805416},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"InterNodeV1","block_num":32,"rdma_block_num":16,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":56,"scale_type_size":4,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"inter-node-v1-vllm-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9895,"throttle_fraction_max":{"power":0.0941,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":true,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":1915,"p50":2068,"max":2088,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0898,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":1937,"p50":2079,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0871,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":1937,"p50":2079,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0861,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":1961,"p50":2086,"max":2099,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0783,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":1928,"p50":2072,"max":2088,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0884,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":1927,"p50":2076,"max":2095,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.088,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":1926,"p50":2077,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0916,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":1916,"p50":2066,"max":2087,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0912,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":1907,"p50":2064,"max":2089,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0941,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":1940,"p50":2079,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0833,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":1937,"p50":2082,"max":2098,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0864,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":1932,"p50":2074,"max":2092,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0829,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":1930,"p50":2079,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0895,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":1945,"p50":2081,"max":2095,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0809,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":1928,"p50":2074,"max":2093,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0886,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":1941,"p50":2079,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0848,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":872.6}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16.json b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16.json new file mode 100644 index 0000000000..08b95c7336 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T06:58:06.151588+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":1,"phase":"prefill","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":278.0529856681824,"p90":282.2990119457245,"p95":283.7409973144531,"p99":286.62601113319397},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":220.48799693584442,"p90":222.8109985589981,"p95":223.8530069589615,"p99":226.25599801540375},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":498.5409826040268,"p90":505.1100105047226,"p95":507.5940042734146,"p99":512.8820091485977},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":488.6859953403473,"p90":493.2129979133606,"p95":494.37499046325684,"p99":498.4210133552551},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":491.129994392395,"p90":494.9750006198883,"p95":496.2579905986786,"p99":499.7420012950897},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":273.08499813079834,"p90":277.37200260162354,"p95":278.57398986816406,"p99":282.2990119457245},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":217.68300235271454,"p90":220.36699950695038,"p95":220.92799842357635,"p99":221.92899882793427},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":6.208986043930054,"p95":6.208986043930054,"p99":6.208986043930054},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.884000539779663,"p90":7.530003786087036,"p95":8.892983198165894,"p99":11.1369788646698},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.840991735458374,"p90":1.6030073165893555,"p95":1.6030073165893555,"p99":1.6030073165893555},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":274.167001247406,"p90":277.89199352264404,"p95":279.4950008392334,"p99":282.8199863433838},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":215.5199944972992,"p90":217.48299896717072,"p95":218.0439978837967,"p99":219.2859947681427},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":485.3610098361969,"p90":489.2460107803345,"p95":490.52900075912476,"p99":493.7739968299866},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.286991596221924,"p90":8.652985095977783,"p95":10.015010833740234,"p99":12.178003787994385},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"roundtrip":{"activation_data_bytes":1041510400,"scale_bytes":0,"total_logical_bytes":1041510400},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"roundtrip":{"activation_data_bytes":1041510400,"scale_bytes":0,"total_logical_bytes":1041510400},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":36325,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":4587,"mean":4540.625,"min":4510,"total":36325},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.009137080051004887,"expert_assignments_per_rank":[6159,6057,6238,6119,6110,6118,6221,6130],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,46,880,3426,3151,689,0,0],"fanout_max":6,"fanout_mean":4.4342041015625,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.12393668293952942,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":36325},"payload_copies_per_rank":[4582,4510,4587,4547,4516,4535,4533,4515],"payload_rank_cv":0.006128025706857443,"routed_copies":36325},"token_rate_at_latency_percentile":{"p50":16679901.642201249,"p90":16550330.804062111,"p95":16507542.760404298,"p99":16392458.466109106},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":497.4589943885803,"p90":506.9130063056946,"p95":510.95902919769287,"p99":519.8519825935364},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":414.57599401474,"p90":417.8209900856018,"p95":418.8219904899597,"p99":421.3449954986572},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":912.0349884033203,"p90":924.7339963912964,"p95":929.7810196876526,"p99":941.1969780921936},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":899.0160226821899,"p90":906.9070219993591,"p95":909.7110033035278,"p99":915.120005607605},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":904.583990573883,"p90":913.7570261955261,"p95":917.9229736328125,"p99":926.9769787788391},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":490.60899019241333,"p90":497.4580109119415,"p95":500.5840063095093,"p99":508.1949830055237},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":410.08999943733215,"p90":412.8130078315735,"p95":413.69500756263733,"p99":415.7780110836029},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.08903169631958,"p90":6.129026412963867,"p95":6.129026412963867,"p99":6.129026412963867},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.3649802207946777,"p90":7.012009620666504,"p95":8.572995662689209,"p99":10.895013809204102},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.8810162544250488,"p90":1.8430352210998535,"p95":1.8430352210998535,"p99":1.8430352210998535},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":493.4529960155487,"p90":502.5460124015808,"p95":506.8730115890503,"p99":514.9250030517578},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":409.20698642730713,"p90":412.7329885959625,"p95":413.5740101337433,"p99":415.0960147380829},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":898.9350199699402,"p90":908.1090092658997,"p95":911.7940068244934,"p99":921.7290282249451},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.286991596221924,"p90":8.333027362823486,"p95":9.894967079162598,"p99":12.099027633666992},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"roundtrip":{"activation_data_bytes":2083221504,"scale_bytes":0,"total_logical_bytes":2083221504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"roundtrip":{"activation_data_bytes":2083221504,"scale_bytes":0,"total_logical_bytes":2083221504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":72657,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":9142,"mean":9082.125,"min":9007,"total":72657},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00735377287492156,"expert_assignments_per_rank":[12308,12293,12414,12290,12351,12153,12356,12139],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,92,1754,6798,6421,1319,0,0],"fanout_max":6,"fanout_mean":4.43463134765625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.12587912380695343,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":72657},"payload_copies_per_rank":[9133,9070,9142,9078,9120,9057,9050,9007],"payload_rank_cv":0.004780265036970377,"routed_copies":72657},"token_rate_at_latency_percentile":{"p50":18112193.196792837,"p90":17930368.28205373,"p95":17848992.203733563,"p99":17674656.84162254},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1001.6880035400391,"p90":1015.7890319824219,"p95":1020.4349756240845,"p99":1028.8889408111572},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":807.3999881744385,"p90":812.1259808540344,"p95":813.5679960250854,"p99":816.1320090293884},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1809.0879917144775,"p90":1827.9150128364563,"p95":1834.00297164917,"p99":1845.0209498405457},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1795.4670190811157,"p90":1809.607982635498,"p95":1814.3349885940552,"p99":1827.5549411773682},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1801.516056060791,"p90":1816.0580396652222,"p95":1821.0649490356445,"p99":1829.4769525527954},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":994.2770004272461,"p90":1008.5779428482056,"p95":1013.8260126113892,"p99":1024.4020223617554},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":801.7110228538513,"p90":806.4780235290527,"p95":807.8799843788147,"p99":811.1249804496765},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.25002384185791,"p90":6.5299272537231445,"p95":6.5299272537231445,"p99":6.5299272537231445},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.892012596130371,"p90":12.256979942321777,"p95":14.140963554382324,"p99":17.14491844177246},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.5659074783325195,"p90":5.768060684204102,"p95":5.768060684204102,"p99":5.768060684204102},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":994.517982006073,"p90":1008.6580514907837,"p95":1012.2640132904053,"p99":1020.8760499954224},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":801.9520044326782,"p90":806.8789839744568,"p95":808.4009885787964,"p99":811.1640214920044},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1790.8999919891357,"p90":1804.92103099823,"p95":1810.1290464401245,"p99":1818.4210062026978},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.655999183654785,"p90":13.86106014251709,"p95":15.022993087768555,"p99":17.7459716796875},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"roundtrip":{"activation_data_bytes":4160278528,"scale_bytes":0,"total_logical_bytes":4160278528},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"roundtrip":{"activation_data_bytes":4160278528,"scale_bytes":0,"total_logical_bytes":4160278528},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":145099,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":18284,"mean":18137.375,"min":17991,"total":145099},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005656693130731583,"expert_assignments_per_rank":[24624,24496,24614,24519,24621,24284,24656,24794],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,192,3636,13520,12793,2627,0,0],"fanout_max":6,"fanout_mean":4.428070068359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.125583216547966,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":145099},"payload_copies_per_rank":[18250,18101,18151,18109,18125,17991,18088,18284],"payload_rank_cv":0.004795215092599392,"routed_copies":145099},"token_rate_at_latency_percentile":{"p50":18189124.593011267,"p90":18043476.190904427,"p95":17993866.730207775,"p99":17911130.257353913},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2008.6228847503662,"p90":2042.4740314483643,"p95":2052.968978881836,"p99":2082.7341079711914},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1596.371054649353,"p90":1603.5419702529907,"p95":1605.8249473571777,"p99":1611.433982849121},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":3604.9939393997192,"p90":3646.016001701355,"p95":3658.7939262390137,"p99":3694.1680908203125},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":3587.409019470215,"p90":3621.1390495300293,"p95":3629.9519538879395,"p99":3671.293020248413},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3597.583055496216,"p90":3632.275104522705,"p95":3648.0588912963867,"p99":3678.4238815307617},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1996.3250160217285,"p90":2031.1369895935059,"p95":2043.7560081481934,"p99":2068.4731006622314},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1584.7140550613403,"p90":1594.6489572525024,"p95":1599.735975265503,"p99":1608.7900400161743},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.248950958251953,"p90":6.489038467407227,"p95":6.489038467407227,"p99":6.489038467407227},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.339996337890625,"p90":19.0279483795166,"p95":20.78986167907715,"p99":24.35612678527832},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-7.812023162841797,"p90":8.972883224487305,"p95":8.972883224487305,"p99":8.972883224487305},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1993.6410188674927,"p90":2027.2510051727295,"p95":2038.9480590820312,"p99":2067.8319931030273},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1590.522050857544,"p90":1598.2539653778076,"p95":1600.6170511245728,"p99":1605.7850122451782},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3579.7970294952393,"p90":3615.0898933410645,"p95":3630.2330493927,"p99":3659.6360206604004},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.865896224975586,"p90":22.11284637451172,"p95":23.715972900390625,"p99":26.278972625732422},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"roundtrip":{"activation_data_bytes":8316542976,"scale_bytes":0,"total_logical_bytes":8316542976},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"roundtrip":{"activation_data_bytes":8316542976,"scale_bytes":0,"total_logical_bytes":8316542976},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":290058,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":36460,"mean":36257.25,"min":36144,"total":290058},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00432177959010005,"expert_assignments_per_rank":[49106,49069,49115,48918,49142,48893,49536,49437],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,413,7235,27175,25451,5262,0,0],"fanout_max":6,"fanout_mean":4.425933837890625,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.1256852149963379,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":290058},"payload_copies_per_rank":[36354,36201,36216,36144,36222,36193,36268,36460],"payload_rank_cv":0.0026462646201252937,"routed_copies":290058},"token_rate_at_latency_percentile":{"p50":18216674.636566687,"p90":18042686.226711806,"p95":17964622.269765746,"p99":17816326.261107095},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"IntraNode","block_num":80,"rdma_block_num":0,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":0,"scale_type_size":0,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"intranode-vllm-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi300x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9934,"throttle_fraction_max":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2079,"p50":2104,"max":2106,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":78,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":270.0},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2081,"p50":2097,"max":2099,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":270.0},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2074,"p50":2097,"max":2099,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":73,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":270.0},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2076,"p50":2094,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":270.0},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2082,"p50":2101,"max":2103,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":76,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":270.0},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2079,"p50":2103,"max":2105,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":270.0},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2082,"p50":2104,"max":2107,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":74,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":270.0},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2108,"max":2110,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":270.0}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8.json b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8.json new file mode 100644 index 0000000000..2e700ebdcf --- /dev/null +++ b/collectivex/results/mi300x/mi300x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T07:00:20.154615+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":1,"phase":"prefill","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":277.8150141239166,"p90":281.61999583244324,"p95":283.1830084323883,"p99":286.3079905509949},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":141.21100306510925,"p90":143.77500116825104,"p95":144.69599723815918,"p99":146.77800238132477},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":419.0260171890259,"p90":425.3949970006943,"p95":427.8790056705475,"p99":433.08599293231964},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":409.4510078430176,"p90":413.7380123138428,"p95":415.461003780365,"p99":419.62701082229614},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":412.53599524497986,"p90":416.90200567245483,"p95":418.4249937534332,"p99":421.3089942932129},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":203.3039927482605,"p90":206.8289965391159,"p95":207.99100399017334,"p99":209.99400317668915},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":273.00798892974854,"p90":277.0940065383911,"p95":278.49599719047546,"p99":281.9010019302368},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":137.64600455760956,"p90":140.80999791622162,"p95":141.3320004940033,"p99":142.6129937171936},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.288975477218628,"p95":6.288975477218628,"p99":6.288975477218628},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.6440024375915527,"p90":6.36899471282959,"p95":8.053004741668701,"p99":11.577010154724121},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.0430216789245605,"p90":1.8030107021331787,"p95":1.8030107021331787,"p99":1.8030107021331787},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":274.24898743629456,"p90":277.6539921760559,"p95":278.8560092449188,"p99":282.6620042324066},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":137.0050013065338,"p90":139.16699588298798,"p95":139.8089975118637,"p99":141.21100306510925},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":407.1280062198639,"p90":411.4550054073334,"p95":412.977010011673,"p99":416.10100865364075},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.086988210678101,"p90":8.45298171043396,"p95":9.894013404846191,"p99":13.058990240097046},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.007518797181546688,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":260377600,"scale_bytes":8136800,"total_logical_bytes":268514400},"roundtrip":{"activation_data_bytes":781132800,"scale_bytes":8136800,"total_logical_bytes":789269600},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":260377600,"scale_bytes":8136800,"total_logical_bytes":268514400},"roundtrip":{"activation_data_bytes":781132800,"scale_bytes":8136800,"total_logical_bytes":789269600},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":36325,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":4587,"mean":4540.625,"min":4510,"total":36325},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.009137080051004887,"expert_assignments_per_rank":[6159,6057,6238,6119,6110,6118,6221,6130],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,46,880,3426,3151,689,0,0],"fanout_max":6,"fanout_mean":4.4342041015625,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.12393668293952942,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":36325},"payload_copies_per_rank":[4582,4510,4587,4547,4516,4535,4533,4515],"payload_rank_cv":0.006128025706857443,"routed_copies":36325},"token_rate_at_latency_percentile":{"p50":19857661.136054985,"p90":19649701.58103812,"p95":19578180.372340113,"p99":19444161.199888177},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":502.7509927749634,"p90":521.49897813797,"p95":530.1520228385925,"p99":550.8220195770264},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":249.41200017929077,"p90":252.49600410461426,"p95":253.73899936676025,"p99":256.02200627326965},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":752.1629929542542,"p90":773.9949822425842,"p95":783.8910222053528,"p99":806.844025850296},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":742.027997970581,"p90":763.1800174713135,"p95":772.3940014839172,"p99":793.8259840011597},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":743.1899905204773,"p90":764.2210125923157,"p95":774.7970223426819,"p99":798.5929846763611},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":400.3570079803467,"p90":405.5649936199188,"p95":407.1280062198639,"p99":410.6929898262024},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":497.3829984664917,"p90":521.2979912757874,"p95":531.9539904594421,"p99":544.8139905929565},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":244.04500424861908,"p90":246.288001537323,"p95":246.8090057373047,"p99":248.17100167274475},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.289005279541016,"p90":6.449997425079346,"p95":6.449997425079346,"p99":6.449997425079346},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.4450292587280273,"p90":7.3310136795043945,"p95":8.53198766708374,"p99":11.377990245819092},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.9630193710327148,"p90":4.807949066162109,"p95":4.807949066162109,"p99":4.807949066162109},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":498.54400753974915,"p90":517.4930095672607,"p95":525.7850289344788,"p99":545.494019985199},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":244.8049932718277,"p90":247.8100061416626,"p95":248.77099692821503,"p99":250.6139874458313},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":737.621009349823,"p90":758.6529850959778,"p95":769.1890001296997,"p99":793.8240170478821},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.20700216293335,"p90":8.491992950439453,"p95":9.895026683807373,"p99":12.338995933532715},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":520805376,"scale_bytes":16275168,"total_logical_bytes":537080544},"roundtrip":{"activation_data_bytes":1562416128,"scale_bytes":16275168,"total_logical_bytes":1578691296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":520805376,"scale_bytes":16275168,"total_logical_bytes":537080544},"roundtrip":{"activation_data_bytes":1562416128,"scale_bytes":16275168,"total_logical_bytes":1578691296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":72657,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":9142,"mean":9082.125,"min":9007,"total":72657},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00735377287492156,"expert_assignments_per_rank":[12308,12293,12414,12290,12351,12153,12356,12139],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,92,1754,6798,6421,1319,0,0],"fanout_max":6,"fanout_mean":4.43463134765625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.12587912380695343,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":72657},"payload_copies_per_rank":[9133,9070,9142,9078,9120,9057,9050,9007],"payload_rank_cv":0.004780265036970377,"routed_copies":72657},"token_rate_at_latency_percentile":{"p50":22045506.813844215,"p90":21438824.27993415,"p95":21146183.48746517,"p99":20516083.054047618},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1007.7430009841919,"p90":1024.2880582809448,"p95":1029.3760299682617,"p99":1044.6380376815796},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":473.74698519706726,"p90":477.55300998687744,"p95":478.7150025367737,"p99":481.2780022621155},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1481.4899861812592,"p90":1501.8410682678223,"p95":1508.0910325050354,"p99":1525.916039943695},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1464.5880460739136,"p90":1479.1699647903442,"p95":1483.9359521865845,"p99":1498.1989860534668},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1472.838044166565,"p90":1489.0220165252686,"p95":1494.6329593658447,"p99":1505.486011505127},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":779.2019844055176,"p90":785.772979259491,"p95":787.3749732971191,"p99":791.6610240936279},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":997.0890283584595,"p90":1013.0339860916138,"p95":1018.2009935379028,"p99":1026.492953300476},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":460.04700660705566,"p90":465.2149975299835,"p95":466.856986284256,"p99":469.46099400520325},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.0890913009643555,"p90":6.410002708435059,"p95":6.410002708435059,"p99":6.410002708435059},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11.81793212890625,"p90":17.064929008483887,"p95":18.587946891784668,"p99":22.272944450378418},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.3249855041503906,"p90":8.453011512756348,"p95":8.453011512756348,"p99":8.453011512756348},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1001.4139413833618,"p90":1016.3969993591309,"p95":1021.8460559844971,"p99":1036.707043647766},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":468.89999508857727,"p90":472.463995218277,"p95":473.347008228302,"p99":475.9899973869324},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1463.8659954071045,"p90":1479.248046875,"p95":1484.616994857788,"p99":1497.6749420166016},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.733034133911133,"p90":13.019084930419922,"p95":14.220952987670898,"p99":16.66402816772461},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":1040069632,"scale_bytes":32502176,"total_logical_bytes":1072571808},"roundtrip":{"activation_data_bytes":3120208896,"scale_bytes":32502176,"total_logical_bytes":3152711072},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":1040069632,"scale_bytes":32502176,"total_logical_bytes":1072571808},"roundtrip":{"activation_data_bytes":3120208896,"scale_bytes":32502176,"total_logical_bytes":3152711072},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":145099,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":18284,"mean":18137.375,"min":17991,"total":145099},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005656693130731583,"expert_assignments_per_rank":[24624,24496,24614,24519,24621,24284,24656,24794],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,192,3636,13520,12793,2627,0,0],"fanout_max":6,"fanout_mean":4.428070068359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.125583216547966,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":145099},"payload_copies_per_rank":[18250,18101,18151,18109,18125,17991,18088,18284],"payload_rank_cv":0.004795215092599392,"routed_copies":145099},"token_rate_at_latency_percentile":{"p50":22248203.140721038,"p90":22006390.527701057,"p95":21923777.202065103,"p99":21765728.641503494},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2005.155086517334,"p90":2041.6059494018555,"p95":2056.2710762023926,"p99":2086.195945739746},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":910.6400012969971,"p90":916.4490103721619,"p95":918.0110096931458,"p99":921.3759899139404},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":2915.795087814331,"p90":2958.0549597740173,"p95":2974.2820858955383,"p99":3007.5719356536865},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":2905.9009552001953,"p90":2944.7989463806152,"p95":2964.2279148101807,"p99":2984.4579696655273},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2908.1430435180664,"p90":2945.3558921813965,"p95":2961.3380432128906,"p99":2986.375093460083},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1567.5779581069946,"p90":1576.3520002365112,"p95":1579.3559551239014,"p99":1584.0859413146973},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1999.4269609451294,"p90":2034.2390537261963,"p95":2048.340082168579,"p99":2068.0088996887207},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":903.1490087509155,"p90":909.2379808425903,"p95":910.7599854469299,"p99":914.3660068511963},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.690025329589844,"p90":7.091045379638672,"p95":7.091045379638672,"p99":7.091045379638672},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11.977910995483398,"p90":16.585111618041992,"p95":18.507003784179688,"p99":21.272897720336914},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-6.6890716552734375,"p90":3.5250186920166016,"p95":3.5250186920166016,"p99":3.5250186920166016},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1996.819019317627,"p90":2032.0351123809814,"p95":2046.5331077575684,"p99":2076.781988143921},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":905.7130217552185,"p90":911.3219976425171,"p95":913.0420088768005,"p99":916.7270064353943},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2897.9270458221436,"p90":2935.4989528656006,"p95":2951.2081146240234,"p99":2977.682113647461},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.774923324584961,"p90":13.980865478515625,"p95":15.342950820922852,"p99":18.22686195373535},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":2079135744,"scale_bytes":64972992,"total_logical_bytes":2144108736},"roundtrip":{"activation_data_bytes":6237407232,"scale_bytes":64972992,"total_logical_bytes":6302380224},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":2079135744,"scale_bytes":64972992,"total_logical_bytes":2144108736},"roundtrip":{"activation_data_bytes":6237407232,"scale_bytes":64972992,"total_logical_bytes":6302380224},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":290058,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":36460,"mean":36257.25,"min":36144,"total":290058},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00432177959010005,"expert_assignments_per_rank":[49106,49069,49115,48918,49142,48893,49536,49437],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,413,7235,27175,25451,5262,0,0],"fanout_max":6,"fanout_mean":4.425933837890625,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.1256852149963379,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":290058},"payload_copies_per_rank":[36354,36201,36216,36144,36222,36193,36268,36460],"payload_rank_cv":0.0026462646201252937,"routed_copies":290058},"token_rate_at_latency_percentile":{"p50":22535342.663447246,"p90":22250621.792079113,"p95":22130536.616784558,"p99":21944999.522504214},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"IntraNode","block_num":80,"rdma_block_num":0,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":56,"scale_type_size":4,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"intranode-vllm-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi300x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9932,"throttle_fraction_max":{"power":0.1607,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":true,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":1840,"p50":2046,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":70,"clock_event_reasons":[],"throttle_fraction":{"power":0.1607,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":354.0},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":1876,"p50":2055,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.1385,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":354.0},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":1840,"p50":2043,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":71,"clock_event_reasons":[],"throttle_fraction":{"power":0.1581,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":354.0},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":1844,"p50":2041,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.15,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":354.0},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":1851,"p50":2049,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":69,"clock_event_reasons":[],"throttle_fraction":{"power":0.1509,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":354.0},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":1881,"p50":2054,"max":2098,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.1316,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":354.0},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":1837,"p50":2041,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":72,"clock_event_reasons":[],"throttle_fraction":{"power":0.1586,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":354.0},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":1858,"p50":2048,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.1447,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":354.0}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-swap-blocks-smoke-contiguous.json b/collectivex/results/mi300x/mi300x-swap-blocks-smoke-contiguous.json new file mode 100644 index 0000000000..d3c9df7d81 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-swap-blocks-smoke-contiguous.json @@ -0,0 +1,4075 @@ +{ + "schema": "collectivex-swap-blocks-v1", + "case_id": "mi300x-swap-blocks-smoke-contiguous", + "profile": "smoke", + "generated_at": "2026-10-28T08:29:15.597236+00:00", + "operation": "vllm._custom_ops.swap_blocks", + "timing": "drained-wall-clock-including-submission-and-synchronization", + "warmup": 4, + "iterations": 20, + "selection": { + "requested_block_bytes": [ + 257, + 4096, + 65536, + 262144 + ], + "requested_num_blocks": [ + 1, + 4, + 16, + 64, + 256, + 1024, + 2048 + ], + "max_payload_bytes": 2147483648, + "skipped_cases": [] + }, + "runtime": { + "sku": "mi300x", + "torch": "2.13.0+git733fca1", + "vllm": "0.32.0", + "cuda": null, + "hip": "7.2.53211", + "device": "", + "device_index": 0, + "image": "vllm/vllm-openai-rocm:v0.32.0", + "source_sha": "6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4", + "source_dirty": false + }, + "cases": [ + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 257, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 16.515, + "p90": 17.185, + "p95": 17.216, + "p99": 22.694 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.015561610656978505, + "p90": 0.014954902531277276, + "p95": 0.014927973977695167, + "p99": 0.011324579183925267 + }, + "samples_us": [ + 17.216, + 16.955, + 17.185, + 16.304, + 16.815, + 16.475, + 16.735, + 16.134, + 16.555, + 16.735, + 16.445, + 16.956, + 16.565, + 16.335, + 22.694, + 16.515, + 16.494, + 16.304, + 16.124, + 16.184 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1028, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 25.238, + "p90": 25.768, + "p95": 26.128, + "p99": 28.362 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.04073222917822331, + "p90": 0.039894442719652284, + "p95": 0.03934476423759951, + "p99": 0.03624568084056132 + }, + "samples_us": [ + 25.238, + 25.608, + 25.438, + 25.098, + 25.418, + 24.797, + 25.768, + 25.188, + 28.362, + 24.818, + 25.619, + 26.128, + 24.837, + 24.857, + 25.388, + 25.007, + 25.248, + 25.238, + 24.827, + 25.048 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4112, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 65.548, + "p90": 69.395, + "p95": 176.644, + "p99": 180.731 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.06273265393299567, + "p90": 0.05925498955256142, + "p95": 0.023278458368243472, + "p99": 0.022752045858209162 + }, + "samples_us": [ + 65.648, + 63.145, + 65.568, + 66.189, + 66.97, + 176.644, + 69.395, + 65.548, + 63.015, + 62.323, + 64.126, + 65.488, + 65.629, + 68.253, + 69.044, + 63.575, + 65.358, + 63.916, + 180.731, + 64.286 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16448, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 212.728, + "p90": 219.258, + "p95": 220.931, + "p99": 224.796 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.07731939377985032, + "p90": 0.07501664705506754, + "p95": 0.07444858349439418, + "p99": 0.07316856171818004 + }, + "samples_us": [ + 224.796, + 216.074, + 216.894, + 219.258, + 218.397, + 202.173, + 217.336, + 203.064, + 215.073, + 202.404, + 214.201, + 211.978, + 213.49, + 201.903, + 209.584, + 209.504, + 220.931, + 201.051, + 212.728, + 211.197 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65792, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 784.736, + "p90": 805.748, + "p95": 904.475, + "p99": 915.252 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08383966072666477, + "p90": 0.08165332088940959, + "p95": 0.07274054009231874, + "p99": 0.07188402756836369 + }, + "samples_us": [ + 915.252, + 904.475, + 785.577, + 797.645, + 784.736, + 782.803, + 771.536, + 788.952, + 775.222, + 786.488, + 775.792, + 778.837, + 791.637, + 805.748, + 777.195, + 778.746, + 776.143, + 787.11, + 782.592, + 788.461 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 263168, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3061.687, + "p90": 3111.361, + "p95": 3163.239, + "p99": 3220.756 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08595522664465702, + "p90": 0.08458292046470982, + "p95": 0.08319573702777439, + "p99": 0.08171000845764162 + }, + "samples_us": [ + 3220.756, + 3163.239, + 3111.361, + 3077.13, + 3066.264, + 3091.902, + 3057.591, + 3042.398, + 3061.678, + 3061.687, + 3067.936, + 3090.46, + 3077.53, + 3075.087, + 3049.999, + 3060.655, + 3060.115, + 3039.634, + 3051.572, + 3056.81 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 526336, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 6310.945, + "p90": 6361.321, + "p95": 6374.35, + "p99": 6390.324 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08340050499568606, + "p90": 0.082740047232328, + "p95": 0.08257092880058359, + "p99": 0.08236452486603184 + }, + "samples_us": [ + 6304.396, + 6323.263, + 6322.892, + 6310.945, + 6313.639, + 6287.75, + 6324.426, + 6308.631, + 6201.03, + 6163.775, + 6200.85, + 6138.957, + 6374.35, + 6302.623, + 6343.694, + 6361.321, + 6309.683, + 6330.975, + 6390.324, + 6317.845 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4096, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 19.729, + "p90": 20.201, + "p95": 20.47, + "p99": 39.689 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.20761315829489585, + "p90": 0.2027622394930944, + "p95": 0.20009770395701026, + "p99": 0.10320239864949986 + }, + "samples_us": [ + 39.689, + 19.93, + 19.979, + 20.47, + 19.539, + 19.479, + 19.75, + 20.04, + 19.639, + 19.729, + 20.15, + 19.599, + 19.529, + 19.77, + 20.201, + 19.63, + 19.729, + 19.519, + 19.439, + 19.499 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16384, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 28.703, + "p90": 29.014, + "p95": 29.054, + "p99": 41.813 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.5708114134411038, + "p90": 0.5646929068725443, + "p95": 0.5639154677497075, + "p99": 0.3918398584172387 + }, + "samples_us": [ + 41.813, + 28.973, + 28.823, + 28.683, + 29.014, + 29.054, + 28.874, + 28.703, + 28.543, + 28.673, + 28.533, + 28.833, + 28.893, + 28.493, + 28.412, + 28.653, + 28.493, + 28.733, + 28.613, + 28.743 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 75.143, + "p90": 80.891, + "p95": 86.75, + "p99": 93.0 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.8721504331740814, + "p90": 0.810176657477346, + "p95": 0.7554582132564841, + "p99": 0.7046881720430107 + }, + "samples_us": [ + 75.143, + 75.032, + 72.769, + 86.75, + 75.854, + 75.914, + 77.616, + 77.376, + 74.923, + 75.833, + 71.608, + 72.959, + 74.741, + 77.547, + 80.891, + 93.0, + 72.478, + 76.414, + 72.679, + 74.382 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 248.292, + "p90": 266.379, + "p95": 271.527, + "p99": 274.702 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.0557891514829314, + "p90": 0.9841015996005691, + "p95": 0.9654435838793195, + "p99": 0.9542850070257952 + }, + "samples_us": [ + 253.91, + 238.727, + 260.14, + 239.399, + 254.03, + 240.29, + 271.527, + 249.274, + 266.379, + 248.292, + 274.702, + 246.8, + 256.374, + 241.151, + 259.779, + 244.396, + 256.905, + 212.859, + 221.672, + 207.391 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 805.967, + "p90": 813.91, + "p95": 909.823, + "p99": 926.397 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.3010160465627005, + "p90": 1.2883193473479868, + "p95": 1.1525054873310523, + "p99": 1.1318862215659162 + }, + "samples_us": [ + 926.397, + 909.823, + 807.44, + 813.91, + 794.611, + 804.646, + 806.859, + 808.181, + 803.954, + 795.141, + 807.91, + 798.216, + 813.088, + 804.825, + 803.494, + 810.624, + 805.967, + 807.289, + 804.996, + 797.064 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3094.607, + "p90": 3168.597, + "p95": 3179.815, + "p99": 3197.71 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.3553591780798013, + "p90": 1.3237101467936756, + "p95": 1.3190402586313983, + "p99": 1.3116586557255037 + }, + "samples_us": [ + 3168.597, + 3167.235, + 3197.71, + 3179.815, + 3120.345, + 3056.298, + 3104.472, + 3074.526, + 3103.489, + 3080.465, + 3087.175, + 3114.366, + 3120.215, + 3110.911, + 3077.651, + 3094.607, + 3081.737, + 3085.072, + 3074.015, + 3081.136 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 8388608, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 6338.796, + "p90": 6390.835, + "p95": 6435.472, + "p99": 6490.815 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.323375606345432, + "p90": 1.3125996837658929, + "p95": 1.3034953768736777, + "p99": 1.29238131112965 + }, + "samples_us": [ + 6344.125, + 6314.791, + 6335.882, + 6309.393, + 6357.635, + 6376.343, + 6329.182, + 6336.924, + 6361.641, + 6390.835, + 6324.325, + 6330.605, + 6317.805, + 6345.446, + 6329.022, + 6339.568, + 6435.472, + 6373.598, + 6338.796, + 6490.815 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 19.319, + "p90": 19.67, + "p95": 19.88, + "p99": 20.38 + }, + "payload_gbps_at_latency_percentile": { + "p50": 3.392308090480874, + "p90": 3.3317742755465174, + "p95": 3.296579476861167, + "p99": 3.2157016683022572 + }, + "samples_us": [ + 20.38, + 19.589, + 19.559, + 19.429, + 19.67, + 19.319, + 19.88, + 18.738, + 19.008, + 19.399, + 19.359, + 18.618, + 19.369, + 19.089, + 18.968, + 19.249, + 18.748, + 18.717, + 19.329, + 18.738 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 63.055, + "p90": 64.056, + "p95": 64.407, + "p99": 65.177 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.157386408690826, + "p90": 4.092419133258399, + "p95": 4.070116602232677, + "p99": 4.0220323120119055 + }, + "samples_us": [ + 62.985, + 62.674, + 62.894, + 62.563, + 65.177, + 63.986, + 63.084, + 63.165, + 63.545, + 64.056, + 64.407, + 62.644, + 63.945, + 62.133, + 62.404, + 63.014, + 63.055, + 63.145, + 63.345, + 62.684 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 240.29, + "p90": 241.672, + "p95": 242.603, + "p99": 246.9 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.363793749219693, + "p90": 4.338839418716277, + "p95": 4.3221889259407344, + "p99": 4.246966383151073 + }, + "samples_us": [ + 240.801, + 239.619, + 241.672, + 238.177, + 240.079, + 246.9, + 240.851, + 240.831, + 240.35, + 241.632, + 240.76, + 238.728, + 240.179, + 238.527, + 242.603, + 240.471, + 240.109, + 239.9, + 240.29, + 239.919 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 948.561, + "p90": 951.897, + "p95": 954.049, + "p99": 954.13 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.421754636760314, + "p90": 4.4062582401247194, + "p95": 4.396319266620478, + "p99": 4.395946045088196 + }, + "samples_us": [ + 950.404, + 951.897, + 954.13, + 948.561, + 951.235, + 949.342, + 948.582, + 948.05, + 948.652, + 944.525, + 949.964, + 954.049, + 948.431, + 951.826, + 940.199, + 941.13, + 937.094, + 934.22, + 935.222, + 932.758 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3544.12, + "p90": 3697.199, + "p95": 3705.251, + "p99": 3716.458 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.733817139374513, + "p90": 4.5378179535372585, + "p95": 4.527956675539659, + "p99": 4.51430259671978 + }, + "samples_us": [ + 3716.458, + 3697.199, + 3705.251, + 3663.168, + 3621.165, + 3604.46, + 3595.858, + 3594.806, + 3602.347, + 3570.169, + 3544.12, + 3540.895, + 3532.613, + 3532.092, + 3527.565, + 3529.608, + 3529.037, + 3529.328, + 3526.383, + 3523.379 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14262.017, + "p90": 14285.883, + "p95": 14301.466, + "p99": 14317.85 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.705425887516471, + "p90": 4.697565001757329, + "p95": 4.692446494646074, + "p99": 4.687076900512298 + }, + "samples_us": [ + 14301.466, + 14273.434, + 14270.09, + 14263.961, + 14265.252, + 14257.891, + 14250.6, + 14258.712, + 14260.885, + 14265.383, + 14264.441, + 14256.85, + 14285.883, + 14262.017, + 14317.85, + 14256.079, + 14258.152, + 14261.707, + 14265.322, + 14261.096 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 134217728, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 28429.163, + "p90": 28499.007, + "p95": 28500.829, + "p99": 28557.194 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.721128370891538, + "p90": 4.709558055829805, + "p95": 4.7092569833670455, + "p99": 4.6999620480919795 + }, + "samples_us": [ + 28422.923, + 28434.57, + 28465.327, + 28412.998, + 28428.591, + 28449.543, + 28442.813, + 28499.007, + 28429.163, + 28425.026, + 28425.998, + 28442.973, + 28453.369, + 28425.326, + 28493.829, + 28414.25, + 28557.194, + 28425.647, + 28415.001, + 28500.829 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 21.853, + "p90": 22.053, + "p95": 22.173, + "p99": 22.784 + }, + "payload_gbps_at_latency_percentile": { + "p50": 11.995790051709147, + "p90": 11.88699950120165, + "p95": 11.822667207865422, + "p99": 11.50561797752809 + }, + "samples_us": [ + 22.784, + 21.823, + 21.853, + 22.053, + 21.863, + 21.772, + 21.843, + 22.173, + 21.853, + 21.863, + 21.823, + 21.933, + 21.913, + 21.752, + 21.642, + 21.893, + 21.663, + 21.622, + 21.882, + 21.813 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 81.212, + "p90": 81.422, + "p95": 81.492, + "p99": 81.563 + }, + "payload_gbps_at_latency_percentile": { + "p50": 12.911589420282716, + "p90": 12.878288423276265, + "p95": 12.86722623079566, + "p99": 12.856025403675686 + }, + "samples_us": [ + 81.402, + 81.392, + 80.36, + 81.322, + 81.563, + 81.251, + 80.21, + 81.422, + 79.108, + 80.681, + 81.141, + 80.58, + 80.861, + 81.492, + 81.292, + 80.43, + 81.212, + 81.352, + 79.389, + 81.252 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 330.415, + "p90": 332.949, + "p95": 336.064, + "p99": 339.468 + }, + "payload_gbps_at_latency_percentile": { + "p50": 12.694048393686728, + "p90": 12.597436844681917, + "p95": 12.480670348505047, + "p99": 12.355520991669318 + }, + "samples_us": [ + 330.415, + 327.07, + 325.828, + 331.356, + 331.397, + 331.437, + 329.514, + 332.298, + 332.949, + 329.353, + 327.781, + 331.937, + 330.916, + 327.851, + 329.854, + 339.468, + 328.743, + 336.064, + 329.183, + 331.666 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 1283.232, + "p90": 1312.757, + "p95": 1317.915, + "p99": 1338.966 + }, + "payload_gbps_at_latency_percentile": { + "p50": 13.074187676117802, + "p90": 12.78013828911215, + "p95": 12.730119924274327, + "p99": 12.529979103278201 + }, + "samples_us": [ + 1338.966, + 1317.915, + 1312.757, + 1305.887, + 1300.468, + 1304.665, + 1298.766, + 1293.558, + 1283.232, + 1284.695, + 1285.146, + 1262.953, + 1243.753, + 1240.518, + 1226.618, + 1229.612, + 1201.91, + 1205.135, + 1190.965, + 1180.198 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 4490.177, + "p90": 4829.676, + "p95": 4929.395, + "p99": 5004.157 + }, + "payload_gbps_at_latency_percentile": { + "p50": 14.94570570380633, + "p90": 13.895106835323944, + "p95": 13.614016324518525, + "p99": 13.410623207864981 + }, + "samples_us": [ + 5004.157, + 4929.395, + 4829.676, + 4746.342, + 4653.882, + 4552.07, + 4519.251, + 4488.264, + 4489.206, + 4486.862, + 4488.935, + 4490.177, + 4489.306, + 4484.569, + 4484.698, + 4484.88, + 4490.337, + 4490.538, + 4493.171, + 4483.037 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 268435456, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 17892.727, + "p90": 17942.942, + "p95": 17946.587, + "p99": 17955.34 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.002489894357636, + "p90": 14.96050402436791, + "p95": 14.957465505836849, + "p99": 14.950173931543485 + }, + "samples_us": [ + 17911.435, + 17895.731, + 17892.727, + 17894.39, + 17882.271, + 17946.587, + 17889.842, + 17889.511, + 17942.942, + 17883.053, + 17880.058, + 17955.34, + 17884.545, + 17895.07, + 17910.022, + 17889.663, + 17895.681, + 17932.826, + 17878.476, + 17889.592 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 536870912, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 35650.571, + "p90": 35717.421, + "p95": 35723.271, + "p99": 35726.856 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.05925142124652, + "p90": 15.031065988778977, + "p95": 15.028604519446162, + "p99": 15.027096478906513 + }, + "samples_us": [ + 35662.79, + 35656.471, + 35641.068, + 35640.957, + 35645.704, + 35684.152, + 35649.61, + 35655.248, + 35717.421, + 35650.171, + 35726.856, + 35723.271, + 35647.046, + 35642.439, + 35663.611, + 35643.2, + 35635.639, + 35717.321, + 35650.571, + 35655.579 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 257, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 16.965, + "p90": 17.546, + "p95": 18.347, + "p99": 18.678 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.01514883583849101, + "p90": 0.014647213040009118, + "p95": 0.01400773968496212, + "p99": 0.013759503158796446 + }, + "samples_us": [ + 18.678, + 17.546, + 16.575, + 16.965, + 16.986, + 16.445, + 16.935, + 17.196, + 17.006, + 17.025, + 18.347, + 17.316, + 17.226, + 16.775, + 16.715, + 17.366, + 16.745, + 16.916, + 16.846, + 16.836 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1028, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 25.608, + "p90": 26.84, + "p95": 27.551, + "p99": 31.467 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.0401437050921587, + "p90": 0.03830104321907601, + "p95": 0.037312620231570545, + "p99": 0.03266914545396765 + }, + "samples_us": [ + 27.551, + 26.469, + 31.467, + 25.608, + 25.659, + 26.089, + 25.347, + 25.358, + 26.84, + 26.099, + 25.799, + 26.019, + 25.258, + 25.589, + 25.479, + 25.548, + 25.578, + 25.779, + 25.178, + 25.608 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4112, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 68.282, + "p90": 73.78, + "p95": 78.517, + "p99": 190.485 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.0602208488327817, + "p90": 0.055733261046354024, + "p95": 0.052370824152731255, + "p99": 0.021587001601175947 + }, + "samples_us": [ + 73.24, + 61.282, + 63.676, + 65.398, + 68.282, + 78.517, + 66.73, + 66.42, + 65.759, + 61.502, + 72.489, + 70.546, + 70.856, + 68.443, + 69.474, + 73.78, + 69.314, + 63.184, + 190.485, + 61.542 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16448, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 202.253, + "p90": 215.473, + "p95": 217.286, + "p99": 229.594 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.0813238864194845, + "p90": 0.0763343899235635, + "p95": 0.07569746785342821, + "p99": 0.07163950277446275 + }, + "samples_us": [ + 229.594, + 187.972, + 212.768, + 199.428, + 208.072, + 189.744, + 202.253, + 194.03, + 215.473, + 196.194, + 202.043, + 194.371, + 211.797, + 191.317, + 202.684, + 201.111, + 217.286, + 207.191, + 211.397, + 205.508 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65792, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 769.353, + "p90": 782.322, + "p95": 895.532, + "p99": 919.627 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08551601150577173, + "p90": 0.08409836358941715, + "p95": 0.07346694478812595, + "p99": 0.07154204911339054 + }, + "samples_us": [ + 895.532, + 919.627, + 779.939, + 774.3, + 760.879, + 779.989, + 772.557, + 782.322, + 761.271, + 773.559, + 767.3, + 769.353, + 754.39, + 764.055, + 768.792, + 771.525, + 766.479, + 768.411, + 766.078, + 775.972 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 263168, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3062.228, + "p90": 3081.206, + "p95": 3081.487, + "p99": 3124.281 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08594004104201254, + "p90": 0.08541071255865398, + "p95": 0.08540292397793663, + "p99": 0.08423314036093424 + }, + "samples_us": [ + 3051.982, + 3064.351, + 3052.283, + 3056.299, + 3055.668, + 3059.253, + 3059.945, + 3081.206, + 3062.228, + 3057.711, + 3043.75, + 3073.776, + 3074.186, + 3043.099, + 3081.487, + 3067.866, + 3071.982, + 3066.995, + 3064.481, + 3124.281 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 526336, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 6153.389, + "p90": 6229.543, + "p95": 6236.784, + "p99": 6258.136 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08553595425220151, + "p90": 0.08449030691336427, + "p95": 0.08439221239664545, + "p99": 0.08410427641713124 + }, + "samples_us": [ + 6195.973, + 6151.196, + 6188.111, + 6164.515, + 6214.03, + 6151.045, + 6149.923, + 6144.325, + 6130.345, + 6189.203, + 6136.174, + 6125.608, + 6155.712, + 6236.784, + 6156.654, + 6229.543, + 6153.389, + 6258.136, + 6152.768, + 6119.929 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4096, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 19.048, + "p90": 19.299, + "p95": 19.63, + "p99": 31.247 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.2150356992860143, + "p90": 0.21223897611275194, + "p95": 0.2086602139582272, + "p99": 0.13108458412007554 + }, + "samples_us": [ + 19.63, + 19.279, + 19.159, + 31.247, + 18.798, + 18.919, + 19.299, + 18.908, + 19.119, + 18.938, + 18.899, + 19.048, + 18.889, + 19.158, + 19.179, + 18.828, + 18.728, + 18.978, + 19.299, + 19.189 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16384, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 28.273, + "p90": 28.983, + "p95": 29.605, + "p99": 40.351 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.5794928023202348, + "p90": 0.5652968981816927, + "p95": 0.5534200304002702, + "p99": 0.4060370251047062 + }, + "samples_us": [ + 28.983, + 28.973, + 28.703, + 40.351, + 28.452, + 28.513, + 28.383, + 28.612, + 28.273, + 28.362, + 27.902, + 27.982, + 28.013, + 27.931, + 28.032, + 27.631, + 28.013, + 29.605, + 27.832, + 27.722 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 78.568, + "p90": 83.975, + "p95": 90.446, + "p99": 120.691 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.8341309438957336, + "p90": 0.780422744864543, + "p95": 0.7245870464144352, + "p99": 0.5430065207844826 + }, + "samples_us": [ + 78.237, + 80.13, + 74.312, + 76.364, + 90.446, + 79.7, + 78.568, + 81.442, + 77.076, + 77.346, + 72.829, + 75.654, + 80.08, + 79.259, + 83.975, + 82.443, + 120.691, + 82.404, + 76.845, + 73.009 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 251.747, + "p90": 264.717, + "p95": 267.25, + "p99": 268.112 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.041299399794238, + "p90": 0.9902801860099654, + "p95": 0.9808942937324603, + "p99": 0.9777406457003044 + }, + "samples_us": [ + 262.383, + 239.118, + 264.065, + 251.747, + 254.411, + 243.004, + 264.567, + 241.071, + 259.92, + 250.515, + 267.25, + 245.498, + 261.151, + 243.325, + 268.112, + 252.708, + 264.717, + 214.13, + 223.796, + 209.794 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 787.49, + "p90": 798.467, + "p95": 804.956, + "p99": 966.878 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.331541987834766, + "p90": 1.3132364894228565, + "p95": 1.3026500827374417, + "p99": 1.084496699687034 + }, + "samples_us": [ + 966.878, + 804.956, + 780.199, + 783.464, + 781.37, + 792.387, + 774.601, + 773.81, + 785.447, + 777.966, + 787.56, + 791.395, + 782.613, + 793.679, + 798.467, + 792.036, + 781.681, + 796.533, + 787.49, + 790.173 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2983.881, + "p90": 3020.906, + "p95": 3043.68, + "p99": 3068.106 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.4056539118014424, + "p90": 1.3884258563490555, + "p95": 1.3780371129685118, + "p99": 1.3670661965394937 + }, + "samples_us": [ + 3068.106, + 3008.518, + 2984.071, + 2969.619, + 2982.008, + 2983.84, + 2983.881, + 2977.551, + 2980.836, + 2972.374, + 2986.103, + 2999.254, + 2984.381, + 2966.474, + 2975.278, + 2985.863, + 3009.199, + 2979.974, + 3020.906, + 3043.68 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 8388608, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5965.217, + "p90": 6111.857, + "p95": 6122.102, + "p99": 6146.849 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.4062536199437505, + "p90": 1.3725137875444402, + "p95": 1.3702169614292607, + "p99": 1.3647005156625778 + }, + "samples_us": [ + 5919.289, + 5931.557, + 5925.677, + 5965.217, + 5938.487, + 5919.108, + 5935.833, + 5938.818, + 5948.372, + 5892.678, + 5990.194, + 6096.453, + 5993.199, + 6044.456, + 5983.164, + 5986.509, + 6084.937, + 6122.102, + 6146.849, + 6111.857 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 15.864, + "p90": 16.374, + "p95": 16.765, + "p99": 22.504 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.131114473020676, + "p90": 4.00244289727617, + "p95": 3.9090963316433047, + "p99": 2.9121933878421613 + }, + "samples_us": [ + 15.924, + 16.264, + 15.694, + 15.743, + 15.794, + 15.903, + 15.904, + 16.765, + 15.984, + 22.504, + 16.374, + 15.683, + 15.523, + 15.723, + 15.864, + 16.094, + 15.674, + 15.964, + 15.654, + 15.253 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 56.655, + "p90": 57.466, + "p95": 57.546, + "p99": 58.528 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.627023210661019, + "p90": 4.561723453868375, + "p95": 4.555381781531297, + "p99": 4.478950246036085 + }, + "samples_us": [ + 57.426, + 57.346, + 56.044, + 55.783, + 55.914, + 57.266, + 56.395, + 57.265, + 56.465, + 56.345, + 56.575, + 56.495, + 57.255, + 56.725, + 56.465, + 57.466, + 56.655, + 56.975, + 57.546, + 58.528 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 216.334, + "p90": 219.579, + "p95": 220.06, + "p99": 221.802 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.847023583902669, + "p90": 4.775392910979647, + "p95": 4.764955012269381, + "p99": 4.7275317625630064 + }, + "samples_us": [ + 220.06, + 218.077, + 217.075, + 215.763, + 214.452, + 218.708, + 219.579, + 215.193, + 214.511, + 219.169, + 215.043, + 216.855, + 218.888, + 216.745, + 215.903, + 216.304, + 216.334, + 215.653, + 214.992, + 221.802 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 871.546, + "p90": 877.325, + "p95": 877.474, + "p99": 988.571 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.81248723532665, + "p90": 4.780787051548742, + "p95": 4.77997524712983, + "p99": 4.242794902945767 + }, + "samples_us": [ + 988.571, + 871.546, + 875.742, + 872.277, + 869.553, + 870.925, + 869.793, + 872.988, + 871.436, + 875.131, + 874.09, + 872.957, + 871.335, + 877.325, + 877.474, + 869.893, + 868.972, + 869.934, + 871.896, + 867.76 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3457.59, + "p90": 3459.994, + "p95": 3460.334, + "p99": 3463.278 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.852286129934434, + "p90": 4.84891476690422, + "p95": 4.848438329941561, + "p99": 4.844316858190419 + }, + "samples_us": [ + 3463.278, + 3457.59, + 3456.638, + 3456.719, + 3458.642, + 3458.131, + 3455.197, + 3458.772, + 3458.281, + 3458.512, + 3459.903, + 3453.504, + 3457.61, + 3455.146, + 3460.334, + 3452.532, + 3459.994, + 3457.31, + 3453.734, + 3455.347 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 13646.595, + "p90": 13653.416, + "p95": 13666.426, + "p99": 13666.886 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.917626997796886, + "p90": 4.915170240180187, + "p95": 4.910491155478397, + "p99": 4.910325878184687 + }, + "samples_us": [ + 13636.35, + 13629.048, + 13644.572, + 13640.566, + 13645.363, + 13636.59, + 13647.696, + 13648.738, + 13653.416, + 13645.343, + 13641.046, + 13666.426, + 13635.258, + 13666.886, + 13652.675, + 13647.636, + 13648.828, + 13646.595, + 13651.602, + 13650.981 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 134217728, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 27221.383, + "p90": 27232.96, + "p95": 27254.712, + "p99": 27258.899 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.930599154348624, + "p90": 4.928503107998543, + "p95": 4.924569667072615, + "p99": 4.923813247189478 + }, + "samples_us": [ + 27212.81, + 27258.899, + 27254.712, + 27232.209, + 27217.818, + 27208.053, + 27215.825, + 27225.079, + 27230.667, + 27223.146, + 27218.699, + 27228.003, + 27221.383, + 27204.768, + 27231.067, + 27227.573, + 27209.745, + 27221.143, + 27232.96, + 27214.593 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 19.239, + "p90": 19.649, + "p95": 20.09, + "p99": 37.617 + }, + "payload_gbps_at_latency_percentile": { + "p50": 13.62565621913821, + "p90": 13.341340526235431, + "p95": 13.048481831757092, + "p99": 6.968764122604142 + }, + "samples_us": [ + 19.509, + 19.53, + 19.159, + 19.649, + 19.028, + 19.339, + 19.179, + 19.169, + 37.617, + 20.09, + 19.259, + 19.279, + 19.119, + 19.239, + 19.279, + 19.039, + 19.239, + 18.988, + 19.048, + 19.139 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 70.796, + "p90": 71.548, + "p95": 71.628, + "p99": 78.257 + }, + "payload_gbps_at_latency_percentile": { + "p50": 14.811232273009775, + "p90": 14.655559903840778, + "p95": 14.639191377673535, + "p99": 13.399133623829178 + }, + "samples_us": [ + 71.477, + 78.257, + 70.215, + 70.706, + 70.646, + 70.365, + 70.666, + 71.026, + 69.864, + 71.157, + 70.796, + 70.867, + 71.628, + 71.427, + 70.956, + 70.846, + 70.135, + 70.296, + 71.548, + 70.706 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 272.879, + "p90": 274.671, + "p95": 275.142, + "p99": 277.636 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.37056350983403, + "p90": 15.270283357180045, + "p95": 15.244143024329256, + "p99": 15.107205117491969 + }, + "samples_us": [ + 270.866, + 271.486, + 277.636, + 272.879, + 274.051, + 273.82, + 272.879, + 273.88, + 269.865, + 272.217, + 272.618, + 273.3, + 273.54, + 270.606, + 274.2, + 275.142, + 270.726, + 269.794, + 269.614, + 274.671 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 1096.092, + "p90": 1097.694, + "p95": 1098.316, + "p99": 1098.356 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.306393988825755, + "p90": 15.284055483586501, + "p95": 15.275399793866246, + "p99": 15.274843493366449 + }, + "samples_us": [ + 1093.469, + 1095.932, + 1096.402, + 1098.356, + 1095.792, + 1097.694, + 1096.312, + 1097.554, + 1093.698, + 1094.189, + 1097.364, + 1097.634, + 1097.124, + 1096.142, + 1096.092, + 1095.121, + 1094.56, + 1098.316, + 1094.08, + 1094.028 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 4324.709, + "p90": 4328.755, + "p95": 4329.596, + "p99": 4330.138 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.517544417439415, + "p90": 15.503040481616539, + "p95": 15.500029102022452, + "p99": 15.498088975455286 + }, + "samples_us": [ + 4324.709, + 4324.759, + 4326.823, + 4328.435, + 4327.143, + 4318.991, + 4322.786, + 4326.722, + 4318.53, + 4325.681, + 4329.596, + 4322.946, + 4323.137, + 4322.897, + 4328.755, + 4330.138, + 4326.132, + 4321.144, + 4323.437, + 4321.284 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 268435456, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 17280.39, + "p90": 17303.995, + "p95": 17309.913, + "p99": 17333.229 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.534108663056795, + "p90": 15.512918028466837, + "p95": 15.507614394133581, + "p99": 15.486754141423967 + }, + "samples_us": [ + 17263.203, + 17333.229, + 17287.39, + 17274.53, + 17279.679, + 17277.886, + 17273.318, + 17275.753, + 17281.501, + 17281.201, + 17280.39, + 17269.323, + 17303.334, + 17303.995, + 17273.138, + 17283.304, + 17283.895, + 17287.32, + 17309.913, + 17269.213 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 536870912, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 34477.254, + "p90": 34516.182, + "p95": 34537.314, + "p99": 34564.545 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.571742227498744, + "p90": 15.554180123398353, + "p95": 15.544663143173207, + "p99": 15.532416584682368 + }, + "samples_us": [ + 34436.943, + 34473.979, + 34490.423, + 34447.579, + 34459.227, + 34492.326, + 34516.182, + 34502.021, + 34537.314, + 34465.496, + 34487.338, + 34458.656, + 34452.227, + 34564.545, + 34495.25, + 34462.112, + 34472.738, + 34484.375, + 34485.446, + 34477.254 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 257, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 16.896, + "p90": 17.627, + "p95": 17.847, + "p99": 18.999 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.015210700757575758, + "p90": 0.014579905826289215, + "p95": 0.014400179301843447, + "p99": 0.013527027738302016 + }, + "samples_us": [ + 17.627, + 17.226, + 16.806, + 16.896, + 17.126, + 16.995, + 17.146, + 16.966, + 16.685, + 16.896, + 18.999, + 16.485, + 16.444, + 16.585, + 16.956, + 16.855, + 16.776, + 16.635, + 16.225, + 17.847 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1028, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 24.667, + "p90": 26.71, + "p95": 27.091, + "p99": 32.468 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.04167511249847975, + "p90": 0.03848745788094347, + "p95": 0.03794618138865306, + "p99": 0.03166194406800542 + }, + "samples_us": [ + 25.058, + 24.297, + 25.388, + 32.468, + 24.667, + 24.477, + 24.597, + 25.128, + 24.546, + 24.467, + 24.607, + 24.898, + 24.848, + 26.71, + 24.447, + 24.187, + 24.097, + 25.448, + 27.091, + 25.748 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4112, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 61.672, + "p90": 68.974, + "p95": 70.145, + "p99": 72.899 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.06667531456738876, + "p90": 0.05961666714994056, + "p95": 0.05862142704398032, + "p99": 0.05640680942125406 + }, + "samples_us": [ + 68.974, + 58.007, + 58.798, + 60.05, + 62.183, + 72.899, + 64.126, + 60.771, + 60.01, + 58.617, + 68.402, + 61.823, + 63.525, + 68.943, + 64.566, + 70.145, + 60.361, + 59.128, + 61.672, + 61.552 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16448, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 209.424, + "p90": 222.874, + "p95": 225.337, + "p99": 225.668 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.07853923141569256, + "p90": 0.07379954593178208, + "p95": 0.0729928950860267, + "p99": 0.07288583228459507 + }, + "samples_us": [ + 205.748, + 225.337, + 201.762, + 206.77, + 219.599, + 197.646, + 209.424, + 207.431, + 222.874, + 210.445, + 214.161, + 203.215, + 220.551, + 209.885, + 215.543, + 207.071, + 222.222, + 205.438, + 225.668, + 207.07 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65792, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 797.104, + "p90": 806.468, + "p95": 824.755, + "p99": 852.578 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.0825387904213252, + "p90": 0.08158042228581915, + "p95": 0.07977156852641087, + "p99": 0.07716830600836522 + }, + "samples_us": [ + 806.358, + 806.158, + 796.043, + 788.141, + 852.578, + 824.755, + 798.416, + 796.093, + 797.104, + 792.156, + 797.104, + 788.641, + 806.468, + 801.5, + 792.187, + 799.979, + 796.142, + 805.327, + 802.583, + 792.928 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 263168, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3069.729, + "p90": 3090.15, + "p95": 3092.012, + "p99": 3137.531 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08573004327092065, + "p90": 0.08516350338980308, + "p95": 0.08511221819320236, + "p99": 0.083877418262959 + }, + "samples_us": [ + 3083.35, + 3071.712, + 3090.15, + 3080.375, + 3058.823, + 3072.623, + 3055.868, + 3060.456, + 3071.292, + 3059.003, + 3056.309, + 3076.569, + 3092.012, + 3137.531, + 3062.858, + 3069.86, + 3059.884, + 3063.8, + 3069.729, + 3054.376 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 526336, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 6111.306, + "p90": 6157.445, + "p95": 6157.996, + "p99": 6176.133 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08612496248755994, + "p90": 0.0854796104553106, + "p95": 0.08547196198243714, + "p99": 0.08522096269623727 + }, + "samples_us": [ + 6111.306, + 6118.006, + 6112.227, + 6117.575, + 6109.854, + 6176.133, + 6157.445, + 6157.996, + 6103.945, + 6114.13, + 6107.941, + 6151.767, + 6109.363, + 6094.28, + 6111.105, + 6095.251, + 6114.5, + 6118.116, + 6086.799, + 6101.04 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4096, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 16.845, + "p90": 17.396, + "p95": 18.007, + "p99": 31.227 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.2431582071831404, + "p90": 0.23545642676477352, + "p95": 0.22746709612928306, + "p99": 0.13116854004547346 + }, + "samples_us": [ + 17.296, + 17.396, + 16.845, + 17.166, + 16.294, + 18.007, + 16.625, + 16.816, + 16.605, + 16.775, + 17.256, + 31.227, + 16.475, + 16.986, + 16.544, + 17.065, + 16.635, + 17.256, + 17.266, + 16.795 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16384, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 24.747, + "p90": 25.679, + "p95": 25.949, + "p99": 27.21 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.6620600476825473, + "p90": 0.6380310759764788, + "p95": 0.6313923465258777, + "p99": 0.6021315692760014 + }, + "samples_us": [ + 27.21, + 25.388, + 24.687, + 24.747, + 24.517, + 24.897, + 24.547, + 25.679, + 25.949, + 24.586, + 25.258, + 25.458, + 24.136, + 24.206, + 24.617, + 24.497, + 24.867, + 25.398, + 24.737, + 24.848 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65536, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 67.472, + "p90": 71.127, + "p95": 72.329, + "p99": 74.752 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.9713066160777805, + "p90": 0.9213941259999718, + "p95": 0.9060819311756004, + "p99": 0.8767123287671232 + }, + "samples_us": [ + 67.472, + 66.58, + 63.695, + 65.248, + 67.411, + 67.671, + 68.793, + 71.127, + 69.534, + 69.134, + 74.752, + 65.268, + 69.574, + 70.867, + 72.329, + 69.875, + 67.11, + 65.448, + 63.595, + 65.919 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 219.569, + "p90": 229.844, + "p95": 231.767, + "p99": 236.094 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.1939026000938202, + "p90": 1.1405300986756235, + "p95": 1.1310669767481998, + "p99": 1.1103374079815667 + }, + "samples_us": [ + 220.34, + 215.493, + 229.844, + 213.58, + 227.381, + 219.569, + 219.7, + 227.57, + 236.094, + 212.098, + 225.358, + 227.942, + 220.421, + 209.914, + 217.095, + 209.795, + 231.767, + 213.81, + 216.033, + 211.638 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 793.288, + "p90": 799.808, + "p95": 801.43, + "p99": 805.037 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.3218099857807, + "p90": 1.3110346483155957, + "p95": 1.3083812684825873, + "p99": 1.302519014653985 + }, + "samples_us": [ + 805.037, + 797.535, + 788.341, + 793.288, + 793.809, + 796.013, + 786.329, + 776.143, + 790.454, + 793.729, + 799.808, + 796.424, + 785.417, + 799.258, + 801.43, + 796.183, + 792.507, + 789.904, + 792.758, + 792.517 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2999.775, + "p90": 3039.875, + "p95": 3040.516, + "p99": 3053.986 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.3982061987982433, + "p90": 1.3797619967926313, + "p95": 1.3794711160868747, + "p99": 1.3733867804240099 + }, + "samples_us": [ + 2961.126, + 3026.595, + 3040.516, + 3039.875, + 3053.986, + 3029.549, + 3000.304, + 2988.377, + 2990.981, + 2980.996, + 2982.839, + 3003.65, + 3009.139, + 2998.883, + 2999.775, + 2999.444, + 3009.989, + 3010.03, + 2987.395, + 2988.308 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 8388608, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5891.407, + "p90": 5958.287, + "p95": 5970.936, + "p99": 5979.779 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.4238717508398249, + "p90": 1.4078892137958443, + "p95": 1.4049067013948902, + "p99": 1.4028291012092586 + }, + "samples_us": [ + 5886.95, + 5889.804, + 5970.936, + 5851.877, + 5949.824, + 5854.892, + 5871.196, + 5952.448, + 5941.611, + 5880.8, + 5950.324, + 5891.637, + 5890.394, + 5856.775, + 5872.859, + 5958.287, + 5891.407, + 5979.779, + 5934.882, + 5928.002 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65536, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 16.875, + "p90": 17.476, + "p95": 17.476, + "p99": 20.842 + }, + "payload_gbps_at_latency_percentile": { + "p50": 3.883614814814815, + "p90": 3.7500572213321126, + "p95": 3.7500572213321126, + "p99": 3.1444199213127337 + }, + "samples_us": [ + 17.436, + 17.476, + 16.965, + 16.925, + 17.035, + 16.495, + 17.436, + 20.842, + 16.875, + 16.455, + 17.476, + 16.344, + 16.755, + 16.926, + 16.614, + 17.276, + 16.495, + 16.334, + 16.705, + 16.745 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 24.497, + "p90": 25.017, + "p95": 25.087, + "p99": 25.178 + }, + "payload_gbps_at_latency_percentile": { + "p50": 10.701065436584072, + "p90": 10.478634528520606, + "p95": 10.449396101566549, + "p99": 10.411629200095321 + }, + "samples_us": [ + 24.857, + 24.397, + 24.537, + 25.017, + 24.667, + 24.427, + 24.497, + 25.087, + 25.178, + 24.497, + 24.036, + 24.807, + 24.006, + 24.306, + 24.477, + 24.717, + 24.346, + 24.597, + 23.916, + 24.867 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 66.991, + "p90": 69.575, + "p95": 75.232, + "p99": 78.057 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.652490633070114, + "p90": 15.071160618038089, + "p95": 13.937898766482348, + "p99": 13.433465288186838 + }, + "samples_us": [ + 66.609, + 67.231, + 75.232, + 64.517, + 67.451, + 67.842, + 66.74, + 69.575, + 66.85, + 66.991, + 63.435, + 63.235, + 68.112, + 67.17, + 69.414, + 68.533, + 78.057, + 64.246, + 60.811, + 64.787 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 202.924, + "p90": 214.101, + "p95": 219.159, + "p99": 222.213 + }, + "payload_gbps_at_latency_percentile": { + "p50": 20.66933433206521, + "p90": 19.5903055100163, + "p95": 19.138178217641073, + "p99": 18.87515131877973 + }, + "samples_us": [ + 213.52, + 201.372, + 208.332, + 214.101, + 219.159, + 205.818, + 209.013, + 207.611, + 213.66, + 193.7, + 202.924, + 193.109, + 222.213, + 191.847, + 202.123, + 193.641, + 206.289, + 201.161, + 200.671, + 196.374 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 759.618, + "p90": 768.141, + "p95": 772.858, + "p99": 776.554 + }, + "payload_gbps_at_latency_percentile": { + "p50": 22.086385525356167, + "p90": 21.841323402864838, + "p95": 21.70801880811223, + "p99": 21.604699737558494 + }, + "samples_us": [ + 776.554, + 772.858, + 754.691, + 761.581, + 752.207, + 753.288, + 754.521, + 763.844, + 758.376, + 760.48, + 759.858, + 759.849, + 754.6, + 768.141, + 755.081, + 757.765, + 759.618, + 758.406, + 760.109, + 764.095 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2967.897, + "p90": 2982.248, + "p95": 2983.981, + "p99": 2995.978 + }, + "payload_gbps_at_latency_percentile": { + "p50": 22.6115879358347, + "p90": 22.502777770326276, + "p95": 22.489708882194627, + "p99": 22.39965179984633 + }, + "samples_us": [ + 2967.887, + 2981.617, + 2983.981, + 2995.978, + 2976.97, + 2967.897, + 2981.287, + 2964.631, + 2967.376, + 2980.155, + 2982.208, + 2982.248, + 2962.739, + 2967.516, + 2956.679, + 2955.999, + 2961.748, + 2958.423, + 2969.399, + 2972.664 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 134217728, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5747.831, + "p90": 5815.332, + "p95": 5827.601, + "p99": 5846.87 + }, + "payload_gbps_at_latency_percentile": { + "p50": 23.351021976811776, + "p90": 23.07997686116631, + "p95": 23.03138598541664, + "p99": 22.95548353221467 + }, + "samples_us": [ + 5797.586, + 5747.831, + 5761.912, + 5725.908, + 5827.601, + 5846.87, + 5751.517, + 5735.163, + 5728.753, + 5735.292, + 5743.845, + 5736.764, + 5815.332, + 5726.45, + 5759.088, + 5747.141, + 5735.984, + 5776.755, + 5753.229, + 5765.168 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 16.925, + "p90": 17.406, + "p95": 17.757, + "p99": 34.242 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.488567208271787, + "p90": 15.06055383201195, + "p95": 14.762854085712677, + "p99": 7.655627591846271 + }, + "samples_us": [ + 17.306, + 17.065, + 16.885, + 16.745, + 17.046, + 34.242, + 17.006, + 16.525, + 16.925, + 16.775, + 16.795, + 17.226, + 17.035, + 17.406, + 17.757, + 17.176, + 16.835, + 16.615, + 16.175, + 16.385 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 24.587, + "p90": 25.207, + "p95": 25.749, + "p99": 29.424 + }, + "payload_gbps_at_latency_percentile": { + "p50": 42.647577988367836, + "p90": 41.59860356250248, + "p95": 40.72297953318576, + "p99": 35.63675910821098 + }, + "samples_us": [ + 24.657, + 24.607, + 24.607, + 24.607, + 24.938, + 24.717, + 24.587, + 24.076, + 24.136, + 24.447, + 29.424, + 24.537, + 24.456, + 24.487, + 25.207, + 24.577, + 25.749, + 25.077, + 24.116, + 24.016 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 64.387, + "p90": 69.645, + "p95": 73.37, + "p99": 88.252 + }, + "payload_gbps_at_latency_percentile": { + "p50": 65.14209390094274, + "p90": 60.224050542034604, + "p95": 57.16647130979965, + "p99": 47.526446992702716 + }, + "samples_us": [ + 64.837, + 88.252, + 64.557, + 64.527, + 66.099, + 73.37, + 69.645, + 67.591, + 62.935, + 64.237, + 62.573, + 62.063, + 64.206, + 64.266, + 64.387, + 66.16, + 65.388, + 62.303, + 60.911, + 62.544 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 206.75, + "p90": 224.076, + "p95": 224.616, + "p99": 229.284 + }, + "payload_gbps_at_latency_percentile": { + "p50": 81.14735671100362, + "p90": 74.87288241489495, + "p95": 74.69288029347865, + "p99": 73.17220564888959 + }, + "samples_us": [ + 224.616, + 200.35, + 224.076, + 201.121, + 207.1, + 199.429, + 203.805, + 214.662, + 229.284, + 196.615, + 208.002, + 198.106, + 218.077, + 212.498, + 203.836, + 197.847, + 206.87, + 211.777, + 203.064, + 206.75 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 780.028, + "p90": 788.571, + "p95": 790.685, + "p99": 804.696 + }, + "payload_gbps_at_latency_percentile": { + "p50": 86.03391673119427, + "p90": 85.1018665408695, + "p95": 84.87433554449623, + "p99": 83.39654229671827 + }, + "samples_us": [ + 782.212, + 790.685, + 788.571, + 780.86, + 763.023, + 780.028, + 783.905, + 775.111, + 772.207, + 782.773, + 773.319, + 770.464, + 780.109, + 773.489, + 780.66, + 804.696, + 782.332, + 768.912, + 766.589, + 775.161 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 268435456, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2977.671, + "p90": 2994.186, + "p95": 2995.267, + "p99": 2995.878 + }, + "payload_gbps_at_latency_percentile": { + "p50": 90.14946782233497, + "p90": 89.65223135770457, + "p95": 89.61987562377578, + "p99": 89.60159792888763 + }, + "samples_us": [ + 2990.27, + 2995.878, + 2975.658, + 2965.042, + 2977.671, + 2972.754, + 2972.734, + 2958.583, + 2995.267, + 2992.544, + 2974.366, + 2979.764, + 2980.405, + 2972.803, + 2965.884, + 2977.991, + 2976.089, + 2981.216, + 2994.186, + 2985.242 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 536870912, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 7164.524, + "p90": 7331.675, + "p95": 7344.524, + "p99": 7677.532 + }, + "payload_gbps_at_latency_percentile": { + "p50": 74.9346239889768, + "p90": 73.22622893131515, + "p95": 73.09812208388182, + "p99": 69.92753817242311 + }, + "samples_us": [ + 7677.532, + 7344.524, + 7152.005, + 7155.801, + 7308.469, + 7324.504, + 7151.344, + 7190.332, + 7164.524, + 7288.62, + 7194.098, + 7178.045, + 7154.528, + 7148.931, + 7128.57, + 7160.458, + 7129.471, + 7134.87, + 7250.893, + 7331.675 + ] + } + } + ] +} diff --git a/collectivex/results/mi300x/mi300x-swap-blocks-smoke-random.json b/collectivex/results/mi300x/mi300x-swap-blocks-smoke-random.json new file mode 100644 index 0000000000..1e43e60b88 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-swap-blocks-smoke-random.json @@ -0,0 +1,4075 @@ +{ + "schema": "collectivex-swap-blocks-v1", + "case_id": "mi300x-swap-blocks-smoke-random", + "profile": "smoke", + "generated_at": "2026-10-28T08:29:44.277056+00:00", + "operation": "vllm._custom_ops.swap_blocks", + "timing": "drained-wall-clock-including-submission-and-synchronization", + "warmup": 4, + "iterations": 20, + "selection": { + "requested_block_bytes": [ + 257, + 4096, + 65536, + 262144 + ], + "requested_num_blocks": [ + 1, + 4, + 16, + 64, + 256, + 1024, + 2048 + ], + "max_payload_bytes": 2147483648, + "skipped_cases": [] + }, + "runtime": { + "sku": "mi300x", + "torch": "2.13.0+git733fca1", + "vllm": "0.32.0", + "cuda": null, + "hip": "7.2.53211", + "device": "", + "device_index": 0, + "image": "vllm/vllm-openai-rocm:v0.32.0", + "source_sha": "6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4", + "source_dirty": false + }, + "cases": [ + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 257, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 18.117, + "p90": 20.22, + "p95": 21.342, + "p99": 32.719 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.014185571562620743, + "p90": 0.012710187932739862, + "p95": 0.012041982944428826, + "p99": 0.007854763287386534 + }, + "samples_us": [ + 18.808, + 19.109, + 19.138, + 32.719, + 18.759, + 17.877, + 21.342, + 17.796, + 18.027, + 18.207, + 18.347, + 18.117, + 17.877, + 20.22, + 18.107, + 17.967, + 18.057, + 17.937, + 17.947, + 18.227 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 1028, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 28.323, + "p90": 29.184, + "p95": 29.354, + "p99": 30.216 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.036295590156409985, + "p90": 0.03522478070175439, + "p95": 0.03502078081351775, + "p99": 0.03402171035213132 + }, + "samples_us": [ + 28.242, + 28.603, + 29.354, + 28.373, + 28.172, + 27.902, + 29.184, + 30.216, + 27.902, + 28.563, + 28.262, + 28.342, + 28.432, + 28.102, + 29.153, + 28.323, + 28.012, + 27.862, + 28.223, + 28.473 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 4112, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 74.031, + "p90": 81.823, + "p95": 189.284, + "p99": 200.14 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.05554429900987424, + "p90": 0.05025481832736517, + "p95": 0.02172397033029733, + "p99": 0.020545618067352854 + }, + "samples_us": [ + 200.14, + 74.832, + 75.403, + 76.605, + 75.583, + 81.823, + 73.911, + 73.199, + 73.11, + 72.859, + 74.762, + 189.284, + 72.599, + 74.522, + 73.01, + 71.437, + 74.963, + 73.84, + 72.149, + 74.031 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 16448, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 242.733, + "p90": 249.514, + "p95": 249.924, + "p99": 254.842 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.06776169700864736, + "p90": 0.06592014876920733, + "p95": 0.06581200685008243, + "p99": 0.06454195148366439 + }, + "samples_us": [ + 249.514, + 240.26, + 242.733, + 254.842, + 246.74, + 235.753, + 244.286, + 240.991, + 246.449, + 238.397, + 243.004, + 247.461, + 249.924, + 235.883, + 239.519, + 237.335, + 246.199, + 234.581, + 240.661, + 246.159 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 65792, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 894.039, + "p90": 930.714, + "p95": 934.82, + "p99": 1010.975 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.0735896308774002, + "p90": 0.0706898144865125, + "p95": 0.07037932436190925, + "p99": 0.06507777145824575 + }, + "samples_us": [ + 1010.975, + 930.714, + 896.433, + 912.767, + 887.901, + 934.82, + 891.576, + 891.126, + 883.513, + 882.091, + 894.17, + 899.869, + 889.392, + 894.039, + 887.48, + 898.886, + 880.5, + 897.734, + 889.403, + 900.709 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 263168, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3508.176, + "p90": 3580.504, + "p95": 3584.38, + "p99": 3613.364 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.0750156206530117, + "p90": 0.0735002670015171, + "p95": 0.0734207868585362, + "p99": 0.0728318541945954 + }, + "samples_us": [ + 3470.009, + 3465.071, + 3568.847, + 3584.38, + 3512.653, + 3437.781, + 3580.504, + 3519.423, + 3498.061, + 3491.541, + 3534.576, + 3514.816, + 3482.037, + 3539.894, + 3613.364, + 3508.176, + 3521.556, + 3467.495, + 3470.199, + 3476.849 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 526336, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 7023.773, + "p90": 7064.324, + "p95": 7069.922, + "p99": 7118.705 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.0749363625504412, + "p90": 0.07450620894511634, + "p95": 0.07444721455201345, + "p99": 0.07393704332459344 + }, + "samples_us": [ + 7024.013, + 6987.228, + 6986.187, + 7118.705, + 7034.589, + 7038.606, + 7008.58, + 7005.626, + 7061.55, + 7069.922, + 7023.773, + 7017.384, + 7038.284, + 7012.086, + 7064.324, + 7019.767, + 7059.666, + 7015.0, + 7063.032, + 6985.275 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 4096, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 21.612, + "p90": 22.164, + "p95": 22.193, + "p99": 23.856 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.1895243383305571, + "p90": 0.18480418696986103, + "p95": 0.1845626999504348, + "p99": 0.17169684775318578 + }, + "samples_us": [ + 22.104, + 21.983, + 21.672, + 21.172, + 21.843, + 21.592, + 21.612, + 21.401, + 21.703, + 21.222, + 21.612, + 21.382, + 21.502, + 21.432, + 21.532, + 22.164, + 21.532, + 22.193, + 21.622, + 23.856 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 16384, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 31.948, + "p90": 32.739, + "p95": 32.759, + "p99": 33.92 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.512833354200576, + "p90": 0.5004428968508507, + "p95": 0.5001373668304894, + "p99": 0.4830188679245283 + }, + "samples_us": [ + 32.739, + 32.048, + 31.487, + 32.549, + 31.317, + 32.759, + 32.058, + 31.517, + 31.217, + 31.417, + 31.588, + 33.92, + 31.858, + 31.677, + 32.539, + 32.058, + 32.458, + 32.159, + 31.907, + 31.948 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 82.865, + "p90": 84.227, + "p95": 87.231, + "p99": 90.967 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.7908767272069028, + "p90": 0.7780877865767509, + "p95": 0.7512925450814504, + "p99": 0.7204370815790342 + }, + "samples_us": [ + 81.622, + 81.502, + 80.822, + 82.653, + 82.924, + 83.095, + 83.575, + 81.853, + 83.916, + 90.967, + 83.094, + 81.973, + 84.086, + 87.231, + 82.865, + 82.965, + 80.711, + 81.863, + 81.061, + 84.227 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 272.247, + "p90": 282.303, + "p95": 285.608, + "p99": 286.269 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.9628903165140479, + "p90": 0.9285909111840823, + "p95": 0.9178454385031232, + "p99": 0.9157261177424031 + }, + "samples_us": [ + 277.816, + 267.771, + 286.269, + 279.76, + 279.468, + 270.706, + 285.608, + 271.397, + 278.337, + 279.869, + 272.247, + 273.67, + 277.596, + 282.303, + 270.835, + 249.664, + 245.167, + 256.113, + 241.281, + 245.177 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 900.88, + "p90": 909.894, + "p95": 933.619, + "p99": 1021.18 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.1639463635556344, + "p90": 1.1524155560977432, + "p95": 1.1231305275492465, + "p99": 1.0268277874615641 + }, + "samples_us": [ + 1021.18, + 933.619, + 895.471, + 909.894, + 901.34, + 893.158, + 892.897, + 901.651, + 896.072, + 902.111, + 892.328, + 898.787, + 908.061, + 896.082, + 889.282, + 890.154, + 905.116, + 908.0, + 900.88, + 903.183 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3537.951, + "p90": 3559.683, + "p95": 3568.466, + "p99": 3579.943 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.1855178322141828, + "p90": 1.1782802007931605, + "p95": 1.175380121318236, + "p99": 1.1716119502461353 + }, + "samples_us": [ + 3537.951, + 3553.073, + 3557.489, + 3528.687, + 3552.372, + 3555.837, + 3503.509, + 3559.683, + 3498.601, + 3543.289, + 3525.823, + 3568.466, + 3522.057, + 3508.446, + 3542.497, + 3526.203, + 3579.943, + 3535.978, + 3536.459, + 3552.673 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 8388608, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 7064.324, + "p90": 7091.484, + "p95": 7097.673, + "p99": 7192.035 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.1874608242770293, + "p90": 1.18291291357352, + "p95": 1.1818814419880994, + "p99": 1.1663747465077687 + }, + "samples_us": [ + 7057.764, + 7050.413, + 7074.74, + 7083.503, + 7070.033, + 7064.324, + 7077.824, + 7023.733, + 7082.612, + 7063.392, + 7192.035, + 7091.484, + 7072.836, + 7036.442, + 7097.673, + 7070.383, + 7057.524, + 7037.293, + 7045.316, + 7042.511 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 19.54, + "p90": 20.039, + "p95": 20.17, + "p99": 21.161 + }, + "payload_gbps_at_latency_percentile": { + "p50": 3.353940634595701, + "p90": 3.270422675782225, + "p95": 3.249181953396133, + "p99": 3.0970180993336798 + }, + "samples_us": [ + 20.17, + 19.71, + 19.289, + 19.499, + 19.54, + 20.0, + 19.659, + 19.279, + 19.099, + 18.918, + 19.039, + 19.409, + 18.828, + 19.94, + 19.749, + 19.81, + 20.039, + 19.269, + 21.161, + 19.57 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 66.029, + "p90": 66.59, + "p95": 66.7, + "p99": 80.951 + }, + "payload_gbps_at_latency_percentile": { + "p50": 3.970134334913447, + "p90": 3.936687190268809, + "p95": 3.9301949025487257, + "p99": 3.238304653432323 + }, + "samples_us": [ + 64.367, + 63.265, + 64.286, + 65.398, + 65.348, + 66.029, + 66.049, + 66.37, + 66.7, + 66.28, + 80.951, + 66.169, + 66.008, + 65.839, + 65.969, + 66.079, + 66.049, + 66.59, + 66.049, + 65.979 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 245.618, + "p90": 248.442, + "p95": 248.563, + "p99": 249.835 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.269133369704175, + "p90": 4.220606821712915, + "p95": 4.21855223826555, + "p99": 4.197074068885464 + }, + "samples_us": [ + 245.408, + 244.706, + 248.563, + 245.548, + 246.42, + 245.298, + 249.835, + 244.276, + 244.897, + 245.618, + 244.426, + 246.45, + 245.658, + 248.442, + 245.017, + 247.371, + 245.758, + 247.231, + 246.419, + 243.715 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 965.216, + "p90": 967.419, + "p95": 968.091, + "p99": 969.844 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.3454563538109605, + "p90": 4.335560910009003, + "p95": 4.332551382049828, + "p99": 4.324720264289927 + }, + "samples_us": [ + 969.844, + 968.091, + 964.505, + 965.387, + 965.076, + 964.354, + 965.656, + 966.178, + 967.069, + 965.216, + 966.738, + 965.106, + 963.984, + 967.419, + 965.306, + 965.156, + 960.449, + 962.842, + 966.578, + 946.528 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3658.552, + "p90": 3780.154, + "p95": 3814.395, + "p99": 3868.326 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.585753052027141, + "p90": 4.438236114190056, + "p95": 4.398395027258582, + "p99": 4.337073969463794 + }, + "samples_us": [ + 3868.326, + 3814.395, + 3780.154, + 3749.958, + 3750.519, + 3745.242, + 3737.981, + 3733.424, + 3721.836, + 3675.667, + 3654.115, + 3648.526, + 3651.701, + 3658.552, + 3647.195, + 3639.222, + 3634.975, + 3637.349, + 3633.483, + 3632.382 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14510.93, + "p90": 14544.831, + "p95": 14549.799, + "p99": 14554.125 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.6247114416512245, + "p90": 4.61393219350572, + "p95": 4.612356775512844, + "p99": 4.6109858201712575 + }, + "samples_us": [ + 14544.831, + 14536.699, + 14475.196, + 14468.046, + 14549.799, + 14505.653, + 14525.552, + 14505.121, + 14480.715, + 14495.217, + 14509.969, + 14484.921, + 14528.387, + 14495.116, + 14521.316, + 14554.125, + 14528.908, + 14543.449, + 14510.93, + 14544.431 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 134217728, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 28849.322, + "p90": 28915.882, + "p95": 28926.958, + "p99": 28987.619 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.652370270608093, + "p90": 4.641661215798294, + "p95": 4.639883944934687, + "p99": 4.630174282337573 + }, + "samples_us": [ + 28855.101, + 28871.857, + 28866.228, + 28926.958, + 28881.24, + 28782.012, + 28987.619, + 28712.047, + 28793.129, + 28847.169, + 28740.359, + 28841.36, + 28915.882, + 28848.572, + 28849.322, + 28901.821, + 28822.693, + 28850.504, + 28874.54, + 28793.709 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 23.756, + "p90": 24.357, + "p95": 24.466, + "p99": 33.571 + }, + "payload_gbps_at_latency_percentile": { + "p50": 11.03485435258461, + "p90": 10.7625733875272, + "p95": 10.714624376686013, + "p99": 7.8086443656727536 + }, + "samples_us": [ + 24.357, + 23.625, + 24.087, + 24.026, + 23.585, + 23.885, + 24.247, + 24.466, + 23.796, + 23.916, + 23.726, + 23.675, + 23.656, + 23.496, + 23.806, + 23.546, + 33.571, + 23.565, + 23.756, + 23.605 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 82.434, + "p90": 83.154, + "p95": 83.455, + "p99": 95.874 + }, + "payload_gbps_at_latency_percentile": { + "p50": 12.720188271829585, + "p90": 12.610048825071553, + "p95": 12.564567731112575, + "p99": 10.937021507395123 + }, + "samples_us": [ + 82.675, + 81.352, + 82.093, + 83.105, + 81.883, + 82.934, + 83.154, + 81.963, + 82.774, + 95.874, + 81.592, + 81.552, + 82.784, + 82.473, + 82.294, + 81.662, + 83.455, + 82.644, + 82.373, + 82.434 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 330.676, + "p90": 334.502, + "p95": 335.422, + "p99": 339.088 + }, + "payload_gbps_at_latency_percentile": { + "p50": 12.68402907982436, + "p90": 12.538950439758208, + "p95": 12.504558436834794, + "p99": 12.369367243901289 + }, + "samples_us": [ + 339.088, + 330.676, + 330.646, + 334.502, + 333.68, + 333.049, + 330.725, + 335.422, + 332.839, + 328.091, + 329.454, + 332.979, + 333.99, + 319.178, + 319.319, + 330.715, + 324.877, + 317.916, + 320.37, + 322.703 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 1237.625, + "p90": 1265.076, + "p95": 1275.791, + "p99": 1300.909 + }, + "payload_gbps_at_latency_percentile": { + "p50": 13.555976972023029, + "p90": 13.261824586032775, + "p95": 13.150442353018637, + "p99": 12.896533116459338 + }, + "samples_us": [ + 1300.909, + 1275.791, + 1265.076, + 1256.853, + 1248.721, + 1236.422, + 1237.223, + 1237.625, + 1238.115, + 1242.301, + 1238.696, + 1247.219, + 1248.541, + 1236.412, + 1233.628, + 1215.101, + 1215.782, + 1213.418, + 1188.05, + 1187.339 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 4548.895, + "p90": 4789.556, + "p95": 4871.839, + "p99": 4994.163 + }, + "payload_gbps_at_latency_percentile": { + "p50": 14.75278369801897, + "p90": 14.011500022131488, + "p95": 13.77485257620377, + "p99": 13.437459690442623 + }, + "samples_us": [ + 4994.163, + 4871.839, + 4789.556, + 4725.81, + 4676.837, + 4623.487, + 4594.914, + 4549.696, + 4548.895, + 4549.746, + 4552.38, + 4547.773, + 4546.452, + 4540.112, + 4544.298, + 4545.37, + 4548.505, + 4548.765, + 4535.906, + 4547.082 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 268435456, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 18096.282, + "p90": 18138.175, + "p95": 18139.436, + "p99": 18155.151 + }, + "payload_gbps_at_latency_percentile": { + "p50": 14.83373523909497, + "p90": 14.799474368286777, + "p95": 14.79844555255191, + "p99": 14.785636098537545 + }, + "samples_us": [ + 18092.857, + 18098.455, + 18096.282, + 18058.295, + 18136.783, + 18138.175, + 18102.622, + 18085.146, + 18087.329, + 18097.925, + 18090.994, + 18099.527, + 18077.765, + 18103.253, + 18082.531, + 18096.172, + 18155.151, + 18096.583, + 18028.09, + 18139.436 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 536870912, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 36023.38, + "p90": 36086.365, + "p95": 36094.477, + "p99": 36149.379 + }, + "payload_gbps_at_latency_percentile": { + "p50": 14.903401957284409, + "p90": 14.877389617934641, + "p95": 14.87404602094664, + "p99": 14.851456009797568 + }, + "samples_us": [ + 36020.427, + 36086.365, + 35987.196, + 36021.899, + 36149.379, + 35971.914, + 36094.477, + 36047.527, + 36040.966, + 36046.235, + 36050.331, + 36040.226, + 35990.331, + 36024.683, + 36009.981, + 36007.167, + 36018.614, + 36037.351, + 36002.159, + 36023.38 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 257, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 18.958, + "p90": 19.309, + "p95": 19.329, + "p99": 19.369 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.013556282308260365, + "p90": 0.013309855507794293, + "p95": 0.013296083604945936, + "p99": 0.013268625122618618 + }, + "samples_us": [ + 19.309, + 19.169, + 18.958, + 19.128, + 19.329, + 18.959, + 19.369, + 18.368, + 19.089, + 19.239, + 19.018, + 18.517, + 18.578, + 18.458, + 18.477, + 18.778, + 19.279, + 18.627, + 18.458, + 18.418 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 1028, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 27.831, + "p90": 28.212, + "p95": 28.232, + "p99": 29.564 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.03693722827063347, + "p90": 0.03643839500921594, + "p95": 0.036412581467837914, + "p99": 0.03477202002435394 + }, + "samples_us": [ + 28.232, + 27.401, + 27.982, + 27.691, + 27.452, + 27.34, + 28.193, + 27.912, + 28.022, + 27.711, + 27.872, + 28.212, + 27.401, + 27.241, + 27.241, + 28.152, + 29.564, + 27.831, + 27.902, + 27.321 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 4112, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 71.327, + "p90": 75.013, + "p95": 75.013, + "p99": 75.563 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.05764997826909866, + "p90": 0.054817165024729046, + "p95": 0.054817165024729046, + "p99": 0.054418167621719624 + }, + "samples_us": [ + 69.153, + 72.839, + 73.63, + 66.77, + 75.563, + 68.152, + 68.293, + 73.421, + 74.462, + 68.493, + 75.013, + 71.818, + 71.327, + 73.721, + 68.723, + 66.63, + 73.46, + 69.845, + 68.243, + 75.013 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 16448, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 233.81, + "p90": 238.237, + "p95": 242.974, + "p99": 244.246 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.07034771823275308, + "p90": 0.06904049329029496, + "p95": 0.06769448582975957, + "p99": 0.06734194214030118 + }, + "samples_us": [ + 218.017, + 237.306, + 237.746, + 220.781, + 233.81, + 220.651, + 233.96, + 222.043, + 227.801, + 218.848, + 224.386, + 235.443, + 230.996, + 221.162, + 234.561, + 235.814, + 244.246, + 236.063, + 238.237, + 242.974 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 65792, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 873.85, + "p90": 883.624, + "p95": 885.397, + "p99": 895.261 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.07528980946386679, + "p90": 0.0744570088634985, + "p95": 0.07430790933332732, + "p99": 0.07348918360120679 + }, + "samples_us": [ + 895.261, + 885.397, + 883.624, + 883.283, + 870.455, + 882.753, + 868.892, + 870.114, + 864.415, + 870.625, + 868.801, + 878.326, + 867.089, + 858.947, + 872.617, + 873.85, + 875.462, + 874.991, + 877.134, + 877.585 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 263168, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3397.751, + "p90": 3415.587, + "p95": 3419.023, + "p99": 3420.024 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.077453586210408, + "p90": 0.07704912801225675, + "p95": 0.07697169630037587, + "p99": 0.07694916760818053 + }, + "samples_us": [ + 3420.024, + 3402.949, + 3400.234, + 3395.607, + 3389.759, + 3415.587, + 3390.069, + 3401.857, + 3387.185, + 3368.507, + 3390.52, + 3401.606, + 3385.111, + 3412.934, + 3381.697, + 3367.665, + 3419.023, + 3400.515, + 3397.751, + 3399.964 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 526336, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 6743.723, + "p90": 6828.56, + "p95": 6831.465, + "p99": 6847.619 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.07804828282537703, + "p90": 0.07707862272572841, + "p95": 0.07704584594958767, + "p99": 0.07686408954703818 + }, + "samples_us": [ + 6823.563, + 6847.619, + 6831.465, + 6721.34, + 6729.532, + 6743.723, + 6761.13, + 6741.941, + 6733.377, + 6725.246, + 6828.56, + 6727.479, + 6733.468, + 6750.374, + 6750.894, + 6735.581, + 6722.541, + 6819.608, + 6750.784, + 6744.885 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 4096, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 18.838, + "p90": 19.249, + "p95": 19.349, + "p99": 20.13 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.21743284849771738, + "p90": 0.21279027481947113, + "p95": 0.21169052664220372, + "p99": 0.20347739692001987 + }, + "samples_us": [ + 20.13, + 18.988, + 19.349, + 18.377, + 18.768, + 18.578, + 18.798, + 18.788, + 18.878, + 18.968, + 18.698, + 18.818, + 18.869, + 18.938, + 19.109, + 18.838, + 19.249, + 18.518, + 18.819, + 18.949 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 16384, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 27.762, + "p90": 28.303, + "p95": 28.383, + "p99": 28.703 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.5901592104315251, + "p90": 0.5788785641098116, + "p95": 0.5772469435929958, + "p99": 0.5708114134411038 + }, + "samples_us": [ + 28.213, + 28.703, + 28.383, + 28.072, + 28.082, + 27.551, + 27.912, + 27.851, + 28.082, + 28.303, + 27.161, + 27.371, + 27.762, + 27.422, + 27.471, + 27.481, + 27.531, + 27.832, + 27.181, + 27.651 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 74.632, + "p90": 76.535, + "p95": 88.192, + "p99": 90.376 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.8781219852074177, + "p90": 0.8562879728228915, + "p95": 0.7431059506531205, + "p99": 0.7251482694520669 + }, + "samples_us": [ + 72.508, + 73.56, + 88.192, + 73.831, + 76.525, + 76.204, + 74.011, + 75.343, + 72.148, + 74.111, + 75.093, + 74.461, + 76.455, + 76.535, + 75.814, + 90.376, + 74.632, + 74.492, + 75.754, + 73.971 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 237.446, + "p90": 248.532, + "p95": 258.307, + "p99": 263.565 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.1040152287256892, + "p90": 1.0547696071330854, + "p95": 1.0148544174180334, + "p99": 0.9946085405877109 + }, + "samples_us": [ + 237.446, + 235.383, + 233.14, + 248.532, + 263.565, + 233.059, + 236.785, + 244.155, + 238.237, + 231.537, + 237.125, + 237.876, + 245.999, + 235.653, + 240.67, + 236.705, + 258.307, + 237.706, + 238.206, + 234.311 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 866.609, + "p90": 878.586, + "p95": 883.233, + "p99": 895.241 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.2099758945499066, + "p90": 1.1934813438866543, + "p95": 1.1872020180405396, + "p99": 1.1712779017046806 + }, + "samples_us": [ + 869.343, + 883.233, + 871.186, + 865.067, + 878.586, + 858.987, + 870.515, + 870.244, + 860.319, + 863.885, + 866.608, + 895.241, + 866.438, + 866.609, + 863.393, + 871.156, + 865.647, + 866.418, + 870.695, + 872.578 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3298.071, + "p90": 3335.778, + "p95": 3340.836, + "p99": 3393.905 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.2717446046492025, + "p90": 1.2573690455419995, + "p95": 1.2554653984810988, + "p99": 1.2358342381416096 + }, + "samples_us": [ + 3393.905, + 3340.836, + 3335.778, + 3324.621, + 3298.371, + 3297.531, + 3321.827, + 3298.071, + 3286.785, + 3311.211, + 3297.801, + 3299.754, + 3276.549, + 3296.078, + 3294.806, + 3297.53, + 3287.465, + 3296.038, + 3305.823, + 3306.043 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 8388608, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 6659.537, + "p90": 6836.012, + "p95": 6848.119, + "p99": 6866.067 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.2596383202015395, + "p90": 1.2271201396369695, + "p95": 1.2249506762367885, + "p99": 1.2217486371746737 + }, + "samples_us": [ + 6559.097, + 6560.89, + 6574.15, + 6553.649, + 6681.39, + 6656.644, + 6682.562, + 6677.254, + 6646.818, + 6537.705, + 6666.237, + 6650.915, + 6675.061, + 6659.537, + 6646.749, + 6663.594, + 6761.711, + 6836.012, + 6866.067, + 6848.119 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 16.555, + "p90": 16.865, + "p95": 16.916, + "p99": 16.965 + }, + "payload_gbps_at_latency_percentile": { + "p50": 3.9586831772878286, + "p90": 3.8859175807886155, + "p95": 3.8742019389926696, + "p99": 3.8630120837017388 + }, + "samples_us": [ + 16.865, + 16.715, + 16.404, + 16.605, + 16.555, + 16.334, + 16.785, + 16.675, + 16.916, + 16.374, + 16.865, + 16.965, + 16.205, + 16.776, + 16.445, + 16.335, + 16.435, + 16.455, + 16.174, + 16.555 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 57.957, + "p90": 58.918, + "p95": 58.938, + "p99": 59.309 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.523077453974499, + "p90": 4.449302420312978, + "p95": 4.447792595608945, + "p99": 4.419969987691581 + }, + "samples_us": [ + 57.746, + 57.696, + 58.007, + 57.015, + 57.356, + 58.778, + 57.957, + 58.838, + 58.938, + 59.309, + 57.826, + 57.937, + 57.447, + 58.398, + 56.645, + 58.858, + 55.493, + 58.377, + 58.918, + 58.507 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 223.254, + "p90": 225.247, + "p95": 225.328, + "p99": 225.408 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.696784828043395, + "p90": 4.655227372617615, + "p95": 4.653553930270539, + "p99": 4.651902328222601 + }, + "samples_us": [ + 223.394, + 224.426, + 222.483, + 222.744, + 223.865, + 224.617, + 220.771, + 223.254, + 224.977, + 224.877, + 225.007, + 220.991, + 222.163, + 225.328, + 222.193, + 222.083, + 225.247, + 225.408, + 220.891, + 221.132 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 877.475, + "p90": 880.269, + "p95": 880.349, + "p99": 882.712 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.779969799709393, + "p90": 4.764798033328448, + "p95": 4.76436504159146, + "p99": 4.751610944453004 + }, + "samples_us": [ + 872.388, + 876.524, + 875.532, + 877.385, + 877.885, + 879.488, + 874.511, + 877.475, + 879.127, + 882.712, + 880.349, + 877.835, + 878.366, + 879.617, + 879.588, + 880.269, + 876.774, + 876.453, + 874.34, + 875.211 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3525.502, + "p90": 3531.06, + "p95": 3531.842, + "p99": 3531.962 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.75881619128283, + "p90": 4.75132566424813, + "p95": 4.750273653238168, + "p99": 4.750112260550934 + }, + "samples_us": [ + 3529.778, + 3531.962, + 3520.144, + 3528.186, + 3523.959, + 3525.442, + 3531.842, + 3525.502, + 3529.748, + 3524.41, + 3518.321, + 3527.515, + 3525.162, + 3529.228, + 3531.06, + 3525.221, + 3530.64, + 3528.947, + 3518.692, + 3519.593 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 13871.512, + "p90": 13886.415, + "p95": 13908.287, + "p99": 13923.32 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.837891067678852, + "p90": 4.832699008347367, + "p95": 4.8250991656988385, + "p99": 4.8198895091113325 + }, + "samples_us": [ + 13836.751, + 13867.727, + 13871.512, + 13881.176, + 13859.695, + 13855.168, + 13864.542, + 13872.524, + 13886.415, + 13874.777, + 13880.827, + 13857.291, + 13879.735, + 13868.097, + 13908.287, + 13827.526, + 13882.71, + 13851.763, + 13923.32, + 13876.62 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 134217728, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 27573.371, + "p90": 28038.678, + "p95": 28050.966, + "p99": 28055.764 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.867657567150567, + "p90": 4.78687789773826, + "p95": 4.784780959058594, + "p99": 4.783962682320824 + }, + "samples_us": [ + 27989.334, + 28050.966, + 28038.678, + 28055.764, + 27993.37, + 27821.492, + 27561.293, + 27567.412, + 27499.34, + 27551.839, + 27528.254, + 27527.442, + 27581.733, + 27542.805, + 27545.128, + 27614.493, + 27573.371, + 27562.724, + 27617.036, + 27602.024 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 19.529, + "p90": 19.89, + "p95": 19.9, + "p99": 20.171 + }, + "payload_gbps_at_latency_percentile": { + "p50": 13.423319166367966, + "p90": 13.179688285570638, + "p95": 13.173065326633166, + "p99": 12.996083486193049 + }, + "samples_us": [ + 19.66, + 20.171, + 19.509, + 19.89, + 19.369, + 19.539, + 19.559, + 19.549, + 19.539, + 19.349, + 19.9, + 19.889, + 19.469, + 19.309, + 19.529, + 19.469, + 19.329, + 19.329, + 19.439, + 19.61 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 71.987, + "p90": 73.33, + "p95": 73.521, + "p99": 82.874 + }, + "payload_gbps_at_latency_percentile": { + "p50": 14.566185561281898, + "p90": 14.299413609709532, + "p95": 14.26226520313924, + "p99": 12.652653425682361 + }, + "samples_us": [ + 73.521, + 73.33, + 82.874, + 72.829, + 71.798, + 71.878, + 71.788, + 71.987, + 72.849, + 73.33, + 72.178, + 71.447, + 71.427, + 72.588, + 72.839, + 70.856, + 71.537, + 72.128, + 71.367, + 71.667 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 276.524, + "p90": 279.028, + "p95": 279.198, + "p99": 285.528 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.16795648840607, + "p90": 15.031839098585088, + "p95": 15.02268640892843, + "p99": 14.689641646353422 + }, + "samples_us": [ + 275.713, + 276.685, + 276.874, + 279.198, + 278.718, + 275.423, + 278.547, + 285.528, + 276.524, + 277.085, + 275.773, + 279.028, + 275.242, + 278.738, + 274.771, + 275.973, + 275.814, + 275.733, + 274.05, + 277.476 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 1112.157, + "p90": 1115.392, + "p95": 1116.112, + "p99": 1117.535 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.085294612181553, + "p90": 15.041542345650678, + "p95": 15.031839098585088, + "p99": 15.012698483716393 + }, + "samples_us": [ + 1113.188, + 1112.157, + 1111.786, + 1117.535, + 1109.262, + 1108.54, + 1116.112, + 1106.718, + 1111.906, + 1110.654, + 1115.392, + 1114.91, + 1111.975, + 1114.019, + 1114.39, + 1112.097, + 1114.209, + 1111.315, + 1113.569, + 1113.759 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 4388.836, + "p90": 4392.701, + "p95": 4392.802, + "p99": 4394.024 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.290811504462686, + "p90": 15.277357598434312, + "p95": 15.277006339006402, + "p99": 15.272757727313278 + }, + "samples_us": [ + 4389.797, + 4379.541, + 4392.701, + 4388.675, + 4392.802, + 4388.836, + 4389.717, + 4381.153, + 4386.863, + 4390.829, + 4391.579, + 4383.357, + 4383.287, + 4385.24, + 4389.466, + 4382.997, + 4391.449, + 4392.571, + 4394.024, + 4385.761 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 268435456, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 17682.522, + "p90": 17713.598, + "p95": 17739.748, + "p99": 17741.271 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.180835403456589, + "p90": 15.15420277687232, + "p95": 15.131864105397664, + "p99": 15.130565110019457 + }, + "samples_us": [ + 17465.307, + 17452.708, + 17462.322, + 17496.222, + 17407.78, + 17669.503, + 17682.522, + 17685.857, + 17690.074, + 17687.56, + 17713.598, + 17682.782, + 17699.968, + 17682.572, + 17675.051, + 17739.748, + 17712.627, + 17661.561, + 17741.271, + 17681.931 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 536870912, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 34841.6, + "p90": 34949.742, + "p95": 35067.749, + "p99": 35188.88 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.408905216752387, + "p90": 15.361226758125998, + "p95": 15.309534467125335, + "p99": 15.256834318114132 + }, + "samples_us": [ + 34819.417, + 34843.953, + 34831.786, + 34789.512, + 34757.004, + 34832.246, + 34841.6, + 34731.226, + 34847.258, + 34813.298, + 34880.688, + 34880.799, + 34904.024, + 34780.889, + 34949.742, + 34873.007, + 34879.648, + 34822.392, + 35067.749, + 35188.88 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 257, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 18.748, + "p90": 19.699, + "p95": 19.81, + "p99": 20.601 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.013708128867079155, + "p90": 0.013046347530331489, + "p95": 0.012973245835436648, + "p99": 0.012475122566865686 + }, + "samples_us": [ + 19.699, + 18.869, + 18.908, + 18.548, + 19.81, + 19.409, + 18.928, + 18.748, + 18.828, + 17.997, + 19.42, + 18.798, + 18.588, + 18.278, + 18.588, + 18.648, + 20.601, + 18.597, + 18.308, + 18.357 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 1028, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 27.992, + "p90": 28.923, + "p95": 28.974, + "p99": 37.245 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.036724778508145185, + "p90": 0.035542647719807766, + "p95": 0.03548008559398081, + "p99": 0.027601020271177338 + }, + "samples_us": [ + 27.992, + 28.162, + 37.245, + 28.923, + 27.772, + 28.032, + 28.062, + 28.082, + 28.332, + 27.882, + 28.092, + 27.992, + 28.974, + 27.431, + 27.671, + 27.581, + 27.24, + 27.862, + 28.072, + 27.982 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 4112, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 71.757, + "p90": 74.432, + "p95": 78.408, + "p99": 83.405 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.05730451384533913, + "p90": 0.05524505588993981, + "p95": 0.052443628201203955, + "p99": 0.04930160062346382 + }, + "samples_us": [ + 69.264, + 72.809, + 73.069, + 83.405, + 70.085, + 71.978, + 67.892, + 71.136, + 78.408, + 70.425, + 74.422, + 73.519, + 73.219, + 73.711, + 68.853, + 74.432, + 68.443, + 68.553, + 71.197, + 71.757 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 16448, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 234.041, + "p90": 241.522, + "p95": 253.26, + "p99": 254.771 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.07027828457407036, + "p90": 0.06810145659608649, + "p95": 0.06494511569138435, + "p99": 0.0645599381405262 + }, + "samples_us": [ + 238.447, + 234.601, + 223.575, + 222.803, + 228.903, + 221.732, + 235.463, + 234.041, + 253.26, + 231.797, + 240.44, + 236.344, + 241.252, + 228.293, + 238.478, + 233.259, + 241.522, + 232.398, + 254.771, + 232.679 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 65792, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 878.546, + "p90": 894.78, + "p95": 913.969, + "p99": 1005.586 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.07488737072390063, + "p90": 0.07352868861619616, + "p95": 0.07198493603174724, + "p99": 0.06542652741784392 + }, + "samples_us": [ + 1005.586, + 913.969, + 860.519, + 863.824, + 863.875, + 881.23, + 880.129, + 878.546, + 876.483, + 874.781, + 892.878, + 894.78, + 875.853, + 879.909, + 890.023, + 881.581, + 875.602, + 876.213, + 873.689, + 891.475 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 263168, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3407.456, + "p90": 3422.748, + "p95": 3423.99, + "p99": 3431.492 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.07723298554698872, + "p90": 0.07688792747815498, + "p95": 0.0768600375585209, + "p99": 0.07669200452747668 + }, + "samples_us": [ + 3422.748, + 3407.496, + 3388.757, + 3414.626, + 3396.408, + 3392.723, + 3414.857, + 3431.492, + 3410.339, + 3401.056, + 3419.112, + 3411.821, + 3382.768, + 3397.871, + 3403.729, + 3391.291, + 3419.703, + 3392.112, + 3407.456, + 3423.99 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 526336, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 6769.402, + "p90": 6863.723, + "p95": 6868.05, + "p99": 6897.354 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.07775221504056046, + "p90": 0.07668374728991831, + "p95": 0.07663543509438632, + "p99": 0.07630984287597824 + }, + "samples_us": [ + 6749.873, + 6777.735, + 6785.877, + 6751.796, + 6740.439, + 6769.192, + 6868.05, + 6761.299, + 6769.402, + 6897.354, + 6768.751, + 6769.191, + 6863.723, + 6793.879, + 6839.106, + 6841.79, + 6771.495, + 6745.215, + 6778.075, + 6751.295 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 4096, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 18.699, + "p90": 19.219, + "p95": 19.219, + "p99": 19.52 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.21904914701320927, + "p90": 0.21312243092772776, + "p95": 0.21312243092772776, + "p99": 0.2098360655737705 + }, + "samples_us": [ + 19.52, + 18.699, + 18.889, + 19.219, + 18.718, + 19.219, + 19.069, + 18.657, + 18.758, + 18.458, + 18.758, + 18.438, + 18.288, + 18.488, + 18.728, + 18.658, + 18.598, + 18.779, + 18.548, + 18.377 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 16384, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 28.263, + "p90": 28.793, + "p95": 28.953, + "p99": 29.445 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.5796978381629693, + "p90": 0.5690271941096794, + "p95": 0.5658826373778192, + "p99": 0.5564272372219392 + }, + "samples_us": [ + 28.623, + 28.953, + 27.932, + 28.793, + 28.153, + 28.763, + 28.263, + 29.445, + 28.603, + 28.493, + 27.791, + 27.732, + 28.343, + 28.572, + 27.751, + 28.242, + 28.322, + 28.051, + 27.531, + 27.782 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 65536, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 74.422, + "p90": 76.785, + "p95": 85.829, + "p99": 85.999 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.8805998226330923, + "p90": 0.8535000325584424, + "p95": 0.7635647624928638, + "p99": 0.7620553727368923 + }, + "samples_us": [ + 72.489, + 85.829, + 71.798, + 72.529, + 74.221, + 74.422, + 76.785, + 75.043, + 72.609, + 73.21, + 75.864, + 74.632, + 76.384, + 75.764, + 85.999, + 75.583, + 74.642, + 72.729, + 74.321, + 71.196 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 244.777, + "p90": 253.84, + "p95": 258.878, + "p99": 264.917 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.0709502935324806, + "p90": 1.0327135203277655, + "p95": 1.012615981273032, + "p99": 0.9895325705787096 + }, + "samples_us": [ + 264.917, + 239.008, + 241.772, + 253.08, + 245.328, + 236.494, + 240.781, + 258.878, + 249.414, + 238.837, + 238.468, + 251.968, + 244.126, + 234.681, + 243.725, + 253.84, + 250.546, + 244.777, + 247.891, + 249.203 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 876.343, + "p90": 884.816, + "p95": 885.516, + "p99": 891.856 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.1965360595109449, + "p90": 1.1850780275221062, + "p95": 1.18414122387399, + "p99": 1.1757234351733912 + }, + "samples_us": [ + 883.674, + 876.343, + 865.377, + 891.856, + 876.343, + 865.056, + 872.718, + 885.516, + 881.681, + 872.457, + 866.519, + 879.217, + 884.816, + 874.841, + 871.225, + 879.678, + 882.542, + 882.292, + 872.277, + 869.192 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3260.695, + "p90": 3273.134, + "p95": 3274.246, + "p99": 3279.574 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.2863220877757655, + "p90": 1.281433635164341, + "p95": 1.280998434448725, + "p99": 1.2789173227986317 + }, + "samples_us": [ + 3260.605, + 3240.425, + 3246.765, + 3241.276, + 3248.116, + 3258.412, + 3260.716, + 3260.695, + 3279.574, + 3266.013, + 3273.134, + 3257.05, + 3271.471, + 3257.02, + 3264.07, + 3259.683, + 3261.077, + 3262.578, + 3268.797, + 3274.246 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 8388608, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 6550.114, + "p90": 6613.208, + "p95": 6619.738, + "p99": 6630.353 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.280681221731408, + "p90": 1.2684627490924223, + "p95": 1.2672114817837201, + "p99": 1.265182713499568 + }, + "samples_us": [ + 6393.188, + 6597.275, + 6533.839, + 6583.804, + 6579.197, + 6533.348, + 6596.614, + 6595.942, + 6502.813, + 6536.123, + 6550.114, + 6619.738, + 6630.353, + 6521.782, + 6516.323, + 6606.828, + 6504.525, + 6507.62, + 6613.208, + 6581.28 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 65536, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 18.607, + "p90": 19.108, + "p95": 19.239, + "p99": 20.01 + }, + "payload_gbps_at_latency_percentile": { + "p50": 3.5221153329392165, + "p90": 3.4297676365920036, + "p95": 3.4064140547845523, + "p99": 3.2751624187906048 + }, + "samples_us": [ + 18.869, + 18.858, + 18.858, + 19.108, + 20.01, + 18.027, + 18.337, + 18.738, + 18.358, + 18.458, + 17.997, + 18.538, + 18.047, + 18.338, + 18.808, + 18.808, + 19.239, + 18.758, + 18.607, + 18.177 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 27.321, + "p90": 28.192, + "p95": 28.272, + "p99": 41.953 + }, + "payload_gbps_at_latency_percentile": { + "p50": 9.59496358112807, + "p90": 9.298524404086265, + "p95": 9.272212790039616, + "p99": 6.248516196696303 + }, + "samples_us": [ + 27.761, + 27.19, + 27.441, + 27.431, + 27.16, + 27.191, + 27.571, + 27.211, + 27.291, + 41.953, + 28.272, + 27.321, + 27.281, + 26.87, + 26.79, + 27.821, + 28.192, + 27.421, + 27.091, + 27.772 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 73.079, + "p90": 80.591, + "p95": 89.634, + "p99": 91.697 + }, + "payload_gbps_at_latency_percentile": { + "p50": 14.348526936602854, + "p90": 13.011080641759005, + "p95": 11.698418011022603, + "p99": 11.435226888556878 + }, + "samples_us": [ + 69.484, + 70.556, + 74.572, + 80.591, + 73.931, + 73.079, + 73.21, + 75.012, + 72.999, + 75.964, + 70.826, + 91.697, + 74.021, + 89.634, + 72.629, + 73.851, + 71.067, + 72.208, + 71.387, + 72.879 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 232.048, + "p90": 239.739, + "p95": 243.234, + "p99": 255.834 + }, + "payload_gbps_at_latency_percentile": { + "p50": 18.07515686409708, + "p90": 17.49529279758404, + "p95": 17.243905046169534, + "p99": 16.394630893469984 + }, + "samples_us": [ + 231.116, + 238.598, + 225.798, + 226.039, + 233.99, + 235.763, + 226.359, + 229.113, + 236.985, + 255.834, + 243.234, + 228.532, + 237.085, + 234.311, + 232.048, + 228.573, + 236.494, + 239.739, + 230.866, + 222.483 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 818.306, + "p90": 828.171, + "p95": 828.232, + "p99": 833.188 + }, + "payload_gbps_at_latency_percentile": { + "p50": 20.502374417393984, + "p90": 20.258154414969855, + "p95": 20.256662384452664, + "p99": 20.13617094821337 + }, + "samples_us": [ + 828.232, + 821.511, + 827.45, + 820.539, + 819.197, + 833.188, + 815.953, + 807.56, + 818.306, + 828.171, + 814.04, + 810.204, + 814.681, + 817.485, + 819.247, + 817.285, + 822.121, + 826.418, + 814.811, + 816.243 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3272.563, + "p90": 3285.863, + "p95": 3292.373, + "p99": 3293.985 + }, + "payload_gbps_at_latency_percentile": { + "p50": 20.506515535377012, + "p90": 20.42351248363063, + "p95": 20.383129129050687, + "p99": 20.373154097544464 + }, + "samples_us": [ + 3274.706, + 3258.812, + 3259.694, + 3273.424, + 3267.155, + 3285.863, + 3268.036, + 3276.619, + 3267.686, + 3269.238, + 3279.583, + 3272.313, + 3292.373, + 3274.296, + 3272.563, + 3271.031, + 3293.985, + 3275.096, + 3264.591, + 3280.625 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 134217728, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 6561.31, + "p90": 6662.042, + "p95": 6704.585, + "p99": 6720.299 + }, + "payload_gbps_at_latency_percentile": { + "p50": 20.455934561848167, + "p90": 20.146634920644452, + "p95": 20.018797285738042, + "p99": 19.971987555910832 + }, + "samples_us": [ + 6528.992, + 6610.043, + 6561.31, + 6593.869, + 6523.273, + 6557.304, + 6555.421, + 6517.775, + 6547.1, + 6536.734, + 6618.196, + 6605.476, + 6540.399, + 6535.101, + 6641.82, + 6662.042, + 6720.299, + 6630.614, + 6704.585, + 6627.94 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 19.319, + "p90": 19.66, + "p95": 20.05, + "p99": 23.575 + }, + "payload_gbps_at_latency_percentile": { + "p50": 13.569232361923495, + "p90": 13.333875890132248, + "p95": 13.074513715710724, + "p99": 11.119575821845174 + }, + "samples_us": [ + 19.66, + 19.499, + 19.329, + 19.6, + 18.909, + 19.489, + 19.158, + 19.149, + 19.509, + 19.039, + 18.828, + 23.575, + 19.329, + 18.828, + 20.05, + 19.319, + 19.119, + 18.808, + 18.918, + 19.369 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 28.012, + "p90": 28.843, + "p95": 29.154, + "p99": 45.188 + }, + "payload_gbps_at_latency_percentile": { + "p50": 37.43310009995716, + "p90": 36.35460943729848, + "p95": 35.96679700898676, + "p99": 23.20474462246614 + }, + "samples_us": [ + 29.154, + 28.443, + 28.202, + 28.262, + 45.188, + 27.982, + 28.613, + 27.862, + 28.012, + 28.182, + 28.242, + 26.9, + 27.771, + 27.702, + 27.602, + 28.183, + 27.732, + 27.821, + 28.843, + 27.171 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 72.188, + "p90": 77.937, + "p95": 85.688, + "p99": 106.57 + }, + "payload_gbps_at_latency_percentile": { + "p50": 58.10251011248407, + "p90": 53.81659545530364, + "p95": 48.948557557651014, + "p99": 39.357267523693345 + }, + "samples_us": [ + 72.188, + 74.061, + 74.802, + 72.108, + 76.374, + 77.937, + 71.106, + 72.058, + 69.684, + 70.977, + 73.32, + 71.558, + 73.12, + 76.334, + 106.57, + 71.808, + 72.528, + 71.037, + 85.688, + 71.077 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 227.821, + "p90": 240.68, + "p95": 245.037, + "p99": 251.477 + }, + "payload_gbps_at_latency_percentile": { + "p50": 73.64209620710997, + "p90": 69.70756190792754, + "p95": 68.46809257377457, + "p99": 66.71471347280269 + }, + "samples_us": [ + 237.375, + 226.169, + 230.315, + 234.762, + 251.477, + 223.004, + 226.96, + 227.881, + 245.037, + 218.888, + 227.821, + 224.396, + 238.938, + 224.366, + 231.687, + 228.803, + 240.68, + 225.939, + 225.428, + 221.271 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 843.974, + "p90": 848.581, + "p95": 851.185, + "p99": 853.299 + }, + "payload_gbps_at_latency_percentile": { + "p50": 79.51532156203864, + "p90": 79.08362784460175, + "p95": 78.84169011436997, + "p99": 78.64636428731312 + }, + "samples_us": [ + 843.414, + 842.061, + 853.299, + 840.73, + 845.677, + 848.581, + 845.026, + 848.241, + 851.185, + 847.279, + 840.169, + 846.007, + 840.61, + 835.963, + 833.348, + 842.623, + 847.389, + 842.021, + 843.974, + 846.428 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 268435456, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3278.893, + "p90": 3295.347, + "p95": 3309.338, + "p99": 3333.865 + }, + "payload_gbps_at_latency_percentile": { + "p50": 81.86770840036561, + "p90": 81.45893467364742, + "p95": 81.11454798512572, + "p99": 80.51779421182322 + }, + "samples_us": [ + 3288.187, + 3278.893, + 3284.661, + 3295.347, + 3265.522, + 3290.85, + 3275.517, + 3291.382, + 3278.001, + 3272.082, + 3271.521, + 3309.338, + 3285.643, + 3277.02, + 3284.941, + 3333.865, + 3271.782, + 3266.834, + 3285.232, + 3276.379 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 536870912, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 6802.121, + "p90": 6918.535, + "p95": 6935.391, + "p99": 7006.437 + }, + "payload_gbps_at_latency_percentile": { + "p50": 78.92698645025574, + "p90": 77.59892983124317, + "p95": 77.4103308667096, + "p99": 76.6253820593834 + }, + "samples_us": [ + 6686.087, + 6649.722, + 6681.069, + 6763.904, + 6815.922, + 6833.748, + 6757.504, + 6793.908, + 6793.438, + 6918.535, + 6904.494, + 6825.977, + 7006.437, + 6935.391, + 6788.611, + 6802.121, + 6824.955, + 6802.04, + 6875.26, + 6875.181 + ] + } + } + ] +} diff --git a/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16.json b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16.json new file mode 100644 index 0000000000..b7e3d96f17 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T10:25:46.263716+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256","mode":"low-latency","nodes":2,"phase":"decode","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-uccl-ep-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256],"ladder_dropped":[512],"ladder_cap":256},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":121.26199901103973,"p90":124.78700280189514,"p95":126.26899778842926,"p99":130.03399968147278},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":204.22600209712982,"p90":206.95100724697113,"p95":208.2729935646057,"p99":211.557999253273},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":325.48800110816956,"p90":331.7380100488663,"p95":334.541991353035,"p99":341.5919989347458},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":295.6010103225708,"p90":298.48799109458923,"p95":299.24800992012024,"p99":302.09100246429443},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":304.6169877052307,"p90":308.102011680603,"p95":310.06500124931335,"p99":315.8339858055115},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":102.67399996519089,"p90":106.11899942159653,"p95":107.68099874258041,"p99":109.04300212860107},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":181.43099546432495,"p90":185.5980008840561,"p95":186.6389960050583,"p99":187.92100250720978},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.20800256729126,"p95":6.20800256729126,"p99":6.20800256729126},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.735018014907837,"p90":14.301985502243042,"p95":15.946000814437866,"p99":20.713001489639282},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.6019940376281738,"p90":-1.082003116607666,"p95":-1.082003116607666,"p99":-1.082003116607666},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":101.8730029463768,"p90":104.35599833726883,"p95":105.39799928665161,"p99":108.68299752473831},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":178.42599749565125,"p90":181.15100264549255,"p95":182.0320039987564,"p99":185.39699912071228},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":290.473997592926,"p90":293.2780086994171,"p95":294.1589951515198,"p99":296.083003282547},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.94200325012207,"p90":17.987996339797974,"p95":19.75199580192566,"p99":24.95899796485901},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.00048828125,"missed_samples":{"combine":1,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003888666396960616,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1376256,"scale_bytes":0,"total_logical_bytes":1376256},"dispatch":{"activation_data_bytes":1376256,"scale_bytes":0,"total_logical_bytes":1376256},"roundtrip":{"activation_data_bytes":2752512,"scale_bytes":0,"total_logical_bytes":2752512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"per-assignment"},"receive":{"max":11,"mean":6.0,"min":2,"total":96},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":52524.976103705514,"p90":51930.85209903322,"p95":51602.08322620363,"p99":50659.52595061349},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":123.22399765253067,"p90":127.14999914169312,"p95":128.55300307273865,"p99":131.91699981689453},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":206.42900466918945,"p90":209.47399735450745,"p95":211.0770046710968,"p99":214.44199979305267},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":329.6530023217201,"p90":336.62399649620056,"p95":339.63000774383545,"p99":346.3589996099472},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":299.24800992012024,"p90":301.2920022010803,"p95":301.81100964546204,"p99":303.21499705314636},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":307.1810007095337,"p90":310.62600016593933,"p95":311.86699867248535,"p99":316.39400124549866},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":102.99400240182877,"p90":106.32000118494034,"p95":107.36099630594254,"p99":110.28499901294708},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":183.43399465084076,"p90":186.95899844169617,"p95":187.96099722385406,"p99":189.80400264263153},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089985370635986,"p90":6.248980760574341,"p95":6.248980760574341,"p99":6.248980760574341},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.052008390426636,"p90":13.06101679801941,"p95":14.301985502243042,"p99":18.42600107192993},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.0010004043579102,"p90":1.0010004043579102,"p95":1.0010004043579102,"p99":1.0010004043579102},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":101.8730029463768,"p90":104.35699671506882,"p95":105.07699847221375,"p99":107.84099996089935},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":179.58900332450867,"p90":182.95300006866455,"p95":183.99499356746674,"p99":186.51899695396423},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":293.43798756599426,"p90":296.1220145225525,"p95":296.96398973464966,"p99":299.20700192451477},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.661980628967285,"p90":17.345011234283447,"p95":18.629014492034912,"p99":22.114992141723633},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003888666396960616,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2752512,"scale_bytes":0,"total_logical_bytes":2752512},"dispatch":{"activation_data_bytes":2752512,"scale_bytes":0,"total_logical_bytes":2752512},"roundtrip":{"activation_data_bytes":5505024,"scale_bytes":0,"total_logical_bytes":5505024},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"per-assignment"},"receive":{"max":17,"mean":12.0,"min":8,"total":192},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":104173.10942436436,"p90":103017.7769501114,"p95":102607.84288242558,"p99":101139.71780131929},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":124.6659979224205,"p90":128.1919926404953,"p95":129.59299981594086,"p99":132.7590048313141},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":211.51700615882874,"p90":214.04099464416504,"p95":215.1630073785782,"p99":218.6869978904724},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":336.18300408124924,"p90":342.23298728466034,"p95":344.75600719451904,"p99":351.4460027217865},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":305.6179881095886,"p90":308.14099311828613,"p95":308.82298946380615,"p99":310.9470009803772},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":315.15198945999146,"p90":318.63799691200256,"p95":319.8400139808655,"p99":325.2060115337372},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":105.6779995560646,"p90":109.00299996137619,"p95":110.16499996185303,"p99":112.40799725055695},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":187.96099722385406,"p90":190.68600237369537,"p95":191.8479949235916,"p99":194.1709965467453},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.208986043930054,"p95":6.208986043930054,"p99":6.208986043930054},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.096997022628784,"p90":14.660000801086426,"p95":16.385018825531006,"p99":19.82802152633667},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.961989164352417,"p90":1.0409951210021973,"p95":1.0409951210021973,"p99":1.0409951210021973},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":104.35599833726883,"p90":106.51999711990356,"p95":107.24099725484848,"p99":110.20500212907791},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":185.8379989862442,"p90":187.92100250720978,"p95":188.56200575828552,"p99":191.00600481033325},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":301.4500141143799,"p90":303.9340078830719,"p95":304.61499094963074,"p99":306.4180016517639},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.701021671295166,"p90":17.627984285354614,"p95":19.06999945640564,"p99":23.91800284385681},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038896200712770224,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":5505024,"scale_bytes":0,"total_logical_bytes":5505024},"dispatch":{"activation_data_bytes":5505024,"scale_bytes":0,"total_logical_bytes":5505024},"roundtrip":{"activation_data_bytes":11010048,"scale_bytes":0,"total_logical_bytes":11010048},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"per-assignment"},"receive":{"max":31,"mean":24.0,"min":17,"total":384},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":203076.61744310457,"p90":200854.8905662206,"p95":200100.04127822738,"p99":196798.3300744137},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":131.4769983291626,"p90":135.40199398994446,"p95":136.72399520874023,"p99":139.8089975118637},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":220.7310050725937,"p90":223.53500127792358,"p95":224.37700629234314,"p99":226.0179966688156},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":352.2080034017563,"p90":358.93699526786804,"p95":361.1010015010834,"p99":365.8269941806793},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":316.2730038166046,"p90":318.91700625419617,"p95":319.7999894618988,"p99":322.2010135650635},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":327.57100462913513,"p90":331.05701208114624,"p95":332.37800002098083,"p99":335.38299798965454},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":109.32300239801407,"p90":112.2480034828186,"p95":113.12899738550186,"p99":114.41200226545334},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":194.9310004711151,"p90":197.61499762535095,"p95":198.05599749088287,"p99":199.4989961385727},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.168991327285767,"p95":6.168991327285767,"p99":6.168991327285767},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":12.017011642456055,"p90":16.82499051094055,"p95":17.94499158859253,"p99":22.431999444961548},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.6019940376281738,"p90":1.4010071754455566,"p95":1.4010071754455566,"p99":1.4010071754455566},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":107.92200267314911,"p90":110.4860007762909,"p95":111.20600253343582,"p99":113.89099806547165},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":191.96699559688568,"p90":194.77100670337677,"p95":195.4919993877411,"p99":197.69500195980072},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":312.3469948768616,"p90":315.15100598335266,"p95":316.1120116710663,"p99":317.95498728752136},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":15.22400975227356,"p90":19.471019506454468,"p95":20.87301015853882,"p99":24.27700161933899},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003889858489856124,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":11010048,"scale_bytes":0,"total_logical_bytes":11010048},"dispatch":{"activation_data_bytes":11010048,"scale_bytes":0,"total_logical_bytes":11010048},"roundtrip":{"activation_data_bytes":22020096,"scale_bytes":0,"total_logical_bytes":22020096},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"per-assignment"},"receive":{"max":58,"mean":48.0,"min":40,"total":768},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":390754.97584078694,"p90":386640.35295716854,"p95":385103.70720059756,"p99":381653.2166724456},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":135.2819949388504,"p90":139.3679976463318,"p95":140.8510059118271,"p99":144.0960019826889},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":227.8210073709488,"p90":230.22499680519104,"p95":231.22599720954895,"p99":233.51000249385834},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":363.1030023097992,"p90":369.5929944515228,"p95":372.07700312137604,"p99":377.60600447654724},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":327.6109993457794,"p90":330.29401302337646,"p95":331.0549855232239,"p99":333.3379924297333},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":343.3949947357178,"p90":348.1619954109192,"p95":349.9230146408081,"p99":354.6920120716095},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":114.17099833488464,"p90":117.8169995546341,"p95":119.01800334453583,"p99":121.26100063323975},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":201.0210007429123,"p90":203.74499261379242,"p95":204.46500182151794,"p99":206.54800534248352},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.247997283935547,"p95":6.247997283935547,"p99":6.247997283935547},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":15.021026134490967,"p90":19.78999376296997,"p95":21.03400230407715,"p99":24.396002292633057},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.1230111122131348,"p90":1.8030107021331787,"p95":1.8030107021331787,"p99":1.8030107021331787},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":113.24899643659592,"p90":116.49499833583832,"p95":117.65599995851517,"p99":120.98199874162674},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":202.22200453281403,"p90":205.1060050725937,"p95":205.8670073747635,"p99":207.43000507354736},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":325.32599568367004,"p90":328.2110095024109,"p95":329.53399419784546,"p99":333.297997713089},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.987996339797974,"p90":22.91598916053772,"p95":24.639010429382324,"p99":28.203994035720825},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038906957488507032,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":22020096,"scale_bytes":0,"total_logical_bytes":22020096},"dispatch":{"activation_data_bytes":22020096,"scale_bytes":0,"total_logical_bytes":22020096},"roundtrip":{"activation_data_bytes":44040192,"scale_bytes":0,"total_logical_bytes":44040192},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"per-assignment"},"receive":{"max":105,"mean":96.0,"min":80,"total":1536},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":745497.1794129431,"p90":735289.9034768436,"p95":731589.4905134521,"p99":721752.9329313332},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":166.20799899101257,"p90":170.414000749588,"p95":171.7360019683838,"p99":174.86099898815155},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":254.22099232673645,"p90":256.8250000476837,"p95":257.8270137310028,"p99":260.78999042510986},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":420.428991317749,"p90":427.23900079727173,"p95":429.5630156993866,"p99":435.6509894132614},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":384.0560019016266,"p90":387.46100664138794,"p95":388.50200176239014,"p99":391.185998916626},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":402.2429883480072,"p90":407.44999051094055,"p95":408.97300839424133,"p99":413.98099064826965},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":139.97000455856323,"p90":143.65500211715698,"p95":144.93800699710846,"p99":146.8999981880188},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":226.05900466442108,"p90":229.50400412082672,"p95":230.42400181293488,"p99":231.905996799469},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.170004606246948,"p95":6.170004606246948,"p99":6.170004606246948},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.4670090675354,"p90":24.91798996925354,"p95":27.401000261306763,"p99":32.65097737312317},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.6420185565948486,"p90":1.762986183166504,"p95":1.762986183166504,"p99":1.762986183166504},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":139.44900035858154,"p90":143.3749943971634,"p95":144.49700713157654,"p99":149.1439938545227},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":229.54300045967102,"p90":231.50600492954254,"p95":232.02699422836304,"p99":233.26900601387024},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":379.40698862075806,"p90":382.73200392723083,"p95":383.69500637054443,"p99":386.6580128669739},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":22.995978593826294,"p90":27.963995933532715,"p95":29.605001211166382,"p99":33.77300500869751},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003891168860718608,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":44040192,"scale_bytes":0,"total_logical_bytes":44040192},"dispatch":{"activation_data_bytes":44040192,"scale_bytes":0,"total_logical_bytes":44040192},"roundtrip":{"activation_data_bytes":88080384,"scale_bytes":0,"total_logical_bytes":88080384},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"per-assignment"},"receive":{"max":211,"mean":192.0,"min":175,"total":3072},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":1272862.46082936,"p90":1256595.9306023158,"p95":1251916.359982473,"p99":1236771.7638392972},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":227.86200046539307,"p90":240.11999368667603,"p95":243.9659982919693,"p99":250.69600343704224},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":301.01099610328674,"p90":305.6989908218384,"p95":307.30101466178894,"p99":310.5059862136841},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":528.8729965686798,"p90":545.8189845085144,"p95":551.2670129537582,"p99":561.2019896507263},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":480.7179868221283,"p90":487.89098858833313,"p95":490.17301201820374,"p99":492.93598532676697},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":501.15102529525757,"p90":509.76401567459106,"p95":512.4480128288269,"p99":518.41801404953},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":192.52799451351166,"p90":197.29599356651306,"p95":198.65700602531433,"p99":202.42300629615784},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":257.26601481437683,"p90":268.202006816864,"p95":271.08699083328247,"p99":276.45498514175415},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.170004606246948,"p90":6.209015846252441,"p95":6.209015846252441,"p99":6.209015846252441},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":22.395998239517212,"p90":29.323995113372803,"p95":31.76701068878174,"p99":36.81197762489319},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.562999725341797,"p90":2.203017473220825,"p95":2.203017473220825,"p99":2.203017473220825},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":191.00600481033325,"p90":195.7329958677292,"p95":197.05499708652496,"p99":199.29799437522888},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":270.5250084400177,"p90":277.97698974609375,"p95":279.6590030193329,"p99":283.02299976348877},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":472.26598858833313,"p90":479.19800877571106,"p95":481.31901025772095,"p99":485.76900362968445},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":29.04599905014038,"p90":35.293012857437134,"p95":37.61899471282959,"p99":41.46197438240051},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003891168860718608,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":88080384,"scale_bytes":0,"total_logical_bytes":88080384},"dispatch":{"activation_data_bytes":88080384,"scale_bytes":0,"total_logical_bytes":88080384},"roundtrip":{"activation_data_bytes":176160768,"scale_bytes":0,"total_logical_bytes":176160768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"per-assignment"},"receive":{"max":421,"mean":384.0,"min":335,"total":6144},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":2043296.2287101005,"p90":2008772.6252016826,"p95":1998251.479886306,"p99":1975240.0037205622},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":337.8260135650635,"p90":348.4019935131073,"p95":352.32800245285034,"p99":358.817994594574},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":398.2369899749756,"p90":404.5659899711609,"p95":406.8099856376648,"p99":411.857008934021},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":736.0630035400391,"p90":752.9679834842682,"p95":759.1379880905151,"p99":770.675003528595},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":667.2800183296204,"p90":675.6929755210876,"p95":679.0949702262878,"p99":685.7479810714722},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":697.1660256385803,"p90":706.496000289917,"p95":709.9850177764893,"p99":719.4790244102478},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":278.0570089817047,"p90":287.83199191093445,"p95":290.15499353408813,"p99":294.68101263046265},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":337.06599473953247,"p90":348.2019901275635,"p95":351.4469861984253,"p99":357.7350080013275},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129026412963867,"p90":6.3700079917907715,"p95":6.3700079917907715,"p99":6.3700079917907715},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":28.84000539779663,"p90":39.81602191925049,"p95":44.82400417327881,"p99":51.232993602752686},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.0049681663513184,"p90":4.486978054046631,"p95":4.486978054046631,"p99":4.486978054046631},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":276.8149971961975,"p90":286.547988653183,"p95":289.07299041748047,"p99":294.281005859375},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":360.9809875488281,"p90":366.50800704956055,"p95":368.353009223938,"p99":372.39599227905273},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":659.8259806632996,"p90":667.4379706382751,"p95":670.0040102005005,"p99":678.134024143219},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":37.617027759552,"p90":46.112000942230225,"p95":48.71600866317749,"p99":54.485023021698},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003891168860718608,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":176160768,"scale_bytes":0,"total_logical_bytes":176160768},"dispatch":{"activation_data_bytes":176160768,"scale_bytes":0,"total_logical_bytes":176160768},"roundtrip":{"activation_data_bytes":352321536,"scale_bytes":0,"total_logical_bytes":352321536},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"per-assignment"},"receive":{"max":802,"mean":768.0,"min":696,"total":12288},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":2937607.2910668617,"p90":2898813.297116452,"p95":2884567.911607301,"p99":2846504.109940845},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":584.276020526886,"p90":601.4209985733032,"p95":606.3089966773987,"p99":620.3699707984924},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":666.038990020752,"p90":676.4910221099854,"p95":681.020975112915,"p99":689.4329786300659},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1250.315010547638,"p90":1277.9120206832886,"p95":1287.3299717903137,"p99":1309.8029494285583},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1212.3719453811646,"p90":1227.4349927902222,"p95":1232.522964477539,"p99":1241.3820028305054},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1229.6830415725708,"p90":1246.108055114746,"p95":1252.1170377731323,"p99":1264.456033706665},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":551.6240000724792,"p90":562.7239942550659,"p95":564.844012260437,"p99":569.6520209312439},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":632.5089931488037,"p90":645.363986492157,"p95":648.9329934120178,"p99":655.4229855537415},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.168961524963379,"p90":6.25002384185791,"p95":6.25002384185791,"p99":6.25002384185791},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":28.247952461242676,"p90":46.3029146194458,"p95":52.95407772064209,"p99":61.00499629974365},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.764986991882324,"p90":7.571935653686523,"p95":7.571935653686523,"p99":7.571935653686523},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":551.9449710845947,"p90":562.3630285263062,"p95":565.405011177063,"p99":569.7339773178101},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":642.7609920501709,"p90":651.6969799995422,"p95":654.4610261917114,"p99":659.6289873123169},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1201.9959688186646,"p90":1216.6179418563843,"p95":1221.1910486221313,"p99":1233.2090139389038},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":27.166008949279785,"p90":44.03102397918701,"p95":49.118995666503906,"p99":58.85303020477295},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003891168860718608,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":352321536,"scale_bytes":0,"total_logical_bytes":352321536},"dispatch":{"activation_data_bytes":352321536,"scale_bytes":0,"total_logical_bytes":352321536},"roundtrip":{"activation_data_bytes":704643072,"scale_bytes":0,"total_logical_bytes":704643072},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"per-assignment"},"receive":{"max":1613,"mean":1536.0,"min":1426,"total":24576},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":3330939.6499132505,"p90":3287034.3652684484,"p95":3271259.695726737,"p99":3239337.620931635},"tokens_per_rank":256}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"receive_capacity":256,"num_qps_per_rank":24,"return_recv_hook":true,"zero_copy":false,"fp8_scales":null,"deep_ep":"uccl deep_ep_wrapper","num_nvl_bytes":1073741824,"num_rdma_bytes":5650787456,"is_intranode":false,"cpu_proxy_threads":4},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-ll-hook-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9938,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2069,"p50":2099,"max":2101,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.4},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2077,"p50":2100,"max":2101,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2073,"p50":2099,"max":2101,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2103,"max":2104,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2069,"p50":2094,"max":2095,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2057,"p50":2104,"max":2106,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.4},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2081,"p50":2107,"max":2108,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2066,"p50":2096,"max":2098,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2080,"p50":2102,"max":2104,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2076,"p50":2102,"max":2104,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2081,"p50":2104,"max":2106,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2073,"p50":2099,"max":2100,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":53,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2076,"p50":2102,"max":2104,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2076,"p50":2100,"max":2101,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2101,"max":2102,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2049,"p50":2101,"max":2102,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":260.5}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-low-latency-decode-ep16-uniform-fp8.json b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-low-latency-decode-ep16-uniform-fp8.json new file mode 100644 index 0000000000..f8bf05fe14 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-low-latency-decode-ep16-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T10:34:24.015206+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256","mode":"low-latency","nodes":2,"phase":"decode","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-uccl-ep-deepseek-v4-pro-low-latency-decode-ep16-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256],"ladder_dropped":[512],"ladder_cap":256},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":122.78299778699875,"p90":127.26999819278717,"p95":128.79300117492676,"p99":132.7590048313141},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":196.493998169899,"p90":200.10000467300415,"p95":201.34200155735016,"p99":204.14599776268005},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":319.27699595689774,"p90":327.3700028657913,"p95":330.1350027322769,"p99":336.90500259399414},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":283.74600410461426,"p90":286.0279977321625,"p95":286.87000274658203,"p99":288.9139950275421},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":292.0379936695099,"p90":296.3249981403351,"p95":297.8450059890747,"p99":302.6939928531647},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":552.266001701355,"p90":556.1509728431702,"p95":557.2729706764221,"p99":559.4369769096375},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":102.51300036907196,"p90":106.19799792766571,"p95":107.12099820375443,"p99":109.16300117969513},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":171.05600237846375,"p90":175.26300251483917,"p95":176.14300549030304,"p99":178.06699872016907},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.168991327285767,"p95":6.168991327285767,"p99":6.168991327285767},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.8539719581604,"p90":13.781994581222534,"p95":14.784008264541626,"p99":18.545985221862793},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.961989164352417,"p90":0.8800029754638672,"p95":0.8800029754638672,"p99":0.8800029754638672},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":98.78899902105331,"p90":102.35399752855301,"p95":103.39599847793579,"p99":107.60200023651123},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":166.80899262428284,"p90":170.9350049495697,"p95":172.13700711727142,"p99":175.14200508594513},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":279.6579897403717,"p90":282.54199028015137,"p95":283.5040092468262,"p99":286.10700368881226},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":12.260019779205322,"p90":16.305983066558838,"p95":18.029004335403442,"p99":22.235006093978882},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":false,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0817006528377533,"oracle_passed":{"pre":false,"chained":false,"post":false},"oracle_failed_checks":{"pre":["combine_values","payload"],"chained":["combine_values","payload"],"post":["combine_values","payload"]},"passed":false},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":587776,"scale_bytes":18368,"total_logical_bytes":606144},"roundtrip":{"activation_data_bytes":1763328,"scale_bytes":18368,"total_logical_bytes":1781696},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1376256,"scale_bytes":0,"total_logical_bytes":1376256},"dispatch":{"activation_data_bytes":688128,"scale_bytes":21504,"total_logical_bytes":709632},"roundtrip":{"activation_data_bytes":2064384,"scale_bytes":21504,"total_logical_bytes":2085888},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"per-assignment"},"receive":{"max":11,"mean":6.0,"min":2,"total":96},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":54787.39186965752,"p90":53994.769595586535,"p95":53719.215290743865,"p99":52858.6637917242},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":122.70300090312958,"p90":126.82999670505524,"p95":128.59199941158295,"p99":131.67700171470642},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":198.77800345420837,"p90":202.0229995250702,"p95":203.5059928894043,"p99":207.11000263690948},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":321.48100435733795,"p90":328.8529962301254,"p95":332.09799230098724,"p99":338.7870043516159},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":287.87198662757874,"p90":289.9929881095886,"p95":290.83698987960815,"p99":292.8380072116852},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":296.2450087070465,"p90":299.8889982700348,"p95":301.3319969177246,"p99":306.098997592926},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":552.2249937057495,"p90":556.2319755554199,"p95":557.4730038642883,"p99":559.9169731140137},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":103.3950001001358,"p90":106.91999644041061,"p95":107.88100212812424,"p99":109.88499969244003},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":173.61900210380554,"p90":177.42499709129333,"p95":178.2670021057129,"p99":179.58900332450867},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129026412963867,"p90":6.208986043930054,"p95":6.208986043930054,"p99":6.208986043930054},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.531004190444946,"p90":13.538002967834473,"p95":14.860004186630249,"p99":19.83100175857544},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.3220012187957764,"p90":-1.0419785976409912,"p95":-1.0419785976409912,"p99":-1.0419785976409912},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":100.832000374794,"p90":103.43500226736069,"p95":104.19599711894989,"p99":107.60100185871124},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":171.57700657844543,"p90":174.50100183486938,"p95":175.38200318813324,"p99":178.2670021057129},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":283.58399868011475,"p90":286.188006401062,"p95":287.1110141277313,"p99":289.03400897979736},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":12.699991464614868,"p90":15.98501205444336,"p95":17.666995525360107,"p99":22.235006093978882},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":false,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0817006528377533,"oracle_passed":{"pre":false,"chained":false,"post":false},"oracle_failed_checks":{"pre":["combine_values","payload"],"chained":["combine_values","payload"],"post":["combine_values","payload"]},"passed":false},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":1189888,"scale_bytes":37184,"total_logical_bytes":1227072},"roundtrip":{"activation_data_bytes":3569664,"scale_bytes":37184,"total_logical_bytes":3606848},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2752512,"scale_bytes":0,"total_logical_bytes":2752512},"dispatch":{"activation_data_bytes":1376256,"scale_bytes":43008,"total_logical_bytes":1419264},"roundtrip":{"activation_data_bytes":4128768,"scale_bytes":43008,"total_logical_bytes":4171776},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"per-assignment"},"receive":{"max":17,"mean":12.0,"min":8,"total":192},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":108018.69756274766,"p90":106706.14855695916,"p95":106195.16124182873,"p99":104541.34202215214},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":124.90600347518921,"p90":128.59299778938293,"p95":130.03399968147278,"p99":133.56000185012817},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":200.58099925518036,"p90":203.70499789714813,"p95":205.14799654483795,"p99":208.43300223350525},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":325.48700273036957,"p90":332.29799568653107,"p95":335.18199622631073,"p99":341.9930040836334},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":293.3200001716614,"p90":295.522004365921,"p95":296.68399691581726,"p99":298.3259856700897},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":302.53300070762634,"p90":306.05900287628174,"p95":307.5009882450104,"p99":312.10801005363464},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":552.1460175514221,"p90":556.0309886932373,"p95":557.3530197143555,"p99":560.2380037307739},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":105.99800199270248,"p90":108.84299874305725,"p95":109.52500253915787,"p99":111.24599725008011},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":176.46299302577972,"p90":179.7100007534027,"p95":180.8300018310547,"p99":182.59400129318237},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.168991327285767,"p95":6.168991327285767,"p99":6.168991327285767},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.534984827041626,"p90":14.421999454498291,"p95":15.665978193283081,"p99":20.10899782180786},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.1210143566131592,"p90":1.0010004043579102,"p95":1.0010004043579102,"p99":1.0010004043579102},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":104.19599711894989,"p90":106.51999711990356,"p95":107.40099847316742,"p99":110.04400253295898},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":173.77999424934387,"p90":176.7439991235733,"p95":177.66499519348145,"p99":180.18899857997894},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":289.4330024719238,"p90":292.0779883861542,"p95":292.95799136161804,"p99":295.16100883483887},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.062000274658203,"p90":16.90700650215149,"p95":18.30899715423584,"p99":22.03500270843506},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":false,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.09439205378293991,"oracle_passed":{"pre":false,"chained":false,"post":false},"oracle_failed_checks":{"pre":["combine_values","payload"],"chained":["combine_values","payload"],"post":["combine_values","payload"]},"passed":false},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":2358272,"scale_bytes":73696,"total_logical_bytes":2431968},"roundtrip":{"activation_data_bytes":7074816,"scale_bytes":73696,"total_logical_bytes":7148512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":5505024,"scale_bytes":0,"total_logical_bytes":5505024},"dispatch":{"activation_data_bytes":2752512,"scale_bytes":86016,"total_logical_bytes":2838528},"roundtrip":{"activation_data_bytes":8257536,"scale_bytes":86016,"total_logical_bytes":8343552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"per-assignment"},"receive":{"max":31,"mean":24.0,"min":17,"total":384},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":211547.16956597677,"p90":209110.00623585877,"p95":208129.41241348512,"p99":205057.21717620073},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":129.87400591373444,"p90":133.88000428676605,"p95":135.36199927330017,"p99":138.96700739860535},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":202.62399315834045,"p90":206.2699943780899,"p95":207.75100588798523,"p99":211.0770046710968},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":332.4979990720749,"p90":340.14999866485596,"p95":343.1130051612854,"p99":350.04401206970215},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":301.813006401062,"p90":304.4160008430481,"p95":305.33701181411743,"p99":306.78001046180725},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":310.4259967803955,"p90":313.9910101890564,"p95":315.5539929866791,"p99":319.0790116786957},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":552.0650148391724,"p90":556.1519861221313,"p95":557.433009147644,"p99":559.6359968185425},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":108.9629977941513,"p90":111.92800104618073,"p95":112.7690002322197,"p99":114.89199846982956},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":180.35000562667847,"p90":183.55399370193481,"p95":184.47500467300415,"p99":186.44000589847565},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.130009889602661,"p90":6.169021129608154,"p95":6.169021129608154,"p99":6.169021129608154},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11.734992265701294,"p90":16.101986169815063,"p95":17.30501651763916,"p99":22.753983736038208},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.2409985065460205,"p90":1.4429986476898193,"p95":1.4429986476898193,"p99":1.4429986476898193},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":107.00000077486038,"p90":109.484001994133,"p95":110.2060005068779,"p99":113.13000321388245},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":174.78099465370178,"p90":177.82600224018097,"p95":178.70700359344482,"p99":181.0300052165985},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":295.44198513031006,"p90":298.3250021934509,"p95":299.20700192451477,"p99":301.45201086997986},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.8639976978302,"p90":18.748998641967773,"p95":20.23199200630188,"p99":23.597002029418945},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.00048828125,"missed_samples":{"combine":1,"dispatch":0,"roundtrip":0,"stage":3}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":false,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.09439205378293991,"oracle_passed":{"pre":false,"chained":false,"post":false},"oracle_failed_checks":{"pre":["combine_values","payload"],"chained":["combine_values","payload"],"post":["combine_values","payload"]},"passed":false},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":4652032,"scale_bytes":145376,"total_logical_bytes":4797408},"roundtrip":{"activation_data_bytes":13956096,"scale_bytes":145376,"total_logical_bytes":14101472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":11010048,"scale_bytes":0,"total_logical_bytes":11010048},"dispatch":{"activation_data_bytes":5505024,"scale_bytes":172032,"total_logical_bytes":5677056},"roundtrip":{"activation_data_bytes":16515072,"scale_bytes":172032,"total_logical_bytes":16687104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"per-assignment"},"receive":{"max":58,"mean":48.0,"min":40,"total":768},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":412336.59979370533,"p90":407654.98325232376,"p95":405635.81144543923,"p99":401154.55832266615},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":135.36199927330017,"p90":139.64800536632538,"p95":141.6109949350357,"p99":146.45899832248688},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":204.82699573040009,"p90":207.67199993133545,"p95":209.074005484581,"p99":212.67899870872498},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":340.18899500370026,"p90":347.3200052976608,"p95":350.6850004196167,"p99":359.13799703121185},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":309.904009103775,"p90":313.3080005645752,"p95":314.15098905563354,"p99":318.8360035419464},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":321.042001247406,"p90":325.28799772262573,"p95":326.8499970436096,"p99":330.81600069999695},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":551.7849922180176,"p90":555.6700229644775,"p95":557.1529865264893,"p99":559.436023235321},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":116.33399873971939,"p90":119.13800239562988,"p95":119.97900158166885,"p99":123.54499846696854},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":182.07299709320068,"p90":184.8369985818863,"p95":185.67800521850586,"p99":186.75899505615234},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.168991327285767,"p90":6.209015846252441,"p95":6.209015846252441,"p99":6.209015846252441},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.100998640060425,"p90":19.14900541305542,"p95":20.471006631851196,"p99":25.59801936149597},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.323014497756958,"p90":1.4420151710510254,"p95":1.4420151710510254,"p99":1.4420151710510254},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":113.89099806547165,"p90":116.73499643802643,"p95":117.65699833631516,"p99":121.62300199270248},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":178.7469983100891,"p90":181.71200156211853,"p95":182.59300291538239,"p99":184.35600399971008},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":303.3739924430847,"p90":306.4979910850525,"p95":307.5000047683716,"p99":310.38400530815125},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.706990242004395,"p90":22.27500081062317,"p95":23.716986179351807,"p99":26.562005281448364},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":false,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.09439205378293991,"oracle_passed":{"pre":false,"chained":false,"post":false},"oracle_failed_checks":{"pre":["combine_values","payload"],"chained":["combine_values","payload"],"post":["combine_values","payload"]},"passed":false},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":9418752,"scale_bytes":294336,"total_logical_bytes":9713088},"roundtrip":{"activation_data_bytes":28256256,"scale_bytes":294336,"total_logical_bytes":28550592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":22020096,"scale_bytes":0,"total_logical_bytes":22020096},"dispatch":{"activation_data_bytes":11010048,"scale_bytes":344064,"total_logical_bytes":11354112},"roundtrip":{"activation_data_bytes":33030144,"scale_bytes":344064,"total_logical_bytes":33374208},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"per-assignment"},"receive":{"max":105,"mean":96.0,"min":80,"total":1536},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":797403.451901353,"p90":786994.9146365129,"p95":783233.9064266336,"p99":773844.0687823792},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":165.56799411773682,"p90":169.85400021076202,"p95":171.33699357509613,"p99":173.77999424934387},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":216.84600412845612,"p90":219.32899951934814,"p95":220.36999464035034,"p99":223.13499450683594},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":382.41399824619293,"p90":389.18299973011017,"p95":391.7069882154465,"p99":396.9149887561798},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":350.7649898529053,"p90":354.16901111602783,"p95":355.2109897136688,"p99":359.4599962234497},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":366.99000000953674,"p90":372.1590042114258,"p95":373.9199936389923,"p99":378.1670033931732},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":551.5050292015076,"p90":555.3500056266785,"p95":556.5119981765747,"p99":559.5960021018982},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":141.13199710845947,"p90":145.25699615478516,"p95":146.74000442028046,"p99":149.74400401115417},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":192.00700521469116,"p90":194.77100670337677,"p95":195.6920027732849,"p99":196.81499898433685},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.209015846252441,"p95":6.209015846252441,"p99":6.209015846252441},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.54799485206604,"p90":23.515015840530396,"p95":25.39801597595215,"p99":29.40499782562256},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.3610124588012695,"p90":2.2430121898651123,"p95":2.2430121898651123,"p99":2.2430121898651123},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":137.96700537204742,"p90":141.5719985961914,"p95":142.69499480724335,"p99":145.41900157928467},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":192.9280012845993,"p90":194.9310004711151,"p95":195.4919993877411,"p99":196.9749927520752},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":343.63600611686707,"p90":347.0410108566284,"p95":347.9999899864197,"p99":351.04501247406006},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":23.396998643875122,"p90":28.205007314682007,"p95":29.80700135231018,"p99":33.609986305236816},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":false,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.09881741553544998,"oracle_passed":{"pre":false,"chained":false,"post":false},"oracle_failed_checks":{"pre":["combine_values","payload"],"chained":["combine_values","payload"],"post":["combine_values","payload"]},"passed":false},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":18808832,"scale_bytes":587776,"total_logical_bytes":19396608},"roundtrip":{"activation_data_bytes":56426496,"scale_bytes":587776,"total_logical_bytes":57014272},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":44040192,"scale_bytes":0,"total_logical_bytes":44040192},"dispatch":{"activation_data_bytes":22020096,"scale_bytes":688128,"total_logical_bytes":22708224},"roundtrip":{"activation_data_bytes":66060288,"scale_bytes":688128,"total_logical_bytes":66748416},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"per-assignment"},"receive":{"max":211,"mean":192.0,"min":175,"total":3072},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":1395133.382344737,"p90":1375756.0456850582,"p95":1369276.8739569448,"p99":1353899.1911139404},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":216.72500669956207,"p90":222.4940061569214,"p95":224.4970053434372,"p99":228.50100696086884},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":242.04300343990326,"p90":244.40699815750122,"p95":245.32799422740936,"p99":248.05299937725067},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":458.76801013946533,"p90":466.9010043144226,"p95":469.82499957084656,"p99":476.5540063381195},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":419.1479980945587,"p90":423.07499051094055,"p95":424.1969883441925,"p99":426.31998658180237},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":439.7389888763428,"p90":445.9089934825897,"p95":447.952002286911,"p99":451.7970085144043},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":551.4640212059021,"p90":555.4710030555725,"p95":556.5119981765747,"p99":559.0360164642334},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":186.71999871730804,"p90":192.44900345802307,"p95":193.9699947834015,"p99":196.13400101661682},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":208.3120048046112,"p90":214.4010066986084,"p95":215.6819999217987,"p99":218.52600574493408},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129026412963867,"p90":6.170004606246948,"p95":6.170004606246948,"p99":6.170004606246948},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":21.87401056289673,"p90":29.44597601890564,"p95":31.890004873275757,"p99":37.73897886276245},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.4839937686920166,"p90":1.801997423171997,"p95":1.801997423171997,"p99":1.801997423171997},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":185.43800711631775,"p90":190.76600670814514,"p95":192.32800602912903,"p99":194.9719935655594},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":218.7269926071167,"p90":220.7300066947937,"p95":221.2110012769699,"p99":222.2920060157776},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":411.8160009384155,"p90":416.2620007991791,"p95":417.26601123809814,"p99":419.4670021533966},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":28.24300527572632,"p90":35.1330041885376,"p95":37.496984004974365,"p99":43.30497980117798},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":false,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.10453245788812637,"oracle_passed":{"pre":false,"chained":false,"post":false},"oracle_failed_checks":{"pre":["combine_values","payload"],"chained":["combine_values","payload"],"post":["combine_values","payload"]},"passed":false},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":37675008,"scale_bytes":1177344,"total_logical_bytes":38852352},"roundtrip":{"activation_data_bytes":113025024,"scale_bytes":1177344,"total_logical_bytes":114202368},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":88080384,"scale_bytes":0,"total_logical_bytes":88080384},"dispatch":{"activation_data_bytes":44040192,"scale_bytes":1376256,"total_logical_bytes":45416448},"roundtrip":{"activation_data_bytes":132120576,"scale_bytes":1376256,"total_logical_bytes":133496832},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"per-assignment"},"receive":{"max":421,"mean":384.0,"min":335,"total":6144},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":2328654.101417318,"p90":2296432.713775219,"p95":2285959.198244934,"p99":2266504.604284808},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":325.9289860725403,"p90":336.34400367736816,"p95":339.7490084171295,"p99":349.76398944854736},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":288.9539897441864,"p90":293.68099570274353,"p95":295.1219975948334,"p99":298.0870008468628},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":614.8829758167267,"p90":630.0249993801117,"p95":634.8710060119629,"p99":647.8509902954102},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":569.6139931678772,"p90":580.3470015525818,"p95":582.7909708023071,"p99":592.4479961395264},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":595.6540107727051,"p90":605.3479909896851,"p95":609.2740297317505,"p99":616.3610219955444},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":551.144003868103,"p90":555.1499724388123,"p95":556.4320087432861,"p99":559.1560006141663},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":285.7489883899689,"p90":296.36499285697937,"p95":299.5699942111969,"p99":312.1879994869232},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":249.37400221824646,"p90":257.94699788093567,"p95":260.468989610672,"p99":263.83399963378906},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.169021129608154,"p90":6.2490105628967285,"p95":6.2490105628967285,"p99":6.2490105628967285},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":27.51898765563965,"p90":39.261043071746826,"p95":43.34801435470581,"p99":49.596965312957764},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.6840567588806152,"p90":7.091045379638672,"p95":7.091045379638672,"p99":7.091045379638672},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":283.3850085735321,"p90":294.16000843048096,"p95":297.92699217796326,"p99":309.90299582481384},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":264.83601331710815,"p90":270.004004240036,"p95":271.7680037021637,"p99":275.25100111961365},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":557.9580068588257,"p90":567.0880079269409,"p95":570.2919960021973,"p99":583.1149816513062},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":37.37598657608032,"p90":45.828044414520264,"p95":48.47598075866699,"p99":54.52299118041992},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":false,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.10609076172113419,"oracle_passed":{"pre":false,"chained":false,"post":false},"oracle_failed_checks":{"pre":["combine_values","payload"],"chained":["combine_values","payload"],"post":["combine_values","payload"]},"passed":false},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":75787264,"scale_bytes":2368352,"total_logical_bytes":78155616},"roundtrip":{"activation_data_bytes":227361792,"scale_bytes":2368352,"total_logical_bytes":229730144},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":176160768,"scale_bytes":0,"total_logical_bytes":176160768},"dispatch":{"activation_data_bytes":88080384,"scale_bytes":2752512,"total_logical_bytes":90832896},"roundtrip":{"activation_data_bytes":264241152,"scale_bytes":2752512,"total_logical_bytes":266993664},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"per-assignment"},"receive":{"max":802,"mean":768.0,"min":696,"total":12288},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":3438237.57241768,"p90":3383177.9909795676,"p95":3361377.475586294,"p99":3322727.9579901868},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":587.1210098266602,"p90":599.8200178146362,"p95":603.344976902008,"p99":610.9970211982727},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":413.0989909172058,"p90":420.59099674224854,"p95":424.235999584198,"p99":431.8479895591736},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1000.220000743866,"p90":1020.4110145568848,"p95":1027.580976486206,"p99":1042.8450107574463},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":954.5519948005676,"p90":966.8099880218506,"p95":971.6920256614685,"p99":980.7860255241394},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":970.6509709358215,"p90":983.115017414093,"p95":987.7210259437561,"p99":997.0560073852539},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":550.1430034637451,"p90":553.9479851722717,"p95":555.5099844932556,"p99":557.7139854431152},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":551.8649816513062,"p90":560.9179735183716,"p95":563.0850195884705,"p99":568.1290030479431},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":373.59800934791565,"p90":386.2980008125305,"p95":390.9060060977936,"p99":397.31600880622864},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.088972091674805,"p90":6.2490105628967285,"p95":6.2490105628967285,"p99":6.2490105628967285},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":30.03901243209839,"p90":46.39500379562378,"p95":51.551997661590576,"p99":62.09897994995117},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.3249855041503906,"p90":4.2479634284973145,"p95":4.2479634284973145,"p99":4.2479634284973145},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":549.3040084838867,"p90":558.9179992675781,"p95":561.2390041351318,"p99":565.6489729881287},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":382.89400935173035,"p90":390.1030123233795,"p95":392.5890028476715,"p99":401.3200104236603},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":942.5740242004395,"p90":954.9480080604553,"p95":958.5530161857605,"p99":968.8529968261719},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":28.20795774459839,"p90":43.10500621795654,"p95":47.797977924346924,"p99":61.4970326423645},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":false,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.12359141558408737,"oracle_passed":{"pre":false,"chained":false,"post":false},"oracle_failed_checks":{"pre":["combine_values","payload"],"chained":["combine_values","payload"],"post":["combine_values","payload"]},"passed":false},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":151825408,"scale_bytes":4744544,"total_logical_bytes":156569952},"roundtrip":{"activation_data_bytes":455476224,"scale_bytes":4744544,"total_logical_bytes":460220768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":352321536,"scale_bytes":0,"total_logical_bytes":352321536},"dispatch":{"activation_data_bytes":176160768,"scale_bytes":5505024,"total_logical_bytes":181665792},"roundtrip":{"activation_data_bytes":528482304,"scale_bytes":5505024,"total_logical_bytes":533987328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"per-assignment"},"receive":{"max":1613,"mean":1536.0,"min":1426,"total":24576},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":4219848.454950779,"p90":4166348.725679921,"p95":4146919.9221372446,"p99":4108094.1989824856},"tokens_per_rank":256}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"receive_capacity":256,"num_qps_per_rank":24,"return_recv_hook":true,"zero_copy":false,"fp8_scales":"fp32","deep_ep":"uccl deep_ep_wrapper","num_nvl_bytes":1073741824,"num_rdma_bytes":5650787456,"is_intranode":false,"cpu_proxy_threads":4},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-ll-hook-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9755,"throttle_fraction_max":{"power":0.2238,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":true,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2007,"p50":2041,"max":2095,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.2072,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.8},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2042,"p50":2068,"max":2098,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.1875,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.8},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2044,"p50":2069,"max":2097,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.1876,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.8},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2064,"p50":2083,"max":2100,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.157,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.9},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2028,"p50":2056,"max":2091,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.1904,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.8},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2023,"p50":2054,"max":2101,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.1933,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.8},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2016,"p50":2049,"max":2103,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.2095,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.8},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2012,"p50":2044,"max":2094,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.2103,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.9},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":1994,"p50":2032,"max":2100,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.2238,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.9},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2045,"p50":2071,"max":2100,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.1728,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.8},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2046,"p50":2073,"max":2102,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.1878,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.8},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2042,"p50":2068,"max":2095,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.1767,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.9},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2032,"p50":2062,"max":2100,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.2016,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.8},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2047,"p50":2073,"max":2099,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.1887,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.9},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2030,"p50":2060,"max":2097,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.1946,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.8},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2050,"p50":2075,"max":2099,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.1748,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":394.8}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":["semantic correctness or routing identity failed"],"status":"invalid"}} diff --git a/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-bf16.json b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-bf16.json new file mode 100644 index 0000000000..7fcf792fb0 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T10:43:56.367832+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":2,"phase":"decode","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":193.45000386238098,"p90":198.37699830532074,"p95":200.30100643634796,"p99":204.18600738048553},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":461.08898520469666,"p90":472.18799591064453,"p95":475.9939908981323,"p99":485.48799753189087},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":654.5389890670776,"p90":670.5649942159653,"p95":676.2949973344803,"p99":689.6740049123764},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":863.008975982666,"p90":883.2849860191345,"p95":895.2980041503906,"p99":932.5129985809326},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":737.1450066566467,"p90":750.0439882278442,"p95":755.1289796829224,"p99":766.3099765777588},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":230.46499490737915,"p90":235.19200086593628,"p95":237.31400072574615,"p99":243.32499504089355},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":522.0999717712402,"p90":535.0000262260437,"p95":544.3339943885803,"p99":563.2020235061646},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.369054317474365,"p90":6.609022617340088,"p95":6.609022617340088,"p99":6.609022617340088},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.013035774230957,"p90":12.29405403137207,"p95":13.864994049072266,"p99":19.143998622894287},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.681983470916748,"p90":30.485033988952637,"p95":30.485033988952637,"p99":30.485033988952637},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":158.5180014371872,"p90":162.76399791240692,"p95":164.64699804782867,"p99":168.77299547195435},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":362.8219962120056,"p90":370.63300609588623,"p95":373.4380006790161,"p99":378.0049979686737},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":639.7600173950195,"p90":652.658998966217,"p95":657.1459770202637,"p99":668.0009961128235},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":95.94500064849854,"p90":109.32397842407227,"p95":112.40702867507935,"p99":118.09700727462769},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":9,"mean":5.125,"min":2,"total":82},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":21705.363063596804,"p90":21332.082186011212,"p95":21188.4332749597,"p99":20879.279259098166},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":196.4540034532547,"p90":201.34200155735016,"p95":203.5049945116043,"p99":207.75200426578522},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":461.01099252700806,"p90":472.2679853439331,"p95":476.0720133781433,"p99":484.8040044307709},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":657.4649959802628,"p90":673.6099869012833,"p95":679.5770078897476,"p99":692.5560086965561},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":863.9310002326965,"p90":892.0980095863342,"p95":898.747980594635,"p99":910.3249907493591},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":736.6610169410706,"p90":749.4440078735352,"p95":754.4879913330078,"p99":766.5449976921082},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":229.6230047941208,"p90":233.94900560379028,"p95":235.59199273586273,"p99":239.60000276565552},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":523.3020186424255,"p90":539.5259857177734,"p95":546.1360216140747,"p99":566.6069984436035},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.289005279541016,"p90":6.569027900695801,"p95":6.569027900695801,"p99":6.569027900695801},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.814002990722656,"p90":12.775003910064697,"p95":14.822006225585938,"p99":18.428027629852295},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.4050345420837402,"p90":2.562999725341797,"p95":2.562999725341797,"p99":2.562999725341797},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":171.4159995317459,"p90":174.30000007152557,"p95":176.5040010213852,"p99":180.18899857997894},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":367.9499924182892,"p90":375.2799928188324,"p95":377.40400433540344,"p99":383.21301341056824},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":652.6550054550171,"p90":664.0719771385193,"p95":668.6000227928162,"p99":680.978000164032},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":82.72695541381836,"p90":96.2669849395752,"p95":99.5519757270813,"p99":106.56201839447021},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":15,"mean":10.375,"min":7,"total":166},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":43439.24717623527,"p90":42698.31990624154,"p95":42412.86855137789,"p99":41745.755430333105},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":197.175994515419,"p90":202.06299424171448,"p95":204.22600209712982,"p99":209.03299748897552},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":461.29000186920166,"p90":472.98699617385864,"p95":476.5940010547638,"p99":483.9630126953125},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":658.4659963846207,"p90":675.0499904155731,"p95":680.8200031518936,"p99":692.996010184288},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":861.2470030784607,"p90":874.5869994163513,"p95":886.9699835777283,"p99":976.0590195655823},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":737.9459738731384,"p90":751.1669993400574,"p95":755.4489970207214,"p99":765.5479907989502},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":231.58599436283112,"p90":236.35399341583252,"p95":238.99699747562408,"p99":242.4830049276352},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":520.4579830169678,"p90":532.5160026550293,"p95":537.4860167503357,"p99":542.9720282554626},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.330013275146484,"p90":6.408989429473877,"p95":6.408989429473877,"p99":6.408989429473877},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.129967212677002,"p90":12.894988059997559,"p95":14.7019624710083,"p99":18.467962741851807},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.3659934997558594,"p90":2.1229982376098633,"p95":2.1229982376098633,"p99":2.1229982376098633},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":173.82000386714935,"p90":176.54399573802948,"p95":177.9859960079193,"p99":182.75299668312073},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":373.4790086746216,"p90":382.2520077228546,"p95":384.7750127315521,"p99":389.8639976978302},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":655.5399894714355,"p90":667.4370169639587,"p95":672.003984451294,"p99":680.4159879684448},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":81.52502775192261,"p90":94.54500675201416,"p95":97.70703315734863,"p99":104.55900430679321},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":26,"mean":20.5625,"min":15,"total":329},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":86727.2161728771,"p90":85200.7610241499,"p95":84717.83039278365,"p99":83600.24553549879},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":198.37799668312073,"p90":202.94399559497833,"p95":204.94699478149414,"p99":208.91299843788147},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":462.1709883213043,"p90":473.3079969882965,"p95":477.8369963169098,"p99":490.05499482154846},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":660.548985004425,"p90":676.2519925832748,"p95":682.7839910984039,"p99":698.9679932594299},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":859.2479825019836,"p90":872.1470236778259,"p95":880.8810114860535,"p99":1020.4499959945679},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":739.0289902687073,"p90":753.0099749565125,"p95":757.535994052887,"p99":767.9920196533203},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":232.22699761390686,"p90":236.71400547027588,"p95":239.07700181007385,"p99":243.28400194644928},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":521.9399929046631,"p90":532.8360199928284,"p95":535.9609723091125,"p99":551.5850186347961},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.330013275146484,"p90":6.3700079917907715,"p95":6.3700079917907715,"p99":6.3700079917907715},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.404995918273926,"p90":12.659013271331787,"p95":15.507042407989502,"p99":19.304990768432617},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.4850239753723145,"p90":1.5619993209838867,"p95":1.5619993209838867,"p99":1.5619993209838867},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":174.02000725269318,"p90":176.5040010213852,"p95":178.2660037279129,"p99":182.75299668312073},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":376.0420083999634,"p90":384.0540051460266,"p95":386.77799701690674,"p99":393.9889967441559},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":656.6200256347656,"p90":669.0000295639038,"p95":673.445999622345,"p99":684.7429871559143},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":81.20203018188477,"p90":95.30502557754517,"p95":98.15001487731934,"p99":104.8399806022644},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":49,"mean":40.5625,"min":36,"total":649},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":173200.2420547262,"p90":169984.46801105418,"p95":168968.86881267288,"p99":166668.39853073025},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":205.58799803256989,"p90":210.43600142002106,"p95":212.3579978942871,"p99":217.00499951839447},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":461.8529975414276,"p90":473.7510085105896,"p95":477.8349995613098,"p99":486.7669939994812},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":667.4409955739975,"p90":684.1870099306107,"p95":690.1929974555969,"p99":703.7719935178757},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":859.8899841308594,"p90":873.989999294281,"p95":884.7669959068298,"p99":1048.130989074707},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":739.3050193786621,"p90":752.5690197944641,"p95":757.0120096206665,"p99":773.2359766960144},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":238.4359985589981,"p90":243.80500614643097,"p95":246.44799530506134,"p99":250.73501467704773},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":515.811026096344,"p90":527.2679924964905,"p95":530.472993850708,"p99":542.9720282554626},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.328999996185303,"p90":6.410002708435059,"p95":6.410002708435059,"p99":6.410002708435059},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.485998630523682,"p90":13.624966144561768,"p95":14.901041984558105,"p99":18.027007579803467},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-4.204988479614258,"p90":2.8839707374572754,"p95":2.8839707374572754,"p99":2.8839707374572754},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":180.2700012922287,"p90":182.71300196647644,"p95":184.5950037240982,"p99":188.48200142383575},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":383.8129937648773,"p90":392.54701137542725,"p95":395.03100514411926,"p99":401.84101462364197},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":663.2710099220276,"p90":675.3699779510498,"p95":679.5740127563477,"p99":691.8339729309082},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":74.35500621795654,"p90":88.6160135269165,"p95":92.14097261428833,"p99":99.83295202255249},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":89,"mean":82.125,"min":72,"total":1314},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":346271.1509995582,"p90":340168.1351032983,"p95":338171.6495201705,"p99":331076.1626662418},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":222.13299572467804,"p90":226.89999639987946,"p95":229.18400168418884,"p99":233.18900167942047},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":462.13299036026,"p90":473.6280143260956,"p95":477.5960147380829,"p99":487.0480000972748},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":684.265986084938,"p90":700.528010725975,"p95":706.7800164222717,"p99":720.2370017766953},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":861.7280125617981,"p90":877.9119849205017,"p95":891.4530277252197,"p99":1027.1799564361572},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":739.0279769897461,"p90":753.4059882164001,"p95":757.8169703483582,"p99":770.7560062408447},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":254.02000546455383,"p90":258.8669955730438,"p95":261.35000586509705,"p99":266.35798811912537},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":499.6669888496399,"p90":509.3640089035034,"p95":512.2860074043274,"p99":522.4609971046448},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.329953670501709,"p90":6.330013275146484,"p95":6.330013275146484,"p99":6.330013275146484},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.8879594802856445,"p90":14.024972915649414,"p95":15.744030475616455,"p99":18.424034118652344},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.884984016418457,"p90":6.929993629455566,"p95":6.929993629455566,"p99":6.929993629455566},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":196.09199464321136,"p90":198.17699491977692,"p95":199.7780054807663,"p99":204.26499843597412},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":404.96599674224854,"p90":412.1760129928589,"p95":414.6600067615509,"p99":421.2290048599243},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":679.3749928474426,"p90":692.5150156021118,"p95":697.6420283317566,"p99":710.502028465271},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":58.369994163513184,"p90":72.55196571350098,"p95":76.11697912216187,"p99":86.65299415588379},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":178,"mean":164.0,"min":148,"total":2624},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":692801.918116158,"p90":679580.4758760938,"p95":675624.8804571381,"p99":664282.8545665734},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":246.20899558067322,"p90":251.37698650360107,"p95":253.17999720573425,"p99":256.02400302886963},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":466.1790132522583,"p90":477.5559902191162,"p95":481.8420112133026,"p99":491.6970133781433},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":712.3880088329315,"p90":728.9329767227173,"p95":735.0220084190369,"p99":747.7210164070129},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":869.2229986190796,"p90":885.6080174446106,"p95":902.3129940032959,"p99":983.9109778404236},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":738.103985786438,"p90":752.8849840164185,"p95":758.0530047416687,"p99":776.7210006713867},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":281.30000829696655,"p90":287.10898756980896,"p95":290.39499163627625,"p99":299.8889982700348},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":476.99299454689026,"p90":489.61201310157776,"p95":501.4690160751343,"p99":522.7410197257996},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.290018558502197,"p90":6.410002708435059,"p95":6.410002708435059,"p99":6.410002708435059},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.131040096282959,"p90":13.06295394897461,"p95":15.262961387634277,"p99":19.593000411987305},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.645956516265869,"p90":2.40403413772583,"p95":2.40403413772583,"p99":2.40403413772583},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":222.17200696468353,"p90":225.13699531555176,"p95":226.97900235652924,"p99":230.78499734401703},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":439.177006483078,"p90":447.7120041847229,"p95":450.9139955043793,"p99":456.2020003795624},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":705.3750157356262,"p90":719.2350029945374,"p95":724.1219878196716,"p99":736.3399863243103},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":31.045973300933838,"p90":44.5060133934021,"p95":48.99698495864868,"p99":57.40898847579956},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":363,"mean":328.5,"min":292,"total":5256},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":1387338.3963764189,"p90":1360101.505195738,"p95":1350829.0232936435,"p99":1318362.705675357},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":323.00499081611633,"p90":328.6130130290985,"p95":331.09599351882935,"p99":336.7049992084503},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":536.8049740791321,"p90":546.218991279602,"p95":550.0640273094177,"p99":564.6070241928101},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":859.8099648952484,"p90":874.8320043087006,"p95":881.1600208282471,"p99":901.3120234012604},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":940.2850270271301,"p90":954.0659785270691,"p95":960.7160091400146,"p99":992.8439855575562},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":849.9950170516968,"p90":861.411988735199,"p95":865.136981010437,"p99":875.993013381958},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":361.42098903656006,"p90":365.94700813293457,"p95":367.6289916038513,"p99":371.1149990558624},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":493.7399923801422,"p90":501.27100944519043,"p95":503.47501039505005,"p99":509.443998336792},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.289005279541016,"p90":6.3700079917907715,"p95":6.3700079917907715,"p99":6.3700079917907715},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11.341989040374756,"p90":16.460001468658447,"p95":18.347978591918945,"p99":21.757006645202637},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.924978733062744,"p90":3.6860108375549316,"p95":3.6860108375549316,"p99":3.6860108375549316},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":301.53000354766846,"p90":305.01601099967957,"p95":306.8990111351013,"p99":311.66601181030273},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":503.6730170249939,"p90":510.60301065444946,"p95":513.0869746208191,"p99":520.2170014381409},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":827.8769850730896,"p90":838.6120200157166,"p95":842.0990109443665,"p99":853.4749746322632},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":21.596968173980713,"p90":26.443004608154297,"p95":28.527021408081055,"p99":32.492995262145996},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0041841003112494946,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":692,"mean":660.8125,"min":603,"total":10573},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":2409425.8894642913,"p90":2377491.8700714326,"p95":2367255.180339231,"p99":2337918.1896591377},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":416.70501232147217,"p90":422.35299944877625,"p95":424.59601163864136,"p99":428.88298630714417},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":668.0819988250732,"p90":679.5759797096252,"p95":682.6630234718323,"p99":688.4289979934692},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1084.7870111465454,"p90":1101.9289791584015,"p95":1107.2590351104736,"p99":1117.3119843006134},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1118.6720132827759,"p90":1128.9680004119873,"p95":1131.8919658660889,"p99":1138.3470296859741},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1066.7999982833862,"p90":1078.4980058670044,"p95":1081.9029808044434,"p99":1088.832974433899},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":468.3400094509125,"p90":474.75001215934753,"p95":476.75201296806335,"p99":482.36098885536194},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":599.5389819145203,"p90":606.8699955940247,"p95":608.9529991149902,"p99":614.6810054779053},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.448984146118164,"p90":6.569027900695801,"p95":6.569027900695801,"p99":6.569027900695801},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.912983894348145,"p90":25.35700798034668,"p95":27.116060256958008,"p99":31.48198127746582},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.0050277709960938,"p90":3.6859512329101562,"p95":3.6859512329101562,"p99":3.6859512329101562},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":383.45301151275635,"p90":387.73998618125916,"p95":389.6220028400421,"p99":393.3880031108856},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":622.8110194206238,"p90":630.9030055999756,"p95":633.6669921875,"p99":639.9999856948853},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1033.5049629211426,"p90":1044.0000295639038,"p95":1047.685980796814,"p99":1055.938959121704},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":33.09500217437744,"p90":38.58292102813721,"p95":40.46595096588135,"p99":45.51398754119873},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0041841003112494946,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":1373,"mean":1323.8125,"min":1247,"total":21181},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":3839520.0661707665,"p90":3797874.4306599125,"p95":3785921.725582492,"p99":3761825.82285366},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":545.619010925293,"p90":555.0320148468018,"p95":558.7980151176453,"p99":569.7739720344543},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":908.161997795105,"p90":917.7759885787964,"p95":920.9759831428528,"p99":928.0670285224915},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1453.781008720398,"p90":1472.8080034255981,"p95":1479.773998260498,"p99":1497.8410005569458},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1529.4129848480225,"p90":1543.0259704589844,"p95":1547.8010177612305,"p99":1556.6539764404297},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1442.6829814910889,"p90":1456.3839435577393,"p95":1462.7139568328857,"p99":1476.4939546585083},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":591.5639996528625,"p90":599.856972694397,"p95":602.8209924697876,"p99":607.2700023651123},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":877.1960139274597,"p90":885.6080174446106,"p95":888.808012008667,"p99":895.8230018615723},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.36899471282959,"p90":6.569027900695801,"p95":6.569027900695801,"p99":6.569027900695801},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":31.728029251098633,"p90":44.89898681640625,"p95":49.58701133728027,"p99":63.246965408325195},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.767060279846191,"p90":7.290959358215332,"p95":7.290959358215332,"p99":7.290959358215332},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":506.2360167503357,"p90":512.4459862709045,"p95":514.2890214920044,"p99":518.8959836959839},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":870.8620071411133,"p90":879.4379830360413,"p95":881.9180130958557,"p99":887.8870010375977},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1402.2159576416016,"p90":1415.31503200531,"p95":1419.1210269927979,"p99":1426.6120195388794},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":39.78705406188965,"p90":49.80003833770752,"p95":53.80702018737793,"p99":64.38195705413818},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0095911156386137,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"roundtrip":{"activation_data_bytes":1215807488,"scale_bytes":0,"total_logical_bytes":1215807488},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"roundtrip":{"activation_data_bytes":1215807488,"scale_bytes":0,"total_logical_bytes":1215807488},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":42404,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":2702,"mean":2650.25,"min":2578,"total":42404},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.017352504655718803,"expert_assignments_per_rank":[3108,3051,3103,2954,3171,3067,3121,2998,3045,3065,3095,3023,3150,3071,3090,3040],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,3,103,1301,3825,2960,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.17626953125,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.06268276274204254,"same_node_fraction":0.49721723794937134,"same_scaleup_domain_fraction":0.49721723794937134,"cross_node_fraction":0.5027827620506287,"cross_domain_fraction":0.5027827620506287,"gpus_per_node":8,"scale_up_domain":8,"copies":42404},"payload_copies_per_rank":[2682,2637,2676,2578,2702,2659,2687,2602,2635,2661,2672,2599,2691,2625,2662,2636],"payload_rank_cv":0.013201921246945858,"routed_copies":42404},"token_rate_at_latency_percentile":{"p50":5678309.167779284,"p90":5624890.356857483,"p95":5600548.187656304,"p99":5548278.727558144},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":36,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":20,"num_max_rdma_chunked_recv_tokens":140},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":12,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":1073741824,"num_qps_per_rank":10,"is_intranode":false,"cpu_proxy_threads":4},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9938,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2095,"p50":2100,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2099,"p50":2102,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2094,"p50":2100,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.8},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2097,"p50":2104,"max":2106,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.8},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2086,"p50":2096,"max":2097,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2100,"p50":2105,"max":2107,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2103,"p50":2109,"max":2111,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2093,"p50":2098,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2095,"p50":2104,"max":2106,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2099,"p50":2104,"max":2106,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2101,"p50":2106,"max":2108,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2096,"p50":2101,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":54,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2094,"p50":2104,"max":2106,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2102,"max":2103,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2099,"p50":2103,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2097,"p50":2102,"max":2103,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":452.9}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-fp8.json b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-fp8.json new file mode 100644 index 0000000000..5841cb8ce8 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T11:11:54.254794+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":2,"phase":"decode","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":190.80699980258942,"p90":195.37299871444702,"p95":197.85700738430023,"p99":202.14399695396423},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":489.97199535369873,"p90":502.672016620636,"p95":506.6779851913452,"p99":520.8590030670166},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":680.7789951562881,"p90":698.045015335083,"p95":704.5349925756454,"p99":723.0030000209808},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1023.8909721374512,"p90":1034.8269939422607,"p95":1038.7200117111206,"p99":1053.9829730987549},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":781.1679840087891,"p90":796.0309982299805,"p95":800.6359934806824,"p99":812.9760026931763},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.508000299334526,"p90":20.671000704169273,"p95":21.553000435233116,"p99":23.274999111890793},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":234.6310019493103,"p90":239.87899720668793,"p95":242.2019988298416,"p99":246.96999788284302},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":634.6330046653748,"p90":649.616003036499,"p95":653.7820100784302,"p99":663.6769771575928},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.530046463012695,"p90":6.7299604415893555,"p95":6.7299604415893555,"p99":6.7299604415893555},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.648016929626465,"p90":10.851025581359863,"p95":13.180017471313477,"p99":16.41702651977539},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.2019872665405273,"p90":3.7660598754882812,"p95":3.7660598754882812,"p99":3.7660598754882812},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":159.63999927043915,"p90":163.64499926567078,"p95":165.569007396698,"p99":169.69500482082367},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":384.89800691604614,"p90":392.630010843277,"p95":394.912987947464,"p99":399.7580111026764},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":675.2949953079224,"p90":689.9570226669312,"p95":695.1249837875366,"p99":704.0979862213135},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":104.95299100875854,"p90":114.53300714492798,"p95":117.2100305557251,"p99":123.42000007629395},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":587776,"scale_bytes":18368,"total_logical_bytes":606144},"roundtrip":{"activation_data_bytes":1763328,"scale_bytes":18368,"total_logical_bytes":1781696},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":587776,"scale_bytes":18368,"total_logical_bytes":606144},"roundtrip":{"activation_data_bytes":1763328,"scale_bytes":18368,"total_logical_bytes":1781696},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":9,"mean":5.125,"min":2,"total":82},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":20482.150225731708,"p90":20099.719779225805,"p95":19984.112793182892,"p99":19680.77771914078},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":196.53500616550446,"p90":201.42300426959991,"p95":203.5049945116043,"p99":207.67199993133545},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":490.4940128326416,"p90":502.8759837150574,"p95":507.3220133781433,"p99":516.6530013084412},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":687.0290189981461,"p90":704.2989879846573,"p95":710.8270078897476,"p99":724.3250012397766},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1023.6510038375854,"p90":1060.1450204849243,"p95":1067.957043647766,"p99":1081.6650390625},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":781.6540002822876,"p90":796.4720129966736,"p95":800.9179830551147,"p99":813.817024230957},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.547000363469124,"p90":21.472999826073647,"p95":22.593999281525612,"p99":23.91500025987625},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":233.71000587940216,"p90":239.4780069589615,"p95":241.60200357437134,"p99":244.08499896526337},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":625.4990100860596,"p90":639.5599842071533,"p95":645.0480222702026,"p99":659.870982170105},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.330013275146484,"p90":6.5299272537231445,"p95":6.5299272537231445,"p99":6.5299272537231445},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.121993064880371,"p90":12.219071388244629,"p95":13.579964637756348,"p99":16.865015029907227},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.2430419921875,"p90":22.392988204956055,"p95":22.392988204956055,"p99":22.392988204956055},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":171.4159995317459,"p90":174.02000725269318,"p95":176.26400291919708,"p99":179.98899519443512},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":387.18199729919434,"p90":393.99001002311707,"p95":396.51599526405334,"p99":401.56298875808716},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":687.628984451294,"p90":702.8509974479675,"p95":707.4180245399475,"p99":718.2340025901794},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":92.70000457763672,"p90":102.4789810180664,"p95":105.19695281982422,"p99":110.65101623535156},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":1189888,"scale_bytes":37184,"total_logical_bytes":1227072},"roundtrip":{"activation_data_bytes":3569664,"scale_bytes":37184,"total_logical_bytes":3606848},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":1189888,"scale_bytes":37184,"total_logical_bytes":1227072},"roundtrip":{"activation_data_bytes":3569664,"scale_bytes":37184,"total_logical_bytes":3606848},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":15,"mean":10.375,"min":7,"total":166},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":40938.82969759443,"p90":40177.180714237664,"p95":39954.15345518336,"p99":39320.87809325376},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":197.2160041332245,"p90":201.66300237178802,"p95":203.86600494384766,"p99":207.47099816799164},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":491.5750026702881,"p90":504.11397218704224,"p95":508.08000564575195,"p99":518.3389782905579},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":688.7910068035126,"p90":705.7769745588303,"p95":711.9460105895996,"p99":725.8099764585495},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1028.5849571228027,"p90":1041.396975517273,"p95":1043.5600280761719,"p99":1066.9560432434082},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":785.1740121841431,"p90":799.5160222053528,"p95":804.8040270805359,"p99":816.461980342865},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.907999619841576,"p90":21.83300070464611,"p95":22.873999550938606,"p99":24.396000429987907},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":236.27400398254395,"p90":241.6819930076599,"p95":243.7639981508255,"p99":247.69000709056854},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":627.2220015525818,"p90":640.0820016860962,"p95":644.3279981613159,"p99":650.738000869751},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.370067596435547,"p90":6.489992141723633,"p95":6.489992141723633,"p99":6.489992141723633},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.130980491638184,"p90":12.658953666687012,"p95":14.382004737854004,"p99":17.83299446105957},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.1239986419677734,"p90":-1.762986183166504,"p95":-1.762986183166504,"p99":-1.762986183166504},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":173.458993434906,"p90":176.2239933013916,"p95":177.9859960079193,"p99":182.11199343204498},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":386.0970139503479,"p90":393.9119875431061,"p95":397.1160054206848,"p99":402.8019905090332},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":692.1150088310242,"p90":706.2960267066956,"p95":711.3839983940125,"p99":724.2029905319214},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":91.93795919418335,"p90":101.63801908493042,"p95":104.39598560333252,"p99":109.92497205734253},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":2358272,"scale_bytes":73696,"total_logical_bytes":2431968},"roundtrip":{"activation_data_bytes":7074816,"scale_bytes":73696,"total_logical_bytes":7148512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":2358272,"scale_bytes":73696,"total_logical_bytes":2431968},"roundtrip":{"activation_data_bytes":7074816,"scale_bytes":73696,"total_logical_bytes":7148512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":26,"mean":20.5625,"min":15,"total":329},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":81510.59383889847,"p90":80048.42707650183,"p95":79522.46490634867,"p99":78386.99356597579},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":198.57800006866455,"p90":203.30500602722168,"p95":205.46899735927582,"p99":210.3559970855713},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":491.7759895324707,"p90":504.11397218704224,"p95":508.44401121139526,"p99":518.8159942626953},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":690.3539896011353,"p90":707.4189782142639,"p95":713.9130085706711,"p99":729.1719913482666},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1030.2259922027588,"p90":1042.7589416503906,"p95":1046.9650030136108,"p99":1074.6879577636719},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":785.1340174674988,"p90":800.1160025596619,"p95":805.525004863739,"p99":815.6599998474121},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":19.869999960064888,"p90":22.433999925851822,"p95":23.594999685883522,"p99":25.318000465631485},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":238.47700655460358,"p90":243.16400289535522,"p95":245.2870011329651,"p99":247.93000519275665},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":628.0630230903625,"p90":639.6809816360474,"p95":643.4869766235352,"p99":651.1380076408386},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.369948387145996,"p90":6.409049034118652,"p95":6.409049034118652,"p99":6.409049034118652},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.536960601806641,"p90":12.778997421264648,"p95":14.909029006958008,"p99":18.58806610107422},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.885103225708008,"p90":3.285050392150879,"p95":3.285050392150879,"p99":3.285050392150879},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":173.85999858379364,"p90":176.62400007247925,"p95":178.14600467681885,"p99":182.15200304985046},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":391.4650082588196,"p90":399.55800771713257,"p95":402.2420048713684,"p99":408.0899953842163},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":691.8749809265137,"p90":706.3760161399841,"p95":712.3860120773315,"p99":721.6799855232239},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":92.30399131774902,"p90":101.6319990158081,"p95":104.08204793930054,"p99":110.64600944519043},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":4652032,"scale_bytes":145376,"total_logical_bytes":4797408},"roundtrip":{"activation_data_bytes":13956096,"scale_bytes":145376,"total_logical_bytes":14101472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":4652032,"scale_bytes":145376,"total_logical_bytes":4797408},"roundtrip":{"activation_data_bytes":13956096,"scale_bytes":145376,"total_logical_bytes":14101472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":49,"mean":40.5625,"min":36,"total":649},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":163029.4919749782,"p90":159976.80285172837,"p95":158902.57810389416,"p99":156928.13184898774},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":205.7889997959137,"p90":210.23599803447723,"p95":212.43900060653687,"p99":218.04699301719666},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":492.89700388908386,"p90":504.8750042915344,"p95":508.6809992790222,"p99":519.0160274505615},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":698.6860036849976,"p90":715.1110023260117,"p95":721.1199998855591,"p99":737.0630204677582},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1032.824993133545,"p90":1043.4800386428833,"p95":1047.613024711609,"p99":1055.2250146865845},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":785.2140069007874,"p90":799.9569773674011,"p95":804.9240112304688,"p99":814.6190047264099},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":23.91600050032139,"p90":26.079000905156136,"p95":27.241000905632973,"p99":28.88300083577633},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":242.88299679756165,"p90":248.69200587272644,"p95":250.33500790596008,"p99":253.3380091190338},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":621.9750046730042,"p90":634.0320110321045,"p95":637.4369859695435,"p99":643.4460282325745},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.530046463012695,"p90":6.570100784301758,"p95":6.570100784301758,"p99":6.570100784301758},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.290959358215332,"p90":12.337923049926758,"p95":14.301896095275879,"p99":17.580032348632812},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-4.325985908508301,"p90":5.568981170654297,"p95":5.568981170654297,"p99":5.568981170654297},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":180.62999844551086,"p90":182.91400372982025,"p95":184.63599681854248,"p99":189.20299410820007},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":393.26798915863037,"p90":401.71998739242554,"p95":404.244989156723,"p99":410.2939963340759},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":698.7249851226807,"p90":712.5459909439087,"p95":717.2729969024658,"p99":726.8069982528687},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":85.76798439025879,"p90":94.86198425292969,"p95":97.98598289489746,"p99":103.35499048233032},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":9418752,"scale_bytes":294336,"total_logical_bytes":9713088},"roundtrip":{"activation_data_bytes":28256256,"scale_bytes":294336,"total_logical_bytes":28550592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":9418752,"scale_bytes":294336,"total_logical_bytes":9713088},"roundtrip":{"activation_data_bytes":28256256,"scale_bytes":294336,"total_logical_bytes":28550592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":89,"mean":82.125,"min":72,"total":1314},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":326025.7684531421,"p90":320017.2099785628,"p95":318042.44429068366,"p99":314257.33811105683},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":222.1740037202835,"p90":226.65999829769135,"p95":229.0239930152893,"p99":232.3089987039566},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":493.2970106601715,"p90":505.2350163459778,"p95":509.48500633239746,"p99":518.8559889793396},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":715.471014380455,"p90":731.8950146436691,"p95":738.5089993476868,"p99":751.1649876832962},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1034.5139503479004,"p90":1045.7299947738647,"p95":1049.7690439224243,"p99":1062.0750188827515},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":786.6569757461548,"p90":801.5589714050293,"p95":807.8089952468872,"p99":817.6630139350891},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":28.04199978709221,"p90":30.525999143719673,"p95":31.567998230457306,"p99":33.649999648332596},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":258.8669955730438,"p90":263.9150023460388,"p95":266.0380005836487,"p99":269.60399746894836},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":603.6660075187683,"p90":615.8450245857239,"p95":620.3709840774536,"p99":628.8250088691711},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.450057029724121,"p90":6.530046463012695,"p95":6.530046463012695,"p99":6.530046463012695},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.2040557861328125,"p90":12.906074523925781,"p95":15.869975090026855,"p99":22.068023681640625},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.3241043090820312,"p90":3.363966941833496,"p95":3.363966941833496,"p99":3.363966941833496},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":196.13300263881683,"p90":198.53700697422028,"p95":200.2590000629425,"p99":204.46600019931793},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":396.03298902511597,"p90":404.44499254226685,"p95":406.8880081176758,"p99":414.7000014781952},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":715.5500054359436,"p90":730.0119996070862,"p95":734.5790266990662,"p99":745.514988899231},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":70.26499509811401,"p90":79.59997653961182,"p95":82.52400159835815,"p99":90.09498357772827},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":18808832,"scale_bytes":587776,"total_logical_bytes":19396608},"roundtrip":{"activation_data_bytes":56426496,"scale_bytes":587776,"total_logical_bytes":57014272},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":18808832,"scale_bytes":587776,"total_logical_bytes":19396608},"roundtrip":{"activation_data_bytes":56426496,"scale_bytes":587776,"total_logical_bytes":57014272},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":178,"mean":164.0,"min":148,"total":2624},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":650855.475494082,"p90":638755.2485408905,"p95":633813.1947187833,"p99":626174.8315310807},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":245.89000642299652,"p90":250.69600343704224,"p95":252.579003572464,"p99":257.26601481437683},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":494.0989911556244,"p90":506.03699684143066,"p95":510.7640027999878,"p99":523.6629843711853},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":739.9889975786209,"p90":756.7330002784729,"p95":763.3430063724518,"p99":780.9289991855621},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1033.5450172424316,"p90":1045.3699827194214,"p95":1049.7699975967407,"p99":1073.2109546661377},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":788.8990044593811,"p90":804.8840165138245,"p95":809.8919987678528,"p99":823.9529728889465},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":39.17799890041351,"p90":40.94099998474121,"p95":41.94299876689911,"p99":43.946001678705215},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":287.0289981365204,"p90":291.87700152397156,"p95":293.599009513855,"p99":296.44399881362915},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":582.7149748802185,"p90":597.0969796180725,"p95":601.1019945144653,"p99":607.1919798851013},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.36899471282959,"p90":6.690025329589844,"p95":6.690025329589844,"p99":6.690025329589844},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.370948791503906,"p90":13.34691047668457,"p95":14.348983764648438,"p99":17.065048217773438},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.7659406661987305,"p90":5.92803955078125,"p95":5.92803955078125,"p99":5.92803955078125},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":222.37299382686615,"p90":225.33699870109558,"p95":227.18000411987305,"p99":231.26600682735443},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":410.97500920295715,"p90":424.5150089263916,"p95":427.7999997138977,"p99":435.1710081100464},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":743.9929842948914,"p90":758.9750289916992,"p95":763.342022895813,"p99":774.6790051460266},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":44.50702667236328,"p90":54.08596992492676,"p95":57.20597505569458,"p99":64.3010139465332},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":37675008,"scale_bytes":1177344,"total_logical_bytes":38852352},"roundtrip":{"activation_data_bytes":113025024,"scale_bytes":1177344,"total_logical_bytes":114202368},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":37675008,"scale_bytes":1177344,"total_logical_bytes":38852352},"roundtrip":{"activation_data_bytes":113025024,"scale_bytes":1177344,"total_logical_bytes":114202368},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":363,"mean":328.5,"min":292,"total":5256},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":1298011.5251910219,"p90":1272232.9913261635,"p95":1264366.1149361706,"p99":1242789.3747499303},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":324.8080015182495,"p90":329.8560082912445,"p95":332.29899406433105,"p99":336.34498715400696},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":497.063010931015,"p90":509.8459720611572,"p95":513.8480067253113,"p99":523.1419801712036},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":821.8710124492645,"p90":839.7019803524017,"p95":846.1470007896423,"p99":859.4869673252106},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1027.703046798706,"p90":1041.524052619934,"p95":1045.1229810714722,"p99":1070.5679655075073},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":847.4720120429993,"p90":864.2190098762512,"p95":869.2259788513184,"p99":879.4419765472412},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":48.07199910283089,"p90":50.276000052690506,"p95":50.996001809835434,"p99":52.43900045752525},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":366.8690025806427,"p90":373.1580078601837,"p95":375.04100799560547,"p99":379.88799810409546},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":493.06100606918335,"p90":509.80502367019653,"p95":513.3709907531738,"p99":520.9019780158997},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.450057029724121,"p90":6.6100358963012695,"p95":6.6100358963012695,"p99":6.6100358963012695},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":12.296915054321289,"p90":17.701029777526855,"p95":19.629955291748047,"p99":23.074984550476074},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-5.407929420471191,"p90":3.365039825439453,"p95":3.365039825439453,"p99":3.365039825439453},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":301.77098512649536,"p90":305.2169978618622,"p95":306.9390058517456,"p99":311.06600165367126},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":468.66101026535034,"p90":477.1130084991455,"p95":480.31800985336304,"p99":487.68898844718933},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":823.2309818267822,"p90":840.4170274734497,"p95":844.8240160942078,"p99":856.4410209655762},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":23.521006107330322,"p90":28.407037258148193,"p95":30.29000759124756,"p99":33.45602750778198},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0041841003112494946,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":75787264,"scale_bytes":2368352,"total_logical_bytes":78155616},"roundtrip":{"activation_data_bytes":227361792,"scale_bytes":2368352,"total_logical_bytes":229730144},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":75787264,"scale_bytes":2368352,"total_logical_bytes":78155616},"roundtrip":{"activation_data_bytes":227361792,"scale_bytes":2368352,"total_logical_bytes":229730144},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":692,"mean":660.8125,"min":603,"total":10573},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":2416598.980139639,"p90":2369769.6724968553,"p95":2356119.1793950186,"p99":2328749.4281778657},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":416.3860082626343,"p90":422.9159951210022,"p95":425.03899335861206,"p99":430.36600947380066},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":595.4139828681946,"p90":607.1919798851013,"p95":611.4789843559265,"p99":623.8170266151428},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1011.7999911308289,"p90":1030.1079750061035,"p95":1036.5179777145386,"p99":1054.1830360889435},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1103.337049484253,"p90":1117.1499490737915,"p95":1124.6490478515625,"p99":1149.7199535369873},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1006.3120126724243,"p90":1016.808032989502,"p95":1019.8520421981812,"p99":1029.505968093872},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":70.54600119590759,"p90":72.86900281906128,"p95":73.55000078678131,"p99":75.51299780607224},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":458.08500051498413,"p90":463.7329876422882,"p95":466.09699726104736,"p99":473.7890064716339},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":544.6140170097351,"p90":553.1110167503357,"p95":556.2760233879089,"p99":559.7209930419922},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.36899471282959,"p90":6.450057029724121,"p95":6.450057029724121,"p99":6.450057029724121},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.34702491760254,"p90":25.398969650268555,"p95":27.48095989227295,"p99":32.369017601013184},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.006028175354004,"p90":5.568027496337891,"p95":5.568027496337891,"p99":5.568027496337891},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":385.21599769592285,"p90":389.54299688339233,"p95":391.50598645210266,"p99":394.910991191864},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":566.2469863891602,"p90":575.3809809684753,"p95":578.3050060272217,"p99":584.7150087356567},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":974.6180176734924,"p90":984.5920205116272,"p95":987.6769781112671,"p99":995.809018611908},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":31.293034553527832,"p90":36.74203157424927,"p95":38.4640097618103,"p99":44.515013694763184},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0041841003112494946,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":151825408,"scale_bytes":4744544,"total_logical_bytes":156569952},"roundtrip":{"activation_data_bytes":455476224,"scale_bytes":4744544,"total_logical_bytes":460220768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":151825408,"scale_bytes":4744544,"total_logical_bytes":156569952},"roundtrip":{"activation_data_bytes":455476224,"scale_bytes":4744544,"total_logical_bytes":460220768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":1373,"mean":1323.8125,"min":1247,"total":21181},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":4070308.1632926255,"p90":4028292.329632185,"p95":4016268.861090393,"p99":3978607.338803227},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":547.182023525238,"p90":557.5979948043823,"p95":563.8070106506348,"p99":578.9099931716919},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":787.2220277786255,"p90":800.5229830741882,"p95":805.4500222206116,"p99":815.185010433197},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1334.4040513038635,"p90":1358.1209778785706,"p95":1369.2570328712463,"p99":1394.095003604889},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1393.0829763412476,"p90":1405.0300121307373,"p95":1410.3180170059204,"p99":1418.0500507354736},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1325.3910541534424,"p90":1339.7740125656128,"p95":1344.2590236663818,"p99":1357.5210571289062},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":117.53600090742111,"p90":120.46000361442566,"p95":121.42200022935867,"p99":123.74600023031235},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":611.0339760780334,"p90":618.9659833908081,"p95":621.4889883995056,"p99":626.9369721412659},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":734.142005443573,"p90":741.7140007019043,"p95":743.6370253562927,"p99":749.0000128746033},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.410002708435059,"p90":6.609916687011719,"p95":6.609916687011719,"p99":6.609916687011719},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":29.836058616638184,"p90":41.934967041015625,"p95":45.31705379486084,"p99":59.19909477233887},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.1239986419677734,"p90":5.528092384338379,"p95":5.528092384338379,"p99":5.528092384338379},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":512.4459862709045,"p90":517.8149938583374,"p95":519.5770263671875,"p99":523.4230160713196},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":756.650984287262,"p90":766.2659883499146,"p95":769.1910266876221,"p99":775.3999829292297},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1289.6889448165894,"p90":1301.78701877594,"p95":1304.751992225647,"p99":1312.883973121643},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":35.061001777648926,"p90":46.07796669006348,"p95":50.24409294128418,"p99":62.38198280334473},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0041841003112494946,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":303951872,"scale_bytes":9498496,"total_logical_bytes":313450368},"roundtrip":{"activation_data_bytes":911855616,"scale_bytes":9498496,"total_logical_bytes":921354112},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":303951872,"scale_bytes":9498496,"total_logical_bytes":313450368},"roundtrip":{"activation_data_bytes":911855616,"scale_bytes":9498496,"total_logical_bytes":921354112},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":42404,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":2702,"mean":2650.25,"min":2578,"total":42404},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.017352504655718803,"expert_assignments_per_rank":[3108,3051,3103,2954,3171,3067,3121,2998,3045,3065,3095,3023,3150,3071,3090,3040],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,3,103,1301,3825,2960,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.17626953125,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.06268276274204254,"same_node_fraction":0.49721723794937134,"same_scaleup_domain_fraction":0.49721723794937134,"cross_node_fraction":0.5027827620506287,"cross_domain_fraction":0.5027827620506287,"gpus_per_node":8,"scale_up_domain":8,"copies":42404},"payload_copies_per_rank":[2682,2637,2676,2578,2702,2659,2687,2602,2635,2661,2672,2599,2691,2625,2662,2636],"payload_rank_cv":0.013201921246945858,"routed_copies":42404},"token_rate_at_latency_percentile":{"p50":6180817.332611633,"p90":6114464.023908519,"p95":6094063.611086527,"p99":6034528.8619873775},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":36,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":20,"num_max_rdma_chunked_recv_tokens":140},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":12,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":1073741824,"num_qps_per_rank":10,"is_intranode":false,"cpu_proxy_threads":4},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9934,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2094,"p50":2101,"max":2103,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.9},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2096,"p50":2102,"max":2105,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.8},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2096,"p50":2100,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.9},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2101,"p50":2104,"max":2105,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.9},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2090,"p50":2095,"max":2097,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.9},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2099,"p50":2106,"max":2107,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.9},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2102,"p50":2109,"max":2110,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.8},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2098,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.8},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2098,"p50":2104,"max":2106,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.9},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2098,"p50":2104,"max":2106,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.8},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2101,"p50":2106,"max":2108,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.9},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2096,"p50":2101,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":54,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.9},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2097,"p50":2103,"max":2105,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.9},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2097,"p50":2102,"max":2103,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.9},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2097,"p50":2103,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.9},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2097,"p50":2102,"max":2103,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":662.9}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-bf16.json b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-bf16.json new file mode 100644 index 0000000000..a372d5144d --- /dev/null +++ b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T06:44:31.605471+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":1,"phase":"decode","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":73.22800159454346,"p90":77.47600227594376,"p95":78.79800349473953,"p99":81.56300336122513},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":210.79599857330322,"p90":220.93099355697632,"p95":225.85800290107727,"p99":238.23699355125427},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":284.0240001678467,"p90":298.4069958329201,"p95":304.6560063958168,"p99":319.7999969124794},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":328.7329971790314,"p90":342.43300557136536,"p95":348.3220040798187,"p99":385.89799404144287},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":291.2369966506958,"p90":302.5740087032318,"p95":307.7009916305542,"p99":317.63601303100586},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":91.61700308322906,"p90":95.5829992890358,"p95":97.8270024061203,"p99":106.07899725437164},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":222.41300344467163,"p90":233.78999531269073,"p95":239.84000086784363,"p99":254.22099232673645},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.328999996185303,"p90":6.4910054206848145,"p95":6.4910054206848145,"p99":6.4910054206848145},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.411003112792969,"p90":15.943974256515503,"p95":20.110011100769043,"p99":33.28999876976013},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.1239986419677734,"p90":8.893996477127075,"p95":8.893996477127075,"p99":8.893996477127075},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":65.81799685955048,"p90":67.781001329422,"p95":69.22399997711182,"p99":71.82800024747849},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":184.71699953079224,"p90":192.28799641132355,"p95":194.73199546337128,"p99":201.90200209617615},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":282.58299827575684,"p90":293.4400141239166,"p95":298.40800166130066,"p99":308.54201316833496},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.932007312774658,"p90":11.8580162525177,"p95":13.660997152328491,"p99":20.391017198562622},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8,"byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"roundtrip":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"roundtrip":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":35,"assignments":48,"wire":"rank-deduplicated"},"receive":{"max":6,"mean":4.375,"min":2,"total":35},"routing":{"empty_expert_count":337,"empty_rank_count":0,"expert_assignment_rank_cv":0.3227486312389374,"expert_assignments_per_rank":[8,6,6,6,6,5,9,2],"expert_load_cv":2.708012819290161,"expert_load_max":2,"expert_load_mean":0.125,"expert_load_min":0,"fanout_histogram":[0,0,0,5,3,0,0,0],"fanout_max":5,"fanout_mean":4.375,"fanout_min":4,"hotspot_ratio":16.0,"locality":{"placement":"packed","local_rank_fraction":0.08571428805589676,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":35},"payload_copies_per_rank":[5,5,5,4,4,4,6,2],"payload_rank_cv":0.2539483904838562,"routed_copies":35},"token_rate_at_latency_percentile":{"p50":27469.03756048223,"p90":26439.812310007423,"p95":25999.266227927274,"p99":25186.061000013513},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":75.39299875497818,"p90":79.51900362968445,"p95":80.68100363016129,"p99":83.2040011882782},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":211.0770046710968,"p90":221.85200452804565,"p95":226.17900371551514,"p99":236.71500384807587},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":286.470003426075,"p90":301.3710081577301,"p95":306.8600073456764,"p99":319.91900503635406},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":330.29499650001526,"p90":346.48001194000244,"p95":352.08699107170105,"p99":371.1560070514679},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":294.20098662376404,"p90":307.3410093784332,"p95":313.14900517463684,"p99":329.57300543785095},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":93.74099969863892,"p90":99.22900050878525,"p95":101.67299956083298,"p99":108.68299752473831},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":224.17600452899933,"p90":232.74900019168854,"p95":236.47500574588776,"p99":243.24500560760498},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.28998875617981,"p90":6.76998496055603,"p95":6.76998496055603,"p99":6.76998496055603},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.971001625061035,"p90":17.705976963043213,"p95":20.55099606513977,"p99":26.960015296936035},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.9629895687103271,"p90":2.2839903831481934,"p95":2.2839903831481934,"p99":2.2839903831481934},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":66.89999997615814,"p90":69.70500200986862,"p95":71.30700349807739,"p99":74.11099970340729},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":184.797003865242,"p90":192.52799451351166,"p95":194.9719935655594,"p99":200.4999965429306},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":285.1470112800598,"p90":297.2460091114044,"p95":303.6560118198395,"p99":319.5990025997162},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.732020854949951,"p90":12.058019638061523,"p95":13.619989156723022,"p99":17.86699891090393},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"roundtrip":{"activation_data_bytes":2007040,"scale_bytes":0,"total_logical_bytes":2007040},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"roundtrip":{"activation_data_bytes":2007040,"scale_bytes":0,"total_logical_bytes":2007040},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":70,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":12,"mean":8.75,"min":5,"total":70},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.33592739701271057,"expert_assignments_per_rank":[18,17,13,7,13,7,13,8],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,3,6,5,2,0,0],"fanout_max":6,"fanout_mean":4.375,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.15714286267757416,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":70},"payload_copies_per_rank":[12,11,10,5,9,6,10,7],"payload_rank_cv":0.26649653911590576,"routed_copies":70},"token_rate_at_latency_percentile":{"p50":54384.58988059561,"p90":52059.43727574272,"p95":51093.88736865737,"p99":48547.66542163658},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":77.71699875593185,"p90":81.84300363063812,"p95":83.48499983549118,"p99":85.9690010547638},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":210.99600195884705,"p90":221.7320054769516,"p95":226.86000168323517,"p99":237.635999917984},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":288.7130007147789,"p90":303.5750091075897,"p95":310.34500151872635,"p99":323.6050009727478},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":328.61199975013733,"p90":342.3529863357544,"p95":348.24201464653015,"p99":438.9370083808899},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":296.6049909591675,"p90":309.54399704933167,"p95":316.2339925765991,"p99":332.33800530433655},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":96.94500267505646,"p90":102.51399874687195,"p95":104.1560024023056,"p99":107.60100185871124},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":217.36599504947662,"p90":228.302001953125,"p95":232.587993144989,"p99":240.679994225502},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.28998875617981,"p90":6.448984146118164,"p95":6.448984146118164,"p99":6.448984146118164},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.608022212982178,"p90":11.496007442474365,"p95":13.65998387336731,"p99":18.507003784179688},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.0439794063568115,"p90":3.4449994564056396,"p95":3.4449994564056396,"p99":3.4449994564056396},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":70.50599902868271,"p90":72.46900349855423,"p95":73.78900051116943,"p99":76.3539969921112},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":185.11800467967987,"p90":193.28999519348145,"p95":196.49499654769897,"p99":203.70499789714813},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":288.672000169754,"p90":301.5710115432739,"p95":307.4190020561218,"p99":323.8050043582916},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.292002439498901,"p90":10.856002569198608,"p95":12.657999992370605,"p99":16.34499430656433},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"roundtrip":{"activation_data_bytes":4100096,"scale_bytes":0,"total_logical_bytes":4100096},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"roundtrip":{"activation_data_bytes":4100096,"scale_bytes":0,"total_logical_bytes":4100096},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":143,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":22,"mean":17.875,"min":13,"total":143},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.20518453419208527,"expert_assignments_per_rank":[31,30,24,17,22,22,28,18],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,4,12,13,3,0,0],"fanout_max":6,"fanout_mean":4.46875,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.1258741319179535,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":143},"payload_copies_per_rank":[22,21,18,13,16,15,22,16],"payload_rank_cv":0.1800614446401596,"routed_copies":143},"token_rate_at_latency_percentile":{"p50":107887.59790089076,"p90":103377.87295193515,"p95":101190.89266549631,"p99":96287.51298153876},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":78.67799699306488,"p90":82.64400064945221,"p95":84.24700051546097,"p99":87.17100322246552},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":211.79699897766113,"p90":223.2149988412857,"p95":228.303000330925,"p99":239.23900723457336},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":290.474995970726,"p90":305.8589994907379,"p95":312.55000084638596,"p99":326.4100104570389},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":329.2129933834076,"p90":342.47300028800964,"p95":347.3210036754608,"p99":437.1350109577179},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":297.72698879241943,"p90":310.38498878479004,"p95":315.43299555778503,"p99":327.892005443573},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":97.94700145721436,"p90":103.83500158786774,"p95":105.99900037050247,"p99":110.92600226402283},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":219.32800114154816,"p90":229.14400696754456,"p95":232.02799260616302,"p99":238.51799964904785},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.209015846252441,"p90":6.248980760574341,"p95":6.248980760574341,"p99":6.248980760574341},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.4870078563690186,"p90":7.69197940826416,"p95":9.294003248214722,"p99":11.416018009185791},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.1619925498962402,"p90":2.4439990520477295,"p95":2.4439990520477295,"p99":2.4439990520477295},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":72.0679983496666,"p90":74.11099970340729,"p95":75.3529965877533,"p99":77.99699902534485},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":185.5980008840561,"p90":194.77200508117676,"p95":198.53800535202026,"p99":207.63200521469116},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":290.2750074863434,"p90":302.45301127433777,"p95":307.82100558280945,"p99":320.67999243736267},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.010012865066528,"p90":10.013997554779053,"p95":11.49699091911316,"p99":14.78201150894165},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"roundtrip":{"activation_data_bytes":8056832,"scale_bytes":0,"total_logical_bytes":8056832},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"roundtrip":{"activation_data_bytes":8056832,"scale_bytes":0,"total_logical_bytes":8056832},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":281,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":42,"mean":35.125,"min":26,"total":281},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.1458333283662796,"expert_assignments_per_rank":[55,56,58,37,46,43,46,43],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,1,9,24,24,6,0,0],"fanout_max":6,"fanout_mean":4.390625,"fanout_min":2,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.11743772029876709,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":281},"payload_copies_per_rank":[38,42,42,26,29,31,38,35],"payload_rank_cv":0.15879225730895996,"routed_copies":281},"token_rate_at_latency_percentile":{"p50":214962.03706484247,"p90":206195.53880672797,"p95":202895.70495574764,"p99":195186.2166124504},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":82.28299766778946,"p90":86.32999658584595,"p95":88.37199956178665,"p99":91.29700064659119},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":211.99800074100494,"p90":223.13499450683594,"p95":228.3419966697693,"p99":238.95800113677979},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":294.2809984087944,"p90":309.4649910926819,"p95":316.71399623155594,"p99":330.255001783371},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":330.255001783371,"p90":343.99500489234924,"p95":351.6469895839691,"p99":423.7549901008606},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":301.45201086997986,"p90":314.63098526000977,"p95":321.48200273513794,"p99":337.0650112628937},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":103.79499942064285,"p90":108.04200172424316,"p95":110.32599955797195,"p99":115.93300104141235},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":216.28400683403015,"p90":227.09999978542328,"p95":232.3479950428009,"p99":246.84999883174896},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.248980760574341,"p90":6.449997425079346,"p95":6.449997425079346,"p99":6.449997425079346},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.60699200630188,"p90":7.8909993171691895,"p95":9.175002574920654,"p99":10.776996612548828},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.004997968673706,"p90":2.6440024375915527,"p95":2.6440024375915527,"p99":2.6440024375915527},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":76.67399942874908,"p90":78.3580020070076,"p95":79.52000200748444,"p99":81.88199996948242},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":187.0799958705902,"p90":196.97600603103638,"p95":201.54200494289398,"p99":209.71499383449554},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":295.3220009803772,"p90":307.4609935283661,"p95":314.43101167678833,"p99":329.4140100479126},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.608975887298584,"p90":9.093999862670898,"p95":10.576993227005005,"p99":13.500005006790161},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"roundtrip":{"activation_data_bytes":16171008,"scale_bytes":0,"total_logical_bytes":16171008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"roundtrip":{"activation_data_bytes":16171008,"scale_bytes":0,"total_logical_bytes":16171008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":564,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":77,"mean":70.5,"min":62,"total":564},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.07384087145328522,"expert_assignments_per_rank":[101,108,101,91,83,97,94,93],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,2,14,52,50,10,0,0],"fanout_max":6,"fanout_mean":4.40625,"fanout_min":2,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.14716312289237976,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":564},"payload_copies_per_rank":[73,76,77,63,62,69,74,70],"payload_rank_cv":0.07472094893455505,"routed_copies":564},"token_rate_at_latency_percentile":{"p50":424611.53146929265,"p90":406825.7927432714,"p95":398156.0364530155,"p99":379748.7004670635},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":91.69799834489822,"p90":95.46300023794174,"p95":97.42599725723267,"p99":101.27200186252594},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":211.31600439548492,"p90":222.2529947757721,"p95":227.5809943675995,"p99":240.1600033044815},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":303.01400274038315,"p90":317.71599501371384,"p95":325.00699162483215,"p99":341.43200516700745},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":328.49299907684326,"p90":345.1569974422455,"p95":355.65298795700073,"p99":456.72500133514404},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":310.1849853992462,"p90":322.6830065250397,"p95":328.8530111312866,"p99":343.2750105857849},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":112.7690002322197,"p90":117.49599874019623,"p95":119.13800239562988,"p99":122.62299656867981},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":219.6889966726303,"p90":229.46399450302124,"p95":232.06700384616852,"p99":243.64599585533142},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.2499940395355225,"p90":6.449013948440552,"p95":6.449013948440552,"p99":6.449013948440552},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.006998777389526,"p90":8.693009614944458,"p95":10.374993085861206,"p99":13.13897967338562},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.841975212097168,"p90":4.607021808624268,"p95":4.607021808624268,"p99":4.607021808624268},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":86.40900254249573,"p90":87.69199997186661,"p95":88.89400213956833,"p99":90.89600294828415},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":196.4540034532547,"p90":206.38899505138397,"p95":211.5979939699173,"p99":220.770001411438},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":304.0960133075714,"p90":316.5139853954315,"p95":323.3250081539154,"p99":337.5459909439087},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.528002977371216,"p90":9.012997150421143,"p95":10.255008935928345,"p99":13.060003519058228},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"roundtrip":{"activation_data_bytes":32628736,"scale_bytes":0,"total_logical_bytes":32628736},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"roundtrip":{"activation_data_bytes":32628736,"scale_bytes":0,"total_logical_bytes":32628736},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1138,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":149,"mean":142.25,"min":126,"total":1138},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.04658475145697594,"expert_assignments_per_rank":[189,194,196,170,193,201,198,195],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,3,26,101,106,20,0,0],"fanout_max":6,"fanout_mean":4.4453125,"fanout_min":2,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.1344463974237442,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":1138},"payload_copies_per_rank":[143,141,149,126,144,145,148,142],"payload_rank_cv":0.04686218500137329,"routed_copies":1138},"token_rate_at_latency_percentile":{"p50":825313.964409001,"p90":793348.2545512815,"p95":778463.2992087709,"p99":745757.751381746},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":105.59800267219543,"p90":109.68499630689621,"p95":111.92700266838074,"p99":115.73299765586853},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":217.6869958639145,"p90":228.38300466537476,"p95":233.8709980249405,"p99":245.5279976129532},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":323.2849985361099,"p90":338.06800097227097,"p95":345.7980006933212,"p99":361.2609952688217},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":332.2179913520813,"p90":346.5990126132965,"p95":356.29400610923767,"p99":443.2240128517151},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":324.5669901371002,"p90":335.62299609184265,"p95":340.87100625038147,"p99":353.52998971939087},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":124.10599738359451,"p90":129.4340044260025,"p95":130.83599507808685,"p99":136.16399466991425},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":224.29600358009338,"p90":233.71100425720215,"p95":236.95500195026398,"p99":244.40699815750122},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.2490105628967285,"p90":6.28998875617981,"p95":6.28998875617981,"p99":6.28998875617981},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.807995796203613,"p90":10.255992412567139,"p95":12.097984552383423,"p99":14.261007308959961},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.7219781875610352,"p90":3.965020179748535,"p95":3.965020179748535,"p99":3.965020179748535},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":97.26600348949432,"p90":99.50900077819824,"p95":100.71100294589996,"p99":103.07399928569794},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":210.315003991127,"p90":220.00999748706818,"p95":225.33799707889557,"p99":235.55199801921844},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":316.15298986434937,"p90":327.129989862442,"p95":332.1380019187927,"p99":344.5959985256195},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.332997560501099,"p90":10.695993900299072,"p95":11.8979811668396,"p99":14.461010694503784},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.004672897048294544,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"roundtrip":{"activation_data_bytes":64569344,"scale_bytes":0,"total_logical_bytes":64569344},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"roundtrip":{"activation_data_bytes":64569344,"scale_bytes":0,"total_logical_bytes":64569344},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2252,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":290,"mean":281.5,"min":268,"total":2252},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03184119611978531,"expert_assignments_per_rank":[396,393,399,359,380,389,380,376],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,3,57,223,191,38,0,0],"fanout_max":6,"fanout_mean":4.3984375,"fanout_min":2,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.11767318099737167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":2252},"payload_copies_per_rank":[290,286,290,268,282,282,280,274],"payload_rank_cv":0.025306938216090202,"routed_copies":2252},"token_rate_at_latency_percentile":{"p50":1577486.3604697639,"p90":1525521.212676059,"p95":1502034.4664454048,"p99":1448250.54419398},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":122.38399684429169,"p90":126.75000727176666,"p95":128.4320056438446,"p99":130.9960037469864},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":235.19200086593628,"p90":245.60800194740295,"p95":249.89500641822815,"p99":259.227991104126},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":357.57599771022797,"p90":372.3580092191696,"p95":378.32701206207275,"p99":390.22399485111237},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":339.58899974823,"p90":350.4050076007843,"p95":360.17999053001404,"p99":393.6299979686737},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":349.8440086841583,"p90":360.17999053001404,"p95":364.7870123386383,"p99":374.80199337005615},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":123.50499629974365,"p90":132.55800306797028,"p95":135.4829967021942,"p99":144.29600536823273},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":223.13399612903595,"p90":233.3499938249588,"p95":236.07400059700012,"p99":244.7270005941391},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.28998875617981,"p90":6.3700079917907715,"p95":6.3700079917907715,"p99":6.3700079917907715},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.205971956253052,"p90":8.011996746063232,"p95":9.45398211479187,"p99":11.0970139503479},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.2420117855072021,"p90":2.164006233215332,"p95":2.164006233215332,"p99":2.164006233215332},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":116.93599820137024,"p90":118.53799968957901,"p95":119.61899697780609,"p99":121.82199954986572},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":229.3040007352829,"p90":238.87799680233002,"p95":243.04400384426117,"p99":251.97800993919373},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":343.71501207351685,"p90":353.769987821579,"p95":358.85798931121826,"p99":367.2710061073303},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.648016929626465,"p90":8.854001760482788,"p95":10.095000267028809,"p99":12.860000133514404},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"roundtrip":{"activation_data_bytes":128651264,"scale_bytes":0,"total_logical_bytes":128651264},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"roundtrip":{"activation_data_bytes":128651264,"scale_bytes":0,"total_logical_bytes":128651264},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":4487,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":569,"mean":560.875,"min":553,"total":4487},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.01615842990577221,"expert_assignments_per_rank":[775,756,779,747,762,761,782,782],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,5,126,437,385,71,0,0],"fanout_max":6,"fanout_mean":4.3818359375,"fanout_min":2,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.12859371304512024,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":4487},"payload_copies_per_rank":[569,555,565,556,553,564,558,567],"payload_rank_cv":0.010142241604626179,"routed_copies":4487},"token_rate_at_latency_percentile":{"p50":2927018.827195279,"p90":2843023.007727769,"p95":2807117.4832545915,"p99":2732109.268663804},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":171.81700468063354,"p90":175.62399804592133,"p95":177.30499804019928,"p99":180.91100454330444},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":281.7020118236542,"p90":290.31500220298767,"p95":294.2410111427307,"p99":304.29598689079285},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":453.5190165042877,"p90":465.939000248909,"p95":471.54600918293,"p99":485.2069914340973},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":437.05400824546814,"p90":442.26300716400146,"p95":444.54601407051086,"p99":466.05899930000305},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":446.58899307250977,"p90":454.92199063301086,"p95":458.48798751831055,"p99":465.81798791885376},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":161.28100454807281,"p90":163.8060063123703,"p95":164.76699709892273,"p99":167.1299934387207},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":268.9230144023895,"p90":273.3289897441864,"p95":274.6109962463379,"p99":280.8609902858734},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.328999996185303,"p90":6.370991468429565,"p95":6.370991468429565,"p99":6.370991468429565},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.607021808624268,"p90":8.69297981262207,"p95":10.215014219284058,"p99":13.220995664596558},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.0020136833190918,"p90":1.7220079898834229,"p95":1.7220079898834229,"p99":1.7220079898834229},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":161.5619957447052,"p90":164.28600251674652,"p95":165.8090054988861,"p99":168.8929945230484},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":274.4100093841553,"p90":282.8240096569061,"p95":287.11000084877014,"p99":298.0459928512573},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":435.932993888855,"p90":444.0250098705292,"p95":447.14999198913574,"p99":454.8420011997223},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.535985231399536,"p90":13.18100094795227,"p95":14.42098617553711,"p99":17.066001892089844},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"roundtrip":{"activation_data_bytes":259768320,"scale_bytes":0,"total_logical_bytes":259768320},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"roundtrip":{"activation_data_bytes":259768320,"scale_bytes":0,"total_logical_bytes":259768320},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":9060,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":1149,"mean":1132.5,"min":1088,"total":9060},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.018761977553367615,"expert_assignments_per_rank":[1542,1490,1556,1492,1545,1541,1581,1541],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,11,225,859,791,162,0,0],"fanout_max":6,"fanout_mean":4.423828125,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.12792494893074036,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":9060},"payload_copies_per_rank":[1144,1088,1148,1126,1135,1149,1140,1130],"payload_rank_cv":0.016329573467373848,"routed_copies":9060},"token_rate_at_latency_percentile":{"p50":4585872.092166588,"p90":4501870.743048202,"p95":4466856.397013475,"p99":4396567.013545138},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":250.2950131893158,"p90":254.1010081768036,"p95":255.82298636436462,"p99":258.98900628089905},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":374.60100650787354,"p90":382.93400406837463,"p95":386.69899106025696,"p99":396.15398645401},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":624.8960196971893,"p90":637.0350122451782,"p95":642.5219774246216,"p99":655.1429927349091},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":611.0349893569946,"p90":616.4429783821106,"p95":618.5669898986816,"p99":622.5730180740356},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":617.1249747276306,"p90":624.8570084571838,"p95":627.9410123825073,"p99":634.2700123786926},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":241.00099503993988,"p90":242.84400045871735,"p95":243.80500614643097,"p99":245.88799476623535},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":355.0119996070862,"p90":359.7390055656433,"p95":361.38200759887695,"p99":364.46601152420044},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.328999996185303,"p90":6.3700079917907715,"p95":6.3700079917907715,"p99":6.3700079917907715},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.0089826583862305,"p90":9.254038333892822,"p95":10.255992412567139,"p99":13.180017471313477},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.082003116607666,"p90":1.7230510711669922,"p95":1.7230510711669922,"p99":1.7230510711669922},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":241.40200018882751,"p90":243.40499937534332,"p95":245.12700736522675,"p99":248.69300425052643},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":358.4569990634918,"p90":366.5489852428436,"p95":369.8340058326721,"p99":380.5299997329712},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":607.8310012817383,"p90":615.3619885444641,"p95":618.3260083198547,"p99":625.1770257949829},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.014010429382324,"p90":12.37797737121582,"p95":13.859987258911133,"p99":16.865015029907227},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"roundtrip":{"activation_data_bytes":520626176,"scale_bytes":0,"total_logical_bytes":520626176},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"roundtrip":{"activation_data_bytes":520626176,"scale_bytes":0,"total_logical_bytes":520626176},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":18158,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":2294,"mean":2269.75,"min":2209,"total":18158},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.016587993130087852,"expert_assignments_per_rank":[3123,2960,3095,3022,3074,3092,3108,3102],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,23,454,1699,1566,354,0,0],"fanout_max":6,"fanout_mean":4.43310546875,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.1253442019224167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":18158},"payload_copies_per_rank":[2289,2209,2294,2272,2259,2292,2267,2276],"payload_rank_cv":0.01135234534740448,"routed_copies":18158},"token_rate_at_latency_percentile":{"p50":6637229.358296151,"p90":6555099.718115212,"p95":6522905.685773143,"p99":6457817.522601829},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":6,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":0,"num_qps_per_rank":1,"is_intranode":true,"cpu_proxy_threads":0},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi300x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9957,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2074,"p50":2108,"max":2109,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":248.2},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2077,"p50":2108,"max":2110,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":248.2},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2073,"p50":2101,"max":2103,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":248.2},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2081,"p50":2106,"max":2107,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":248.2},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2082,"p50":2109,"max":2110,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":248.2},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2110,"max":2111,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":248.2},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2073,"p50":2107,"max":2109,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":248.2},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2103,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":248.2}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-fp8.json b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-fp8.json new file mode 100644 index 0000000000..d8a6b359e4 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T06:59:47.216004+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":1,"phase":"decode","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":72.5490003824234,"p90":76.67499780654907,"p95":78.39799672365189,"p99":81.1220034956932},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":221.25199437141418,"p90":231.346994638443,"p95":235.99399626255035,"p99":250.014990568161},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":293.8009947538376,"p90":308.02199244499207,"p95":314.39199298620224,"p99":331.1369940638542},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":427.5209903717041,"p90":444.1849887371063,"p95":448.91300797462463,"p99":463.1739854812622},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":301.3310134410858,"p90":312.42701411247253,"p95":318.95700097084045,"p99":336.9449973106384},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.78700016438961,"p90":19.750000908970833,"p95":20.750999450683594,"p99":22.794000804424286},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":94.62200105190277,"p90":100.34999996423721,"p95":102.95400023460388,"p99":108.00199955701828},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":320.3999996185303,"p90":334.2210054397583,"p95":339.66898918151855,"p99":365.7880127429962},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.4900219440460205,"p90":6.76998496055603,"p95":6.76998496055603,"p99":6.76998496055603},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.616004467010498,"p90":19.91000771522522,"p95":22.914975881576538,"p99":31.567007303237915},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.044009208679199,"p90":2.6440024375915527,"p95":2.6440024375915527,"p99":2.6440024375915527},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":65.81900268793106,"p90":68.2620033621788,"p95":69.58399713039398,"p99":71.90799713134766},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":199.01899993419647,"p90":205.50699532032013,"p95":207.46999979019165,"p99":212.9189968109131},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":293.4800088405609,"p90":303.77501249313354,"p95":309.7040057182312,"p99":325.8480131626129},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.051020860671997,"p90":10.856002569198608,"p95":12.980014085769653,"p99":18.22701096534729},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8,"byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":250880,"scale_bytes":7840,"total_logical_bytes":258720},"roundtrip":{"activation_data_bytes":752640,"scale_bytes":7840,"total_logical_bytes":760480},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":250880,"scale_bytes":7840,"total_logical_bytes":258720},"roundtrip":{"activation_data_bytes":752640,"scale_bytes":7840,"total_logical_bytes":760480},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":35,"assignments":48,"wire":"rank-deduplicated"},"receive":{"max":6,"mean":4.375,"min":2,"total":35},"routing":{"empty_expert_count":337,"empty_rank_count":0,"expert_assignment_rank_cv":0.3227486312389374,"expert_assignments_per_rank":[8,6,6,6,6,5,9,2],"expert_load_cv":2.708012819290161,"expert_load_max":2,"expert_load_mean":0.125,"expert_load_min":0,"fanout_histogram":[0,0,0,5,3,0,0,0],"fanout_max":5,"fanout_mean":4.375,"fanout_min":4,"hotspot_ratio":16.0,"locality":{"placement":"packed","local_rank_fraction":0.08571428805589676,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":35},"payload_copies_per_rank":[5,5,5,4,4,4,6,2],"payload_rank_cv":0.2539483904838562,"routed_copies":35},"token_rate_at_latency_percentile":{"p50":26548.876959736193,"p90":25605.98040065777,"p95":25081.750755273035,"p99":23742.74752215594},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":76.1139988899231,"p90":80.48100024461746,"p95":82.12299644947052,"p99":84.40600335597992},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":221.61200642585754,"p90":231.0660034418106,"p95":236.79399490356445,"p99":249.89500641822815},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":297.72600531578064,"p90":311.54700368642807,"p95":318.916991353035,"p99":334.30100977420807},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":430.16499280929565,"p90":455.0420045852661,"p95":463.9750123023987,"p99":479.47901487350464},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":304.6160042285919,"p90":317.0349895954132,"p95":324.80698823928833,"p99":337.14500069618225},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.827000468969345,"p90":19.70900036394596,"p95":20.871000364422798,"p99":22.51400053501129},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":100.14999657869339,"p90":104.99700158834457,"p95":106.51999711990356,"p99":114.69200253486633},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":314.35099244117737,"p90":327.6509940624237,"p95":334.06001329421997,"p99":344.5970118045807},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.2490105628967285,"p90":6.450027227401733,"p95":6.450027227401733,"p99":6.450027227401733},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.132010698318481,"p90":19.870012998580933,"p95":23.474007844924927,"p99":33.84998440742493},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.805994987487793,"p90":5.366981029510498,"p95":5.366981029510498,"p99":5.366981029510498},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":67.38100200891495,"p90":70.18599659204483,"p95":72.10800051689148,"p99":74.91199672222137},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":199.09800589084625,"p90":205.66900074481964,"p95":208.3120048046112,"p99":213.6400043964386},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":295.522004365921,"p90":307.54101276397705,"p95":313.87099623680115,"p99":326.97001099586487},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.612006902694702,"p90":11.738002300262451,"p95":13.299018144607544,"p99":17.545998096466064},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":501760,"scale_bytes":15680,"total_logical_bytes":517440},"roundtrip":{"activation_data_bytes":1505280,"scale_bytes":15680,"total_logical_bytes":1520960},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":501760,"scale_bytes":15680,"total_logical_bytes":517440},"roundtrip":{"activation_data_bytes":1505280,"scale_bytes":15680,"total_logical_bytes":1520960},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":70,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":12,"mean":8.75,"min":5,"total":70},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.33592739701271057,"expert_assignments_per_rank":[18,17,13,7,13,7,13,8],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,3,6,5,2,0,0],"fanout_max":6,"fanout_mean":4.375,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.15714286267757416,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":70},"payload_copies_per_rank":[12,11,10,5,9,6,10,7],"payload_rank_cv":0.26649653911590576,"routed_copies":70},"token_rate_at_latency_percentile":{"p50":52525.14568470663,"p90":50467.61564210477,"p95":49260.023889057,"p99":47457.325385104494},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":77.79599726200104,"p90":82.04299956560135,"p95":83.36500078439713,"p99":85.84800362586975},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":222.2130000591278,"p90":232.54899680614471,"p95":238.4369969367981,"p99":255.8639943599701},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":300.00899732112885,"p90":314.59199637174606,"p95":321.8019977211952,"p99":341.71199798583984},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":429.3240010738373,"p90":445.82799077033997,"p95":454.72100377082825,"p99":550.104022026062},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":306.85898661613464,"p90":318.43799352645874,"p95":324.80600476264954,"p99":342.5939977169037},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.986999824643135,"p90":20.029999315738678,"p95":21.07200026512146,"p99":22.91399985551834},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":99.90999847650528,"p90":105.9190034866333,"p95":109.92500185966492,"p99":120.94199657440186},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":317.2360062599182,"p90":331.6569924354553,"p95":338.2669985294342,"p99":348.76298904418945},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.289005279541016,"p90":6.650000810623169,"p95":6.650000810623169,"p99":6.650000810623169},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.449984550476074,"p90":14.180988073348999,"p95":17.225980758666992,"p99":24.79800581932068},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.2430121898651123,"p90":5.407005548477173,"p95":5.407005548477173,"p99":5.407005548477173},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":70.58600336313248,"p90":73.10900092124939,"p95":74.39199835062027,"p99":76.67399942874908},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":199.37899708747864,"p90":206.38999342918396,"p95":208.87300372123718,"p99":216.60499274730682},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":299.5690107345581,"p90":310.34499406814575,"p95":316.43399596214294,"p99":333.4999978542328},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.050991058349609,"p90":10.494977235794067,"p95":12.378007173538208,"p99":16.54300093650818},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.00048828125,"host_held_samples":{"dispatch":0,"roundtrip":1},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":1025024,"scale_bytes":32032,"total_logical_bytes":1057056},"roundtrip":{"activation_data_bytes":3075072,"scale_bytes":32032,"total_logical_bytes":3107104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":1025024,"scale_bytes":32032,"total_logical_bytes":1057056},"roundtrip":{"activation_data_bytes":3075072,"scale_bytes":32032,"total_logical_bytes":3107104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":143,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":22,"mean":17.875,"min":13,"total":143},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.20518453419208527,"expert_assignments_per_rank":[31,30,24,17,22,22,28,18],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,4,12,13,3,0,0],"fanout_max":6,"fanout_mean":4.46875,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.1258741319179535,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":143},"payload_copies_per_rank":[22,21,18,13,16,15,22,16],"payload_rank_cv":0.1800614446401596,"routed_copies":143},"token_rate_at_latency_percentile":{"p50":104282.42742009186,"p90":100490.5213904419,"p95":98520.34608591627,"p99":93405.02230994316},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":78.1169980764389,"p90":82.40299671888351,"p95":84.04599875211716,"p99":87.69100159406662},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":221.2509959936142,"p90":231.18700087070465,"p95":236.71500384807587,"p99":249.21299517154694},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":299.3679940700531,"p90":313.58999758958817,"p95":320.761002600193,"p99":336.90399676561356},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":429.6039938926697,"p90":445.10701298713684,"p95":450.67501068115234,"p99":468.8220024108887},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":307.09999799728394,"p90":317.4760043621063,"p95":322.48300313949585,"p99":337.5059962272644},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":19.148999825119972,"p90":21.15200087428093,"p95":22.273000329732895,"p99":24.035999551415443},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":103.55500131845474,"p90":110.28499901294708,"p95":114.41099643707275,"p99":119.29900199174881},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":315.67299365997314,"p90":329.85401153564453,"p95":335.2630138397217,"p99":349.8849868774414},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.2490105628967285,"p90":6.488978862762451,"p95":6.488978862762451,"p99":6.488978862762451},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.727989435195923,"p90":9.053021669387817,"p95":10.776996612548828,"p99":15.86499810218811},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.5239884853363037,"p90":2.5640130043029785,"p95":2.5640130043029785,"p99":2.5640130043029785},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":71.86800241470337,"p90":73.99100065231323,"p95":75.39299875497818,"p99":78.47800105810165},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":198.85799288749695,"p90":205.62799274921417,"p95":208.23100209236145,"p99":214.1609936952591},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":300.00799894332886,"p90":310.5039894580841,"p95":315.15198945999146,"p99":329.7339975833893},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.408989429473877,"p90":9.894996881484985,"p95":11.216998100280762,"p99":13.78101110458374},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":2014208,"scale_bytes":62944,"total_logical_bytes":2077152},"roundtrip":{"activation_data_bytes":6042624,"scale_bytes":62944,"total_logical_bytes":6105568},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":2014208,"scale_bytes":62944,"total_logical_bytes":2077152},"roundtrip":{"activation_data_bytes":6042624,"scale_bytes":62944,"total_logical_bytes":6105568},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":281,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":42,"mean":35.125,"min":26,"total":281},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.1458333283662796,"expert_assignments_per_rank":[55,56,58,37,46,43,46,43],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,1,9,24,24,6,0,0],"fanout_max":6,"fanout_mean":4.390625,"fanout_min":2,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.11743772029876709,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":281},"payload_copies_per_rank":[38,42,42,26,29,31,38,35],"payload_rank_cv":0.15879225730895996,"routed_copies":281},"token_rate_at_latency_percentile":{"p50":208401.1736156574,"p90":201590.038682114,"p95":198460.0719322737,"p99":189626.26061584015},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":81.64200186729431,"p90":86.2089991569519,"p95":87.61099725961685,"p99":90.57600051164627},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":222.77399897575378,"p90":232.42799937725067,"p95":237.31499910354614,"p99":249.21399354934692},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":304.4160008430481,"p90":318.6369985342026,"p95":324.925996363163,"p99":339.7899940609932},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":427.0800054073334,"p90":441.6219890117645,"p95":449.0729868412018,"p99":485.6080114841461},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":311.7870092391968,"p90":323.3250081539154,"p95":328.57298851013184,"p99":338.42700719833374},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":22.714000195264816,"p90":23.794999346137047,"p95":24.796999990940094,"p99":26.921000331640244},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":107.28099942207336,"p90":115.17299711704254,"p95":119.37899887561798,"p99":125.98900496959686},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":310.6650114059448,"p90":322.6430118083954,"p95":330.7349979877472,"p99":343.0339992046356},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.289005279541016,"p90":6.408989429473877,"p95":6.408989429473877,"p99":6.408989429473877},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.526019096374512,"p90":8.45298171043396,"p95":9.615004062652588,"p99":13.419002294540405},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.2050013542175293,"p90":4.767000675201416,"p95":4.767000675201416,"p99":4.767000675201416},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":76.63500308990479,"p90":78.79800349473953,"p95":80.23999631404877,"p99":82.36400038003922},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":199.93999600410461,"p90":206.94999396800995,"p95":209.43400263786316,"p99":215.48299491405487},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":306.01799488067627,"p90":317.15500354766846,"p95":322.6430118083954,"p99":332.4170112609863},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.086988210678101,"p90":8.733004331588745,"p95":10.17501950263977,"p99":13.099998235702515},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":4042752,"scale_bytes":126336,"total_logical_bytes":4169088},"roundtrip":{"activation_data_bytes":12128256,"scale_bytes":126336,"total_logical_bytes":12254592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":4042752,"scale_bytes":126336,"total_logical_bytes":4169088},"roundtrip":{"activation_data_bytes":12128256,"scale_bytes":126336,"total_logical_bytes":12254592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":564,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":77,"mean":70.5,"min":62,"total":564},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.07384087145328522,"expert_assignments_per_rank":[101,108,101,91,83,97,94,93],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,2,14,52,50,10,0,0],"fanout_max":6,"fanout_mean":4.40625,"fanout_min":2,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.14716312289237976,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":564},"payload_copies_per_rank":[73,76,77,63,62,69,74,70],"payload_rank_cv":0.07472094893455505,"routed_copies":564},"token_rate_at_latency_percentile":{"p50":410536.6683247568,"p90":395886.48193605547,"p95":389563.36788485886,"p99":378220.40581113},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":91.0160019993782,"p90":95.18200159072876,"p95":97.06500172615051,"p99":100.6309986114502},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":223.0139970779419,"p90":233.6300015449524,"p95":238.55799436569214,"p99":249.37400221824646},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":314.0299990773201,"p90":328.81200313568115,"p95":335.62299609184265,"p99":350.00500082969666},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":432.6080083847046,"p90":449.2729902267456,"p95":456.24300837516785,"p99":564.005970954895},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":321.1219906806946,"p90":331.7379951477051,"p95":337.3059928417206,"p99":354.4510006904602},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":26.079000905156136,"p90":28.28199975192547,"p95":29.16399948298931,"p99":31.045999377965927},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":116.73499643802643,"p90":123.66499751806259,"p95":126.91099941730499,"p99":133.35900008678436},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":303.1739890575409,"p90":317.67600774765015,"p95":321.20099663734436,"p99":335.82401275634766},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.2490105628967285,"p90":6.569981575012207,"p95":6.569981575012207,"p99":6.569981575012207},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.647987127304077,"p90":9.373992681503296,"p95":10.73598861694336,"p99":13.219982385635376},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.2439956665039062,"p90":3.044992685317993,"p95":3.044992685317993,"p99":3.044992685317993},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":86.28900349140167,"p90":87.65199780464172,"p95":88.85300159454346,"p99":91.09699726104736},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":200.05899667739868,"p90":209.15299654006958,"p95":213.3989930152893,"p99":222.89399802684784},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":315.5530095100403,"p90":325.6079852581024,"p95":331.57700300216675,"p99":348.3220040798187},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.848003387451172,"p90":8.492976427078247,"p95":10.095000267028809,"p99":12.620002031326294},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":8157184,"scale_bytes":254912,"total_logical_bytes":8412096},"roundtrip":{"activation_data_bytes":24471552,"scale_bytes":254912,"total_logical_bytes":24726464},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":8157184,"scale_bytes":254912,"total_logical_bytes":8412096},"roundtrip":{"activation_data_bytes":24471552,"scale_bytes":254912,"total_logical_bytes":24726464},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1138,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":149,"mean":142.25,"min":126,"total":1138},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.04658475145697594,"expert_assignments_per_rank":[189,194,196,170,193,201,198,195],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,3,26,101,106,20,0,0],"fanout_max":6,"fanout_mean":4.4453125,"fanout_min":2,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.1344463974237442,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":1138},"payload_copies_per_rank":[143,141,149,126,144,145,148,142],"payload_rank_cv":0.04686218500137329,"routed_copies":1138},"token_rate_at_latency_percentile":{"p50":797204.8238033996,"p90":771693.3355373327,"p95":758954.7930745687,"p99":722243.6937723958},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":105.23799806833267,"p90":109.76400226354599,"p95":112.12799698114395,"p99":116.17399752140045},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":222.45299816131592,"p90":232.42799937725067,"p95":237.31599748134613,"p99":250.61601400375366},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":327.6909962296486,"p90":342.19200164079666,"p95":349.4439944624901,"p99":366.7900115251541},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":430.8460056781769,"p90":445.7069933414459,"p95":459.76901054382324,"p99":582.431972026825},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":334.90198850631714,"p90":345.19699215888977,"p95":351.4060080051422,"p99":363.90599608421326},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":34.651998430490494,"p90":39.379000663757324,"p95":39.980001747608185,"p99":40.98200052976608},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":125.34800171852112,"p90":130.79600036144257,"p95":134.16099548339844,"p99":138.36699724197388},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":298.44701290130615,"p90":309.58399176597595,"p95":314.5110011100769,"p99":327.77100801467896},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.248980760574341,"p90":6.3700079917907715,"p95":6.3700079917907715,"p99":6.3700079917907715},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.688982248306274,"p90":10.536015033721924,"p95":11.898010969161987,"p99":14.622986316680908},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.804011106491089,"p90":3.3649802207946777,"p95":3.3649802207946777,"p99":3.3649802207946777},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":97.78700023889542,"p90":99.90900009870529,"p95":101.27200186252594,"p99":103.3950001001358},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":206.46899938583374,"p90":214.84200656414032,"p95":219.9299931526184,"p99":229.8240065574646},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":326.73001289367676,"p90":337.024986743927,"p95":342.67300367355347,"p99":356.89499974250793},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.771998643875122,"p90":10.856002569198608,"p95":12.338995933532715,"p99":15.141993761062622},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":16142336,"scale_bytes":504448,"total_logical_bytes":16646784},"roundtrip":{"activation_data_bytes":48427008,"scale_bytes":504448,"total_logical_bytes":48931456},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":16142336,"scale_bytes":504448,"total_logical_bytes":16646784},"roundtrip":{"activation_data_bytes":48427008,"scale_bytes":504448,"total_logical_bytes":48931456},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2252,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":290,"mean":281.5,"min":268,"total":2252},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03184119611978531,"expert_assignments_per_rank":[396,393,399,359,380,389,380,376],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,3,57,223,191,38,0,0],"fanout_max":6,"fanout_mean":4.3984375,"fanout_min":2,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.11767318099737167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":2252},"payload_copies_per_rank":[290,286,290,268,282,282,280,274],"payload_rank_cv":0.025306938216090202,"routed_copies":2252},"token_rate_at_latency_percentile":{"p50":1528805.4940597713,"p90":1483211.0697080842,"p95":1457004.115855947,"p99":1406956.7567155876},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":122.02300131320953,"p90":126.34900212287903,"p95":127.71199643611908,"p99":130.4759979248047},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":226.33999586105347,"p90":236.43499612808228,"p95":241.84200167655945,"p99":253.45999002456665},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":348.362997174263,"p90":362.7839982509613,"p95":369.5539981126785,"p99":383.93598794937134},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":432.24701285362244,"p90":451.7970085144043,"p95":462.0929956436157,"p99":563.7249946594238},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":351.7270088195801,"p90":363.22399973869324,"p95":367.75100231170654,"p99":381.5320134162903},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":44.54699903726578,"p90":46.390000730752945,"p95":47.35099896788597,"p99":49.11300167441368},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":146.6190069913864,"p90":152.42800116539001,"p95":155.7130068540573,"p99":160.67999601364136},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":295.72299122810364,"p90":306.4990043640137,"p95":310.2250099182129,"p99":325.80798864364624},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.25002384185791,"p90":6.529003381729126,"p95":6.529003381729126,"p99":6.529003381729126},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.36899471282959,"p90":10.095000267028809,"p95":11.49800419807434,"p99":14.301002025604248},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.1249821186065674,"p90":3.0839741230010986,"p95":3.0839741230010986,"p99":3.0839741230010986},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":117.01600253582001,"p90":118.89799684286118,"p95":120.02000212669373,"p99":122.1029981970787},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":220.56999802589417,"p90":229.3030023574829,"p95":234.3509942293167,"p99":242.32299625873566},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":345.9179997444153,"p90":356.5340042114258,"p95":361.2610101699829,"p99":374.00001287460327},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.448013544082642,"p90":8.772999048233032,"p95":10.055005550384521,"p99":12.058019638061523},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.006756756920367479,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":32162816,"scale_bytes":1005088,"total_logical_bytes":33167904},"roundtrip":{"activation_data_bytes":96488448,"scale_bytes":1005088,"total_logical_bytes":97493536},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":32162816,"scale_bytes":1005088,"total_logical_bytes":33167904},"roundtrip":{"activation_data_bytes":96488448,"scale_bytes":1005088,"total_logical_bytes":97493536},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":4487,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":569,"mean":560.875,"min":553,"total":4487},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.01615842990577221,"expert_assignments_per_rank":[775,756,779,747,762,761,782,782],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,5,126,437,385,71,0,0],"fanout_max":6,"fanout_mean":4.3818359375,"fanout_min":2,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.12859371304512024,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":4487},"payload_copies_per_rank":[569,555,565,556,553,564,558,567],"payload_rank_cv":0.010142241604626179,"routed_copies":4487},"token_rate_at_latency_percentile":{"p50":2911348.785629554,"p90":2819196.9713914148,"p95":2784492.750701072,"p99":2683916.326787267},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":171.97799682617188,"p90":175.9839951992035,"p95":177.7060031890869,"p99":180.59000372886658},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":263.7549936771393,"p90":271.9680070877075,"p95":276.6149938106537,"p99":286.2290143966675},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":435.73299050331116,"p90":447.952002286911,"p95":454.3209969997406,"p99":466.81901812553406},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":421.79200053215027,"p90":444.78699564933777,"p95":457.5650095939636,"p99":473.34998846054077},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":427.8010129928589,"p90":435.932993888855,"p95":440.3800070285797,"p99":449.3130147457123},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":62.373001128435135,"p90":64.7369995713234,"p95":65.65800309181213,"p99":67.70099699497223},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":169.0939962863922,"p90":177.2249937057495,"p95":179.748997092247,"p99":189.6049976348877},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":275.49299597740173,"p90":287.07098960876465,"p95":291.83700680732727,"p99":303.8559854030609},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.28998875617981,"p90":6.569981575012207,"p95":6.569981575012207,"p99":6.569981575012207},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.847019910812378,"p90":8.533000946044922,"p95":10.334998369216919,"p99":13.459980487823486},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.7219781875610352,"p90":7.651001214981079,"p95":7.651001214981079,"p99":7.651001214981079},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":161.40200197696686,"p90":164.0859991312027,"p95":165.4070019721985,"p99":168.21199655532837},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":257.8260004520416,"p90":265.03700017929077,"p95":268.76300573349,"p99":278.0170142650604},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":416.9049859046936,"p90":424.35601353645325,"p95":429.0030002593994,"p99":438.81699442863464},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.616004467010498,"p90":13.4199857711792,"p95":14.82301950454712,"p99":17.58599281311035},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.006756756920367479,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":64942080,"scale_bytes":2029440,"total_logical_bytes":66971520},"roundtrip":{"activation_data_bytes":194826240,"scale_bytes":2029440,"total_logical_bytes":196855680},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":64942080,"scale_bytes":2029440,"total_logical_bytes":66971520},"roundtrip":{"activation_data_bytes":194826240,"scale_bytes":2029440,"total_logical_bytes":196855680},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":9060,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":1149,"mean":1132.5,"min":1088,"total":9060},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.018761977553367615,"expert_assignments_per_rank":[1542,1490,1556,1492,1545,1541,1581,1541],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,11,225,859,791,162,0,0],"fanout_max":6,"fanout_mean":4.423828125,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.12792494893074036,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":9060},"payload_copies_per_rank":[1144,1088,1148,1126,1135,1149,1140,1130],"payload_rank_cv":0.016329573467373848,"routed_copies":9060},"token_rate_at_latency_percentile":{"p50":4787272.441625066,"p90":4697969.7079825895,"p95":4650529.014290809,"p99":4558069.614696252},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":249.2929995059967,"p90":253.65900993347168,"p95":254.98199462890625,"p99":257.8260004520416},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":328.57298851013184,"p90":335.7830047607422,"p95":338.8670086860657,"p99":347.55998849868774},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":577.8659880161285,"p90":589.4420146942139,"p95":593.8490033149719,"p99":605.3859889507294},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":556.5140247344971,"p90":562.362015247345,"p95":564.8459792137146,"p99":569.1720247268677},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":566.2490129470825,"p90":574.5810270309448,"p95":577.9460072517395,"p99":587.9610180854797},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":101.27200186252594,"p90":103.79599779844284,"p95":104.63699698448181,"p99":106.64000362157822},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":240.1999980211258,"p90":243.4850037097931,"p95":244.5259988307953,"p99":247.45100736618042},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":307.82100558280945,"p90":312.7880096435547,"p95":315.71298837661743,"p99":324.4459927082062},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.328999996185303,"p90":6.488978862762451,"p95":6.488978862762451,"p99":6.488978862762451},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.769014358520508,"p90":8.733034133911133,"p95":9.775996208190918,"p99":11.81703805923462},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.48100948333740234,"p90":1.5619993209838867,"p95":1.5619993209838867,"p99":1.5619993209838867},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":240.60000479221344,"p90":243.28500032424927,"p95":244.84699964523315,"p99":247.89099395275116},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":316.3140118122101,"p90":323.96501302719116,"p95":327.77100801467896,"p99":335.38299798965454},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":557.4749708175659,"p90":565.2070045471191,"p95":568.5319900512695,"p99":579.4280171394348},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.534014225006104,"p90":11.696994304656982,"p95":13.218998908996582,"p99":16.224980354309082},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.006756756920367479,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":130156544,"scale_bytes":4067392,"total_logical_bytes":134223936},"roundtrip":{"activation_data_bytes":390469632,"scale_bytes":4067392,"total_logical_bytes":394537024},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":130156544,"scale_bytes":4067392,"total_logical_bytes":134223936},"roundtrip":{"activation_data_bytes":390469632,"scale_bytes":4067392,"total_logical_bytes":394537024},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":18158,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":2294,"mean":2269.75,"min":2209,"total":18158},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.016587993130087852,"expert_assignments_per_rank":[3123,2960,3095,3022,3074,3092,3108,3102],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,23,454,1699,1566,354,0,0],"fanout_max":6,"fanout_mean":4.43310546875,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.1253442019224167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":18158},"payload_copies_per_rank":[2289,2209,2294,2272,2259,2292,2267,2276],"payload_rank_cv":0.01135234534740448,"routed_copies":18158},"token_rate_at_latency_percentile":{"p50":7233566.693003281,"p90":7128672.558447366,"p95":7087167.224283428,"p99":6966448.240628956},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":6,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":0,"num_qps_per_rank":1,"is_intranode":true,"cpu_proxy_threads":0},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi300x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9962,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2086,"p50":2108,"max":2109,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":367.5},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2108,"max":2109,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":367.5},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2079,"p50":2102,"max":2103,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":367.5},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2106,"max":2107,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":367.5},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2109,"max":2110,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":367.5},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2110,"max":2111,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":367.5},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2089,"p50":2108,"max":2110,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":367.5},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2082,"p50":2104,"max":2105,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":367.5}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16.json b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16.json new file mode 100644 index 0000000000..711dd5eb78 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T10:58:35.008857+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":2,"phase":"prefill","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":816.8650269508362,"p90":831.246018409729,"p95":836.9349837303162,"p99":846.7900156974792},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1327.6699781417847,"p90":1345.0570106506348,"p95":1350.0640392303467,"p99":1359.639048576355},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":2144.535005092621,"p90":2176.3030290603638,"p95":2186.999022960663,"p99":2206.4290642738342},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":2227.6899814605713,"p90":2258.9468955993652,"p95":2268.8019275665283,"p99":2288.501024246216},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2142.4129009246826,"p90":2161.0000133514404,"p95":2165.2069091796875,"p99":2178.705930709839},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":837.4149799346924,"p90":848.6679792404175,"p95":851.8329858779907,"p99":861.9289994239807},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1305.9519529342651,"p90":1317.8499937057495,"p95":1320.0589418411255,"p99":1328.3050060272217},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.450176239013672,"p90":6.649971008300781,"p95":6.649971008300781,"p99":6.649971008300781},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":52.58893966674805,"p90":74.56302642822266,"p95":83.84513854980469,"p99":96.6651439666748},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.928993225097656,"p90":6.289005279541016,"p95":6.289005279541016,"p99":6.289005279541016},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":755.3690075874329,"p90":763.221025466919,"p95":765.8650279045105,"p99":770.6320285797119},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1285.7619524002075,"p90":1294.854998588562,"p95":1297.3790168762207,"p99":1302.5130033493042},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2080.63006401062,"p90":2094.810962677002,"p95":2099.3781089782715,"p99":2109.231948852539},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":61.181068420410156,"p90":76.6448974609375,"p95":81.8929672241211,"p99":90.8651351928711},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0095911156386137,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"roundtrip":{"activation_data_bytes":2428403712,"scale_bytes":0,"total_logical_bytes":2428403712},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"roundtrip":{"activation_data_bytes":2428403712,"scale_bytes":0,"total_logical_bytes":2428403712},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":84696,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":5367,"mean":5293.5,"min":5181,"total":84696},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.012361490167677402,"expert_assignments_per_rank":[6171,6137,6243,6050,6268,6146,6227,6063,6176,6175,6160,5993,6212,6144,6110,6029],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,5,241,2597,7671,5870,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16943359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.06316709518432617,"same_node_fraction":0.49927976727485657,"same_scaleup_domain_fraction":0.49927976727485657,"cross_node_fraction":0.500720202922821,"cross_domain_fraction":0.500720202922821,"gpus_per_node":8,"scale_up_domain":8,"copies":84696},"payload_copies_per_rank":[5326,5273,5366,5236,5367,5310,5348,5225,5315,5334,5305,5181,5318,5287,5280,5225],"payload_rank_cv":0.009882226586341858,"routed_copies":84696},"token_rate_at_latency_percentile":{"p50":7647452.082149306,"p90":7581675.103551003,"p95":7566944.263173102,"p99":7520060.3115180265},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1326.0619640350342,"p90":1348.9760160446167,"p95":1357.314944267273,"p99":1379.4610500335693},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2234.4300746917725,"p90":2255.2618980407715,"p95":2261.1498832702637,"p99":2271.646022796631},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":3560.4920387268066,"p90":3604.237914085388,"p95":3618.4648275375366,"p99":3651.1070728302},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":3645.970106124878,"p90":3682.921886444092,"p95":3699.9709606170654,"p99":3720.659017562866},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3544.5549488067627,"p90":3572.3800659179688,"p95":3582.9319953918457,"p99":3611.495018005371},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1335.8429670333862,"p90":1354.4310331344604,"p95":1359.5930337905884,"p99":1367.0899868011475},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2198.085069656372,"p90":2222.522974014282,"p95":2231.775999069214,"p99":2244.035005569458},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.450176239013672,"p90":6.6089630126953125,"p95":6.6089630126953125,"p99":6.6089630126953125},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":86.66491508483887,"p90":122.98393249511719,"p95":134.0010166168213,"p99":156.43310546875},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.894132614135742,"p90":23.275136947631836,"p95":23.275136947631836,"p99":23.275136947631836},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1239.292025566101,"p90":1252.6320219039917,"p95":1255.916953086853,"p99":1261.3259553909302},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2168.642044067383,"p90":2178.096055984497,"p95":2180.938959121704,"p99":2187.7899169921875},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3458.609104156494,"p90":3481.6839694976807,"p95":3489.0949726104736,"p99":3501.7940998077393},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":85.98494529724121,"p90":111.38296127319336,"p95":120.49984931945801,"p99":141.53099060058594},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0095911156386137,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"roundtrip":{"activation_data_bytes":4850069504,"scale_bytes":0,"total_logical_bytes":4850069504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"roundtrip":{"activation_data_bytes":4850069504,"scale_bytes":0,"total_logical_bytes":4850069504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":169157,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":10735,"mean":10572.3125,"min":10405,"total":169157},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.008082338608801365,"expert_assignments_per_rank":[12320,12304,12326,12170,12384,12230,12308,12211,12262,12359,12237,12047,12413,12243,12482,12312],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,16,505,5263,15346,11638,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.162261962890625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.0630006417632103,"same_node_fraction":0.5010611414909363,"same_scaleup_domain_fraction":0.5010611414909363,"cross_node_fraction":0.4989388585090637,"cross_domain_fraction":0.4989388585090637,"gpus_per_node":8,"scale_up_domain":8,"copies":169157},"payload_copies_per_rank":[10630,10585,10602,10505,10644,10564,10577,10498,10504,10645,10499,10405,10642,10518,10735,10604],"payload_rank_cv":0.00741199404001236,"routed_copies":169157},"token_rate_at_latency_percentile":{"p50":9244602.065213013,"p90":9172596.251059823,"p95":9145582.456531202,"p99":9073250.783022752},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2396.0320949554443,"p90":2430.8440685272217,"p95":2445.626974105835,"p99":2500.067949295044},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4047.5687980651855,"p90":4073.3680725097656,"p95":4081.6612243652344,"p99":4097.76496887207},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":6443.60089302063,"p90":6504.212141036987,"p95":6527.288198471069,"p99":6597.832918167114},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":6524.433135986328,"p90":6601.988792419434,"p95":6623.490810394287,"p99":6684.903144836426},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6430.120944976807,"p90":6473.185062408447,"p95":6488.698959350586,"p99":6524.902820587158},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2314.340114593506,"p90":2341.339111328125,"p95":2347.9089736938477,"p99":2360.848903656006},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":3990.8230304718018,"p90":4021.9497680664062,"p95":4029.1829109191895,"p99":4047.57022857666},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.609916687011719,"p90":7.010936737060547,"p95":7.010936737060547,"p99":7.010936737060547},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":139.7991180419922,"p90":194.6120262145996,"p95":213.59968185424805,"p99":284.06381607055664},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":15.623092651367188,"p90":25.397777557373047,"p95":25.397777557373047,"p99":25.397777557373047},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2249.682903289795,"p90":2268.589973449707,"p95":2274.0890979766846,"p99":2286.1359119415283},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3944.4739818573,"p90":3957.4129581451416,"p95":3961.940050125122,"p99":3971.034049987793},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6277.462005615234,"p90":6310.260772705078,"p95":6320.806980133057,"p99":6341.639041900635},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":151.13592147827148,"p90":193.6798095703125,"p95":206.70795440673828,"p99":238.54780197143555},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"roundtrip":{"activation_data_bytes":9698332672,"scale_bytes":0,"total_logical_bytes":9698332672},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"roundtrip":{"activation_data_bytes":9698332672,"scale_bytes":0,"total_logical_bytes":9698332672},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":338251,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":21435,"mean":21140.6875,"min":20921,"total":338251},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.0070151351392269135,"expert_assignments_per_rank":[24582,24524,24711,24358,24680,24435,24611,24307,24504,24638,24478,24415,25004,24532,24846,24591],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,29,1048,10572,30561,23326,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1613006591796875,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.06308037787675858,"same_node_fraction":0.5009504556655884,"same_scaleup_domain_fraction":0.5009504556655884,"cross_node_fraction":0.49904951453208923,"cross_domain_fraction":0.49904951453208923,"gpus_per_node":8,"scale_up_domain":8,"copies":338251},"payload_copies_per_rank":[21145,21167,21213,20988,21198,21055,21170,20921,21054,21231,21048,21005,21435,21070,21396,21155],"payload_rank_cv":0.00635375315323472,"routed_copies":338251},"token_rate_at_latency_percentile":{"p50":10192032.243374296,"p90":10124227.774760442,"p95":10100021.654658347,"p99":10043981.006617136},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4591.944217681885,"p90":4664.812088012695,"p95":4708.316802978516,"p99":4821.846961975098},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7650.030136108398,"p90":7696.098804473877,"p95":7712.362766265869,"p99":7740.005016326904},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":12241.974353790283,"p90":12360.910892486572,"p95":12420.679569244385,"p99":12561.851978302002},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":12339.381217956543,"p90":12424.132347106934,"p95":12460.322380065918,"p99":12670.539855957031},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":12220.64208984375,"p90":12311.4013671875,"p95":12353.143692016602,"p99":12505.81169128418},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":4403.302192687988,"p90":4445.324897766113,"p95":4457.783222198486,"p99":4490.030765533447},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":7600.6760597229,"p90":7650.997161865234,"p95":7663.455009460449,"p99":7689.174175262451},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.409645080566406,"p90":7.091522216796875,"p95":7.091522216796875,"p99":7.091522216796875},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":232.22923278808594,"p90":356.2765121459961,"p95":409.35707092285156,"p99":522.1662521362305},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-18.026351928710938,"p90":51.23615264892578,"p95":51.23615264892578,"p99":51.23615264892578},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4340.928077697754,"p90":4366.686820983887,"p95":4374.878883361816,"p99":4389.260768890381},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7513.7457847595215,"p90":7534.056186676025,"p95":7539.308071136475,"p99":7550.9209632873535},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11929.308891296387,"p90":11970.170021057129,"p95":11980.266571044922,"p99":12008.564949035645},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":290.1945114135742,"p90":386.37638092041016,"p95":431.90956115722656,"p99":578.6647796630859},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":131072,"byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"roundtrip":{"activation_data_bytes":19408105472,"scale_bytes":0,"total_logical_bytes":19408105472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"roundtrip":{"activation_data_bytes":19408105472,"scale_bytes":0,"total_logical_bytes":19408105472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":676901,"assignments":786432,"wire":"rank-deduplicated"},"receive":{"max":42692,"mean":42306.3125,"min":41888,"total":676901},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005902660544961691,"expert_assignments_per_rank":[49152,49158,48937,48637,49532,49147,49291,48640,48855,49080,49131,49180,49736,49110,49535,49311],"expert_load_cv":0.02157296985387802,"expert_load_max":2171,"expert_load_mean":2048.0,"expert_load_min":1922,"fanout_histogram":[0,51,2031,21028,61178,46784,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.164344787597656,"fanout_min":2,"hotspot_ratio":1.06005859375,"locality":{"placement":"packed","local_rank_fraction":0.06220702826976776,"same_node_fraction":0.5000066757202148,"same_scaleup_domain_fraction":0.5000066757202148,"cross_node_fraction":0.49999335408210754,"cross_domain_fraction":0.49999335408210754,"gpus_per_node":8,"scale_up_domain":8,"copies":676901},"payload_copies_per_rank":[42365,42433,42070,41888,42530,42296,42419,41908,42030,42386,42324,42267,42692,42172,42625,42496],"payload_rank_cv":0.005482310429215431,"routed_copies":676901},"token_rate_at_latency_percentile":{"p50":10725459.352821605,"p90":10646391.591889348,"p95":10610416.527795041,"p99":10480887.065599231},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":36,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":20,"num_max_rdma_chunked_recv_tokens":140},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":12,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":1073741824,"num_qps_per_rank":10,"is_intranode":false,"cpu_proxy_threads":4},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9938,"throttle_fraction_max":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2098,"p50":2099,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2099,"p50":2101,"max":2102,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2098,"p50":2099,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2100,"p50":2101,"max":2103,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2093,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2102,"p50":2105,"max":2106,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2105,"p50":2106,"max":2108,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2095,"p50":2097,"max":2098,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2101,"p50":2103,"max":2104,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2101,"p50":2102,"max":2104,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2103,"p50":2104,"max":2106,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2096,"p50":2098,"max":2100,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2100,"p50":2103,"max":2104,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2099,"p50":2100,"max":2102,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2098,"p50":2100,"max":2102,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2098,"p50":2100,"max":2102,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":830.4}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8.json b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8.json new file mode 100644 index 0000000000..42cbd7767b --- /dev/null +++ b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T11:30:05.825069+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":2,"phase":"prefill","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":812.4189972877502,"p90":828.482985496521,"p95":834.9729776382446,"p99":843.1059718132019},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1119.6800470352173,"p90":1130.5359601974487,"p95":1135.4639530181885,"p99":1149.1249799728394},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1932.0990443229675,"p90":1959.0189456939697,"p95":1970.436930656433,"p99":1992.2309517860413},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":2038.7389659881592,"p90":2056.955099105835,"p95":2061.8019104003906,"p99":2077.346086502075},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1936.1850023269653,"p90":1954.77294921875,"p95":1961.3829851150513,"p99":1973.0809926986694},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":218.92699599266052,"p90":222.21200168132782,"p95":223.37499260902405,"p99":225.89799761772156},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":857.7669858932495,"p90":869.1040277481079,"p95":872.4240064620972,"p99":881.4380168914795},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1119.4339990615845,"p90":1132.2189569473267,"p95":1136.3450288772583,"p99":1149.07705783844},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.448984146118164,"p90":6.489992141723633,"p95":6.489992141723633,"p99":6.489992141723633},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":49.18408393859863,"p90":68.78304481506348,"p95":76.82299613952637,"p99":88.48094940185547},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.927158355712891,"p90":7.893085479736328,"p95":7.893085479736328,"p99":7.893085479736328},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":747.838020324707,"p90":758.9340209960938,"p95":762.0589733123779,"p99":766.5460109710693},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1084.7020149230957,"p90":1093.3539867401123,"p95":1096.439003944397,"p99":1104.4909954071045},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1871.2780475616455,"p90":1885.8979940414429,"p95":1890.5459642410278,"p99":1900.6110429763794},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":64.628005027771,"p90":82.09407329559326,"p95":87.9819393157959,"p99":99.92098808288574},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.005859375,"host_held_samples":{"dispatch":7,"roundtrip":12},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.009900989942252636,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":607100928,"scale_bytes":18971904,"total_logical_bytes":626072832},"roundtrip":{"activation_data_bytes":1821302784,"scale_bytes":18971904,"total_logical_bytes":1840274688},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":607100928,"scale_bytes":18971904,"total_logical_bytes":626072832},"roundtrip":{"activation_data_bytes":1821302784,"scale_bytes":18971904,"total_logical_bytes":1840274688},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":84696,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":5367,"mean":5293.5,"min":5181,"total":84696},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.012361490167677402,"expert_assignments_per_rank":[6171,6137,6243,6050,6268,6146,6227,6063,6176,6175,6160,5993,6212,6144,6110,6029],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,5,241,2597,7671,5870,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16943359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.06316709518432617,"same_node_fraction":0.49927976727485657,"same_scaleup_domain_fraction":0.49927976727485657,"cross_node_fraction":0.500720202922821,"cross_domain_fraction":0.500720202922821,"gpus_per_node":8,"scale_up_domain":8,"copies":84696},"payload_copies_per_rank":[5326,5273,5366,5236,5367,5310,5348,5225,5315,5334,5305,5181,5318,5287,5280,5225],"payload_rank_cv":0.009882226586341858,"routed_copies":84696},"token_rate_at_latency_percentile":{"p50":8462001.29652343,"p90":8381536.0789334,"p95":8353289.553513152,"p99":8303764.5492651},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1332.4799537658691,"p90":1356.7960262298584,"p95":1364.7600412368774,"p99":1377.9079914093018},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1820.7720518112183,"p90":1832.669973373413,"p95":1836.637020111084,"p99":1845.3689813613892},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":3153.2520055770874,"p90":3189.4659996032715,"p95":3201.3970613479614,"p99":3223.276972770691},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":3265.058994293213,"p90":3291.137933731079,"p95":3299.791097640991,"p99":3315.3340816497803},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3157.360076904297,"p90":3183.617115020752,"p95":3192.2121047973633,"p99":3213.7820720672607},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":452.19600200653076,"p90":457.8849971294403,"p95":459.9680006504059,"p99":463.61300349235535},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1351.4599800109863,"p90":1368.7340021133423,"p95":1372.418999671936,"p99":1378.7809610366821},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1821.763038635254,"p90":1835.513949394226,"p95":1840.9520387649536,"p99":1848.1329679489136},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.409883499145508,"p90":6.690025329589844,"p95":6.690025329589844,"p99":6.690025329589844},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":81.00104331970215,"p90":110.382080078125,"p95":124.62806701660156,"p99":142.27581024169922},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-8.013010025024414,"p90":11.456966400146484,"p95":11.456966400146484,"p99":11.456966400146484},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1246.3430166244507,"p90":1259.1220140457153,"p95":1263.0480527877808,"p99":1269.3779468536377},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1780.661940574646,"p90":1788.7639999389648,"p95":1791.5979623794556,"p99":1797.366976737976},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3069.7500705718994,"p90":3087.775945663452,"p95":3092.6640033721924,"p99":3102.518081665039},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":87.7079963684082,"p90":113.46817016601562,"p95":122.52092361450195,"p99":142.63010025024414},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.00390625,"host_held_samples":{"dispatch":8,"roundtrip":7},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.009900989942252636,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":1212517376,"scale_bytes":37891168,"total_logical_bytes":1250408544},"roundtrip":{"activation_data_bytes":3637552128,"scale_bytes":37891168,"total_logical_bytes":3675443296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":1212517376,"scale_bytes":37891168,"total_logical_bytes":1250408544},"roundtrip":{"activation_data_bytes":3637552128,"scale_bytes":37891168,"total_logical_bytes":3675443296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":169157,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":10735,"mean":10572.3125,"min":10405,"total":169157},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.008082338608801365,"expert_assignments_per_rank":[12320,12304,12326,12170,12384,12230,12308,12211,12262,12359,12237,12047,12413,12243,12482,12312],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,16,505,5263,15346,11638,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.162261962890625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.0630006417632103,"same_node_fraction":0.5010611414909363,"same_scaleup_domain_fraction":0.5010611414909363,"cross_node_fraction":0.4989388585090637,"cross_domain_fraction":0.4989388585090637,"gpus_per_node":8,"scale_up_domain":8,"copies":169157},"payload_copies_per_rank":[10630,10585,10602,10505,10644,10564,10577,10498,10504,10645,10499,10405,10642,10518,10735,10604],"payload_rank_cv":0.00741199404001236,"routed_copies":169157},"token_rate_at_latency_percentile":{"p50":10378290.471110316,"p90":10292695.012033949,"p95":10264982.06392838,"p99":10196086.500327645},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2383.5620880126953,"p90":2419.1880226135254,"p95":2432.049036026001,"p99":2481.843948364258},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3252.2220611572266,"p90":3274.432897567749,"p95":3282.0260524749756,"p99":3297.0669269561768},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":5635.784149169922,"p90":5693.620920181274,"p95":5714.075088500977,"p99":5778.910875320435},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":5745.549201965332,"p90":5785.84098815918,"p95":5797.378063201904,"p99":5836.637020111084},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5630.857944488525,"p90":5667.34504699707,"p95":5679.370880126953,"p99":5721.75407409668},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":929.1149973869324,"p90":936.8410110473633,"p95":939.5250082015991,"p99":944.9719786643982},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2355.1740646362305,"p90":2380.851984024048,"p95":2388.57102394104,"p99":2400.9900093078613},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":3229.5479774475098,"p90":3249.315023422241,"p95":3255.002975463867,"p99":3267.8630352020264},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.490230560302734,"p90":6.968975067138672,"p95":6.968975067138672,"p99":6.968975067138672},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":132.39717483520508,"p90":190.7639503479004,"p95":212.03899383544922,"p99":257.74621963500977},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-18.147945404052734,"p90":14.300823211669922,"p95":14.300823211669922,"p99":14.300823211669922},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2245.8510398864746,"p90":2267.2300338745117,"p95":2273.2789516448975,"p99":2283.1459045410156},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3175.724983215332,"p90":3188.0240440368652,"p95":3191.8299198150635,"p99":3198.601007461548},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5492.321968078613,"p90":5521.415233612061,"p95":5529.386043548584,"p99":5543.406963348389},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":139.68896865844727,"p90":184.36527252197266,"p95":200.06895065307617,"p99":230.9861183166504},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.00537109375,"host_held_samples":{"dispatch":6,"roundtrip":11},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":2424583168,"scale_bytes":75768224,"total_logical_bytes":2500351392},"roundtrip":{"activation_data_bytes":7273749504,"scale_bytes":75768224,"total_logical_bytes":7349517728},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":2424583168,"scale_bytes":75768224,"total_logical_bytes":2500351392},"roundtrip":{"activation_data_bytes":7273749504,"scale_bytes":75768224,"total_logical_bytes":7349517728},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":338251,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":21435,"mean":21140.6875,"min":20921,"total":338251},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.0070151351392269135,"expert_assignments_per_rank":[24582,24524,24711,24358,24680,24435,24611,24307,24504,24638,24478,24415,25004,24532,24846,24591],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,29,1048,10572,30561,23326,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1613006591796875,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.06308037787675858,"same_node_fraction":0.5009504556655884,"same_scaleup_domain_fraction":0.5009504556655884,"cross_node_fraction":0.49904951453208923,"cross_domain_fraction":0.49904951453208923,"gpus_per_node":8,"scale_up_domain":8,"copies":338251},"payload_copies_per_rank":[21145,21167,21213,20988,21198,21055,21170,20921,21054,21231,21048,21005,21435,21070,21396,21155],"payload_rank_cv":0.00635375315323472,"routed_copies":338251},"token_rate_at_latency_percentile":{"p50":11638723.733768942,"p90":11563792.120743603,"p95":11539306.268819524,"p99":11453830.268010337},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4551.403999328613,"p90":4619.906902313232,"p95":4673.226833343506,"p99":4804.944038391113},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6102.9181480407715,"p90":6144.618988037109,"p95":6156.317234039307,"p99":6177.1087646484375},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":10654.322147369385,"p90":10764.525890350342,"p95":10829.544067382812,"p99":10982.05280303955},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":10768.196105957031,"p90":10862.278938293457,"p95":10899.934768676758,"p99":11025.201797485352},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10633.27407836914,"p90":10713.094711303711,"p95":10746.465682983398,"p99":10897.452354431152},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1866.67001247406,"p90":1877.9269456863403,"p95":1881.3430070877075,"p99":1889.4250392913818},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":4414.543151855469,"p90":4456.2859535217285,"p95":4464.474201202393,"p99":4485.304832458496},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":6057.775974273682,"p90":6089.463233947754,"p95":6098.357200622559,"p99":6115.856170654297},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.9713592529296875,"p90":7.330894470214844,"p95":7.330894470214844,"p99":7.330894470214844},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":225.2788543701172,"p90":321.7020034790039,"p95":352.9510498046875,"p99":486.30523681640625},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":23.23627471923828,"p90":25.438308715820312,"p95":25.438308715820312,"p99":25.438308715820312},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4319.0178871154785,"p90":4348.542213439941,"p95":4358.076095581055,"p99":4375.741958618164},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5987.310886383057,"p90":6002.493858337402,"p95":6007.133007049561,"p99":6017.635822296143},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10380.134582519531,"p90":10419.257164001465,"p95":10430.048942565918,"p99":10450.544357299805},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":254.07791137695312,"p90":339.3716812133789,"p95":373.7821578979492,"p99":519.7200775146484},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.005859375,"host_held_samples":{"dispatch":12,"roundtrip":3},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":131072,"byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":4852026368,"scale_bytes":151625824,"total_logical_bytes":5003652192},"roundtrip":{"activation_data_bytes":14556079104,"scale_bytes":151625824,"total_logical_bytes":14707704928},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":4852026368,"scale_bytes":151625824,"total_logical_bytes":5003652192},"roundtrip":{"activation_data_bytes":14556079104,"scale_bytes":151625824,"total_logical_bytes":14707704928},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":676901,"assignments":786432,"wire":"rank-deduplicated"},"receive":{"max":42692,"mean":42306.3125,"min":41888,"total":676901},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005902660544961691,"expert_assignments_per_rank":[49152,49158,48937,48637,49532,49147,49291,48640,48855,49080,49131,49180,49736,49110,49535,49311],"expert_load_cv":0.02157296985387802,"expert_load_max":2171,"expert_load_mean":2048.0,"expert_load_min":1922,"fanout_histogram":[0,51,2031,21028,61178,46784,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.164344787597656,"fanout_min":2,"hotspot_ratio":1.06005859375,"locality":{"placement":"packed","local_rank_fraction":0.06220702826976776,"same_node_fraction":0.5000066757202148,"same_scaleup_domain_fraction":0.5000066757202148,"cross_node_fraction":0.49999335408210754,"cross_domain_fraction":0.49999335408210754,"gpus_per_node":8,"scale_up_domain":8,"copies":676901},"payload_copies_per_rank":[42365,42433,42070,41888,42530,42296,42419,41908,42030,42386,42324,42267,42692,42172,42625,42496],"payload_rank_cv":0.005482310429215431,"routed_copies":676901},"token_rate_at_latency_percentile":{"p50":12326589.066921044,"p90":12234746.684513297,"p95":12196754.157745771,"p99":12027765.365425356},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":36,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":20,"num_max_rdma_chunked_recv_tokens":140},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":12,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":1073741824,"num_qps_per_rank":10,"is_intranode":false,"cpu_proxy_threads":4},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi300x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9943,"throttle_fraction_max":{"power":0.0577,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":true,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":1936,"p50":2091,"max":2099,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0515,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":1958,"p50":2094,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.048,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":1958,"p50":2093,"max":2100,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0474,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":1978,"p50":2096,"max":2102,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0435,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":1948,"p50":2086,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.0499,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":1944,"p50":2096,"max":2105,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0496,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":1947,"p50":2098,"max":2106,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0533,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":1934,"p50":2089,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0522,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":1933,"p50":2094,"max":2103,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0577,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":1959,"p50":2095,"max":2103,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0459,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":1956,"p50":2097,"max":2105,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0477,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":1952,"p50":2091,"max":2098,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.045,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":1948,"p50":2096,"max":2104,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0509,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":1965,"p50":2094,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0448,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":1948,"p50":2092,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0496,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":1959,"p50":2094,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0462,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1030.3}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16.json b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16.json new file mode 100644 index 0000000000..5b027968f5 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T06:51:32.741166+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":1,"phase":"prefill","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":410.77500581741333,"p90":414.98199105262756,"p95":416.74500703811646,"p99":420.7099974155426},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":536.8850231170654,"p90":543.533980846405,"p95":545.7779765129089,"p99":552.2270202636719},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":947.6600289344788,"p90":958.5159718990326,"p95":962.5229835510254,"p99":972.9370176792145},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":933.6389899253845,"p90":940.1689767837524,"p95":941.9310092926025,"p99":948.4210014343262},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":939.6880269050598,"p90":946.3779926300049,"p95":948.1809735298157,"p99":954.0299773216248},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":381.0110092163086,"p90":383.21399688720703,"p95":383.976012468338,"p99":385.4979872703552},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":520.6199884414673,"p90":526.3479948043823,"p95":528.5120010375977,"p99":534.400999546051},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.3700079917907715,"p90":6.448984146118164,"p95":6.448984146118164,"p99":6.448984146118164},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.133994579315186,"p90":13.45902681350708,"p95":14.50204849243164,"p99":17.82602071762085},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.8810162544250488,"p90":1.963973045349121,"p95":1.963973045349121,"p99":1.963973045349121},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":383.37400555610657,"p90":386.8190050125122,"p95":388.66299390792847,"p99":392.54799485206604},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":518.297016620636,"p90":524.586021900177,"p95":526.5489816665649,"p99":531.715989112854},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":911.6449952125549,"p90":918.0160164833069,"p95":920.0180172920227,"p99":925.1059889793396},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":27.922987937927246,"p90":31.64595365524292,"p95":32.607972621917725,"p99":35.333991050720215},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"roundtrip":{"activation_data_bytes":1041510400,"scale_bytes":0,"total_logical_bytes":1041510400},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"roundtrip":{"activation_data_bytes":1041510400,"scale_bytes":0,"total_logical_bytes":1041510400},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":36325,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":4587,"mean":4540.625,"min":4510,"total":36325},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.009137080051004887,"expert_assignments_per_rank":[6159,6057,6238,6119,6110,6118,6221,6130],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,46,880,3426,3151,689,0,0],"fanout_max":6,"fanout_mean":4.4342041015625,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.12393668293952942,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":36325},"payload_copies_per_rank":[4582,4510,4587,4547,4516,4535,4533,4515],"payload_rank_cv":0.006128025706857443,"routed_copies":36325},"token_rate_at_latency_percentile":{"p50":8717786.930818975,"p90":8656160.713579418,"p95":8639700.88906493,"p99":8586732.27753125},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":691.5559768676758,"p90":696.4830160140991,"p95":698.3669996261597,"p99":701.6109824180603},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":860.489010810852,"p90":868.5010075569153,"p95":870.864987373352,"p99":876.7920136451721},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1552.0449876785278,"p90":1564.9840235710144,"p95":1569.2319869995117,"p99":1578.4029960632324},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1540.0680303573608,"p90":1547.7190017700195,"p95":1550.2029657363892,"p99":1554.1269779205322},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1542.3909425735474,"p90":1550.9239435195923,"p95":1553.6079406738281,"p99":1558.655023574829},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":680.1390051841736,"p90":685.2269768714905,"p95":686.3489747047424,"p99":688.5120272636414},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":841.4999842643738,"p90":848.8720059394836,"p95":851.5160083770752,"p99":854.6810150146484},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.369948387145996,"p90":6.410002708435059,"p95":6.410002708435059,"p99":6.410002708435059},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":12.25900650024414,"p90":18.46790313720703,"p95":19.87004280090332,"p99":23.674964904785156},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.6820430755615234,"p90":2.164006233215332,"p95":2.164006233215332,"p99":2.164006233215332},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":672.7280020713806,"p90":676.5329837799072,"p95":677.6959896087646,"p99":680.8199882507324},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":833.8090181350708,"p90":840.0179743766785,"p95":842.5419926643372,"p99":846.9889760017395},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1523.121953010559,"p90":1531.7349433898926,"p95":1534.6989631652832,"p99":1539.5870208740234},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":19.148945808410645,"p90":24.716973304748535,"p95":26.23903751373291,"p99":29.4039249420166},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"roundtrip":{"activation_data_bytes":2083221504,"scale_bytes":0,"total_logical_bytes":2083221504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"roundtrip":{"activation_data_bytes":2083221504,"scale_bytes":0,"total_logical_bytes":2083221504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":72657,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":9142,"mean":9082.125,"min":9007,"total":72657},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00735377287492156,"expert_assignments_per_rank":[12308,12293,12414,12290,12351,12153,12356,12139],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,92,1754,6798,6421,1319,0,0],"fanout_max":6,"fanout_mean":4.43463134765625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.12587912380695343,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":72657},"payload_copies_per_rank":[9133,9070,9142,9078,9120,9057,9050,9007],"payload_rank_cv":0.004780265036970377,"routed_copies":72657},"token_rate_at_latency_percentile":{"p50":10622469.017266514,"p90":10564025.443323119,"p95":10545775.141245712,"p99":10511626.852760999},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1268.6599493026733,"p90":1276.7130136489868,"p95":1278.756022453308,"p99":1282.4009656906128},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1505.776047706604,"p90":1515.2299404144287,"p95":1518.1139707565308,"p99":1521.9199657440186},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":2774.4359970092773,"p90":2791.9429540634155,"p95":2796.869993209839,"p99":2804.3209314346313},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":2761.0559463500977,"p90":2772.9549407958984,"p95":2776.400089263916,"p99":2785.2931022644043},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2762.5789642333984,"p90":2773.555040359497,"p95":2777.200937271118,"p99":2786.6148948669434},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1227.5999784469604,"p90":1237.1740341186523,"p95":1240.3379678726196,"p99":1244.1439628601074},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1491.9960498809814,"p90":1502.771019935608,"p95":1504.7739744186401,"p99":1512.3460292816162},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.329059600830078,"p90":6.44993782043457,"p95":6.44993782043457,"p99":6.44993782043457},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":21.312952041625977,"p90":29.80494499206543,"p95":31.808137893676758,"p99":35.41302680969238},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.4039745330810547,"p90":3.8061141967773438,"p95":3.8061141967773438,"p99":3.8061141967773438},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1218.9459800720215,"p90":1225.3559827804565,"p95":1226.999044418335,"p99":1229.8829555511475},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1468.1589603424072,"p90":1475.7720232009888,"p95":1478.4150123596191,"p99":1483.7830066680908},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2712.5020027160645,"p90":2722.759008407593,"p95":2726.124048233032,"p99":2733.175039291382},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":50.395965576171875,"p90":59.56912040710449,"p95":61.491966247558594,"p99":64.77785110473633},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"roundtrip":{"activation_data_bytes":4160278528,"scale_bytes":0,"total_logical_bytes":4160278528},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"roundtrip":{"activation_data_bytes":4160278528,"scale_bytes":0,"total_logical_bytes":4160278528},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":145099,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":18284,"mean":18137.375,"min":17991,"total":145099},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005656693130731583,"expert_assignments_per_rank":[24624,24496,24614,24519,24621,24284,24656,24794],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,192,3636,13520,12793,2627,0,0],"fanout_max":6,"fanout_mean":4.428070068359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.125583216547966,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":145099},"payload_copies_per_rank":[18250,18101,18151,18109,18125,17991,18088,18284],"payload_rank_cv":0.004795215092599392,"routed_copies":145099},"token_rate_at_latency_percentile":{"p50":11861380.407308267,"p90":11814440.14024425,"p95":11798930.196314095,"p99":11759070.1393149},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2384.171962738037,"p90":2393.626928329468,"p95":2396.1100578308105,"p99":2401.4370441436768},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2784.532070159912,"p90":2794.0258979797363,"p95":2797.070026397705,"p99":2802.3180961608887},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":5168.704032897949,"p90":5187.652826309204,"p95":5193.180084228516,"p99":5203.755140304565},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":5158.810138702393,"p90":5175.233840942383,"p95":5180.5620193481445,"p99":5186.651229858398},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5158.888816833496,"p90":5174.912929534912,"p95":5179.521083831787,"p99":5186.450004577637},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2348.6790657043457,"p90":2360.57710647583,"p95":2363.8219833374023,"p99":2368.9889907836914},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2769.550085067749,"p90":2783.289909362793,"p95":2786.454916000366,"p99":2796.0290908813477},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.410121917724609,"p90":6.489753723144531,"p95":6.489753723144531,"p99":6.489753723144531},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":35.21299362182617,"p90":50.354957580566406,"p95":55.36317825317383,"p99":62.89386749267578},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.6458969116210938,"p90":5.729198455810547,"p95":5.729198455810547,"p99":5.729198455810547},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2340.5470848083496,"p90":2347.3548889160156,"p95":2349.2801189422607,"p99":2352.8449535369873},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2754.326105117798,"p90":2763.2598876953125,"p95":2766.184091567993,"p99":2772.0329761505127},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5115.664958953857,"p90":5127.801895141602,"p95":5131.689071655273,"p99":5138.498783111572},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":43.784141540527344,"p90":56.20384216308594,"p95":59.689998626708984,"p99":64.77737426757812},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"roundtrip":{"activation_data_bytes":8316542976,"scale_bytes":0,"total_logical_bytes":8316542976},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"roundtrip":{"activation_data_bytes":8316542976,"scale_bytes":0,"total_logical_bytes":8316542976},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":290058,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":36460,"mean":36257.25,"min":36144,"total":290058},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00432177959010005,"expert_assignments_per_rank":[49106,49069,49115,48918,49142,48893,49536,49437],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,413,7235,27175,25451,5262,0,0],"fanout_max":6,"fanout_mean":4.425933837890625,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.1256852149963379,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":290058},"payload_copies_per_rank":[36354,36201,36216,36144,36222,36193,36268,36460],"payload_rank_cv":0.0026462646201252937,"routed_copies":290058},"token_rate_at_latency_percentile":{"p50":12703510.838643294,"p90":12664174.430059437,"p95":12652907.274492018,"p99":12636003.420867255},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":6,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":0,"num_qps_per_rank":1,"is_intranode":true,"cpu_proxy_threads":0},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi300x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9957,"throttle_fraction_max":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2101,"p50":2106,"max":2107,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":377.5},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2101,"p50":2106,"max":2107,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":377.5},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2094,"p50":2099,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":377.5},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2101,"p50":2105,"max":2106,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":377.6},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2103,"p50":2108,"max":2109,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":70,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":377.5},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2104,"p50":2108,"max":2110,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":377.5},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2101,"p50":2106,"max":2108,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":70,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":377.5},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2096,"p50":2101,"max":2103,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":377.5}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8.json b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8.json new file mode 100644 index 0000000000..8593c2abf6 --- /dev/null +++ b/collectivex/results/mi300x/mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T07:08:40.603492+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":1,"phase":"prefill","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi300x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi300x"},"case_id":"mi300x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":408.45200419425964,"p90":412.97900676727295,"p95":414.50101137161255,"p99":418.34598779678345},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":450.4750072956085,"p90":458.2059979438782,"p95":460.6899917125702,"p99":467.9810106754303},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":858.9270114898682,"p90":871.1850047111511,"p95":875.1910030841827,"p99":886.3269984722137},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":832.5679898262024,"p90":847.9509949684143,"p95":852.9580235481262,"p99":864.175021648407},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":839.057981967926,"p90":846.2679982185364,"p95":848.7120270729065,"p99":856.2830090522766},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":184.95699763298035,"p90":187.96199560165405,"p95":188.96299600601196,"p99":191.24700129032135},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":386.17900013923645,"p90":389.3029987812042,"p95":390.3850018978119,"p99":393.1100070476532},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":420.79100012779236,"p90":430.12499809265137,"p95":434.6120059490204,"p99":443.4640109539032},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.290018558502197,"p90":6.3700079917907715,"p95":6.3700079917907715,"p99":6.3700079917907715},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.012997150421143,"p90":12.539029121398926,"p95":13.499975204467773,"p99":18.186986446380615},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.362966537475586,"p90":5.849003791809082,"p95":5.849003791809082,"p99":5.849003791809082},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":385.0170075893402,"p90":388.46200704574585,"p95":390.02498984336853,"p99":393.669992685318},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":431.16599321365356,"p90":437.8969967365265,"p95":440.53998589515686,"p99":448.1920003890991},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":815.1810169219971,"p90":822.0319747924805,"p95":824.7150182723999,"p99":832.3280215263367},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":23.555994033813477,"p90":27.240991592407227,"p95":28.522014617919922,"p99":31.648993492126465},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.006756756920367479,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":260377600,"scale_bytes":8136800,"total_logical_bytes":268514400},"roundtrip":{"activation_data_bytes":781132800,"scale_bytes":8136800,"total_logical_bytes":789269600},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":260377600,"scale_bytes":8136800,"total_logical_bytes":268514400},"roundtrip":{"activation_data_bytes":781132800,"scale_bytes":8136800,"total_logical_bytes":789269600},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":36325,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":4587,"mean":4540.625,"min":4510,"total":36325},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.009137080051004887,"expert_assignments_per_rank":[6159,6057,6238,6119,6110,6118,6221,6130],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,46,880,3426,3151,689,0,0],"fanout_max":6,"fanout_mean":4.4342041015625,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.12393668293952942,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":36325},"payload_copies_per_rank":[4582,4510,4587,4547,4516,4535,4533,4515],"payload_rank_cv":0.006128025706857443,"routed_copies":36325},"token_rate_at_latency_percentile":{"p50":9763330.03922624,"p90":9680148.625783833,"p95":9652272.783564886,"p99":9566930.458034901},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":702.8530240058899,"p90":707.8610062599182,"p95":709.9440097808838,"p99":714.4309878349304},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":664.7570133209229,"p90":671.6070175170898,"p95":674.2100119590759,"p99":682.9040050506592},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1367.6100373268127,"p90":1379.468023777008,"p95":1384.1540217399597,"p99":1397.3349928855896},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1347.0180034637451,"p90":1354.9909591674805,"p95":1357.193946838379,"p99":1363.5629415512085},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1358.515977859497,"p90":1366.5670156478882,"p95":1369.9729442596436,"p99":1378.3060312271118},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":375.883013010025,"p90":380.65099716186523,"p95":382.0919990539551,"p99":385.5380117893219},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":675.9330034255981,"p90":681.1800003051758,"p95":682.8629970550537,"p99":686.1479878425598},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":648.9719748497009,"p90":654.9410223960876,"p95":656.5840244293213,"p99":660.5899930000305},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.328940391540527,"p90":6.44993782043457,"p95":6.44993782043457,"p99":6.44993782043457},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11.898994445800781,"p90":17.625927925109863,"p95":19.308924674987793,"p99":24.396061897277832},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.5619993209838867,"p90":2.4030208587646484,"p95":2.4030208587646484,"p99":2.4030208587646484},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":676.2539744377136,"p90":680.3399920463562,"p95":681.9819808006287,"p99":685.6279969215393},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":647.130012512207,"p90":653.6189913749695,"p95":655.9029817581177,"p99":663.5140180587769},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1331.6349983215332,"p90":1339.406967163086,"p95":1342.3320055007935,"p99":1351.6260385513306},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":26.759028434753418,"p90":32.48894214630127,"p95":34.33096408843994,"p99":37.175893783569336},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":520805376,"scale_bytes":16275168,"total_logical_bytes":537080544},"roundtrip":{"activation_data_bytes":1562416128,"scale_bytes":16275168,"total_logical_bytes":1578691296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":520805376,"scale_bytes":16275168,"total_logical_bytes":537080544},"roundtrip":{"activation_data_bytes":1562416128,"scale_bytes":16275168,"total_logical_bytes":1578691296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":72657,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":9142,"mean":9082.125,"min":9007,"total":72657},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00735377287492156,"expert_assignments_per_rank":[12308,12293,12414,12290,12351,12153,12356,12139],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,92,1754,6798,6421,1319,0,0],"fanout_max":6,"fanout_mean":4.43463134765625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.12587912380695343,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":72657},"payload_copies_per_rank":[9133,9070,9142,9078,9120,9057,9050,9007],"payload_rank_cv":0.004780265036970377,"routed_copies":72657},"token_rate_at_latency_percentile":{"p50":12060218.846903026,"p90":11989166.877580725,"p95":11959360.269595828,"p99":11887055.29019071},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1274.469017982483,"p90":1283.4030389785767,"p95":1285.9660387039185,"p99":1290.1740074157715},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1115.8720254898071,"p90":1123.242974281311,"p95":1125.6870031356812,"p99":1133.2980394363403},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":2390.34104347229,"p90":2406.6460132598877,"p95":2411.6530418395996,"p99":2423.472046852112},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":2378.88503074646,"p90":2390.4619216918945,"p95":2393.4669494628906,"p99":2399.6360301971436},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2383.0509185791016,"p90":2393.5070037841797,"p95":2396.791934967041,"p99":2403.8829803466797},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":760.7399821281433,"p90":766.1489844322205,"p95":767.9510116577148,"p99":772.2780108451843},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1234.8910570144653,"p90":1242.1009540557861,"p95":1244.5850372314453,"p99":1250.7940530776978},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1104.6149730682373,"p90":1114.0300035476685,"p95":1116.9929504394531,"p99":1128.0499696731567},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.370067596435547,"p90":6.448984146118164,"p95":6.448984146118164,"p99":6.448984146118164},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":23.03481101989746,"p90":32.00793266296387,"p95":35.01105308532715,"p99":41.66293144226074},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.0438899993896484,"p90":3.284931182861328,"p95":3.284931182861328,"p99":3.284931182861328},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1224.0749597549438,"p90":1231.0049533843994,"p95":1232.4869632720947,"p99":1235.7720136642456},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1090.8750295639038,"p90":1097.6849794387817,"p95":1099.768042564392,"p99":1104.8150062561035},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2331.7739963531494,"p90":2341.468095779419,"p95":2344.6340560913086,"p99":2350.761890411377},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":51.71704292297363,"p90":61.45215034484863,"p95":64.13602828979492,"p99":68.14289093017578},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":1040069632,"scale_bytes":32502176,"total_logical_bytes":1072571808},"roundtrip":{"activation_data_bytes":3120208896,"scale_bytes":32502176,"total_logical_bytes":3152711072},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":1040069632,"scale_bytes":32502176,"total_logical_bytes":1072571808},"roundtrip":{"activation_data_bytes":3120208896,"scale_bytes":32502176,"total_logical_bytes":3152711072},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":145099,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":18284,"mean":18137.375,"min":17991,"total":145099},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005656693130731583,"expert_assignments_per_rank":[24624,24496,24614,24519,24621,24284,24656,24794],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,192,3636,13520,12793,2627,0,0],"fanout_max":6,"fanout_mean":4.428070068359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.125583216547966,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":145099},"payload_copies_per_rank":[18250,18101,18151,18109,18125,17991,18088,18284],"payload_rank_cv":0.004795215092599392,"routed_copies":145099},"token_rate_at_latency_percentile":{"p50":13750440.556905087,"p90":13690371.47089738,"p95":13671608.086602896,"p99":13631279.171199221},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2389.3001079559326,"p90":2399.8770713806152,"p95":2402.7209281921387,"p99":2407.2868824005127},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2018.2650089263916,"p90":2028.439998626709,"p95":2031.9249629974365,"p99":2040.4179096221924},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":4407.565116882324,"p90":4428.317070007324,"p95":4434.645891189575,"p99":4447.704792022705},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":4407.485008239746,"p90":4421.064853668213,"p95":4425.110816955566,"p99":4481.956005096436},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4401.515960693359,"p90":4416.819095611572,"p95":4421.145915985107,"p99":4431.760787963867},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1552.1659851074219,"p90":1562.5410079956055,"p95":1565.98699092865,"p99":1572.4760293960571},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2348.159074783325,"p90":2358.0939769744873,"p95":2361.098051071167,"p99":2366.06502532959},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2007.9290866851807,"p90":2017.1830654144287,"p95":2020.0269222259521,"p99":2025.8350372314453},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.449699401855469,"p90":6.451129913330078,"p95":6.451129913330078,"p99":6.451129913330078},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":36.253929138183594,"p90":51.19609832763672,"p95":54.882049560546875,"p99":66.65992736816406},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.322673797607422,"p90":3.6859512329101562,"p95":3.6859512329101562,"p99":3.6859512329101562},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2343.7929153442383,"p90":2351.963996887207,"p95":2354.127883911133,"p99":2358.0141067504883},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1987.6580238342285,"p90":1995.9110021591187,"p95":1998.5140562057495,"p99":2003.6420822143555},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4355.766773223877,"p90":4367.505073547363,"p95":4371.06990814209,"p99":4379.723072052002},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":46.35000228881836,"p90":57.36589431762695,"p95":60.690879821777344,"p99":66.45965576171875},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":2079135744,"scale_bytes":64972992,"total_logical_bytes":2144108736},"roundtrip":{"activation_data_bytes":6237407232,"scale_bytes":64972992,"total_logical_bytes":6302380224},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":2079135744,"scale_bytes":64972992,"total_logical_bytes":2144108736},"roundtrip":{"activation_data_bytes":6237407232,"scale_bytes":64972992,"total_logical_bytes":6302380224},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":290058,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":36460,"mean":36257.25,"min":36144,"total":290058},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00432177959010005,"expert_assignments_per_rank":[49106,49069,49115,48918,49142,48893,49536,49437],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,413,7235,27175,25451,5262,0,0],"fanout_max":6,"fanout_mean":4.425933837890625,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.1256852149963379,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":290058},"payload_copies_per_rank":[36354,36201,36216,36144,36222,36193,36268,36460],"payload_rank_cv":0.0026462646201252937,"routed_copies":290058},"token_rate_at_latency_percentile":{"p50":14889415.507123662,"p90":14837827.536363157,"p95":14823306.275200704,"p99":14787801.764478793},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":6,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":0,"num_qps_per_rank":1,"is_intranode":true,"cpu_proxy_threads":0},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi300x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9952,"throttle_fraction_max":{"power":0.0872,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":true,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":1907,"p50":2091,"max":2107,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":69,"clock_event_reasons":[],"throttle_fraction":{"power":0.0855,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":479.0},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":1932,"p50":2091,"max":2106,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0764,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":479.0},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":1906,"p50":2082,"max":2098,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":70,"clock_event_reasons":[],"throttle_fraction":{"power":0.0829,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":479.0},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":1917,"p50":2090,"max":2104,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0795,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":479.0},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":1900,"p50":2091,"max":2106,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":71,"clock_event_reasons":[],"throttle_fraction":{"power":0.0855,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":479.0},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":1903,"p50":2092,"max":2108,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":69,"clock_event_reasons":[],"throttle_fraction":{"power":0.0817,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":479.0},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":1885,"p50":2087,"max":2105,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":72,"clock_event_reasons":[],"throttle_fraction":{"power":0.0865,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":479.0},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":1893,"p50":2084,"max":2102,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":750.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.0872,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":479.0}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16.json b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16.json new file mode 100644 index 0000000000..0409658eab --- /dev/null +++ b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T09:58:25.372884+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256","mode":"low-latency","nodes":2,"phase":"decode","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256],"ladder_dropped":[512],"ladder_cap":256},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":61.452001333236694,"p90":64.25700336694717,"p95":65.37800282239914,"p99":67.90199875831604},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":61.73200160264969,"p90":63.335999846458435,"p95":64.77800011634827,"p99":67.7419975399971},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":123.18400293588638,"p90":127.59300321340561,"p95":130.1560029387474,"p99":135.64399629831314},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":105.23899644613266,"p90":109.1649979352951,"p95":111.36700212955475,"p99":115.2539998292923},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":115.05299806594849,"p90":116.73600226640701,"p95":117.81799793243408,"p99":120.5419972538948},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":51.07700079679489,"p90":53.24000120162964,"p95":53.52000147104263,"p99":54.482001811265945},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":48.67300018668175,"p90":52.23799869418144,"p95":54.00199815630913,"p99":56.1240017414093},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048999726772308,"p90":6.089001893997192,"p95":6.089001893997192,"p99":6.089001893997192},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.571002304553986,"p90":13.42099905014038,"p95":15.624001622200012,"p99":19.7489932179451},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.840999186038971,"p90":1.1610016226768494,"p95":1.1610016226768494,"p99":1.1610016226768494},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":50.71600154042244,"p90":51.87699943780899,"p95":52.39899829030037,"p99":53.679000586271286},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":51.15700140595436,"p90":52.198998630046844,"p95":52.51999944448471,"p99":54.00099977850914},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":101.03199630975723,"p90":103.27599942684174,"p95":103.67599874734879,"p99":104.51699793338776},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.541000127792358,"p90":18.267005681991577,"p95":19.02800053358078,"p99":20.391002297401428},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":9,"mean":5.125,"min":2,"total":82},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":139066.34567513646,"p90":137061.40084775118,"p95":135802.68109101322,"p99":132733.82194173845},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":60.610998421907425,"p90":63.415996730327606,"p95":64.37800079584122,"p99":66.01899862289429},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":62.33400106430054,"p90":65.01799821853638,"p95":66.2589967250824,"p99":68.82300227880478},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":122.94499948620796,"p90":128.43399494886398,"p95":130.63699752092361,"p99":134.84200090169907},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":108.4040030837059,"p90":111.44799739122391,"p95":112.81000077724457,"p99":116.25500023365021},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":115.29400199651718,"p90":117.49599874019623,"p95":118.5780018568039,"p99":120.38099765777588},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":54.120998829603195,"p90":56.1240017414093,"p95":56.48500099778175,"p99":57.28600174188614},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":50.436001271009445,"p90":51.95799842476845,"p95":52.47899889945984,"p99":57.40699917078018},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048999726772308,"p90":6.09000027179718,"p95":6.09000027179718,"p99":6.09000027179718},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.972002029418945,"p90":13.21999728679657,"p95":14.541998505592346,"p99":17.826996743679047},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.7220059633255005,"p90":1.0810047388076782,"p95":1.0810047388076782,"p99":1.0810047388076782},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":54.56199869513512,"p90":55.44300004839897,"p95":55.68400025367737,"p99":56.405000388622284},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":49.515001475811005,"p90":50.5560003221035,"p95":51.03699862957001,"p99":52.799999713897705},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":99.75100308656693,"p90":101.43200308084488,"p95":102.31500118970871,"p99":104.75900024175644},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":15.542998909950256,"p90":18.02700012922287,"p95":18.787004053592682,"p99":20.63100039958954},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":15,"mean":10.375,"min":7,"total":166},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":277551.2988174932,"p90":272349.6999311226,"p95":269864.55749729666,"p99":265822.6848307981},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":63.77600133419037,"p90":66.74099713563919,"p95":67.90199875831604,"p99":69.70500200986862},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":65.49900025129318,"p90":68.14300268888474,"p95":69.46499645709991,"p99":72.50899821519852},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":129.27500158548355,"p90":134.88399982452393,"p95":137.36699521541595,"p99":142.21400022506714},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":109.96600240468979,"p90":115.41400104761124,"p95":117.61700361967087,"p99":121.5829998254776},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":122.30399996042252,"p90":124.34700131416321,"p95":125.50899386405945,"p99":127.4310052394867},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":54.80299890041351,"p90":56.56500160694122,"p95":56.88599869608879,"p99":57.80699849128723},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":52.3190014064312,"p90":54.722998291254044,"p95":56.405000388622284,"p99":59.36900153756142},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.050005555152893,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.13199782371521,"p90":13.940997421741486,"p95":15.102997422218323,"p99":17.667993903160095},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.001998782157898,"p90":2.3640021681785583,"p95":2.3640021681785583,"p99":2.3640021681785583},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":54.68200147151947,"p90":56.08399957418442,"p95":56.52499943971634,"p99":57.52699822187424},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":56.244999170303345,"p90":57.60699883103371,"p95":57.96699970960617,"p99":58.688998222351074},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":105.3989976644516,"p90":106.9210022687912,"p95":107.40199685096741,"p99":108.76400023698807},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":16.9450044631958,"p90":18.98900419473648,"p95":19.910000264644623,"p99":22.55299687385559},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":26,"mean":20.5625,"min":15,"total":329},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":523286.237741287,"p90":514688.7285066387,"p95":509923.61606626614,"p99":502232.56011927384},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":65.33899903297424,"p90":67.86300241947174,"p95":68.9840018749237,"p99":71.06699794530869},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":66.05999916791916,"p90":68.62299889326096,"p95":69.82500106096268,"p99":71.82800024747849},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":131.3989982008934,"p90":136.4860013127327,"p95":138.80900293588638,"p99":142.89499819278717},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":110.60699820518494,"p90":114.29200321435928,"p95":115.73400348424911,"p99":119.53999847173691},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":120.02000212669373,"p90":122.22400307655334,"p95":123.30599874258041,"p99":125.6289929151535},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":55.88499829173088,"p90":57.60699883103371,"p95":57.96800181269646,"p99":58.84899944067001},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":52.3190014064312,"p90":54.802000522613525,"p95":55.48299849033356,"p99":57.24700167775154},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.17299634218216,"p90":12.457005679607391,"p95":14.661997556686401,"p99":17.026998102664948},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.8400008082389832,"p90":1.3629943132400513,"p95":1.3629943132400513,"p99":1.3629943132400513},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":56.52499943971634,"p90":57.486001402139664,"p95":57.767000049352646,"p99":58.6090013384819},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":49.39499869942665,"p90":52.23900079727173,"p95":52.799999713897705,"p99":53.92099916934967},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":106.04000091552734,"p90":107.20200091600418,"p95":107.52200335264206,"p99":108.56399685144424},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.98099958896637,"p90":16.383998095989227,"p95":17.264999449253082,"p99":19.26898956298828},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":49,"mean":40.5625,"min":36,"total":649},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":1066488.899615937,"p90":1047257.4680754723,"p95":1038067.9067140847,"p99":1018873.0883678087},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":66.58100336790085,"p90":69.54500079154968,"p95":70.46599686145782,"p99":72.10899889469147},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":69.50400024652481,"p90":71.66799902915955,"p95":72.91000336408615,"p99":74.75300133228302},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":136.08500361442566,"p90":141.21299982070923,"p95":143.37600022554398,"p99":146.8620002269745},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":112.4889999628067,"p90":115.57400226593018,"p95":116.53599888086319,"p99":121.50300294160843},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":125.1479983329773,"p90":127.11100280284882,"p95":128.152996301651,"p99":129.99600172042847},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":57.56700038909912,"p90":59.289999306201935,"p95":59.73000079393387,"p99":61.211999505758286},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":53.60100045800209,"p90":55.52399903535843,"p95":55.96499890089035,"p99":57.165998965501785},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048999726772308,"p90":6.048999726772308,"p95":6.048999726772308,"p99":6.048999726772308},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.208000540733337,"p90":9.734004735946655,"p95":12.1380016207695,"p99":14.58200067281723},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.4810020327568054,"p90":0.6410032510757446,"p95":0.6410032510757446,"p99":0.6410032510757446},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":58.48800018429756,"p90":59.408999979496,"p95":59.689998626708984,"p99":60.210999101400375},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":52.51900106668472,"p90":53.720999509096146,"p95":54.24100160598755,"p99":55.88499829173088},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":110.6870025396347,"p90":111.76799982786179,"p95":112.08900064229965,"p99":112.81000077724457},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.462009072303772,"p90":16.783997416496277,"p95":17.746999859809875,"p99":19.548997282981873},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":89,"mean":82.125,"min":72,"total":1314},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":2045578.062853782,"p90":2013987.7300555962,"p95":1997612.286781171,"p99":1969291.3367486317},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":80.84200322628021,"p90":83.60599726438522,"p95":84.80799943208694,"p99":87.41199970245361},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":77.03600078821182,"p90":78.91800254583359,"p95":79.80000227689743,"p99":81.6819965839386},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":157.87800401449203,"p90":162.5239998102188,"p95":164.60800170898438,"p99":169.0939962863922},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":128.3939927816391,"p90":130.59699535369873,"p95":132.27899372577667,"p99":135.08300483226776},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":134.68299806118011,"p90":137.36699521541595,"p95":138.68799805641174,"p99":141.5729969739914},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":63.29599767923355,"p90":67.38200038671494,"p95":67.78199970722198,"p99":68.66300106048584},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":58.04700031876564,"p90":61.69300153851509,"p95":62.33299896121025,"p99":63.534997403621674},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009995937347412,"p90":6.089001893997192,"p95":6.089001893997192,"p99":6.089001893997192},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.609005689620972,"p90":11.135995388031006,"p95":13.179004192352295,"p99":17.867006361484528},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.8410066366195679,"p90":0.3210008144378662,"p95":0.3210008144378662,"p99":0.3210008144378662},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":59.48999896645546,"p90":61.45299971103668,"p95":62.57399916648865,"p99":64.4569993019104},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":55.723998695611954,"p90":57.1260005235672,"p95":57.88800120353699,"p99":59.689998626708984},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":126.23000144958496,"p90":127.3919939994812,"p95":127.75300443172455,"p99":128.47399711608887},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.413001894950867,"p90":11.057987809181213,"p95":12.299999594688416,"p99":15.341997146606445},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.004048583097755909,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":178,"mean":164.0,"min":148,"total":2624},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":3801519.177405174,"p90":3727241.752628371,"p95":3691739.7840853003,"p99":3616508.874881418},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":100.99299997091293,"p90":104.07699644565582,"p95":105.23899644613266,"p99":107.48200118541718},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":97.50699996948242,"p90":99.78999942541122,"p95":100.99200159311295,"p99":102.95400023460388},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":198.49999994039536,"p90":203.86699587106705,"p95":206.2309980392456,"p99":210.43600142002106},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":174.86299574375153,"p90":177.1470010280609,"p95":177.9080033302307,"p99":180.19099533557892},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":180.95199763774872,"p90":183.1970065832138,"p95":184.07699465751648,"p99":186.12000346183777},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":76.43599808216095,"p90":82.92499929666519,"p95":85.56900173425674,"p99":90.25599807500839},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":68.34299862384796,"p90":76.27400010824203,"p95":78.35700362920761,"p99":83.9259997010231},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048992276191711,"p90":6.09000027179718,"p95":6.09000027179718,"p99":6.09000027179718},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.849989295005798,"p90":11.978000402450562,"p95":13.781994581222534,"p99":17.06700026988983},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.6420016288757324,"p90":0.4409998655319214,"p95":0.4409998655319214,"p99":0.4409998655319214},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":85.80899983644485,"p90":88.37199956178665,"p95":89.53499794006348,"p99":91.45700186491013},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":78.63900065422058,"p90":79.96100187301636,"p95":80.48100024461746,"p99":82.44500309228897},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":170.21700739860535,"p90":171.6189980506897,"p95":172.09899425506592,"p99":172.82100021839142},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.81499457359314,"p90":13.458997011184692,"p95":14.260008931159973,"p99":16.063004732131958},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.006103515625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":363,"mean":328.5,"min":292,"total":5256},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":5658959.355894845,"p90":5589610.982725677,"p95":5562889.60445708,"p99":5501826.676088377},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":137.9680037498474,"p90":147.5829929113388,"p95":149.66599643230438,"p99":152.54999697208405},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":129.6750009059906,"p90":134.56200063228607,"p95":135.68399846553802,"p99":137.84700632095337},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":267.643004655838,"p90":282.1449935436249,"p95":285.3499948978424,"p99":290.3970032930374},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":236.55599355697632,"p90":243.1669980287552,"p95":244.76799368858337,"p99":251.13800168037415},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":246.9329982995987,"p90":251.41900777816772,"p95":252.7810037136078,"p99":255.5060088634491},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":111.7279976606369,"p90":121.30200117826462,"p95":122.74400144815445,"p99":124.42699819803238},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":99.99100118875504,"p90":110.56599766016006,"p95":112.32899874448776,"p99":116.61600321531296},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.050005555152893,"p90":6.170004606246948,"p95":6.170004606246948,"p99":6.170004606246948},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.619989156723022,"p90":20.672008395195007,"p95":22.47600257396698,"p99":29.362976551055908},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.9260069131851196,"p90":2.5639981031417847,"p95":2.5639981031417847,"p99":2.5639981031417847},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":116.09499901533127,"p90":121.66400253772736,"p95":122.5849986076355,"p99":124.1070032119751},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":107.80200362205505,"p90":113.7700006365776,"p95":114.65299874544144,"p99":116.13599956035614},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":228.2630056142807,"p90":235.71400344371796,"p95":237.0380014181137,"p99":239.52099680900574},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":16.66700839996338,"p90":22.713005542755127,"p95":24.3149995803833,"p99":27.639001607894897},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.006103515625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":692,"mean":660.8125,"min":603,"total":10573},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":8293747.753855092,"p90":8145764.387897805,"p95":8101874.62630821,"p99":8015467.069091589},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":200.78200101852417,"p90":212.36099302768707,"p95":217.20699965953827,"p99":223.77699613571167},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":195.1739937067032,"p90":205.7500034570694,"p95":208.5929960012436,"p99":213.52200210094452},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":395.95599472522736,"p90":418.11099648475647,"p95":425.79999566078186,"p99":437.2989982366562},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":366.19099974632263,"p90":377.3680031299591,"p95":384.89899039268494,"p99":440.7829940319061},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":375.20700693130493,"p90":383.49801301956177,"p95":385.4199945926666,"p99":390.38899540901184},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":161.7240011692047,"p90":177.78800427913666,"p95":179.35100197792053,"p99":183.15599858760834},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":156.71700239181519,"p90":172.05999791622162,"p95":176.50599777698517,"p99":186.12000346183777},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129980087280273,"p90":6.448984146118164,"p95":6.448984146118164,"p99":6.448984146118164},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":21.751999855041504,"p90":38.45599293708801,"p95":46.3080108165741,"p99":97.50699996948242},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-4.4870078563690186,"p90":3.1239986419677734,"p95":3.1239986419677734,"p99":3.1239986419677734},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":171.3789999485016,"p90":178.6690056324005,"p95":179.75099384784698,"p99":181.4740002155304},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":174.18299615383148,"p90":183.2360029220581,"p95":185.15899777412415,"p99":187.96400725841522},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":353.45199704170227,"p90":360.7049882411957,"p95":363.62698674201965,"p99":367.0719861984253},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":22.03500270843506,"p90":33.689022064208984,"p95":36.93300485610962,"p99":41.58100485801697},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.006103515625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":1373,"mean":1323.8125,"min":1247,"total":21181},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":10916640.479344564,"p90":10680629.001827627,"p95":10627367.696190443,"p99":10492099.029862786},"tokens_per_rank":256}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"InterNodeV1LL","block_num":32,"rdma_block_num":16,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":0,"scale_type_size":0,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"inter-node-v1ll-vllm-cudagraph-spin-ar","rank_alignment":{"method":"in-graph-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":true,"cuda_graph_supported":true,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9948,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2070,"p50":2095,"max":2097,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2073,"p50":2099,"max":2101,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2072,"p50":2099,"max":2100,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2071,"p50":2099,"max":2100,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2067,"p50":2093,"max":2095,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2070,"p50":2097,"max":2098,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2067,"p50":2095,"max":2097,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2071,"p50":2097,"max":2098,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2073,"p50":2100,"max":2101,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2066,"p50":2094,"max":2096,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2068,"p50":2095,"max":2096,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2060,"p50":2090,"max":2094,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2071,"p50":2099,"max":2101,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2068,"p50":2097,"max":2098,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2058,"p50":2089,"max":2093,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2069,"p50":2098,"max":2099,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":712.5}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-fp8.json b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-fp8.json new file mode 100644 index 0000000000..4592fb8da5 --- /dev/null +++ b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T10:12:08.782390+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256","mode":"low-latency","nodes":2,"phase":"decode","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-mori-deepseek-v4-pro-low-latency-decode-ep16-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256],"ladder_dropped":[512],"ladder_cap":256},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":62.49399855732918,"p90":65.17799943685532,"p95":66.21900200843811,"p99":68.14199686050415},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":63.81600350141525,"p90":67.34099984169006,"p95":68.46299767494202,"p99":70.62599807977676},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":126.31000205874443,"p90":132.51899927854538,"p95":134.68199968338013,"p99":138.76799494028091},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":107.84299671649933,"p90":110.9670028090477,"p95":112.44899779558182,"p99":116.7759969830513},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":117.93699860572815,"p90":120.02100050449371,"p95":121.3420033454895,"p99":123.54599684476852},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":50.074998289346695,"p90":52.960000932216644,"p95":53.44000086188316,"p99":54.56199869513512},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":51.47799849510193,"p90":54.24199998378754,"p95":55.52399903535843,"p99":57.64700099825859},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.611997425556183,"p90":11.737994849681854,"p95":13.499997556209564,"p99":16.58499985933304},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.8020028471946716,"p90":1.1209994554519653,"p95":1.1209994554519653,"p99":1.1209994554519653},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":49.435000866651535,"p90":50.7970005273819,"p95":51.11800134181976,"p99":52.23799869418144},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":55.723998695611954,"p90":56.80600181221962,"p95":57.165998965501785,"p99":58.88799950480461},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":100.79199820756912,"p90":102.67499834299088,"p95":103.55599969625473,"p99":106.20000213384628},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.305999994277954,"p90":19.78999376296997,"p95":20.911000669002533,"p99":22.873997688293457},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":587776,"scale_bytes":18368,"total_logical_bytes":606144},"roundtrip":{"activation_data_bytes":1763328,"scale_bytes":18368,"total_logical_bytes":1781696},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":587776,"scale_bytes":18368,"total_logical_bytes":606144},"roundtrip":{"activation_data_bytes":1763328,"scale_bytes":18368,"total_logical_bytes":1781696},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":9,"mean":5.125,"min":2,"total":82},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":135665.6536045075,"p90":133310.00352226646,"p95":131858.70975316106,"p99":129506.42196932917},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":60.89099869132042,"p90":63.975997269153595,"p95":64.85799700021744,"p99":66.82100147008896},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":66.01899862289429,"p90":69.2640021443367,"p95":70.94600051641464,"p99":74.35200363397598},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":126.9099973142147,"p90":133.2399994134903,"p95":135.80399751663208,"p99":141.17300510406494},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":111.52800172567368,"p90":113.57100307941437,"p95":114.33199793100357,"p99":118.41899901628494},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":117.5369992852211,"p90":120.26099860668182,"p95":121.38299643993378,"p99":123.30500036478043},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":53.640998899936676,"p90":55.3629994392395,"p95":55.64400181174278,"p99":56.48500099778175},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":52.15800181031227,"p90":54.16199937462807,"p95":55.925000458955765,"p99":58.40900167822838},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049007177352905,"p90":6.129004061222076,"p95":6.129004061222076,"p99":6.129004061222076},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.2510018944740295,"p90":11.578001081943512,"p95":12.860000133514404,"p99":15.983998775482178},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.7610023021697998,"p90":2.2030025720596313,"p95":2.2030025720596313,"p99":2.2030025720596313},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":54.402001202106476,"p90":55.323000997304916,"p95":55.6039996445179,"p99":56.164998561143875},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":52.8389997780323,"p90":53.881000727415085,"p95":54.52200025320053,"p99":56.324999779462814},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":102.95599699020386,"p90":104.31700199842453,"p95":105.3989976644516,"p99":108.00299793481827},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.58200067281723,"p90":16.944997012615204,"p95":17.986997961997986,"p99":19.86999809741974},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":1189888,"scale_bytes":37184,"total_logical_bytes":1227072},"roundtrip":{"activation_data_bytes":3569664,"scale_bytes":37184,"total_logical_bytes":3606848},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":1189888,"scale_bytes":37184,"total_logical_bytes":1227072},"roundtrip":{"activation_data_bytes":3569664,"scale_bytes":37184,"total_logical_bytes":3606848},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":15,"mean":10.375,"min":7,"total":166},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":272254.69592215144,"p90":266087.92851169663,"p95":263628.35766568966,"p99":259519.07793952004},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":65.69799780845642,"p90":69.10400092601776,"p95":70.0249969959259,"p99":72.06899672746658},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":68.86400282382965,"p90":72.22899794578552,"p95":73.67099821567535,"p99":76.43499970436096},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":134.56200063228607,"p90":141.33299887180328,"p95":143.69599521160126,"p99":148.50399643182755},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":113.73099684715271,"p90":116.29500240087509,"p95":118.21799725294113,"p99":119.8199987411499},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":125.34800171852112,"p90":127.6320070028305,"p95":128.712996840477,"p99":130.4360032081604},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":54.68299984931946,"p90":56.64600059390068,"p95":57.04599991440773,"p99":57.686999440193176},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":53.8409985601902,"p90":57.24700167775154,"p95":57.8869991004467,"p99":59.769000858068466},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049998104572296,"p90":6.130002439022064,"p95":6.130002439022064,"p99":6.130002439022064},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.292995393276215,"p90":11.617004871368408,"p95":12.45899498462677,"p99":14.781005680561066},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.881001353263855,"p90":0.961005687713623,"p95":0.961005687713623,"p99":0.961005687713623},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":52.71900072693825,"p90":53.720999509096146,"p95":54.322000592947006,"p99":55.403999984264374},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":60.210999101400375,"p90":61.531998217105865,"p95":61.85400113463402,"p99":63.17500025033951},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":108.00299793481827,"p90":109.20500010251999,"p95":109.72599685192108,"p99":111.64800077676773},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.266005277633667,"p90":19.227996468544006,"p95":20.229004323482513,"p99":22.31299877166748},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":2358272,"scale_bytes":73696,"total_logical_bytes":2431968},"roundtrip":{"activation_data_bytes":7074816,"scale_bytes":73696,"total_logical_bytes":7148512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":2358272,"scale_bytes":73696,"total_logical_bytes":2431968},"roundtrip":{"activation_data_bytes":7074816,"scale_bytes":73696,"total_logical_bytes":7148512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":26,"mean":20.5625,"min":15,"total":329},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":510578.5423186648,"p90":501441.6172158185,"p95":497230.2842060284,"p99":490662.0750856923},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":64.49700146913528,"p90":67.50199943780899,"p95":68.38399916887283,"p99":70.06599754095078},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":68.86299699544907,"p90":71.98800146579742,"p95":73.47100228071213,"p99":76.91600173711777},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":133.35999846458435,"p90":139.49000090360641,"p95":141.85500144958496,"p99":146.98199927806854},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":114.41200226545334,"p90":117.41700023412704,"p95":118.41800063848495,"p99":121.98399752378464},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":123.02500009536743,"p90":125.78900158405304,"p95":127.03099846839905,"p99":129.03399765491486},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":54.802000522613525,"p90":56.68500065803528,"p95":57.1260005235672,"p99":58.20799991488457},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":53.801000118255615,"p90":55.96499890089035,"p95":56.765999644994736,"p99":60.21000072360039},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048999726772308,"p90":6.0889944434165955,"p95":6.0889944434165955,"p99":6.0889944434165955},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.49300080537796,"p90":12.579001486301422,"p95":14.222003519535065,"p99":16.343995928764343},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.8010044693946838,"p90":0.9230002760887146,"p95":0.9230002760887146,"p99":0.9230002760887146},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":55.723998695611954,"p90":57.006001472473145,"p95":57.36599862575531,"p99":58.16800147294998},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":52.51900106668472,"p90":56.76500126719475,"p95":57.406000792980194,"p99":58.92900004982948},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":107.36200213432312,"p90":109.00499671697617,"p95":109.88599807024002,"p99":112.16899752616882},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":15.74300229549408,"p90":17.986997961997986,"p95":19.02899146080017,"p99":21.230995655059814},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":4652032,"scale_bytes":145376,"total_logical_bytes":4797408},"roundtrip":{"activation_data_bytes":13956096,"scale_bytes":145376,"total_logical_bytes":14101472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":4652032,"scale_bytes":145376,"total_logical_bytes":4797408},"roundtrip":{"activation_data_bytes":13956096,"scale_bytes":145376,"total_logical_bytes":14101472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":49,"mean":40.5625,"min":36,"total":649},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":1040438.9343692422,"p90":1017577.0408231564,"p95":1007628.0714414917,"p99":991986.6262093178},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":65.81900268793106,"p90":69.023996591568,"p95":70.30600309371948,"p99":72.10899889469147},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":69.90499794483185,"p90":72.90899753570557,"p95":74.19200241565704,"p99":77.27599889039993},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":135.7240006327629,"p90":141.93299412727356,"p95":144.49800550937653,"p99":149.3849977850914},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":114.21199887990952,"p90":118.09799820184708,"p95":119.13999915122986,"p99":121.78400158882141},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":127.95299291610718,"p90":129.87500429153442,"p95":130.79699873924255,"p99":133.4809958934784},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":56.96599930524826,"p90":58.688998222351074,"p95":59.20900031924248,"p99":60.13000011444092},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":55.08299916982651,"p90":57.28600174188614,"p95":58.56899917125702,"p99":60.77200174331665},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.09000027179718,"p90":6.1689987778663635,"p95":6.1689987778663635,"p99":6.1689987778663635},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.49099999666214,"p90":11.858001351356506,"p95":12.818999588489532,"p99":15.262991189956665},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.0410025715827942,"p90":1.6020014882087708,"p95":1.6020014882087708,"p99":1.6020014882087708},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":55.883001536130905,"p90":57.52699822187424,"p95":58.24799835681915,"p99":59.20900031924248},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":55.64400181174278,"p90":57.006001472473145,"p95":57.927001267671585,"p99":60.531001538038254},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":110.60699820518494,"p90":111.56900227069855,"p95":112.00899630784988,"p99":113.45099657773972},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.346002161502838,"p90":19.268997013568878,"p95":20.2300027012825,"p99":22.153005003929138},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":9418752,"scale_bytes":294336,"total_logical_bytes":9713088},"roundtrip":{"activation_data_bytes":28256256,"scale_bytes":294336,"total_logical_bytes":28550592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":9418752,"scale_bytes":294336,"total_logical_bytes":9713088},"roundtrip":{"activation_data_bytes":28256256,"scale_bytes":294336,"total_logical_bytes":28550592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":89,"mean":82.125,"min":72,"total":1314},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":2000734.7555195314,"p90":1971126.017638844,"p95":1957231.4538375817,"p99":1917876.0113858848},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":75.19400119781494,"p90":77.43699848651886,"p95":78.39799672365189,"p99":80.76199889183044},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":73.71100038290024,"p90":75.8339986205101,"p95":77.15599983930588,"p99":79.27899807691574},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":148.90500158071518,"p90":153.27099710702896,"p95":155.55399656295776,"p99":160.04099696874619},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":125.26899576187134,"p90":129.8760026693344,"p95":131.07800483703613,"p99":133.16099345684052},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":137.9680037498474,"p90":139.97100293636322,"p95":140.97200334072113,"p99":143.09599995613098},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":65.9789964556694,"p90":67.54200160503387,"p95":67.86300241947174,"p99":68.70400160551071},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":58.6090013384819,"p90":60.531001538038254,"p95":61.85299903154373,"p99":63.134998083114624},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":6.089001893997192,"p95":6.089001893997192,"p99":6.089001893997192},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.450004875659943,"p90":12.017995119094849,"p95":13.82099837064743,"p99":17.185993492603302},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.2010037899017334,"p90":-0.6809979677200317,"p95":-0.6809979677200317,"p99":-0.6809979677200317},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":66.61999970674515,"p90":67.58199632167816,"p95":67.82200187444687,"p99":68.30299645662308},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":57.006001472473145,"p90":58.20799991488457,"p95":59.08900126814842,"p99":61.33300065994263},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":122.14499711990356,"p90":123.46599996089935,"p95":123.90699982643127,"p99":124.94900077581406},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":15.782997012138367,"p90":18.10799539089203,"p95":18.98900419473648,"p99":20.74999362230301},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":18808832,"scale_bytes":587776,"total_logical_bytes":19396608},"roundtrip":{"activation_data_bytes":56426496,"scale_bytes":587776,"total_logical_bytes":57014272},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":18808832,"scale_bytes":587776,"total_logical_bytes":19396608},"roundtrip":{"activation_data_bytes":56426496,"scale_bytes":587776,"total_logical_bytes":57014272},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":178,"mean":164.0,"min":148,"total":2624},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":3711005.3496774337,"p90":3657900.4883802757,"p95":3631926.821402444,"p99":3578017.555745542},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":101.03300213813782,"p90":103.8770005106926,"p95":105.03800213336945,"p99":107.40199685096741},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":91.53799712657928,"p90":93.74099969863892,"p95":94.86299753189087,"p99":97.30599820613861},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":192.5709992647171,"p90":197.6180002093315,"p95":199.90099966526031,"p99":204.70799505710602},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":166.41199588775635,"p90":169.0949946641922,"p95":170.01600563526154,"p99":171.93900048732758},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":172.33900725841522,"p90":175.06299912929535,"p95":176.1849969625473,"p99":179.10900712013245},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":85.32799780368805,"p90":89.17500078678131,"p95":89.93600308895111,"p99":90.85699915885925},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":73.39099794626236,"p90":77.3560032248497,"p95":77.8769999742508,"p99":78.91900092363358},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049007177352905,"p90":6.1289966106414795,"p95":6.1289966106414795,"p99":6.1289966106414795},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.411003112792969,"p90":11.136993765830994,"p95":12.74000108242035,"p99":15.784993767738342},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.6809979677200317,"p90":0.7620006799697876,"p95":0.7620006799697876,"p99":0.7620006799697876},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":82.92499929666519,"p90":85.48799902200699,"p95":86.57100051641464,"p99":88.29200267791748},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":70.86700201034546,"p90":72.14900106191635,"p95":72.75000214576721,"p99":74.87300038337708},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":161.64299845695496,"p90":162.96499967575073,"p95":163.2869988679886,"p99":164.08699750900269},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.695993900299072,"p90":13.619989156723022,"p95":14.82100784778595,"p99":17.22399890422821},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":37675008,"scale_bytes":1177344,"total_logical_bytes":38852352},"roundtrip":{"activation_data_bytes":113025024,"scale_bytes":1177344,"total_logical_bytes":114202368},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":37675008,"scale_bytes":1177344,"total_logical_bytes":38852352},"roundtrip":{"activation_data_bytes":113025024,"scale_bytes":1177344,"total_logical_bytes":114202368},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":363,"mean":328.5,"min":292,"total":5256},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":5941777.292847894,"p90":5849322.844307665,"p95":5812072.637590577,"p99":5717188.747035933},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":139.69099521636963,"p90":147.5829929113388,"p95":149.06500279903412,"p99":151.30800008773804},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":115.49299955368042,"p90":119.74100023508072,"p95":120.38099765777588,"p99":122.22400307655334},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":255.18399477005005,"p90":267.3239931464195,"p95":269.44600045681,"p99":273.5320031642914},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":219.08999979496002,"p90":225.09899735450745,"p95":226.86199843883514,"p99":230.786994099617},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":228.10399532318115,"p90":232.47100412845612,"p95":233.71300101280212,"p99":235.63499748706818},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":113.69200050830841,"p90":120.34200131893158,"p95":121.58399820327759,"p99":123.70699644088745},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":86.36999875307083,"p90":94.26199644804001,"p95":95.94400227069855,"p99":98.78899902105331},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.09000027179718,"p90":6.248995661735535,"p95":6.248995661735535,"p99":6.248995661735535},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.45999538898468,"p90":19.870996475219727,"p95":22.032007575035095,"p99":26.520997285842896},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.4050047397613525,"p90":3.685995936393738,"p95":3.685995936393738,"p99":3.685995936393738},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":111.08700186014175,"p90":119.22000348567963,"p95":120.34200131893158,"p99":122.10399657487869},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":89.89600092172623,"p90":94.22200173139572,"p95":95.02299875020981,"p99":96.22500091791153},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":210.91899275779724,"p90":216.68599545955658,"p95":217.80900657176971,"p99":219.53000128269196},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":16.103997826576233,"p90":21.03099226951599,"p95":22.231996059417725,"p99":24.67900514602661},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00495867757126689,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":75787264,"scale_bytes":2368352,"total_logical_bytes":78155616},"roundtrip":{"activation_data_bytes":227361792,"scale_bytes":2368352,"total_logical_bytes":229730144},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":75787264,"scale_bytes":2368352,"total_logical_bytes":78155616},"roundtrip":{"activation_data_bytes":227361792,"scale_bytes":2368352,"total_logical_bytes":229730144},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":692,"mean":660.8125,"min":603,"total":10573},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":8978360.93181254,"p90":8809700.838511199,"p95":8762884.354421586,"p99":8691408.414882835},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":198.21900129318237,"p90":209.07600224018097,"p95":213.360995054245,"p99":220.9330052137375},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":159.7999930381775,"p90":176.58600211143494,"p95":178.10800671577454,"p99":180.31099438667297},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":358.01899433135986,"p90":385.6620043516159,"p95":391.46900177001953,"p99":401.24399960041046},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":341.7940139770508,"p90":351.72900557518005,"p95":354.0950119495392,"p99":358.4209978580475},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":340.4330015182495,"p90":350.0080108642578,"p95":352.41198539733887,"p99":356.1779856681824},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":156.03500604629517,"p90":162.92600333690643,"p95":167.3319935798645,"p99":175.26400089263916},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":119.7810024023056,"p90":131.47799670696259,"p95":134.80299711227417,"p99":140.57199656963348},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.369978189468384,"p90":6.449013948440552,"p95":6.449013948440552,"p99":6.449013948440552},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.945021390914917,"p90":25.239020586013794,"p95":26.921004056930542,"p99":35.85100173950195},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.7260055541992188,"p90":-2.3640096187591553,"p95":-2.3640096187591553,"p99":-2.3640096187591553},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":172.73999750614166,"p90":178.30799520015717,"p95":179.189994931221,"p99":181.11200630664825},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":140.41200280189514,"p90":148.18300306797028,"p95":149.8659998178482,"p99":153.1910002231598},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":318.32098960876465,"p90":326.0509967803955,"p95":328.575998544693,"p99":333.5820138454437},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":22.47297763824463,"p90":33.93000364303589,"p95":36.893993616104126,"p99":41.54101014137268},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00495867757126689,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":151825408,"scale_bytes":4744544,"total_logical_bytes":156569952},"roundtrip":{"activation_data_bytes":455476224,"scale_bytes":4744544,"total_logical_bytes":460220768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":151825408,"scale_bytes":4744544,"total_logical_bytes":156569952},"roundtrip":{"activation_data_bytes":455476224,"scale_bytes":4744544,"total_logical_bytes":460220768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":1373,"mean":1323.8125,"min":1247,"total":21181},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":12031736.000131663,"p90":11702589.29184491,"p95":11622760.20601804,"p99":11499868.506235696},"tokens_per_rank":256}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"InterNodeV1LL","block_num":32,"rdma_block_num":16,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":56,"scale_type_size":4,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"inter-node-v1ll-vllm-cudagraph-spin-ar","rank_alignment":{"method":"in-graph-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":true,"cuda_graph_supported":true,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9948,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2072,"p50":2095,"max":2097,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.5},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2100,"max":2101,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.4},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2099,"max":2100,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.4},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2099,"max":2100,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.5},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2067,"p50":2093,"max":2094,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":54,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.4},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2072,"p50":2097,"max":2098,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.4},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2068,"p50":2095,"max":2097,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.4},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2069,"p50":2097,"max":2098,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.5},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2076,"p50":2100,"max":2102,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.5},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2069,"p50":2094,"max":2096,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.5},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2071,"p50":2094,"max":2096,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.5},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2067,"p50":2090,"max":2094,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.5},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2076,"p50":2099,"max":2100,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.5},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2070,"p50":2097,"max":2098,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.5},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2063,"p50":2089,"max":2091,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.5},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2071,"p50":2098,"max":2100,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":734.5}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-bf16.json b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-bf16.json new file mode 100644 index 0000000000..ae2afd3160 --- /dev/null +++ b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T10:28:10.055687+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":2,"phase":"decode","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":132.55999982357025,"p90":135.28400659561157,"p95":136.4859938621521,"p99":138.64800333976746},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":76.71499997377396,"p90":78.31700146198273,"p95":79.40000295639038,"p99":81.92399889230728},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":209.2749997973442,"p90":213.6010080575943,"p95":215.88599681854248,"p99":220.57200223207474},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":189.72599506378174,"p90":191.36899709701538,"p95":192.41000711917877,"p99":196.1749941110611},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":199.7399926185608,"p90":201.58299803733826,"p95":202.62500643730164,"p99":205.14899492263794},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":121.98399752378464,"p90":123.62600117921829,"p95":124.06700104475021,"p99":125.98900496959686},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":63.735999166965485,"p90":65.61899930238724,"p95":66.01999700069427,"p99":66.82100147008896},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048992276191711,"p90":6.049007177352905,"p95":6.049007177352905,"p99":6.049007177352905},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.690993905067444,"p90":10.73598861694336,"p95":12.01900839805603,"p99":15.744000673294067},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.7599890232086182,"p90":-0.5200058221817017,"p95":-0.5200058221817017,"p99":-0.5200058221817017},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":120.8610013127327,"p90":122.78500199317932,"p95":123.54499846696854,"p99":124.82599914073944},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":62.734998762607574,"p90":63.81700187921524,"p95":64.1769990324974,"p99":65.82000106573105},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":182.11500346660614,"p90":183.5169941186905,"p95":183.9579939842224,"p99":185.7600063085556},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.746001482009888,"p90":20.309001207351685,"p95":21.310999989509583,"p99":24.23599362373352},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":9,"mean":5.125,"min":2,"total":82},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":80104.13833625628,"p90":79371.77319406866,"p95":78963.60020573714,"p99":77992.09548179181},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":141.41300320625305,"p90":143.89599859714508,"p95":144.97800171375275,"p99":147.06100523471832},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":78.75800132751465,"p90":81.00199699401855,"p95":82.28400349617004,"p99":85.16799658536911},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":220.1710045337677,"p90":224.89799559116364,"p95":227.2620052099228,"p99":232.22900182008743},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":199.74100589752197,"p90":201.86400413513184,"p95":202.9470056295395,"p99":205.1900029182434},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":207.91299641132355,"p90":209.99599993228912,"p95":210.75700223445892,"p99":213.23999762535095},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":125.30900537967682,"p90":126.6700029373169,"p95":126.99100375175476,"p99":127.59299576282501},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":70.2660009264946,"p90":71.82800024747849,"p95":72.18900322914124,"p99":73.27000051736832},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.008997559547424,"p90":6.050005555152893,"p95":6.050005555152893,"p99":6.050005555152893},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.8890005350112915,"p90":11.49800419807434,"p95":13.899996876716614,"p99":17.66599714756012},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.4809945821762085,"p90":0.8800029754638672,"p95":0.8800029754638672,"p99":0.8800029754638672},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":124.74799901247025,"p90":125.9090006351471,"p95":126.23099982738495,"p99":127.67200171947479},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":66.17999821901321,"p90":67.38100200891495,"p95":67.82200187444687,"p99":70.2660009264946},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":189.20500576496124,"p90":190.72799384593964,"p95":191.24899804592133,"p99":192.21000373363495},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.66699755191803,"p90":21.07200026512146,"p95":22.0319926738739,"p99":24.435997009277344},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":15,"mean":10.375,"min":7,"total":166},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":153910.53254166458,"p90":152383.8549797046,"p95":151833.6266920387,"p99":150065.6553946411},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":140.61200618743896,"p90":143.25599372386932,"p95":144.25800740718842,"p99":146.139994263649},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":92.37899631261826,"p90":94.34200078248978,"p95":94.94300186634064,"p99":97.38700091838837},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":232.99100250005722,"p90":237.5979945063591,"p95":239.20100927352905,"p99":243.52699518203735},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":205.34999668598175,"p90":208.07400345802307,"p95":208.83500576019287,"p99":211.7989957332611},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":212.67999708652496,"p90":214.68299627304077,"p95":215.68399667739868,"p99":217.72800385951996},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":127.43200361728668,"p90":129.23400104045868,"p95":129.6350061893463,"p99":130.4360032081604},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":73.15000146627426,"p90":75.39299875497818,"p95":76.03400200605392,"p99":77.03600078821182},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048992276191711,"p90":6.089001893997192,"p95":6.089001893997192,"p99":6.089001893997192},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.289005279541016,"p90":10.57499647140503,"p95":12.13701069355011,"p99":14.502003788948059},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.7210075855255127,"p90":1.1219978332519531,"p95":1.1219978332519531,"p99":1.1219978332519531},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":125.02799928188324,"p90":128.23300063610077,"p95":128.712996840477,"p99":129.6350061893463},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":67.90299713611603,"p90":69.30500268936157,"p95":69.7460025548935,"p99":70.86700201034546},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":201.2230008840561,"p90":203.02599668502808,"p95":203.4270018339157,"p99":204.38799262046814},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11.656001210212708,"p90":15.140995383262634,"p95":16.26300811767578,"p99":17.705991864204407},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":26,"mean":20.5625,"min":15,"total":329},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":300921.5764375,"p90":298113.9685539079,"p95":296730.40645535523,"p99":293944.73317861935},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":148.46299588680267,"p90":151.10699832439423,"p95":152.06800401210785,"p99":154.4319987297058},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":96.38500213623047,"p90":98.26699644327164,"p95":99.43000227212906,"p99":102.27400064468384},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":244.84799802303314,"p90":249.37399476766586,"p95":251.4980062842369,"p99":256.70599937438965},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":224.13699328899384,"p90":227.743998169899,"p95":228.70400547981262,"p99":229.7469973564148},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":232.3099970817566,"p90":234.5529943704605,"p95":235.71500182151794,"p99":237.91800439357758},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":137.88799941539764,"p90":139.09000158309937,"p95":139.44999873638153,"p99":140.45199751853943},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":79.71999794244766,"p90":82.32399821281433,"p95":83.04499834775925,"p99":85.6890007853508},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":6.1690062284469604,"p95":6.1690062284469604,"p99":6.1690062284469604},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.808994174003601,"p90":9.5750093460083,"p95":10.97799837589264,"p99":14.142006635665894},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.881001353263855,"p90":1.0409951210021973,"p95":1.0409951210021973,"p99":1.0409951210021973},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":137.68799602985382,"p90":138.7680023908615,"p95":139.12999629974365,"p99":139.69099521636963},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":77.63700187206268,"p90":79.31900024414062,"p95":79.88099753856659,"p99":81.44299685955048},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":213.96300196647644,"p90":215.96600115299225,"p95":216.52700006961823,"p99":218.20999681949615},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.387004733085632,"p90":20.510002970695496,"p95":21.431997418403625,"p99":23.79399538040161},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":49,"mean":40.5625,"min":36,"total":649},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":550987.9110150955,"p90":545718.89113397,"p95":543028.6532925931,"p99":538000.4776278095},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":165.3289943933487,"p90":168.49400103092194,"p95":169.5760041475296,"p99":172.30099439620972},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":112.24900186061859,"p90":114.09199982881546,"p95":114.97300118207932,"p99":116.89600348472595},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":277.5779962539673,"p90":282.5860008597374,"p95":284.5490053296089,"p99":289.19699788093567},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":251.85999274253845,"p90":254.42400574684143,"p95":255.18399477005005,"p99":258.22898745536804},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":254.90298867225647,"p90":257.1069896221161,"p95":258.26799869537354,"p99":260.3920102119446},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":152.34999358654022,"p90":153.99199724197388,"p95":154.4329971075058,"p99":155.95600008964539},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":92.73900091648102,"p90":94.38300132751465,"p95":94.98299658298492,"p99":96.02399915456772},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048977375030518,"p90":6.089001893997192,"p95":6.089001893997192,"p99":6.089001893997192},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.728006362915039,"p90":9.776011109352112,"p95":11.175990104675293,"p99":13.179987668991089},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.840991735458374,"p90":1.0020136833190918,"p95":1.0020136833190918,"p99":1.0020136833190918},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":146.57899737358093,"p90":148.30300211906433,"p95":149.02399480342865,"p99":150.82700550556183},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":86.4500030875206,"p90":88.05300295352936,"p95":88.65399658679962,"p99":90.57699888944626},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":241.72499775886536,"p90":243.4069961309433,"p95":243.8880056142807,"p99":245.01000344753265},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.179004192352295,"p90":16.142994165420532,"p95":17.225012183189392,"p99":19.747987389564514},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":89,"mean":82.125,"min":72,"total":1314},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":1004303.6424698575,"p90":995694.4397982215,"p95":991218.429279546,"p99":983133.0838132486},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":168.69500279426575,"p90":171.5790033340454,"p95":172.82100021839142,"p99":175.1839965581894},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":126.75100564956665,"p90":129.6350061893463,"p95":130.4759979248047,"p99":132.4400007724762},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":295.4460084438324,"p90":301.2140095233917,"p95":303.2969981431961,"p99":307.6239973306656},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":272.33099937438965,"p90":275.09400248527527,"p95":275.85500478744507,"p99":277.6179909706116},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":273.5309898853302,"p90":276.2160003185272,"p95":277.17700600624084,"p99":279.1000008583069},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":155.1949977874756,"p90":156.79599344730377,"p95":157.19600021839142,"p99":158.51899981498718},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":106.27999901771545,"p90":108.6840033531189,"p95":109.40500348806381,"p99":111.88799887895584},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049007177352905,"p90":6.168007850646973,"p95":6.168007850646973,"p99":6.168007850646973},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.769014358520508,"p90":9.694010019302368,"p95":10.616004467010498,"p99":12.138992547988892},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.7210075855255127,"p90":0.9220242500305176,"p95":0.9220242500305176,"p99":0.9220242500305176},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":149.6250033378601,"p90":152.6699960231781,"p95":153.7110060453415,"p99":155.55499494075775},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":103.19600254297256,"p90":105.47900199890137,"p95":106.20000213384628,"p99":108.00299793481827},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":261.99498772621155,"p90":263.7979984283447,"p95":264.27900791168213,"p99":265.3610110282898},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11.416018009185791,"p90":14.701008796691895,"p95":15.701979398727417,"p99":18.025994300842285},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.004048583097755909,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":178,"mean":164.0,"min":148,"total":2624},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":1871817.1575902274,"p90":1853621.801088898,"p95":1847195.0735641902,"p99":1834467.9269991529},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":177.9080033302307,"p90":182.39599466323853,"p95":183.83799493312836,"p99":186.48099899291992},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":143.85700225830078,"p90":147.0620036125183,"p95":148.1429934501648,"p99":151.26800537109375},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":321.7650055885315,"p90":329.45799827575684,"p95":331.98098838329315,"p99":337.7490043640137},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":293.2409942150116,"p90":298.36899042129517,"p95":300.65199732780457,"p99":304.17799949645996},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":296.12699151039124,"p90":299.2919981479645,"p95":300.2929985523224,"p99":302.9769957065582},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":151.187002658844,"p90":153.11099588871002,"p95":154.07200157642365,"p99":158.59900414943695},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":114.09299820661545,"p90":117.49699711799622,"p95":119.3000003695488,"p99":123.74699860811234},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009995937347412,"p90":6.168007850646973,"p95":6.168007850646973,"p99":6.168007850646973},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009995937347412,"p90":9.73498821258545,"p95":10.856002569198608,"p99":13.458997011184692},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.2829899787902832,"p90":1.1210143566131592,"p95":1.1210143566131592,"p99":1.1210143566131592},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":152.9099941253662,"p90":157.11599588394165,"p95":158.5180014371872,"p99":160.96200048923492},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":119.58000063896179,"p90":122.42499738931656,"p95":123.54599684476852,"p99":125.26899576187134},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":285.2709889411926,"p90":287.83300518989563,"p95":288.673996925354,"p99":290.3969883918762},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.857999324798584,"p90":13.62299919128418,"p95":14.422982931137085,"p99":16.304999589920044},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.006103515625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":363,"mean":328.5,"min":292,"total":5256},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":3457975.9000593075,"p90":3421407.877045056,"p95":3410002.9136096574,"p99":3379794.5537481434},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":206.55100047588348,"p90":211.11799776554108,"p95":213.04099261760712,"p99":218.20899844169617},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":177.34700441360474,"p90":183.03599953651428,"p95":184.55800414085388,"p99":188.44300508499146},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":383.8980048894882,"p90":394.15399730205536,"p95":397.598996758461,"p99":406.6520035266876},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":343.67701411247253,"p90":353.25199365615845,"p95":355.93700408935547,"p99":361.9860112667084},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":351.8509864807129,"p90":356.97901248931885,"p95":358.3810031414032,"p99":361.62498593330383},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":156.31499886512756,"p90":157.55699574947357,"p95":157.87799656391144,"p99":159.04000401496887},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":121.18300050497055,"p90":126.06999278068542,"p95":127.59199738502502,"p99":132.11999833583832},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.36899471282959,"p90":6.69097900390625,"p95":6.69097900390625,"p99":6.69097900390625},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.8110058307647705,"p90":9.815007448196411,"p95":10.976016521453857,"p99":14.622986316680908},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.3619959354400635,"p90":3.004997968673706,"p95":3.004997968673706,"p99":3.004997968673706},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":169.0949946641922,"p90":173.98199439048767,"p95":175.98499357700348,"p99":179.4700026512146},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":147.58199453353882,"p90":152.71100401878357,"p95":154.31299805641174,"p99":157.03700482845306},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":333.5820138454437,"p90":338.7100100517273,"p95":340.2320146560669,"p99":342.35599637031555},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.229007720947266,"p90":23.797988891601562,"p95":26.721984148025513,"p99":29.006004333496094},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.006103515625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":692,"mean":660.8125,"min":603,"total":10573},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":5820645.894685486,"p90":5737031.949633953,"p95":5714588.613928119,"p99":5663325.488183281},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":271.7289924621582,"p90":290.83898663520813,"p95":298.2499897480011,"p99":311.3099932670593},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":278.90101075172424,"p90":296.7660129070282,"p95":301.89400911331177,"p99":310.1069927215576},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":550.6300032138824,"p90":587.6049995422363,"p95":600.1439988613129,"p99":621.416985988617},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":522.5890278816223,"p90":539.0520095825195,"p95":543.9010262489319,"p99":556.8379759788513},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":522.7090120315552,"p90":536.2870097160339,"p95":539.3739938735962,"p99":546.1840033531189},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":221.6939926147461,"p90":226.14100575447083,"p95":227.5030016899109,"p99":237.79800534248352},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":215.16400575637817,"p90":229.2259931564331,"p95":237.71800100803375,"p99":256.3459873199463},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.408989429473877,"p90":6.449997425079346,"p95":6.449997425079346,"p99":6.449997425079346},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":19.991040229797363,"p90":25.878965854644775,"p95":27.201026678085327,"p99":30.24601936340332},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.364039421081543,"p90":4.84699010848999,"p95":4.84699010848999,"p99":4.84699010848999},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":231.34900629520416,"p90":243.20800602436066,"p95":244.8890060186386,"p99":247.1729964017868},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":251.78000330924988,"p90":266.32198691368103,"p95":269.76698637008667,"p99":275.57599544525146},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":498.3519911766052,"p90":511.52998208999634,"p95":515.2170062065125,"p99":522.026002407074},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":24.074018001556396,"p90":34.810006618499756,"p95":38.0139946937561,"p99":42.020976543426514},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.006103515625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":1373,"mean":1323.8125,"min":1247,"total":21181},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":7836099.829387926,"p90":7637701.316257591,"p95":7593988.67302436,"p99":7499304.217725054},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":458.13098549842834,"p90":494.26600337028503,"p95":502.2779703140259,"p99":522.3479866981506},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":469.3470001220703,"p90":494.78501081466675,"p95":499.63200092315674,"p99":509.3259811401367},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":927.4779856204987,"p90":989.0510141849518,"p95":1001.9099712371826,"p99":1031.6739678382874},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":828.0460238456726,"p90":857.9710125923157,"p95":864.6209836006165,"p99":871.9930052757263},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":852.5270223617554,"p90":877.4440288543701,"p95":882.972002029419,"p99":892.5870060920715},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":363.1860017776489,"p90":372.079998254776,"p95":381.1349868774414,"p99":389.9480104446411},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":387.1839940547943,"p90":420.99499702453613,"p95":426.8440008163452,"p99":442.0680105686188},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1280131340026855,"p90":6.449997425079346,"p95":6.449997425079346,"p99":6.449997425079346},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":26.880979537963867,"p90":47.1949577331543,"p95":53.31599712371826,"p99":65.62197208404541},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.769014358520508,"p90":7.3310136795043945,"p95":7.3310136795043945,"p99":7.3310136795043945},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":361.58400774002075,"p90":367.87399649620056,"p95":373.1609880924225,"p99":383.6599886417389},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":396.19800448417664,"p90":424.8009920120239,"p95":430.56800961494446,"p99":441.3059949874878},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":820.7550048828125,"p90":844.7510004043579,"p95":850.4400253295898,"p99":861.2160086631775},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":31.651973724365234,"p90":36.858975887298584,"p95":39.3030047416687,"p99":46.07301950454712},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"roundtrip":{"activation_data_bytes":1215807488,"scale_bytes":0,"total_logical_bytes":1215807488},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"roundtrip":{"activation_data_bytes":1215807488,"scale_bytes":0,"total_logical_bytes":1215807488},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":42404,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":2702,"mean":2650.25,"min":2578,"total":42404},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.017352504655718803,"expert_assignments_per_rank":[3108,3051,3103,2954,3171,3067,3121,2998,3045,3065,3095,3023,3150,3071,3090,3040],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,3,103,1301,3825,2960,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.17626953125,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.06268276274204254,"same_node_fraction":0.49721723794937134,"same_scaleup_domain_fraction":0.49721723794937134,"cross_node_fraction":0.5027827620506287,"cross_domain_fraction":0.5027827620506287,"gpus_per_node":8,"scale_up_domain":8,"copies":42404},"payload_copies_per_rank":[2682,2637,2676,2578,2702,2659,2687,2602,2635,2661,2672,2599,2691,2625,2662,2636],"payload_rank_cv":0.013201921246945858,"routed_copies":42404},"token_rate_at_latency_percentile":{"p50":9609079.57768389,"p90":9336208.043601185,"p95":9277757.370756427,"p99":9177816.777622892},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"InterNodeV1","block_num":32,"rdma_block_num":16,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":0,"scale_type_size":0,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"inter-node-v1-vllm-cudagraph-spin-ar","rank_alignment":{"method":"in-graph-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":true,"cuda_graph_supported":true,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9948,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2077,"p50":2096,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2079,"p50":2098,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2080,"p50":2098,"max":2100,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2080,"p50":2098,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2072,"p50":2093,"max":2094,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2077,"p50":2096,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2094,"max":2097,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2077,"p50":2097,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2080,"p50":2099,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2071,"p50":2094,"max":2095,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2076,"p50":2094,"max":2095,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2069,"p50":2088,"max":2093,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2081,"p50":2098,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2096,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2072,"p50":2088,"max":2091,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2077,"p50":2097,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":887.6}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-fp8.json b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-fp8.json new file mode 100644 index 0000000000..f33fef3201 --- /dev/null +++ b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T10:55:51.673427+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":2,"phase":"decode","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-mori-deepseek-v4-pro-normal-decode-ep16-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":132.88000226020813,"p90":135.68399846553802,"p95":136.68499886989594,"p99":138.68899643421173},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":84.56700295209885,"p90":86.41000092029572,"p95":87.57100254297256,"p99":90.45600146055222},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":217.44700521230698,"p90":222.09399938583374,"p95":224.2560014128685,"p99":229.14499789476395},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":199.3809938430786,"p90":200.94400644302368,"p95":202.06399261951447,"p99":205.5490016937256},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":207.07200467586517,"p90":209.19500291347504,"p95":209.95600521564484,"p99":212.67999708652496},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":121.02299928665161,"p90":123.22600185871124,"p95":124.06700104475021,"p99":125.9090006351471},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":72.10899889469147,"p90":74.07200336456299,"p95":74.35200363397598,"p99":74.91199672222137},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009995937347412,"p90":6.089001893997192,"p95":6.089001893997192,"p99":6.089001893997192},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.408989429473877,"p90":10.857000946998596,"p95":12.338995933532715,"p99":15.182003378868103},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.7610023021697998,"p90":0.7219910621643066,"p95":0.7219910621643066,"p99":0.7219910621643066},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":119.41999942064285,"p90":121.74399942159653,"p95":122.66500294208527,"p99":123.94600361585617},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":68.54300200939178,"p90":69.58500295877457,"p95":70.10599970817566,"p99":71.74800336360931},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":189.36599791049957,"p90":191.12899899482727,"p95":191.72999262809753,"p99":193.65200400352478},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.784997820854187,"p90":20.63000202178955,"p95":21.751999855041504,"p99":24.877995252609253},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":587776,"scale_bytes":18368,"total_logical_bytes":606144},"roundtrip":{"activation_data_bytes":1763328,"scale_bytes":18368,"total_logical_bytes":1781696},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":587776,"scale_bytes":18368,"total_logical_bytes":606144},"roundtrip":{"activation_data_bytes":1763328,"scale_bytes":18368,"total_logical_bytes":1781696},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":9,"mean":5.125,"min":2,"total":82},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":77267.80848548401,"p90":76483.66250229097,"p95":76206.44136168658,"p99":75230.394109375},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":139.41000401973724,"p90":142.0540064573288,"p95":143.0550068616867,"p99":145.0989991426468},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":88.45300227403641,"p90":90.01599997282028,"p95":91.17700159549713,"p99":94.42199766635895},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":227.86300629377365,"p90":232.07000643014908,"p95":234.23200845718384,"p99":239.52099680900574},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":207.31300115585327,"p90":209.3150019645691,"p95":210.0359946489334,"p99":213.80199491977692},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":212.5599980354309,"p90":214.56299722194672,"p95":215.52500128746033,"p99":217.44699776172638},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":125.109001994133,"p90":126.43100321292877,"p95":126.95099413394928,"p99":128.75400483608246},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":77.27599889039993,"p90":79.03899997472763,"p95":79.71999794244766,"p99":80.96200227737427},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048992276191711,"p90":6.050005555152893,"p95":6.050005555152893,"p99":6.050005555152893},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.529003381729126,"p90":10.857000946998596,"p95":12.498006224632263,"p99":14.822989702224731},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.6809979677200317,"p90":0.6409883499145508,"p95":0.6409883499145508,"p99":0.6409883499145508},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":125.18900632858276,"p90":126.47099792957306,"p95":126.8710047006607,"p99":128.31400334835052},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":70.90699672698975,"p90":72.14800268411636,"p95":72.75000214576721,"p99":74.35200363397598},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":193.8530057668686,"p90":195.77600061893463,"p95":196.57699763774872,"p99":198.50000739097595},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.706992268562317,"p90":21.071001887321472,"p95":21.991997957229614,"p99":24.195000529289246},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":1189888,"scale_bytes":37184,"total_logical_bytes":1227072},"roundtrip":{"activation_data_bytes":3569664,"scale_bytes":37184,"total_logical_bytes":3606848},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":1189888,"scale_bytes":37184,"total_logical_bytes":1227072},"roundtrip":{"activation_data_bytes":3569664,"scale_bytes":37184,"total_logical_bytes":3606848},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":15,"mean":10.375,"min":7,"total":166},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":150545.72965636756,"p90":149140.34765695778,"p95":148474.65402549485,"p99":147162.29853430716},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":140.2910053730011,"p90":143.41700077056885,"p95":144.49800550937653,"p99":146.8220055103302},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":98.86900335550308,"p90":102.31500118970871,"p95":103.3559963107109,"p99":106.08000308275223},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":239.16000872850418,"p90":245.73200196027756,"p95":247.85400182008743,"p99":252.90200859308243},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":213.8029932975769,"p90":216.28700196743011,"p95":217.3279970884323,"p99":218.8889980316162},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":220.73200345039368,"p90":223.01599383354187,"p95":223.9380031824112,"p99":226.1810004711151},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":127.31100618839264,"p90":129.47499752044678,"p95":129.99500334262848,"p99":130.55700063705444},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":80.68200200796127,"p90":83.1649973988533,"p95":83.68600159883499,"p99":84.64699983596802},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049007177352905,"p90":6.089001893997192,"p95":6.089001893997192,"p99":6.089001893997192},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.130995392799377,"p90":11.336997151374817,"p95":12.256994843482971,"p99":15.623003244400024},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.6410032510757446,"p90":0.8819997310638428,"p95":0.8819997310638428,"p99":0.8819997310638428},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":125.58899819850922,"p90":128.75400483608246,"p95":129.23599779605865,"p99":130.1569938659668},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":74.59300011396408,"p90":76.43599808216095,"p95":76.95599645376205,"p99":78.79800349473953},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":207.39300549030304,"p90":209.83700454235077,"p95":210.63700318336487,"p99":212.0400071144104},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.340011239051819,"p90":16.423001885414124,"p95":17.34599471092224,"p99":19.748002290725708},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":2358272,"scale_bytes":73696,"total_logical_bytes":2431968},"roundtrip":{"activation_data_bytes":7074816,"scale_bytes":73696,"total_logical_bytes":7148512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":2358272,"scale_bytes":73696,"total_logical_bytes":2431968},"roundtrip":{"activation_data_bytes":7074816,"scale_bytes":73696,"total_logical_bytes":7148512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":26,"mean":20.5625,"min":15,"total":329},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":289944.36239230295,"p90":286974.9335008202,"p95":285793.38518021925,"p99":282959.2223338549},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":148.26400578022003,"p90":151.1480063199997,"p95":152.1490067243576,"p99":155.51500022411346},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":105.92000186443329,"p90":108.00299793481827,"p95":109.04400050640106,"p99":111.44699901342392},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":254.18400764465332,"p90":259.15100425481796,"p95":261.19300723075867,"p99":266.9619992375374},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":230.62799870967865,"p90":233.55300724506378,"p95":234.2730015516281,"p99":238.6000007390976},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":237.47800290584564,"p90":239.72100019454956,"p95":240.76299369335175,"p99":242.9260015487671},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":137.28800415992737,"p90":139.5699977874756,"p95":139.97100293636322,"p99":141.5330022573471},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":88.53299915790558,"p90":90.61700105667114,"p95":91.09699726104736,"p99":91.89800173044205},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.0079991817474365,"p90":6.09000027179718,"p95":6.09000027179718,"p99":6.09000027179718},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.732003927230835,"p90":12.018993496894836,"p95":13.499990105628967,"p99":16.145989298820496},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.3229995965957642,"p90":1.681998372077942,"p95":1.681998372077942,"p99":1.681998372077942},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":137.60800659656525,"p90":139.2499953508377,"p95":139.69099521636963,"p99":140.37099480628967},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":82.1240022778511,"p90":83.52600038051605,"p95":84.08699929714203,"p99":85.92899888753891},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":221.73500061035156,"p90":223.97799789905548,"p95":224.77999329566956,"p99":226.9420027732849},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":15.744000673294067,"p90":18.987998366355896,"p95":19.829005002975464,"p99":21.39100432395935},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":4652032,"scale_bytes":145376,"total_logical_bytes":4797408},"roundtrip":{"activation_data_bytes":13956096,"scale_bytes":145376,"total_logical_bytes":14101472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":4652032,"scale_bytes":145376,"total_logical_bytes":4797408},"roundtrip":{"activation_data_bytes":13956096,"scale_bytes":145376,"total_logical_bytes":14101472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":49,"mean":40.5625,"min":36,"total":649},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":538997.2899963664,"p90":533954.0544888411,"p95":531643.1650747268,"p99":526909.4258496003},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":162.60500252246857,"p90":165.69000482559204,"p95":167.13200509548187,"p99":169.45600509643555},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":113.81100118160248,"p90":115.81499874591827,"p95":116.89600348472595,"p99":119.3000003695488},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":276.41600370407104,"p90":281.5050035715103,"p95":284.0280085802078,"p99":288.75600546598434},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":255.7860016822815,"p90":258.19000601768494,"p95":259.2709958553314,"p99":262.8769874572754},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":260.5929970741272,"p90":262.71599531173706,"p95":263.79698514938354,"p99":266.60099625587463},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":153.71200442314148,"p90":155.27400374412537,"p95":155.99599480628967,"p99":157.39700198173523},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":96.46499902009964,"p90":98.38800132274628,"p95":98.66800159215927,"p99":99.78999942541122},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009012460708618,"p90":6.049007177352905,"p95":6.049007177352905,"p99":6.049007177352905},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":10.3759765625,"p95":11.537984013557434,"p99":15.943005681037903},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.8009970188140869,"p90":0.5609989166259766,"p95":0.5609989166259766,"p99":0.5609989166259766},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":150.10599792003632,"p90":152.30900049209595,"p95":152.83100306987762,"p99":153.9520025253296},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":91.89900010824203,"p90":93.2610034942627,"p95":93.90199929475784,"p99":96.38500213623047},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":243.44700574874878,"p90":245.21000683307648,"p95":245.60999870300293,"p99":246.73299491405487},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.22601056098938,"p90":19.94900405406952,"p95":21.029993891716003,"p99":22.553011775016785},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":9418752,"scale_bytes":294336,"total_logical_bytes":9713088},"roundtrip":{"activation_data_bytes":28256256,"scale_bytes":294336,"total_logical_bytes":28550592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":9418752,"scale_bytes":294336,"total_logical_bytes":9713088},"roundtrip":{"activation_data_bytes":28256256,"scale_bytes":294336,"total_logical_bytes":28550592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":89,"mean":82.125,"min":72,"total":1314},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":982374.8253955548,"p90":974436.2907794483,"p95":970443.2363206568,"p99":960236.4717133308},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":169.49599981307983,"p90":171.90000414848328,"p95":173.02100360393524,"p99":175.54500699043274},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":129.23499941825867,"p90":131.1579942703247,"p95":131.9579929113388,"p99":134.16199386119843},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":298.7309992313385,"p90":303.057998418808,"p95":304.97899651527405,"p99":309.70700085163116},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":271.6900110244751,"p90":274.4539976119995,"p95":275.2160131931305,"p99":277.21700072288513},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":273.171991109848,"p90":275.41500329971313,"p95":276.41600370407104,"p99":278.3389985561371},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":154.55299615859985,"p90":156.03500604629517,"p95":156.63599967956543,"p99":158.27900171279907},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":104.63699698448181,"p90":107.04100131988525,"p95":107.80200362205505,"p99":110.68599671125412},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049007177352905,"p90":6.049990653991699,"p95":6.049990653991699,"p99":6.049990653991699},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.930006504058838,"p90":9.535014629364014,"p95":11.056005954742432,"p99":13.89998197555542},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.6020069122314453,"p90":1.202017068862915,"p95":1.202017068862915,"p99":1.202017068862915},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":148.50400388240814,"p90":150.22599697113037,"p95":150.98699927330017,"p99":153.07100117206573},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":101.95399820804596,"p90":103.75600308179855,"p95":104.63699698448181,"p99":106.72099888324738},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":261.5149915218353,"p90":263.357013463974,"p95":263.8390064239502,"p99":264.67999815940857},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11.49800419807434,"p90":14.461010694503784,"p95":15.542984008789062,"p99":17.825007438659668},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":18808832,"scale_bytes":587776,"total_logical_bytes":19396608},"roundtrip":{"activation_data_bytes":56426496,"scale_bytes":587776,"total_logical_bytes":57014272},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":18808832,"scale_bytes":587776,"total_logical_bytes":19396608},"roundtrip":{"activation_data_bytes":56426496,"scale_bytes":587776,"total_logical_bytes":57014272},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":178,"mean":164.0,"min":148,"total":2624},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":1874277.0732820642,"p90":1859012.7402857188,"p95":1852280.5956928004,"p99":1839483.5170635881},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":178.74999344348907,"p90":182.83599615097046,"p95":183.9579939842224,"p99":186.60199642181396},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":135.36399602890015,"p90":138.7690007686615,"p95":139.85100388526917,"p99":141.97400212287903},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":314.1139894723892,"p90":321.60499691963196,"p95":323.8089978694916,"p99":328.575998544693},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":281.8639874458313,"p90":284.22901034355164,"p95":284.94900465011597,"p99":286.8730127811432},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":286.8730127811432,"p90":290.03700613975525,"p95":291.1190092563629,"p99":293.04200410842896},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":159.24100577831268,"p90":163.0460023880005,"p95":163.72700035572052,"p99":164.64799642562866},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":113.9720007777214,"p90":118.69999766349792,"p95":119.86000090837479,"p99":121.22300267219543},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049007177352905,"p90":6.089985370635986,"p95":6.089985370635986,"p99":6.089985370635986},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.850987672805786,"p90":10.976999998092651,"p95":12.860983610153198,"p99":17.425984144210815},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.719994306564331,"p90":1.0809898376464844,"p95":1.0809898376464844,"p99":1.0809898376464844},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":152.78999507427216,"p90":156.7160040140152,"p95":158.11799466609955,"p99":160.4820042848587},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":108.84500294923782,"p90":111.24800145626068,"p95":112.08900064229965,"p99":113.97100239992142},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":275.89499950408936,"p90":278.18000316619873,"p95":278.93999218940735,"p99":280.38299083709717},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11.137992143630981,"p90":13.740986585617065,"p95":14.542996883392334,"p99":16.62600040435791},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":37675008,"scale_bytes":1177344,"total_logical_bytes":38852352},"roundtrip":{"activation_data_bytes":113025024,"scale_bytes":1177344,"total_logical_bytes":114202368},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":37675008,"scale_bytes":1177344,"total_logical_bytes":38852352},"roundtrip":{"activation_data_bytes":113025024,"scale_bytes":1177344,"total_logical_bytes":114202368},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":363,"mean":328.5,"min":292,"total":5256},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":3569523.6372102196,"p90":3530583.9541957704,"p95":3517461.819534613,"p99":3494379.596247602},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":202.30600237846375,"p90":208.515003323555,"p95":210.11699736118317,"p99":213.0420058965683},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":149.1049975156784,"p90":154.231995344162,"p95":156.1560034751892,"p99":159.52099859714508},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":351.41099989414215,"p90":362.746998667717,"p95":366.2730008363724,"p99":372.5630044937134},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":316.7169988155365,"p90":320.80399990081787,"p95":321.60601019859314,"p99":324.33000206947327},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":323.5290050506592,"p90":328.93699407577515,"p95":330.5389881134033,"p99":333.22298526763916},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":162.20399737358093,"p90":168.49400103092194,"p95":170.01600563526154,"p99":172.1390038728714},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":109.00499671697617,"p90":113.0099967122078,"p95":114.05199766159058,"p99":117.0559972524643},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.248980760574341,"p90":6.289005279541016,"p95":6.289005279541016,"p99":6.289005279541016},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.414970874786377,"p90":14.622002840042114,"p95":16.26601815223694,"p99":18.267005681991577},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.882016658782959,"p90":2.003997564315796,"p95":2.003997564315796,"p99":2.003997564315796},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":173.58100414276123,"p90":178.62899601459503,"p95":180.15199899673462,"p99":183.07599425315857},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":124.22700226306915,"p90":127.51300632953644,"p95":128.6340057849884,"p99":130.9570074081421},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":310.9480142593384,"p90":314.31299448013306,"p95":315.07399678230286,"p99":316.6370093822479},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":12.460976839065552,"p90":18.709003925323486,"p95":20.671993494033813,"p99":24.517983198165894},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00495867757126689,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":75787264,"scale_bytes":2368352,"total_logical_bytes":78155616},"roundtrip":{"activation_data_bytes":227361792,"scale_bytes":2368352,"total_logical_bytes":229730144},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":75787264,"scale_bytes":2368352,"total_logical_bytes":78155616},"roundtrip":{"activation_data_bytes":227361792,"scale_bytes":2368352,"total_logical_bytes":229730144},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":692,"mean":660.8125,"min":603,"total":10573},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":6330189.775965583,"p90":6226116.359317782,"p95":6195940.792610401,"p99":6146034.609092408},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":271.8900144100189,"p90":283.8680148124695,"p95":287.9149913787842,"p99":294.08299922943115},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":226.1810004711151,"p90":237.99799382686615,"p95":240.48200249671936,"p99":246.05000019073486},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":498.07101488113403,"p90":521.8660086393356,"p95":528.3969938755035,"p99":540.132999420166},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":461.6149961948395,"p90":472.31099009513855,"p95":474.3939936161041,"p99":478.08000445365906},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":466.1419987678528,"p90":475.75798630714417,"p95":478.56199741363525,"p99":483.57000946998596},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":220.1710045337677,"p90":224.13800656795502,"p95":224.77899491786957,"p99":226.86199843883514},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":170.93800008296967,"p90":176.50699615478516,"p95":178.6690056324005,"p99":182.55600333213806},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.410002708435059,"p90":6.570011377334595,"p95":6.570011377334595,"p99":6.570011377334595},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.985999584197998,"p90":24.556010961532593,"p95":26.078999042510986,"p99":29.284000396728516},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.2430121898651123,"p90":4.326999187469482,"p95":4.326999187469482,"p99":4.326999187469482},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":230.42699694633484,"p90":242.20499396324158,"p95":243.96799504756927,"p99":246.9329982995987},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":196.49699330329895,"p90":204.1880041360855,"p95":207.15199410915375,"p99":213.96300196647644},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":444.95001435279846,"p90":453.002005815506,"p95":455.2049934864044,"p99":459.852010011673},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":21.434009075164795,"p90":28.365015983581543,"p95":30.04598617553711,"p99":34.37000513076782},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00495867757126689,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":151825408,"scale_bytes":4744544,"total_logical_bytes":156569952},"roundtrip":{"activation_data_bytes":455476224,"scale_bytes":4744544,"total_logical_bytes":460220768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":151825408,"scale_bytes":4744544,"total_logical_bytes":156569952},"roundtrip":{"activation_data_bytes":455476224,"scale_bytes":4744544,"total_logical_bytes":460220768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":1373,"mean":1323.8125,"min":1247,"total":21181},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":8787022.00365319,"p90":8609419.322192244,"p95":8558974.640979916,"p99":8470335.049291821},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":429.3679893016815,"p90":453.08399200439453,"p95":460.4550004005432,"p99":472.8330075740814},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":377.24798917770386,"p90":397.1180021762848,"p95":402.6469886302948,"p99":412.8609895706177},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":806.6159784793854,"p90":850.2019941806793,"p95":863.101989030838,"p99":885.6939971446991},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":741.9589757919312,"p90":762.1110081672668,"p95":768.2790160179138,"p99":782.2610139846802},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":761.5500092506409,"p90":778.4950137138367,"p95":782.5809717178345,"p99":792.1159863471985},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":361.54401302337646,"p90":366.9930100440979,"p95":372.16100096702576,"p99":380.6929886341095},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":304.41999435424805,"p90":328.3349871635437,"p95":339.31198716163635,"p99":351.73100233078003},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.25002384185791,"p90":6.290018558502197,"p95":6.290018558502197,"p99":6.290018558502197},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":24.317026138305664,"p90":40.583014488220215,"p95":44.82501745223999,"p99":62.97200918197632},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-5.127966403961182,"p90":5.928993225097656,"p95":5.928993225097656,"p99":5.928993225097656},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":362.86699771881104,"p90":375.3249943256378,"p95":379.81098890304565,"p99":385.86199283599854},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":333.22298526763916,"p90":349.2079973220825,"p95":355.1360070705414,"p99":362.6269996166229},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":731.141984462738,"p90":746.9249963760376,"p95":751.0510087013245,"p99":760.2249979972839},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":30.36898374557495,"p90":35.85696220397949,"p95":38.10000419616699,"p99":44.745028018951416},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.005181347019970417,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":303951872,"scale_bytes":9498496,"total_logical_bytes":313450368},"roundtrip":{"activation_data_bytes":911855616,"scale_bytes":9498496,"total_logical_bytes":921354112},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":303951872,"scale_bytes":9498496,"total_logical_bytes":313450368},"roundtrip":{"activation_data_bytes":911855616,"scale_bytes":9498496,"total_logical_bytes":921354112},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":42404,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":2702,"mean":2650.25,"min":2578,"total":42404},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.017352504655718803,"expert_assignments_per_rank":[3108,3051,3103,2954,3171,3067,3121,2998,3045,3065,3095,3023,3150,3071,3090,3040],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,3,103,1301,3825,2960,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.17626953125,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.06268276274204254,"same_node_fraction":0.49721723794937134,"same_scaleup_domain_fraction":0.49721723794937134,"cross_node_fraction":0.5027827620506287,"cross_domain_fraction":0.5027827620506287,"gpus_per_node":8,"scale_up_domain":8,"copies":42404},"payload_copies_per_rank":[2682,2637,2676,2578,2702,2659,2687,2602,2635,2661,2672,2599,2691,2625,2662,2636],"payload_rank_cv":0.013201921246945858,"routed_copies":42404},"token_rate_at_latency_percentile":{"p50":10757008.601524228,"p90":10522867.655785987,"p95":10467926.382132491,"p99":10341919.796085646},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"InterNodeV1","block_num":32,"rdma_block_num":16,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":56,"scale_type_size":4,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"inter-node-v1-vllm-cudagraph-spin-ar","rank_alignment":{"method":"in-graph-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":true,"cuda_graph_supported":true,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9948,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2078,"p50":2095,"max":2097,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.6},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2082,"p50":2099,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.6},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2081,"p50":2098,"max":2100,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.6},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2080,"p50":2098,"max":2100,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.6},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2093,"max":2094,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.6},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2080,"p50":2097,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.6},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2078,"p50":2095,"max":2096,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.6},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2079,"p50":2097,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.6},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2081,"p50":2100,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.6},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2077,"p50":2094,"max":2096,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.7},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2078,"p50":2094,"max":2095,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.6},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2072,"p50":2089,"max":2093,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.6},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2084,"p50":2099,"max":2100,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.7},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2080,"p50":2097,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.7},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2071,"p50":2089,"max":2091,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.7},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2080,"p50":2097,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":909.6}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-bf16.json b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-bf16.json new file mode 100644 index 0000000000..765f0e8a6d --- /dev/null +++ b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T06:52:26.741047+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":1,"phase":"decode","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.282999992370605,"p90":32.12900087237358,"p95":33.13000127673149,"p99":35.0940003991127},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.803000062704086,"p90":30.32599948346615,"p95":31.44799917936325,"p99":33.65100175142288},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":56.08600005507469,"p90":62.45500035583973,"p95":64.57800045609474,"p99":68.74500215053558},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":45.708999037742615,"p90":49.43399876356125,"p95":51.718998700380325,"p99":54.28199842572212},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":48.79400134086609,"p90":51.35799944400787,"p95":52.358999848365784,"p99":53.92099916934967},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":27.28099934756756,"p90":27.962999418377876,"p95":28.123000636696815,"p99":28.442999348044395},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":23.91600050032139,"p90":24.35700036585331,"p95":24.476999416947365,"p99":24.75699968636036},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048999726772308,"p90":6.049998104572296,"p95":6.049998104572296,"p99":6.049998104572296},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.2420006096363068,"p90":6.250001490116119,"p95":8.090998977422714,"p99":10.656002908945084},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.1619979739189148,"p90":0.39999932050704956,"p95":0.39999932050704956,"p99":0.39999932050704956},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.4020005017519,"p90":27.761999517679214,"p95":27.88200043141842,"p99":28.24299968779087},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":24.075999855995178,"p90":24.5169997215271,"p95":24.59700033068657,"p99":25.036999955773354},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":45.26799917221069,"p90":45.74999958276749,"p95":45.949000865221024,"p99":46.27000167965889},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.4850016236305237,"p90":6.289001554250717,"p95":7.130000740289688,"p99":8.772999048233032},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8,"byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"roundtrip":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"roundtrip":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":35,"assignments":48,"wire":"rank-deduplicated"},"receive":{"max":6,"mean":4.375,"min":2,"total":35},"routing":{"empty_expert_count":337,"empty_rank_count":0,"expert_assignment_rank_cv":0.3227486312389374,"expert_assignments_per_rank":[8,6,6,6,6,5,9,2],"expert_load_cv":2.708012819290161,"expert_load_max":2,"expert_load_mean":0.125,"expert_load_min":0,"fanout_histogram":[0,0,0,5,3,0,0,0],"fanout_max":5,"fanout_mean":4.375,"fanout_min":4,"hotspot_ratio":16.0,"locality":{"placement":"packed","local_rank_fraction":0.08571428805589676,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":35},"payload_copies_per_rank":[5,5,5,4,4,4,6,2],"payload_rank_cv":0.2539483904838562,"routed_copies":35},"token_rate_at_latency_percentile":{"p50":163954.58007457605,"p90":155769.3073446495,"p95":152791.30661716976,"p99":148365.20322767762},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.000999242067337,"p90":31.808000057935715,"p95":32.930001616477966,"p99":34.97299924492836},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":29.525000602006912,"p90":31.968001276254654,"p95":33.3699993789196,"p99":35.413000732660294},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":57.52599984407425,"p90":63.77600133419037,"p95":66.30000099539757,"p99":70.38599997758865},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":47.71199822425842,"p90":51.83799937367439,"p95":53.60100045800209,"p99":56.04499951004982},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":50.195999443531036,"p90":52.83999815583229,"p95":53.8019984960556,"p99":55.48400059342384},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":27.000999078154564,"p90":27.441000565886497,"p95":27.641000226140022,"p99":27.883000671863556},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":25.95899999141693,"p90":26.559999212622643,"p95":26.639999821782112,"p99":26.840999722480774},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009001284837723,"p90":6.048001348972321,"p95":6.048001348972321,"p99":6.048001348972321},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.0819993913173676,"p90":6.209999322891235,"p95":7.651001214981079,"p99":10.336000472307205},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.2809986472129822,"p90":0.2410002052783966,"p95":0.2410002052783966,"p99":0.2410002052783966},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":26.799999177455902,"p90":27.161000296473503,"p95":27.320999652147293,"p99":27.80199982225895},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":25.638999417424202,"p90":26.239000260829926,"p95":26.399999856948853,"p99":26.799999177455902},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":46.831000596284866,"p90":47.5119985640049,"p95":47.75200039148331,"p99":48.31299930810928},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.3250004053115845,"p90":6.048999726772308,"p95":6.970997899770737,"p99":8.533000946044922},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"roundtrip":{"activation_data_bytes":2007040,"scale_bytes":0,"total_logical_bytes":2007040},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"roundtrip":{"activation_data_bytes":2007040,"scale_bytes":0,"total_logical_bytes":2007040},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":70,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":12,"mean":8.75,"min":5,"total":70},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.33592739701271057,"expert_assignments_per_rank":[18,17,13,7,13,7,13,8],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,3,6,5,2,0,0],"fanout_max":6,"fanout_mean":4.375,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.15714286267757416,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":70},"payload_copies_per_rank":[12,11,10,5,9,6,10,7],"payload_rank_cv":0.26649653911590576,"routed_copies":70},"token_rate_at_latency_percentile":{"p50":318750.50158129656,"p90":302800.91897077364,"p95":297386.7225614865,"p99":288371.41930779186},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.363000601530075,"p90":32.36899897456169,"p95":33.45099836587906,"p99":36.614999175071716},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":31.647000461816788,"p90":34.331999719142914,"p95":35.53399816155434,"p99":38.01799938082695},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":60.01000106334686,"p90":66.7009986937046,"p95":68.9849965274334,"p99":74.63299855589867},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":50.11599883437157,"p90":53.881000727415085,"p95":55.56400120258331,"p99":57.88800120353699},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":52.51999944448471,"p90":55.323999375104904,"p95":56.48500099778175,"p99":58.28800052404404},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":27.4020005017519,"p90":27.922000735998154,"p95":28.04199978709221,"p99":28.24299968779087},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.282999992370605,"p90":28.88300083577633,"p95":29.044000431895256,"p99":29.203999787569046},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009999662637711,"p90":6.049003452062607,"p95":6.049003452062607,"p99":6.049003452062607},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.1220015585422516,"p90":5.769003182649612,"p95":6.730999797582626,"p99":8.772000670433044},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.23899972438812256,"p90":0.3210008144378662,"p95":0.3210008144378662,"p99":0.3210008144378662},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.401000261306763,"p90":27.761999517679214,"p95":27.883000671863556,"p99":28.40300090610981},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.842000126838684,"p90":28.363000601530075,"p95":28.522999957203865,"p99":29.00400012731552},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":48.9949993789196,"p90":49.674998968839645,"p95":50.074998289346695,"p99":50.75699836015701},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.5250000655651093,"p90":6.087999790906906,"p95":7.250998169183731,"p99":9.053997695446014},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"roundtrip":{"activation_data_bytes":4100096,"scale_bytes":0,"total_logical_bytes":4100096},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"roundtrip":{"activation_data_bytes":4100096,"scale_bytes":0,"total_logical_bytes":4100096},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":143,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":22,"mean":17.875,"min":13,"total":143},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.20518453419208527,"expert_assignments_per_rank":[31,30,24,17,22,22,28,18],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,4,12,13,3,0,0],"fanout_max":6,"fanout_mean":4.46875,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.1258741319179535,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":143},"payload_copies_per_rank":[22,21,18,13,16,15,22,16],"payload_rank_cv":0.1800614446401596,"routed_copies":143},"token_rate_at_latency_percentile":{"p50":609291.7048452182,"p90":578410.8228155248,"p95":566522.0755020733,"p99":548998.0735709037},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.524000197649002,"p90":32.28899836540222,"p95":33.28999876976013,"p99":35.454001277685165},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":31.727999448776245,"p90":34.3719981610775,"p95":35.37299856543541,"p99":37.85699978470802},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":60.25199964642525,"p90":66.66099652647972,"p95":68.66299733519554,"p99":73.31100106239319},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":50.195999443531036,"p90":53.84200066328049,"p95":55.44300004839897,"p99":58.768998831510544},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":52.960000932216644,"p90":55.28400093317032,"p95":56.366000324487686,"p99":58.24799835681915},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":27.64200046658516,"p90":28.08300033211708,"p95":28.201999142766,"p99":28.762999922037125},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.082000091671944,"p90":28.643999248743057,"p95":28.804000467061996,"p99":29.16399948298931},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048999726772308,"p90":6.088998168706894,"p95":6.088998168706894,"p99":6.088998168706894},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.082003116607666,"p90":5.929000675678253,"p95":6.929997354745865,"p99":10.494999587535858},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.19999966025352478,"p90":0.19999966025352478,"p95":0.19999966025352478,"p99":0.19999966025352478},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.441000565886497,"p90":27.842000126838684,"p95":28.04199978709221,"p99":28.603000566363335},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.96199917793274,"p90":28.442999348044395,"p95":28.603000566363335,"p99":29.044000431895256},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":49.59600046277046,"p90":50.15600100159645,"p95":50.31700059771538,"p99":50.957001745700836},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.2860003411769867,"p90":5.688000470399857,"p95":6.730999797582626,"p99":8.533000946044922},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"roundtrip":{"activation_data_bytes":8056832,"scale_bytes":0,"total_logical_bytes":8056832},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"roundtrip":{"activation_data_bytes":8056832,"scale_bytes":0,"total_logical_bytes":8056832},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":281,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":42,"mean":35.125,"min":26,"total":281},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.1458333283662796,"expert_assignments_per_rank":[55,56,58,37,46,43,46,43],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,1,9,24,24,6,0,0],"fanout_max":6,"fanout_mean":4.390625,"fanout_min":2,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.11743772029876709,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":281},"payload_copies_per_rank":[38,42,42,26,29,31,38,35],"payload_rank_cv":0.15879225730895996,"routed_copies":281},"token_rate_at_latency_percentile":{"p50":1208459.193229876,"p90":1157658.6158691002,"p95":1135436.249362469,"p99":1098750.2026755475},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":30.686000362038612,"p90":34.41200032830238,"p95":35.374000668525696,"p99":37.6570001244545},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":32.04900026321411,"p90":34.69200059771538,"p95":35.73400154709816,"p99":38.65800052881241},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":62.735000625252724,"p90":69.10400092601776,"p95":71.10800221562386,"p99":76.3150006532669},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":52.07899957895279,"p90":55.96499890089035,"p95":58.00800025463104,"p99":60.01099944114685},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":55.48400059342384,"p90":57.92799964547157,"p95":58.96899849176407,"p99":60.892000794410706},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":29.444999992847443,"p90":29.924999922513962,"p95":30.16499988734722,"p99":30.446000397205353},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.123000636696815,"p90":28.64300087094307,"p95":28.723999857902527,"p99":29.044000431895256},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048999726772308,"p90":6.048999726772308,"p95":6.048999726772308,"p99":6.048999726772308},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.1210031807422638,"p90":6.328999996185303,"p95":8.813001215457916,"p99":11.737998574972153},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.2400018274784088,"p90":0.19999966025352478,"p95":0.19999966025352478,"p99":0.19999966025352478},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":29.485000297427177,"p90":29.965000227093697,"p95":30.124999582767487,"p99":30.486000701785088},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.32300029695034,"p90":28.683999553322792,"p95":28.84400077164173,"p99":29.245000332593918},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":51.798999309539795,"p90":52.358999848365784,"p95":52.560001611709595,"p99":53.36099863052368},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.6449991166591644,"p90":6.009001284837723,"p95":7.089998573064804,"p99":8.973997086286545},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"roundtrip":{"activation_data_bytes":16171008,"scale_bytes":0,"total_logical_bytes":16171008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"roundtrip":{"activation_data_bytes":16171008,"scale_bytes":0,"total_logical_bytes":16171008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":564,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":77,"mean":70.5,"min":62,"total":564},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.07384087145328522,"expert_assignments_per_rank":[101,108,101,91,83,97,94,93],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,2,14,52,50,10,0,0],"fanout_max":6,"fanout_mean":4.40625,"fanout_min":2,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.14716312289237976,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":564},"payload_copies_per_rank":[73,76,77,63,62,69,74,70],"payload_rank_cv":0.07472094893455505,"routed_copies":564},"token_rate_at_latency_percentile":{"p50":2306971.354462335,"p90":2209639.5660713306,"p95":2170632.0825149706,"p99":2102082.347928846},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":33.65100175142288,"p90":37.33700141310692,"p95":38.378000259399414,"p99":40.421001613140106},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":32.00799971818924,"p90":34.61199998855591,"p95":35.773999989032745,"p99":37.93700039386749},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":65.65900146961212,"p90":71.94900140166283,"p95":74.15200024843216,"p99":78.3580020070076},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":55.68400025367737,"p90":59.93000045418739,"p95":61.25300005078316,"p99":63.85599821805954},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":58.48899856209755,"p90":61.292000114917755,"p95":62.37399950623512,"p99":64.01699781417847},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":32.409001141786575,"p90":32.80999884009361,"p95":32.930001616477966,"p99":33.3699993789196},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.522999957203865,"p90":29.16399948298931,"p95":29.364999383687973,"p99":29.684999957680702},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.047997623682022,"p90":6.088998168706894,"p95":6.088998168706894,"p99":6.088998168706894},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.722000539302826,"p90":5.609001964330673,"p95":6.8509988486766815,"p99":10.69599762558937},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.20100176334381104,"p90":0.3210008144378662,"p95":0.3210008144378662,"p99":0.3210008144378662},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":32.409001141786575,"p90":32.77000039815903,"p95":32.88999944925308,"p99":33.41099992394447},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.1629990786314,"p90":28.883999213576317,"p95":29.16399948298931,"p99":29.565000906586647},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":54.8430010676384,"p90":55.68400025367737,"p95":56.005001068115234,"p99":56.926000863313675},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.605000674724579,"p90":6.250001490116119,"p95":7.2109997272491455,"p99":9.052995592355728},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"roundtrip":{"activation_data_bytes":32628736,"scale_bytes":0,"total_logical_bytes":32628736},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"roundtrip":{"activation_data_bytes":32628736,"scale_bytes":0,"total_logical_bytes":32628736},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1138,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":149,"mean":142.25,"min":126,"total":1138},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.04658475145697594,"expert_assignments_per_rank":[189,194,196,170,193,201,198,195],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,3,26,101,106,20,0,0],"fanout_max":6,"fanout_mean":4.4453125,"fanout_min":2,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.1344463974237442,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":1138},"payload_copies_per_rank":[143,141,149,126,144,145,148,142],"payload_rank_cv":0.04686218500137329,"routed_copies":1138},"token_rate_at_latency_percentile":{"p50":4376891.4888875345,"p90":4176727.786987206,"p95":4104274.2493113554,"p99":3998937.918692919},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":43.063998222351074,"p90":46.74899950623512,"p95":47.47200012207031,"p99":48.87399822473526},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":34.69200059771538,"p90":37.4159999191761,"p95":38.458000868558884,"p99":40.502000600099564},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":77.75599882006645,"p90":84.16499942541122,"p95":85.9300009906292,"p99":89.37599882483482},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":68.42400133609772,"p90":71.62799686193466,"p95":73.51099699735641,"p99":75.314000248909},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":70.70700079202652,"p90":73.43100011348724,"p95":74.51300323009491,"p99":76.03500038385391},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":41.944000869989395,"p90":42.583998292684555,"p95":42.865000665187836,"p99":44.867999851703644},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":31.48699924349785,"p90":32.04900026321411,"p95":32.248999923467636,"p99":32.52900019288063},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.010003387928009,"p90":6.048999726772308,"p95":6.048999726772308,"p99":6.048999726772308},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.5619993209838867,"p90":7.050998508930206,"p95":8.172005414962769,"p99":11.055998504161835},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.20100176334381104,"p90":-0.160999596118927,"p95":-0.160999596118927,"p99":-0.160999596118927},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":41.78300127387047,"p90":42.22400113940239,"p95":42.424000799655914,"p99":43.143998831510544},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":30.806999653577805,"p90":31.408000737428665,"p95":31.567998230457306,"p99":31.92799910902977},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":66.9810026884079,"p90":67.7030012011528,"p95":67.9830014705658,"p99":68.743996322155},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.724999725818634,"p90":6.409995257854462,"p95":7.451996207237244,"p99":9.333997964859009},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.007518797181546688,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"roundtrip":{"activation_data_bytes":64569344,"scale_bytes":0,"total_logical_bytes":64569344},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"roundtrip":{"activation_data_bytes":64569344,"scale_bytes":0,"total_logical_bytes":64569344},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2252,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":290,"mean":281.5,"min":268,"total":2252},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03184119611978531,"expert_assignments_per_rank":[396,393,399,359,380,389,380,376],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,3,57,223,191,38,0,0],"fanout_max":6,"fanout_mean":4.3984375,"fanout_min":2,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.11767318099737167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":2252},"payload_copies_per_rank":[290,286,290,268,282,282,280,274],"payload_rank_cv":0.025306938216090202,"routed_copies":2252},"token_rate_at_latency_percentile":{"p50":7241150.017181003,"p90":6972532.026102144,"p95":6871283.907574528,"p99":6733741.00631587},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":68.38300079107285,"p90":71.22699916362762,"p95":72.42900133132935,"p99":76.67600363492966},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":46.95099964737892,"p90":48.87399822473526,"p95":50.074998289346695,"p99":51.67900025844574},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":115.33400043845177,"p90":120.10099738836288,"p95":122.50399962067604,"p99":128.3550038933754},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":106.84099793434143,"p90":108.28399658203125,"p95":109.6860021352768,"p99":111.52900010347366},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":108.72499644756317,"p90":110.60699820518494,"p95":111.88899725675583,"p99":113.69200050830841},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":66.98200106620789,"p90":67.66299903392792,"p95":67.82200187444687,"p99":68.14300268888474},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":44.14699971675873,"p90":44.90799829363823,"p95":45.06799951195717,"p99":45.42800039052963},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.087996065616608,"p90":6.1289966106414795,"p95":6.1289966106414795,"p99":6.1289966106414795},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.6430020332336426,"p90":5.888998508453369,"p95":6.771005690097809,"p99":9.853996336460114},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.2809986472129822,"p90":0.19999593496322632,"p95":0.19999593496322632,"p99":0.19999593496322632},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":66.7399987578392,"p90":67.34199821949005,"p95":67.50299781560898,"p99":68.06299835443497},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":43.30499842762947,"p90":43.90700161457062,"p95":44.14600133895874,"p99":44.54699903726578},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":104.95900362730026,"p90":105.80000281333923,"p95":106.00099712610245,"p99":106.48100078105927},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.68499755859375,"p90":6.328999996185303,"p95":7.531002163887024,"p99":9.975001215934753},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"roundtrip":{"activation_data_bytes":128651264,"scale_bytes":0,"total_logical_bytes":128651264},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"roundtrip":{"activation_data_bytes":128651264,"scale_bytes":0,"total_logical_bytes":128651264},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":4487,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":569,"mean":560.875,"min":553,"total":4487},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.01615842990577221,"expert_assignments_per_rank":[775,756,779,747,762,761,782,782],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,5,126,437,385,71,0,0],"fanout_max":6,"fanout_mean":4.3818359375,"fanout_min":2,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.12859371304512024,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":4487},"payload_copies_per_rank":[569,555,565,556,553,564,558,567],"payload_rank_cv":0.010142241604626179,"routed_copies":4487},"token_rate_at_latency_percentile":{"p50":9418257.378319286,"p90":9258003.712390758,"p95":9151927.58095945,"p99":9006790.23521244},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":87.29200065135956,"p90":90.29699862003326,"p95":91.45700186491013,"p99":94.54300254583359},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":69.14500147104263,"p90":71.22799754142761,"p95":72.34899699687958,"p99":74.23199713230133},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":156.4370021224022,"p90":161.52499616146088,"p95":163.8059988617897,"p99":168.77499967813492},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":149.46700632572174,"p90":150.70800483226776,"p95":151.06800198554993,"p99":152.6709944009781},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":149.1049975156784,"p90":151.42899751663208,"p95":152.34999358654022,"p99":154.03300523757935},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":86.20999753475189,"p90":86.97199821472168,"p95":87.29200065135956,"p99":87.6929983496666},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":67.90299713611603,"p90":69.06399875879288,"p95":69.34499740600586,"p99":69.90599632263184},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009995937347412,"p90":6.09000027179718,"p95":6.09000027179718,"p99":6.09000027179718},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.042993903160095,"p90":5.969002842903137,"p95":7.250994443893433,"p99":9.613990783691406},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.48100948333740234,"p90":0.3609955310821533,"p95":0.3609955310821533,"p99":0.3609955310821533},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":85.52999794483185,"p90":86.25099807977676,"p95":86.49100363254547,"p99":86.93200349807739},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":65.57899713516235,"p90":66.54000282287598,"p95":66.90099835395813,"p99":67.62100011110306},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":145.94100415706635,"p90":147.1019983291626,"p95":147.42299914360046,"p99":148.14400672912598},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.0039995908737183,"p90":5.970001220703125,"p95":6.970003247261047,"p99":8.251994848251343},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"roundtrip":{"activation_data_bytes":259768320,"scale_bytes":0,"total_logical_bytes":259768320},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"roundtrip":{"activation_data_bytes":259768320,"scale_bytes":0,"total_logical_bytes":259768320},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":9060,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":1149,"mean":1132.5,"min":1088,"total":9060},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.018761977553367615,"expert_assignments_per_rank":[1542,1490,1556,1492,1545,1541,1581,1541],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,11,225,859,791,162,0,0],"fanout_max":6,"fanout_mean":4.423828125,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.12792494893074036,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":9060},"payload_copies_per_rank":[1144,1088,1148,1126,1135,1149,1140,1130],"payload_rank_cv":0.016329573467373848,"routed_copies":9060},"token_rate_at_latency_percentile":{"p50":13735287.442559749,"p90":13524490.246823827,"p95":13442731.120540962,"p99":13295851.735419823},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":162.84699738025665,"p90":165.32999277114868,"p95":166.45100712776184,"p99":168.41499507427216},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":122.42600321769714,"p90":124.02699887752533,"p95":125.22900104522705,"p99":127.15199589729309},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":285.2730005979538,"p90":289.356991648674,"p95":291.6800081729889,"p99":295.56699097156525},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":277.0189940929413,"p90":281.34599328041077,"p95":283.14799070358276,"p99":286.23300790786743},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":278.34099531173706,"p90":281.0249924659729,"p95":282.0670008659363,"p99":284.91100668907166},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":162.04500198364258,"p90":166.0510003566742,"p95":167.89400577545166,"p99":171.25900089740753},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":117.05700308084488,"p90":119.18000131845474,"p95":119.62100118398666,"p99":120.26199698448181},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.090998649597168,"p90":6.209999322891235,"p95":6.209999322891235,"p99":6.209999322891235},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.0039846897125244,"p90":7.12999701499939,"p95":8.573025465011597,"p99":13.981014490127563},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.4010200500488281,"p90":1.0419785976409912,"p95":1.0419785976409912,"p99":1.0419785976409912},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":159.32099521160126,"p90":161.28399968147278,"p95":161.92500293254852,"p99":163.52799534797668},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":118.85999888181686,"p90":119.82099711894989,"p95":120.10200321674347,"p99":120.74200063943863},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":274.09499883651733,"p90":276.4590084552765,"p95":277.33999490737915,"p99":279.7029912471771},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.166007041931152,"p90":6.689995527267456,"p95":7.651001214981079,"p99":9.213000535964966},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"roundtrip":{"activation_data_bytes":520626176,"scale_bytes":0,"total_logical_bytes":520626176},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"roundtrip":{"activation_data_bytes":520626176,"scale_bytes":0,"total_logical_bytes":520626176},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":18158,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":2294,"mean":2269.75,"min":2209,"total":18158},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.016587993130087852,"expert_assignments_per_rank":[3123,2960,3095,3022,3074,3092,3108,3102],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,23,454,1699,1566,354,0,0],"fanout_max":6,"fanout_mean":4.43310546875,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.1253442019224167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":18158},"payload_copies_per_rank":[2289,2209,2294,2272,2259,2292,2267,2276],"payload_rank_cv":0.01135234534740448,"routed_copies":18158},"token_rate_at_latency_percentile":{"p50":14715762.568185661,"p90":14575216.11888648,"p95":14521372.537111456,"p99":14376418.965343926},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"IntraNode","block_num":80,"rdma_block_num":0,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":0,"scale_type_size":0,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"intranode-vllm-cudagraph-spin-ar","rank_alignment":{"method":"in-graph-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":true,"cuda_graph_supported":true,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi325x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9957,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2061,"p50":2099,"max":2100,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":726.4},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2054,"p50":2093,"max":2097,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":70,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":726.4},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2045,"p50":2094,"max":2095,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":70,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":726.4},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2060,"p50":2100,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":726.4},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2058,"p50":2098,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":726.4},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2062,"p50":2102,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":71,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":726.4},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2056,"p50":2096,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":69,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":726.4},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2054,"p50":2096,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":726.4}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-fp8.json b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-fp8.json new file mode 100644 index 0000000000..23ad0acd87 --- /dev/null +++ b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T06:53:13.509284+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":1,"phase":"decode","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-mori-deepseek-v4-pro-normal-decode-ep8-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.800999581813812,"p90":31.60800039768219,"p95":32.568998634815216,"p99":34.3719981610775},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.401000261306763,"p90":30.12399934232235,"p95":31.046999618411064,"p99":33.13000127673149},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":55.201999843120575,"p90":61.73199974000454,"p95":63.61599825322628,"p99":67.50199943780899},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":45.708999037742615,"p90":49.47499930858612,"p95":51.51699855923653,"p99":54.60299924015999},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":47.79199883341789,"p90":50.67700147628784,"p95":51.5579991042614,"p99":53.88199910521507},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":26.96100063621998,"p90":27.442000806331635,"p95":27.64200046658516,"p99":28.28199975192547},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":24.397000670433044,"p90":24.99699965119362,"p95":25.19799955189228,"p99":25.638999417424202},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.008997559547424,"p90":6.048999726772308,"p95":6.048999726772308,"p99":6.048999726772308},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.722000539302826,"p90":6.889998912811279,"p95":8.892998099327087,"p99":13.620998710393906},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.1600012183189392,"p90":0.360000878572464,"p95":0.360000878572464,"p99":0.360000878572464},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":26.760000735521317,"p90":27.080999687314034,"p95":27.161000296473503,"p99":27.481000870466232},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":23.956000804901123,"p90":24.637000635266304,"p95":24.83700029551983,"p99":25.35800077021122},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":44.58700120449066,"p90":45.34799978137016,"p95":45.58899998664856,"p99":46.310000121593475},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.1250007450580597,"p90":5.968999117612839,"p95":6.970997899770737,"p99":8.852999657392502},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8,"byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":250880,"scale_bytes":7840,"total_logical_bytes":258720},"roundtrip":{"activation_data_bytes":752640,"scale_bytes":7840,"total_logical_bytes":760480},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":250880,"scale_bytes":7840,"total_logical_bytes":258720},"roundtrip":{"activation_data_bytes":752640,"scale_bytes":7840,"total_logical_bytes":760480},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":35,"assignments":48,"wire":"rank-deduplicated"},"receive":{"max":6,"mean":4.375,"min":2,"total":35},"routing":{"empty_expert_count":337,"empty_rank_count":0,"expert_assignment_rank_cv":0.3227486312389374,"expert_assignments_per_rank":[8,6,6,6,6,5,9,2],"expert_load_cv":2.708012819290161,"expert_load_max":2,"expert_load_mean":0.125,"expert_load_min":0,"fanout_histogram":[0,0,0,5,3,0,0,0],"fanout_max":5,"fanout_mean":4.375,"fanout_min":4,"hotspot_ratio":16.0,"locality":{"placement":"packed","local_rank_fraction":0.08571428805589676,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":35},"payload_copies_per_rank":[5,5,5,4,4,4,6,2],"payload_rank_cv":0.2539483904838562,"routed_copies":35},"token_rate_at_latency_percentile":{"p50":167392.0362252376,"p90":157862.53659351298,"p95":155165.05952494926,"p99":148472.59071398678},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.602000162005424,"p90":31.24699927866459,"p95":32.16800093650818,"p99":34.651998430490494},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.32300029695034,"p90":30.967000871896744,"p95":32.047998160123825,"p99":34.53199937939644},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":55.925000458955765,"p90":62.21400015056133,"p95":64.215999096632,"p99":69.18399780988693},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":46.549998223781586,"p90":49.07400161027908,"p95":50.916001200675964,"p99":54.88299950957298},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":48.79400134086609,"p90":51.39800161123276,"p95":52.43900045752525,"p99":55.52399903535843},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":26.839999482035637,"p90":27.320999652147293,"p95":27.520999312400818,"p99":28.241999447345734},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":25.157999247312546,"p90":25.59800073504448,"p95":25.759000331163406,"p99":26.000000536441803},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048001348972321,"p90":6.048999726772308,"p95":6.048999726772308,"p99":6.048999726772308},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.2020021677017212,"p90":5.447998642921448,"p95":6.8910010159015656,"p99":9.895000606775284},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.07999688386917114,"p90":0.12000277638435364,"p95":0.12000277638435364,"p99":0.12000277638435364},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":26.520000770688057,"p90":26.920000091195107,"p95":27.0409993827343,"p99":27.320999652147293},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":25.03800019621849,"p90":25.558000430464745,"p95":25.71900002658367,"p99":26.158999651670456},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":45.90899869799614,"p90":46.470001339912415,"p95":46.67099937796593,"p99":47.11100086569786},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.8839968144893646,"p90":5.648002028465271,"p95":6.650000810623169,"p99":9.174000471830368},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":501760,"scale_bytes":15680,"total_logical_bytes":517440},"roundtrip":{"activation_data_bytes":1505280,"scale_bytes":15680,"total_logical_bytes":1520960},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":501760,"scale_bytes":15680,"total_logical_bytes":517440},"roundtrip":{"activation_data_bytes":1505280,"scale_bytes":15680,"total_logical_bytes":1520960},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":70,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":12,"mean":8.75,"min":5,"total":70},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.33592739701271057,"expert_assignments_per_rank":[18,17,13,7,13,7,13,8],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,3,6,5,2,0,0],"fanout_max":6,"fanout_mean":4.375,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.15714286267757416,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":70},"payload_copies_per_rank":[12,11,10,5,9,6,10,7],"payload_rank_cv":0.26649653911590576,"routed_copies":70},"token_rate_at_latency_percentile":{"p50":327909.1601491521,"p90":311296.1496250719,"p95":305116.4183222703,"p99":288163.6819749058},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.202999383211136,"p90":31.647998839616776,"p95":32.489001750946045,"p99":34.13200005888939},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":31.127000227570534,"p90":33.81099924445152,"p95":34.81300175189972,"p99":37.49600052833557},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":59.32999961078167,"p90":65.4589980840683,"p95":67.30200350284576,"p99":71.62800058722496},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":49.153998494148254,"p90":51.51800066232681,"p95":52.71900072693825,"p99":55.44399842619896},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":51.79800093173981,"p90":54.76300045847893,"p95":55.52399903535843,"p99":57.40699917078018},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":27.241000905632973,"p90":27.761999517679214,"p95":27.88200043141842,"p99":28.202999383211136},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":27.760999277234077,"p90":28.201999142766,"p95":28.282999992370605,"p99":28.522999957203865},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.008997559547424,"p90":6.048999726772308,"p95":6.048999726772308,"p99":6.048999726772308},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.921003520488739,"p90":4.406999796628952,"p95":5.447998642921448,"p99":6.811000406742096},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.12000277638435364,"p90":0.08100271224975586,"p95":0.08100271224975586,"p99":0.08100271224975586},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.080999687314034,"p90":27.520999312400818,"p95":27.64200046658516,"p99":27.921000495553017},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.761999517679214,"p90":28.322000056505203,"p95":28.48299965262413,"p99":28.76400016248226},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":48.59299957752228,"p90":49.233999103307724,"p95":49.43399876356125,"p99":50.356000661849976},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.1239986419677734,"p90":6.088998168706894,"p95":7.09100067615509,"p99":9.494997560977936},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":1025024,"scale_bytes":32032,"total_logical_bytes":1057056},"roundtrip":{"activation_data_bytes":3075072,"scale_bytes":32032,"total_logical_bytes":3107104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":1025024,"scale_bytes":32032,"total_logical_bytes":1057056},"roundtrip":{"activation_data_bytes":3075072,"scale_bytes":32032,"total_logical_bytes":3107104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":143,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":22,"mean":17.875,"min":13,"total":143},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.20518453419208527,"expert_assignments_per_rank":[31,30,24,17,22,22,28,18],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,4,12,13,3,0,0],"fanout_max":6,"fanout_mean":4.46875,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.1258741319179535,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":143},"payload_copies_per_rank":[22,21,18,13,16,15,22,16],"payload_rank_cv":0.1800614446401596,"routed_copies":143},"token_rate_at_latency_percentile":{"p50":617784.4593301986,"p90":584336.1344720741,"p95":576327.3639498116,"p99":557423.3187978202},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.241999447345734,"p90":31.727999448776245,"p95":32.77000039815903,"p99":34.73300114274025},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":31.408000737428665,"p90":33.77100080251694,"p95":34.8920002579689,"p99":38.25800120830536},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":59.6500001847744,"p90":65.49900025129318,"p95":67.66200065612793,"p99":72.99100235104561},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":50.07600039243698,"p90":53.20100113749504,"p95":55.00299856066704,"p99":56.88599869608879},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":52.3190014064312,"p90":54.96300011873245,"p95":56.04499951004982,"p99":58.48800018429756},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":27.28099934756756,"p90":27.80199982225895,"p95":27.96199917793274,"p99":28.40300090610981},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.363000601530075,"p90":29.08400073647499,"p95":29.284000396728516,"p99":29.64399941265583},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009999662637711,"p90":6.09000027179718,"p95":6.09000027179718,"p99":6.09000027179718},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.2819990515708923,"p90":5.848001688718796,"p95":7.051002234220505,"p99":11.296000331640244},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.12099742889404297,"p90":0.23999810218811035,"p95":0.23999810218811035,"p99":0.23999810218811035},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.12099999189377,"p90":27.56199985742569,"p95":27.682000771164894,"p99":27.96199917793274},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":27.96199917793274,"p90":28.80300022661686,"p95":28.963999822735786,"p99":29.48400005698204},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":49.11400005221367,"p90":49.956001341342926,"p95":50.23600161075592,"p99":50.71600154042244},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.0849985778331757,"p90":5.847997963428497,"p95":6.850000470876694,"p99":9.613998234272003},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":2014208,"scale_bytes":62944,"total_logical_bytes":2077152},"roundtrip":{"activation_data_bytes":6042624,"scale_bytes":62944,"total_logical_bytes":6105568},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":2014208,"scale_bytes":62944,"total_logical_bytes":2077152},"roundtrip":{"activation_data_bytes":6042624,"scale_bytes":62944,"total_logical_bytes":6105568},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":281,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":42,"mean":35.125,"min":26,"total":281},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.1458333283662796,"expert_assignments_per_rank":[55,56,58,37,46,43,46,43],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,1,9,24,24,6,0,0],"fanout_max":6,"fanout_mean":4.390625,"fanout_min":2,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.11743772029876709,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":281},"payload_copies_per_rank":[38,42,42,26,29,31,38,35],"payload_rank_cv":0.15879225730895996,"routed_copies":281},"token_rate_at_latency_percentile":{"p50":1223264.937777137,"p90":1164419.6980103997,"p95":1141939.5228743595,"p99":1094241.5503750164},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":30.206000432372093,"p90":33.97100046277046,"p95":34.81300175189972,"p99":36.695998162031174},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":31.40699863433838,"p90":34.331999719142914,"p95":35.29300168156624,"p99":37.77699917554855},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":61.61299906671047,"p90":68.30300018191338,"p95":70.10600343346596,"p99":74.47299733757973},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":51.75799876451492,"p90":56.28500133752823,"p95":57.686999440193176,"p99":60.610998421907425},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":54.402001202106476,"p90":56.88599869608879,"p95":57.96699970960617,"p99":60.57199835777283},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":29.24400009214878,"p90":29.725000262260437,"p95":29.884999617934227,"p99":30.605999752879143},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.08300033211708,"p90":28.5630002617836,"p95":28.683999553322792,"p99":29.203999787569046},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009999662637711,"p90":6.048999726772308,"p95":6.048999726772308,"p99":6.048999726772308},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.563001424074173,"p90":7.532000541687012,"p95":10.175999253988266,"p99":12.89999857544899},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.20000338554382324,"p90":0.2800002694129944,"p95":0.2800002694129944,"p99":0.2800002694129944},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":29.08400073647499,"p90":29.48400005698204,"p95":29.565000906586647,"p99":29.844999313354492},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.12200039625168,"p90":28.603000566363335,"p95":28.723999857902527,"p99":29.123999178409576},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":51.15700140595436,"p90":51.67799815535545,"p95":51.83799937367439,"p99":52.23900079727173},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.244999796152115,"p90":5.85000216960907,"p95":7.050998508930206,"p99":9.493999183177948},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":4042752,"scale_bytes":126336,"total_logical_bytes":4169088},"roundtrip":{"activation_data_bytes":12128256,"scale_bytes":126336,"total_logical_bytes":12254592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":4042752,"scale_bytes":126336,"total_logical_bytes":4169088},"roundtrip":{"activation_data_bytes":12128256,"scale_bytes":126336,"total_logical_bytes":12254592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":564,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":77,"mean":70.5,"min":62,"total":564},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.07384087145328522,"expert_assignments_per_rank":[101,108,101,91,83,97,94,93],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,2,14,52,50,10,0,0],"fanout_max":6,"fanout_mean":4.40625,"fanout_min":2,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.14716312289237976,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":564},"payload_copies_per_rank":[73,76,77,63,62,69,74,70],"payload_rank_cv":0.07472094893455505,"routed_copies":564},"token_rate_at_latency_percentile":{"p50":2352854.622470097,"p90":2250114.3151908955,"p95":2208152.925651388,"p99":2113187.668730341},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":33.4089994430542,"p90":36.93599998950958,"p95":37.77699917554855,"p99":39.65900093317032},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":32.0889987051487,"p90":34.692998975515366,"p95":35.693999379873276,"p99":39.218999445438385},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":65.4979981482029,"p90":71.62899896502495,"p95":73.47099855542183,"p99":78.8780003786087},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":55.52399903535843,"p90":59.73000079393387,"p95":61.211999505758286,"p99":62.81500309705734},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":58.1279993057251,"p90":60.892000794410706,"p95":62.09300085902214,"p99":64.01699781417847},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":32.207999378442764,"p90":32.68900141119957,"p95":32.84899890422821,"p99":33.13000127673149},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":28.522999957203865,"p90":29.32400070130825,"p95":29.524000361561775,"p99":30.08499927818775},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048999726772308,"p90":6.088998168706894,"p95":6.088998168706894,"p99":6.088998168706894},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.962997019290924,"p90":6.72999769449234,"p95":7.892001420259476,"p99":11.01600006222725},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.2400018274784088,"p90":0.40099769830703735,"p95":0.40099769830703735,"p99":0.40099769830703735},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":31.968001276254654,"p90":32.409001141786575,"p95":32.52900019288063,"p99":32.77000039815903},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":28.762999922037125,"p90":29.48400005698204,"p95":29.7240000218153,"p99":30.12399934232235},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":54.84199896454811,"p90":55.48400059342384,"p95":55.68400025367737,"p99":56.405000388622284},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.244001418352127,"p90":6.1689987778663635,"p95":7.332000881433487,"p99":9.695004671812057},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":8157184,"scale_bytes":254912,"total_logical_bytes":8412096},"roundtrip":{"activation_data_bytes":24471552,"scale_bytes":254912,"total_logical_bytes":24726464},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":8157184,"scale_bytes":254912,"total_logical_bytes":8412096},"roundtrip":{"activation_data_bytes":24471552,"scale_bytes":254912,"total_logical_bytes":24726464},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1138,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":149,"mean":142.25,"min":126,"total":1138},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.04658475145697594,"expert_assignments_per_rank":[189,194,196,170,193,201,198,195],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,3,26,101,106,20,0,0],"fanout_max":6,"fanout_mean":4.4453125,"fanout_min":2,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.1344463974237442,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":1138},"payload_copies_per_rank":[143,141,149,126,144,145,148,142],"payload_rank_cv":0.04686218500137329,"routed_copies":1138},"token_rate_at_latency_percentile":{"p50":4404073.820837426,"p90":4204164.695857692,"p95":4122847.9290480786,"p99":3998937.918692919},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":42.74500161409378,"p90":46.30900174379349,"p95":47.07099869847298,"p99":49.0339994430542},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":33.250998705625534,"p90":36.093998700380325,"p95":37.17600181698799,"p99":39.659999310970306},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":75.99600031971931,"p90":82.40300044417381,"p95":84.24700051546097,"p99":88.6939987540245},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":66.57999753952026,"p90":69.82500106096268,"p95":71.66799902915955,"p99":74.39199835062027},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":68.9840018749237,"p90":71.90799713134766,"p95":73.0699971318245,"p99":74.67199862003326},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":41.822999715805054,"p90":42.38399863243103,"p95":42.54399985074997,"p99":43.14500093460083},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":30.205000191926956,"p90":30.805999413132668,"p95":30.967000871896744,"p99":31.327001750469208},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048999726772308,"p90":6.049007177352905,"p95":6.049007177352905,"p99":6.049007177352905},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.4420002698898315,"p90":5.768001079559326,"p95":7.570996880531311,"p99":11.096999049186707},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.1600012183189392,"p90":0.23999810218811035,"p95":0.23999810218811035,"p99":0.23999810218811035},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":41.423000395298004,"p90":41.822999715805054,"p95":41.98300093412399,"p99":42.344000190496445},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":29.965000227093697,"p90":30.605999752879143,"p95":30.76699934899807,"p99":31.207000836730003},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":65.53799659013748,"p90":66.26000255346298,"p95":66.54000282287598,"p99":67.50199943780899},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.4449994564056396,"p90":6.129004061222076,"p95":7.1709975600242615,"p99":8.892998099327087},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.007518797181546688,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":16142336,"scale_bytes":504448,"total_logical_bytes":16646784},"roundtrip":{"activation_data_bytes":48427008,"scale_bytes":504448,"total_logical_bytes":48931456},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":16142336,"scale_bytes":504448,"total_logical_bytes":16646784},"roundtrip":{"activation_data_bytes":48427008,"scale_bytes":504448,"total_logical_bytes":48931456},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2252,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":290,"mean":281.5,"min":268,"total":2252},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03184119611978531,"expert_assignments_per_rank":[396,393,399,359,380,389,380,376],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,3,57,223,191,38,0,0],"fanout_max":6,"fanout_mean":4.3984375,"fanout_min":2,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.11767318099737167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":2252},"payload_copies_per_rank":[290,286,290,268,282,282,280,274],"payload_rank_cv":0.025306938216090202,"routed_copies":2252},"token_rate_at_latency_percentile":{"p50":7422010.699354867,"p90":7120209.440193101,"p95":7006979.883635529,"p99":6856653.222920953},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":67.78199970722198,"p90":71.0270032286644,"p95":71.78899645805359,"p99":73.19100201129913},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":41.22199863195419,"p90":43.74599829316139,"p95":44.62800174951553,"p99":46.310000121593475},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":109.00399833917618,"p90":114.77300152182579,"p95":116.41699820756912,"p99":119.50100213289261},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":99.55000132322311,"p90":101.95399820804596,"p95":103.91700267791748,"p99":107.08200186491013},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":102.3550033569336,"p90":104.47700321674347,"p95":105.55999726057053,"p99":107.44199901819229},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":66.62099808454514,"p90":67.34199821949005,"p95":67.70200282335281,"p99":69.023996591568},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":37.57600113749504,"p90":38.217999041080475,"p95":38.378000259399414,"p99":38.617998361587524},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.010003387928009,"p90":6.087996065616608,"p95":6.087996065616608,"p99":6.087996065616608},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.76200270652771,"p90":6.610997021198273,"p95":8.011005818843842,"p99":9.814999997615814},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.20100176334381104,"p90":0.3599971532821655,"p95":0.3599971532821655,"p99":0.3599971532821655},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":66.17899984121323,"p90":66.78099930286407,"p95":66.9810026884079,"p99":67.58199632167816},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":37.93799877166748,"p90":38.458000868558884,"p95":38.65800052881241,"p99":39.058998227119446},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":98.90999644994736,"p90":99.74999725818634,"p95":99.95099902153015,"p99":100.47200322151184},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.3650025725364685,"p90":6.087996065616608,"p95":7.09100067615509,"p99":8.571997284889221},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.007518797181546688,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":32162816,"scale_bytes":1005088,"total_logical_bytes":33167904},"roundtrip":{"activation_data_bytes":96488448,"scale_bytes":1005088,"total_logical_bytes":97493536},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":32162816,"scale_bytes":1005088,"total_logical_bytes":33167904},"roundtrip":{"activation_data_bytes":96488448,"scale_bytes":1005088,"total_logical_bytes":97493536},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":4487,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":569,"mean":560.875,"min":553,"total":4487},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.01615842990577221,"expert_assignments_per_rank":[775,756,779,747,762,761,782,782],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,5,126,437,385,71,0,0],"fanout_max":6,"fanout_mean":4.3818359375,"fanout_min":2,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.12859371304512024,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":4487},"payload_copies_per_rank":[569,555,565,556,553,564,558,567],"payload_rank_cv":0.010142241604626179,"routed_copies":4487},"token_rate_at_latency_percentile":{"p50":10004396.135175679,"p90":9801199.962403726,"p95":9700644.43514808,"p99":9530723.640264869},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":87.4520018696785,"p90":90.2170017361641,"p95":91.17700159549713,"p99":92.5000011920929},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":52.03799903392792,"p90":54.281000047922134,"p95":55.48400059342384,"p99":57.085998356342316},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":139.49000090360641,"p90":144.49800178408623,"p95":146.66100218892097,"p99":149.5859995484352},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":132.07900524139404,"p90":133.36099684238434,"p95":134.28300619125366,"p99":136.245995759964},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":131.87900185585022,"p90":134.1630071401596,"p95":135.12299954891205,"p99":136.92599534988403},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":86.13000065088272,"p90":87.13199943304062,"p95":87.37199753522873,"p99":87.81299740076065},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":50.39599910378456,"p90":51.277000457048416,"p95":51.357001066207886,"p99":51.75799876451492},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048992276191711,"p90":6.089001893997192,"p95":6.089001893997192,"p99":6.089001893997192},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.003997564315796,"p90":6.4500123262405396,"p95":8.092001080513,"p99":11.096999049186707},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.12099742889404297,"p90":0.48100948333740234,"p95":0.48100948333740234,"p99":0.48100948333740234},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":85.44900268316269,"p90":86.25099807977676,"p95":86.52999997138977,"p99":87.09199726581573},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":48.67200180888176,"p90":49.5540015399456,"p95":49.795001745224,"p99":50.39599910378456},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":128.83399426937103,"p90":129.9159973859787,"p95":130.23699820041656,"p99":130.79799711704254},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.925008535385132,"p90":5.968004465103149,"p95":6.930992007255554,"p99":8.854001760482788},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.007518797181546688,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":64942080,"scale_bytes":2029440,"total_logical_bytes":66971520},"roundtrip":{"activation_data_bytes":194826240,"scale_bytes":2029440,"total_logical_bytes":196855680},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":64942080,"scale_bytes":2029440,"total_logical_bytes":66971520},"roundtrip":{"activation_data_bytes":194826240,"scale_bytes":2029440,"total_logical_bytes":196855680},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":9060,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":1149,"mean":1132.5,"min":1088,"total":9060},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.018761977553367615,"expert_assignments_per_rank":[1542,1490,1556,1492,1545,1541,1581,1541],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,11,225,859,791,162,0,0],"fanout_max":6,"fanout_mean":4.423828125,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.12792494893074036,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":9060},"payload_copies_per_rank":[1144,1088,1148,1126,1135,1149,1140,1130],"payload_rank_cv":0.016329573467373848,"routed_copies":9060},"token_rate_at_latency_percentile":{"p50":15529386.567837067,"p90":15265012.641379317,"p95":15156561.10978103,"p99":14956984.572336245},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":168.93500089645386,"p90":175.70500075817108,"p95":177.8690069913864,"p99":181.79500102996826},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":83.6469978094101,"p90":85.6499969959259,"p95":86.57100051641464,"p99":88.17300200462341},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":252.58199870586395,"p90":261.354997754097,"p95":264.44000750780106,"p99":269.9680030345917},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":240.04200100898743,"p90":244.1689968109131,"p95":246.411994099617,"p99":250.37801265716553},"sample_count":448},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":244.84899640083313,"p90":252.2200047969818,"p95":254.46400046348572,"p99":259.1109871864319},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":163.60700130462646,"p90":168.2939976453781,"p95":170.29699683189392,"p99":173.58200252056122},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":79.31999862194061,"p90":81.04199916124344,"p95":81.3629999756813,"p99":82.08400011062622},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":6.129994988441467,"p95":6.129994988441467,"p99":6.129994988441467},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.125995397567749,"p90":6.689995527267456,"p95":7.61200487613678,"p99":10.175004601478577},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.961005687713623,"p90":0.7209926843643188,"p95":0.7209926843643188,"p99":0.7209926843643188},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":164.60800170898438,"p90":171.05799913406372,"p95":173.26200008392334,"p99":177.02700197696686},"sample_count":2048},"dispatch":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":80.24100214242935,"p90":81.1230018734932,"p95":81.3629999756813,"p99":81.8839967250824},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"cuda-graph-replay","percentiles_us":{"p50":240.00200629234314,"p90":247.25399911403656,"p95":249.61699545383453,"p99":254.2240023612976},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.566997289657593,"p90":7.010996341705322,"p95":8.011996746063232,"p99":9.893998503684998},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":true,"cuda_graph_last_output_passed":true,"cuda_graph_last_output_error":0.0,"max_relative_error":0.007518797181546688,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":130156544,"scale_bytes":4067392,"total_logical_bytes":134223936},"roundtrip":{"activation_data_bytes":390469632,"scale_bytes":4067392,"total_logical_bytes":394537024},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":130156544,"scale_bytes":4067392,"total_logical_bytes":134223936},"roundtrip":{"activation_data_bytes":390469632,"scale_bytes":4067392,"total_logical_bytes":394537024},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":18158,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":2294,"mean":2269.75,"min":2209,"total":18158},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.016587993130087852,"expert_assignments_per_rank":[3123,2960,3095,3022,3074,3092,3108,3102],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,23,454,1699,1566,354,0,0],"fanout_max":6,"fanout_mean":4.43310546875,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.1253442019224167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":18158},"payload_copies_per_rank":[2289,2209,2294,2272,2259,2292,2267,2276],"payload_rank_cv":0.01135234534740448,"routed_copies":18158},"token_rate_at_latency_percentile":{"p50":16728677.920715638,"p90":16239790.350083344,"p95":16096579.44754254,"p99":15807897.783403927},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"IntraNode","block_num":80,"rdma_block_num":0,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":56,"scale_type_size":4,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"intranode-vllm-cudagraph-spin-ar","rank_alignment":{"method":"in-graph-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":true,"cuda_graph_supported":true,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi325x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9953,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2067,"p50":2102,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":753.3},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2066,"p50":2103,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":753.3},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2068,"p50":2103,"max":2105,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":753.3},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2068,"p50":2107,"max":2109,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":753.3},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2068,"p50":2101,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":753.3},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2063,"p50":2099,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":753.3},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2067,"p50":2103,"max":2105,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":753.3},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2061,"p50":2097,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":753.3}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16.json b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16.json new file mode 100644 index 0000000000..42be4b81cd --- /dev/null +++ b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T10:39:13.183528+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":2,"phase":"prefill","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":938.8120174407959,"p90":991.3709759712219,"p95":1003.350019454956,"p99":1029.9899578094482},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":838.9459848403931,"p90":875.0399947166443,"p95":883.4130167961121,"p99":894.8310017585754},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1777.758002281189,"p90":1866.4109706878662,"p95":1886.7630362510681,"p99":1924.8209595680237},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1684.9830150604248,"p90":1727.3600101470947,"p95":1739.1040325164795,"p99":1762.4520063400269},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1723.3530282974243,"p90":1763.8540267944336,"p95":1772.7080583572388,"p99":1787.25004196167},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":850.4030108451843,"p90":858.8560223579407,"p95":860.979974269867,"p99":864.2640113830566},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":775.2460241317749,"p90":814.149022102356,"p95":821.0359811782837,"p99":839.8669958114624},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.209969520568848,"p90":7.089972496032715,"p95":7.089972496032715,"p99":7.089972496032715},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":32.209038734436035,"p90":49.63493347167969,"p95":55.24301528930664,"p99":64.6580457687378},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-9.974956512451172,"p90":12.820005416870117,"p95":12.820005416870117,"p99":12.820005416870117},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":851.4050245285034,"p90":858.5759997367859,"p95":860.4580163955688,"p99":863.1830215454102},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":779.653012752533,"p90":813.103973865509,"p95":824.4410157203674,"p99":843.7129855155945},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1671.923041343689,"p90":1712.5849723815918,"p95":1722.1989631652832,"p99":1736.822009086609},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":50.26900768280029,"p90":60.72402000427246,"p95":64.44907188415527,"p99":74.06401634216309},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"roundtrip":{"activation_data_bytes":2428403712,"scale_bytes":0,"total_logical_bytes":2428403712},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"roundtrip":{"activation_data_bytes":2428403712,"scale_bytes":0,"total_logical_bytes":2428403712},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":84696,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":5367,"mean":5293.5,"min":5181,"total":84696},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.012361490167677402,"expert_assignments_per_rank":[6171,6137,6243,6050,6268,6146,6227,6063,6176,6175,6160,5993,6212,6144,6110,6029],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,5,241,2597,7671,5870,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16943359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.06316709518432617,"same_node_fraction":0.49927976727485657,"same_scaleup_domain_fraction":0.49927976727485657,"cross_node_fraction":0.500720202922821,"cross_domain_fraction":0.500720202922821,"gpus_per_node":8,"scale_up_domain":8,"copies":84696},"payload_copies_per_rank":[5326,5273,5366,5236,5367,5310,5348,5225,5315,5334,5305,5181,5318,5287,5280,5225],"payload_rank_cv":0.009882226586341858,"routed_copies":84696},"token_rate_at_latency_percentile":{"p50":9507048.022648308,"p90":9288750.515129479,"p95":9242356.587007893,"p99":9167156.03039912},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1427.2680282592773,"p90":1535.436987876892,"p95":1545.091986656189,"p99":1607.224941253662},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1223.6820459365845,"p90":1308.4490299224854,"p95":1333.6470127105713,"p99":1368.5790300369263},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":2650.950074195862,"p90":2843.8860177993774,"p95":2878.7389993667603,"p99":2975.8039712905884},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":2563.617944717407,"p90":2650.1200199127197,"p95":2666.934013366699,"p99":2701.637029647827},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2548.0339527130127,"p90":2609.2050075531006,"p95":2627.713918685913,"p99":2665.1699542999268},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1287.3369455337524,"p90":1374.1129636764526,"p95":1382.0860385894775,"p99":1389.0559673309326},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1121.0520267486572,"p90":1205.4200172424316,"p95":1225.0889539718628,"p99":1270.1170444488525},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009101867675781,"p90":7.411003112792969,"p95":7.411003112792969,"p99":7.411003112792969},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":46.14901542663574,"p90":67.2907829284668,"p95":75.78420639038086,"p99":92.1480655670166},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-14.50204849243164,"p90":27.52089500427246,"p95":27.52089500427246,"p99":27.52089500427246},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1286.8560552597046,"p90":1363.692045211792,"p95":1377.0780563354492,"p99":1385.0500583648682},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1096.1339473724365,"p90":1153.8219451904297,"p95":1170.727014541626,"p99":1204.3379545211792},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2507.8248977661133,"p90":2569.7579383850098,"p95":2589.0679359436035,"p99":2622.157096862793},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":39.92891311645508,"p90":57.47509002685547,"p95":61.56206130981445,"p99":69.69404220581055},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"roundtrip":{"activation_data_bytes":4850069504,"scale_bytes":0,"total_logical_bytes":4850069504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"roundtrip":{"activation_data_bytes":4850069504,"scale_bytes":0,"total_logical_bytes":4850069504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":169157,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":10735,"mean":10572.3125,"min":10405,"total":169157},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.008082338608801365,"expert_assignments_per_rank":[12320,12304,12326,12170,12384,12230,12308,12211,12262,12359,12237,12047,12413,12243,12482,12312],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,16,505,5263,15346,11638,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.162261962890625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.0630006417632103,"same_node_fraction":0.5010611414909363,"same_scaleup_domain_fraction":0.5010611414909363,"cross_node_fraction":0.4989388585090637,"cross_domain_fraction":0.4989388585090637,"gpus_per_node":8,"scale_up_domain":8,"copies":169157},"payload_copies_per_rank":[10630,10585,10602,10505,10644,10564,10577,10498,10504,10645,10499,10405,10642,10518,10735,10604],"payload_rank_cv":0.00741199404001236,"routed_copies":169157},"token_rate_at_latency_percentile":{"p50":12860111.210492449,"p90":12558614.560812019,"p95":12470155.05264244,"p99":12294900.723735396},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2690.540075302124,"p90":2714.0159606933594,"p95":2723.4981060028076,"p99":2745.971918106079},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2120.4299926757812,"p90":2144.9480056762695,"p95":2152.117967605591,"p99":2167.701005935669},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":4810.970067977905,"p90":4858.963966369629,"p95":4875.616073608398,"p99":4913.672924041748},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":4736.248016357422,"p90":4766.473770141602,"p95":4776.488780975342,"p99":4801.16605758667},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4779.15096282959,"p90":4811.598777770996,"p95":4821.616172790527,"p99":4839.321136474609},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2585.4110717773438,"p90":2615.8969402313232,"p95":2625.241994857788,"p99":2642.698049545288},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2046.4489459991455,"p90":2080.259084701538,"p95":2090.153932571411,"p99":2102.1640300750732},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.249904632568359,"p90":6.489753723144531,"p95":6.489753723144531,"p99":6.489753723144531},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":87.51106262207031,"p90":128.65304946899414,"p95":139.75095748901367,"p99":164.5669937133789},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.45193862915039,"p90":11.576652526855469,"p95":11.576652526855469,"p99":11.576652526855469},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2600.3921031951904,"p90":2623.6400604248047,"p95":2627.7129650115967,"p99":2636.2180709838867},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2082.8239917755127,"p90":2106.137990951538,"p95":2111.7870807647705,"p99":2125.519037246704},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4703.858852386475,"p90":4730.379104614258,"p95":4739.031791687012,"p99":4752.613067626953},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":74.33032989501953,"p90":99.609375,"p95":106.69994354248047,"p99":125.44727325439453},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"roundtrip":{"activation_data_bytes":9698332672,"scale_bytes":0,"total_logical_bytes":9698332672},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"roundtrip":{"activation_data_bytes":9698332672,"scale_bytes":0,"total_logical_bytes":9698332672},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":338251,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":21435,"mean":21140.6875,"min":20921,"total":338251},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.0070151351392269135,"expert_assignments_per_rank":[24582,24524,24711,24358,24680,24435,24611,24307,24504,24638,24478,24415,25004,24532,24846,24591],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,29,1048,10572,30561,23326,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1613006591796875,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.06308037787675858,"same_node_fraction":0.5009504556655884,"same_scaleup_domain_fraction":0.5009504556655884,"cross_node_fraction":0.49904951453208923,"cross_domain_fraction":0.49904951453208923,"gpus_per_node":8,"scale_up_domain":8,"copies":338251},"payload_copies_per_rank":[21145,21167,21213,20988,21198,21055,21170,20921,21054,21231,21048,21005,21435,21070,21396,21155],"payload_rank_cv":0.00635375315323472,"routed_copies":338251},"token_rate_at_latency_percentile":{"p50":13712895.97455991,"p90":13620420.784619113,"p95":13592122.983541181,"p99":13542395.338479694},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5376.570224761963,"p90":5567.562103271484,"p95":5638.628959655762,"p99":5744.949817657471},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4005.146026611328,"p90":4058.1860542297363,"p95":4072.667121887207,"p99":4098.12593460083},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":9381.716251373291,"p90":9625.74815750122,"p95":9711.296081542969,"p99":9843.0757522583},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":9221.118927001953,"p90":9286.094665527344,"p95":9309.652328491211,"p99":9342.580795288086},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9277.91976928711,"p90":9341.498374938965,"p95":9357.398986816406,"p99":9392.130851745605},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":5160.888195037842,"p90":5191.935062408447,"p95":5200.387001037598,"p99":5216.210842132568},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":3845.163106918335,"p90":3942.373037338257,"p95":3959.6400260925293,"p99":3998.955011367798},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.008148193359375,"p90":6.3686370849609375,"p95":6.3686370849609375,"p99":6.3686370849609375},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":186.5224838256836,"p90":278.86104583740234,"p95":312.5925064086914,"p99":352.84996032714844},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-24.517059326171875,"p90":36.69452667236328,"p95":36.69452667236328,"p99":36.69452667236328},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5159.042835235596,"p90":5194.1351890563965,"p95":5203.709125518799,"p99":5217.353820800781},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3897.402048110962,"p90":3982.0098876953125,"p95":3999.3960857391357,"p99":4024.913787841797},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9145.563125610352,"p90":9205.411911010742,"p95":9222.516059875488,"p99":9251.078605651855},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":132.51972198486328,"p90":186.7227554321289,"p95":208.8766098022461,"p99":252.65979766845703},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":131072,"byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"roundtrip":{"activation_data_bytes":19408105472,"scale_bytes":0,"total_logical_bytes":19408105472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"roundtrip":{"activation_data_bytes":19408105472,"scale_bytes":0,"total_logical_bytes":19408105472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":676901,"assignments":786432,"wire":"rank-deduplicated"},"receive":{"max":42692,"mean":42306.3125,"min":41888,"total":676901},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005902660544961691,"expert_assignments_per_rank":[49152,49158,48937,48637,49532,49147,49291,48640,48855,49080,49131,49180,49736,49110,49535,49311],"expert_load_cv":0.02157296985387802,"expert_load_max":2171,"expert_load_mean":2048.0,"expert_load_min":1922,"fanout_histogram":[0,51,2031,21028,61178,46784,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.164344787597656,"fanout_min":2,"hotspot_ratio":1.06005859375,"locality":{"placement":"packed","local_rank_fraction":0.06220702826976776,"same_node_fraction":0.5000066757202148,"same_scaleup_domain_fraction":0.5000066757202148,"cross_node_fraction":0.49999335408210754,"cross_domain_fraction":0.49999335408210754,"gpus_per_node":8,"scale_up_domain":8,"copies":676901},"payload_copies_per_rank":[42365,42433,42070,41888,42530,42296,42419,41908,42030,42386,42324,42267,42692,42172,42625,42496],"payload_rank_cv":0.005482310429215431,"routed_copies":676901},"token_rate_at_latency_percentile":{"p50":14127304.747114798,"p90":14031153.754908876,"p95":14007311.239444498,"p99":13955512.552898386},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"InterNodeV1","block_num":32,"rdma_block_num":16,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":0,"scale_type_size":0,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"inter-node-v1-vllm-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9938,"throttle_fraction_max":{"power":0.0014,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2089,"p50":2093,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":633.0},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2096,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0008,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":632.9},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2091,"p50":2094,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":632.9},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2093,"p50":2095,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":633.0},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2090,"max":2092,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":632.9},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2090,"p50":2092,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":71,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":633.0},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2088,"p50":2092,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":70,"clock_event_reasons":[],"throttle_fraction":{"power":0.0005,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":633.0},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2091,"p50":2095,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":632.9},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2093,"p50":2096,"max":2098,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":632.9},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2088,"p50":2092,"max":2093,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":71,"clock_event_reasons":[],"throttle_fraction":{"power":0.0006,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":632.9},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2090,"max":2093,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":73,"clock_event_reasons":[],"throttle_fraction":{"power":0.0006,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":632.9},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2082,"p50":2084,"max":2088,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0007,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":632.9},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2094,"p50":2097,"max":2098,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0006,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":633.0},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2093,"max":2095,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":73,"clock_event_reasons":[],"throttle_fraction":{"power":0.0006,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":633.0},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2085,"max":2087,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":72,"clock_event_reasons":[],"throttle_fraction":{"power":0.0005,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":632.9},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2089,"p50":2093,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0014,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":632.9}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8.json b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8.json new file mode 100644 index 0000000000..03f10d0447 --- /dev/null +++ b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T11:10:32.648340+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":2,"phase":"prefill","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-mori-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":903.4029841423035,"p90":915.6990051269531,"p95":924.5110154151917,"p99":949.1090178489685},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":682.0269823074341,"p90":707.385003566742,"p95":714.4749760627747,"p99":729.2180061340332},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1585.4299664497375,"p90":1623.084008693695,"p95":1638.9859914779663,"p99":1678.3270239830017},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1530.9510231018066,"p90":1563.1990432739258,"p95":1571.2050199508667,"p99":1585.787057876587},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1533.5079431533813,"p90":1567.7189826965332,"p95":1576.0530233383179,"p99":1592.918038368225},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":224.29800033569336,"p90":227.38300263881683,"p95":228.74599695205688,"p99":230.90800642967224},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":854.0449738502502,"p90":866.6679859161377,"p95":868.8709735870361,"p99":871.7560172080994},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":621.8590140342712,"p90":658.873975276947,"p95":664.4420027732849,"p99":674.2169857025146},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129026412963867,"p90":6.209969520568848,"p95":6.209969520568848,"p99":6.209969520568848},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":38.34497928619385,"p90":56.49304389953613,"p95":64.45693969726562,"p99":71.51401042938232},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.727005958557129,"p90":7.852077484130859,"p95":7.852077484130859,"p99":7.852077484130859},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":828.5269737243652,"p90":860.6590032577515,"p95":864.1049861907959,"p99":869.7530031204224},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":632.3140263557434,"p90":663.8420224189758,"p95":670.2520251274109,"p99":683.9889883995056},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1476.0680198669434,"p90":1510.5199813842773,"p95":1521.3359594345093,"p99":1538.1619930267334},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":57.199954986572266,"p90":68.17710399627686,"p95":71.62106037139893,"p99":82.59797096252441},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.009900989942252636,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":607100928,"scale_bytes":18971904,"total_logical_bytes":626072832},"roundtrip":{"activation_data_bytes":1821302784,"scale_bytes":18971904,"total_logical_bytes":1840274688},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":607100928,"scale_bytes":18971904,"total_logical_bytes":626072832},"roundtrip":{"activation_data_bytes":1821302784,"scale_bytes":18971904,"total_logical_bytes":1840274688},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":84696,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":5367,"mean":5293.5,"min":5181,"total":84696},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.012361490167677402,"expert_assignments_per_rank":[6171,6137,6243,6050,6268,6146,6227,6063,6176,6175,6160,5993,6212,6144,6110,6029],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,5,241,2597,7671,5870,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16943359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.06316709518432617,"same_node_fraction":0.49927976727485657,"same_scaleup_domain_fraction":0.49927976727485657,"cross_node_fraction":0.500720202922821,"cross_domain_fraction":0.500720202922821,"gpus_per_node":8,"scale_up_domain":8,"copies":84696},"payload_copies_per_rank":[5326,5273,5366,5236,5367,5310,5348,5225,5315,5334,5305,5181,5318,5287,5280,5225],"payload_rank_cv":0.009882226586341858,"routed_copies":84696},"token_rate_at_latency_percentile":{"p50":10684000.74036087,"p90":10450852.595928213,"p95":10395589.334485853,"p99":10285526.063088384},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1470.0989723205566,"p90":1530.0689935684204,"p95":1548.656940460205,"p99":1576.2189626693726},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1106.9060564041138,"p90":1155.4189920425415,"p95":1169.0399646759033,"p99":1189.5099878311157},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":2577.0050287246704,"p90":2685.487985610962,"p95":2717.6969051361084,"p99":2765.7289505004883},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":2446.4519023895264,"p90":2498.3699321746826,"p95":2517.3590183258057,"p99":2555.88698387146},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2447.484016418457,"p90":2494.473934173584,"p95":2510.8590126037598,"p99":2556.8079948425293},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":424.1600036621094,"p90":428.1260073184967,"p95":429.1270077228546,"p99":431.61100149154663},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1353.6779880523682,"p90":1426.2269735336304,"p95":1433.763027191162,"p99":1442.451000213623},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":966.1779999732971,"p90":1019.5790529251099,"p95":1035.2380275726318,"p99":1064.3260478973389},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.088972091674805,"p90":6.44993782043457,"p95":6.44993782043457,"p99":6.44993782043457},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":52.369117736816406,"p90":73.66108894348145,"p95":80.45196533203125,"p99":92.71097183227539},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-5.527973175048828,"p90":18.74685287475586,"p95":18.74685287475586,"p99":18.74685287475586},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1322.7109909057617,"p90":1382.7619552612305,"p95":1404.3190479278564,"p99":1422.2660064697266},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":962.332010269165,"p90":1018.0120468139648,"p95":1040.410041809082,"p99":1079.8250436782837},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2404.6690464019775,"p90":2451.4989852905273,"p95":2466.0420417785645,"p99":2507.5840950012207},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":42.33288764953613,"p90":60.80198287963867,"p95":66.89000129699707,"p99":78.74798774719238},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.011583011597394943,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":1212517376,"scale_bytes":37891168,"total_logical_bytes":1250408544},"roundtrip":{"activation_data_bytes":3637552128,"scale_bytes":37891168,"total_logical_bytes":3675443296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":1212517376,"scale_bytes":37891168,"total_logical_bytes":1250408544},"roundtrip":{"activation_data_bytes":3637552128,"scale_bytes":37891168,"total_logical_bytes":3675443296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":169157,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":10735,"mean":10572.3125,"min":10405,"total":169157},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.008082338608801365,"expert_assignments_per_rank":[12320,12304,12326,12170,12384,12230,12308,12211,12262,12359,12237,12047,12413,12243,12482,12312],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,16,505,5263,15346,11638,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.162261962890625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.0630006417632103,"same_node_fraction":0.5010611414909363,"same_scaleup_domain_fraction":0.5010611414909363,"cross_node_fraction":0.4989388585090637,"cross_domain_fraction":0.4989388585090637,"gpus_per_node":8,"scale_up_domain":8,"copies":169157},"payload_copies_per_rank":[10630,10585,10602,10505,10644,10564,10577,10498,10504,10645,10499,10405,10642,10518,10735,10604],"payload_rank_cv":0.00741199404001236,"routed_copies":169157},"token_rate_at_latency_percentile":{"p50":13388442.898986235,"p90":13136236.683449648,"p95":13050513.722799433,"p99":12815979.95082073},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2789.931058883667,"p90":2908.989906311035,"p95":2937.3528957366943,"p99":2989.311933517456},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1880.1089525222778,"p90":1975.0930070877075,"p95":2005.5789947509766,"p99":2053.3699989318848},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":4670.040011405945,"p90":4884.082913398743,"p95":4942.931890487671,"p99":5042.681932449341},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":4423.114776611328,"p90":4504.238128662109,"p95":4532.080173492432,"p99":4591.910839080811},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4483.9887619018555,"p90":4575.48713684082,"p95":4606.092929840088,"p99":4656.167984008789},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":864.5049929618835,"p90":868.7909841537476,"p95":869.9929714202881,"p99":872.9979991912842},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2568.7060356140137,"p90":2627.634048461914,"p95":2634.335994720459,"p99":2646.3940143585205},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1678.8140535354614,"p90":1766.4589881896973,"p95":1791.944980621338,"p99":1822.6699829101562},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049156188964844,"p90":6.0901641845703125,"p95":6.0901641845703125,"p99":6.0901641845703125},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":82.92531967163086,"p90":124.28617477416992,"p95":135.02168655395508,"p99":158.9198112487793},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-26.400089263916016,"p90":24.958133697509766,"p95":24.958133697509766,"p99":24.958133697509766},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2554.8059940338135,"p90":2587.6550674438477,"p95":2610.24808883667,"p99":2635.737895965576},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1687.587022781372,"p90":1769.9110507965088,"p95":1799.0750074386597,"p99":1844.9440002441406},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4416.224956512451,"p90":4507.483005523682,"p95":4537.248134613037,"p99":4596.41695022583},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":64.65673446655273,"p90":91.0787582397461,"p95":104.41923141479492,"p99":147.40324020385742},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":2424583168,"scale_bytes":75768224,"total_logical_bytes":2500351392},"roundtrip":{"activation_data_bytes":7273749504,"scale_bytes":75768224,"total_logical_bytes":7349517728},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":2424583168,"scale_bytes":75768224,"total_logical_bytes":2500351392},"roundtrip":{"activation_data_bytes":7273749504,"scale_bytes":75768224,"total_logical_bytes":7349517728},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":338251,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":21435,"mean":21140.6875,"min":20921,"total":338251},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.0070151351392269135,"expert_assignments_per_rank":[24582,24524,24711,24358,24680,24435,24611,24307,24504,24638,24478,24415,25004,24532,24846,24591],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,29,1048,10572,30561,23326,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1613006591796875,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.06308037787675858,"same_node_fraction":0.5009504556655884,"same_scaleup_domain_fraction":0.5009504556655884,"cross_node_fraction":0.49904951453208923,"cross_domain_fraction":0.49904951453208923,"gpus_per_node":8,"scale_up_domain":8,"copies":338251},"payload_copies_per_rank":[21145,21167,21213,20988,21198,21055,21170,20921,21054,21231,21048,21005,21435,21070,21396,21155],"payload_rank_cv":0.00635375315323472,"routed_copies":338251},"token_rate_at_latency_percentile":{"p50":14615558.485968042,"p90":14323283.628603932,"p95":14228110.678234894,"p99":14075093.558711326},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5355.741024017334,"p90":5440.9308433532715,"p95":5473.620891571045,"p99":5535.11381149292},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3063.5039806365967,"p90":3136.159896850586,"p95":3162.092924118042,"p99":3219.6459770202637},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":8419.24500465393,"p90":8577.090740203857,"p95":8635.713815689087,"p99":8754.759788513184},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":8301.374435424805,"p90":8377.082824707031,"p95":8404.244422912598,"p99":8448.229789733887},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8351.605415344238,"p90":8425.75740814209,"p95":8451.35498046875,"p99":8499.067306518555},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1846.940040588379,"p90":1863.9730215072632,"p95":1867.6179647445679,"p99":1872.465968132019},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":5190.330982208252,"p90":5221.097946166992,"p95":5231.513977050781,"p99":5241.448879241943},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2963.179111480713,"p90":3017.5940990448,"p95":3028.2371044158936,"p99":3065.747022628784},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.450653076171875,"p90":6.8912506103515625,"p95":6.8912506103515625,"p99":6.8912506103515625},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":156.71539306640625,"p90":222.7315902709961,"p95":251.2521743774414,"p99":303.2569885253906},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-20.311355590820312,"p90":25.038719177246094,"p95":25.038719177246094,"p99":25.038719177246094},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5187.928199768066,"p90":5218.655109405518,"p95":5226.466178894043,"p99":5238.163948059082},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2986.907958984375,"p90":3029.5729637145996,"p95":3045.196056365967,"p99":3075.9499073028564},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8235.876083374023,"p90":8300.288200378418,"p95":8325.130462646484,"p99":8375.004768371582},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":115.45944213867188,"p90":165.48919677734375,"p95":182.55901336669922,"p99":218.40858459472656},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01253132801502943,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":131072,"byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":4852026368,"scale_bytes":151625824,"total_logical_bytes":5003652192},"roundtrip":{"activation_data_bytes":14556079104,"scale_bytes":151625824,"total_logical_bytes":14707704928},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":4852026368,"scale_bytes":151625824,"total_logical_bytes":5003652192},"roundtrip":{"activation_data_bytes":14556079104,"scale_bytes":151625824,"total_logical_bytes":14707704928},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":676901,"assignments":786432,"wire":"rank-deduplicated"},"receive":{"max":42692,"mean":42306.3125,"min":41888,"total":676901},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005902660544961691,"expert_assignments_per_rank":[49152,49158,48937,48637,49532,49147,49291,48640,48855,49080,49131,49180,49736,49110,49535,49311],"expert_load_cv":0.02157296985387802,"expert_load_max":2171,"expert_load_mean":2048.0,"expert_load_min":1922,"fanout_histogram":[0,51,2031,21028,61178,46784,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.164344787597656,"fanout_min":2,"hotspot_ratio":1.06005859375,"locality":{"placement":"packed","local_rank_fraction":0.06220702826976776,"same_node_fraction":0.5000066757202148,"same_scaleup_domain_fraction":0.5000066757202148,"cross_node_fraction":0.49999335408210754,"cross_domain_fraction":0.49999335408210754,"gpus_per_node":8,"scale_up_domain":8,"copies":676901},"payload_copies_per_rank":[42365,42433,42070,41888,42530,42296,42419,41908,42030,42386,42324,42267,42692,42172,42625,42496],"payload_rank_cv":0.005482310429215431,"routed_copies":676901},"token_rate_at_latency_percentile":{"p50":15694228.053349363,"p90":15556108.923019877,"p95":15508992.38085609,"p99":15421927.521325937},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"InterNodeV1","block_num":32,"rdma_block_num":16,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":56,"scale_type_size":4,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"inter-node-v1-vllm-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9943,"throttle_fraction_max":{"power":0.0923,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":true,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2003,"p50":2087,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0837,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":1993,"p50":2089,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0883,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":1998,"p50":2087,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0859,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2003,"p50":2088,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0827,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.6},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":1997,"p50":2083,"max":2089,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0837,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2004,"p50":2087,"max":2093,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.0837,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":1996,"p50":2085,"max":2092,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0868,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2000,"p50":2088,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0851,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2003,"p50":2089,"max":2095,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0828,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":1998,"p50":2085,"max":2091,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":72,"clock_event_reasons":[],"throttle_fraction":{"power":0.0918,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":1998,"p50":2083,"max":2090,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":74,"clock_event_reasons":[],"throttle_fraction":{"power":0.0923,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.6},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":1996,"p50":2078,"max":2084,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0897,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2005,"p50":2090,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0918,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2001,"p50":2087,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":73,"clock_event_reasons":[],"throttle_fraction":{"power":0.0897,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":1981,"p50":2078,"max":2085,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":73,"clock_event_reasons":[],"throttle_fraction":{"power":0.0906,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":1989,"p50":2085,"max":2092,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0919,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":842.5}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16.json b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16.json new file mode 100644 index 0000000000..b6b0034d6f --- /dev/null +++ b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T06:57:23.759303+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":1,"phase":"prefill","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":279.62300181388855,"p90":284.2699885368347,"p95":286.19301319122314,"p99":290.3589904308319},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":219.13200616836548,"p90":221.73599898815155,"p95":222.85699844360352,"p99":224.9400019645691},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":498.755007982254,"p90":506.00598752498627,"p95":509.05001163482666,"p99":515.298992395401},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":488.2979989051819,"p90":492.86600947380066,"p95":494.18801069259644,"p99":496.47098779678345},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":491.70398712158203,"p90":496.5910017490387,"p95":498.3139932155609,"p99":502.27999687194824},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":273.25400710105896,"p90":276.8990099430084,"p95":277.90001034736633,"p99":279.5029878616333},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":217.20799803733826,"p90":219.8130041360855,"p95":220.53399682044983,"p99":222.53599762916565},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.970001220703125,"p90":6.049007177352905,"p95":6.049007177352905,"p99":6.049007177352905},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.763986587524414,"p90":6.931006908416748,"p95":8.772999048233032,"p99":11.579006910324097},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.5199909210205078,"p90":1.3620257377624512,"p95":1.3620257377624512,"p99":1.3620257377624512},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":276.09699964523315,"p90":280.4639935493469,"p95":282.14699029922485,"p99":286.91399097442627},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":214.72500264644623,"p90":216.48700535297394,"p95":217.00899302959442,"p99":218.97099912166595},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":486.6560101509094,"p90":491.38298630714417,"p95":492.94498562812805,"p99":496.8720078468323},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.566997289657593,"p90":7.851988077163696,"p95":9.255021810531616,"p99":12.13797926902771},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"roundtrip":{"activation_data_bytes":1041510400,"scale_bytes":0,"total_logical_bytes":1041510400},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"roundtrip":{"activation_data_bytes":1041510400,"scale_bytes":0,"total_logical_bytes":1041510400},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":36325,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":4587,"mean":4540.625,"min":4510,"total":36325},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.009137080051004887,"expert_assignments_per_rank":[6159,6057,6238,6119,6110,6118,6221,6130],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,46,880,3426,3151,689,0,0],"fanout_max":6,"fanout_mean":4.4342041015625,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.12393668293952942,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":36325},"payload_copies_per_rank":[4582,4510,4587,4547,4516,4535,4533,4515],"payload_rank_cv":0.006128025706857443,"routed_copies":36325},"token_rate_at_latency_percentile":{"p50":16660430.288466202,"p90":16496472.894488687,"p95":16439433.99449411,"p99":16309628.197454333},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":493.38701367378235,"p90":499.23598766326904,"p95":501.118004322052,"p99":505.28401136398315},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":416.6699945926666,"p90":419.9959933757782,"p95":420.9969937801361,"p99":423.7610101699829},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":910.057008266449,"p90":919.2319810390472,"p95":922.1149981021881,"p99":929.0450215339661},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":900.0419974327087,"p90":906.3310027122498,"p95":908.5350036621094,"p99":912.2599959373474},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":903.2859802246094,"p90":909.3350172042847,"p95":911.2579822540283,"p99":917.2279834747314},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":488.4589910507202,"p90":493.9480125904083,"p95":495.5500066280365,"p99":500.0759959220886},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":412.9450023174286,"p90":416.5099859237671,"p95":417.5119996070862,"p99":419.87600922584534},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009042263031006,"p90":6.048977375030518,"p95":6.048977375030518,"p99":6.048977375030518},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.004014492034912,"p90":7.210969924926758,"p95":9.21398401260376,"p99":13.060033321380615},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.48100948333740234,"p90":2.283036708831787,"p95":2.283036708831787,"p99":2.283036708831787},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":489.300012588501,"p90":494.94901299476624,"p95":496.832013130188,"p99":500.79697370529175},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":412.02399134635925,"p90":415.2680039405823,"p95":416.06900095939636,"p99":418.07299852371216},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":897.5980281829834,"p90":903.8079977035522,"p95":905.5699706077576,"p99":910.2569818496704},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.248010158538818,"p90":8.893013000488281,"p95":10.335981845855713,"p99":13.579964637756348},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"roundtrip":{"activation_data_bytes":2083221504,"scale_bytes":0,"total_logical_bytes":2083221504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"roundtrip":{"activation_data_bytes":2083221504,"scale_bytes":0,"total_logical_bytes":2083221504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":72657,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":9142,"mean":9082.125,"min":9007,"total":72657},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00735377287492156,"expert_assignments_per_rank":[12308,12293,12414,12290,12351,12153,12356,12139],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,92,1754,6798,6421,1319,0,0],"fanout_max":6,"fanout_mean":4.43463134765625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.12587912380695343,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":72657},"payload_copies_per_rank":[9133,9070,9142,9078,9120,9057,9050,9007],"payload_rank_cv":0.004780265036970377,"routed_copies":72657},"token_rate_at_latency_percentile":{"p50":18138220.18573341,"p90":18017561.943640944,"p95":17979540.72179824,"p99":17862516.5119064},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1007.36403465271,"p90":1019.5020437240601,"p95":1023.5079526901245,"p99":1033.884048461914},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":816.7960047721863,"p90":821.6829895973206,"p95":823.166012763977,"p99":825.8500099182129},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1824.1600394248962,"p90":1841.1850333213806,"p95":1846.6739654541016,"p99":1859.734058380127},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1812.5419616699219,"p90":1824.6400356292725,"p95":1832.772970199585,"p99":1841.8259620666504},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1817.149043083191,"p90":1830.1689624786377,"p95":1834.8159790039062,"p99":1843.9099788665771},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1001.073956489563,"p90":1013.0130052566528,"p95":1016.3370370864868,"p99":1026.6330242156982},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":810.10502576828,"p90":815.7140016555786,"p95":817.4369931221008,"p99":820.0799822807312},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049036979675293,"p90":6.049990653991699,"p95":6.049990653991699,"p99":6.049990653991699},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.134054183959961,"p90":13.459920883178711,"p95":15.342950820922852,"p99":17.38595962524414},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.162933349609375,"p90":4.405975341796875,"p95":4.405975341796875,"p99":4.405975341796875},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1000.633955001831,"p90":1012.6919746398926,"p95":1016.8190002441406,"p99":1025.3510475158691},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":811.8680119514465,"p90":816.474974155426,"p95":817.878007888794,"p99":820.2810287475586},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1808.0960512161255,"p90":1820.9550380706787,"p95":1825.0809907913208,"p99":1833.894968032837},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.733034133911133,"p90":12.738943099975586,"p95":14.263033866882324,"p99":16.98601245880127},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"roundtrip":{"activation_data_bytes":4160278528,"scale_bytes":0,"total_logical_bytes":4160278528},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"roundtrip":{"activation_data_bytes":4160278528,"scale_bytes":0,"total_logical_bytes":4160278528},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":145099,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":18284,"mean":18137.375,"min":17991,"total":145099},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005656693130731583,"expert_assignments_per_rank":[24624,24496,24614,24519,24621,24284,24656,24794],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,192,3636,13520,12793,2627,0,0],"fanout_max":6,"fanout_mean":4.428070068359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.125583216547966,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":145099},"payload_copies_per_rank":[18250,18101,18151,18109,18125,17991,18088,18284],"payload_rank_cv":0.004795215092599392,"routed_copies":145099},"token_rate_at_latency_percentile":{"p50":18032643.015567902,"p90":17904357.833509307,"p95":17859011.680173643,"p99":17770932.624455985},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2058.474063873291,"p90":2099.0960597991943,"p95":2110.9530925750732,"p99":2129.8229694366455},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1588.3229970932007,"p90":1595.1329469680786,"p95":1597.5370407104492,"p99":1601.7030477523804},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":3646.7970609664917,"p90":3694.229006767273,"p95":3708.4901332855225,"p99":3731.526017189026},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":3637.903928756714,"p90":3675.601005554199,"p95":3688.900947570801,"p99":3703.562021255493},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3640.2270793914795,"p90":3680.2470684051514,"p95":3691.6251182556152,"p99":3710.1330757141113},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2051.9440174102783,"p90":2093.8880443573,"p95":2103.5830974578857,"p99":2122.370958328247},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1578.0270099639893,"p90":1586.4399671554565,"p95":1588.0019664764404,"p99":1593.2899713516235},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.170034408569336,"p90":6.6509246826171875,"p95":6.6509246826171875,"p99":6.6509246826171875},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.542102813720703,"p90":19.78898048400879,"p95":22.11284637451172,"p99":25.719165802001953},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-5.92803955078125,"p90":6.208896636962891,"p95":6.208896636962891,"p99":6.208896636962891},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2049.220085144043,"p90":2090.2419090270996,"p95":2102.4999618530273,"p99":2120.60809135437},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1583.3959579467773,"p90":1590.2849435806274,"p95":1592.4899578094482,"p99":1596.4950323104858},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3628.890037536621,"p90":3668.3499813079834,"p95":3680.567979812622,"p99":3698.915958404541},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.456085205078125,"p90":16.18504524230957,"p95":18.22686195373535,"p99":21.3930606842041},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"roundtrip":{"activation_data_bytes":8316542976,"scale_bytes":0,"total_logical_bytes":8316542976},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"roundtrip":{"activation_data_bytes":8316542976,"scale_bytes":0,"total_logical_bytes":8316542976},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":290058,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":36460,"mean":36257.25,"min":36144,"total":290058},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00432177959010005,"expert_assignments_per_rank":[49106,49069,49115,48918,49142,48893,49536,49437],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,413,7235,27175,25451,5262,0,0],"fanout_max":6,"fanout_mean":4.425933837890625,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.1256852149963379,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":290058},"payload_copies_per_rank":[36354,36201,36216,36144,36222,36193,36268,36460],"payload_rank_cv":0.0026462646201252937,"routed_copies":290058},"token_rate_at_latency_percentile":{"p50":18003272.48017598,"p90":17807500.08949814,"p95":17752615.149332225,"p99":17664056.426705368},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"IntraNode","block_num":80,"rdma_block_num":0,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":0,"scale_type_size":0,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"intranode-vllm-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi325x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9938,"throttle_fraction_max":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2079,"p50":2092,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":70,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":269.7},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2073,"p50":2086,"max":2090,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":79,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":269.7},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2070,"p50":2084,"max":2087,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":79,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":269.7},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2078,"p50":2092,"max":2095,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":269.7},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2077,"p50":2092,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":71,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":269.7},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2084,"p50":2097,"max":2099,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":79,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":269.7},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2087,"max":2090,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":78,"clock_event_reasons":[],"throttle_fraction":{"power":0.0002,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":269.7},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2075,"p50":2088,"max":2091,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":74,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":269.7}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8.json b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8.json new file mode 100644 index 0000000000..7baffb771f --- /dev/null +++ b/collectivex/results/mi325x/mi325x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T06:59:36.707513+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"mori","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":1,"phase":"prefill","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-mori-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":278.34001183509827,"p90":282.74598717689514,"p95":284.18898582458496,"p99":288.0750000476837},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":138.44899833202362,"p90":141.41400158405304,"p95":142.5749957561493,"p99":145.5399990081787},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":416.7890101671219,"p90":424.1599887609482,"p95":426.76398158073425,"p99":433.6149990558624},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":406.574010848999,"p90":411.06000542640686,"p95":412.9030108451843,"p99":416.66901111602783},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":409.4580113887787,"p90":414.2250120639801,"p95":415.9870147705078,"p99":420.0339913368225},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":201.54400169849396,"p90":204.708993434906,"p95":205.82999289035797,"p99":207.87400007247925},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":273.934006690979,"p90":278.66101264953613,"p95":280.70300817489624,"p99":285.23099422454834},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":134.52300429344177,"p90":136.08600199222565,"p95":136.52600347995758,"p99":137.08700239658356},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.969017744064331,"p90":5.969017744064331,"p95":5.969017744064331,"p99":5.969017744064331},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.723991870880127,"p90":7.12999701499939,"p95":8.733004331588745,"p99":10.89698076248169},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.6810128688812256,"p90":-0.4799962043762207,"p95":-0.4799962043762207,"p99":-0.4799962043762207},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":274.6550142765045,"p90":278.34001183509827,"p95":279.7819972038269,"p99":282.94700384140015},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":134.36299562454224,"p90":135.76500117778778,"p95":136.2060010433197,"p99":136.92699372768402},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":404.6500027179718,"p90":408.4959924221039,"p95":409.9389910697937,"p99":412.54299879074097},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.525989294052124,"p90":8.412003517150879,"p95":9.69502329826355,"p99":12.017995119094849},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.007518797181546688,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":260377600,"scale_bytes":8136800,"total_logical_bytes":268514400},"roundtrip":{"activation_data_bytes":781132800,"scale_bytes":8136800,"total_logical_bytes":789269600},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":260377600,"scale_bytes":8136800,"total_logical_bytes":268514400},"roundtrip":{"activation_data_bytes":781132800,"scale_bytes":8136800,"total_logical_bytes":789269600},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":36325,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":4587,"mean":4540.625,"min":4510,"total":36325},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.009137080051004887,"expert_assignments_per_rank":[6159,6057,6238,6119,6110,6118,6221,6130],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,46,880,3426,3151,689,0,0],"fanout_max":6,"fanout_mean":4.4342041015625,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.12393668293952942,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":36325},"payload_copies_per_rank":[4582,4510,4587,4547,4516,4535,4533,4515],"payload_rank_cv":0.006128025706857443,"routed_copies":36325},"token_rate_at_latency_percentile":{"p50":20006935.441841263,"p90":19776690.8356917,"p95":19692922.396915134,"p99":19503183.477907836},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":495.74801325798035,"p90":503.07899713516235,"p95":505.88297843933105,"p99":513.2539868354797},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":246.53099477291107,"p90":249.89700317382812,"p95":251.2190043926239,"p99":254.34398651123047},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":742.2790080308914,"p90":752.9760003089905,"p95":757.101982831955,"p99":767.5979733467102},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":728.4989953041077,"p90":735.9899878501892,"p95":738.635003566742,"p99":750.3319978713989},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":734.7490191459656,"p90":741.9989705085754,"p95":744.8040246963501,"p99":752.4160146713257},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":375.96699595451355,"p90":379.3720006942749,"p95":380.57398796081543,"p99":383.05801153182983},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":488.17598819732666,"p90":495.1469898223877,"p95":498.03099036216736,"p99":509.7290277481079},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":241.80500209331512,"p90":244.28899586200714,"p95":244.76900696754456,"p99":245.57100236415863},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009042263031006,"p90":6.08903169631958,"p95":6.08903169631958,"p99":6.08903169631958},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.3249855041503906,"p90":7.169961929321289,"p95":8.572995662689209,"p99":11.29603385925293},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.0809898376464844,"p90":0.961005687713623,"p95":0.961005687713623,"p99":0.961005687713623},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":492.1019971370697,"p90":499.03300404548645,"p95":501.99800729751587,"p99":509.92900133132935},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":242.24600195884705,"p90":244.60899829864502,"p95":245.48999965190887,"p99":247.61299788951874},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":729.9010157585144,"p90":736.7110252380371,"p95":739.7159934043884,"p99":747.2079992294312},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.3669939041137695,"p90":7.77202844619751,"p95":9.253978729248047,"p99":11.8979811668396},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":520805376,"scale_bytes":16275168,"total_logical_bytes":537080544},"roundtrip":{"activation_data_bytes":1562416128,"scale_bytes":16275168,"total_logical_bytes":1578691296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":520805376,"scale_bytes":16275168,"total_logical_bytes":537080544},"roundtrip":{"activation_data_bytes":1562416128,"scale_bytes":16275168,"total_logical_bytes":1578691296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":72657,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":9142,"mean":9082.125,"min":9007,"total":72657},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00735377287492156,"expert_assignments_per_rank":[12308,12293,12414,12290,12351,12153,12356,12139],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,92,1754,6798,6421,1319,0,0],"fanout_max":6,"fanout_mean":4.43463134765625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.12587912380695343,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":72657},"payload_copies_per_rank":[9133,9070,9142,9078,9120,9057,9050,9007],"payload_rank_cv":0.004780265036970377,"routed_copies":72657},"token_rate_at_latency_percentile":{"p50":22298770.83612023,"p90":22080893.16993284,"p95":21997732.902530447,"p99":21775187.769171745},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1012.8879547119141,"p90":1025.8270502090454,"p95":1030.5149555206299,"p99":1040.5700206756592},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":470.0689911842346,"p90":473.71500730514526,"p95":474.83599185943604,"p99":477.84098982810974},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1482.9569458961487,"p90":1499.5420575141907,"p95":1505.350947380066,"p99":1518.411010503769},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1477.7499437332153,"p90":1492.3319816589355,"p95":1497.6199865341187,"p99":1507.4750185012817},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1474.3050336837769,"p90":1487.2039556503296,"p95":1491.3309812545776,"p99":1502.7480125427246},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":730.0209999084473,"p90":734.5079779624939,"p95":735.8310222625732,"p99":738.394021987915},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1007.7610015869141,"p90":1020.6999778747559,"p95":1023.5840082168579,"p99":1030.9959650039673},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":458.65198969841003,"p90":464.0600085258484,"p95":465.38299322128296,"p99":467.6249921321869},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049036979675293,"p90":6.0890913009643555,"p95":6.0890913009643555,"p99":6.0890913009643555},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.140963554382324,"p90":19.549012184143066,"p95":20.831942558288574,"p99":23.836970329284668},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.4879913330078125,"p90":6.770968437194824,"p95":6.770968437194824,"p99":6.770968437194824},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1005.4370164871216,"p90":1018.2160139083862,"p95":1022.5830078125,"p99":1033.519983291626},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":465.70301055908203,"p90":468.94800662994385,"p95":469.7889983654022,"p99":471.71199321746826},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1465.6909704208374,"p90":1478.6310195922852,"p95":1482.7569723129272,"p99":1493.6929941177368},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.292078971862793,"p90":12.097954750061035,"p95":13.460040092468262,"p99":15.823960304260254},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":1040069632,"scale_bytes":32502176,"total_logical_bytes":1072571808},"roundtrip":{"activation_data_bytes":3120208896,"scale_bytes":32502176,"total_logical_bytes":3152711072},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":1040069632,"scale_bytes":32502176,"total_logical_bytes":1072571808},"roundtrip":{"activation_data_bytes":3120208896,"scale_bytes":32502176,"total_logical_bytes":3152711072},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":145099,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":18284,"mean":18137.375,"min":17991,"total":145099},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005656693130731583,"expert_assignments_per_rank":[24624,24496,24614,24519,24621,24284,24656,24794],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,192,3636,13520,12793,2627,0,0],"fanout_max":6,"fanout_mean":4.428070068359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.125583216547966,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":145099},"payload_copies_per_rank":[18250,18101,18151,18109,18125,17991,18088,18284],"payload_rank_cv":0.004795215092599392,"routed_copies":145099},"token_rate_at_latency_percentile":{"p50":22226065.33338907,"p90":22033292.660031352,"p95":21972318.96331559,"p99":21805385.684426833},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2063.2729530334473,"p90":2105.1769256591797,"p95":2117.3551082611084,"p99":2137.6659870147705},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":903.0020236968994,"p90":907.8900218009949,"p95":909.492015838623,"p99":912.4969840049744},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":2966.2749767303467,"p90":3013.0669474601746,"p95":3026.8471240997314,"p99":3050.162971019745},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":2951.8940448760986,"p90":2990.19193649292,"p95":3001.209020614624,"p99":3020.556926727295},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2958.6639404296875,"p90":3001.2080669403076,"p95":3011.4641189575195,"p99":3031.373977661133},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1464.208960533142,"p90":1469.7370529174805,"p95":1471.5800285339355,"p99":1474.6650457382202},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2053.899049758911,"p90":2091.6359424591064,"p95":2104.214906692505,"p99":2122.4420070648193},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":893.5880064964294,"p90":899.5169997215271,"p95":901.2789726257324,"p99":904.8849940299988},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009101867675781,"p90":6.408929824829102,"p95":6.408929824829102,"p99":6.408929824829102},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.459920883178711,"p90":18.908977508544922,"p95":20.389795303344727,"p99":23.075103759765625},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-14.101028442382812,"p90":12.700080871582031,"p95":12.700080871582031,"p99":12.700080871582031},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2053.4980297088623,"p90":2094.640016555786,"p95":2107.4600219726562,"p99":2126.5690326690674},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":898.7159729003906,"p90":903.6830067634583,"p95":905.16597032547,"p99":907.7699780464172},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2947.808027267456,"p90":2990.5519485473633,"p95":3000.727891921997,"p99":3021.238088607788},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.255098342895508,"p90":15.263080596923828,"p95":16.624927520751953,"p99":19.428014755249023},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":2079135744,"scale_bytes":64972992,"total_logical_bytes":2144108736},"roundtrip":{"activation_data_bytes":6237407232,"scale_bytes":64972992,"total_logical_bytes":6302380224},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":2079135744,"scale_bytes":64972992,"total_logical_bytes":2144108736},"roundtrip":{"activation_data_bytes":6237407232,"scale_bytes":64972992,"total_logical_bytes":6302380224},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":290058,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":36460,"mean":36257.25,"min":36144,"total":290058},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00432177959010005,"expert_assignments_per_rank":[49106,49069,49115,48918,49142,48893,49536,49437],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,413,7235,27175,25451,5262,0,0],"fanout_max":6,"fanout_mean":4.425933837890625,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.1256852149963379,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":290058},"payload_copies_per_rank":[36354,36201,36216,36144,36222,36193,36268,36460],"payload_rank_cv":0.0026462646201252937,"routed_copies":290058},"token_rate_at_latency_percentile":{"p50":22150538.66188067,"p90":21836539.999312043,"p95":21762171.957302496,"p99":21619239.487753514},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"contract":"vllm","kernel":"IntraNode","block_num":80,"rdma_block_num":0,"warp_num_per_block":16,"gpu_per_node":8,"scale_dim":56,"scale_type_size":4,"use_external_inp_buffer":true,"realized_from":"handle","num_qp_per_pe":1,"shmem_num_qp_per_pe":4,"quant_type":"none","sdma":false,"heap_mode":"STATIC_HEAP","launch_config_mode":"MANUAL"},"fp8_consume":"native","kernel_generation":"intranode-vllm-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"0.0.0","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"production","name":"mori"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi325x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9957,"throttle_fraction_max":{"power":0.1519,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":true,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2005,"p50":2079,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":72,"clock_event_reasons":[],"throttle_fraction":{"power":0.1486,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":346.8},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2005,"p50":2082,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":79,"clock_event_reasons":[],"throttle_fraction":{"power":0.1501,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":346.9},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2003,"p50":2079,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":79,"clock_event_reasons":[],"throttle_fraction":{"power":0.15,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":346.9},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2014,"p50":2086,"max":2100,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":69,"clock_event_reasons":[],"throttle_fraction":{"power":0.1435,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":346.8},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2022,"p50":2085,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.1258,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":346.8},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2001,"p50":2078,"max":2093,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":72,"clock_event_reasons":[],"throttle_fraction":{"power":0.1519,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":346.9},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2010,"p50":2083,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":78,"clock_event_reasons":[],"throttle_fraction":{"power":0.1466,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":346.8},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2007,"p50":2077,"max":2090,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":70,"clock_event_reasons":[],"throttle_fraction":{"power":0.1413,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":346.9}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-swap-blocks-smoke-contiguous.json b/collectivex/results/mi325x/mi325x-swap-blocks-smoke-contiguous.json new file mode 100644 index 0000000000..46907f6494 --- /dev/null +++ b/collectivex/results/mi325x/mi325x-swap-blocks-smoke-contiguous.json @@ -0,0 +1,4075 @@ +{ + "schema": "collectivex-swap-blocks-v1", + "case_id": "mi325x-swap-blocks-smoke-contiguous", + "profile": "smoke", + "generated_at": "2026-10-28T08:28:55.335202+00:00", + "operation": "vllm._custom_ops.swap_blocks", + "timing": "drained-wall-clock-including-submission-and-synchronization", + "warmup": 4, + "iterations": 20, + "selection": { + "requested_block_bytes": [ + 257, + 4096, + 65536, + 262144 + ], + "requested_num_blocks": [ + 1, + 4, + 16, + 64, + 256, + 1024, + 2048 + ], + "max_payload_bytes": 2147483648, + "skipped_cases": [] + }, + "runtime": { + "sku": "mi325x", + "torch": "2.13.0+git733fca1", + "vllm": "0.32.0", + "cuda": null, + "hip": "7.2.53211", + "device": "", + "device_index": 0, + "image": "vllm/vllm-openai-rocm:v0.32.0", + "source_sha": "6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4", + "source_dirty": false + }, + "cases": [ + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 257, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 15.824, + "p90": 16.115, + "p95": 16.254, + "p99": 16.395 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.016241152679474215, + "p90": 0.015947874650946325, + "p95": 0.0158114925556786, + "p99": 0.015675510826471485 + }, + "samples_us": [ + 16.395, + 16.254, + 16.114, + 15.894, + 15.854, + 16.085, + 16.115, + 16.025, + 15.554, + 15.704, + 15.684, + 15.504, + 15.864, + 15.604, + 15.824, + 15.934, + 15.714, + 15.654, + 15.603, + 15.714 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1028, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 24.337, + "p90": 24.647, + "p95": 24.668, + "p99": 26.069 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.04224021037925792, + "p90": 0.04170893009291192, + "p95": 0.04167342305821307, + "p99": 0.039433810272737735 + }, + "samples_us": [ + 24.647, + 24.166, + 26.069, + 24.447, + 23.996, + 24.337, + 24.457, + 24.517, + 24.346, + 24.668, + 24.027, + 24.307, + 24.166, + 24.257, + 24.527, + 24.477, + 24.296, + 24.237, + 24.326, + 24.447 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4112, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 64.437, + "p90": 69.796, + "p95": 74.893, + "p99": 183.878 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.06381426819994723, + "p90": 0.05891455097713336, + "p95": 0.05490499779685685, + "p99": 0.02236265349851532 + }, + "samples_us": [ + 64.437, + 63.686, + 64.157, + 67.502, + 66.952, + 63.897, + 69.796, + 65.289, + 63.126, + 63.396, + 74.893, + 65.89, + 67.903, + 183.878, + 63.566, + 66.981, + 61.232, + 61.974, + 62.544, + 66.1 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16448, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 206.392, + "p90": 214.425, + "p95": 214.555, + "p99": 215.005 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.07969301135702934, + "p90": 0.0767074734755742, + "p95": 0.07666099601500781, + "p99": 0.07650054649891863 + }, + "samples_us": [ + 214.425, + 199.623, + 210.85, + 215.005, + 204.74, + 214.555, + 209.447, + 199.522, + 203.357, + 197.098, + 206.392, + 207.915, + 201.886, + 212.101, + 207.124, + 208.205, + 204.099, + 196.607, + 202.056, + 209.407 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65792, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 694.501, + "p90": 711.156, + "p95": 777.907, + "p99": 1977.353 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.09473276496362136, + "p90": 0.09251416004364725, + "p95": 0.08457566264347795, + "p99": 0.03327276414479357 + }, + "samples_us": [ + 1977.353, + 777.907, + 694.502, + 711.156, + 686.419, + 695.934, + 686.179, + 698.558, + 686.129, + 688.292, + 678.207, + 699.239, + 694.672, + 687.25, + 707.421, + 696.033, + 681.722, + 694.501, + 692.859, + 689.013 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 263168, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2779.528, + "p90": 2865.137, + "p95": 2873.65, + "p99": 2881.151 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.09468082350672488, + "p90": 0.09185180324710476, + "p95": 0.09157969829311155, + "p99": 0.09134127298430385 + }, + "samples_us": [ + 2881.151, + 2865.137, + 2828.501, + 2738.506, + 2789.322, + 2779.528, + 2768.57, + 2757.444, + 2769.563, + 2784.254, + 2802.994, + 2772.757, + 2873.65, + 2800.399, + 2723.272, + 2780.399, + 2774.229, + 2770.424, + 2756.602, + 2794.811 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 526336, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5744.355, + "p90": 5823.305, + "p95": 5826.4, + "p99": 6018.361 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.09162664911900466, + "p90": 0.09038441228821091, + "p95": 0.09033639983523273, + "p99": 0.08745503966943824 + }, + "samples_us": [ + 5778.157, + 5712.537, + 5709.653, + 5779.128, + 5720.63, + 5744.355, + 5794.882, + 5706.168, + 5530.853, + 5503.4, + 5518.954, + 5464.612, + 5752.147, + 5663.263, + 5796.574, + 6018.361, + 5795.804, + 5823.305, + 5810.576, + 5826.4 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4096, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 16.966, + "p90": 17.446, + "p95": 17.517, + "p99": 36.966 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.2414240245196275, + "p90": 0.23478161183079216, + "p95": 0.2338299937203859, + "p99": 0.11080452307525834 + }, + "samples_us": [ + 16.685, + 17.186, + 17.086, + 16.935, + 16.926, + 17.146, + 17.446, + 16.925, + 17.166, + 16.966, + 15.954, + 16.795, + 36.966, + 17.517, + 16.925, + 16.976, + 17.015, + 17.316, + 16.856, + 16.675 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16384, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 25.519, + "p90": 25.638, + "p95": 25.99, + "p99": 37.617 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.642031427563776, + "p90": 0.6390514080661518, + "p95": 0.6303963062716429, + "p99": 0.4355477576627589 + }, + "samples_us": [ + 25.528, + 25.629, + 24.968, + 25.128, + 25.99, + 25.609, + 25.539, + 25.519, + 25.358, + 25.038, + 25.248, + 25.238, + 25.618, + 25.589, + 24.897, + 37.617, + 25.489, + 25.638, + 24.988, + 25.619 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 77.517, + "p90": 82.395, + "p95": 82.565, + "p99": 82.635 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.845440355018899, + "p90": 0.7953880696644214, + "p95": 0.7937503784896748, + "p99": 0.7930779935862527 + }, + "samples_us": [ + 70.777, + 76.206, + 78.368, + 79.711, + 77.387, + 77.728, + 77.598, + 77.517, + 78.429, + 75.975, + 82.635, + 77.938, + 80.261, + 82.565, + 82.395, + 77.427, + 70.647, + 71.058, + 75.765, + 76.116 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 238.23, + "p90": 247.545, + "p95": 255.466, + "p99": 294.075 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.100381983797171, + "p90": 1.0589751358338888, + "p95": 1.0261404648759522, + "p99": 0.8914188557340814 + }, + "samples_us": [ + 243.779, + 234.135, + 238.23, + 239.212, + 294.075, + 255.466, + 243.649, + 246.873, + 245.952, + 236.979, + 246.753, + 238.461, + 247.545, + 228.085, + 219.011, + 201.856, + 220.343, + 198.51, + 213.523, + 201.335 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 722.563, + "p90": 742.954, + "p95": 762.033, + "p99": 831.709 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.4511897232490454, + "p90": 1.4113605956761792, + "p95": 1.3760243978935296, + "p99": 1.2607486512710575 + }, + "samples_us": [ + 831.709, + 762.033, + 718.618, + 732.769, + 718.107, + 742.954, + 718.378, + 742.254, + 714.782, + 725.928, + 718.948, + 713.42, + 714.782, + 723.886, + 717.987, + 722.563, + 716.955, + 727.732, + 741.573, + 733.75 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2850.525, + "p90": 2893.95, + "p95": 2915.152, + "p99": 3013.152 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.4714145639838276, + "p90": 1.4493353375144697, + "p95": 1.4387942721340088, + "p99": 1.3919988105478913 + }, + "samples_us": [ + 3013.152, + 2863.445, + 2840.399, + 2851.076, + 2847.13, + 2893.95, + 2797.485, + 2850.525, + 2821.712, + 2819.168, + 2841.462, + 2854.782, + 2833.789, + 2833.81, + 2851.546, + 2849.153, + 2872.658, + 2871.427, + 2850.946, + 2915.152 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 8388608, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5703.884, + "p90": 5726.929, + "p95": 5733.149, + "p99": 5799.84 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.4706834851480148, + "p90": 1.4647654964816221, + "p95": 1.4631763451464457, + "p99": 1.4463516234930618 + }, + "samples_us": [ + 5711.205, + 5703.884, + 5695.672, + 5688.772, + 5712.417, + 5720.449, + 5672.687, + 5706.007, + 5649.732, + 5682.482, + 5692.828, + 5668.731, + 5700.059, + 5714.089, + 5636.703, + 5799.84, + 5723.283, + 5733.149, + 5726.929, + 5723.535 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 17.787, + "p90": 18.177, + "p95": 18.238, + "p99": 18.719 + }, + "payload_gbps_at_latency_percentile": { + "p50": 3.6844886715016587, + "p90": 3.605435440391704, + "p95": 3.593376466717842, + "p99": 3.501041722314226 + }, + "samples_us": [ + 17.707, + 18.719, + 18.177, + 18.047, + 17.847, + 18.238, + 17.737, + 17.857, + 18.057, + 17.407, + 17.587, + 17.336, + 17.386, + 17.847, + 17.426, + 17.947, + 17.556, + 17.787, + 17.446, + 17.927 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 61.833, + "p90": 62.305, + "p95": 62.465, + "p99": 66.4 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.239548461177688, + "p90": 4.20743118529813, + "p95": 4.196654126310734, + "p99": 3.9479518072289155 + }, + "samples_us": [ + 61.252, + 61.533, + 61.983, + 61.684, + 61.983, + 62.305, + 62.145, + 61.163, + 61.833, + 62.144, + 61.993, + 60.892, + 61.934, + 61.773, + 61.774, + 62.465, + 61.924, + 61.002, + 66.4, + 61.203 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 238.301, + "p90": 242.617, + "p95": 245.161, + "p99": 249.047 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.4002165328722915, + "p90": 4.321939517840876, + "p95": 4.277091380766109, + "p99": 4.210353868948431 + }, + "samples_us": [ + 237.49, + 238.872, + 240.784, + 232.973, + 237.69, + 239.172, + 245.161, + 236.307, + 235.497, + 241.625, + 242.617, + 238.742, + 237.961, + 239.101, + 249.047, + 238.301, + 236.327, + 237.75, + 239.502, + 233.584 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 927.764, + "p90": 942.977, + "p95": 946.843, + "p99": 956.307 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.520873842916949, + "p90": 4.447938815050632, + "p95": 4.429777692817077, + "p99": 4.385938825084414 + }, + "samples_us": [ + 929.017, + 931.049, + 936.277, + 939.873, + 942.977, + 946.843, + 929.687, + 930.068, + 927.764, + 929.557, + 922.506, + 926.923, + 922.676, + 923.508, + 923.498, + 922.897, + 922.367, + 924.55, + 956.307, + 923.478 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3780.092, + "p90": 3788.996, + "p95": 3792.1, + "p99": 3809.517 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.438308908883699, + "p90": 4.427879047642172, + "p95": 4.4242546346351626, + "p99": 4.404027072198391 + }, + "samples_us": [ + 3770.628, + 3777.028, + 3779.771, + 3809.517, + 3783.197, + 3788.996, + 3778.74, + 3769.546, + 3784.359, + 3771.038, + 3778.219, + 3780.272, + 3786.302, + 3780.092, + 3766.722, + 3781.083, + 3780.713, + 3770.207, + 3780.172, + 3792.1 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 15152.958, + "p90": 15166.829, + "p95": 15169.022, + "p99": 15192.798 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.428763281730207, + "p90": 4.4247129047212175, + "p95": 4.424073219750094, + "p99": 4.4171497574047915 + }, + "samples_us": [ + 15192.798, + 15143.203, + 15153.459, + 15155.001, + 15150.524, + 15142.582, + 15159.228, + 15140.759, + 15169.022, + 15152.958, + 15162.742, + 15164.385, + 15142.632, + 15159.087, + 15166.829, + 15130.414, + 15149.813, + 15143.704, + 15158.857, + 15140.109 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 134217728, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 30175.908, + "p90": 30198.412, + "p95": 30198.613, + "p99": 30200.396 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.447843889237732, + "p90": 4.444529334853766, + "p95": 4.444499752356176, + "p99": 4.444237353708872 + }, + "samples_us": [ + 30168.498, + 30170.621, + 30187.266, + 30172.583, + 30181.297, + 30198.342, + 30200.396, + 30198.613, + 30179.494, + 30198.412, + 30175.588, + 30173.716, + 30181.047, + 30175.908, + 30169.819, + 30152.754, + 30171.663, + 30170.361, + 30178.543, + 30177.001 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 20.892, + "p90": 21.032, + "p95": 21.042, + "p99": 21.112 + }, + "payload_gbps_at_latency_percentile": { + "p50": 12.54757802029485, + "p90": 12.464054773678205, + "p95": 12.458131356334949, + "p99": 12.416824554755589 + }, + "samples_us": [ + 20.982, + 20.871, + 20.641, + 20.882, + 20.852, + 21.112, + 20.771, + 20.911, + 21.042, + 20.892, + 20.781, + 20.761, + 20.912, + 21.001, + 21.032, + 21.021, + 20.842, + 20.641, + 21.021, + 20.942 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 80.172, + "p90": 80.572, + "p95": 80.642, + "p99": 87.132 + }, + "payload_gbps_at_latency_percentile": { + "p50": 13.079079978047199, + "p90": 13.014148835823859, + "p95": 13.002852111802783, + "p99": 12.034338704494331 + }, + "samples_us": [ + 80.371, + 80.211, + 80.352, + 78.269, + 78.829, + 80.352, + 80.572, + 78.329, + 87.132, + 79.11, + 79.66, + 80.332, + 80.121, + 80.642, + 80.542, + 80.172, + 80.222, + 79.821, + 80.052, + 79.451 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 306.794, + "p90": 311.962, + "p95": 312.503, + "p99": 312.653 + }, + "payload_gbps_at_latency_percentile": { + "p50": 13.671401657138015, + "p90": 13.444919573537803, + "p95": 13.421643952218059, + "p99": 13.415204715771159 + }, + "samples_us": [ + 306.985, + 305.682, + 312.503, + 305.051, + 307.125, + 303.96, + 303.429, + 303.79, + 305.382, + 305.593, + 306.935, + 311.962, + 312.653, + 306.273, + 307.145, + 306.794, + 308.106, + 308.046, + 305.943, + 310.41 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 1326.568, + "p90": 1337.294, + "p95": 1337.325, + "p99": 1345.536 + }, + "payload_gbps_at_latency_percentile": { + "p50": 12.647083300667587, + "p90": 12.545645161049103, + "p95": 12.545354345428374, + "p99": 12.468797564687975 + }, + "samples_us": [ + 1337.094, + 1334.6, + 1345.536, + 1337.325, + 1321.401, + 1326.568, + 1332.588, + 1324.836, + 1326.539, + 1329.823, + 1325.587, + 1322.322, + 1331.415, + 1324.245, + 1326.789, + 1325.336, + 1337.294, + 1324.645, + 1325.155, + 1329.363 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5434.957, + "p90": 5476.981, + "p95": 5486.565, + "p99": 5494.618 + }, + "payload_gbps_at_latency_percentile": { + "p50": 12.347634765095657, + "p90": 12.252893336675807, + "p95": 12.23148983015785, + "p99": 12.213563163080673 + }, + "samples_us": [ + 5424.321, + 5434.957, + 5476.981, + 5452.433, + 5422.368, + 5494.618, + 5486.565, + 5421.927, + 5408.908, + 5464.952, + 5427.346, + 5432.193, + 5439.784, + 5449.93, + 5415.347, + 5455.828, + 5423.109, + 5465.523, + 5464.642, + 5427.365 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 268435456, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 20489.716, + "p90": 20530.508, + "p95": 20533.903, + "p99": 20536.798 + }, + "payload_gbps_at_latency_percentile": { + "p50": 13.100984708621633, + "p90": 13.074954404440454, + "p95": 13.07279263956784, + "p99": 13.070949814084942 + }, + "samples_us": [ + 20525.361, + 20536.798, + 20530.508, + 20494.704, + 20502.786, + 20495.475, + 20473.672, + 20489.716, + 20525.841, + 20511.529, + 20533.903, + 20470.798, + 20455.765, + 20456.035, + 20480.011, + 20512.842, + 20468.174, + 20477.678, + 20459.059, + 20479.942 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 536870912, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 38049.257, + "p90": 38120.905, + "p95": 38126.194, + "p99": 38151.642 + }, + "payload_gbps_at_latency_percentile": { + "p50": 14.109892132716285, + "p90": 14.083372679636016, + "p95": 14.081418984543802, + "p99": 14.072026362587486 + }, + "samples_us": [ + 38115.337, + 38151.642, + 38120.905, + 38126.194, + 38075.516, + 38049.257, + 38110.74, + 38020.393, + 38066.433, + 38033.773, + 38118.381, + 38106.884, + 37975.636, + 37999.942, + 38026.212, + 38036.187, + 38067.685, + 38007.564, + 37981.985, + 37942.895 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 257, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14.492, + "p90": 15.093, + "p95": 15.173, + "p99": 15.203 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.017733922163952524, + "p90": 0.017027761213807724, + "p95": 0.0169379819416068, + "p99": 0.0169045583108597 + }, + "samples_us": [ + 15.203, + 14.482, + 14.823, + 14.272, + 14.562, + 14.733, + 14.342, + 14.161, + 14.422, + 15.173, + 14.141, + 14.492, + 14.402, + 14.672, + 14.362, + 14.883, + 15.093, + 14.843, + 14.152, + 14.683 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1028, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 21.893, + "p90": 22.193, + "p95": 23.085, + "p99": 26.039 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.04695564792399397, + "p90": 0.046320911999279055, + "p95": 0.04453108078839073, + "p99": 0.03947924267444986 + }, + "samples_us": [ + 22.133, + 21.613, + 22.193, + 21.913, + 26.039, + 21.943, + 21.653, + 21.513, + 21.934, + 21.763, + 21.412, + 23.085, + 22.093, + 21.923, + 21.963, + 21.863, + 21.623, + 21.663, + 21.643, + 21.893 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4112, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 64.517, + "p90": 72.891, + "p95": 73.141, + "p99": 95.515 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.06373513957561573, + "p90": 0.05641300023322495, + "p95": 0.05622017746544346, + "p99": 0.04305082971261059 + }, + "samples_us": [ + 66.03, + 64.548, + 64.517, + 66.651, + 95.515, + 59.4, + 60.501, + 56.055, + 72.891, + 62.094, + 63.627, + 66.211, + 69.425, + 73.141, + 64.027, + 59.38, + 62.324, + 60.481, + 64.668, + 64.638 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16448, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 203.598, + "p90": 212.151, + "p95": 215.115, + "p99": 219.803 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08078664819890176, + "p90": 0.07752968404579758, + "p95": 0.07646142760848848, + "p99": 0.07483064380376973 + }, + "samples_us": [ + 206.312, + 199.021, + 206.072, + 195.005, + 206.512, + 192.432, + 204.62, + 191.88, + 208.355, + 193.243, + 203.579, + 197.97, + 203.598, + 194.534, + 212.151, + 208.655, + 219.803, + 206.262, + 215.115, + 201.034 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65792, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 772.76, + "p90": 794.563, + "p95": 799.58, + "p99": 871.299 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08513898234898287, + "p90": 0.08280274817729998, + "p95": 0.08228319867930664, + "p99": 0.07551024390019959 + }, + "samples_us": [ + 799.58, + 794.563, + 773.872, + 787.492, + 766.38, + 760.441, + 772.76, + 765.559, + 768.813, + 871.299, + 772.77, + 759.93, + 794.152, + 776.225, + 769.665, + 771.037, + 767.823, + 768.843, + 779.51, + 782.514 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 263168, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3001.153, + "p90": 3025.881, + "p95": 3026.601, + "p99": 3029.636 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.0876889648744999, + "p90": 0.08697235615015925, + "p95": 0.08695166624209798, + "p99": 0.08686456062708524 + }, + "samples_us": [ + 3015.604, + 2989.685, + 3006.371, + 2998.73, + 3026.601, + 3001.153, + 3017.538, + 3005.519, + 3024.558, + 3025.881, + 2993.641, + 2981.614, + 2994.523, + 3011.829, + 2991.908, + 2991.308, + 3029.636, + 3005.7, + 2984.488, + 2967.041 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 526336, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5974.094, + "p90": 6014.044, + "p95": 6014.224, + "p99": 6032.982 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08810306633943155, + "p90": 0.08751781663054012, + "p95": 0.08751519730558756, + "p99": 0.08724309139327782 + }, + "samples_us": [ + 5957.077, + 5991.219, + 5995.015, + 5957.458, + 5974.094, + 6014.224, + 5958.359, + 6032.982, + 6006.573, + 5982.476, + 5977.728, + 5988.365, + 5972.29, + 5989.257, + 5973.151, + 5947.663, + 5952.581, + 5924.789, + 5940.051, + 6014.044 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4096, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14.262, + "p90": 14.472, + "p95": 14.552, + "p99": 14.822 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.2871967465993549, + "p90": 0.2830292979546711, + "p95": 0.28147333699835075, + "p99": 0.27634597220348134 + }, + "samples_us": [ + 14.472, + 14.361, + 14.262, + 13.962, + 14.302, + 14.291, + 14.552, + 14.131, + 14.131, + 14.212, + 14.232, + 14.381, + 14.412, + 14.112, + 13.961, + 14.822, + 14.272, + 13.921, + 14.352, + 14.002 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16384, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 21.953, + "p90": 22.394, + "p95": 22.944, + "p99": 23.105 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.7463216872409237, + "p90": 0.7316245422881129, + "p95": 0.7140864714086471, + "p99": 0.7091105821250812 + }, + "samples_us": [ + 22.104, + 23.105, + 21.573, + 22.944, + 22.073, + 21.993, + 21.883, + 21.873, + 21.953, + 22.183, + 22.344, + 22.184, + 22.394, + 21.663, + 21.653, + 21.592, + 21.923, + 21.292, + 21.813, + 22.033 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 70.767, + "p90": 79.48, + "p95": 85.019, + "p99": 91.729 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.9260813656082637, + "p90": 0.8245596376446905, + "p95": 0.7708394594149544, + "p99": 0.714452354217314 + }, + "samples_us": [ + 70.827, + 75.705, + 69.094, + 71.478, + 71.468, + 79.48, + 71.909, + 68.514, + 68.063, + 67.603, + 91.729, + 69.295, + 85.019, + 74.734, + 70.767, + 70.237, + 70.848, + 66.551, + 68.864, + 68.984 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 213.353, + "p90": 223.158, + "p95": 224.8, + "p99": 231.31 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.2286867304420375, + "p90": 1.1747013326880507, + "p95": 1.166120996441281, + "p99": 1.133301629847391 + }, + "samples_us": [ + 219.573, + 205.23, + 213.353, + 221.195, + 219.653, + 224.8, + 209.878, + 222.237, + 218.781, + 207.794, + 208.986, + 202.917, + 231.31, + 216.668, + 213.242, + 204.099, + 217.139, + 223.158, + 212.071, + 204.289 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 773.681, + "p90": 782.344, + "p95": 798.499, + "p99": 799.24 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.3553079369921195, + "p90": 1.3403004305011605, + "p95": 1.3131838612196134, + "p99": 1.3119663680496472 + }, + "samples_us": [ + 799.24, + 781.984, + 768.383, + 774.262, + 774.292, + 777.066, + 771.989, + 777.628, + 772.75, + 771.247, + 766.37, + 774.382, + 773.681, + 771.718, + 774.592, + 798.499, + 770.206, + 769.935, + 782.344, + 772.068 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2979.71, + "p90": 2994.152, + "p95": 2999.571, + "p99": 3002.756 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.407621547063305, + "p90": 1.4008320218880004, + "p95": 1.398301290417863, + "p99": 1.39681812308426 + }, + "samples_us": [ + 2980.462, + 2958.598, + 2976.606, + 2983.676, + 2969.485, + 2982.295, + 2973.572, + 2994.152, + 2979.71, + 2976.145, + 2985.96, + 2999.571, + 2959.41, + 2987.342, + 2992.57, + 2960.331, + 2966.621, + 2968.945, + 3002.756, + 2990.306 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 8388608, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5958.279, + "p90": 6000.933, + "p95": 6184.902, + "p99": 6217.441 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.4078911041258726, + "p90": 1.3978839623771837, + "p95": 1.3563041095881552, + "p99": 1.3492058871165806 + }, + "samples_us": [ + 5971.249, + 5973.231, + 5959.912, + 5936.767, + 6000.933, + 5958.279, + 5980.893, + 5987.734, + 5939.521, + 5949.886, + 5952.711, + 5972.28, + 5948.895, + 5952.731, + 6184.902, + 5962.105, + 5954.293, + 6217.441, + 5941.444, + 5951.369 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14.762, + "p90": 15.093, + "p95": 15.163, + "p99": 15.333 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.439506841891343, + "p90": 4.342145365401179, + "p95": 4.322099848314977, + "p99": 4.274179873475511 + }, + "samples_us": [ + 14.692, + 15.093, + 14.713, + 14.873, + 15.163, + 14.422, + 14.762, + 14.902, + 14.682, + 14.722, + 15.042, + 14.893, + 14.783, + 14.903, + 14.753, + 14.672, + 15.333, + 14.672, + 14.682, + 14.882 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 55.674, + "p90": 57.177, + "p95": 57.216, + "p99": 59.34 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.708553364227467, + "p90": 4.584780593595327, + "p95": 4.58165548098434, + "p99": 4.417660936973374 + }, + "samples_us": [ + 54.823, + 57.177, + 55.374, + 55.674, + 56.035, + 54.753, + 57.216, + 55.454, + 55.033, + 55.614, + 55.604, + 55.194, + 55.954, + 57.117, + 55.905, + 56.476, + 59.34, + 57.046, + 55.484, + 55.684 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 217.128, + "p90": 223.388, + "p95": 224.079, + "p99": 228.196 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.8292988467632, + "p90": 4.693967446774223, + "p95": 4.679492500412801, + "p99": 4.595067398201546 + }, + "samples_us": [ + 218.751, + 218.231, + 215.947, + 228.196, + 215.637, + 217.128, + 215.807, + 217.329, + 223.388, + 214.926, + 214.014, + 218.18, + 219.001, + 215.596, + 215.957, + 217.649, + 213.924, + 213.944, + 218.17, + 224.079 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 845.69, + "p90": 851.109, + "p95": 851.599, + "p99": 853.022 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.959623502701936, + "p90": 4.928045643977446, + "p95": 4.925210104755877, + "p99": 4.9169939345057925 + }, + "samples_us": [ + 848.945, + 844.899, + 843.987, + 849.226, + 844.409, + 849.186, + 845.69, + 845.189, + 851.599, + 844.99, + 851.109, + 847.654, + 841.624, + 849.777, + 853.022, + 847.483, + 842.325, + 844.668, + 840.192, + 846.141 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3375.18, + "p90": 3378.845, + "p95": 3380.467, + "p99": 3392.245 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.970761855663994, + "p90": 4.965370119079153, + "p95": 4.962987658214087, + "p99": 4.945755981658165 + }, + "samples_us": [ + 3375.18, + 3375.961, + 3369.71, + 3375.86, + 3371.704, + 3372.395, + 3371.674, + 3378.655, + 3376.751, + 3376.001, + 3374.378, + 3375.81, + 3371.193, + 3376.741, + 3369.891, + 3392.245, + 3373.396, + 3378.845, + 3372.695, + 3380.467 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 13548.309, + "p90": 13556.541, + "p95": 13562.2, + "p99": 13562.641 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.953301847485173, + "p90": 4.950294031493727, + "p95": 4.948228458509681, + "p99": 4.948067562947364 + }, + "samples_us": [ + 13552.816, + 13548.309, + 13547.317, + 13544.173, + 13546.415, + 13549.73, + 13554.749, + 13545.324, + 13546.456, + 13550.042, + 13549.931, + 13556.541, + 13547.457, + 13562.641, + 13562.2, + 13547.708, + 13549.13, + 13548.139, + 13552.806, + 13548.248 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 134217728, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 27036.716, + "p90": 27045.119, + "p95": 27049.777, + "p99": 27054.083 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.964276282666874, + "p90": 4.962733867061187, + "p95": 4.9618792790787145, + "p99": 4.96108953313997 + }, + "samples_us": [ + 27033.141, + 27037.698, + 27054.083, + 27039.341, + 27041.173, + 27034.743, + 27024.138, + 27045.119, + 27036.576, + 27032.5, + 27049.777, + 27033.853, + 27034.633, + 27036.877, + 27033.282, + 27039.2, + 27027.482, + 27041.865, + 27039.651, + 27036.716 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 18.468, + "p90": 18.668, + "p95": 18.669, + "p99": 18.728 + }, + "payload_gbps_at_latency_percentile": { + "p50": 14.19449859215941, + "p90": 14.042425541032783, + "p95": 14.041673362258289, + "p99": 13.997436992738146 + }, + "samples_us": [ + 18.648, + 18.669, + 18.588, + 18.668, + 18.548, + 18.598, + 18.438, + 18.348, + 18.267, + 18.068, + 18.338, + 18.508, + 18.398, + 18.468, + 18.638, + 18.278, + 18.418, + 18.548, + 18.458, + 18.728 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 69.755, + "p90": 70.697, + "p95": 70.757, + "p99": 71.037 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.032270088165722, + "p90": 14.831973068164137, + "p95": 14.819395960823664, + "p99": 14.760983712713093 + }, + "samples_us": [ + 69.876, + 69.495, + 69.425, + 69.755, + 69.575, + 70.757, + 71.037, + 69.075, + 70.167, + 69.745, + 70.296, + 70.557, + 69.895, + 70.697, + 70.247, + 69.205, + 68.794, + 69.164, + 69.165, + 69.795 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 271.671, + "p90": 274.966, + "p95": 275.978, + "p99": 278.502 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.438909563405737, + "p90": 15.253900482241441, + "p95": 15.197965055185557, + "p99": 15.06022936998657 + }, + "samples_us": [ + 269.047, + 272.493, + 271.671, + 273.614, + 271.871, + 269.648, + 272.543, + 270.349, + 269.928, + 275.978, + 269.678, + 271.781, + 271.952, + 271.41, + 271.421, + 274.966, + 270.82, + 267.484, + 273.644, + 278.502 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 1081.157, + "p90": 1082.218, + "p95": 1083.05, + "p99": 1083.451 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.51783506003291, + "p90": 15.502621468133038, + "p95": 15.4907123401505, + "p99": 15.484979016125326 + }, + "samples_us": [ + 1081.287, + 1081.938, + 1083.451, + 1077.952, + 1081.837, + 1081.157, + 1079.894, + 1080.806, + 1080.846, + 1082.218, + 1079.965, + 1083.05, + 1081.947, + 1081.207, + 1079.494, + 1076.97, + 1081.557, + 1077.291, + 1078.262, + 1081.607 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 4309.534, + "p90": 4313.059, + "p95": 4313.459, + "p99": 4314.211 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.572185763008251, + "p90": 15.559458843479767, + "p95": 15.558015968159197, + "p99": 15.555304086888658 + }, + "samples_us": [ + 4310.235, + 4305.948, + 4309.293, + 4308.092, + 4309.383, + 4306.789, + 4310.255, + 4314.211, + 4310.585, + 4310.625, + 4309.534, + 4310.285, + 4313.459, + 4307.11, + 4313.049, + 4310.184, + 4313.059, + 4309.463, + 4308.442, + 4305.367 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 268435456, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 17198.905, + "p90": 17210.893, + "p95": 17215.409, + "p99": 17217.322 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.60770618827187, + "p90": 15.596834864989283, + "p95": 15.592743454425044, + "p99": 15.591010959776439 + }, + "samples_us": [ + 17205.475, + 17199.726, + 17215.409, + 17206.617, + 17202.71, + 17190.662, + 17197.703, + 17195.129, + 17195.951, + 17201.299, + 17198.324, + 17197.423, + 17209.2, + 17210.893, + 17198.905, + 17195.549, + 17196.571, + 17198.614, + 17205.675, + 17217.322 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 536870912, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 34316.446, + "p90": 34345.991, + "p95": 34351.689, + "p99": 34355.505 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.64471192617091, + "p90": 15.631254081444323, + "p95": 15.62866128649453, + "p99": 15.62692535010037 + }, + "samples_us": [ + 34295.695, + 34309.056, + 34322.395, + 34313.302, + 34316.125, + 34305.41, + 34298.619, + 34316.446, + 34302.795, + 34305.249, + 34339.682, + 34339.942, + 34355.505, + 34312.521, + 34336.547, + 34321.665, + 34322.956, + 34339.481, + 34351.689, + 34345.991 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 257, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14.552, + "p90": 15.123, + "p95": 15.383, + "p99": 18.918 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.017660802638812534, + "p90": 0.016993982675395095, + "p95": 0.016706754209191967, + "p99": 0.013584945554498361 + }, + "samples_us": [ + 14.622, + 15.052, + 14.682, + 14.542, + 15.383, + 14.842, + 14.552, + 15.123, + 14.312, + 14.452, + 14.332, + 14.512, + 14.622, + 18.918, + 14.302, + 14.632, + 14.632, + 13.801, + 14.282, + 14.272 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1028, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 21.813, + "p90": 22.094, + "p95": 22.354, + "p99": 22.704 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.047127859533305824, + "p90": 0.04652846926767448, + "p95": 0.045987295338641855, + "p99": 0.04527836504580691 + }, + "samples_us": [ + 21.883, + 21.823, + 21.623, + 22.094, + 21.502, + 21.813, + 21.342, + 21.813, + 22.094, + 21.613, + 22.013, + 22.013, + 22.704, + 22.354, + 21.923, + 21.643, + 21.503, + 21.362, + 21.833, + 21.773 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4112, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 65.179, + "p90": 70.616, + "p95": 72.86, + "p99": 78.739 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.06308780435416315, + "p90": 0.058230429364449984, + "p95": 0.056437002470491354, + "p99": 0.052223167680564904 + }, + "samples_us": [ + 70.296, + 63.136, + 63.976, + 62.004, + 70.616, + 57.517, + 65.179, + 57.016, + 62.624, + 62.965, + 72.86, + 67.753, + 78.739, + 60.902, + 69.595, + 69.655, + 68.574, + 61.894, + 68.193, + 66.03 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16448, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 193.984, + "p90": 229.728, + "p95": 230.79, + "p99": 233.684 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08479049818541735, + "p90": 0.0715977155592701, + "p95": 0.07126825252393951, + "p99": 0.07038564899608017 + }, + "samples_us": [ + 223.859, + 180.763, + 227.895, + 187.814, + 224.56, + 190.358, + 221.085, + 190.568, + 233.684, + 193.984, + 222.687, + 189.788, + 230.79, + 190.338, + 229.728, + 193.573, + 223.668, + 191.951, + 226.643, + 192.231 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65792, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 768.543, + "p90": 829.105, + "p95": 830.527, + "p99": 834.914 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.0856061404501765, + "p90": 0.07935303731131763, + "p95": 0.07921717174757714, + "p99": 0.07880093039522633 + }, + "samples_us": [ + 756.836, + 828.995, + 737.567, + 834.914, + 718.528, + 820.192, + 708.322, + 830.527, + 715.262, + 829.105, + 819.671, + 739.8, + 779.38, + 771.458, + 768.543, + 767.632, + 768.163, + 824.929, + 763.315, + 779.32 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 263168, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3014.723, + "p90": 3041.284, + "p95": 3043.527, + "p99": 3043.547 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08729425555847088, + "p90": 0.08653187272217919, + "p95": 0.0864681009894113, + "p99": 0.08646753278329528 + }, + "samples_us": [ + 2981.724, + 3008.484, + 3012.871, + 3012.79, + 3027.373, + 3014.242, + 3004.689, + 3016.206, + 3043.547, + 3009.205, + 3016.486, + 3010.507, + 3002.165, + 3014.723, + 3043.527, + 3034.634, + 3018.259, + 3028.425, + 3024.799, + 3041.284 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 526336, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5926.141, + "p90": 5967.263, + "p95": 5990.809, + "p99": 6107.715 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08881597653515162, + "p90": 0.08820392196556445, + "p95": 0.08785724932976498, + "p99": 0.08617559922164017 + }, + "samples_us": [ + 5927.403, + 5922.866, + 5949.907, + 5948.354, + 5920.963, + 5944.218, + 5918.69, + 5990.809, + 5908.013, + 5891.068, + 5906.631, + 5950.778, + 5951.939, + 5916.146, + 5948.915, + 5925.439, + 5926.141, + 5967.263, + 6107.715, + 5916.616 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4096, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14.332, + "p90": 15.143, + "p95": 15.293, + "p99": 16.094 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.28579402735138154, + "p90": 0.27048801426401636, + "p95": 0.26783495716994704, + "p99": 0.25450478439169877 + }, + "samples_us": [ + 14.572, + 14.582, + 14.612, + 14.663, + 14.242, + 14.152, + 15.293, + 14.362, + 14.221, + 14.622, + 16.094, + 14.311, + 14.281, + 14.272, + 14.332, + 14.331, + 14.221, + 15.083, + 15.143, + 13.911 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16384, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 21.853, + "p90": 22.674, + "p95": 23.465, + "p99": 36.175 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.7497368782318217, + "p90": 0.7225897503748787, + "p95": 0.698231408480716, + "p99": 0.45290946786454733 + }, + "samples_us": [ + 21.182, + 22.174, + 21.943, + 21.973, + 21.372, + 21.623, + 21.643, + 22.013, + 21.933, + 22.113, + 21.583, + 21.563, + 36.175, + 22.674, + 22.194, + 21.753, + 21.853, + 21.452, + 23.465, + 21.602 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65536, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 69.686, + "p90": 73.522, + "p95": 78.269, + "p99": 80.712 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.9404471486381769, + "p90": 0.8913794510486657, + "p95": 0.8373174564642452, + "p99": 0.8119734364158985 + }, + "samples_us": [ + 68.053, + 71.298, + 71.118, + 69.355, + 71.318, + 71.799, + 66.39, + 67.072, + 69.295, + 70.186, + 73.522, + 70.046, + 78.269, + 80.712, + 72.159, + 69.686, + 65.168, + 68.093, + 65.98, + 67.932 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 214.996, + "p90": 228.375, + "p95": 230.729, + "p99": 237.91 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.21929710320192, + "p90": 1.1478664477285168, + "p95": 1.136155403091939, + "p99": 1.1018620486738682 + }, + "samples_us": [ + 220.424, + 204.91, + 208.856, + 220.634, + 217.599, + 204.52, + 214.996, + 209.787, + 230.729, + 219.002, + 208.476, + 204.349, + 223.078, + 212.933, + 216.478, + 237.91, + 228.375, + 218.882, + 214.735, + 205.762 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 769.975, + "p90": 781.302, + "p95": 781.493, + "p99": 782.144 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.3618312282866327, + "p90": 1.3420879506260064, + "p95": 1.341759938988577, + "p99": 1.3406431552246134 + }, + "samples_us": [ + 769.405, + 767.572, + 765.929, + 769.975, + 768.343, + 773.191, + 769.605, + 769.074, + 764.017, + 781.302, + 781.493, + 782.144, + 772.259, + 767.632, + 775.153, + 767.542, + 775.013, + 778.819, + 771.247, + 772.709 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2959.179, + "p90": 2972.37, + "p95": 2982.665, + "p99": 3001.183 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.417387728150274, + "p90": 1.4110975416923195, + "p95": 1.406226981575202, + "p99": 1.3975502326915752 + }, + "samples_us": [ + 2934.922, + 2937.207, + 2942.704, + 2946.771, + 2972.069, + 2964.828, + 2972.37, + 2959.179, + 2935.894, + 2962.645, + 2958.839, + 2942.975, + 2948.764, + 2948.773, + 2960.602, + 2962.134, + 2982.665, + 3001.183, + 2962.795, + 2962.655 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 8388608, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5910.327, + "p90": 6088.526, + "p95": 6088.636, + "p99": 6111.171 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.4193136860278628, + "p90": 1.3777732081623697, + "p95": 1.3777483167001607, + "p99": 1.372667856945911 + }, + "samples_us": [ + 5904.388, + 5896.205, + 5891.709, + 6080.374, + 6088.526, + 6087.966, + 5910.687, + 5941.504, + 5918.078, + 5956.597, + 5893.211, + 5898.579, + 5856.605, + 6088.636, + 5866.1, + 6087.826, + 5884.738, + 5910.327, + 5888.794, + 6111.171 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 65536, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14.592, + "p90": 15.192, + "p95": 16.285, + "p99": 25.108 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.491228070175438, + "p90": 4.313849394418114, + "p95": 4.024316856002456, + "p99": 2.610164091126334 + }, + "samples_us": [ + 14.172, + 14.441, + 14.292, + 14.292, + 14.502, + 14.361, + 16.285, + 14.853, + 15.192, + 14.322, + 14.592, + 14.632, + 14.542, + 25.108, + 14.733, + 15.173, + 14.993, + 14.752, + 14.752, + 14.302 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 22.304, + "p90": 23.356, + "p95": 24.867, + "p99": 24.897 + }, + "payload_gbps_at_latency_percentile": { + "p50": 11.7532281205165, + "p90": 11.223839698578523, + "p95": 10.541842602646078, + "p99": 10.529140057034985 + }, + "samples_us": [ + 23.356, + 22.243, + 22.294, + 22.645, + 22.204, + 22.494, + 22.654, + 22.153, + 22.444, + 24.897, + 21.793, + 24.867, + 22.344, + 21.863, + 21.803, + 23.255, + 22.304, + 22.354, + 21.903, + 22.104 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 70.857, + "p90": 75.404, + "p95": 77.127, + "p99": 77.978 + }, + "payload_gbps_at_latency_percentile": { + "p50": 14.798481448551307, + "p90": 13.906105776881864, + "p95": 13.595446471404307, + "p99": 13.447074815973735 + }, + "samples_us": [ + 69.645, + 71.789, + 77.978, + 71.328, + 75.404, + 72.78, + 68.624, + 68.344, + 69.004, + 65.349, + 77.127, + 69.315, + 74.183, + 73.341, + 72.299, + 74.603, + 66.951, + 67.843, + 69.936, + 70.857 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 212.932, + "p90": 229.087, + "p95": 233.614, + "p99": 238.411 + }, + "payload_gbps_at_latency_percentile": { + "p50": 19.697856592715045, + "p90": 18.308782253030508, + "p95": 17.953992483327198, + "p99": 17.592745301181573 + }, + "samples_us": [ + 223.278, + 212.301, + 229.087, + 212.932, + 218.761, + 213.633, + 212.421, + 208.115, + 238.411, + 205.431, + 213.643, + 208.596, + 233.614, + 206.673, + 212.701, + 202.156, + 217.159, + 226.422, + 217.95, + 210.148 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 766.3, + "p90": 782.624, + "p95": 783.055, + "p99": 793.181 + }, + "payload_gbps_at_latency_percentile": { + "p50": 21.893796163382486, + "p90": 21.437134562701885, + "p95": 21.425335385126203, + "p99": 21.151812764047552 + }, + "samples_us": [ + 783.055, + 774.322, + 759.259, + 770.436, + 765.719, + 765.579, + 768.594, + 762.354, + 778.759, + 761.533, + 766.24, + 766.3, + 782.624, + 766.13, + 762.844, + 777.267, + 774.192, + 793.181, + 768.223, + 761.984 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2873.64, + "p90": 2889.534, + "p95": 2893.33, + "p99": 2898.517 + }, + "payload_gbps_at_latency_percentile": { + "p50": 23.35326067287482, + "p90": 23.224805106982647, + "p95": 23.19433455568497, + "p99": 23.152827463147535 + }, + "samples_us": [ + 2867.461, + 2877.847, + 2873.64, + 2862.313, + 2876.844, + 2878.227, + 2868.241, + 2888.122, + 2856.454, + 2867.461, + 2868.332, + 2898.517, + 2879.058, + 2879.389, + 2859.408, + 2872.488, + 2879.729, + 2893.33, + 2866.51, + 2889.534 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 134217728, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5733.759, + "p90": 5877.227, + "p95": 5893.651, + "p99": 5913.742 + }, + "payload_gbps_at_latency_percentile": { + "p50": 23.408330904734573, + "p90": 22.836914075294352, + "p95": 22.77327381617948, + "p99": 22.69590523225396 + }, + "samples_us": [ + 5729.583, + 5704.895, + 5712.578, + 5714.37, + 5756.003, + 5752.578, + 5768.352, + 5721.751, + 5692.507, + 5753.189, + 5776.854, + 5736.995, + 5765.557, + 5732.778, + 5729.042, + 5893.651, + 5877.227, + 5913.742, + 5733.759, + 5706.658 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 1, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 262144, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 15.393, + "p90": 16.195, + "p95": 20.23, + "p99": 22.363 + }, + "payload_gbps_at_latency_percentile": { + "p50": 17.030078607159098, + "p90": 16.186724297622725, + "p95": 12.958180919426594, + "p99": 11.722219737960023 + }, + "samples_us": [ + 16.195, + 15.634, + 15.484, + 15.393, + 15.884, + 15.483, + 15.193, + 15.273, + 15.113, + 15.173, + 14.852, + 15.874, + 16.094, + 15.844, + 14.882, + 15.023, + 20.23, + 15.143, + 15.263, + 22.363 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 4, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 23.275, + "p90": 24.066, + "p95": 24.988, + "p99": 25.529 + }, + "payload_gbps_at_latency_percentile": { + "p50": 45.05160042964554, + "p90": 43.57084683786255, + "p95": 41.96318232751721, + "p99": 41.07391593873634 + }, + "samples_us": [ + 23.546, + 23.326, + 23.104, + 22.695, + 24.988, + 23.696, + 25.529, + 23.405, + 23.025, + 23.275, + 24.066, + 22.945, + 23.205, + 23.786, + 23.065, + 23.105, + 23.225, + 23.285, + 23.095, + 23.275 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 16, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 70.747, + "p90": 77.737, + "p95": 79.841, + "p99": 86.641 + }, + "payload_gbps_at_latency_percentile": { + "p50": 59.28596265566031, + "p90": 53.95505357809023, + "p95": 52.533209754386846, + "p99": 48.41015223739338 + }, + "samples_us": [ + 71.509, + 70.747, + 67.853, + 71.899, + 77.737, + 86.641, + 71.227, + 73.952, + 69.185, + 66.561, + 66.511, + 69.736, + 76.646, + 72.64, + 70.006, + 68.283, + 65.249, + 68.553, + 72.64, + 79.841 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 64, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 210.128, + "p90": 235.506, + "p95": 242.196, + "p99": 246.824 + }, + "payload_gbps_at_latency_percentile": { + "p50": 79.84283865072717, + "p90": 71.2390172649529, + "p95": 69.27123486762787, + "p99": 67.9723851813438 + }, + "samples_us": [ + 246.824, + 242.196, + 235.506, + 206.313, + 217.219, + 209.307, + 213.693, + 208.586, + 216.618, + 210.128, + 222.788, + 204.46, + 214.665, + 198.58, + 207.554, + 210.208, + 210.248, + 201.365, + 205.23, + 201.655 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 256, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 757.637, + "p90": 767.482, + "p95": 777.117, + "p99": 785.369 + }, + "payload_gbps_at_latency_percentile": { + "p50": 88.57653995251025, + "p90": 87.44031000075572, + "p95": 86.35619089532207, + "p99": 85.44883233231768 + }, + "samples_us": [ + 751.978, + 751.808, + 785.369, + 759.029, + 759.951, + 755.363, + 755.734, + 767.482, + 763.726, + 777.117, + 763.556, + 757.917, + 761.673, + 747.622, + 752.178, + 761.723, + 757.637, + 751.017, + 754.983, + 751.688 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 1024, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 268435456, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2884.045, + "p90": 2906.14, + "p95": 2920.18, + "p99": 2921.473 + }, + "payload_gbps_at_latency_percentile": { + "p50": 93.07602898013033, + "p90": 92.36838417970229, + "p95": 91.92428411947209, + "p99": 91.88359981420331 + }, + "samples_us": [ + 2904.337, + 2867.14, + 2894.641, + 2883.434, + 2879.198, + 2884.045, + 2877.686, + 2876.795, + 2861.171, + 2873.109, + 2877.496, + 2890.456, + 2884.606, + 2890.956, + 2881.832, + 2921.473, + 2890.576, + 2906.14, + 2896.275, + 2920.18 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 2048, + "layout": "contiguous", + "seed": 0, + "payload_bytes": 536870912, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5773.439, + "p90": 5953.282, + "p95": 6011.38, + "p99": 6158.262 + }, + "payload_gbps_at_latency_percentile": { + "p50": 92.98979551009373, + "p90": 90.18066202810483, + "p95": 89.30909574839721, + "p99": 87.17896575364932 + }, + "samples_us": [ + 6158.262, + 6011.38, + 5928.584, + 5768.071, + 5762.664, + 5753.469, + 5783.475, + 5753.93, + 5806.29, + 5743.203, + 5756.384, + 5759.197, + 5758.007, + 5796.465, + 5792.017, + 5760.079, + 5793.169, + 5798.748, + 5773.439, + 5953.282 + ] + } + } + ] +} diff --git a/collectivex/results/mi325x/mi325x-swap-blocks-smoke-random.json b/collectivex/results/mi325x/mi325x-swap-blocks-smoke-random.json new file mode 100644 index 0000000000..656433d40e --- /dev/null +++ b/collectivex/results/mi325x/mi325x-swap-blocks-smoke-random.json @@ -0,0 +1,4075 @@ +{ + "schema": "collectivex-swap-blocks-v1", + "case_id": "mi325x-swap-blocks-smoke-random", + "profile": "smoke", + "generated_at": "2026-10-28T08:29:16.973529+00:00", + "operation": "vllm._custom_ops.swap_blocks", + "timing": "drained-wall-clock-including-submission-and-synchronization", + "warmup": 4, + "iterations": 20, + "selection": { + "requested_block_bytes": [ + 257, + 4096, + 65536, + 262144 + ], + "requested_num_blocks": [ + 1, + 4, + 16, + 64, + 256, + 1024, + 2048 + ], + "max_payload_bytes": 2147483648, + "skipped_cases": [] + }, + "runtime": { + "sku": "mi325x", + "torch": "2.13.0+git733fca1", + "vllm": "0.32.0", + "cuda": null, + "hip": "7.2.53211", + "device": "", + "device_index": 0, + "image": "vllm/vllm-openai-rocm:v0.32.0", + "source_sha": "6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4", + "source_dirty": false + }, + "cases": [ + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 257, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14.612, + "p90": 15.113, + "p95": 15.183, + "p99": 19.67 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.01758828360251848, + "p90": 0.0170052272877655, + "p95": 0.01692682605545676, + "p99": 0.013065582104728012 + }, + "samples_us": [ + 15.113, + 15.183, + 14.732, + 14.702, + 14.833, + 14.482, + 14.822, + 14.692, + 14.612, + 14.632, + 14.562, + 19.67, + 14.572, + 14.673, + 14.552, + 14.482, + 14.502, + 14.562, + 14.482, + 14.572 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 1028, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 21.282, + "p90": 21.543, + "p95": 21.672, + "p99": 24.046 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.0483037308523635, + "p90": 0.04771851645546117, + "p95": 0.047434477667035804, + "p99": 0.04275139316310405 + }, + "samples_us": [ + 21.543, + 21.543, + 21.672, + 21.383, + 21.413, + 21.272, + 21.142, + 21.112, + 21.142, + 21.282, + 21.202, + 21.323, + 21.302, + 21.332, + 21.282, + 21.282, + 21.132, + 21.122, + 21.232, + 24.046 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 4112, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 56.376, + "p90": 59.801, + "p95": 60.511, + "p99": 184.93 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.07293883922236413, + "p90": 0.06876139194996739, + "p95": 0.06795458676934772, + "p99": 0.022235440436922077 + }, + "samples_us": [ + 57.998, + 60.511, + 59.801, + 55.444, + 58.038, + 53.35, + 56.275, + 59.58, + 57.947, + 59.63, + 59.37, + 54.692, + 54.623, + 54.833, + 53.381, + 184.93, + 53.681, + 56.376, + 56.445, + 55.664 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 16448, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 173.633, + "p90": 184.058, + "p95": 185.141, + "p99": 197.329 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.09472853662610219, + "p90": 0.08936313553336449, + "p95": 0.08884039731880027, + "p99": 0.08335318174216663 + }, + "samples_us": [ + 170.498, + 185.141, + 197.329, + 184.009, + 166.432, + 178.941, + 171.339, + 184.058, + 173.633, + 182.937, + 169.046, + 179.982, + 168.184, + 183.728, + 168.165, + 173.122, + 167.995, + 180.243, + 172.702, + 182.947 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 65792, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 608.03, + "p90": 636.824, + "p95": 650.054, + "p99": 657.785 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.10820518724405045, + "p90": 0.10331268922025552, + "p95": 0.10121005331864738, + "p99": 0.10002052342330701 + }, + "samples_us": [ + 636.824, + 631.826, + 607.71, + 624.916, + 606.508, + 614.58, + 603.343, + 657.785, + 608.732, + 608.03, + 604.095, + 610.114, + 603.824, + 610.584, + 604.044, + 612.287, + 605.336, + 607.8, + 606.748, + 650.054 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 263168, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2428.456, + "p90": 2494.897, + "p95": 2527.205, + "p99": 2560.216 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.10836844480608254, + "p90": 0.10548251090125163, + "p95": 0.1041340136633158, + "p99": 0.1027913269817859 + }, + "samples_us": [ + 2560.216, + 2527.205, + 2432.703, + 2419.803, + 2448.587, + 2428.466, + 2475.007, + 2426.783, + 2414.174, + 2414.355, + 2401.655, + 2429.558, + 2379.502, + 2410.579, + 2426.763, + 2428.456, + 2398.541, + 2474.625, + 2494.897, + 2460.235 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 257, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 526336, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5097.437, + "p90": 5226.041, + "p95": 5231.99, + "p99": 5250.959 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.10325502796797685, + "p90": 0.10071409696173451, + "p95": 0.10059958065669086, + "p99": 0.10023616638408336 + }, + "samples_us": [ + 5189.816, + 5135.955, + 5030.314, + 5113.531, + 5152.339, + 5061.602, + 5226.041, + 5250.959, + 4878.324, + 4816.781, + 4847.178, + 4870.163, + 5092.259, + 5097.437, + 5165.87, + 5039.358, + 5182.325, + 5231.99, + 5118.178, + 5083.845 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 4096, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 16.144, + "p90": 16.305, + "p95": 16.886, + "p99": 18.798 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.25371655104063434, + "p90": 0.25121128488193806, + "p95": 0.24256780765130878, + "p99": 0.21789552080008512 + }, + "samples_us": [ + 16.295, + 16.245, + 16.886, + 16.195, + 16.254, + 16.204, + 15.834, + 16.134, + 16.104, + 16.305, + 16.074, + 16.144, + 18.798, + 15.904, + 16.144, + 15.834, + 15.904, + 16.284, + 15.854, + 15.924 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 16384, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 23.686, + "p90": 24.407, + "p95": 24.928, + "p99": 28.774 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.6917166258549354, + "p90": 0.6712828286966854, + "p95": 0.6572528883183568, + "p99": 0.5694029332035866 + }, + "samples_us": [ + 23.746, + 23.205, + 23.736, + 23.686, + 24.928, + 23.075, + 23.936, + 22.944, + 23.325, + 23.666, + 23.635, + 23.956, + 24.407, + 23.716, + 23.837, + 23.686, + 23.535, + 28.774, + 23.595, + 23.666 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 65.649, + "p90": 68.033, + "p95": 68.404, + "p99": 111.398 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.99827872473305, + "p90": 0.9632972234062881, + "p95": 0.958072627331735, + "p99": 0.5883049964990394 + }, + "samples_us": [ + 66.601, + 66.832, + 68.404, + 111.398, + 64.818, + 65.259, + 63.256, + 62.855, + 65.649, + 65.97, + 68.033, + 66.911, + 63.547, + 65.509, + 63.055, + 62.735, + 66.2, + 65.539, + 67.622, + 67.952 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 204.319, + "p90": 215.176, + "p95": 215.456, + "p99": 220.303 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.2830133271991346, + "p90": 1.2182771312785812, + "p95": 1.2166938957374127, + "p99": 1.1899247854091863 + }, + "samples_us": [ + 196.317, + 220.303, + 200.203, + 209.347, + 196.687, + 209.947, + 213.613, + 205.871, + 194.124, + 215.176, + 196.798, + 215.456, + 198.811, + 210.408, + 195.446, + 208.756, + 207.754, + 204.319, + 175.736, + 182.807 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 644.836, + "p90": 666.689, + "p95": 729.374, + "p99": 739.539 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.6261126860162893, + "p90": 1.5728113108210875, + "p95": 1.4376383035315214, + "p99": 1.417877894201658 + }, + "samples_us": [ + 729.374, + 739.539, + 646.999, + 644.877, + 662.322, + 642.953, + 639.087, + 644.395, + 644.836, + 638.297, + 644.837, + 638.657, + 640.209, + 647.21, + 638.597, + 666.689, + 642.102, + 639.919, + 655.172, + 654.711 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2588.028, + "p90": 2628.549, + "p95": 2648.89, + "p99": 2661.959 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.6206563452945641, + "p90": 1.5956727456859279, + "p95": 1.5834194700421687, + "p99": 1.5756456053605634 + }, + "samples_us": [ + 2617.232, + 2587.808, + 2623.402, + 2610.943, + 2618.955, + 2549.469, + 2628.549, + 2648.89, + 2529.419, + 2556.81, + 2588.028, + 2579.545, + 2558.453, + 2615.289, + 2606.335, + 2597.632, + 2661.959, + 2543.811, + 2532.724, + 2563.551 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 4096, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 8388608, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5183.186, + "p90": 5258.791, + "p95": 5265.04, + "p99": 5312.853 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.618426967506086, + "p90": 1.595159039406586, + "p95": 1.5932657681613056, + "p99": 1.5789271790505026 + }, + "samples_us": [ + 5227.373, + 5258.791, + 5139.109, + 5191.529, + 5146.05, + 5236.847, + 5231.91, + 5192.7, + 5130.176, + 5183.186, + 5133.431, + 5251.499, + 5265.04, + 5124.327, + 5312.853, + 5237.899, + 5143.096, + 5079.039, + 5066.199, + 5141.954 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 16.905, + "p90": 17.236, + "p95": 17.357, + "p99": 17.647 + }, + "payload_gbps_at_latency_percentile": { + "p50": 3.8767228630582666, + "p90": 3.8022743095845906, + "p95": 3.7757677017917843, + "p99": 3.7137190457301523 + }, + "samples_us": [ + 17.236, + 16.465, + 17.136, + 17.357, + 17.016, + 16.795, + 16.936, + 16.676, + 17.236, + 16.836, + 16.505, + 16.786, + 17.016, + 16.765, + 17.647, + 16.905, + 17.016, + 16.805, + 17.146, + 16.866 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 61.773, + "p90": 61.953, + "p95": 62.535, + "p99": 66.941 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.243666326712318, + "p90": 4.2313366584346195, + "p95": 4.1919565043575595, + "p99": 3.9160454728791025 + }, + "samples_us": [ + 61.092, + 59.76, + 61.773, + 62.535, + 60.922, + 61.453, + 61.774, + 61.853, + 61.844, + 61.953, + 61.864, + 61.613, + 61.813, + 61.834, + 61.744, + 61.764, + 61.533, + 61.483, + 66.941, + 61.943 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 249.017, + "p90": 260.615, + "p95": 268.747, + "p99": 272.443 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.210861105868274, + "p90": 4.023467567100896, + "p95": 3.9017216936375103, + "p99": 3.8487903891823243 + }, + "samples_us": [ + 254.515, + 245.952, + 255.206, + 246.694, + 249.017, + 268.747, + 246.423, + 244.119, + 247.675, + 247.475, + 260.615, + 245.502, + 244.4, + 258.701, + 246.804, + 251.23, + 252.292, + 254.244, + 272.443, + 252.952 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 920.694, + "p90": 927.564, + "p95": 927.855, + "p99": 928.936 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.55558958785438, + "p90": 4.521848627156724, + "p95": 4.520430455189658, + "p99": 4.515170044007338 + }, + "samples_us": [ + 924.289, + 924.68, + 927.855, + 926.492, + 927.564, + 928.936, + 918.53, + 920.644, + 918.13, + 922.026, + 918.7, + 922.556, + 924.139, + 920.052, + 919.122, + 921.775, + 918.971, + 918.41, + 920.694, + 918.961 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3726.391, + "p90": 3729.816, + "p95": 3730.236, + "p99": 3746.481 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.502269353913746, + "p90": 4.498135028644844, + "p95": 4.4976285682728925, + "p99": 4.478126540612378 + }, + "samples_us": [ + 3716.566, + 3724.538, + 3722.044, + 3725.209, + 3725.279, + 3722.555, + 3726.391, + 3724.458, + 3727.632, + 3730.236, + 3729.766, + 3746.481, + 3726.631, + 3724.198, + 3729.816, + 3725.649, + 3727.783, + 3729.706, + 3726.902, + 3728.063 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14982.62, + "p90": 14992.665, + "p95": 14992.806, + "p99": 14994.728 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.479114066832103, + "p90": 4.476113085965704, + "p95": 4.476070990313621, + "p99": 4.475497254768476 + }, + "samples_us": [ + 14992.204, + 14985.094, + 14981.378, + 14982.62, + 14985.304, + 14978.444, + 14978.373, + 14981.568, + 14986.116, + 14992.806, + 14977.231, + 14988.078, + 14994.728, + 14983.361, + 14972.455, + 14980.066, + 14985.143, + 14992.665, + 14975.038, + 14976.741 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 65536, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 134217728, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 29853.731, + "p90": 29885.739, + "p95": 29895.404, + "p99": 29904.808 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.495844355266684, + "p90": 4.491029249770267, + "p95": 4.4895773276721735, + "p99": 4.488165515056977 + }, + "samples_us": [ + 29895.404, + 29885.739, + 29880.261, + 29904.808, + 29881.623, + 29878.849, + 29881.343, + 29860.06, + 29846.981, + 29856.294, + 29848.823, + 29853.731, + 29855.894, + 29846.82, + 29852.229, + 29843.065, + 29843.606, + 29835.293, + 29829.794, + 29804.276 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 20.451, + "p90": 20.702, + "p95": 20.762, + "p99": 20.771 + }, + "payload_gbps_at_latency_percentile": { + "p50": 12.81815070167718, + "p90": 12.662737899719835, + "p95": 12.626143916771024, + "p99": 12.620673053776901 + }, + "samples_us": [ + 20.241, + 20.24, + 19.96, + 20.551, + 20.771, + 20.381, + 20.521, + 20.451, + 20.26, + 20.702, + 20.561, + 20.431, + 20.511, + 20.351, + 20.762, + 20.21, + 20.461, + 20.421, + 20.551, + 20.651 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 82.505, + "p90": 83.246, + "p95": 83.667, + "p99": 83.747 + }, + "payload_gbps_at_latency_percentile": { + "p50": 12.709241864129448, + "p90": 12.596112726136992, + "p95": 12.53273094529504, + "p99": 12.520758952559495 + }, + "samples_us": [ + 82.274, + 82.585, + 82.234, + 82.104, + 81.854, + 83.747, + 82.075, + 83.667, + 82.264, + 82.395, + 83.095, + 81.914, + 82.996, + 83.246, + 82.295, + 83.015, + 83.026, + 82.625, + 82.985, + 82.505 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 304.181, + "p90": 307.255, + "p95": 307.565, + "p99": 308.748 + }, + "payload_gbps_at_latency_percentile": { + "p50": 13.78884282713253, + "p90": 13.650889326455225, + "p95": 13.637130362687563, + "p99": 13.584878282612356 + }, + "samples_us": [ + 303.9, + 301.176, + 299.042, + 303.259, + 301.026, + 300.985, + 302.618, + 304.07, + 305.252, + 304.181, + 303.038, + 308.748, + 307.565, + 307.255, + 307.085, + 307.145, + 307.195, + 304.311, + 304.24, + 305.422 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 1320.238, + "p90": 1327.38, + "p95": 1331.095, + "p99": 1332.837 + }, + "payload_gbps_at_latency_percentile": { + "p50": 12.707720880629099, + "p90": 12.639346682939324, + "p95": 12.604071084332825, + "p99": 12.587597733256205 + }, + "samples_us": [ + 1332.837, + 1331.095, + 1327.38, + 1326.208, + 1322.332, + 1300.108, + 1327.019, + 1320.238, + 1317.324, + 1317.124, + 1309.993, + 1323.503, + 1319.427, + 1311.906, + 1319.387, + 1320.739, + 1320.939, + 1312.918, + 1320.76, + 1316.023 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5394.566, + "p90": 5410.961, + "p95": 5419.344, + "p99": 5425.502 + }, + "payload_gbps_at_latency_percentile": { + "p50": 12.440085819693373, + "p90": 12.402392846668088, + "p95": 12.383208004511246, + "p99": 12.369152937368744 + }, + "samples_us": [ + 5394.746, + 5425.502, + 5396.789, + 5387.215, + 5403.059, + 5419.344, + 5399.945, + 5388.277, + 5394.566, + 5391.021, + 5401.847, + 5397.43, + 5405.162, + 5377.491, + 5410.961, + 5376.489, + 5384.411, + 5383.379, + 5365.452, + 5357.41 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 268435456, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 20031.683, + "p90": 20153.747, + "p95": 20157.363, + "p99": 20191.065 + }, + "payload_gbps_at_latency_percentile": { + "p50": 13.400544327703269, + "p90": 13.319382048410155, + "p95": 13.316992703857146, + "p99": 13.294764590178874 + }, + "samples_us": [ + 20191.065, + 20139.276, + 20153.747, + 20093.166, + 20157.363, + 20094.658, + 20044.362, + 20004.843, + 20047.527, + 20045.724, + 20034.537, + 20027.827, + 20023.891, + 19982.649, + 19991.192, + 19965.253, + 20031.683, + 19938.602, + 19930.009, + 19921.476 + ] + } + }, + { + "direction": "h2d", + "block_bytes": 262144, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 536870912, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 37738.236, + "p90": 37935.986, + "p95": 37939.781, + "p99": 37975.165 + }, + "payload_gbps_at_latency_percentile": { + "p50": 14.226179305254226, + "p90": 14.15202209321777, + "p95": 14.150606509826718, + "p99": 14.13742144372513 + }, + "samples_us": [ + 37911.729, + 37939.781, + 37935.986, + 37871.437, + 37975.165, + 37875.163, + 37893.431, + 37860.481, + 37899.059, + 37756.113, + 37666.046, + 37666.017, + 37725.697, + 37689.432, + 37704.465, + 37713.128, + 37717.564, + 37738.236, + 37645.596, + 37649.642 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 257, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 13.961, + "p90": 14.873, + "p95": 14.992, + "p99": 15.213 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.01840842346536781, + "p90": 0.017279634236536005, + "p95": 0.017142475987193168, + "p99": 0.016893446394530994 + }, + "samples_us": [ + 14.992, + 13.831, + 13.631, + 14.042, + 15.213, + 13.821, + 13.911, + 14.873, + 14.693, + 13.972, + 14.853, + 13.851, + 13.901, + 13.77, + 13.961, + 14.111, + 14.251, + 14.111, + 13.661, + 13.731 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 1028, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 21.072, + "p90": 21.492, + "p95": 23.074, + "p99": 24.217 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.04878511769172361, + "p90": 0.04783175134933929, + "p95": 0.04455230995926151, + "p99": 0.04244951893298096 + }, + "samples_us": [ + 21.382, + 21.132, + 20.591, + 20.811, + 21.262, + 21.021, + 20.651, + 20.831, + 21.162, + 21.072, + 24.217, + 21.402, + 21.492, + 21.312, + 20.671, + 20.651, + 21.342, + 23.074, + 20.691, + 20.892 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 4112, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 57.176, + "p90": 62.114, + "p95": 62.174, + "p99": 65.499 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.07191828739331188, + "p90": 0.06620085648968027, + "p95": 0.06613697043780359, + "p99": 0.06277958442113621 + }, + "samples_us": [ + 56.185, + 60.101, + 58.098, + 60.161, + 62.174, + 55.124, + 53.2, + 57.888, + 54.192, + 65.499, + 62.114, + 55.804, + 57.176, + 53.031, + 55.814, + 61.974, + 54.732, + 61.323, + 59.68, + 56.636 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 16448, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 186.913, + "p90": 198.12, + "p95": 199.832, + "p99": 209.998 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08799815957156538, + "p90": 0.08302039168180901, + "p95": 0.08230913967732896, + "p99": 0.07832455547195688 + }, + "samples_us": [ + 177.228, + 194.064, + 185.08, + 186.913, + 178.5, + 196.979, + 186.322, + 196.077, + 180.333, + 199.832, + 185.741, + 189.937, + 178.82, + 198.12, + 180.573, + 189.607, + 181.084, + 192.812, + 209.998, + 196.147 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 65792, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 692.949, + "p90": 707.491, + "p95": 788.183, + "p99": 822.335 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.09494493822777722, + "p90": 0.09299340910343736, + "p95": 0.08347300055951473, + "p99": 0.08000632345698529 + }, + "samples_us": [ + 822.335, + 788.183, + 707.491, + 698.888, + 694.231, + 693.93, + 695.443, + 691.516, + 703.876, + 699.488, + 688.362, + 691.176, + 692.949, + 684.366, + 689.984, + 694.381, + 689.844, + 690.475, + 691.447, + 689.734 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 263168, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2705.465, + "p90": 2715.341, + "p95": 2715.921, + "p99": 2721.49 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.09727274239363658, + "p90": 0.0969189505111881, + "p95": 0.09689825293151016, + "p99": 0.09669996950200074 + }, + "samples_us": [ + 2712.205, + 2715.921, + 2696.602, + 2703.022, + 2692.535, + 2715.341, + 2704.063, + 2711.735, + 2710.553, + 2707.118, + 2708.67, + 2705.465, + 2702.771, + 2707.348, + 2721.49, + 2699.977, + 2698.645, + 2690.293, + 2708.911, + 2694.679 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 257, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 526336, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5385.132, + "p90": 5507.917, + "p95": 5516.28, + "p99": 5573.216 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.09773873695203757, + "p90": 0.09555990041244267, + "p95": 0.0954150260682924, + "p99": 0.09444026572808231 + }, + "samples_us": [ + 5466.385, + 5507.917, + 5394.836, + 5385.132, + 5375.337, + 5391.832, + 5385.423, + 5403.319, + 5516.28, + 5363.369, + 5361.295, + 5473.365, + 5404.191, + 5372.973, + 5378.782, + 5384.801, + 5367.485, + 5358.531, + 5573.216, + 5363.92 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 4096, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 15.914, + "p90": 16.104, + "p95": 16.214, + "p99": 16.214 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.2573834359683298, + "p90": 0.25434674615002484, + "p95": 0.25262119156284696, + "p99": 0.25262119156284696 + }, + "samples_us": [ + 15.854, + 16.214, + 16.104, + 15.914, + 15.824, + 15.864, + 15.904, + 15.955, + 15.774, + 15.954, + 16.054, + 16.024, + 15.694, + 15.914, + 16.034, + 16.214, + 15.684, + 15.945, + 15.934, + 15.904 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 16384, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 23.636, + "p90": 23.987, + "p95": 24.016, + "p99": 24.557 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.6931798950753089, + "p90": 0.6830366448492934, + "p95": 0.6822118587608261, + "p99": 0.6671824734291648 + }, + "samples_us": [ + 23.987, + 23.977, + 23.755, + 23.886, + 23.986, + 23.596, + 24.557, + 23.585, + 23.636, + 23.555, + 23.576, + 23.475, + 23.636, + 23.365, + 23.505, + 23.616, + 24.016, + 23.706, + 23.556, + 23.946 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 69.185, + "p90": 73.522, + "p95": 77.638, + "p99": 80.352 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.9472573534725736, + "p90": 0.8913794510486657, + "p95": 0.8441227234086401, + "p99": 0.8156113102349661 + }, + "samples_us": [ + 69.185, + 69.055, + 71.137, + 72.34, + 80.352, + 68.494, + 65.509, + 65.119, + 71.108, + 67.522, + 71.218, + 71.068, + 67.712, + 68.453, + 65.59, + 65.429, + 69.635, + 69.495, + 77.638, + 73.522 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 216.728, + "p90": 233.774, + "p95": 235.126, + "p99": 236.598 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.2095529880772211, + "p90": 1.121356523822153, + "p95": 1.1149086021962693, + "p99": 1.107972172207711 + }, + "samples_us": [ + 236.598, + 235.126, + 213.954, + 217.549, + 206.252, + 233.774, + 206.313, + 220.855, + 205.09, + 223.088, + 219.392, + 218.621, + 210.318, + 229.678, + 218.931, + 216.728, + 198.801, + 195.326, + 183.468, + 196.367 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 692.358, + "p90": 699.779, + "p95": 775.664, + "p99": 801.884 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.5144997241311575, + "p90": 1.4984387928188756, + "p95": 1.3518430660698446, + "p99": 1.3076405066069408 + }, + "samples_us": [ + 801.884, + 775.664, + 695.333, + 699.679, + 690.345, + 695.443, + 687.481, + 699.779, + 696.404, + 692.188, + 688.442, + 690.856, + 689.123, + 692.358, + 697.336, + 689.744, + 692.949, + 687.0, + 691.387, + 699.448 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2689.712, + "p90": 2699.216, + "p95": 2699.636, + "p99": 2711.385 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.5593877708840203, + "p90": 1.5538971316115495, + "p95": 1.5536553816884944, + "p99": 1.5469230669934368 + }, + "samples_us": [ + 2697.534, + 2680.378, + 2692.536, + 2690.783, + 2693.988, + 2699.216, + 2690.753, + 2688.119, + 2693.627, + 2687.729, + 2699.636, + 2689.221, + 2687.498, + 2689.231, + 2711.385, + 2685.635, + 2688.249, + 2696.091, + 2689.712, + 2682.942 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 4096, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 8388608, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5387.786, + "p90": 5501.408, + "p95": 5510.711, + "p99": 5511.863 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.556967555875456, + "p90": 1.5248111029031113, + "p95": 1.5222369672443357, + "p99": 1.5219188140198696 + }, + "samples_us": [ + 5356.589, + 5362.948, + 5360.855, + 5358.962, + 5485.364, + 5511.863, + 5387.786, + 5378.942, + 5375.027, + 5395.377, + 5476.56, + 5510.711, + 5494.828, + 5376.258, + 5372.713, + 5397.931, + 5450.15, + 5501.408, + 5482.079, + 5361.656 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 65536, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 16.054, + "p90": 16.255, + "p95": 16.325, + "p99": 16.345 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.0822224990656535, + "p90": 4.031744078745002, + "p95": 4.014456355283308, + "p99": 4.009544203120221 + }, + "samples_us": [ + 16.064, + 16.345, + 15.904, + 16.255, + 16.194, + 15.694, + 16.154, + 15.834, + 15.664, + 16.184, + 15.994, + 15.924, + 16.154, + 16.325, + 15.844, + 16.054, + 16.054, + 15.774, + 16.104, + 16.035 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 56.436, + "p90": 56.976, + "p95": 57.037, + "p99": 57.347 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.644978382592671, + "p90": 4.600954787980904, + "p95": 4.59603415326893, + "p99": 4.571189425776414 + }, + "samples_us": [ + 56.906, + 55.484, + 56.255, + 56.466, + 56.505, + 56.976, + 56.255, + 56.436, + 56.675, + 57.347, + 56.646, + 56.305, + 56.555, + 56.395, + 57.037, + 56.296, + 56.436, + 56.105, + 55.984, + 56.476 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 228.045, + "p90": 229.327, + "p95": 230.779, + "p99": 230.859 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.598110022144752, + "p90": 4.572405342589403, + "p95": 4.543636986034258, + "p99": 4.542062471032102 + }, + "samples_us": [ + 227.284, + 226.623, + 228.045, + 228.866, + 228.947, + 226.092, + 225.461, + 228.055, + 228.466, + 226.863, + 228.396, + 228.085, + 228.256, + 227.254, + 226.382, + 226.133, + 229.327, + 227.104, + 230.859, + 230.779 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 841.885, + "p90": 844.439, + "p95": 844.449, + "p99": 846.641 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.982039114605914, + "p90": 4.966970971260209, + "p95": 4.966912152184443, + "p99": 4.954052544112558 + }, + "samples_us": [ + 841.774, + 841.885, + 842.115, + 841.885, + 840.582, + 840.392, + 841.965, + 842.385, + 846.641, + 840.062, + 844.198, + 840.953, + 844.288, + 843.036, + 842.626, + 844.449, + 839.52, + 844.439, + 840.071, + 841.784 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 3333.076, + "p90": 3335.88, + "p95": 3336.0, + "p99": 3338.434 + }, + "payload_gbps_at_latency_percentile": { + "p50": 5.033553390321733, + "p90": 5.0293223976881665, + "p95": 5.029141486810552, + "p99": 5.025474818432834 + }, + "samples_us": [ + 3332.805, + 3333.076, + 3335.88, + 3334.227, + 3335.679, + 3332.134, + 3335.77, + 3328.368, + 3329.379, + 3331.594, + 3332.975, + 3330.462, + 3334.808, + 3328.609, + 3336.0, + 3338.434, + 3334.267, + 3335.659, + 3330.181, + 3335.239 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 13370.439, + "p90": 13376.699, + "p95": 13379.003, + "p99": 13386.443 + }, + "payload_gbps_at_latency_percentile": { + "p50": 5.019196751879276, + "p90": 5.016847878538644, + "p95": 5.015983926455506, + "p99": 5.013196111917109 + }, + "samples_us": [ + 13375.768, + 13379.003, + 13362.387, + 13369.859, + 13373.664, + 13370.439, + 13371.401, + 13367.796, + 13372.042, + 13374.686, + 13376.699, + 13365.251, + 13371.121, + 13364.651, + 13368.807, + 13371.932, + 13368.346, + 13386.443, + 13369.838, + 13363.289 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 65536, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 134217728, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 26695.95, + "p90": 26702.902, + "p95": 26703.152, + "p99": 26703.933 + }, + "payload_gbps_at_latency_percentile": { + "p50": 5.027643818631665, + "p90": 5.026334890492427, + "p95": 5.026287832986907, + "p99": 5.026140831015416 + }, + "samples_us": [ + 26702.902, + 26703.933, + 26691.074, + 26697.943, + 26693.477, + 26695.139, + 26697.423, + 26699.276, + 26693.607, + 26695.95, + 26682.431, + 26703.152, + 26702.3, + 26694.298, + 26699.406, + 26692.315, + 26700.348, + 26694.959, + 26697.594, + 26692.786 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 18.038, + "p90": 18.298, + "p95": 19.139, + "p99": 35.623 + }, + "payload_gbps_at_latency_percentile": { + "p50": 14.532875041578889, + "p90": 14.32637446715488, + "p95": 13.696849365170594, + "p99": 7.35884119810235 + }, + "samples_us": [ + 18.278, + 18.298, + 18.038, + 17.948, + 18.018, + 35.623, + 18.208, + 18.168, + 18.017, + 18.277, + 19.139, + 17.937, + 18.097, + 18.038, + 18.077, + 17.706, + 17.747, + 17.847, + 18.098, + 18.018 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 73.01, + "p90": 73.822, + "p95": 73.882, + "p99": 73.952 + }, + "payload_gbps_at_latency_percentile": { + "p50": 14.362087385289687, + "p90": 14.204112595161334, + "p95": 14.192577353076528, + "p99": 14.179143228039809 + }, + "samples_us": [ + 73.181, + 73.712, + 73.882, + 72.71, + 73.701, + 73.01, + 73.021, + 71.489, + 73.431, + 73.822, + 73.501, + 73.281, + 72.549, + 73.001, + 72.96, + 72.98, + 72.941, + 71.889, + 73.952, + 72.9 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 270.769, + "p90": 274.456, + "p95": 274.746, + "p99": 275.507 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.490340474721995, + "p90": 15.282245605853033, + "p95": 15.26611488429313, + "p99": 15.223947122940615 + }, + "samples_us": [ + 270.769, + 267.365, + 271.01, + 271.291, + 269.147, + 267.606, + 271.171, + 271.732, + 269.347, + 266.493, + 274.746, + 274.456, + 275.507, + 269.117, + 272.883, + 270.8, + 269.918, + 268.797, + 268.526, + 271.3 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 1071.031, + "p90": 1074.627, + "p95": 1075.698, + "p99": 1076.78 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.66454752476819, + "p90": 15.612129604039355, + "p95": 15.596585658800146, + "p99": 15.580913464217389 + }, + "samples_us": [ + 1067.796, + 1064.761, + 1069.359, + 1065.022, + 1071.031, + 1071.171, + 1072.353, + 1071.101, + 1069.559, + 1073.916, + 1071.512, + 1071.342, + 1074.627, + 1072.864, + 1069.549, + 1076.78, + 1068.768, + 1066.965, + 1075.698, + 1070.711 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 4264.796, + "p90": 4269.022, + "p95": 4270.795, + "p99": 4271.496 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.73553905040241, + "p90": 15.719962089677683, + "p95": 15.713436023035523, + "p99": 15.710857273423644 + }, + "samples_us": [ + 4266.178, + 4262.272, + 4266.688, + 4266.018, + 4265.386, + 4264.155, + 4265.027, + 4264.685, + 4265.337, + 4263.895, + 4268.191, + 4264.626, + 4264.385, + 4263.554, + 4264.796, + 4263.564, + 4269.022, + 4261.861, + 4271.496, + 4270.795 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 268435456, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 17034.926, + "p90": 17044.471, + "p95": 17045.303, + "p99": 17045.633 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.757946703143881, + "p90": 15.74912216401436, + "p95": 15.74835343202758, + "p99": 15.748048547097078 + }, + "samples_us": [ + 17045.303, + 17045.633, + 17026.854, + 17031.962, + 17039.744, + 17034.926, + 17033.935, + 17042.649, + 17036.659, + 17037.76, + 17023.399, + 17031.24, + 17043.65, + 17044.471, + 17022.517, + 17034.486, + 17038.422, + 17037.551, + 17029.408, + 17033.745 + ] + } + }, + { + "direction": "d2h", + "block_bytes": 262144, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 536870912, + "host_memory": "pinned", + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 34013.097, + "p90": 34022.12, + "p95": 34023.072, + "p99": 34026.577 + }, + "payload_gbps_at_latency_percentile": { + "p50": 15.78424075878771, + "p90": 15.780054623286262, + "p95": 15.779613081381951, + "p99": 15.777987659469831 + }, + "samples_us": [ + 34014.038, + 34019.868, + 34011.124, + 34008.32, + 34023.072, + 34003.853, + 34015.931, + 34016.302, + 34026.577, + 34012.436, + 34022.12, + 34011.555, + 34012.646, + 34014.099, + 34015.771, + 34002.921, + 34014.249, + 34006.357, + 34013.097, + 34004.664 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 257, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14.241, + "p90": 14.852, + "p95": 14.973, + "p99": 15.063 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.018046485499613792, + "p90": 0.0173040667923512, + "p95": 0.017164228945435116, + "p99": 0.01706167430126801 + }, + "samples_us": [ + 14.973, + 15.063, + 14.162, + 14.852, + 14.362, + 14.091, + 14.411, + 13.891, + 14.021, + 14.783, + 14.242, + 14.091, + 13.991, + 14.241, + 14.342, + 14.261, + 14.161, + 14.332, + 14.141, + 14.091 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 1028, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 20.721, + "p90": 21.693, + "p95": 23.987, + "p99": 26.551 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.049611505236233776, + "p90": 0.04738855852118195, + "p95": 0.042856547296452246, + "p99": 0.03871793906067568 + }, + "samples_us": [ + 20.912, + 20.711, + 23.987, + 20.792, + 20.471, + 21.142, + 20.451, + 20.551, + 20.791, + 21.693, + 26.551, + 20.351, + 21.262, + 20.461, + 20.622, + 21.683, + 20.441, + 20.721, + 20.491, + 21.162 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 4112, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 58.218, + "p90": 62.505, + "p95": 63.847, + "p99": 64.168 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.07063107629942629, + "p90": 0.06578673706103512, + "p95": 0.06440396573057465, + "p99": 0.06408178531355192 + }, + "samples_us": [ + 60.151, + 58.468, + 62.505, + 54.793, + 63.847, + 51.468, + 55.825, + 53.911, + 58.399, + 57.096, + 59.25, + 54.813, + 58.599, + 52.088, + 55.544, + 55.304, + 60.221, + 58.218, + 64.168, + 59.911 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 16448, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 189.166, + "p90": 196.277, + "p95": 196.968, + "p99": 200.544 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.08695008616770455, + "p90": 0.08379993580501027, + "p95": 0.08350595020510945, + "p99": 0.0820169139939365 + }, + "samples_us": [ + 181.375, + 192.882, + 179.622, + 189.166, + 183.698, + 196.277, + 180.944, + 192.111, + 182.907, + 200.544, + 183.468, + 190.449, + 196.968, + 193.773, + 185.761, + 190.027, + 179.462, + 195.856, + 186.352, + 195.697 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 65792, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 702.644, + "p90": 714.441, + "p95": 734.421, + "p99": 737.687 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.09363489903848891, + "p90": 0.0920887798992499, + "p95": 0.08958349502533289, + "p99": 0.08918687736126568 + }, + "samples_us": [ + 734.421, + 737.687, + 714.282, + 714.441, + 694.201, + 702.644, + 691.977, + 711.367, + 706.94, + 697.626, + 693.509, + 699.309, + 707.641, + 710.315, + 691.747, + 697.496, + 692.278, + 705.188, + 706.018, + 700.04 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 263168, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2711.494, + "p90": 2740.479, + "p95": 2745.195, + "p99": 2745.486 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.09705645669877934, + "p90": 0.09602992761484397, + "p95": 0.09586495676991981, + "p99": 0.09585479583578281 + }, + "samples_us": [ + 2707.058, + 2740.479, + 2703.983, + 2705.215, + 2718.665, + 2708.84, + 2712.135, + 2725.406, + 2715.01, + 2745.195, + 2708.611, + 2729.241, + 2745.486, + 2711.494, + 2704.935, + 2706.867, + 2723.112, + 2710.923, + 2705.926, + 2712.126 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 257, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 526336, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5400.114, + "p90": 5417.1, + "p95": 5418.342, + "p99": 5513.717 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.0974675719808878, + "p90": 0.0971619501209134, + "p95": 0.09713967852158464, + "p99": 0.0954593788545912 + }, + "samples_us": [ + 5410.49, + 5374.215, + 5418.342, + 5409.458, + 5409.218, + 5391.552, + 5417.1, + 5398.732, + 5400.114, + 5407.215, + 5414.977, + 5389.659, + 5513.717, + 5396.559, + 5379.763, + 5388.427, + 5366.524, + 5369.418, + 5408.697, + 5402.999 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 4096, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14.442, + "p90": 14.862, + "p95": 15.003, + "p99": 15.133 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.2836172275308129, + "p90": 0.275602206970798, + "p95": 0.2730120642538159, + "p99": 0.2706667547743342 + }, + "samples_us": [ + 14.172, + 14.842, + 15.003, + 15.133, + 14.572, + 14.582, + 14.732, + 14.071, + 14.251, + 14.071, + 14.132, + 13.801, + 14.742, + 14.842, + 14.021, + 14.862, + 14.442, + 13.932, + 13.921, + 14.702 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 16384, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 20.601, + "p90": 20.841, + "p95": 20.921, + "p99": 20.962 + }, + "payload_gbps_at_latency_percentile": { + "p50": 0.7953011989709238, + "p90": 0.7861426994865889, + "p95": 0.7831365613498399, + "p99": 0.7816048087014598 + }, + "samples_us": [ + 20.921, + 20.401, + 20.801, + 20.321, + 20.792, + 20.751, + 20.841, + 20.831, + 20.541, + 20.601, + 20.841, + 20.751, + 20.771, + 20.421, + 20.572, + 20.341, + 20.401, + 20.501, + 20.551, + 20.962 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 65536, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 60.061, + "p90": 64.378, + "p95": 69.685, + "p99": 77.428 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.091157323387889, + "p90": 1.017987511261611, + "p95": 0.9404606443280477, + "p99": 0.8464121506431782 + }, + "samples_us": [ + 61.523, + 61.002, + 64.378, + 77.428, + 59.5, + 60.471, + 56.646, + 58.458, + 58.629, + 59.64, + 62.154, + 60.762, + 57.127, + 58.408, + 59.59, + 55.103, + 61.744, + 60.061, + 61.403, + 69.685 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 193.884, + "p90": 209.417, + "p95": 211.53, + "p99": 213.203 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.3520661839037775, + "p90": 1.2517799414565198, + "p95": 1.2392757528482958, + "p99": 1.2295511789233735 + }, + "samples_us": [ + 183.938, + 203.007, + 194.364, + 193.884, + 179.181, + 197.208, + 182.196, + 211.53, + 183.648, + 198.55, + 183.859, + 194.544, + 209.417, + 198.43, + 183.428, + 196.056, + 185.13, + 213.203, + 184.74, + 191.15 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 699.339, + "p90": 723.786, + "p95": 729.554, + "p99": 733.13 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.4993815588720205, + "p90": 1.4487376102881238, + "p95": 1.4372836006656122, + "p99": 1.4302729393149918 + }, + "samples_us": [ + 702.203, + 696.054, + 713.23, + 697.426, + 700.881, + 729.554, + 701.812, + 691.597, + 693.98, + 702.654, + 702.894, + 723.786, + 695.312, + 699.339, + 698.497, + 698.668, + 698.708, + 733.13, + 697.165, + 700.6 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2705.336, + "p90": 2711.204, + "p95": 2712.095, + "p99": 2714.389 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.550381911895602, + "p90": 1.5470263395893484, + "p95": 1.5465180976330106, + "p99": 1.5452110953883176 + }, + "samples_us": [ + 2704.374, + 2704.825, + 2710.122, + 2685.896, + 2690.904, + 2698.525, + 2706.206, + 2714.389, + 2708.85, + 2707.489, + 2705.336, + 2709.933, + 2706.277, + 2705.285, + 2701.85, + 2696.211, + 2710.573, + 2711.204, + 2712.095, + 2695.029 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 4096, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 8388608, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5401.947, + "p90": 5525.845, + "p95": 5531.393, + "p99": 5532.315 + }, + "payload_gbps_at_latency_percentile": { + "p50": 1.5528860242427407, + "p90": 1.5180679154047934, + "p95": 1.516545289766972, + "p99": 1.5162925466102346 + }, + "samples_us": [ + 5500.877, + 5378.602, + 5514.077, + 5401.947, + 5531.393, + 5381.276, + 5380.925, + 5410.05, + 5422.628, + 5525.845, + 5367.495, + 5368.477, + 5397.37, + 5397.7, + 5532.315, + 5511.473, + 5496.56, + 5369.979, + 5410.901, + 5374.065 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 65536, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14.361, + "p90": 14.822, + "p95": 16.064, + "p99": 18.047 + }, + "payload_gbps_at_latency_percentile": { + "p50": 4.563470510410139, + "p90": 4.421535555255701, + "p95": 4.079681274900398, + "p99": 3.6314068820302543 + }, + "samples_us": [ + 14.452, + 14.793, + 14.822, + 14.352, + 14.202, + 14.391, + 14.352, + 18.047, + 14.261, + 14.702, + 14.642, + 14.312, + 13.801, + 14.712, + 14.262, + 14.162, + 14.361, + 16.064, + 14.332, + 14.422 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 21.112, + "p90": 21.522, + "p95": 21.753, + "p99": 21.823 + }, + "payload_gbps_at_latency_percentile": { + "p50": 12.416824554755589, + "p90": 12.180280643062913, + "p95": 12.050935503148992, + "p99": 12.012280621362782 + }, + "samples_us": [ + 21.312, + 21.823, + 21.262, + 21.522, + 21.212, + 21.252, + 21.022, + 21.182, + 21.193, + 21.002, + 21.042, + 21.753, + 21.002, + 21.012, + 20.811, + 21.142, + 20.991, + 21.102, + 21.032, + 21.112 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 61.603, + "p90": 67.563, + "p95": 68.063, + "p99": 69.164 + }, + "payload_gbps_at_latency_percentile": { + "p50": 17.02150869275847, + "p90": 15.519973950239036, + "p95": 15.405962123326917, + "p99": 15.160719449424557 + }, + "samples_us": [ + 67.563, + 69.164, + 64.568, + 67.002, + 59.631, + 61.423, + 58.338, + 56.706, + 61.603, + 59.139, + 63.887, + 62.184, + 57.787, + 62.224, + 55.704, + 58.559, + 57.807, + 68.063, + 65.259, + 62.435 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 191.941, + "p90": 200.313, + "p95": 201.214, + "p99": 206.322 + }, + "payload_gbps_at_latency_percentile": { + "p50": 21.852048285671117, + "p90": 20.938750854912062, + "p95": 20.844990905205403, + "p99": 20.328922751815124 + }, + "samples_us": [ + 182.045, + 198.8, + 183.268, + 198.22, + 182.406, + 200.313, + 189.377, + 193.112, + 191.941, + 201.214, + 189.407, + 193.623, + 185.701, + 206.322, + 185.33, + 194.986, + 188.646, + 199.722, + 191.39, + 192.251 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 686.83, + "p90": 707.952, + "p95": 709.293, + "p99": 712.538 + }, + "payload_gbps_at_latency_percentile": { + "p50": 24.42702852234177, + "p90": 23.69823942866183, + "p95": 23.65343518122976, + "p99": 23.545714053145236 + }, + "samples_us": [ + 709.293, + 707.952, + 686.709, + 691.036, + 680.26, + 685.778, + 685.518, + 683.444, + 686.249, + 712.538, + 687.281, + 688.302, + 670.606, + 687.962, + 686.46, + 689.224, + 694.231, + 686.83, + 683.915, + 692.168 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2599.145, + "p90": 2623.481, + "p95": 2632.374, + "p99": 2660.077 + }, + "payload_gbps_at_latency_percentile": { + "p50": 25.81959221205435, + "p90": 25.580083865673128, + "p95": 25.49366617357564, + "p99": 25.228165951587116 + }, + "samples_us": [ + 2632.374, + 2605.434, + 2599.145, + 2660.077, + 2586.345, + 2623.481, + 2598.143, + 2594.958, + 2602.95, + 2612.765, + 2595.228, + 2610.522, + 2602.93, + 2594.658, + 2599.275, + 2596.2, + 2588.959, + 2600.357, + 2597.091, + 2589.37 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 65536, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 134217728, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5193.772, + "p90": 5288.095, + "p95": 5295.976, + "p99": 5303.138 + }, + "payload_gbps_at_latency_percentile": { + "p50": 25.842052365795034, + "p90": 25.381111345389975, + "p95": 25.34334143508203, + "p99": 25.30911471660741 + }, + "samples_us": [ + 5195.425, + 5183.927, + 5303.138, + 5189.466, + 5177.677, + 5175.804, + 5196.166, + 5174.723, + 5197.628, + 5193.772, + 5186.932, + 5174.112, + 5203.286, + 5208.915, + 5177.878, + 5288.095, + 5295.976, + 5198.039, + 5194.122, + 5173.932 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 1, + "layout": "random", + "seed": 0, + "payload_bytes": 262144, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 14.052, + "p90": 14.322, + "p95": 14.411, + "p99": 14.462 + }, + "payload_gbps_at_latency_percentile": { + "p50": 18.655280387133505, + "p90": 18.303588884234046, + "p95": 18.190548886267436, + "p99": 18.126400221269535 + }, + "samples_us": [ + 14.462, + 14.011, + 14.101, + 14.052, + 14.411, + 13.981, + 14.051, + 14.092, + 13.74, + 14.062, + 14.182, + 13.871, + 14.322, + 14.221, + 14.031, + 14.252, + 13.881, + 14.082, + 13.891, + 13.972 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 4, + "layout": "random", + "seed": 0, + "payload_bytes": 1048576, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 21.693, + "p90": 22.053, + "p95": 22.765, + "p99": 36.515 + }, + "payload_gbps_at_latency_percentile": { + "p50": 48.337067256718754, + "p90": 47.5479980048066, + "p95": 46.060882934329015, + "p99": 28.71630836642476 + }, + "samples_us": [ + 21.863, + 21.703, + 21.403, + 21.623, + 21.262, + 21.122, + 22.053, + 21.863, + 21.222, + 21.973, + 36.515, + 21.563, + 21.903, + 21.512, + 21.362, + 21.693, + 21.763, + 21.382, + 21.773, + 22.765 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 16, + "layout": "random", + "seed": 0, + "payload_bytes": 4194304, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 61.563, + "p90": 65.8, + "p95": 66.882, + "p99": 71.738 + }, + "payload_gbps_at_latency_percentile": { + "p50": 68.13027305361987, + "p90": 63.74322188449848, + "p95": 62.712000239227294, + "p99": 58.46697705539602 + }, + "samples_us": [ + 60.532, + 64.498, + 65.369, + 62.955, + 58.408, + 59.46, + 66.882, + 62.004, + 60.762, + 62.074, + 65.8, + 61.563, + 71.738, + 60.611, + 57.397, + 56.335, + 58.138, + 58.769, + 62.926, + 61.993 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 64, + "layout": "random", + "seed": 0, + "payload_bytes": 16777216, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 187.013, + "p90": 197.569, + "p95": 199.673, + "p99": 214.875 + }, + "payload_gbps_at_latency_percentile": { + "p50": 89.71149599225723, + "p90": 84.9182614681453, + "p95": 84.02345835440946, + "p99": 78.07895753344968 + }, + "samples_us": [ + 199.673, + 214.875, + 184.859, + 189.967, + 195.837, + 191.59, + 184.188, + 189.347, + 183.658, + 197.569, + 181.235, + 182.526, + 177.779, + 187.013, + 189.677, + 193.953, + 181.325, + 188.154, + 177.589, + 186.472 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 256, + "layout": "random", + "seed": 0, + "payload_bytes": 67108864, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 674.281, + "p90": 683.054, + "p95": 683.815, + "p99": 699.048 + }, + "payload_gbps_at_latency_percentile": { + "p50": 99.52655346954757, + "p90": 98.24825562839834, + "p95": 98.13891768972603, + "p99": 96.00036621233448 + }, + "samples_us": [ + 674.982, + 668.091, + 680.901, + 679.168, + 683.815, + 671.256, + 673.92, + 679.699, + 673.82, + 680.55, + 672.919, + 674.281, + 676.745, + 699.048, + 669.333, + 683.054, + 670.595, + 668.823, + 671.446, + 681.051 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 1024, + "layout": "random", + "seed": 0, + "payload_bytes": 268435456, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 2604.402, + "p90": 2622.049, + "p95": 2626.906, + "p99": 2634.688 + }, + "payload_gbps_at_latency_percentile": { + "p50": 103.069900883197, + "p90": 102.37621646277395, + "p95": 102.18692865294761, + "p99": 101.8851021449219 + }, + "samples_us": [ + 2599.816, + 2599.585, + 2622.049, + 2600.617, + 2604.102, + 2604.402, + 2611.563, + 2605.224, + 2610.642, + 2608.288, + 2601.849, + 2634.688, + 2610.411, + 2598.554, + 2599.134, + 2610.863, + 2598.123, + 2626.906, + 2609.17, + 2598.293 + ] + } + }, + { + "direction": "d2d", + "block_bytes": 262144, + "num_blocks": 2048, + "layout": "random", + "seed": 0, + "payload_bytes": 536870912, + "host_memory": null, + "api": "explicit-block-size", + "correctness_passed": true, + "latency": { + "sample_count": 20, + "percentiles_us": { + "p50": 5213.972, + "p90": 5317.87, + "p95": 5339.643, + "p99": 5398.522 + }, + "payload_gbps_at_latency_percentile": { + "p50": 102.96773975771255, + "p90": 100.95600531791864, + "p95": 100.54434575495029, + "p99": 99.44775847908001 + }, + "samples_us": [ + 5398.522, + 5227.824, + 5208.664, + 5339.643, + 5230.818, + 5207.463, + 5211.358, + 5217.759, + 5221.815, + 5205.57, + 5197.628, + 5199.23, + 5238.71, + 5204.268, + 5217.358, + 5317.87, + 5213.972, + 5217.899, + 5207.282, + 5197.528 + ] + } + } + ] +} diff --git a/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16.json b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16.json new file mode 100644 index 0000000000..6803f2019f --- /dev/null +++ b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T11:15:36.027685+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256","mode":"low-latency","nodes":2,"phase":"decode","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-uccl-ep-deepseek-v4-pro-low-latency-decode-ep16-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256],"ladder_dropped":[512],"ladder_cap":256},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":120.62200158834457,"p90":123.94800037145615,"p95":125.26899576187134,"p99":128.59399616718292},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":200.30200481414795,"p90":203.10600101947784,"p95":204.10799980163574,"p99":206.95100724697113},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":320.9240064024925,"p90":327.054001390934,"p95":329.3769955635071,"p99":335.54500341415405},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":293.32199692726135,"p90":295.4860031604767,"p95":296.4470088481903,"p99":298.33000898361206},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":299.12999272346497,"p90":302.3349940776825,"p95":304.0579855442047,"p99":308.74499678611755},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":101.67399793863297,"p90":104.67799752950668,"p95":105.59999942779541,"p99":107.32199996709824},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":179.10900712013245,"p90":182.2350025177002,"p95":183.39699506759644,"p99":185.43900549411774},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049007177352905,"p90":6.169021129608154,"p95":6.169021129608154,"p99":6.169021129608154},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.493023157119751,"p90":14.059990644454956,"p95":15.542000532150269,"p99":21.351993083953857},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.6810128688812256,"p90":1.042008399963379,"p95":1.042008399963379,"p99":1.042008399963379},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":101.27300024032593,"p90":103.47600281238556,"p95":104.19700294733047,"p99":106.23999685049057},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":174.82399940490723,"p90":177.7079999446869,"p95":178.6700040102005,"p99":181.11300468444824},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":287.47400641441345,"p90":289.557009935379,"p95":290.1580035686493,"p99":291.31999611854553},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11.777013540267944,"p90":15.42198657989502,"p95":16.984999179840088,"p99":22.59400486946106},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003888666396960616,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1376256,"scale_bytes":0,"total_logical_bytes":1376256},"dispatch":{"activation_data_bytes":1376256,"scale_bytes":0,"total_logical_bytes":1376256},"roundtrip":{"activation_data_bytes":2752512,"scale_bytes":0,"total_logical_bytes":2752512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"per-assignment"},"receive":{"max":11,"mean":6.0,"min":2,"total":96},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":53488.45113900508,"p90":52921.429253700386,"p95":52621.54181335875,"p99":51822.70212166051},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":122.42499738931656,"p90":125.67000091075897,"p95":126.83199346065521,"p99":129.07500565052032},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":204.18700575828552,"p90":207.47199654579163,"p95":208.63500237464905,"p99":210.71800589561462},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":326.6120031476021,"p90":333.1419974565506,"p95":335.46699583530426,"p99":339.79301154613495},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":296.7279851436615,"p90":298.6510097980499,"p95":299.33100938796997,"p99":300.6939888000488},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":302.93598771095276,"p90":305.98101019859314,"p95":307.02200531959534,"p99":309.02498960494995},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":103.51599752902985,"p90":106.48100078105927,"p95":107.32199996709824,"p99":107.84299671649933},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":182.79500305652618,"p90":185.88000535964966,"p95":186.92199885845184,"p99":188.64400684833527},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.969017744064331,"p90":6.009995937347412,"p95":6.009995937347412,"p99":6.009995937347412},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.372979402542114,"p90":13.499975204467773,"p95":15.66198468208313,"p99":17.906993627548218},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.8809864521026611,"p90":1.202017068862915,"p95":1.202017068862915,"p99":1.202017068862915},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":101.39299929141998,"p90":103.6359965801239,"p95":104.19700294733047,"p99":105.51899671554565},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":177.7079999446869,"p90":180.11200428009033,"p95":181.03399872779846,"p99":182.55600333213806},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":291.238009929657,"p90":293.32301020622253,"p95":293.9639985561371,"p99":295.0040102005005},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11.73698902130127,"p90":14.981985092163086,"p95":16.183018684387207,"p99":18.747001886367798},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003888666396960616,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2752512,"scale_bytes":0,"total_logical_bytes":2752512},"dispatch":{"activation_data_bytes":2752512,"scale_bytes":0,"total_logical_bytes":2752512},"roundtrip":{"activation_data_bytes":5505024,"scale_bytes":0,"total_logical_bytes":5505024},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"per-assignment"},"receive":{"max":17,"mean":12.0,"min":8,"total":192},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":105632.87723521609,"p90":104581.65354520138,"p95":104227.05684138022,"p99":103551.49608097399},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":125.14899671077728,"p90":128.79399955272675,"p95":130.19700348377228,"p99":132.59999454021454},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":208.39400589466095,"p90":211.03699505329132,"p95":211.95900440216064,"p99":214.48299288749695},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":333.54300260543823,"p90":339.83099460601807,"p95":342.1560078859329,"p99":347.0829874277115},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":303.45800518989563,"p90":305.98101019859314,"p95":306.5820038318634,"p99":308.74601006507874},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":310.3480041027069,"p90":313.713014125824,"p95":315.03400206565857,"p99":318.36000084877014},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":105.59900104999542,"p90":108.28299820423126,"p95":109.16399955749512,"p99":111.24800145626068},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":186.5209937095642,"p90":189.12500143051147,"p95":190.00700116157532,"p99":191.20800495147705},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049007177352905,"p90":6.049007177352905,"p95":6.049007177352905,"p99":6.049007177352905},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11.216014623641968,"p90":15.343993902206421,"p95":17.34602451324463,"p99":19.428998231887817},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.8410215377807617,"p90":1.2409985065460205,"p95":1.2409985065460205,"p99":1.2409985065460205},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":103.99699956178665,"p90":106.39999806880951,"p95":107.04100131988525,"p99":108.88399928808212},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":182.87600576877594,"p90":185.84099411964417,"p95":186.7620050907135,"p99":188.68499994277954},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":297.85001277923584,"p90":300.1730144023895,"p95":300.93300342559814,"p99":302.3360073566437},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":12.496978044509888,"p90":16.342997550964355,"p95":17.664998769760132,"p99":20.59200406074524},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038896200712770224,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":5505024,"scale_bytes":0,"total_logical_bytes":5505024},"dispatch":{"activation_data_bytes":5505024,"scale_bytes":0,"total_logical_bytes":5505024},"roundtrip":{"activation_data_bytes":11010048,"scale_bytes":0,"total_logical_bytes":11010048},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"per-assignment"},"receive":{"max":31,"mean":24.0,"min":17,"total":384},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":206220.1114682206,"p90":204008.11288731202,"p95":203152.67425216307,"p99":201030.27964999215},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":128.59399616718292,"p90":132.76000320911407,"p95":133.72200727462769,"p99":135.88500022888184},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":215.56499600410461,"p90":218.6889946460724,"p95":219.650998711586,"p99":221.49400413036346},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":344.15899217128754,"p90":351.44899785518646,"p95":353.3730059862137,"p99":357.3790043592453},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":312.4699890613556,"p90":315.27599692344666,"p95":316.07601046562195,"p99":317.4780011177063},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":322.284996509552,"p90":326.1309862136841,"p95":327.3729979991913,"p99":330.21700382232666},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":107.64200240373611,"p90":110.56700348854065,"p95":111.40699684619904,"p99":113.16999793052673},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":192.3699975013733,"p90":195.49499452114105,"p95":196.33600115776062,"p99":197.97800481319427},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009012460708618,"p90":6.089001893997192,"p95":6.089001893997192,"p99":6.089001893997192},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":12.699007987976074,"p90":17.787009477615356,"p95":19.62900161743164,"p99":22.392988204956055},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.1610090732574463,"p90":2.2439956665039062,"p95":2.2439956665039062,"p99":2.2439956665039062},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":107.60200023651123,"p90":110.28700321912766,"p95":111.00699752569199,"p99":112.85000294446945},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":190.12700021266937,"p90":193.21100413799286,"p95":194.01299953460693,"p99":195.8550065755844},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":307.42400884628296,"p90":310.3089928627014,"p95":311.0699951648712,"p99":312.6710057258606},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.941006898880005,"p90":19.187986850738525,"p95":20.631998777389526,"p99":23.914992809295654},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003889858489856124,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":11010048,"scale_bytes":0,"total_logical_bytes":11010048},"dispatch":{"activation_data_bytes":11010048,"scale_bytes":0,"total_logical_bytes":11010048},"roundtrip":{"activation_data_bytes":22020096,"scale_bytes":0,"total_logical_bytes":22020096},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"per-assignment"},"receive":{"max":58,"mean":48.0,"min":40,"total":768},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":397164.00510815054,"p90":392480.3389155215,"p95":390991.31810594903,"p99":387623.8913150288},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":135.88599860668182,"p90":140.13199508190155,"p95":141.5330022573471,"p99":145.01899480819702},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":222.97599911689758,"p90":225.82000494003296,"p95":226.90099477767944,"p99":229.0250062942505},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":358.8619977235794,"p90":365.9520000219345,"p95":368.43399703502655,"p99":374.0440011024475},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":323.1259882450104,"p90":326.8919885158539,"p95":328.2949924468994,"p99":330.8579921722412},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":335.9859883785248,"p90":340.59301018714905,"p95":342.195987701416,"p99":345.3199863433838},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":114.49199914932251,"p90":117.73800104856491,"p95":118.45800280570984,"p99":120.22200226783752},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":197.8989988565445,"p90":200.50199329853058,"p95":201.26299560070038,"p99":202.3860067129135},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049007177352905,"p90":6.089001893997192,"p95":6.089001893997192,"p99":6.089001893997192},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":14.70300555229187,"p90":19.830018281936646,"p95":21.632999181747437,"p99":23.71498942375183},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.9620060920715332,"p90":2.003014087677002,"p95":2.003014087677002,"p99":2.003014087677002},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":113.13000321388245,"p90":116.09499901533127,"p95":116.93599820137024,"p99":118.93899738788605},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":197.85800576210022,"p90":199.82099533081055,"p95":200.422003865242,"p99":201.58499479293823},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":320.2430009841919,"p90":323.08799028396606,"p95":324.0489959716797,"p99":326.5329897403717},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":15.823006629943848,"p90":20.350009202957153,"p95":22.192001342773438,"p99":25.2780020236969},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038906957488507032,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":22020096,"scale_bytes":0,"total_logical_bytes":22020096},"dispatch":{"activation_data_bytes":22020096,"scale_bytes":0,"total_logical_bytes":22020096},"roundtrip":{"activation_data_bytes":44040192,"scale_bytes":0,"total_logical_bytes":44040192},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"per-assignment"},"receive":{"max":105,"mean":96.0,"min":80,"total":1536},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":761936.535614063,"p90":751630.222415114,"p95":748109.2975975319,"p99":741341.3938498057},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":158.23900699615479,"p90":162.32499480247498,"p95":163.7669950723648,"p99":167.5329953432083},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":237.03700304031372,"p90":239.68100547790527,"p95":240.52199721336365,"p99":242.60500073432922},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":395.2760100364685,"p90":402.00600028038025,"p95":404.28899228572845,"p99":410.13799607753754},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":362.86699771881104,"p90":367.63399839401245,"p95":369.4770038127899,"p99":373.20199608802795},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":376.48698687553406,"p90":380.9739947319031,"p95":382.1749985218048,"p99":384.4990134239197},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":135.44400036334991,"p90":139.32999968528748,"p95":140.61200618743896,"p99":142.2950029373169},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":213.56099843978882,"p90":216.20599925518036,"p95":216.8869972229004,"p99":218.40900182724},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.047993898391724,"p90":6.049007177352905,"p95":6.049007177352905,"p99":6.049007177352905},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.386019229888916,"p90":24.3149995803833,"p95":25.75799822807312,"p99":30.80698847770691},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.2419989109039307,"p90":1.8430054187774658,"p95":1.8430054187774658,"p99":1.8430054187774658},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":131.35799765586853,"p90":135.04299521446228,"p95":136.28600537776947,"p99":139.2499953508377},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":210.63800156116486,"p90":213.0420058965683,"p95":213.64299952983856,"p99":214.8440033197403},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":354.73498702049255,"p90":357.901006937027,"p95":358.74199867248535,"p99":360.4240119457245},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":21.753013134002686,"p90":26.99899673461914,"p95":28.162002563476562,"p99":31.08498454093933},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003891168860718608,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":44040192,"scale_bytes":0,"total_logical_bytes":44040192},"dispatch":{"activation_data_bytes":44040192,"scale_bytes":0,"total_logical_bytes":44040192},"roundtrip":{"activation_data_bytes":88080384,"scale_bytes":0,"total_logical_bytes":88080384},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"per-assignment"},"receive":{"max":211,"mean":192.0,"min":175,"total":3072},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":1359940.7624924532,"p90":1343923.7509119273,"p95":1339700.4042136162,"p99":1331602.8965606405},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":191.97000563144684,"p90":198.78000020980835,"p95":201.06300711631775,"p99":206.67199790477753},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":264.1980051994324,"p90":268.56499910354614,"p95":269.8870003223419,"p99":272.5310027599335},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":456.1680108308792,"p90":467.3449993133545,"p95":470.95000743865967,"p99":479.203000664711},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":419.5930063724518,"p90":424.3200123310089,"p95":426.04199051856995,"p99":437.2580051422119},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":435.4960024356842,"p90":440.744012594223,"p95":442.74699687957764,"p99":446.1919963359833},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":167.49300062656403,"p90":170.5780029296875,"p95":171.3380068540573,"p99":172.78100550174713},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":236.71700060367584,"p90":240.76299369335175,"p95":241.80400371551514,"p99":245.4099953174591},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":6.168991327285767,"p95":6.168991327285767,"p99":6.168991327285767},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":20.871996879577637,"p90":28.122007846832275,"p95":30.64700961112976,"p99":48.593997955322266},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.8430054187774658,"p90":2.204000949859619,"p95":2.204000949859619,"p99":2.204000949859619},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":167.45199263095856,"p90":170.77800631523132,"p95":171.77900671958923,"p99":173.06099832057953},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":236.71700060367584,"p90":239.96199667453766,"p95":241.20399355888367,"p99":243.367001414299},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":410.81899404525757,"p90":414.90599513053894,"p95":415.9480035305023,"p99":418.43000054359436},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":24.75598454475403,"p90":30.32401204109192,"p95":32.24700689315796,"p99":36.17298603057861},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003891168860718608,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":88080384,"scale_bytes":0,"total_logical_bytes":88080384},"dispatch":{"activation_data_bytes":88080384,"scale_bytes":0,"total_logical_bytes":88080384},"roundtrip":{"activation_data_bytes":176160768,"scale_bytes":0,"total_logical_bytes":176160768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"per-assignment"},"receive":{"max":421,"mean":384.0,"min":335,"total":6144},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":2351341.9050298366,"p90":2323344.097115982,"p95":2312833.3048377894,"p99":2294976.172609171},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":265.1999890804291,"p90":275.77701210975647,"p95":279.0620028972626,"p99":286.27198934555054},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":320.7229971885681,"p90":329.97700572013855,"p95":332.4199914932251,"p99":337.9490077495575},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":585.9229862689972,"p90":605.754017829895,"p95":611.4819943904877,"p99":624.220997095108},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":532.3209762573242,"p90":538.2900238037109,"p95":540.0930047035217,"p99":543.658971786499},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":560.3240132331848,"p90":571.0999965667725,"p95":574.0650296211243,"p99":580.5140137672424},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":228.06400060653687,"p90":233.5519939661026,"p95":234.5539927482605,"p99":235.95599830150604},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":282.5469970703125,"p90":288.1140112876892,"p95":289.4760072231293,"p99":293.08098554611206},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.049990653991699,"p90":6.25002384185791,"p95":6.25002384185791,"p99":6.25002384185791},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":24.477005004882812,"p90":34.37399864196777,"p95":37.17601299285889,"p99":41.060030460357666},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.5640130043029785,"p90":5.047023296356201,"p95":5.047023296356201,"p99":5.047023296356201},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":232.22999274730682,"p90":238.19799721240997,"p95":239.84099924564362,"p99":243.52599680423737},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":291.55999422073364,"p90":297.61001467704773,"p95":299.4920015335083,"p99":302.45599150657654},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":535.5290174484253,"p90":545.5030202865601,"p95":548.1060147285461,"p99":553.354024887085},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":24.595022201538086,"p90":31.88800811767578,"p95":34.012019634246826,"p99":39.337992668151855},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003891168860718608,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":176160768,"scale_bytes":0,"total_logical_bytes":176160768},"dispatch":{"activation_data_bytes":176160768,"scale_bytes":0,"total_logical_bytes":176160768},"roundtrip":{"activation_data_bytes":352321536,"scale_bytes":0,"total_logical_bytes":352321536},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"per-assignment"},"receive":{"max":802,"mean":768.0,"min":696,"total":12288},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":3655028.075956657,"p90":3586062.007199732,"p95":3567540.077039102,"p99":3527907.9426688007},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":409.9780023097992,"p90":430.3700029850006,"p95":437.7810060977936,"p99":454.9669921398163},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":467.9050147533417,"p90":485.0510060787201,"p95":489.93799090385437,"p99":500.1140236854553},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":877.8830170631409,"p90":915.4210090637207,"p95":927.718997001648,"p99":955.0810158252716},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1197.4430084228516,"p90":1224.848985671997,"p95":1231.6190004348755,"p99":1253.648042678833},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":855.5690050125122,"p90":872.5950121879578,"p95":878.2830238342285,"p99":890.1439905166626},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":355.89599609375,"p90":377.3680031299591,"p95":381.6159963607788,"p99":387.78600096702576},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":420.9139943122864,"p90":445.9129869937897,"p95":450.639009475708,"p99":454.7260105609894},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.410002708435059,"p90":6.690025329589844,"p95":6.690025329589844,"p99":6.690025329589844},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":32.330989837646484,"p90":48.38907718658447,"p95":54.80694770812988,"p99":262.9129886627197},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.532000541687012,"p90":366.39201641082764,"p95":366.39201641082764,"p99":366.39201641082764},"sample_count":4},"host_retried_siblings":null,"host_stalled_trials":null},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":377.24900245666504,"p90":388.5050117969513,"p95":391.62999391555786,"p99":396.9990015029907},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":439.5830035209656,"p90":449.1580128669739,"p95":451.6400098800659,"p99":456.80901408195496},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":830.4139971733093,"p90":846.1139798164368,"p95":850.8049845695496,"p99":859.3350052833557},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":24.394989013671875,"p90":34.37197208404541,"p95":38.215041160583496,"p99":46.590983867645264},"sample_count":2048},"rank_alignment":{"host_held_fraction":null,"host_held_samples":null,"host_stalled_fraction":null,"host_stalled_samples":null,"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.003891168860718608,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":352321536,"scale_bytes":0,"total_logical_bytes":352321536},"dispatch":{"activation_data_bytes":352321536,"scale_bytes":0,"total_logical_bytes":352321536},"roundtrip":{"activation_data_bytes":704643072,"scale_bytes":0,"total_logical_bytes":704643072},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"per-assignment"},"receive":{"max":1613,"mean":1536.0,"min":1426,"total":24576},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":4787457.208013396,"p90":4694044.708930467,"p95":4663644.735063328,"p99":4601502.727241439},"tokens_per_rank":256}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"receive_capacity":256,"num_qps_per_rank":24,"return_recv_hook":true,"zero_copy":false,"fp8_scales":null,"deep_ep":"uccl deep_ep_wrapper","num_nvl_bytes":1073741824,"num_rdma_bytes":5650787456,"is_intranode":false,"cpu_proxy_threads":4},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-ll-hook-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9943,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2055,"p50":2091,"max":2093,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2061,"p50":2096,"max":2097,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2059,"p50":2094,"max":2095,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2059,"p50":2094,"max":2096,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2056,"p50":2089,"max":2090,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2066,"p50":2093,"max":2095,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2063,"p50":2091,"max":2093,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2062,"p50":2093,"max":2094,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2068,"p50":2095,"max":2098,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2059,"p50":2089,"max":2091,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2036,"p50":2091,"max":2092,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.7},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2057,"p50":2088,"max":2089,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2066,"p50":2095,"max":2099,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2066,"p50":2093,"max":2096,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2058,"p50":2084,"max":2087,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2062,"p50":2092,"max":2093,"samples":2340},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":54,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":235.8}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-bf16.json b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-bf16.json new file mode 100644 index 0000000000..9df44f64b9 --- /dev/null +++ b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T11:23:41.049347+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":2,"phase":"decode","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":172.46000468730927,"p90":176.74599587917328,"p95":178.7090003490448,"p99":183.35600197315216},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":380.9340000152588,"p90":389.5869851112366,"p95":393.0310010910034,"p99":414.18299078941345},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":553.394004702568,"p90":566.3329809904099,"p95":571.7400014400482,"p99":597.5389927625656},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":638.3230090141296,"p90":699.7330188751221,"p95":704.6639919281006,"p99":712.5530242919922},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":626.3830065727234,"p90":638.0410194396973,"p95":642.687976360321,"p99":781.6179990768433},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":170.45700550079346,"p90":193.53200495243073,"p95":194.8940008878708,"p99":197.176992893219},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":377.5689899921417,"p90":399.603009223938,"p95":402.8880000114441,"p99":410.09899973869324},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.289958953857422,"p90":6.489992141723633,"p95":6.489992141723633,"p99":6.489992141723633},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.052958011627197,"p90":13.339996337890625,"p95":15.345990657806396,"p99":22.552013397216797},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.8810162544250488,"p90":5.167961120605469,"p95":5.167961120605469,"p99":5.167961120605469},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":148.30300211906433,"p90":151.98899805545807,"p95":153.79199385643005,"p99":157.51700103282928},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":294.964998960495,"p90":302.0159900188446,"p95":304.5380115509033,"p99":311.830997467041},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":528.0749797821045,"p90":538.2099747657776,"p95":542.7770018577576,"p99":582.8369855880737},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":97.54699468612671,"p90":104.83700037002563,"p95":107.72198438644409,"p99":113.37101459503174},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"roundtrip":{"activation_data_bytes":2351104,"scale_bytes":0,"total_logical_bytes":2351104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":9,"mean":5.125,"min":2,"total":82},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":25543.47712519304,"p90":25076.757626101495,"p95":24895.440071263525,"p99":20470.357666913183},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":178.4690022468567,"p90":182.95499682426453,"p95":184.87900495529175,"p99":189.5260065793991},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":381.01398944854736,"p90":390.0279998779297,"p95":394.3130075931549,"p99":409.7369909286499},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":559.482991695404,"p90":572.9829967021942,"p95":579.1920125484467,"p99":599.262997508049},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":642.1300172805786,"p90":701.7790079116821,"p95":706.5070271492004,"p99":717.5999879837036},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":627.7049779891968,"p90":639.7240161895752,"p95":644.4510221481323,"p99":800.4480004310608},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":181.9940060377121,"p90":196.1359977722168,"p95":198.78000020980835,"p99":205.7500034570694},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":380.77399134635925,"p90":405.129998922348,"p95":414.0230119228363,"p99":435.2949857711792},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.209015846252441,"p90":6.530046463012695,"p95":6.530046463012695,"p99":6.530046463012695},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.20900297164917,"p90":12.982010841369629,"p95":15.626013278961182,"p99":18.78798007965088},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.3620257377624512,"p90":3.1249523162841797,"p95":3.1249523162841797,"p99":3.1249523162841797},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":157.51700103282928,"p90":160.08199751377106,"p95":161.60400211811066,"p99":165.009006857872},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":300.6939888000488,"p90":308.94601345062256,"p95":311.99100613594055,"p99":321.725994348526},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":536.5700125694275,"p90":546.4249849319458,"p95":551.8730282783508,"p99":756.1410069465637},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":90.09295701980591,"p90":98.30695390701294,"p95":100.62998533248901,"p99":105.7579517364502},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"roundtrip":{"activation_data_bytes":4759552,"scale_bytes":0,"total_logical_bytes":4759552},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":15,"mean":10.375,"min":7,"total":166},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":50979.363111806866,"p90":50021.57053693785,"p95":49654.66559946667,"p99":39977.612515450375},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":178.58900129795074,"p90":183.35600197315216,"p95":185.31900644302368,"p99":191.68899953365326},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":381.4939856529236,"p90":390.86800813674927,"p95":395.19500732421875,"p99":445.2719986438751},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":560.0829869508743,"p90":574.2240101099014,"p95":580.5140137672424,"p99":636.9609981775284},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":641.8460011482239,"p90":701.8970251083374,"p95":709.388017654419,"p99":733.2260012626648},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":628.026008605957,"p90":639.6030187606812,"p95":645.2109813690186,"p99":784.4619750976562},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":186.8010014295578,"p90":196.73700630664825,"p95":198.05899262428284,"p99":202.54600048065186},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":381.3750147819519,"p90":397.038996219635,"p95":399.04099702835083,"p99":408.8970124721527},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.290018558502197,"p90":6.410956382751465,"p95":6.410956382751465,"p99":6.410956382751465},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.3310136795043945,"p90":13.181984424591064,"p95":14.744043350219727,"p99":19.190967082977295},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-5.567014217376709,"p90":0.8010268211364746,"p95":0.8010268211364746,"p99":0.8010268211364746},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":159.4810038805008,"p90":162.16500103473663,"p95":163.68700563907623,"p99":168.614000082016},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":302.89700627326965,"p90":311.9089901447296,"p95":315.5969977378845,"p99":339.711993932724},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":539.0939712524414,"p90":548.9490032196045,"p95":554.6370148658752,"p99":590.3710126876831},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":87.88996934890747,"p90":95.5820083618164,"p95":98.46603870391846,"p99":103.43396663665771},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"roundtrip":{"activation_data_bytes":9433088,"scale_bytes":0,"total_logical_bytes":9433088},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":26,"mean":20.5625,"min":15,"total":329},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":101906.60756560415,"p90":100062.06681764699,"p95":99192.36009313392,"p99":81584.57902568541},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":180.59200048446655,"p90":185.19900739192963,"p95":187.00100481510162,"p99":192.20900535583496},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":381.3740015029907,"p90":390.747994184494,"p95":394.47399973869324,"p99":418.19000244140625},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":561.9660019874573,"p90":575.9470015764236,"p95":581.4750045537949,"p99":610.3990077972412},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":638.9219760894775,"p90":647.8949785232544,"p95":651.3810157775879,"p99":677.299976348877},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":628.305971622467,"p90":640.2040123939514,"p95":644.9310183525085,"p99":799.4059920310974},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":187.20200657844543,"p90":190.00600278377533,"p95":192.0890063047409,"p99":201.62299275398254},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":379.8530101776123,"p90":384.57998633384705,"p95":386.5019977092743,"p99":390.14700055122375},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.209969520568848,"p90":6.36899471282959,"p95":6.36899471282959,"p99":6.36899471282959},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.533013820648193,"p90":12.978971004486084,"p95":15.179991722106934,"p99":22.394001483917236},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.2819766998291016,"p90":2.523958683013916,"p95":2.523958683013916,"p99":2.523958683013916},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":160.44199466705322,"p90":162.96599805355072,"p95":164.12700712680817,"p99":168.37400197982788},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":318.4800148010254,"p90":325.2510130405426,"p95":327.97399163246155,"p99":339.231014251709},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":539.6940112113953,"p90":549.3080019950867,"p95":554.6770095825195,"p99":585.8439803123474},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":87.65000104904175,"p90":95.50100564956665,"p95":98.1060266494751,"p99":104.27403450012207},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"roundtrip":{"activation_data_bytes":18608128,"scale_bytes":0,"total_logical_bytes":18608128},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":49,"mean":40.5625,"min":36,"total":649},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":203722.39924676684,"p90":199936.26644319564,"p95":198470.83852003122,"p99":160118.8898706938},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":188.60399723052979,"p90":192.65000522136688,"p95":194.73299384117126,"p99":199.37999546527863},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":380.4929852485657,"p90":389.06601071357727,"p95":392.7119970321655,"p99":403.20900082588196},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":569.0969824790955,"p90":581.7160159349442,"p95":587.4449908733368,"p99":602.5889962911606},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":645.8939909934998,"p90":655.826985836029,"p95":660.9150171279907,"p99":680.7050108909607},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":627.705991268158,"p90":641.0449743270874,"p95":645.8929777145386,"p99":683.351993560791},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":193.3310031890869,"p90":196.33600115776062,"p95":198.49899411201477,"p99":205.38899302482605},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":380.854994058609,"p90":386.1829936504364,"p95":387.9449963569641,"p99":392.91301369667053},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.289958953857422,"p90":6.330013275146484,"p95":6.330013275146484,"p99":6.330013275146484},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.892012596130371,"p90":15.183031558990479,"p95":16.743957996368408,"p99":23.07802438735962},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.1619925498962402,"p90":2.2829771041870117,"p95":2.2829771041870117,"p99":2.2829771041870117},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":166.45100712776184,"p90":168.97399723529816,"p95":170.45800387859344,"p99":174.26300048828125},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":326.45100355148315,"p90":333.0230116844177,"p95":335.7050120830536,"p99":342.15599298477173},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":545.183002948761,"p90":555.3190112113953,"p95":560.9260201454163,"p99":593.0150151252747},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":81.92098140716553,"p90":89.974045753479,"p95":92.41700172424316,"p99":99.70802068710327},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"roundtrip":{"activation_data_bytes":37675008,"scale_bytes":0,"total_logical_bytes":37675008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":89,"mean":82.125,"min":72,"total":1314},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":407834.2465439939,"p90":399347.9556855216,"p95":396350.4927795372,"p99":374623.91624270024},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":203.1860053539276,"p90":207.7530026435852,"p95":209.39600467681885,"p99":214.40300345420837},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":380.1319897174835,"p90":389.5060122013092,"p95":394.23298835754395,"p99":416.6269898414612},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":583.3179950714111,"p90":597.2590148448944,"p95":603.6289930343628,"p99":631.0299932956696},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":661.8790030479431,"p90":670.85200548172,"p95":675.1769781112671,"p99":691.2410259246826},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":626.783013343811,"p90":639.3629908561707,"p95":646.1330056190491,"p99":688.8800263404846},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":209.11499857902527,"p90":212.64100074768066,"p95":213.80099654197693,"p99":218.48900616168976},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":381.1750113964081,"p90":386.30199432373047,"p95":388.10500502586365,"p99":393.4740126132965},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.209969520568848,"p90":6.249964237213135,"p95":6.249964237213135,"p99":6.249964237213135},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.7729949951171875,"p90":12.982010841369629,"p95":17.109036445617676,"p99":23.993968963623047},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.3219714164733887,"p90":2.483963966369629,"p95":2.483963966369629,"p99":2.483963966369629},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":181.15399777889252,"p90":183.7570071220398,"p95":184.83899533748627,"p99":188.40500712394714},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":343.9989984035492,"p90":351.20999813079834,"p95":354.21401262283325,"p99":362.6669943332672},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":558.5629940032959,"p90":569.0180063247681,"p95":574.9080181121826,"p99":606.4350008964539},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":67.05904006958008,"p90":75.111985206604,"p95":77.83502340316772,"p99":84.44398641586304},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"roundtrip":{"activation_data_bytes":75235328,"scale_bytes":0,"total_logical_bytes":75235328},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":178,"mean":164.0,"min":148,"total":2624},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":816869.6169166142,"p90":800797.05475974,"p95":792406.510033428,"p99":743235.368166909},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":224.41799938678741,"p90":229.14500534534454,"p95":230.98799586296082,"p99":237.15800046920776},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":400.52300691604614,"p90":410.2979898452759,"p95":415.18500447273254,"p99":436.37698888778687},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":624.9410063028336,"p90":639.4429951906204,"p95":646.1730003356934,"p99":673.5349893569946},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":685.8350038528442,"p90":696.5680122375488,"p95":707.0670127868652,"p99":723.3290076255798},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":625.3820061683655,"p90":637.3990178108215,"p95":644.6099877357483,"p99":862.4989986419678},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":232.87099599838257,"p90":235.91600358486176,"p95":237.5980019569397,"p99":253.7420094013214},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":382.4169933795929,"p90":389.5860016345978,"p95":391.7509913444519,"p99":395.5569863319397},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.2490105628967285,"p90":6.329953670501709,"p95":6.329953670501709,"p99":6.329953670501709},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.650971412658691,"p90":13.62001895904541,"p95":15.381991863250732,"p99":22.073984146118164},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.4830231666564941,"p90":6.489992141723633,"p95":6.489992141723633,"p99":6.489992141723633},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":205.87100088596344,"p90":208.59499275684357,"p95":209.95700359344482,"p99":213.48300576210022},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":381.33400678634644,"p90":388.386994600296,"p95":391.2310004234314,"p99":398.0399966239929},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":605.1530241966248,"p90":615.0490045547485,"p95":620.4559803009033,"p99":639.6049857139587},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":19.308030605316162,"p90":26.197969913482666,"p95":29.844999313354492,"p99":53.000032901763916},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"roundtrip":{"activation_data_bytes":150700032,"scale_bytes":0,"total_logical_bytes":150700032},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":363,"mean":328.5,"min":292,"total":5256},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":1637399.205445509,"p90":1606528.9895126899,"p95":1588557.4525410226,"p99":1187247.755200088},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":290.51700234413147,"p90":295.6849932670593,"p95":297.64801263809204,"p99":302.8559982776642},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":480.8039963245392,"p90":489.29500579833984,"p95":493.78401041030884,"p99":588.7680053710938},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":771.3209986686707,"p90":784.9799990653992,"p95":791.4320230484009,"p99":891.6240036487579},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":821.960985660553,"p90":830.5339813232422,"p95":833.9359760284424,"p99":1004.7160387039185},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":758.5819959640503,"p90":768.9579725265503,"p95":773.6849784851074,"p99":1000.2659559249878},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":326.73200964927673,"p90":337.7479910850525,"p95":341.715008020401,"p99":348.2840061187744},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":435.85801124572754,"p90":441.90698862075806,"p95":443.3870017528534,"p99":454.2039930820465},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.290018558502197,"p90":6.489992141723633,"p95":6.489992141723633,"p99":6.489992141723633},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.618984699249268,"p90":15.50602912902832,"p95":17.229020595550537,"p99":21.113991737365723},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.1229982376098633,"p90":3.004014492034912,"p95":3.004014492034912,"p99":3.004014492034912},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":272.33099937438965,"p90":275.0549912452698,"p95":276.37699246406555,"p99":280.94398975372314},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":439.7439956665039,"p90":446.47398591041565,"p95":449.55798983573914,"p99":463.41899037361145},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":739.1149997711182,"p90":749.4519948959351,"p95":753.5369992256165,"p99":920.7490086555481},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.504977226257324,"p90":23.553013801574707,"p95":25.154948234558105,"p99":30.16197681427002},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0041841003112494946,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"roundtrip":{"activation_data_bytes":303149056,"scale_bytes":0,"total_logical_bytes":303149056},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":692,"mean":660.8125,"min":603,"total":10573},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":2699774.066476864,"p90":2663344.5170883997,"p95":2647072.2024486377,"p99":2047455.4670873794},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":374.96501207351685,"p90":380.8940052986145,"p95":383.05699825286865,"p99":389.2660140991211},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":626.3030171394348,"p90":635.4359984397888,"p95":638.8819813728333,"p99":720.0070023536682},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1001.2680292129517,"p90":1016.3300037384033,"p95":1021.9389796257019,"p99":1109.2730164527893},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1032.4339866638184,"p90":1043.0940389633179,"p95":1046.0950136184692,"p99":1056.033968925476},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":977.512001991272,"p90":989.1700148582458,"p95":994.0969944000244,"p99":1205.3749561309814},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":429.6090006828308,"p90":437.7410113811493,"p95":439.8620128631592,"p99":444.1080093383789},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":556.2790036201477,"p90":563.2879734039307,"p95":565.6909942626953,"p99":581.957995891571},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.410002708435059,"p90":6.410956382751465,"p95":6.410956382751465,"p99":6.410956382751465},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":13.33606243133545,"p90":18.503904342651367,"p95":20.470023155212402,"p99":26.479005813598633},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.644062042236328,"p90":4.205942153930664,"p95":4.205942153930664,"p99":4.205942153930664},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":349.2879867553711,"p90":353.69399189949036,"p95":355.1360070705414,"p99":359.9030077457428},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":563.9309883117676,"p90":571.7430114746094,"p95":576.187014579773,"p99":590.8899903297424},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":951.636016368866,"p90":962.332010269165,"p95":966.4549827575684,"p99":1155.7040214538574},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":25.836050510406494,"p90":31.00299835205078,"p95":32.96703100204468,"p99":42.94198751449585},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.001953125,"host_stalled_samples":{"dispatch":4,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0041841003112494946,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"roundtrip":{"activation_data_bytes":607301632,"scale_bytes":0,"total_logical_bytes":607301632},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":1373,"mean":1323.8125,"min":1247,"total":21181},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":4190229.8812251026,"p90":4140845.292997465,"p95":4120322.285525159,"p99":3398112.744226379},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":493.06198954582214,"p90":499.35099482536316,"p95":501.91497802734375,"p99":549.0670204162598},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":860.9769940376282,"p90":869.949996471405,"p95":873.4750151634216,"p99":1032.5590372085571},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1354.0389835834503,"p90":1369.3009912967682,"p95":1375.3899931907654,"p99":1581.626057624817},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1416.4530038833618,"p90":1427.1140098571777,"p95":1431.3950538635254,"p99":1443.613052368164},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1336.1719846725464,"p90":1347.5090265274048,"p95":1352.0359992980957,"p99":1579.9380540847778},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":543.2180166244507,"p90":550.3910183906555,"p95":553.3949732780457,"p99":558.1629872322083},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":829.0910124778748,"p90":835.8209729194641,"p95":837.7050161361694,"p99":842.8289890289307},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.289958953857422,"p90":6.330013275146484,"p95":6.330013275146484,"p99":6.330013275146484},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.786026000976562,"p90":25.117993354797363,"p95":26.965975761413574,"p99":32.89508819580078},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-4.045963287353516,"p90":4.285931587219238,"p95":4.285931587219238,"p99":4.285931587219238},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":464.260995388031,"p90":468.90801191329956,"p95":470.75098752975464,"p99":477.9219925403595},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":819.2769885063171,"p90":827.6900053024292,"p95":830.6540250778198,"p99":841.1089777946472},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1307.0529699325562,"p90":1317.1080350875854,"p95":1321.6350078582764,"p99":1510.640025138855},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":28.799057006835938,"p90":34.568071365356445,"p95":36.77105903625488,"p99":73.54998588562012},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0095911156386137,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"roundtrip":{"activation_data_bytes":1215807488,"scale_bytes":0,"total_logical_bytes":1215807488},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"roundtrip":{"activation_data_bytes":1215807488,"scale_bytes":0,"total_logical_bytes":1215807488},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":42404,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":2702,"mean":2650.25,"min":2578,"total":42404},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.017352504655718803,"expert_assignments_per_rank":[3108,3051,3103,2954,3171,3067,3121,2998,3045,3065,3095,3023,3150,3071,3090,3040],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,3,103,1301,3825,2960,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.17626953125,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.06268276274204254,"same_node_fraction":0.49721723794937134,"same_scaleup_domain_fraction":0.49721723794937134,"cross_node_fraction":0.5027827620506287,"cross_domain_fraction":0.5027827620506287,"gpus_per_node":8,"scale_up_domain":8,"copies":42404},"payload_copies_per_rank":[2682,2637,2676,2578,2702,2659,2687,2602,2635,2661,2672,2599,2691,2625,2662,2636],"payload_rank_cv":0.013201921246945858,"routed_copies":42404},"token_rate_at_latency_percentile":{"p50":6130947.283711836,"p90":6079365.583999964,"p95":6059010.266185846,"p99":5185013.411646344},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":36,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":20,"num_max_rdma_chunked_recv_tokens":140},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":12,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":1073741824,"num_qps_per_rank":10,"is_intranode":false,"cpu_proxy_threads":4},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9948,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2084,"p50":2097,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2089,"p50":2101,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2089,"p50":2100,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2088,"p50":2100,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2082,"p50":2094,"max":2095,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":54,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2097,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2096,"max":2097,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2098,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2101,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2085,"p50":2095,"max":2097,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2082,"p50":2095,"max":2096,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2077,"p50":2091,"max":2095,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2086,"p50":2100,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2085,"p50":2098,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2076,"p50":2090,"max":2094,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2084,"p50":2098,"max":2100,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":410.2}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-fp8.json b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-fp8.json new file mode 100644 index 0000000000..4cfbad3485 --- /dev/null +++ b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T11:48:39.153519+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":2,"phase":"decode","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep16-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":173.78099262714386,"p90":177.82799899578094,"p95":179.59000170230865,"p99":184.19699370861053},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":372.0400035381317,"p90":388.7470066547394,"p95":417.67001152038574,"p99":431.32901191711426},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":545.8209961652756,"p90":566.5750056505203,"p95":597.2600132226944,"p99":615.5260056257248},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":719.9670076370239,"p90":736.6690039634705,"p95":744.7620034217834,"p99":844.4340229034424},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":589.7279977798462,"p90":621.0169792175293,"p95":679.3419718742371,"p99":699.413001537323},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.22800002992153,"p90":20.591000095009804,"p95":21.472999826073647,"p99":23.195000365376472},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":172.58000373840332,"p90":186.92199885845184,"p95":190.3270035982132,"p99":196.73599302768707},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":434.6959888935089,"p90":448.3970105648041,"p95":451.8420100212097,"p99":458.2520127296448},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.25002384185791,"p90":6.36899471282959,"p95":6.36899471282959,"p99":6.36899471282959},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.8549981117248535,"p90":15.586018562316895,"p95":18.747985363006592,"p99":27.15998888015747},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.7649998664855957,"p90":22.31299877166748,"p95":22.31299877166748,"p99":22.31299877166748},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":149.82600510120392,"p90":153.63100171089172,"p95":155.55399656295776,"p99":159.28100049495697},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":312.4319911003113,"p90":323.12700152397156,"p95":326.41100883483887,"p99":330.69801330566406},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":533.0449938774109,"p90":561.3269805908203,"p95":575.7870078086853,"p99":589.7279977798462},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":57.443976402282715,"p90":68.70400905609131,"p95":100.03000497817993,"p99":113.92998695373535},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":587776,"scale_bytes":18368,"total_logical_bytes":606144},"roundtrip":{"activation_data_bytes":1763328,"scale_bytes":18368,"total_logical_bytes":1781696},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1175552,"scale_bytes":0,"total_logical_bytes":1175552},"dispatch":{"activation_data_bytes":587776,"scale_bytes":18368,"total_logical_bytes":606144},"roundtrip":{"activation_data_bytes":1763328,"scale_bytes":18368,"total_logical_bytes":1781696},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":82,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":9,"mean":5.125,"min":2,"total":82},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.48591265082359314,"expert_assignments_per_rank":[10,8,6,11,4,9,4,3,9,4,5,2,3,10,5,3],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,0,3,8,5,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.060975611209869385,"same_node_fraction":0.5121951103210449,"same_scaleup_domain_fraction":0.5121951103210449,"cross_node_fraction":0.4878048896789551,"cross_domain_fraction":0.4878048896789551,"gpus_per_node":8,"scale_up_domain":8,"copies":82},"payload_copies_per_rank":[9,7,5,9,4,8,4,3,6,3,4,2,2,8,5,3],"payload_rank_cv":0.4569510817527771,"routed_copies":82},"token_rate_at_latency_percentile":{"p50":27131.152090854313,"p90":25764.19089242894,"p95":23552.202958780228,"p99":22876.32624047837},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":180.2310049533844,"p90":184.39799547195435,"p95":185.87900698184967,"p99":190.0860071182251},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":372.1599876880646,"p90":389.54800367355347,"p95":422.435998916626,"p99":436.4570081233978},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":552.390992641449,"p90":573.9459991455078,"p95":608.3150058984756,"p99":626.5430152416229},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":717.9610133171082,"p90":737.5929951667786,"p95":756.3419938087463,"p99":812.703013420105},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":590.4489755630493,"p90":620.3359961509705,"p95":681.7060112953186,"p99":705.7819962501526},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.588999286293983,"p90":21.271999925374985,"p95":22.474000230431557,"p99":23.716000840067863},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":191.24899804592133,"p90":195.4140067100525,"p95":196.57599925994873,"p99":202.1850049495697},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":432.01199173927307,"p90":444.67100501060486,"p95":447.47498631477356,"p99":457.89098739624023},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.248950958251953,"p90":6.410002708435059,"p95":6.410002708435059,"p99":6.410002708435059},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.8119635581970215,"p90":14.660000801086426,"p95":16.98601245880127,"p99":19.668996334075928},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.0449628829956055,"p90":2.1229982376098633,"p95":2.1229982376098633,"p99":2.1229982376098633},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":159.19999778270721,"p90":161.96399927139282,"p95":163.16600143909454,"p99":166.2909984588623},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":312.4319911003113,"p90":323.0479955673218,"p95":327.37401127815247,"p99":333.46301317214966},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":531.8019986152649,"p90":559.8850250244141,"p95":579.675018787384,"p99":599.465012550354},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":58.60704183578491,"p90":69.78297233581543,"p95":98.18500280380249,"p99":110.96400022506714},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0009765625,"host_stalled_samples":{"dispatch":2,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0038910505827516317,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":1189888,"scale_bytes":37184,"total_logical_bytes":1227072},"roundtrip":{"activation_data_bytes":3569664,"scale_bytes":37184,"total_logical_bytes":3606848},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2379776,"scale_bytes":0,"total_logical_bytes":2379776},"dispatch":{"activation_data_bytes":1189888,"scale_bytes":37184,"total_logical_bytes":1227072},"roundtrip":{"activation_data_bytes":3569664,"scale_bytes":37184,"total_logical_bytes":3606848},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":166,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":15,"mean":10.375,"min":7,"total":166},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.2465033382177353,"expert_assignments_per_rank":[17,14,14,16,10,14,9,8,13,9,11,11,11,17,10,8],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,0,5,16,11,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1875,"fanout_min":4,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.05421686917543411,"same_node_fraction":0.5481927990913391,"same_scaleup_domain_fraction":0.5481927990913391,"cross_node_fraction":0.4518072307109833,"cross_domain_fraction":0.4518072307109833,"gpus_per_node":8,"scale_up_domain":8,"copies":166},"payload_copies_per_rank":[15,13,11,14,9,13,9,7,10,8,9,8,9,14,9,8],"payload_rank_cv":0.24066247045993805,"routed_copies":166},"token_rate_at_latency_percentile":{"p50":54196.046270526516,"p90":51584.94783238759,"p95":46941.055924087246,"p99":45339.7793794929},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":180.31099438667297,"p90":184.7579926252365,"p95":186.24000251293182,"p99":190.12700021266937},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":372.5210130214691,"p90":389.7489905357361,"p95":421.27400636672974,"p99":437.33900785446167},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":552.8320074081421,"p90":574.5069831609726,"p95":607.5140088796616,"p99":627.466008067131},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":715.516984462738,"p90":729.9389839172363,"p95":736.7900013923645,"p99":838.1850123405457},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":591.0900235176086,"p90":620.2970147132874,"p95":682.1460127830505,"p99":701.0149955749512},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.70799995958805,"p90":21.511999890208244,"p95":22.633999586105347,"p99":23.95699918270111},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":192.8510069847107,"p90":196.05499505996704,"p95":197.69899547100067,"p99":201.1840045452118},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":429.96901273727417,"p90":441.2260055541992,"p95":445.51199674606323,"p99":448.5569894313812},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.210029125213623,"p90":6.328999996185303,"p95":6.328999996185303,"p99":6.328999996185303},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.812989711761475,"p90":14.581024646759033,"p95":17.305970191955566,"p99":20.71303129196167},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.1619925498962402,"p90":6.851017475128174,"p95":6.851017475128174,"p99":6.851017475128174},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":161.68400645256042,"p90":164.60900008678436,"p95":165.76999425888062,"p99":169.17499899864197},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":313.71399760246277,"p90":324.2489993572235,"p95":328.65598797798157,"p99":335.4659974575043},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":534.8870158195496,"p90":561.6080164909363,"p95":583.6009979248047,"p99":599.0239977836609},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":56.40304088592529,"p90":68.340003490448,"p95":97.58400917053223,"p99":108.48003625869751},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":2358272,"scale_bytes":73696,"total_logical_bytes":2431968},"roundtrip":{"activation_data_bytes":7074816,"scale_bytes":73696,"total_logical_bytes":7148512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4716544,"scale_bytes":0,"total_logical_bytes":4716544},"dispatch":{"activation_data_bytes":2358272,"scale_bytes":73696,"total_logical_bytes":2431968},"roundtrip":{"activation_data_bytes":7074816,"scale_bytes":73696,"total_logical_bytes":7148512},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":329,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":26,"mean":20.5625,"min":15,"total":329},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.17116330564022064,"expert_assignments_per_rank":[26,29,28,28,31,27,18,19,26,20,21,22,22,24,26,17],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,0,1,12,28,23,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.140625,"fanout_min":3,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.054711245000362396,"same_node_fraction":0.5075987577438354,"same_scaleup_domain_fraction":0.5075987577438354,"cross_node_fraction":0.49240121245384216,"cross_domain_fraction":0.49240121245384216,"gpus_per_node":8,"scale_up_domain":8,"copies":329},"payload_copies_per_rank":[22,25,24,25,26,24,17,15,21,17,18,18,19,21,21,16],"payload_rank_cv":0.1666751354932785,"routed_copies":329},"token_rate_at_latency_percentile":{"p50":108274.5393318137,"p90":103176.3791892211,"p95":93821.55550376946,"p99":91296.19252653667},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":182.07399547100067,"p90":186.15999817848206,"p95":188.24300169944763,"p99":192.96999275684357},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":373.0419874191284,"p90":389.4279897212982,"p95":420.1529920101166,"p99":436.2170100212097},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":555.1159828901291,"p90":575.5879878997803,"p95":608.3959937095642,"p99":629.1870027780533},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":713.0740284919739,"p90":727.1339893341064,"p95":736.3510131835938,"p99":848.3200073242188},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":590.408980846405,"p90":619.6150183677673,"p95":682.6270222663879,"p99":701.1749744415283},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":19.509999081492424,"p90":22.192999720573425,"p95":22.995000705122948,"p99":24.5169997215271},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":194.132998585701,"p90":197.53700494766235,"p95":198.86000454425812,"p99":202.9850035905838},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":428.52601408958435,"p90":438.38098645210266,"p95":440.86501002311707,"p99":448.1959939002991},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.290018558502197,"p90":6.408989429473877,"p95":6.408989429473877,"p99":6.408989429473877},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.173001766204834,"p90":14.104008674621582,"p95":16.705036163330078,"p99":19.909024238586426},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.8040409088134766,"p90":3.7260055541992188,"p95":3.7260055541992188,"p99":3.7260055541992188},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":162.08499670028687,"p90":164.88899290561676,"p95":165.76999425888062,"p99":169.3349927663803},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":315.79598784446716,"p90":326.21100544929504,"p95":329.97798919677734,"p99":335.6660008430481},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":535.0080132484436,"p90":562.8100037574768,"p95":581.637978553772,"p99":598.1829762458801},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":55.239975452423096,"p90":67.65997409820557,"p95":96.54301404953003,"p99":106.63801431655884},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":4652032,"scale_bytes":145376,"total_logical_bytes":4797408},"roundtrip":{"activation_data_bytes":13956096,"scale_bytes":145376,"total_logical_bytes":14101472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9304064,"scale_bytes":0,"total_logical_bytes":9304064},"dispatch":{"activation_data_bytes":4652032,"scale_bytes":145376,"total_logical_bytes":4797408},"roundtrip":{"activation_data_bytes":13956096,"scale_bytes":145376,"total_logical_bytes":14101472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":649,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":49,"mean":40.5625,"min":36,"total":649},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.10390592366456985,"expert_assignments_per_rank":[55,46,58,50,46,55,43,48,43,40,44,53,46,48,50,43],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,0,3,26,58,41,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.0703125,"fanout_min":3,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.07858243584632874,"same_node_fraction":0.5146378874778748,"same_scaleup_domain_fraction":0.5146378874778748,"cross_node_fraction":0.48536208271980286,"cross_domain_fraction":0.48536208271980286,"gpus_per_node":8,"scale_up_domain":8,"copies":649},"payload_copies_per_rank":[44,40,49,41,40,48,39,41,36,37,37,40,39,42,39,37],"payload_rank_cv":0.0888754278421402,"routed_copies":649},"token_rate_at_latency_percentile":{"p50":216798.8701941836,"p90":206579.88622868835,"p95":187510.88929211677,"p99":182550.72509105082},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":190.88700413703918,"p90":195.05299627780914,"p95":196.49599492549896,"p99":200.4619985818863},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":372.68099188804626,"p90":389.9889886379242,"p95":421.5550124645233,"p99":437.45899200439453},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":563.5679960250854,"p90":585.0419849157333,"p95":618.0510073900223,"p99":637.9209905862808},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":712.87602186203,"p90":730.2600145339966,"p95":738.4349703788757,"p99":839.506983757019},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":591.3299918174744,"p90":620.1760172843933,"p95":682.466983795166,"p99":702.6979923248291},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":23.876000195741653,"p90":25.558000430464745,"p95":26.480000466108322,"p99":28.442999348044395},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":199.62100684642792,"p90":203.2260000705719,"p95":204.7879993915558,"p99":209.11499857902527},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":419.7939932346344,"p90":430.1289916038513,"p95":433.4140121936798,"p99":442.90900230407715},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.36899471282959,"p90":6.4490437507629395,"p95":6.4490437507629395,"p99":6.4490437507629395},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.011996746063232,"p90":14.705002307891846,"p95":16.501963138580322,"p99":20.312011241912842},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-3.363966941833496,"p90":3.4050345420837402,"p95":3.4050345420837402,"p99":3.4050345420837402},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":168.25400292873383,"p90":171.45900428295135,"p95":172.46000468730927,"p99":176.06499791145325},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":319.28199529647827,"p90":330.17900586128235,"p95":334.8639905452728,"p99":343.7570035457611},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":543.9410209655762,"p90":570.0600147247314,"p95":590.9320116043091,"p99":606.5160036087036},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":47.23101854324341,"p90":59.325993061065674,"p95":89.77198600769043,"p99":102.07200050354004},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0009765625,"host_stalled_samples":{"dispatch":0,"roundtrip":2},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":1}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":9418752,"scale_bytes":294336,"total_logical_bytes":9713088},"roundtrip":{"activation_data_bytes":28256256,"scale_bytes":294336,"total_logical_bytes":28550592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":18837504,"scale_bytes":0,"total_logical_bytes":18837504},"dispatch":{"activation_data_bytes":9418752,"scale_bytes":294336,"total_logical_bytes":9713088},"roundtrip":{"activation_data_bytes":28256256,"scale_bytes":294336,"total_logical_bytes":28550592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1314,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":89,"mean":82.125,"min":72,"total":1314},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.07547592371702194,"expert_assignments_per_rank":[105,84,105,89,102,94,80,90,97,96,96,105,101,97,93,102],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,0,4,47,116,89,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.0669710785150528,"same_node_fraction":0.5350075960159302,"same_scaleup_domain_fraction":0.5350075960159302,"cross_node_fraction":0.4649924039840698,"cross_domain_fraction":0.4649924039840698,"gpus_per_node":8,"scale_up_domain":8,"copies":1314},"payload_copies_per_rank":[85,77,89,77,87,82,72,77,84,83,84,85,84,86,73,89],"payload_rank_cv":0.06281190365552902,"routed_copies":1314},"token_rate_at_latency_percentile":{"p50":432922.40126900154,"p90":412786.0363271778,"p95":375109.7211712666,"p99":364310.1343623329},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":204.86900210380554,"p90":209.39500629901886,"p95":210.91799437999725,"p99":214.48299288749695},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":372.96101450920105,"p90":389.4680142402649,"p95":419.5519983768463,"p99":435.2549910545349},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":577.8300166130066,"p90":598.8630205392838,"p95":630.4699927568436,"p99":649.7379839420319},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":721.6070294380188,"p90":734.3860268592834,"p95":743.6419725418091,"p99":811.5019798278809},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":591.0900235176086,"p90":620.8959817886353,"p95":682.6670169830322,"p99":702.7379870414734},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":27.842000126838684,"p90":30.446000397205353,"p95":31.40699863433838,"p99":33.00999850034714},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":214.6030068397522,"p90":218.0889993906021,"p95":219.4100022315979,"p99":223.13599288463593},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":416.0279929637909,"p90":421.43601179122925,"p95":422.9179918766022,"p99":429.4080138206482},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.208956241607666,"p90":6.3700079917907715,"p95":6.3700079917907715,"p99":6.3700079917907715},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.092999458312988,"p90":14.584958553314209,"p95":16.781985759735107,"p99":20.027995109558105},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.044009208679199,"p90":5.207955837249756,"p95":5.207955837249756,"p99":5.207955837249756},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":182.75600671768188,"p90":185.59999763965607,"p95":186.64099276065826,"p99":189.68600034713745},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":331.0990035533905,"p90":342.7169919013977,"p95":348.12599420547485,"p99":357.01900720596313},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":556.5999746322632,"p90":582.3990106582642,"p95":604.3130159378052,"p99":621.8580007553101},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":34.53195095062256,"p90":46.549975872039795,"p95":75.03104209899902,"p99":86.7680311203003},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":18808832,"scale_bytes":587776,"total_logical_bytes":19396608},"roundtrip":{"activation_data_bytes":56426496,"scale_bytes":587776,"total_logical_bytes":57014272},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":37617664,"scale_bytes":0,"total_logical_bytes":37617664},"dispatch":{"activation_data_bytes":18808832,"scale_bytes":587776,"total_logical_bytes":19396608},"roundtrip":{"activation_data_bytes":56426496,"scale_bytes":587776,"total_logical_bytes":57014272},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2624,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":178,"mean":164.0,"min":148,"total":2624},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.061073120683431625,"expert_assignments_per_rank":[205,191,207,186,202,197,181,178,184,196,178,211,175,205,178,198],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,0,7,95,237,173,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.125,"fanout_min":3,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.059832315891981125,"same_node_fraction":0.4878048896789551,"same_scaleup_domain_fraction":0.4878048896789551,"cross_node_fraction":0.5121951103210449,"cross_domain_fraction":0.5121951103210449,"gpus_per_node":8,"scale_up_domain":8,"copies":2624},"payload_copies_per_rank":[171,162,178,161,167,167,161,155,157,176,151,172,152,174,148,172],"payload_rank_cv":0.05613400414586067,"routed_copies":2624},"token_rate_at_latency_percentile":{"p50":866196.3146545097,"p90":824614.7744829416,"p95":749999.6151311436,"p99":728578.8009774736},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":226.2209951877594,"p90":230.58700561523438,"p95":232.791006565094,"p99":237.23700642585754},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":374.72400069236755,"p90":390.3470039367676,"p95":419.1089868545532,"p99":434.9740147590637},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":600.944995880127,"p90":620.934009552002,"p95":651.8999934196472,"p99":672.2110211849213},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":746.766984462738,"p90":758.5039734840393,"p95":770.0790166854858,"p99":801.2089729309082},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":606.7129969596863,"p90":631.7510008811951,"p95":686.7139935493469,"p99":705.9429883956909},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":38.49799931049347,"p90":40.54100066423416,"p95":41.30199924111366,"p99":42.62400045990944},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":237.91900277137756,"p90":241.1630004644394,"p95":242.12400615215302,"p99":244.32800710201263},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":407.8559875488281,"p90":413.58399391174316,"p95":415.3069853782654,"p99":417.5100028514862},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.249964237213135,"p90":6.289958953857422,"p95":6.289958953857422,"p99":6.289958953857422},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.3310136795043945,"p90":13.257026672363281,"p95":14.658987522125244,"p99":21.432995796203613},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.8420219421386719,"p90":4.287004470825195,"p95":4.287004470825195,"p99":4.287004470825195},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":207.39300549030304,"p90":210.5170041322708,"p95":211.47899329662323,"p99":214.52400088310242},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":354.49498891830444,"p90":365.51201343536377,"p95":369.1979944705963,"p99":376.6080141067505},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":587.0869755744934,"p90":610.9619736671448,"p95":632.6749920845032,"p99":649.058997631073},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.665015697479248,"p90":25.637030601501465,"p95":50.752997398376465,"p99":61.930954456329346},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":37675008,"scale_bytes":1177344,"total_logical_bytes":38852352},"roundtrip":{"activation_data_bytes":113025024,"scale_bytes":1177344,"total_logical_bytes":114202368},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":75350016,"scale_bytes":0,"total_logical_bytes":75350016},"dispatch":{"activation_data_bytes":37675008,"scale_bytes":1177344,"total_logical_bytes":38852352},"roundtrip":{"activation_data_bytes":113025024,"scale_bytes":1177344,"total_logical_bytes":114202368},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":5256,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":363,"mean":328.5,"min":292,"total":5256},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.04799197241663933,"expert_assignments_per_rank":[383,392,421,335,392,387,382,365,377,385,368,393,374,408,391,391],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,0,15,181,481,347,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1328125,"fanout_min":3,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.0641172006726265,"same_node_fraction":0.5055174827575684,"same_scaleup_domain_fraction":0.5055174827575684,"cross_node_fraction":0.49448248744010925,"cross_domain_fraction":0.49448248744010925,"gpus_per_node":8,"scale_up_domain":8,"copies":5256},"payload_copies_per_rank":[328,330,363,292,334,320,334,318,322,334,323,331,317,344,325,341],"payload_rank_cv":0.044492896646261215,"routed_copies":5256},"token_rate_at_latency_percentile":{"p50":1687783.1942473466,"p90":1620891.7731379583,"p95":1491159.3612755118,"p99":1450542.0647736976},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":293.36199164390564,"p90":298.3289957046509,"p95":300.37200450897217,"p99":304.17799949645996},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":432.41098523139954,"p90":442.26598739624023,"p95":447.55300879478455,"p99":459.21099185943604},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":725.7729768753052,"p90":740.5949831008911,"p95":747.9250133037567,"p99":763.388991355896},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":814.3100142478943,"p90":826.3239860534668,"p95":830.8939933776855,"p99":865.3830289840698},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":727.0950078964233,"p90":741.2350177764893,"p95":745.3219890594482,"p99":756.738007068634},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":48.07300120592117,"p90":50.03499984741211,"p95":50.7970005273819,"p99":52.43900045752525},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":305.7810068130493,"p90":309.6660077571869,"p95":310.8690083026886,"p99":315.35500288009644},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":416.14800691604614,"p90":422.3569929599762,"p95":425.00099539756775,"p99":429.72901463508606},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.25002384185791,"p90":6.329953670501709,"p95":6.329953670501709,"p99":6.329953670501709},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.57499647140503,"p90":15.900969505310059,"p95":18.065035343170166,"p99":21.232008934020996},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.4049878120422363,"p90":2.68399715423584,"p95":2.68399715423584,"p99":2.68399715423584},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":274.1740047931671,"p90":277.61900424957275,"p95":278.62000465393066,"p99":281.94499015808105},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":411.78199648857117,"p90":420.5949902534485,"p95":423.1970012187958,"p99":429.76701259613037},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":707.3469758033752,"p90":720.9680080413818,"p95":724.8539924621582,"p99":735.2290153503418},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":19.145965576171875,"p90":23.350954055786133,"p95":25.07603168487549,"p99":28.400003910064697},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0041841003112494946,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":75787264,"scale_bytes":2368352,"total_logical_bytes":78155616},"roundtrip":{"activation_data_bytes":227361792,"scale_bytes":2368352,"total_logical_bytes":229730144},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":151574528,"scale_bytes":0,"total_logical_bytes":151574528},"dispatch":{"activation_data_bytes":75787264,"scale_bytes":2368352,"total_logical_bytes":78155616},"roundtrip":{"activation_data_bytes":227361792,"scale_bytes":2368352,"total_logical_bytes":229730144},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":10573,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":692,"mean":660.8125,"min":603,"total":10573},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03450827673077583,"expert_assignments_per_rank":[801,741,794,696,766,790,752,740,759,786,758,783,779,802,773,768],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,1,23,348,946,730,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16259765625,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.06195025146007538,"same_node_fraction":0.5033575892448425,"same_scaleup_domain_fraction":0.5033575892448425,"cross_node_fraction":0.4966423809528351,"cross_domain_fraction":0.4966423809528351,"gpus_per_node":8,"scale_up_domain":8,"copies":10573},"payload_copies_per_rank":[689,638,684,603,655,661,656,647,649,688,663,667,660,692,662,659],"payload_rank_cv":0.03241767734289169,"routed_copies":10573},"token_rate_at_latency_percentile":{"p50":2816688.2976203067,"p90":2762956.3510686033,"p95":2747805.6867535245,"p99":2706352.76789296},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":377.3680031299591,"p90":382.81700015068054,"p95":384.89899039268494,"p99":389.2660140991211},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":545.3010201454163,"p90":555.2759766578674,"p95":559.6830248832703,"p99":572.3040103912354},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":922.6690232753754,"p90":938.092976808548,"p95":944.5820152759552,"p99":961.5700244903564},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":991.1320209503174,"p90":998.6280202865601,"p95":1002.189040184021,"p99":1006.598949432373},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":913.815975189209,"p90":925.432026386261,"p95":929.718017578125,"p99":937.2509717941284},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":69.46399807929993,"p90":71.58800214529037,"p95":72.42999970912933,"p99":73.75100255012512},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":448.7549960613251,"p90":455.4069936275482,"p95":457.5279951095581,"p99":461.5350067615509},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":506.28501176834106,"p90":511.9730234146118,"p95":514.136016368866,"p99":517.3810124397278},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.330013275146484,"p90":6.3700079917907715,"p95":6.3700079917907715,"p99":6.3700079917907715},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":12.419044971466064,"p90":17.342984676361084,"p95":19.429028034210205,"p99":22.350013256072998},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.083944320678711,"p90":6.489038467407227,"p95":6.489038467407227,"p99":6.489038467407227},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":349.808007478714,"p90":354.1739881038666,"p95":355.65701127052307,"p99":359.02100801467896},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":512.0530128479004,"p90":522.6290225982666,"p95":525.1529812812805,"p99":530.7610034942627},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":886.2580060958862,"p90":896.9939947128296,"p95":901.0400176048279,"p99":908.2099795341492},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":27.356982231140137,"p90":32.40501880645752,"p95":34.2479944229126,"p99":39.134979248046875},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0041841003112494946,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":151825408,"scale_bytes":4744544,"total_logical_bytes":156569952},"roundtrip":{"activation_data_bytes":455476224,"scale_bytes":4744544,"total_logical_bytes":460220768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":303650816,"scale_bytes":0,"total_logical_bytes":303650816},"dispatch":{"activation_data_bytes":151825408,"scale_bytes":4744544,"total_logical_bytes":156569952},"roundtrip":{"activation_data_bytes":455476224,"scale_bytes":4744544,"total_logical_bytes":460220768},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":21181,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":1373,"mean":1323.8125,"min":1247,"total":21181},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.027300119400024414,"expert_assignments_per_rank":[1613,1510,1534,1426,1574,1521,1542,1480,1511,1563,1557,1535,1576,1532,1570,1532],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,2,50,672,1893,1479,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.171142578125,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.0642084851861,"same_node_fraction":0.4972380995750427,"same_scaleup_domain_fraction":0.4972380995750427,"cross_node_fraction":0.5027619004249573,"cross_domain_fraction":0.5027619004249573,"gpus_per_node":8,"scale_up_domain":8,"copies":21181},"payload_copies_per_rank":[1373,1292,1338,1247,1349,1309,1342,1292,1296,1358,1337,1317,1343,1320,1354,1314],"payload_rank_cv":0.023345857858657837,"routed_copies":21181},"token_rate_at_latency_percentile":{"p50":4482302.904752687,"p90":4426040.901128694,"p95":4405636.894797308,"p99":4370227.530582605},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":498.1899857521057,"p90":503.7990212440491,"p95":506.0020089149475,"p99":509.365975856781},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":736.8289828300476,"p90":745.881974697113,"p95":749.2870092391968,"p99":756.3380002975464},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1235.0189685821533,"p90":1249.680995941162,"p95":1255.2890181541443,"p99":1265.7039761543274},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1321.7500448226929,"p90":1332.5669765472412,"p95":1336.5720510482788,"p99":1347.9950428009033},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1223.04105758667,"p90":1235.298991203308,"p95":1238.3840084075928,"p99":1247.4770545959473},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":116.57500267028809,"p90":119.41999942064285,"p95":120.34200131893158,"p99":122.6240023970604},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":568.6590075492859,"p90":576.2680172920227,"p95":579.3520212173462,"p99":584.3999981880188},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":713.6350274085999,"p90":720.0440168380737,"p95":721.8090295791626,"p99":726.6939878463745},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.329059600830078,"p90":6.5299272537231445,"p95":6.5299272537231445,"p99":6.5299272537231445},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.78697967529297,"p90":25.51901340484619,"p95":28.757929801940918,"p99":38.73896598815918},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-6.7299604415893555,"p90":3.165006637573242,"p95":3.165006637573242,"p99":3.165006637573242},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":469.74900364875793,"p90":474.1159975528717,"p95":475.6369888782501,"p99":479.6440005302429},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":701.8600106239319,"p90":712.315022945404,"p95":714.6779894828796,"p99":719.6059823036194},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1194.5630311965942,"p90":1205.8199644088745,"p95":1208.7440490722656,"p99":1216.7969942092896},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":28.558015823364258,"p90":34.36708450317383,"p95":35.768985748291016,"p99":39.37399387359619},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0041841003112494946,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":303951872,"scale_bytes":9498496,"total_logical_bytes":313450368},"roundtrip":{"activation_data_bytes":911855616,"scale_bytes":9498496,"total_logical_bytes":921354112},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":607903744,"scale_bytes":0,"total_logical_bytes":607903744},"dispatch":{"activation_data_bytes":303951872,"scale_bytes":9498496,"total_logical_bytes":313450368},"roundtrip":{"activation_data_bytes":911855616,"scale_bytes":9498496,"total_logical_bytes":921354112},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":42404,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":2702,"mean":2650.25,"min":2578,"total":42404},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.017352504655718803,"expert_assignments_per_rank":[3108,3051,3103,2954,3171,3067,3121,2998,3045,3065,3095,3023,3150,3071,3090,3040],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,3,103,1301,3825,2960,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.17626953125,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.06268276274204254,"same_node_fraction":0.49721723794937134,"same_scaleup_domain_fraction":0.49721723794937134,"cross_node_fraction":0.5027827620506287,"cross_domain_fraction":0.5027827620506287,"gpus_per_node":8,"scale_up_domain":8,"copies":42404},"payload_copies_per_rank":[2682,2637,2676,2578,2702,2659,2687,2602,2635,2661,2672,2599,2691,2625,2662,2636],"payload_rank_cv":0.013201921246945858,"routed_copies":42404},"token_rate_at_latency_percentile":{"p50":6698058.048978851,"p90":6631592.884262093,"p95":6615072.5012461115,"p99":6566854.251802936},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":36,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":20,"num_max_rdma_chunked_recv_tokens":140},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":12,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":1073741824,"num_qps_per_rank":10,"is_intranode":false,"cpu_proxy_threads":4},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9948,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2096,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2091,"p50":2101,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2089,"p50":2100,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2088,"p50":2099,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2085,"p50":2094,"max":2096,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":54,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2085,"p50":2097,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2096,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2089,"p50":2098,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2090,"p50":2100,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":54,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2084,"p50":2094,"max":2096,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2085,"p50":2095,"max":2097,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2085,"p50":2093,"max":2095,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2089,"p50":2100,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2085,"p50":2097,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2083,"p50":2090,"max":2093,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2098,"max":2100,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":53,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":552.1}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-bf16.json b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-bf16.json new file mode 100644 index 0000000000..0c1fc0a9d6 --- /dev/null +++ b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T06:43:36.133653+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":1,"phase":"decode","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":72.26800173521042,"p90":76.67499780654907,"p95":78.3580020070076,"p99":80.80100268125534},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":164.44799304008484,"p90":175.10299384593964,"p95":180.7519942522049,"p99":189.04399871826172},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":236.71599477529526,"p90":251.7779916524887,"p95":259.1099962592125,"p99":269.84500139951706},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":263.31600546836853,"p90":274.3319869041443,"p95":277.8179943561554,"p99":318.7600076198578},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":238.5990023612976,"p90":250.41699409484863,"p95":256.1050057411194,"p99":264.47799801826477},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":80.11999726295471,"p90":82.28400349617004,"p95":83.80600064992905,"p99":87.65199780464172},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":174.06199872493744,"p90":181.31299316883087,"p95":183.35600197315216,"p99":192.96999275684357},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.009995937347412,"p90":6.2490105628967285,"p95":6.2490105628967285,"p99":6.2490105628967285},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.845006227493286,"p90":8.851975202560425,"p95":11.055976152420044,"p99":17.266005277633667},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.3220012187957764,"p90":4.166007041931152,"p95":4.166007041931152,"p99":4.166007041931152},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":65.1789978146553,"p90":67.34099984169006,"p95":68.70300322771072,"p99":70.62599807977676},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":152.9099941253662,"p90":160.2409929037094,"p95":165.36900401115417,"p99":171.8990057706833},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":230.5469959974289,"p90":242.60400235652924,"p95":248.17200005054474,"p99":256.46600127220154},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.531002163887024,"p90":10.737001895904541,"p95":12.138009071350098,"p99":14.82100784778595},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8,"byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"roundtrip":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"roundtrip":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":35,"assignments":48,"wire":"rank-deduplicated"},"receive":{"max":6,"mean":4.375,"min":2,"total":35},"routing":{"empty_expert_count":337,"empty_rank_count":0,"expert_assignment_rank_cv":0.3227486312389374,"expert_assignments_per_rank":[8,6,6,6,6,5,9,2],"expert_load_cv":2.708012819290161,"expert_load_max":2,"expert_load_mean":0.125,"expert_load_min":0,"fanout_histogram":[0,0,0,5,3,0,0,0],"fanout_max":5,"fanout_mean":4.375,"fanout_min":4,"hotspot_ratio":16.0,"locality":{"placement":"packed","local_rank_fraction":0.08571428805589676,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":35},"payload_copies_per_rank":[5,5,5,4,4,4,6,2],"payload_rank_cv":0.2539483904838562,"routed_copies":35},"token_rate_at_latency_percentile":{"p50":33529.05888468901,"p90":31946.713636255445,"p95":31237.187171916128,"p99":30248.26284206644},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":75.63400268554688,"p90":79.83999699354172,"p95":81.24200254678726,"p99":83.88599753379822},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":164.0859991312027,"p90":176.3859987258911,"p95":182.71400034427643,"p99":190.28599560260773},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":239.72000181674957,"p90":256.22599571943283,"p95":263.9560028910637,"p99":274.17199313640594},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":264.1170024871826,"p90":276.45599842071533,"p95":281.54298663139343,"p99":294.2430078983307},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":242.00299382209778,"p90":254.86299395561218,"p95":260.23098826408386,"p99":268.40299367904663},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":83.64500105381012,"p90":86.52999997138977,"p95":88.41300010681152,"p99":91.01700037717819},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":173.78099262714386,"p90":183.2759976387024,"p95":185.47900021076202,"p99":197.97800481319427},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.2490105628967285,"p95":6.2490105628967285,"p99":6.2490105628967285},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.4050047397613525,"p90":9.133994579315186,"p95":11.177003383636475,"p99":18.187016248703003},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.9219815731048584,"p90":2.763986587524414,"p95":2.763986587524414,"p99":2.763986587524414},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":66.7010024189949,"p90":69.46399807929993,"p95":71.10700011253357,"p99":73.99100065231323},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":153.07000279426575,"p90":162.64399886131287,"p95":168.17300021648407,"p99":174.02100563049316},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":232.42899775505066,"p90":245.28899788856506,"p95":250.93701481819153,"p99":259.0689957141876},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.332999587059021,"p90":12.217998504638672,"p95":13.900995254516602,"p99":16.62398874759674},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"roundtrip":{"activation_data_bytes":2007040,"scale_bytes":0,"total_logical_bytes":2007040},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"roundtrip":{"activation_data_bytes":2007040,"scale_bytes":0,"total_logical_bytes":2007040},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":70,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":12,"mean":8.75,"min":5,"total":70},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.33592739701271057,"expert_assignments_per_rank":[18,17,13,7,13,7,13,8],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,3,6,5,2,0,0],"fanout_max":6,"fanout_mean":4.375,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.15714286267757416,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":70},"payload_copies_per_rank":[12,11,10,5,9,6,10,7],"payload_rank_cv":0.26649653911590576,"routed_copies":70},"token_rate_at_latency_percentile":{"p50":66114.88456114715,"p90":62778.82776024602,"p95":61483.838288171544,"p99":59611.853730411916},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":76.95599645376205,"p90":81.28199726343155,"p95":82.80500024557114,"p99":86.13000065088272},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":164.0460044145584,"p90":175.54399371147156,"p95":182.75399506092072,"p99":190.5670017004013},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":241.00200086832047,"p90":256.8259909749031,"p95":265.55899530649185,"p99":276.697002351284},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":262.1940076351166,"p90":272.2499966621399,"p95":278.73900532722473,"p99":360.5419993400574},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":243.60600113868713,"p90":257.46798515319824,"p95":263.0760073661804,"p99":273.09098839759827},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":85.24899929761887,"p90":89.09399807453156,"p95":90.61600267887115,"p99":96.38500213623047},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":174.54299330711365,"p90":181.43300712108612,"p95":184.11600589752197,"p99":193.77200305461884},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":6.2490105628967285,"p95":6.2490105628967285,"p99":6.2490105628967285},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.4849941730499268,"p90":8.893996477127075,"p95":10.976999998092651,"p99":21.75202965736389},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.3620257377624512,"p90":1.7220079898834229,"p95":1.7220079898834229,"p99":1.7220079898834229},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":69.5049986243248,"p90":71.5079978108406,"p95":73.15000146627426,"p99":75.27299970388412},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":154.03200685977936,"p90":165.24900496006012,"p95":171.45800590515137,"p99":177.946999669075},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":235.5549931526184,"p90":249.41499531269073,"p95":255.10400533676147,"p99":264.6380066871643},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.61200487613678,"p90":10.856002569198608,"p95":12.339010834693909,"p99":17.106011509895325},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"roundtrip":{"activation_data_bytes":4100096,"scale_bytes":0,"total_logical_bytes":4100096},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"roundtrip":{"activation_data_bytes":4100096,"scale_bytes":0,"total_logical_bytes":4100096},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":143,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":22,"mean":17.875,"min":13,"total":143},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.20518453419208527,"expert_assignments_per_rank":[31,30,24,17,22,22,28,18],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,4,12,13,3,0,0],"fanout_max":6,"fanout_mean":4.46875,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.1258741319179535,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":143},"payload_copies_per_rank":[22,21,18,13,16,15,22,16],"payload_rank_cv":0.1800614446401596,"routed_copies":143},"token_rate_at_latency_percentile":{"p50":131359.65391009438,"p90":124287.29723798245,"p95":121637.85029418742,"p99":117177.06317504187},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":77.95699685811996,"p90":82.24300295114517,"p95":83.80600064992905,"p99":86.52999997138977},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":164.20699656009674,"p90":176.34500563144684,"p95":181.83299899101257,"p99":189.88600373268127},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":242.1639934182167,"p90":258.588008582592,"p95":265.6389996409416,"p99":276.41600370407104},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":262.51500844955444,"p90":271.5680003166199,"p95":280.46199679374695,"p99":364.0669882297516},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":244.7279989719391,"p90":257.18700885772705,"p95":262.55500316619873,"p99":269.60501074790955},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":88.29200267791748,"p90":90.85600078105927,"p95":91.57799929380417,"p99":94.82300281524658},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":177.10599303245544,"p90":183.67600440979004,"p95":186.39999628067017,"p99":190.5670017004013},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.050020456314087,"p90":6.089001893997192,"p95":6.089001893997192,"p99":6.089001893997192},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.564983606338501,"p90":8.973002433776855,"p95":10.135024785995483,"p99":14.541998505592346},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.7200241088867188,"p90":1.5210211277008057,"p95":1.5210211277008057,"p99":1.5210211277008057},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":70.82699984312057,"p90":73.15000146627426,"p95":74.47200268507004,"p99":76.95599645376205},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":154.87299859523773,"p90":168.29299926757812,"p95":173.05999994277954,"p99":179.63099479675293},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":237.19699680805206,"p90":249.17499721050262,"p95":254.86299395561218,"p99":261.3129913806915},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.2109997272491455,"p90":9.77499783039093,"p95":12.057989835739136,"p99":18.14599335193634},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"roundtrip":{"activation_data_bytes":8056832,"scale_bytes":0,"total_logical_bytes":8056832},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"roundtrip":{"activation_data_bytes":8056832,"scale_bytes":0,"total_logical_bytes":8056832},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":281,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":42,"mean":35.125,"min":26,"total":281},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.1458333283662796,"expert_assignments_per_rank":[55,56,58,37,46,43,46,43],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,1,9,24,24,6,0,0],"fanout_max":6,"fanout_mean":4.390625,"fanout_min":2,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.11743772029876709,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":281},"payload_copies_per_rank":[38,42,42,26,29,31,38,35],"payload_rank_cv":0.15879225730895996,"routed_copies":281},"token_rate_at_latency_percentile":{"p50":261514.8257202003,"p90":248846.1617258595,"p95":243758.44767082063,"p99":237384.31204397132},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":81.52200281620026,"p90":85.72900295257568,"p95":87.53199875354767,"p99":90.33600240945816},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":164.00599479675293,"p90":177.06699669361115,"p95":183.47500264644623,"p99":191.32700562477112},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":245.5279976129532,"p90":262.79599964618683,"p95":271.0070013999939,"p99":281.6630080342293},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":263.4359896183014,"p90":273.5710144042969,"p95":279.0600061416626,"p99":374.72400069236755},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":249.2150068283081,"p90":262.51500844955444,"p95":267.7629888057709,"p99":274.7330069541931},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":92.13899821043015,"p90":94.26199644804001,"p95":95.06300091743469,"p99":98.30799698829651},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":175.46400427818298,"p90":182.71499872207642,"p95":186.0000044107437,"p99":193.97200644016266},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":6.20800256729126,"p95":6.20800256729126,"p99":6.20800256729126},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.764986991882324,"p90":9.373992681503296,"p95":11.978000402450562,"p99":18.10699701309204},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.242995262145996,"p90":1.76200270652771,"p95":1.76200270652771,"p99":1.76200270652771},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":75.99300146102905,"p90":77.83699780702591,"p95":78.95900309085846,"p99":81.04199916124344},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":157.87799656391144,"p90":170.6160008907318,"p95":176.86599493026733,"p99":184.11700427532196},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":243.00600588321686,"p90":256.5859854221344,"p95":261.2319886684418,"p99":267.4030065536499},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.768001079559326,"p90":9.254008531570435,"p95":10.575994849205017,"p99":15.262991189956665},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"roundtrip":{"activation_data_bytes":16171008,"scale_bytes":0,"total_logical_bytes":16171008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"roundtrip":{"activation_data_bytes":16171008,"scale_bytes":0,"total_logical_bytes":16171008},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":564,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":77,"mean":70.5,"min":62,"total":564},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.07384087145328522,"expert_assignments_per_rank":[101,108,101,91,83,97,94,93],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,2,14,52,50,10,0,0],"fanout_max":6,"fanout_mean":4.40625,"fanout_min":2,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.14716312289237976,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":564},"payload_copies_per_rank":[73,76,77,63,62,69,74,70],"payload_rank_cv":0.07472094893455505,"routed_copies":564},"token_rate_at_latency_percentile":{"p50":513612.729943599,"p90":487591.16957153636,"p95":478034.6999071192,"p99":465906.88690471667},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":90.41599929332733,"p90":95.22300213575363,"p95":96.86599671840668,"p99":99.74999725818634},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":173.1799989938736,"p90":186.7610067129135,"p95":192.32900440692902,"p99":198.97900521755219},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":263.5959982872009,"p90":281.98400884866714,"p95":289.1950011253357,"p99":298.7290024757385},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":264.8789882659912,"p90":277.3370146751404,"p95":292.1999990940094,"p99":370.0369894504547},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":257.0669949054718,"p90":270.5659866333008,"p95":275.93499422073364,"p99":287.83300518989563},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":101.71200335025787,"p90":103.67599874734879,"p95":105.35799711942673,"p99":109.04400050640106},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":177.66700685024261,"p90":185.7990026473999,"p95":188.52399289608002,"p99":193.53100657463074},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.048977375030518,"p90":6.28998875617981,"p95":6.28998875617981,"p99":6.28998875617981},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.4870078563690186,"p90":11.25800609588623,"p95":14.421999454498291,"p99":19.950002431869507},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.4839937686920166,"p90":5.2490234375,"p95":5.2490234375,"p99":5.2490234375},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":85.80899983644485,"p90":87.2109979391098,"p95":88.29300105571747,"p99":90.65599739551544},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":168.8939929008484,"p90":181.87299370765686,"p95":187.64199316501617,"p99":192.65000522136688},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":251.81901454925537,"p90":265.4789984226227,"p95":271.0070013999939,"p99":280.8220088481903},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.567012190818787,"p90":8.172020316123962,"p95":9.73498821258545,"p99":12.578994035720825},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"roundtrip":{"activation_data_bytes":32628736,"scale_bytes":0,"total_logical_bytes":32628736},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"roundtrip":{"activation_data_bytes":32628736,"scale_bytes":0,"total_logical_bytes":32628736},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1138,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":149,"mean":142.25,"min":126,"total":1138},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.04658475145697594,"expert_assignments_per_rank":[189,194,196,170,193,201,198,195],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,3,26,101,106,20,0,0],"fanout_max":6,"fanout_mean":4.4453125,"fanout_min":2,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.1344463974237442,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":1138},"payload_copies_per_rank":[143,141,149,126,144,145,148,142],"payload_rank_cv":0.04686218500137329,"routed_copies":1138},"token_rate_at_latency_percentile":{"p50":995849.3508438758,"p90":946164.7533211848,"p95":927754.7442757962,"p99":889404.6040033037},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":104.27699983119965,"p90":108.20300132036209,"p95":110.00499874353409,"p99":113.04999887943268},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":190.6069964170456,"p90":203.66600155830383,"p95":208.99300277233124,"p99":215.40400385856628},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":294.88399624824524,"p90":311.8690028786659,"p95":318.9980015158653,"p99":328.45400273799896},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":285.6299877166748,"p90":294.2819893360138,"p95":298.007994890213,"p99":310.78699231147766},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":286.6309881210327,"p90":299.40998554229736,"p95":303.8969933986664,"p99":311.7089867591858},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":98.90899807214737,"p90":102.59400308132172,"p95":105.31900078058243,"p99":108.28299820423126},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":185.8000010251999,"p90":190.16599655151367,"p95":192.20900535583496,"p99":195.8940029144287},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":6.288975477218628,"p95":6.288975477218628,"p99":6.288975477218628},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.926008939743042,"p90":11.137008666992188,"p95":15.343993902206421,"p99":18.908023834228516},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.2820065021514893,"p90":2.884000539779663,"p95":2.884000539779663,"p99":2.884000539779663},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":96.70499712228775,"p90":98.66800159215927,"p95":99.87100213766098,"p99":101.83300077915192},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":183.19499492645264,"p90":195.73399424552917,"p95":201.14299654960632,"p99":206.83099329471588},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":278.5789966583252,"p90":291.6789948940277,"p95":295.7650125026703,"p99":302.8959929943085},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.771998643875122,"p90":10.416001081466675,"p95":11.777997016906738,"p99":14.180988073348999},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.004672897048294544,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"roundtrip":{"activation_data_bytes":64569344,"scale_bytes":0,"total_logical_bytes":64569344},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"roundtrip":{"activation_data_bytes":64569344,"scale_bytes":0,"total_logical_bytes":64569344},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2252,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":290,"mean":281.5,"min":268,"total":2252},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03184119611978531,"expert_assignments_per_rank":[396,393,399,359,380,389,380,376],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,3,57,223,191,38,0,0],"fanout_max":6,"fanout_mean":4.3984375,"fanout_min":2,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.11767318099737167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":2252},"payload_copies_per_rank":[290,286,290,268,282,282,280,274],"payload_rank_cv":0.025306938216090202,"routed_copies":2252},"token_rate_at_latency_percentile":{"p50":1786268.830723226,"p90":1710029.807698816,"p95":1684781.3934385797,"p99":1642557.7116759592},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":121.54199928045273,"p90":125.98900496959686,"p95":127.83199548721313,"p99":130.31600415706635},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":208.91299843788147,"p90":220.09100019931793,"p95":225.4589945077896,"p99":231.7889928817749},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":330.4549977183342,"p90":346.0800051689148,"p95":353.29098999500275,"p99":362.10499703884125},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":324.768990278244,"p90":331.4580023288727,"p95":338.90900015830994,"p99":370.39700150489807},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":323.92698526382446,"p90":335.82499623298645,"p95":340.1919901371002,"p99":345.7599878311157},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":116.09400063753128,"p90":118.13700199127197,"p95":119.0589964389801,"p99":121.8630000948906},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":205.34899830818176,"p90":209.75500345230103,"p95":210.87700128555298,"p99":213.8810008764267},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.168991327285767,"p90":6.288975477218628,"p95":6.288975477218628,"p99":6.288975477218628},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.727989435195923,"p90":9.97498631477356,"p95":11.377006769180298,"p99":14.18200135231018},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.7610023021697998,"p90":1.5230178833007812,"p95":1.5230178833007812,"p99":1.5230178833007812},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":115.85400253534317,"p90":117.61700361967087,"p95":118.69800090789795,"p99":120.78200280666351},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":203.06600630283356,"p90":213.8410061597824,"p95":219.1299945116043,"p99":225.4589945077896},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":317.59798526763916,"p90":329.4160068035126,"p95":333.7019979953766,"p99":339.4300043582916},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.768001079559326,"p90":8.693009614944458,"p95":10.295987129211426,"p99":13.300001621246338},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"roundtrip":{"activation_data_bytes":128651264,"scale_bytes":0,"total_logical_bytes":128651264},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"roundtrip":{"activation_data_bytes":128651264,"scale_bytes":0,"total_logical_bytes":128651264},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":4487,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":569,"mean":560.875,"min":553,"total":4487},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.01615842990577221,"expert_assignments_per_rank":[775,756,779,747,762,761,782,782],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,5,126,437,385,71,0,0],"fanout_max":6,"fanout_mean":4.3818359375,"fanout_min":2,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.12859371304512024,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":4487},"payload_copies_per_rank":[569,555,565,556,553,564,558,567],"payload_rank_cv":0.010142241604626179,"routed_copies":4487},"token_rate_at_latency_percentile":{"p50":3161206.218018534,"p90":3049207.2105602766,"p95":3010064.9917927803,"p99":2961591.9598544366},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":171.49800062179565,"p90":175.6639927625656,"p95":177.54699289798737,"p99":180.59100210666656},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":256.3849985599518,"p90":265.27899503707886,"p95":269.4860100746155,"p99":275.6139934062958},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":427.88299918174744,"p90":440.94298779964447,"p95":447.03300297260284,"p99":456.20499551296234},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":420.1120138168335,"p90":429.7269880771637,"p95":439.46099281311035,"p99":449.4360089302063},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":421.03299498558044,"p90":429.7260046005249,"p95":434.2130124568939,"p99":441.74501299858093},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":161.8030071258545,"p90":164.5669937133789,"p95":165.40899872779846,"p99":167.69200563430786},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":251.57800316810608,"p90":255.38399815559387,"p95":256.8660080432892,"p99":260.4309916496277},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":6.168991327285767,"p95":6.168991327285767,"p99":6.168991327285767},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.767000675201416,"p90":9.6149742603302,"p95":10.576009750366211,"p99":15.302985906600952},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.5619993209838867,"p90":0.9620189666748047,"p95":0.9620189666748047,"p99":0.9620189666748047},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":161.40300035476685,"p90":164.16700184345245,"p95":165.72900116443634,"p99":168.29299926757812},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":248.29299747943878,"p90":256.6260099411011,"p95":261.03299856185913,"p99":266.52100682258606},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":410.09700298309326,"p90":418.0290102958679,"p95":423.1559932231903,"p99":430.0059974193573},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.697007179260254,"p90":13.220012187957764,"p95":14.701992273330688,"p99":19.2679762840271},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"roundtrip":{"activation_data_bytes":259768320,"scale_bytes":0,"total_logical_bytes":259768320},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"roundtrip":{"activation_data_bytes":259768320,"scale_bytes":0,"total_logical_bytes":259768320},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":9060,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":1149,"mean":1132.5,"min":1088,"total":9060},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.018761977553367615,"expert_assignments_per_rank":[1542,1490,1556,1492,1545,1541,1581,1541],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,11,225,859,791,162,0,0],"fanout_max":6,"fanout_mean":4.423828125,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.12792494893074036,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":9060},"payload_copies_per_rank":[1144,1088,1148,1126,1135,1149,1140,1130],"payload_rank_cv":0.016329573467373848,"routed_copies":9060},"token_rate_at_latency_percentile":{"p50":4864226.852506274,"p90":4765827.476286499,"p95":4716579.055086041,"p99":4636158.733514846},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":250.37699937820435,"p90":254.74199652671814,"p95":256.5859854221344,"p99":259.99099016189575},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":349.1249978542328,"p90":354.73400354385376,"p95":356.9769859313965,"p99":362.2249960899353},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":599.5019972324371,"p90":609.4760000705719,"p95":613.5629713535309,"p99":622.215986251831},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":592.9319858551025,"p90":597.137987613678,"p95":599.2220044136047,"p99":602.0259857177734},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":591.5679931640625,"p90":597.7389812469482,"p95":599.9420285224915,"p99":606.5919995307922},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":240.60100317001343,"p90":242.48400330543518,"p95":243.32599341869354,"p99":244.52799558639526},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":338.2290005683899,"p90":341.9150114059448,"p95":343.39699149131775,"p99":346.5209901332855},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.128966808319092,"p90":6.210029125213623,"p95":6.210029125213623,"p99":6.210029125213623},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.168961524963379,"p90":10.295987129211426,"p95":11.336982250213623,"p99":14.42098617553711},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.042008399963379,"p90":1.562952995300293,"p95":1.562952995300293,"p99":1.562952995300293},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":241.08299612998962,"p90":243.52599680423737,"p95":245.12800574302673,"p99":247.77300655841827},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":333.58100056648254,"p90":339.34998512268066,"p95":342.2349989414215,"p99":347.68301248550415},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":581.475019454956,"p90":587.0029926300049,"p95":589.5259976387024,"p99":596.9380140304565},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":9.893953800201416,"p90":12.858986854553223,"p95":14.742016792297363,"p99":18.747985363006592},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"roundtrip":{"activation_data_bytes":520626176,"scale_bytes":0,"total_logical_bytes":520626176},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"roundtrip":{"activation_data_bytes":520626176,"scale_bytes":0,"total_logical_bytes":520626176},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":18158,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":2294,"mean":2269.75,"min":2209,"total":18158},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.016587993130087852,"expert_assignments_per_rank":[3123,2960,3095,3022,3074,3092,3108,3102],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,23,454,1699,1566,354,0,0],"fanout_max":6,"fanout_mean":4.43310546875,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.1253442019224167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":18158},"payload_copies_per_rank":[2289,2209,2294,2272,2259,2292,2267,2276],"payload_rank_cv":0.01135234534740448,"routed_copies":18158},"token_rate_at_latency_percentile":{"p50":6923971.626815239,"p90":6852489.344856347,"p95":6827326.316990049,"p99":6752479.431262391},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":6,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":0,"num_qps_per_rank":1,"is_intranode":true,"cpu_proxy_threads":0},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi325x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9943,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2059,"p50":2096,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":232.0},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2059,"p50":2100,"max":2101,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":232.0},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2058,"p50":2097,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":232.0},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2074,"p50":2104,"max":2105,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":231.9},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2073,"p50":2103,"max":2104,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":232.0},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2060,"p50":2097,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":231.9},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2059,"p50":2092,"max":2093,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":232.0},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2065,"p50":2098,"max":2099,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":232.0}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-fp8.json b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-fp8.json new file mode 100644 index 0000000000..028a94ab8f --- /dev/null +++ b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T06:57:28.262947+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1 2 4 8 16 32 64 128 256 512","mode":"normal","nodes":1,"phase":"decode","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-uccl-ep-deepseek-v4-pro-normal-decode-ep8-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1,2,4,8,16,32,64,128,256,512],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":72.42900133132935,"p90":76.99599862098694,"p95":78.79900187253952,"p99":81.48299902677536},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":173.34100604057312,"p90":183.07499587535858,"p95":187.76200711727142,"p99":195.97500562667847},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":245.77000737190247,"p90":260.0709944963455,"p95":266.56100898981094,"p99":277.4580046534538},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":305.0589859485626,"p90":313.83201479911804,"p95":315.5539929866791,"p99":321.40299677848816},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":248.65399301052094,"p90":258.46898555755615,"p95":264.2369866371155,"p99":276.095986366272},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.626000568270683,"p90":19.189000129699707,"p95":20.549999549984932,"p99":22.794000804424286},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":84.32699739933014,"p90":86.40900254249573,"p95":87.37099915742874,"p99":90.4960036277771},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":217.8879976272583,"p90":225.13900697231293,"p95":227.34199464321136,"p99":250.7370114326477},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089001893997192,"p90":6.168991327285767,"p95":6.168991327285767,"p99":6.168991327285767},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.4050047397613525,"p90":7.450997829437256,"p95":8.412986993789673,"p99":11.015981435775757},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-0.8410215377807617,"p90":2.804011106491089,"p95":2.804011106491089,"p99":2.804011106491089},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":65.09800255298615,"p90":67.66100227832794,"p95":69.14400309324265,"p99":71.5470016002655},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":159.11899507045746,"p90":166.44999384880066,"p95":169.4139987230301,"p99":175.26400089263916},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":240.84199965000153,"p90":250.41699409484863,"p95":256.14601373672485,"p99":267.9229974746704},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.2109997272491455,"p90":10.496005415916443,"p95":12.29800283908844,"p99":14.942005276679993},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00146484375,"host_stalled_samples":{"dispatch":3,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8,"byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":250880,"scale_bytes":7840,"total_logical_bytes":258720},"roundtrip":{"activation_data_bytes":752640,"scale_bytes":7840,"total_logical_bytes":760480},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":501760,"scale_bytes":0,"total_logical_bytes":501760},"dispatch":{"activation_data_bytes":250880,"scale_bytes":7840,"total_logical_bytes":258720},"roundtrip":{"activation_data_bytes":752640,"scale_bytes":7840,"total_logical_bytes":760480},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":35,"assignments":48,"wire":"rank-deduplicated"},"receive":{"max":6,"mean":4.375,"min":2,"total":35},"routing":{"empty_expert_count":337,"empty_rank_count":0,"expert_assignment_rank_cv":0.3227486312389374,"expert_assignments_per_rank":[8,6,6,6,6,5,9,2],"expert_load_cv":2.708012819290161,"expert_load_max":2,"expert_load_mean":0.125,"expert_load_min":0,"fanout_histogram":[0,0,0,5,3,0,0,0],"fanout_max":5,"fanout_mean":4.375,"fanout_min":4,"hotspot_ratio":16.0,"locality":{"placement":"packed","local_rank_fraction":0.08571428805589676,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":35},"payload_copies_per_rank":[5,5,5,4,4,4,6,2],"payload_rank_cv":0.2539483904838562,"routed_copies":35},"token_rate_at_latency_percentile":{"p50":32173.221524182434,"p90":30951.489141889917,"p95":30275.852377118717,"p99":28975.430267164815},"tokens_per_rank":1},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":75.0729963183403,"p90":79.76000010967255,"p95":81.52300119400024,"p99":84.48600023984909},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":173.33999276161194,"p90":181.0320019721985,"p95":187.6019984483719,"p99":197.09700345993042},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":248.41298907995224,"p90":260.79200208187103,"p95":269.12499964237213,"p99":281.5830036997795},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":306.100994348526,"p90":319.5599913597107,"p95":325.5699872970581,"p99":341.11300110816956},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":251.3369917869568,"p90":261.79400086402893,"p95":266.88098907470703,"p99":276.61600708961487},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":17.746999859809875,"p90":19.62899975478649,"p95":20.95099911093712,"p99":22.83399924635887},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":85.84900200366974,"p90":90.49700200557709,"p95":91.498002409935,"p99":95.94400227069855},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":212.5599980354309,"p90":221.9330072402954,"p95":225.2189964056015,"p99":240.16200006008148},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.209015846252441,"p95":6.209015846252441,"p99":6.209015846252441},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.5659968852996826,"p90":8.412003517150879,"p95":9.654998779296875,"p99":14.421999454498291},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":0.8009970188140869,"p90":2.6040077209472656,"p95":2.6040077209472656,"p99":2.6040077209472656},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":66.17899984121323,"p90":69.66499984264374,"p95":71.42700254917145,"p99":74.43200051784515},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":159.35899317264557,"p90":165.88899493217468,"p95":169.4139987230301,"p99":175.9449988603592},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":241.9240027666092,"p90":251.29801034927368,"p95":257.1069896221161,"p99":266.00000262260437},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.97301733493805,"p90":11.537998914718628,"p95":13.139992952346802,"p99":16.105011105537415},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16,"byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":501760,"scale_bytes":15680,"total_logical_bytes":517440},"roundtrip":{"activation_data_bytes":1505280,"scale_bytes":15680,"total_logical_bytes":1520960},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1003520,"scale_bytes":0,"total_logical_bytes":1003520},"dispatch":{"activation_data_bytes":501760,"scale_bytes":15680,"total_logical_bytes":517440},"roundtrip":{"activation_data_bytes":1505280,"scale_bytes":15680,"total_logical_bytes":1520960},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":70,"assignments":96,"wire":"rank-deduplicated"},"receive":{"max":12,"mean":8.75,"min":5,"total":70},"routing":{"empty_expert_count":300,"empty_rank_count":0,"expert_assignment_rank_cv":0.33592739701271057,"expert_assignments_per_rank":[18,17,13,7,13,7,13,8],"expert_load_cv":2.0,"expert_load_max":2,"expert_load_mean":0.25,"expert_load_min":0,"fanout_histogram":[0,0,3,6,5,2,0,0],"fanout_max":6,"fanout_mean":4.375,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.15714286267757416,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":70},"payload_copies_per_rank":[12,11,10,5,9,6,10,7],"payload_rank_cv":0.26649653911590576,"routed_copies":70},"token_rate_at_latency_percentile":{"p50":63659.55081360342,"p90":61116.75572088495,"p95":59951.81618395898,"p99":57841.91655552495},"tokens_per_rank":2},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":76.79499685764313,"p90":81.80300146341324,"p95":83.20499956607819,"p99":86.60999685525894},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":173.26100170612335,"p90":181.63399398326874,"p95":187.16099858283997,"p99":196.8970000743866},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":250.05599856376648,"p90":263.436995446682,"p95":270.36599814891815,"p99":283.50699692964554},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":305.5790066719055,"p90":315.87499380111694,"p95":319.60099935531616,"p99":451.55900716781616},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":253.0210018157959,"p90":263.3959949016571,"p95":268.7250077724457,"p99":278.6189913749695},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":18.46799999475479,"p90":19.910000264644623,"p95":21.15200087428093,"p99":23.034999147057533},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":85.72900295257568,"p90":88.65399658679962,"p95":90.53599834442139,"p99":93.7810018658638},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":212.35999464988708,"p90":219.29000318050385,"p95":222.25500643253326,"p99":226.33999586105347},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1289966106414795,"p90":6.168991327285767,"p95":6.168991327285767,"p99":6.168991327285767},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.524988889694214,"p90":8.33401083946228,"p95":9.89598035812378,"p99":15.141993761062622},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.5220046043395996,"p90":3.805011510848999,"p95":3.805011510848999,"p99":3.805011510848999},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":69.7849988937378,"p90":72.06899672746658,"p95":73.63100349903107,"p99":75.8339986205101},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":159.84000265598297,"p90":166.89099371433258,"p95":170.6569939851761,"p99":178.62899601459503},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":245.4490065574646,"p90":255.5440068244934,"p95":260.8720064163208,"p99":270.8069980144501},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.0099979639053345,"p90":10.095998644828796,"p95":12.057989835739136,"p99":16.785994172096252},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32,"byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":1025024,"scale_bytes":32032,"total_logical_bytes":1057056},"roundtrip":{"activation_data_bytes":3075072,"scale_bytes":32032,"total_logical_bytes":3107104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2050048,"scale_bytes":0,"total_logical_bytes":2050048},"dispatch":{"activation_data_bytes":1025024,"scale_bytes":32032,"total_logical_bytes":1057056},"roundtrip":{"activation_data_bytes":3075072,"scale_bytes":32032,"total_logical_bytes":3107104},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":143,"assignments":192,"wire":"rank-deduplicated"},"receive":{"max":22,"mean":17.875,"min":13,"total":143},"routing":{"empty_expert_count":229,"empty_rank_count":0,"expert_assignment_rank_cv":0.20518453419208527,"expert_assignments_per_rank":[31,30,24,17,22,22,28,18],"expert_load_cv":1.3768926858901978,"expert_load_max":4,"expert_load_mean":0.5,"expert_load_min":0,"fanout_histogram":[0,0,4,12,13,3,0,0],"fanout_max":6,"fanout_mean":4.46875,"fanout_min":3,"hotspot_ratio":8.0,"locality":{"placement":"packed","local_rank_fraction":0.1258741319179535,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":143},"payload_copies_per_rank":[22,21,18,13,16,15,22,16],"payload_rank_cv":0.1800614446401596,"routed_copies":143},"token_rate_at_latency_percentile":{"p50":126471.7148788171,"p90":121490.07813101976,"p95":119080.8412855173,"p99":114852.18520848759},"tokens_per_rank":4},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":77.71699875593185,"p90":82.48399943113327,"p95":84.40700173377991,"p99":87.2109979391098},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":173.4199970960617,"p90":181.71299993991852,"p95":186.72099709510803,"p99":197.53700494766235},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":251.13699585199356,"p90":264.1969993710518,"p95":271.12799882888794,"p99":284.74800288677216},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":306.3009977340698,"p90":314.6730065345764,"p95":320.0809955596924,"p99":357.09699988365173},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":254.14198637008667,"p90":265.6390070915222,"p95":271.04800939559937,"p99":282.46501088142395},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":19.189000129699707,"p90":20.951999351382256,"p95":22.51400053501129,"p99":24.11700040102005},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":87.25199848413467,"p90":89.85500037670135,"p95":91.29799902439117,"p99":94.94200348854065},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":212.5989943742752,"p90":219.650000333786,"p95":221.77299857139587,"p99":226.14000737667084},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.130009889602661,"p90":6.2499940395355225,"p95":6.2499940395355225,"p99":6.2499940395355225},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.2849907875061035,"p90":7.450997829437256,"p95":8.853018283843994,"p99":12.138992547988892},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.2820065021514893,"p90":2.7649998664855957,"p95":2.7649998664855957,"p99":2.7649998664855957},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":70.82699984312057,"p90":73.50999861955643,"p95":74.7930034995079,"p99":77.31600105762482},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":160.16100347042084,"p90":167.49200224876404,"p95":171.73799872398376,"p99":180.63199520111084},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":246.85099720954895,"p90":257.94801115989685,"p95":263.79698514938354,"p99":274.93399381637573},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.929993629455566,"p90":9.53499972820282,"p95":11.537015438079834,"p99":16.023993492126465},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":64,"byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":2014208,"scale_bytes":62944,"total_logical_bytes":2077152},"roundtrip":{"activation_data_bytes":6042624,"scale_bytes":62944,"total_logical_bytes":6105568},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4028416,"scale_bytes":0,"total_logical_bytes":4028416},"dispatch":{"activation_data_bytes":2014208,"scale_bytes":62944,"total_logical_bytes":2077152},"roundtrip":{"activation_data_bytes":6042624,"scale_bytes":62944,"total_logical_bytes":6105568},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":281,"assignments":384,"wire":"rank-deduplicated"},"receive":{"max":42,"mean":35.125,"min":26,"total":281},"routing":{"empty_expert_count":143,"empty_rank_count":0,"expert_assignment_rank_cv":0.1458333283662796,"expert_assignments_per_rank":[55,56,58,37,46,43,46,43],"expert_load_cv":0.9816015958786011,"expert_load_max":5,"expert_load_mean":1.0,"expert_load_min":0,"fanout_histogram":[0,1,9,24,24,6,0,0],"fanout_max":6,"fanout_mean":4.390625,"fanout_min":2,"hotspot_ratio":5.0,"locality":{"placement":"packed","local_rank_fraction":0.11743772029876709,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":281},"payload_copies_per_rank":[38,42,42,26,29,31,38,35],"payload_rank_cv":0.15879225730895996,"routed_copies":281},"token_rate_at_latency_percentile":{"p50":251827.73186797206,"p90":240928.47169071707,"p95":236120.53135055816,"p99":226576.73529294776},"tokens_per_rank":8},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":81.44199848175049,"p90":85.92899888753891,"p95":87.33099699020386,"p99":89.57400172948837},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":173.78099262714386,"p90":183.19499492645264,"p95":189.04399871826172,"p99":197.09700345993042},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":255.22299110889435,"p90":269.12399381399155,"p95":276.3749957084656,"p99":286.6710051894188},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":301.3330101966858,"p90":311.54900789260864,"p95":316.4759874343872,"p99":330.45700192451477},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":258.2690119743347,"p90":268.7639892101288,"p95":273.5710144042969,"p99":285.22899746894836},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":23.354999721050262,"p90":24.43699911236763,"p95":25.838999077677727,"p99":27.600999921560287},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":93.18000078201294,"p90":94.70199793577194,"p95":96.34499996900558,"p99":101.27200186252594},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":208.11299979686737,"p90":217.32600033283234,"p95":220.33099830150604,"p99":227.5020033121109},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.248980760574341,"p90":6.448984146118164,"p95":6.448984146118164,"p99":6.448984146118164},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.3650100231170654,"p90":8.69297981262207,"p95":10.055005550384521,"p99":14.02100920677185},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.0419955253601074,"p90":1.4019906520843506,"p95":1.4019906520843506,"p99":1.4019906520843506},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":76.03500038385391,"p90":77.91700214147568,"p95":78.95799726247787,"p99":81.40300214290619},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":162.7649962902069,"p90":170.73699831962585,"p95":175.6639927625656,"p99":183.07499587535858},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":252.09900736808777,"p90":262.1540129184723,"p95":267.6829993724823,"p99":278.8189947605133},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.528002977371216,"p90":9.133011102676392,"p95":10.816007852554321,"p99":14.101982116699219},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":128,"byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":4042752,"scale_bytes":126336,"total_logical_bytes":4169088},"roundtrip":{"activation_data_bytes":12128256,"scale_bytes":126336,"total_logical_bytes":12254592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":8085504,"scale_bytes":0,"total_logical_bytes":8085504},"dispatch":{"activation_data_bytes":4042752,"scale_bytes":126336,"total_logical_bytes":4169088},"roundtrip":{"activation_data_bytes":12128256,"scale_bytes":126336,"total_logical_bytes":12254592},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":564,"assignments":768,"wire":"rank-deduplicated"},"receive":{"max":77,"mean":70.5,"min":62,"total":564},"routing":{"empty_expert_count":56,"empty_rank_count":0,"expert_assignment_rank_cv":0.07384087145328522,"expert_assignments_per_rank":[101,108,101,91,83,97,94,93],"expert_load_cv":0.6978389024734497,"expert_load_max":7,"expert_load_mean":2.0,"expert_load_min":0,"fanout_histogram":[0,2,14,52,50,10,0,0],"fanout_max":6,"fanout_mean":4.40625,"fanout_min":2,"hotspot_ratio":3.5,"locality":{"placement":"packed","local_rank_fraction":0.14716312289237976,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":564},"payload_copies_per_rank":[73,76,77,63,62,69,74,70],"payload_rank_cv":0.07472094893455505,"routed_copies":564},"token_rate_at_latency_percentile":{"p50":495607.2701928325,"p90":476254.27936301864,"p95":467885.82583838806,"p99":448762.2266173508},"tokens_per_rank":16},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":90.41599929332733,"p90":95.10300308465958,"p95":96.62500023841858,"p99":99.50999915599823},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":173.78099262714386,"p90":183.87700617313385,"p95":188.64400684833527,"p99":197.81699776649475},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":264.1969919204712,"p90":278.9800092577934,"p95":285.26900708675385,"p99":297.326996922493},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":301.8139898777008,"p90":312.2299909591675,"p95":326.21100544929504,"p99":428.9650022983551},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":267.60199666023254,"p90":276.977002620697,"p95":282.3449969291687,"p99":291.07800126075745},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":26.000000536441803,"p90":27.922000735998154,"p95":28.80300022661686,"p99":31.286999583244324},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":103.3950001001358,"p90":105.6390032172203,"p95":107.76200145483017,"p99":111.4870011806488},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":204.9880027770996,"p90":215.48399329185486,"p95":219.29000318050385,"p99":237.55699396133423},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.169021129608154,"p90":6.170004606246948,"p95":6.170004606246948,"p99":6.170004606246948},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.446983337402344,"p90":10.495990514755249,"p95":13.420015573501587,"p99":18.869012594223022},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.4829933643341064,"p90":2.163022756576538,"p95":2.163022756576538,"p99":2.163022756576538},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":85.9690010547638,"p90":87.53199875354767,"p95":88.45300227403641,"p99":90.77700227499008},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":168.77399384975433,"p90":176.9859939813614,"p95":183.15599858760834,"p99":190.24600088596344},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":262.51500844955444,"p90":271.08800411224365,"p95":277.17700600624084,"p99":285.6700122356415},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.527002573013306,"p90":8.293002843856812,"p95":9.975999593734741,"p99":13.26000690460205},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":256,"byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":8157184,"scale_bytes":254912,"total_logical_bytes":8412096},"roundtrip":{"activation_data_bytes":24471552,"scale_bytes":254912,"total_logical_bytes":24726464},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":16314368,"scale_bytes":0,"total_logical_bytes":16314368},"dispatch":{"activation_data_bytes":8157184,"scale_bytes":254912,"total_logical_bytes":8412096},"roundtrip":{"activation_data_bytes":24471552,"scale_bytes":254912,"total_logical_bytes":24726464},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":1138,"assignments":1536,"wire":"rank-deduplicated"},"receive":{"max":149,"mean":142.25,"min":126,"total":1138},"routing":{"empty_expert_count":10,"empty_rank_count":0,"expert_assignment_rank_cv":0.04658475145697594,"expert_assignments_per_rank":[189,194,196,170,193,201,198,195],"expert_load_cv":0.4824390113353729,"expert_load_max":11,"expert_load_mean":4.0,"expert_load_min":0,"fanout_histogram":[0,3,26,101,106,20,0,0],"fanout_max":6,"fanout_mean":4.4453125,"fanout_min":2,"hotspot_ratio":2.75,"locality":{"placement":"packed","local_rank_fraction":0.1344463974237442,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":1138},"payload_copies_per_rank":[143,141,149,126,144,145,148,142],"payload_rank_cv":0.04686218500137329,"routed_copies":1138},"token_rate_at_latency_percentile":{"p50":956644.5811128857,"p90":924264.4608678082,"p95":906692.1772452097,"p99":879489.3426888231},"tokens_per_rank":32},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":104.27699983119965,"p90":108.96400362253189,"p95":110.80700159072876,"p99":114.05199766159058},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":184.5169961452484,"p90":194.73299384117126,"p95":199.57999885082245,"p99":206.7510038614273},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":288.79399597644806,"p90":303.69699746370316,"p95":310.3870004415512,"p99":320.8030015230179},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":305.5799901485443,"p90":318.63901019096375,"p95":332.29899406433105,"p99":422.9159951210022},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":281.904011964798,"p90":290.677011013031,"p95":295.2440083026886,"p99":305.8199882507324},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":35.1720005273819,"p90":39.980001747608185,"p95":40.94099998474121,"p99":42.06300154328346},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":115.13300240039825,"p90":117.85700172185898,"p95":118.81899833679199,"p99":122.3440021276474},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":201.34299993515015,"p90":208.2729935646057,"p95":209.51500535011292,"p99":215.44399857521057},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.1280131340026855,"p90":6.208986043930054,"p95":6.208986043930054,"p99":6.208986043930054},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.767000675201416,"p90":9.894996881484985,"p95":14.102011919021606,"p99":18.38800311088562},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.086000680923462,"p90":4.446983337402344,"p95":4.446983337402344,"p99":4.446983337402344},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":96.42499685287476,"p90":98.98900240659714,"p95":100.27100145816803,"p99":102.71500051021576},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":178.26800048351288,"p90":187.76200711727142,"p95":193.0910050868988,"p99":200.02099871635437},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":273.6920118331909,"p90":281.6230058670044,"p95":286.79201006889343,"p99":297.325998544693},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.971018552780151,"p90":10.576009750366211,"p95":12.13797926902771,"p99":14.862984418869019},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.00413223123177886,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":512,"byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":16142336,"scale_bytes":504448,"total_logical_bytes":16646784},"roundtrip":{"activation_data_bytes":48427008,"scale_bytes":504448,"total_logical_bytes":48931456},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":32284672,"scale_bytes":0,"total_logical_bytes":32284672},"dispatch":{"activation_data_bytes":16142336,"scale_bytes":504448,"total_logical_bytes":16646784},"roundtrip":{"activation_data_bytes":48427008,"scale_bytes":504448,"total_logical_bytes":48931456},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":2252,"assignments":3072,"wire":"rank-deduplicated"},"receive":{"max":290,"mean":281.5,"min":268,"total":2252},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.03184119611978531,"expert_assignments_per_rank":[396,393,399,359,380,389,380,376],"expert_load_cv":0.3360888659954071,"expert_load_max":16,"expert_load_mean":8.0,"expert_load_min":1,"fanout_histogram":[0,3,57,223,191,38,0,0],"fanout_max":6,"fanout_mean":4.3984375,"fanout_min":2,"hotspot_ratio":2.0,"locality":{"placement":"packed","local_rank_fraction":0.11767318099737167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":2252},"payload_copies_per_rank":[290,286,290,268,282,282,280,274],"payload_rank_cv":0.025306938216090202,"routed_copies":2252},"token_rate_at_latency_percentile":{"p50":1816221.047836434,"p90":1761405.2044076065,"p95":1734158.816442737,"p99":1674187.4948351216},"tokens_per_rank":64},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":122.10399657487869,"p90":126.75100564956665,"p95":128.1919926404953,"p99":131.1980038881302},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":198.73900711536407,"p90":208.87400209903717,"p95":213.0800038576126,"p99":220.4120010137558},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":320.84300369024277,"p90":335.6250077486038,"p95":341.2719964981079,"p99":351.610004901886},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":310.1859986782074,"p90":321.0029900074005,"p95":327.69298553466797,"p99":344.5979952812195},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":314.2319917678833,"p90":323.8070011138916,"p95":328.4139931201935,"p99":336.78698539733887},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":43.70599985122681,"p90":45.82900181412697,"p95":46.67000100016594,"p99":48.2730008661747},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":123.1049969792366,"p90":126.79000198841095,"p95":129.1940063238144,"p99":134.68299806118011},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":195.77400386333466,"p90":201.1830061674118,"p95":204.42700386047363,"p99":213.20000290870667},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.089985370635986,"p90":6.209015846252441,"p95":6.209015846252441,"p99":6.209015846252441},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4.005998373031616,"p90":7.611989974975586,"p95":9.013980627059937,"p99":11.417984962463379},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.122981309890747,"p90":2.4830102920532227,"p95":2.4830102920532227,"p99":2.4830102920532227},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":115.89399725198746,"p90":117.8169995546341,"p95":119.25899982452393,"p99":121.3029995560646},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":193.01000237464905,"p90":202.224001288414,"p95":207.07100629806519,"p99":213.28100562095642},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":307.22200870513916,"p90":316.756010055542,"p95":321.5230107307434,"p99":329.49599623680115},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.328999996185303,"p90":9.775996208190918,"p95":11.658012866973877,"p99":15.662014484405518},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0009765625,"host_stalled_samples":{"dispatch":1,"roundtrip":2},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.006756756920367479,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":1024,"byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":32162816,"scale_bytes":1005088,"total_logical_bytes":33167904},"roundtrip":{"activation_data_bytes":96488448,"scale_bytes":1005088,"total_logical_bytes":97493536},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":64325632,"scale_bytes":0,"total_logical_bytes":64325632},"dispatch":{"activation_data_bytes":32162816,"scale_bytes":1005088,"total_logical_bytes":33167904},"roundtrip":{"activation_data_bytes":96488448,"scale_bytes":1005088,"total_logical_bytes":97493536},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":4487,"assignments":6144,"wire":"rank-deduplicated"},"receive":{"max":569,"mean":560.875,"min":553,"total":4487},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.01615842990577221,"expert_assignments_per_rank":[775,756,779,747,762,761,782,782],"expert_load_cv":0.2481621950864792,"expert_load_max":30,"expert_load_mean":16.0,"expert_load_min":4,"fanout_histogram":[0,5,126,437,385,71,0,0],"fanout_max":6,"fanout_mean":4.3818359375,"fanout_min":2,"hotspot_ratio":1.875,"locality":{"placement":"packed","local_rank_fraction":0.12859371304512024,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":4487},"payload_copies_per_rank":[569,555,565,556,553,564,558,567],"payload_rank_cv":0.010142241604626179,"routed_copies":4487},"token_rate_at_latency_percentile":{"p50":3258738.851632929,"p90":3162377.5782409096,"p95":3118015.740654616,"p99":3040497.5382047263},"tokens_per_rank":128},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":171.49800062179565,"p90":176.18300020694733,"p95":178.14800143241882,"p99":180.7119995355606},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":235.0739985704422,"p90":241.4029985666275,"p95":243.9669966697693,"p99":251.13800168037415},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":406.57199919223785,"p90":417.58599877357483,"p95":422.1149981021881,"p99":431.85000121593475},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":394.5539891719818,"p90":400.6020128726959,"p95":404.7290086746216,"p99":413.22198510169983},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":398.27901124954224,"p90":405.80999851226807,"p95":408.5739850997925,"p99":415.3839945793152},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":62.61499971151352,"p90":64.81699645519257,"p95":65.4980018734932,"p99":67.38100200891495},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":161.5229994058609,"p90":165.96999764442444,"p95":167.8919941186905,"p99":169.854998588562},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":229.46499288082123,"p90":232.91000723838806,"p95":234.15200412273407,"p99":237.31699585914612},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.130009889602661,"p90":6.208986043930054,"p95":6.208986043930054,"p99":6.208986043930054},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5.20700216293335,"p90":8.451998233795166,"p95":9.734004735946655,"p99":12.659013271331787},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.203017473220825,"p90":-1.3220012187957764,"p95":-1.3220012187957764,"p99":-1.3220012187957764},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":161.44299507141113,"p90":164.40699994564056,"p95":165.92900454998016,"p99":169.05400156974792},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":227.86299884319305,"p90":233.5910052061081,"p95":236.51599884033203,"p99":244.00700628757477},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":387.62301206588745,"p90":395.03398537635803,"p95":397.9189991950989,"p99":404.6890139579773},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.255992412567139,"p90":13.29997181892395,"p95":14.78201150894165,"p99":17.987996339797974},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0009765625,"host_stalled_samples":{"dispatch":0,"roundtrip":2},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.006756756920367479,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":2048,"byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":64942080,"scale_bytes":2029440,"total_logical_bytes":66971520},"roundtrip":{"activation_data_bytes":194826240,"scale_bytes":2029440,"total_logical_bytes":196855680},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":129884160,"scale_bytes":0,"total_logical_bytes":129884160},"dispatch":{"activation_data_bytes":64942080,"scale_bytes":2029440,"total_logical_bytes":66971520},"roundtrip":{"activation_data_bytes":194826240,"scale_bytes":2029440,"total_logical_bytes":196855680},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":9060,"assignments":12288,"wire":"rank-deduplicated"},"receive":{"max":1149,"mean":1132.5,"min":1088,"total":9060},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.018761977553367615,"expert_assignments_per_rank":[1542,1490,1556,1492,1545,1541,1581,1541],"expert_load_cv":0.1781381368637085,"expert_load_max":52,"expert_load_mean":32.0,"expert_load_min":18,"fanout_histogram":[0,11,225,859,791,162,0,0],"fanout_max":6,"fanout_mean":4.423828125,"fanout_min":2,"hotspot_ratio":1.625,"locality":{"placement":"packed","local_rank_fraction":0.12792494893074036,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":9060},"payload_copies_per_rank":[1144,1088,1148,1126,1135,1149,1140,1130],"payload_rank_cv":0.016329573467373848,"routed_copies":9060},"token_rate_at_latency_percentile":{"p50":5142123.843219101,"p90":5046696.74849839,"p95":5012556.047834971,"p99":4930377.738974115},"tokens_per_rank":256},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":250.0160038471222,"p90":254.58300113677979,"p95":256.66600465774536,"p99":260.43200492858887},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":303.97701263427734,"p90":310.06601452827454,"p95":312.7110004425049,"p99":319.35998797416687},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":553.9930164813995,"p90":564.6490156650543,"p95":569.3770051002502,"p99":579.7919929027557},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":538.9699935913086,"p90":545.5800294876099,"p95":551.6690015792847,"p99":565.2099847793579},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":542.2559976577759,"p90":548.425018787384,"p95":551.0690212249756,"p99":558.9609742164612},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":100.95199942588806,"p90":103.3559963107109,"p95":104.3969988822937,"p99":106.20000213384628},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":242.04400181770325,"p90":244.6880042552948,"p95":245.60999870300293,"p99":248.61399829387665},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":290.4770076274872,"p90":294.5629954338074,"p95":296.2450087070465,"p99":300.8129894733429},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.130039691925049,"p90":6.209969520568848,"p95":6.209969520568848,"p99":6.209969520568848},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.289958953857422,"p90":9.93502140045166,"p95":10.775983333587646,"p99":13.019979000091553},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.0409951210021973,"p90":0.841975212097168,"p95":0.841975212097168,"p99":0.841975212097168},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":241.4029985666275,"p90":243.80700290203094,"p95":245.52899599075317,"p99":248.57400357723236},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":291.31901264190674,"p90":296.84698581695557,"p95":300.01100897789,"p99":306.05998635292053},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":533.1220030784607,"p90":538.7300252914429,"p95":541.854977607727,"p99":550.025999546051},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.814036846160889,"p90":12.057960033416748,"p95":13.339996337890625,"p99":16.18504524230957},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.006756756920367479,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":4096,"byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":130156544,"scale_bytes":4067392,"total_logical_bytes":134223936},"roundtrip":{"activation_data_bytes":390469632,"scale_bytes":4067392,"total_logical_bytes":394537024},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":260313088,"scale_bytes":0,"total_logical_bytes":260313088},"dispatch":{"activation_data_bytes":130156544,"scale_bytes":4067392,"total_logical_bytes":134223936},"roundtrip":{"activation_data_bytes":390469632,"scale_bytes":4067392,"total_logical_bytes":394537024},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":18158,"assignments":24576,"wire":"rank-deduplicated"},"receive":{"max":2294,"mean":2269.75,"min":2209,"total":18158},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.016587993130087852,"expert_assignments_per_rank":[3123,2960,3095,3022,3074,3092,3108,3102],"expert_load_cv":0.13393747806549072,"expert_load_max":88,"expert_load_mean":64.0,"expert_load_min":38,"fanout_histogram":[0,23,454,1699,1566,354,0,0],"fanout_max":6,"fanout_mean":4.43310546875,"fanout_min":2,"hotspot_ratio":1.375,"locality":{"placement":"packed","local_rank_fraction":0.1253442019224167,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":18158},"payload_copies_per_rank":[2289,2209,2294,2272,2259,2292,2267,2276],"payload_rank_cv":0.01135234534740448,"routed_copies":18158},"token_rate_at_latency_percentile":{"p50":7553627.839419553,"p90":7468659.998511039,"p95":7432825.730060038,"p99":7327881.889682335},"tokens_per_rank":512}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":6,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":0,"num_qps_per_rank":1,"is_intranode":true,"cpu_proxy_threads":0},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi325x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9943,"throttle_fraction_max":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2061,"p50":2095,"max":2097,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":56,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":323.2},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2062,"p50":2100,"max":2102,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":323.2},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2063,"p50":2096,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":323.2},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2073,"p50":2103,"max":2105,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":55,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":323.2},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2073,"p50":2102,"max":2103,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":323.2},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2062,"p50":2097,"max":2098,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":323.2},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2052,"p50":2091,"max":2093,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":323.2},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2063,"p50":2097,"max":2100,"samples":2600},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":323.2}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16.json b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16.json new file mode 100644 index 0000000000..a49a9192c5 --- /dev/null +++ b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T11:37:56.623623+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":2,"phase":"prefill","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":755.6170225143433,"p90":761.0260248184204,"p95":763.5089755058289,"p99":767.6349878311157},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1299.921989440918,"p90":1312.9420280456543,"p95":1317.1080350875854,"p99":1332.4910402297974},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":2055.5390119552612,"p90":2073.9680528640747,"p95":2080.6170105934143,"p99":2100.126028060913},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":2151.918888092041,"p90":2167.4230098724365,"p95":2171.7569828033447,"p99":2194.993019104004},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2056.2150478363037,"p90":2073.4410285949707,"p95":2078.528881072998,"p99":2269.6969509124756},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":796.6430187225342,"p90":811.7049932479858,"p95":814.9499893188477,"p99":819.9970126152039},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1271.994948387146,"p90":1282.4499607086182,"p95":1285.4549884796143,"p99":1291.0629510879517},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.328821182250977,"p90":6.369829177856445,"p95":6.369829177856445,"p99":6.369829177856445},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":20.26987075805664,"p90":27.753114700317383,"p95":31.126976013183594,"p99":35.70199012756348},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-4.245996475219727,"p90":5.768775939941406,"p95":5.768775939941406,"p99":5.768775939941406},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":708.9899778366089,"p90":713.3960127830505,"p95":715.0390148162842,"p99":718.0439829826355},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1255.6140422821045,"p90":1264.1429901123047,"p95":1266.5469646453857,"p99":1272.0799446105957},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2009.0720653533936,"p90":2025.0170230865479,"p95":2030.263900756836,"p99":2168.632984161377},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":47.023892402648926,"p90":52.31213569641113,"p95":54.07392978668213,"p99":58.721065521240234},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0009765625,"host_stalled_samples":{"dispatch":1,"roundtrip":2},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0095911156386137,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"roundtrip":{"activation_data_bytes":2428403712,"scale_bytes":0,"total_logical_bytes":2428403712},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"roundtrip":{"activation_data_bytes":2428403712,"scale_bytes":0,"total_logical_bytes":2428403712},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":84696,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":5367,"mean":5293.5,"min":5181,"total":84696},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.012361490167677402,"expert_assignments_per_rank":[6171,6137,6243,6050,6268,6146,6227,6063,6176,6175,6160,5993,6212,6144,6110,6029],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,5,241,2597,7671,5870,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16943359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.06316709518432617,"same_node_fraction":0.49927976727485657,"same_scaleup_domain_fraction":0.49927976727485657,"cross_node_fraction":0.500720202922821,"cross_domain_fraction":0.500720202922821,"gpus_per_node":8,"scale_up_domain":8,"copies":84696},"payload_copies_per_rank":[5326,5273,5366,5236,5367,5310,5348,5225,5315,5334,5305,5181,5318,5287,5280,5225],"payload_rank_cv":0.009882226586341858,"routed_copies":84696},"token_rate_at_latency_percentile":{"p50":7968038.176376744,"p90":7901840.358152032,"p95":7882498.121239525,"p99":7218584.839448817},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1244.2330121994019,"p90":1251.8450021743774,"p95":1254.8890113830566,"p99":1482.470989227295},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2201.4338970184326,"p90":2218.859910964966,"p95":2224.7889041900635,"p99":2317.096948623657},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":3445.6669092178345,"p90":3470.7049131393433,"p95":3479.67791557312,"p99":3799.567937850952},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":3536.4580154418945,"p90":3552.480936050415,"p95":3558.0360889434814,"p99":3583.16707611084},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3438.015937805176,"p90":3458.9269161224365,"p95":3527.8708934783936,"p99":3689.7940635681152},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1277.2070169448853,"p90":1288.619041442871,"p95":1292.4710512161255,"p99":1297.9140281677246},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2171.950101852417,"p90":2183.4869384765625,"p95":2186.5720748901367,"p99":2194.183111190796},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.329059600830078,"p90":6.409168243408203,"p95":6.409168243408203,"p99":6.409168243408203},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":32.80997276306152,"p90":47.05691337585449,"p95":51.82504653930664,"p99":60.61196327209473},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-6.6089630126953125,"p90":-4.56690788269043,"p95":-4.56690788269043,"p99":-4.56690788269043},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1183.6659908294678,"p90":1189.555048942566,"p95":1192.6000118255615,"p99":1377.4789571762085},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2144.507884979248,"p90":2153.641939163208,"p95":2156.0449600219727,"p99":2161.5428924560547},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3377.135992050171,"p90":3396.044969558716,"p95":3411.0679626464844,"p99":3590.2979373931885},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":60.6381893157959,"p90":68.88985633850098,"p95":72.41511344909668,"p99":134.2298984527588},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0095911156386137,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"roundtrip":{"activation_data_bytes":4850069504,"scale_bytes":0,"total_logical_bytes":4850069504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"roundtrip":{"activation_data_bytes":4850069504,"scale_bytes":0,"total_logical_bytes":4850069504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":169157,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":10735,"mean":10572.3125,"min":10405,"total":169157},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.008082338608801365,"expert_assignments_per_rank":[12320,12304,12326,12170,12384,12230,12308,12211,12262,12359,12237,12047,12413,12243,12482,12312],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,16,505,5263,15346,11638,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.162261962890625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.0630006417632103,"same_node_fraction":0.5010611414909363,"same_scaleup_domain_fraction":0.5010611414909363,"cross_node_fraction":0.4989388585090637,"cross_domain_fraction":0.4989388585090637,"gpus_per_node":8,"scale_up_domain":8,"copies":169157},"payload_copies_per_rank":[10630,10585,10602,10505,10644,10564,10577,10498,10504,10645,10499,10405,10642,10518,10735,10604],"payload_rank_cv":0.00741199404001236,"routed_copies":169157},"token_rate_at_latency_percentile":{"p50":9531078.561816977,"p90":9473458.328149337,"p95":9288321.763864653,"p99":8880712.428788668},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2294.6949005126953,"p90":2310.837984085083,"p95":2317.1679973602295,"p99":2326.9429206848145},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4045.546054840088,"p90":4085.1659774780273,"p95":4105.035781860352,"p99":4192.567825317383},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":6340.240955352783,"p90":6396.00396156311,"p95":6422.203779220581,"p99":6519.510746002197},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":6430.544853210449,"p90":6460.951805114746,"p95":6468.817234039307,"p99":6490.009784698486},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6329.927921295166,"p90":6381.886959075928,"p95":6458.562850952148,"p99":6585.554122924805},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2277.596950531006,"p90":2294.6629524230957,"p95":2298.708915710449,"p99":2305.6800365448},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":3963.1659984588623,"p90":3986.3998889923096,"p95":3991.447925567627,"p99":3998.698949813843},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.450176239013672,"p90":6.570339202880859,"p95":6.570339202880859,"p99":6.570339202880859},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":51.33199691772461,"p90":74.51200485229492,"p95":83.4202766418457,"p99":103.22093963623047},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-6.649971008300781,"p90":12.33816146850586,"p95":12.33816146850586,"p99":12.33816146850586},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2184.8089694976807,"p90":2193.5110092163086,"p95":2195.995092391968,"p99":2200.5209922790527},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3929.410934448242,"p90":3942.214012145996,"p95":3945.420026779175,"p99":3953.3910751342773},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6203.739166259766,"p90":6251.811981201172,"p95":6309.938907623291,"p99":6420.330047607422},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":125.28419494628906,"p90":143.8584327697754,"p95":151.3080596923828,"p99":218.66512298583984},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"roundtrip":{"activation_data_bytes":9698332672,"scale_bytes":0,"total_logical_bytes":9698332672},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"roundtrip":{"activation_data_bytes":9698332672,"scale_bytes":0,"total_logical_bytes":9698332672},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":338251,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":21435,"mean":21140.6875,"min":20921,"total":338251},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.0070151351392269135,"expert_assignments_per_rank":[24582,24524,24711,24358,24680,24435,24611,24307,24504,24638,24478,24415,25004,24532,24846,24591],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,29,1048,10572,30561,23326,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1613006591796875,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.06308037787675858,"same_node_fraction":0.5009504556655884,"same_scaleup_domain_fraction":0.5009504556655884,"cross_node_fraction":0.49904951453208923,"cross_domain_fraction":0.49904951453208923,"gpus_per_node":8,"scale_up_domain":8,"copies":338251},"payload_copies_per_rank":[21145,21167,21213,20988,21198,21055,21170,20921,21054,21231,21048,21005,21435,21070,21396,21155],"payload_rank_cv":0.00635375315323472,"routed_copies":338251},"token_rate_at_latency_percentile":{"p50":10353356.438629191,"p90":10269063.118831763,"p95":10147149.065885209,"p99":9951478.459779764},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4428.545951843262,"p90":4524.2109298706055,"p95":4624.482154846191,"p99":4682.5690269470215},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7670.347213745117,"p90":7718.297958374023,"p95":7743.45588684082,"p99":7842.405796051025},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":12098.893165588379,"p90":12242.508888244629,"p95":12367.938041687012,"p99":12524.974822998047},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":12152.934074401855,"p90":12192.39330291748,"p95":12203.859329223633,"p99":12229.888916015625},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":12063.799858093262,"p90":12134.547233581543,"p95":12219.515800476074,"p99":12330.321311950684},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":4291.518211364746,"p90":4320.961952209473,"p95":4328.0930519104,"p99":4346.983909606934},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":7539.309024810791,"p90":7578.688144683838,"p95":7589.865207672119,"p99":7603.206157684326},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.688117980957031,"p90":6.8492889404296875,"p95":6.8492889404296875,"p99":6.8492889404296875},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":122.8170394897461,"p90":176.03397369384766,"p95":199.74994659423828,"p99":255.89847564697266},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-6.930351257324219,"p90":16.223907470703125,"p95":16.223907470703125,"p99":16.223907470703125},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4229.103088378906,"p90":4262.577056884766,"p95":4361.142158508301,"p99":4411.056995391846},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7529.973983764648,"p90":7549.084186553955,"p95":7554.85200881958,"p99":7579.329967498779},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":11843.596458435059,"p90":11894.271850585938,"p95":11975.51441192627,"p99":12041.373252868652},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":221.24576568603516,"p90":256.2589645385742,"p95":269.44541931152344,"p99":372.91526794433594},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":131072,"byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"roundtrip":{"activation_data_bytes":19408105472,"scale_bytes":0,"total_logical_bytes":19408105472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"roundtrip":{"activation_data_bytes":19408105472,"scale_bytes":0,"total_logical_bytes":19408105472},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":676901,"assignments":786432,"wire":"rank-deduplicated"},"receive":{"max":42692,"mean":42306.3125,"min":41888,"total":676901},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005902660544961691,"expert_assignments_per_rank":[49152,49158,48937,48637,49532,49147,49291,48640,48855,49080,49131,49180,49736,49110,49535,49311],"expert_load_cv":0.02157296985387802,"expert_load_max":2171,"expert_load_mean":2048.0,"expert_load_min":1922,"fanout_histogram":[0,51,2031,21028,61178,46784,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.164344787597656,"fanout_min":2,"hotspot_ratio":1.06005859375,"locality":{"placement":"packed","local_rank_fraction":0.06220702826976776,"same_node_fraction":0.5000066757202148,"same_scaleup_domain_fraction":0.5000066757202148,"cross_node_fraction":0.49999335408210754,"cross_domain_fraction":0.49999335408210754,"gpus_per_node":8,"scale_up_domain":8,"copies":676901},"payload_copies_per_rank":[42365,42433,42070,41888,42530,42296,42419,41908,42030,42386,42324,42267,42692,42172,42625,42496],"payload_rank_cv":0.005482310429215431,"routed_copies":676901},"token_rate_at_latency_percentile":{"p50":10864901.734263066,"p90":10801556.702277863,"p95":10726447.932977298,"p99":10630055.509823866},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":36,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":20,"num_max_rdma_chunked_recv_tokens":140},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":12,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":1073741824,"num_qps_per_rank":10,"is_intranode":false,"cpu_proxy_threads":4},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9948,"throttle_fraction_max":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2094,"p50":2096,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.9},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2096,"p50":2098,"max":2100,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2096,"p50":2097,"max":2099,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2095,"p50":2097,"max":2098,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2090,"p50":2092,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.9},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2094,"p50":2096,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2094,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2095,"p50":2096,"max":2099,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2097,"p50":2099,"max":2100,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2094,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2090,"p50":2092,"max":2095,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":69,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2087,"p50":2088,"max":2089,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2097,"p50":2098,"max":2100,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2093,"p50":2095,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":68,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2086,"p50":2088,"max":2090,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":67,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2094,"p50":2095,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":814.8}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8.json b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8.json new file mode 100644 index 0000000000..b03b4dc6cd --- /dev/null +++ b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T12:07:02.291729+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":16,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":2,"phase":"prefill","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep16-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":760.0650191307068,"p90":765.513002872467,"p95":767.5549983978271,"p99":771.3599801063538},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1069.409966468811,"p90":1078.9450407028198,"p95":1082.5940370559692,"p99":1141.922950744629},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1829.4749855995178,"p90":1844.4580435752869,"p95":1850.1490354537964,"p99":1913.2829308509827},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1939.8080110549927,"p90":1955.9919834136963,"p95":1960.0379467010498,"p99":1969.8129892349243},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1838.608980178833,"p90":1851.626992225647,"p95":1856.835961341858,"p99":1888.7239694595337},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":210.7580006122589,"p90":213.88299763202667,"p95":214.92399275302887,"p99":217.849001288414},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":819.4770216941833,"p90":828.6899924278259,"p95":831.0139775276184,"p99":837.3039960861206},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1061.0369443893433,"p90":1070.451021194458,"p95":1073.3000040054321,"p99":1086.1949920654297},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.369948387145996,"p90":6.489992141723633,"p95":6.489992141723633,"p99":6.489992141723633},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":20.42996883392334,"p90":27.513980865478516,"p95":29.572010040283203,"p99":35.566091537475586},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-4.006028175354004,"p90":3.283977508544922,"p95":3.283977508544922,"p99":3.283977508544922},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":714.3980264663696,"p90":718.7250256538391,"p95":720.4470038414001,"p99":724.1730093955994},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1038.6430025100708,"p90":1045.8990335464478,"p95":1048.4189987182617,"p99":1055.7490587234497},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1792.70601272583,"p90":1805.4449558258057,"p95":1809.6920251846313,"p99":1827.3169994354248},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":45.42195796966553,"p90":50.74894428253174,"p95":52.674055099487305,"p99":61.685919761657715},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0078125,"host_held_samples":{"dispatch":15,"roundtrip":16},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.009900989942252636,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":607100928,"scale_bytes":18971904,"total_logical_bytes":626072832},"roundtrip":{"activation_data_bytes":1821302784,"scale_bytes":18971904,"total_logical_bytes":1840274688},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1214201856,"scale_bytes":0,"total_logical_bytes":1214201856},"dispatch":{"activation_data_bytes":607100928,"scale_bytes":18971904,"total_logical_bytes":626072832},"roundtrip":{"activation_data_bytes":1821302784,"scale_bytes":18971904,"total_logical_bytes":1840274688},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":84696,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":5367,"mean":5293.5,"min":5181,"total":84696},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.012361490167677402,"expert_assignments_per_rank":[6171,6137,6243,6050,6268,6146,6227,6063,6176,6175,6160,5993,6212,6144,6110,6029],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,5,241,2597,7671,5870,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.16943359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.06316709518432617,"same_node_fraction":0.49927976727485657,"same_scaleup_domain_fraction":0.49927976727485657,"cross_node_fraction":0.500720202922821,"cross_domain_fraction":0.500720202922821,"gpus_per_node":8,"scale_up_domain":8,"copies":84696},"payload_copies_per_rank":[5326,5273,5366,5236,5367,5310,5348,5225,5315,5334,5305,5181,5318,5287,5280,5225],"payload_rank_cv":0.009882226586341858,"routed_copies":84696},"token_rate_at_latency_percentile":{"p50":8911084.508249495,"p90":8848434.414053615,"p95":8823611.96201735,"p99":8674639.738218788},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1256.2509775161743,"p90":1263.5010480880737,"p95":1265.8660411834717,"p99":1271.1540460586548},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1773.7510204315186,"p90":1784.4940423965454,"p95":1788.7400388717651,"p99":1903.2319784164429},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":3030.001997947693,"p90":3047.995090484619,"p95":3054.606080055237,"p99":3174.3860244750977},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":3140.1679515838623,"p90":3154.7889709472656,"p95":3161.3199710845947,"p99":3182.9919815063477},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3037.2519493103027,"p90":3050.071954727173,"p95":3055.2799701690674,"p99":3284.558057785034},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":419.9939966201782,"p90":423.72000217437744,"p95":425.00200867652893,"p99":428.04598808288574},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1310.73796749115,"p90":1321.5500116348267,"p95":1324.8800039291382,"p99":1331.7699432373047},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1761.8919610977173,"p90":1772.547960281372,"p95":1775.913953781128,"p99":1780.400037765503},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.448984146118164,"p90":6.569147109985352,"p95":6.569147109985352,"p99":6.569147109985352},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":31.527042388916016,"p90":41.58306121826172,"p95":45.187950134277344,"p99":55.00316619873047},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.724170684814453,"p90":3.6859512329101562,"p95":3.6859512329101562,"p99":3.6859512329101562},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1198.3280181884766,"p90":1203.93705368042,"p95":1205.6200504302979,"p99":1210.3869915008545},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1738.664984703064,"p90":1745.5480098724365,"p95":1747.7580308914185,"p99":1762.4520063400269},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2979.257106781006,"p90":2991.0740852355957,"p95":2997.0040321350098,"p99":3212.2490406036377},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":57.836055755615234,"p90":66.04790687561035,"p95":69.25392150878906,"p99":76.94387435913086},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0107421875,"host_held_samples":{"dispatch":22,"roundtrip":19},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":1,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.009900989942252636,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":1212517376,"scale_bytes":37891168,"total_logical_bytes":1250408544},"roundtrip":{"activation_data_bytes":3637552128,"scale_bytes":37891168,"total_logical_bytes":3675443296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2425034752,"scale_bytes":0,"total_logical_bytes":2425034752},"dispatch":{"activation_data_bytes":1212517376,"scale_bytes":37891168,"total_logical_bytes":1250408544},"roundtrip":{"activation_data_bytes":3637552128,"scale_bytes":37891168,"total_logical_bytes":3675443296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":169157,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":10735,"mean":10572.3125,"min":10405,"total":169157},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.008082338608801365,"expert_assignments_per_rank":[12320,12304,12326,12170,12384,12230,12308,12211,12262,12359,12237,12047,12413,12243,12482,12312],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,16,505,5263,15346,11638,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.162261962890625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.0630006417632103,"same_node_fraction":0.5010611414909363,"same_scaleup_domain_fraction":0.5010611414909363,"cross_node_fraction":0.4989388585090637,"cross_domain_fraction":0.4989388585090637,"gpus_per_node":8,"scale_up_domain":8,"copies":169157},"payload_copies_per_rank":[10630,10585,10602,10505,10644,10564,10577,10498,10504,10645,10499,10405,10642,10518,10735,10604],"payload_rank_cv":0.00741199404001236,"routed_copies":169157},"token_rate_at_latency_percentile":{"p50":10788699.96525673,"p90":10743353.103265092,"p95":10725040.035590174,"p99":9976380.208087215},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2301.4259338378906,"p90":2314.724922180176,"p95":2318.7310695648193,"p99":2329.9880027770996},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3212.1291160583496,"p90":3240.330934524536,"p95":3248.944044113159,"p99":3276.6659259796143},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":5513.55504989624,"p90":5555.055856704712,"p95":5567.6751136779785,"p99":5606.653928756714},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":5622.866153717041,"p90":5651.669025421143,"p95":5661.945819854736,"p99":5692.451000213623},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5522.554874420166,"p90":5555.203914642334,"p95":5565.299034118652,"p99":5661.284923553467},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":862.9429936408997,"p90":867.4269914627075,"p95":868.7509894371033,"p99":871.3949918746948},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2302.706003189087,"p90":2317.0979022979736,"p95":2321.2640285491943,"p99":2328.7949562072754},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":3174.2990016937256,"p90":3194.770097732544,"p95":3202.3539543151855,"p99":3210.674047470093},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.289958953857422,"p90":6.810188293457031,"p95":6.810188293457031,"p99":6.810188293457031},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":50.495147705078125,"p90":70.80793380737305,"p95":75.79517364501953,"p99":88.6530876159668},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.6122283935546875,"p90":13.260364532470703,"p95":13.260364532470703,"p99":13.260364532470703},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2195.794105529785,"p90":2203.526020050049,"p95":2205.4810523986816,"p99":2209.605932235718},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3134.6800327301025,"p90":3145.7769870758057,"p95":3149.3139266967773,"p99":3162.601947784424},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5400.37202835083,"p90":5430.778026580811,"p95":5440.011978149414,"p99":5535.955905914307},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":121.92344665527344,"p90":138.14926147460938,"p95":142.37499237060547,"p99":151.54790878295898},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0078125,"host_held_samples":{"dispatch":16,"roundtrip":15},"host_stalled_fraction":0.0009765625,"host_stalled_samples":{"dispatch":0,"roundtrip":2},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":2424583168,"scale_bytes":75768224,"total_logical_bytes":2500351392},"roundtrip":{"activation_data_bytes":7273749504,"scale_bytes":75768224,"total_logical_bytes":7349517728},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4849166336,"scale_bytes":0,"total_logical_bytes":4849166336},"dispatch":{"activation_data_bytes":2424583168,"scale_bytes":75768224,"total_logical_bytes":2500351392},"roundtrip":{"activation_data_bytes":7273749504,"scale_bytes":75768224,"total_logical_bytes":7349517728},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":338251,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":21435,"mean":21140.6875,"min":20921,"total":338251},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.0070151351392269135,"expert_assignments_per_rank":[24582,24524,24711,24358,24680,24435,24611,24307,24504,24638,24478,24415,25004,24532,24846,24591],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,29,1048,10572,30561,23326,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.1613006591796875,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.06308037787675858,"same_node_fraction":0.5009504556655884,"same_scaleup_domain_fraction":0.5009504556655884,"cross_node_fraction":0.49904951453208923,"cross_domain_fraction":0.49904951453208923,"gpus_per_node":8,"scale_up_domain":8,"copies":338251},"payload_copies_per_rank":[21145,21167,21213,20988,21198,21055,21170,20921,21054,21231,21048,21005,21435,21070,21396,21155],"payload_rank_cv":0.00635375315323472,"routed_copies":338251},"token_rate_at_latency_percentile":{"p50":11866971.264252197,"p90":11797226.709763266,"p95":11775827.246339262,"p99":11576170.584056113},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4391.211986541748,"p90":4414.246082305908,"p95":4421.577930450439,"p99":4435.959815979004},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6039.3548011779785,"p90":6070.380210876465,"p95":6080.436229705811,"p99":6200.935840606689},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":10430.566787719727,"p90":10484.626293182373,"p95":10502.01416015625,"p99":10636.895656585693},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":10496.282577514648,"p90":10527.931213378906,"p95":10536.422729492188,"p99":10558.937072753906},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10412.199020385742,"p90":10446.571350097656,"p95":10459.067344665527,"p99":10559.142112731934},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1746.4359998703003,"p90":1754.4080018997192,"p95":1756.932020187378,"p99":1761.4190578460693},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":4291.982173919678,"p90":4311.491966247559,"p95":4315.033912658691,"p99":4326.169013977051},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":5970.869064331055,"p90":5994.746208190918,"p95":6002.117156982422,"p99":6021.866798400879},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.449699401855469,"p90":6.730079650878906,"p95":6.730079650878906,"p99":6.730079650878906},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":106.64081573486328,"p90":147.50289916992188,"p95":157.79876708984375,"p99":176.46503448486328},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-13.501167297363281,"p90":14.38140869140625,"p95":14.38140869140625,"p99":14.38140869140625},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4190.4449462890625,"p90":4210.154056549072,"p95":4215.561866760254,"p99":4226.539134979248},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5947.032928466797,"p90":5962.897777557373,"p95":5967.145919799805,"p99":5993.36576461792},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10207.44800567627,"p90":10236.20891571045,"p95":10244.38190460205,"p99":10370.575904846191},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":205.71422576904297,"p90":241.0449981689453,"p95":250.86212158203125,"p99":267.84515380859375},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0078125,"host_held_samples":{"dispatch":15,"roundtrip":16},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":131072,"byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":4852026368,"scale_bytes":151625824,"total_logical_bytes":5003652192},"roundtrip":{"activation_data_bytes":14556079104,"scale_bytes":151625824,"total_logical_bytes":14707704928},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":9704052736,"scale_bytes":0,"total_logical_bytes":9704052736},"dispatch":{"activation_data_bytes":4852026368,"scale_bytes":151625824,"total_logical_bytes":5003652192},"roundtrip":{"activation_data_bytes":14556079104,"scale_bytes":151625824,"total_logical_bytes":14707704928},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":676901,"assignments":786432,"wire":"rank-deduplicated"},"receive":{"max":42692,"mean":42306.3125,"min":41888,"total":676901},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005902660544961691,"expert_assignments_per_rank":[49152,49158,48937,48637,49532,49147,49291,48640,48855,49080,49131,49180,49736,49110,49535,49311],"expert_load_cv":0.02157296985387802,"expert_load_max":2171,"expert_load_mean":2048.0,"expert_load_min":1922,"fanout_histogram":[0,51,2031,21028,61178,46784,0,0,0,0,0,0,0,0,0,0],"fanout_max":6,"fanout_mean":5.164344787597656,"fanout_min":2,"hotspot_ratio":1.06005859375,"locality":{"placement":"packed","local_rank_fraction":0.06220702826976776,"same_node_fraction":0.5000066757202148,"same_scaleup_domain_fraction":0.5000066757202148,"cross_node_fraction":0.49999335408210754,"cross_domain_fraction":0.49999335408210754,"gpus_per_node":8,"scale_up_domain":8,"copies":676901},"payload_copies_per_rank":[42365,42433,42070,41888,42530,42296,42419,41908,42030,42386,42324,42267,42692,42172,42625,42496],"payload_rank_cv":0.005482310429215431,"routed_copies":676901},"token_rate_at_latency_percentile":{"p50":12588311.051621078,"p90":12546891.760689953,"p95":12531901.333138568,"p99":12413129.646390198},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":36,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":20,"num_max_rdma_chunked_recv_tokens":140},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":288,"num_max_rdma_chunked_send_tokens":12,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":1073741824,"num_qps_per_rank":10,"is_intranode":false,"cpu_proxy_threads":4},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":2,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","scope":"scale-out","topology_class":"mi325x-xgmi-rdma","transport":"xgmi-rdma","world_size":16},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9943,"throttle_fraction_max":{"power":0.0548,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":true,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2023,"p50":2093,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0496,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2019,"p50":2095,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0536,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2025,"p50":2095,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0509,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2026,"p50":2095,"max":2100,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":59,"clock_event_reasons":[],"throttle_fraction":{"power":0.0486,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2017,"p50":2089,"max":2095,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0494,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2027,"p50":2093,"max":2099,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0484,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2017,"p50":2091,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":66,"clock_event_reasons":[],"throttle_fraction":{"power":0.0515,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2024,"p50":2094,"max":2099,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0504,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":8,"source":"amdsmi","sm_clock_mhz":{"min":2032,"p50":2097,"max":2101,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0458,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":9,"source":"amdsmi","sm_clock_mhz":{"min":2020,"p50":2091,"max":2095,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.053,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":10,"source":"amdsmi","sm_clock_mhz":{"min":2017,"p50":2092,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":65,"clock_event_reasons":[],"throttle_fraction":{"power":0.0525,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":11,"source":"amdsmi","sm_clock_mhz":{"min":2009,"p50":2086,"max":2095,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0528,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":12,"source":"amdsmi","sm_clock_mhz":{"min":2012,"p50":2096,"max":2100,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":58,"clock_event_reasons":[],"throttle_fraction":{"power":0.0546,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":13,"source":"amdsmi","sm_clock_mhz":{"min":2023,"p50":2094,"max":2099,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":64,"clock_event_reasons":[],"throttle_fraction":{"power":0.0507,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":14,"source":"amdsmi","sm_clock_mhz":{"min":2013,"p50":2085,"max":2095,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0527,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3},{"rank":15,"source":"amdsmi","sm_clock_mhz":{"min":2015,"p50":2093,"max":2098,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":57,"clock_event_reasons":[],"throttle_fraction":{"power":0.0548,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":1052.3}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"6cbbb96e7ad6cfd9e60e0b84f83957c2c40df0c4","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16.json b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16.json new file mode 100644 index 0000000000..69d8e5ddc5 --- /dev/null +++ b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T06:50:25.368454+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":1,"phase":"prefill","precision":"bf16","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-bf16"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"bf16","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":410.3370010852814,"p90":415.0640070438385,"p95":416.50599241256714,"p99":419.8710024356842},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":516.2969827651978,"p90":522.3050117492676,"p95":524.6289968490601,"p99":529.3560028076172},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":926.6339838504791,"p90":937.3690187931061,"p95":941.1349892616272,"p99":949.2270052433014},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":917.6599979400635,"p90":922.8280186653137,"p95":924.310028553009,"p99":927.1939992904663},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":917.900025844574,"p90":924.0300059318542,"p95":926.0730147361755,"p99":931.3610196113586},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":380.6520104408264,"p90":383.1759989261627,"p95":384.2970132827759,"p99":387.3020112514496},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":500.5919933319092,"p90":504.35900688171387,"p95":505.6009888648987,"p99":508.8449716567993},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.129026412963867,"p90":6.328999996185303,"p95":6.328999996185303,"p99":6.328999996185303},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":7.651984691619873,"p90":11.137008666992188,"p95":12.539982795715332,"p99":14.460980892181396},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.4820098876953125,"p90":1.4820098876953125,"p95":1.4820098876953125,"p99":1.4820098876953125},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":383.2570016384125,"p90":386.5810036659241,"p95":388.14300298690796,"p99":391.42900705337524},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":491.4189875125885,"p90":496.82700634002686,"p95":498.82999062538147,"p99":503.2370090484619},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":890.097975730896,"p90":895.9869742393494,"p95":897.8700041770935,"p99":903.2390117645264},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":27.72200107574463,"p90":31.889021396636963,"p95":33.20997953414917,"p99":36.81600093841553},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00146484375,"host_stalled_samples":{"dispatch":0,"roundtrip":3},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"roundtrip":{"activation_data_bytes":1041510400,"scale_bytes":0,"total_logical_bytes":1041510400},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"roundtrip":{"activation_data_bytes":1041510400,"scale_bytes":0,"total_logical_bytes":1041510400},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":36325,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":4587,"mean":4540.625,"min":4510,"total":36325},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.009137080051004887,"expert_assignments_per_rank":[6159,6057,6238,6119,6110,6118,6221,6130],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,46,880,3426,3151,689,0,0],"fanout_max":6,"fanout_mean":4.4342041015625,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.12393668293952942,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":36325},"payload_copies_per_rank":[4582,4510,4587,4547,4516,4535,4533,4515],"payload_rank_cv":0.006128025706857443,"routed_copies":36325},"token_rate_at_latency_percentile":{"p50":8924719.217065513,"p90":8865512.967556324,"p95":8845954.767760703,"p99":8795729.934476305},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":693.5229897499084,"p90":697.609007358551,"p95":699.1320252418518,"p99":702.8970122337341},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":836.6190195083618,"p90":842.9880142211914,"p95":845.3919887542725,"p99":849.9979972839355},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1530.1420092582703,"p90":1540.5970215797424,"p95":1544.5240139961243,"p99":1552.8950095176697},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1519.204020500183,"p90":1525.6550312042236,"p95":1527.8170108795166,"p99":1531.3830375671387},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1520.5659866333008,"p90":1527.0969867706299,"p95":1529.1390419006348,"p99":1533.1449508666992},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":680.7439923286438,"p90":685.9109997749329,"p95":686.9940161705017,"p99":689.6770000457764},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":815.9869909286499,"p90":821.9559788703918,"p95":823.8790035247803,"p99":827.8049826622009},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.209969520568848,"p90":6.249070167541504,"p95":6.249070167541504,"p99":6.249070167541504},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":12.018084526062012,"p90":18.108010292053223,"p95":19.67000961303711,"p99":24.235963821411133},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1.443028450012207,"p90":3.285050392150879,"p95":3.285050392150879,"p99":3.285050392150879},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":672.1310019493103,"p90":676.0169863700867,"p95":677.3779988288879,"p99":680.2629828453064},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":807.0930242538452,"p90":812.3019933700562,"p95":813.86399269104,"p99":817.0289993286133},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1498.6939430236816,"p90":1505.905032157898,"p95":1507.9079866409302,"p99":1511.7939710617065},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":21.87204360961914,"p90":27.119994163513184,"p95":28.561949729919434,"p99":31.206965446472168},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01149425283074379,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"roundtrip":{"activation_data_bytes":2083221504,"scale_bytes":0,"total_logical_bytes":2083221504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"roundtrip":{"activation_data_bytes":2083221504,"scale_bytes":0,"total_logical_bytes":2083221504},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":72657,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":9142,"mean":9082.125,"min":9007,"total":72657},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00735377287492156,"expert_assignments_per_rank":[12308,12293,12414,12290,12351,12153,12356,12139],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,92,1754,6798,6421,1319,0,0],"fanout_max":6,"fanout_mean":4.43463134765625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.12587912380695343,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":72657},"payload_copies_per_rank":[9133,9070,9142,9078,9120,9057,9050,9007],"payload_rank_cv":0.004780265036970377,"routed_copies":72657},"token_rate_at_latency_percentile":{"p50":10774935.217560645,"p90":10728853.59733925,"p95":10714525.985573947,"p99":10686530.31843988},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1274.5970487594604,"p90":1282.6889753341675,"p95":1284.4120264053345,"p99":1288.89799118042},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1485.9939813613892,"p90":1494.6080446243286,"p95":1496.770977973938,"p99":1500.3759860992432},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":2760.5910301208496,"p90":2777.297019958496,"p95":2781.1830043792725,"p99":2789.273977279663},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":2745.4090118408203,"p90":2757.7879428863525,"p95":2760.751962661743,"p99":2767.3211097717285},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2746.609926223755,"p90":2757.58695602417,"p95":2761.193037033081,"p99":2766.439914703369},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1237.902045249939,"p90":1249.0390539169312,"p95":1253.0039548873901,"p99":1256.971001625061},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1462.6799821853638,"p90":1473.4959602355957,"p95":1476.2200117111206,"p99":1480.7469844818115},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.208896636962891,"p90":6.409168243408203,"p95":6.409168243408203,"p99":6.409168243408203},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":23.233890533447266,"p90":33.77103805541992,"p95":35.9339714050293,"p99":42.22416877746582},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2.7239322662353516,"p90":4.167079925537109,"p95":4.167079925537109,"p99":4.167079925537109},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1228.4870147705078,"p90":1233.6549758911133,"p95":1235.3769540786743,"p99":1239.3840551376343},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1452.5450468063354,"p90":1459.2349529266357,"p95":1461.3169431686401,"p99":1464.8430347442627},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2700.261116027832,"p90":2709.2740535736084,"p95":2711.3170623779297,"p99":2716.404914855957},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":46.47088050842285,"p90":56.205034255981445,"p95":58.567047119140625,"p99":63.7362003326416},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0009765625,"host_stalled_samples":{"dispatch":2,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"roundtrip":{"activation_data_bytes":4160278528,"scale_bytes":0,"total_logical_bytes":4160278528},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"roundtrip":{"activation_data_bytes":4160278528,"scale_bytes":0,"total_logical_bytes":4160278528},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":145099,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":18284,"mean":18137.375,"min":17991,"total":145099},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005656693130731583,"expert_assignments_per_rank":[24624,24496,24614,24519,24621,24284,24656,24794],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,192,3636,13520,12793,2627,0,0],"fanout_max":6,"fanout_mean":4.428070068359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.125583216547966,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":145099},"payload_copies_per_rank":[18250,18101,18151,18109,18125,17991,18088,18284],"payload_rank_cv":0.004795215092599392,"routed_copies":145099},"token_rate_at_latency_percentile":{"p50":11930343.543559497,"p90":11882852.842923295,"p95":11867333.996760117,"p99":11844826.206360439},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2420.799970626831,"p90":2431.4169883728027,"p95":2434.180974960327,"p99":2439.3088817596436},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2772.329092025757,"p90":2783.345937728882,"p95":2787.029981613159,"p99":2793.360948562622},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":5193.129062652588,"p90":5214.762926101685,"p95":5221.210956573486,"p99":5232.669830322266},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":5175.502777099609,"p90":5191.327095031738,"p95":5195.093154907227,"p99":5201.9829750061035},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5180.39083480835,"p90":5196.094036102295,"p95":5200.541019439697,"p99":5209.555149078369},"sample_count":2048},"stage":{"availability":"unavailable","origin":null,"percentiles_us":null,"sample_count":0}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2371.846914291382,"p90":2380.419969558716,"p95":2382.9829692840576,"p99":2387.3510360717773},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2743.32594871521,"p90":2755.263090133667,"p95":2758.268117904663,"p99":2761.6729736328125},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.330013275146484,"p90":6.410121917724609,"p95":6.410121917724609,"p99":6.410121917724609},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":43.46609115600586,"p90":59.088706970214844,"p95":63.7359619140625,"p99":72.58987426757812},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.04412841796875,"p90":3.8061141967773438,"p95":3.8061141967773438,"p99":3.8061141967773438},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2370.2850341796875,"p90":2377.013921737671,"p95":2379.218101501465,"p99":2383.2640647888184},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2735.39400100708,"p90":2744.1670894622803,"p95":2746.2499141693115,"p99":2750.214099884033},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":5130.795001983643,"p90":5143.734931945801,"p95":5147.420883178711,"p99":5154.070854187012},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":50.436973571777344,"p90":61.371803283691406,"p95":63.93575668334961,"p99":69.22388076782227},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0009765625,"host_stalled_samples":{"dispatch":0,"roundtrip":2},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0}},"correctness":{"chain_last_output_passed":true,"chain_last_output_error":0.0,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"roundtrip":{"activation_data_bytes":8316542976,"scale_bytes":0,"total_logical_bytes":8316542976},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"roundtrip":{"activation_data_bytes":8316542976,"scale_bytes":0,"total_logical_bytes":8316542976},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":290058,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":36460,"mean":36257.25,"min":36144,"total":290058},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00432177959010005,"expert_assignments_per_rank":[49106,49069,49115,48918,49142,48893,49536,49437],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,413,7235,27175,25451,5262,0,0],"fanout_max":6,"fanout_mean":4.425933837890625,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.1256852149963379,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":290058},"payload_copies_per_rank":[36354,36201,36216,36144,36222,36193,36268,36460],"payload_rank_cv":0.0026462646201252937,"routed_copies":290058},"token_rate_at_latency_percentile":{"p50":12650782.940863676,"p90":12612550.801555548,"p95":12601765.807639146,"p99":12579960.88429817},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":6,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":0,"num_qps_per_rank":1,"is_intranode":true,"cpu_proxy_threads":0},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":false,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi325x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9952,"throttle_fraction_max":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":false,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2095,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":373.6},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2093,"p50":2097,"max":2099,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":72,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":373.6},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2091,"p50":2095,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":71,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":373.6},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2099,"p50":2102,"max":2104,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":373.6},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2097,"p50":2101,"max":2103,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":62,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":373.6},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":2092,"p50":2096,"max":2098,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":70,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":373.6},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2086,"p50":2092,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":72,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":373.6},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2090,"p50":2096,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0001,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":373.6}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}} diff --git a/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8.json b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8.json new file mode 100644 index 0000000000..fa69a98835 --- /dev/null +++ b/collectivex/results/mi325x/mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8.json @@ -0,0 +1 @@ +{"version":1,"record_type":"case-attempt","generated_at":"2026-10-28T07:05:59.207868+00:00","identity":{"allocation_factors":{"run_attempt":null,"run_id":null,"source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5"},"attempt_ordinal":1,"case_factors":{"case":{"backend":"uccl-ep","ep":8,"experts":384,"gpus_per_node":8,"hidden":7168,"ladder":"1024 2048 4096 8192","mode":"normal","nodes":1,"phase":"prefill","precision":"fp8","routing":"uniform","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","suite":"ep-core","topk":6,"topology_class":"mi325x-xgmi","transport":"xgmi","workload":"deepseek-v4-pro"},"sku":"mi325x"},"case_id":"mi325x-uccl-ep-deepseek-v4-pro-normal-prefill-ep8-uniform-fp8"},"workload":{"cross_rank_consistent":true,"ladder_measured":[1024,2048,4096,8192],"ladder_dropped":[],"ladder_cap":null},"measurement":{"combine_dtype":"bf16","combine_semantics":"activation-only","dispatch_dtype":"fp8-e4m3fnuz","payload_unit":"token-rank","rows":[{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":411.69899702072144,"p90":416.265994310379,"p95":418.067991733551,"p99":421.31200432777405},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":426.80200934410095,"p90":432.68999457359314,"p95":435.2149963378906,"p99":441.5439963340759},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":838.5010063648224,"p90":848.9559888839722,"p95":853.2829880714417,"p99":862.85600066185},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":815.4259920120239,"p90":820.4740285873413,"p95":821.9969868659973,"p99":825.3610134124756},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":817.2289729118347,"p90":823.7590193748474,"p95":826.0030150413513,"p99":832.131028175354},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":183.03599953651428,"p90":186.039999127388,"p95":187.24200129508972,"p99":190.20600616931915},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":388.4640038013458,"p90":391.0279870033264,"p95":391.94899797439575,"p99":393.2720124721527},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":401.52400732040405,"p90":404.6880006790161,"p95":406.33198618888855,"p99":409.09498929977417},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.209969520568848,"p90":6.2490105628967285,"p95":6.2490105628967285,"p99":6.2490105628967285},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":8.252978324890137,"p90":11.656999588012695,"p95":12.657999992370605,"p99":16.303956508636475},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.0419955253601074,"p90":-1.1230111122131348,"p95":-1.1230111122131348,"p99":-1.1230111122131348},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":386.70098781585693,"p90":390.70698618888855,"p95":392.15001463890076,"p99":395.3939974308014},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":406.5310060977936,"p90":412.42098808288574,"p95":414.66400027275085,"p99":419.99199986457825},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":791.9110059738159,"p90":798.6810207366943,"p95":800.8840084075928,"p99":806.3330054283142},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":25.119006633758545,"p90":29.36500310897827,"p95":30.764997005462646,"p99":33.49095582962036},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.006756756920367479,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":8192,"byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":260377600,"scale_bytes":8136800,"total_logical_bytes":268514400},"roundtrip":{"activation_data_bytes":781132800,"scale_bytes":8136800,"total_logical_bytes":789269600},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":520755200,"scale_bytes":0,"total_logical_bytes":520755200},"dispatch":{"activation_data_bytes":260377600,"scale_bytes":8136800,"total_logical_bytes":268514400},"roundtrip":{"activation_data_bytes":781132800,"scale_bytes":8136800,"total_logical_bytes":789269600},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":36325,"assignments":49152,"wire":"rank-deduplicated"},"receive":{"max":4587,"mean":4540.625,"min":4510,"total":36325},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.009137080051004887,"expert_assignments_per_rank":[6159,6057,6238,6119,6110,6118,6221,6130],"expert_load_cv":0.09066266566514969,"expert_load_max":158,"expert_load_mean":128.0,"expert_load_min":88,"fanout_histogram":[0,46,880,3426,3151,689,0,0],"fanout_max":6,"fanout_mean":4.4342041015625,"fanout_min":2,"hotspot_ratio":1.234375,"locality":{"placement":"packed","local_rank_fraction":0.12393668293952942,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":36325},"payload_copies_per_rank":[4582,4510,4587,4547,4516,4535,4533,4515],"payload_rank_cv":0.006128025706857443,"routed_copies":36325},"token_rate_at_latency_percentile":{"p50":10024118.419115055,"p90":9944655.90946358,"p95":9917639.343713405,"p99":9844603.46102334},"tokens_per_rank":1024},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":706.3819766044617,"p90":711.7099761962891,"p95":713.1929993629456,"p99":716.8779969215393},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":642.4859762191772,"p90":647.9349732398987,"p95":649.97798204422,"p99":654.744029045105},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":1348.867952823639,"p90":1359.6449494361877,"p95":1363.1709814071655,"p99":1371.6220259666443},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":1335.2080583572388,"p90":1341.0569429397583,"p95":1342.9399728775024,"p99":1347.465991973877},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1339.4540548324585,"p90":1346.9059467315674,"p95":1349.1489887237549,"p99":1353.5159826278687},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":356.33599758148193,"p90":359.62098836898804,"p95":360.82300543785095,"p99":363.3059859275818},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":682.4259757995605,"p90":688.1150007247925,"p95":690.3190016746521,"p99":694.2039728164673},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":628.1440258026123,"p90":632.070004940033,"p95":633.8729858398438,"p99":635.7560157775879},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.25002384185791,"p90":6.289005279541016,"p95":6.289005279541016,"p99":6.289005279541016},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":10.975003242492676,"p90":15.984058380126953,"p95":17.746925354003906,"p99":21.75307273864746},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-1.0418891906738281,"p90":1.7229318618774414,"p95":1.7229318618774414,"p99":1.7229318618774414},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":679.1809797286987,"p90":683.1079721450806,"p95":684.6290230751038,"p99":687.8349781036377},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":625.34099817276,"p90":629.9070119857788,"p95":631.3899755477905,"p99":634.7950100898743},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1312.013030052185,"p90":1318.261981010437,"p95":1320.4259872436523,"p99":1324.8720169067383},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":27.400970458984375,"p90":32.76801109313965,"p95":34.49296951293945,"p99":37.81700134277344},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0009765625,"host_stalled_samples":{"dispatch":2,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":16384,"byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":520805376,"scale_bytes":16275168,"total_logical_bytes":537080544},"roundtrip":{"activation_data_bytes":1562416128,"scale_bytes":16275168,"total_logical_bytes":1578691296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":1041610752,"scale_bytes":0,"total_logical_bytes":1041610752},"dispatch":{"activation_data_bytes":520805376,"scale_bytes":16275168,"total_logical_bytes":537080544},"roundtrip":{"activation_data_bytes":1562416128,"scale_bytes":16275168,"total_logical_bytes":1578691296},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":72657,"assignments":98304,"wire":"rank-deduplicated"},"receive":{"max":9142,"mean":9082.125,"min":9007,"total":72657},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00735377287492156,"expert_assignments_per_rank":[12308,12293,12414,12290,12351,12153,12356,12139],"expert_load_cv":0.05904974043369293,"expert_load_max":302,"expert_load_mean":256.0,"expert_load_min":216,"fanout_histogram":[0,92,1754,6798,6421,1319,0,0],"fanout_max":6,"fanout_mean":4.43463134765625,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.12587912380695343,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":72657},"payload_copies_per_rank":[9133,9070,9142,9078,9120,9057,9050,9007],"payload_rank_cv":0.004780265036970377,"routed_copies":72657},"token_rate_at_latency_percentile":{"p50":12231849.193251608,"p90":12164175.263876285,"p95":12143951.584990371,"p99":12104770.250433434},"tokens_per_rank":2048},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1290.6609773635864,"p90":1299.0729808807373,"p95":1301.3570308685303,"p99":1305.122971534729},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1094.6060419082642,"p90":1101.4569997787476,"p95":1104.0199995040894,"p99":1107.8660488128662},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":2385.2670192718506,"p90":2400.529980659485,"p95":2405.3770303726196,"p99":2412.989020347595},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":2378.777027130127,"p90":2389.272928237915,"p95":2393.1190967559814,"p99":2404.2959213256836},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2376.5740394592285,"p90":2387.5908851623535,"p95":2390.6350135803223,"p99":2395.7619667053223},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":730.9799790382385,"p90":735.7059717178345,"p95":737.2279763221741,"p99":740.032970905304},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1259.0140104293823,"p90":1266.9459581375122,"p95":1269.389033317566,"p99":1276.2800455093384},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1082.6679468154907,"p90":1091.4809703826904,"p95":1094.0049886703491,"p99":1100.8960008621216},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.289958953857422,"p90":6.448984146118164,"p95":6.448984146118164,"p99":6.448984146118164},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":23.916006088256836,"p90":32.608985900878906,"p95":35.77399253845215,"p99":40.34113883972168},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":-2.562999725341797,"p90":4.126071929931641,"p95":4.126071929931641,"p99":4.126071929931641},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1243.8709735870361,"p90":1249.5189905166626,"p95":1251.162052154541,"p99":1255.8480501174927},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1067.5660371780396,"p90":1074.0150213241577,"p95":1075.5380392074585,"p99":1079.463005065918},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2329.0629386901855,"p90":2337.594985961914,"p95":2340.359926223755,"p99":2345.607042312622},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":47.029972076416016,"p90":57.524919509887695,"p95":60.36996841430664,"p99":65.41800498962402},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.0,"host_stalled_samples":{"dispatch":0,"roundtrip":0},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.01220703125,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":32768,"byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":1040069632,"scale_bytes":32502176,"total_logical_bytes":1072571808},"roundtrip":{"activation_data_bytes":3120208896,"scale_bytes":32502176,"total_logical_bytes":3152711072},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":2080139264,"scale_bytes":0,"total_logical_bytes":2080139264},"dispatch":{"activation_data_bytes":1040069632,"scale_bytes":32502176,"total_logical_bytes":1072571808},"roundtrip":{"activation_data_bytes":3120208896,"scale_bytes":32502176,"total_logical_bytes":3152711072},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":145099,"assignments":196608,"wire":"rank-deduplicated"},"receive":{"max":18284,"mean":18137.375,"min":17991,"total":145099},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.005656693130731583,"expert_assignments_per_rank":[24624,24496,24614,24519,24621,24284,24656,24794],"expert_load_cv":0.044396016746759415,"expert_load_max":604,"expert_load_mean":512.0,"expert_load_min":444,"fanout_histogram":[0,192,3636,13520,12793,2627,0,0],"fanout_max":6,"fanout_mean":4.428070068359375,"fanout_min":2,"hotspot_ratio":1.1796875,"locality":{"placement":"packed","local_rank_fraction":0.125583216547966,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":145099},"payload_copies_per_rank":[18250,18101,18151,18109,18125,17991,18088,18284],"payload_rank_cv":0.004795215092599392,"routed_copies":145099},"token_rate_at_latency_percentile":{"p50":13787914.64349081,"p90":13724294.310066365,"p95":13706818.403418753,"p99":13677485.683213724},"tokens_per_rank":4096},{"components":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2431.335926055908,"p90":2445.276975631714,"p95":2447.92103767395,"p99":2453.4099102020264},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1987.1879816055298,"p90":1997.4839687347412,"p95":2000.8490085601807,"p99":2005.897045135498},"sample_count":2048},"isolated_sum":{"availability":"derived","origin":"derived-percentile-sum","percentiles_us":{"p50":4418.523907661438,"p90":4442.760944366455,"p95":4448.770046234131,"p99":4459.306955337524},"sample_count":0},"pair_period":{"availability":"measured","origin":"chained-median","percentiles_us":{"p50":4413.9580726623535,"p90":4428.218841552734,"p95":4432.145118713379,"p99":4437.833786010742},"sample_count":448},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4413.999080657959,"p90":4433.267116546631,"p95":4438.55619430542,"p99":4449.010848999023},"sample_count":2048},"stage":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1483.109951019287,"p90":1489.240050315857,"p95":1491.4029836654663,"p99":1495.3689575195312},"sample_count":2048}},"chain_floor_us":{"combine":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":2383.3041191101074,"p90":2391.8368816375732,"p95":2393.7599658966064,"p99":2401.4909267425537},"sample_count":448},"dispatch":{"availability":"measured","origin":"chained-cross-rank-min","percentiles_us":{"p50":1979.377031326294,"p90":1992.7159547805786,"p95":1995.4010248184204,"p99":2001.5699863433838},"sample_count":448}},"chain_health":{"interpair_gap_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":6.248950958251953,"p90":6.409168243408203,"p95":6.409168243408203,"p99":6.409168243408203},"sample_count":4},"pair_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":35.653114318847656,"p90":51.31673812866211,"p95":55.4041862487793,"p99":63.456058502197266},"sample_count":448},"settle_drift_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":3.6458969116210938,"p90":7.811069488525391,"p95":7.811069488525391,"p99":7.811069488525391},"sample_count":4},"host_retried_siblings":0,"host_stalled_trials":0},"cross_rank_min_us":{"combine":{"availability":"measured","origin":"measured","percentiles_us":{"p50":2374.411106109619,"p90":2382.1818828582764,"p95":2384.4659328460693,"p99":2388.9520168304443},"sample_count":2048},"dispatch":{"availability":"measured","origin":"measured","percentiles_us":{"p50":1960.6690406799316,"p90":1967.43905544281,"p95":1969.2020416259766,"p99":1973.0470180511475},"sample_count":2048},"roundtrip":{"availability":"measured","origin":"measured","percentiles_us":{"p50":4355.3900718688965,"p90":4368.410110473633,"p95":4371.614933013916,"p99":4378.544807434082},"sample_count":2048}},"cross_rank_spread_us":{"availability":"measured","origin":"measured","percentiles_us":{"p50":57.88707733154297,"p90":76.11370086669922,"p95":80.40189743041992,"p99":87.33034133911133},"sample_count":2048},"rank_alignment":{"host_held_fraction":0.0,"host_held_samples":{"dispatch":0,"roundtrip":0},"host_stalled_fraction":0.00048828125,"host_stalled_samples":{"dispatch":0,"roundtrip":1},"missed_fraction":0.0,"missed_samples":{"combine":0,"dispatch":0,"roundtrip":0,"stage":0}},"correctness":{"chain_last_output_passed":null,"chain_last_output_error":null,"post_chain_state_passed":true,"cuda_graph_output_rewritten":null,"cuda_graph_last_output_passed":null,"cuda_graph_last_output_error":null,"max_relative_error":0.0244140625,"oracle_passed":{"pre":true,"chained":true,"post":true},"oracle_failed_checks":{"pre":[],"chained":[],"post":[]},"passed":true},"global_tokens":65536,"byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":2079135744,"scale_bytes":64972992,"total_logical_bytes":2144108736},"roundtrip":{"activation_data_bytes":6237407232,"scale_bytes":64972992,"total_logical_bytes":6302380224},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"wire_byte_provenance":{"combine":{"activation_data_bytes":4158271488,"scale_bytes":0,"total_logical_bytes":4158271488},"dispatch":{"activation_data_bytes":2079135744,"scale_bytes":64972992,"total_logical_bytes":2144108736},"roundtrip":{"activation_data_bytes":6237407232,"scale_bytes":64972992,"total_logical_bytes":6302380224},"stage":{"activation_data_bytes":0,"scale_bytes":0,"total_logical_bytes":0}},"logical_copies":{"routed":290058,"assignments":393216,"wire":"rank-deduplicated"},"receive":{"max":36460,"mean":36257.25,"min":36144,"total":290058},"routing":{"empty_expert_count":0,"empty_rank_count":0,"expert_assignment_rank_cv":0.00432177959010005,"expert_assignments_per_rank":[49106,49069,49115,48918,49142,48893,49536,49437],"expert_load_cv":0.03081125020980835,"expert_load_max":1132,"expert_load_mean":1024.0,"expert_load_min":924,"fanout_histogram":[0,413,7235,27175,25451,5262,0,0],"fanout_max":6,"fanout_mean":4.425933837890625,"fanout_min":2,"hotspot_ratio":1.10546875,"locality":{"placement":"packed","local_rank_fraction":0.1256852149963379,"same_node_fraction":1.0,"same_scaleup_domain_fraction":1.0,"cross_node_fraction":0.0,"cross_domain_fraction":0.0,"gpus_per_node":8,"scale_up_domain":8,"copies":290058},"payload_copies_per_rank":[36354,36201,36216,36144,36222,36193,36268,36460],"payload_rank_cv":0.0026462646201252937,"routed_copies":290058},"token_rate_at_latency_percentile":{"p50":14847307.125000825,"p90":14782777.18827157,"p95":14765161.717245216,"p99":14730465.315620627},"tokens_per_rank":8192}],"sampling":{"iterations_per_trial":8,"samples_per_component":2048,"trials":256,"warmup_iterations":32,"chain_drop":16,"chain_iterations_per_trial":128,"chain_trials":4}},"implementation":{"library_config":{"deep_ep":"uccl deep_ep_wrapper","num_sms":20,"dispatch_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":6,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"combine_config":{"num_sms":20,"num_max_nvl_chunked_send_tokens":4,"num_max_nvl_chunked_recv_tokens":256,"num_max_rdma_chunked_send_tokens":6,"num_max_rdma_chunked_recv_tokens":132},"num_nvl_bytes":1073741824,"num_rdma_bytes":0,"num_qps_per_rank":1,"is_intranode":true,"cpu_proxy_threads":0},"fp8_consume":"native","kernel_generation":"uccl-legacy-buffer-vllm-ht-eager-spin-ar","rank_alignment":{"method":"gpu-spin-then-all-reduce","spin_us":2000.0},"combine_reduction":"domain-fp32","library_version":"uccl@df000fd3001d deep_ep==0.1.0 USE_DMABUF=1 PER_EXPERT_BATCHING=1","stage_excluded_from_roundtrip":true,"chained_period":true,"cuda_graph_replay":false,"cuda_graph_supported":false,"maturity":"candidate","name":"uccl-ep"},"topology":{"device_product":"","gpus_per_node":8,"nodes":1,"placement":"packed","scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":null,"scope":"scale-up","topology_class":"mi325x-xgmi","transport":"xgmi","world_size":8},"runtime":{"accelerator_runtime":"7.2.53211","collective_library":{"kind":"rccl","version":"2.27.7"},"framework":"2.13.0+git733fca1","vendor":"amd","gpu_clocks":{"source":"amdsmi","unavailable_reason":null,"sm_clock_ratio_min":0.9938,"throttle_fraction_max":{"power":0.091,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"slow_gpu":true,"per_rank":[{"rank":0,"source":"amdsmi","sm_clock_mhz":{"min":2006,"p50":2089,"max":2094,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0835,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":466.5},{"rank":1,"source":"amdsmi","sm_clock_mhz":{"min":2006,"p50":2091,"max":2097,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":72,"clock_event_reasons":[],"throttle_fraction":{"power":0.0861,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":466.5},{"rank":2,"source":"amdsmi","sm_clock_mhz":{"min":2001,"p50":2087,"max":2093,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":72,"clock_event_reasons":[],"throttle_fraction":{"power":0.0866,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":466.5},{"rank":3,"source":"amdsmi","sm_clock_mhz":{"min":2019,"p50":2098,"max":2103,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":60,"clock_event_reasons":[],"throttle_fraction":{"power":0.0781,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":466.5},{"rank":4,"source":"amdsmi","sm_clock_mhz":{"min":2013,"p50":2095,"max":2100,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":61,"clock_event_reasons":[],"throttle_fraction":{"power":0.0818,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":466.5},{"rank":5,"source":"amdsmi","sm_clock_mhz":{"min":1999,"p50":2090,"max":2096,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":70,"clock_event_reasons":[],"throttle_fraction":{"power":0.091,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":466.5},{"rank":6,"source":"amdsmi","sm_clock_mhz":{"min":2000,"p50":2085,"max":2091,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":73,"clock_event_reasons":[],"throttle_fraction":{"power":0.0862,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":466.5},{"rank":7,"source":"amdsmi","sm_clock_mhz":{"min":2011,"p50":2090,"max":2095,"samples":1040},"sm_clock_limit_mhz":2100,"power_limit_w":1000.0,"temperature_max_c":63,"clock_event_reasons":[],"throttle_fraction":{"power":0.0804,"thermal":0.0,"hbm_thermal":0.0,"vr_thermal":0.0,"prochot":0.0,"sync_boost":null,"board_limit":null,"reliability":null},"span_s":466.5}]}},"provenance":{"image":"vllm/vllm-openai-rocm:v0.32.0","source_sha":"068e7c018c42a9554bf2b26fa2695f3ea67f66c5","source_dirty":false},"outcome":{"reasons":[],"status":"success"}}