perf(amd): DSv4.1-Flash ATOM AgentX DPA + prefill TBO from c32 on the 1022 nightly / DSv4.1-Flash ATOM AgentX 更新至 1022 nightly 并从 c32 起开启 DPA + prefill TBO

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/dsv41flash/atom/mi355x-fp4-mtp/agentic.yaml

@@ -6,7 +6,7 @@ base:
     gpus_per_node: 8
   frontend:
     type: atomesh
-    container_image: rocm/atom-dev@sha256:a60811a8467b43e3e50e9e9eb8cec34f68b7bca03f8c6dbd6a0ff813dde49d4c
+    container_image: rocm/atom-dev@sha256:96a9614276518c2619e88800dec5f276bc76a95691331b4abb2210184e1431af
     enable_multiple_frontends: false
   observability:
     enabled: false
@@ -20,10 +20,9 @@ base:
       args:
         served-model-name: deepseek-ai/DeepSeek-V4.1-Flash
         trust-remote-code: true
-        kv_cache_dtype: bf16
+        kv_cache_dtype: fp8
         index-cache-dtype: fp4
         gpu-memory-utilization: 0.9
-        max-num-seqs: 128
         max-num-batched-tokens: 16384
         attn-prefill-chunk-size: 16384
         enable_prefix_caching: true
@@ -50,13 +49,26 @@ base:
       AIPERF_TIMING_CANCEL_DRAIN_TIMEOUT: '300'
       AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'atom:'
 
-# One variant per point. Every point admits 128 sequences and captures every
-# decode size through 32 plus 48, 64 and 128.
+# One variant per point, on an MXFP8 KV cache. Plain TP everywhere, plus DEP2
+# (DP attention and expert parallelism over the TP group) from concurrency 64.
+# max-num-seqs is ATOM's default except at concurrency 1; decode graphs cover
+# every size through 32 plus 48, 64 and 128 (and 256 on the c256 points).
+#
+# Concurrency 1 follows the C1 recipe of ROCm/ATOM#2659: the prefill delayer
+# with a decode interval no step reaches, so a new turn's prefill waits for the
+# running decode instead of interleaving with it, and max-num-seqs 2 (DSpark's
+# block-drafter warmup needs two sequences) with decode graphs at 1 and 2.
 
 override_tp2_c1:
   roles:
     agg:
       gpus: 2
+      args: &c1_args
+        max-num-seqs: 2
+        cudagraph-capture-sizes: '[1,2]'
+      env: &c1_env
+        ATOM_ENABLE_PREFILL_DELAYER: '1'
+        ATOM_PREFILL_DECODE_INTERVAL: '1000000'
   benchmark:
     env:
       CONC: '1'
@@ -77,14 +89,6 @@ override_tp2_c4:
     env:
       CONC: '4'
 
-override_tp2_c6:
-  roles:
-    agg:
-      gpus: 2
-  benchmark:
-    env:
-      CONC: '6'
-
 override_tp2_c8:
   roles:
     agg:
@@ -113,18 +117,34 @@ override_tp2_c64:
   roles:
     agg:
       gpus: 2
-      # Run at least 4 decode passes after each prefill and batch the prefills
-      # that arrive meanwhile; at this load the batching also raises throughput.
-      env:
-        ATOM_PREFILL_DECODE_INTERVAL: '4'
   benchmark:
     env:
       CONC: '64'
 
+override_tp2_c128:
+  roles:
+    agg:
+      gpus: 2
+  benchmark:
+    env:
+      CONC: '128'
+
+override_tp2_c256:
+  roles:
+    agg:
+      gpus: 2
+      args:
+        cudagraph-capture-sizes: '[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,48,64,128,256]'
+  benchmark:
+    env:
+      CONC: '256'
+
 override_tp4_c1:
   roles:
     agg:
       gpus: 4
+      args: *c1_args
+      env: *c1_env
   benchmark:
     env:
       CONC: '1'
@@ -137,14 +157,57 @@ override_tp4_c8:
     env:
       CONC: '8'
 
-override_tp4_c32:
+# DEP2: the DSv4.1 DEP2 cell of ROCm/ATOM#2784 (DP attention without TBO, which
+# does not support SWA bounded replay, plus expert parallelism). Each AIPerf
+# session stays on one DP rank and at least 10 decode passes run between
+# prefills.
+
+override_tp2_c64_dep:
   roles:
     agg:
-      gpus: 4
-      # Run at least 16 decode passes after each prefill and batch the prefills
-      # that arrive meanwhile, so new turns stall decode less often.
-      env:
-        ATOM_PREFILL_DECODE_INTERVAL: '16'
+      gpus: 2
+      args:
+        enable-dp-attention: true
+        enable-expert-parallel: true
+        # ATOM_DP_LB_REQ_EQUIV weights in-flight requests under least_tokens.
+        dp-load-balance: least_tokens
+      env: &dep_env
+        GPU_MAX_HW_QUEUES: '5'
+        ATOM_DP_SESSION_AFFINITY: '1'
+        ATOM_DP_LB_REQ_EQUIV: '512'
+        ATOM_ENABLE_PREFILL_DELAYER: '1'
+        ATOM_PREFILL_DECODE_INTERVAL: '10'
   benchmark:
     env:
-      CONC: '32'
+      CONC: '64'
+      # Session affinity keys on the AIPerf session id.
+      AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: '1'
+
+override_tp2_c128_dep:
+  roles:
+    agg:
+      gpus: 2
+      args:
+        enable-dp-attention: true
+        enable-expert-parallel: true
+        dp-load-balance: least_tokens
+      env: *dep_env
+  benchmark:
+    env:
+      CONC: '128'
+      AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: '1'
+
+override_tp2_c256_dep:
+  roles:
+    agg:
+      gpus: 2
+      args:
+        enable-dp-attention: true
+        enable-expert-parallel: true
+        dp-load-balance: least_tokens
+        cudagraph-capture-sizes: '[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,48,64,128,256]'
+      env: *dep_env
+  benchmark:
+    env:
+      CONC: '256'
+      AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: '1'

inferencex-e2e/configs/amd-master.yaml

@@ -1628,10 +1628,12 @@ minimaxm31flash-fp4-mi355x-sglang-agentic-dspark:
       # Low-Latency: TP4, GPU-resident.
       - { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 8, 16, 32, 48, 64], srt-recipe: agentic.yaml }
 
-# DeepSeek-V4.1-Flash on ATOM, following the upstream TP2 / TP4 AgentX recipe.
+# DeepSeek-V4.1-Flash on ATOM with an MXFP8 KV cache, following the upstream
+# TP2 / TP4 AgentX recipe, with DEP2 (DP attention, no TBO, plus expert
+# parallelism) from concurrency 64.
 # Five-token DSpark uses golden AL 3.51 for throughput and real acceptance for eval.
 dsv41flash-fp4-mi355x-atom-agentic-dspark:
-  image: rocm/atom-dev:nightly_202610121516
+  image: rocm/atom-dev:nightly_202610281457
   model: deepseek-ai/DeepSeek-V4.1-Flash
   model-prefix: dsv41flash
   runner: cluster:mi355x-amds
@@ -1642,8 +1644,9 @@ dsv41flash-fp4-mi355x-atom-agentic-dspark:
   scenarios:
     agentic-coding:
     - search-space:
-      - { tp: 2, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: draft_model, conc-list: [1, 2, 4, 6, 8, 16, 32, 64], srt-recipe: agentic.yaml }
-      - { tp: 4, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: draft_model, conc-list: [1, 8, 32], srt-recipe: agentic.yaml }
+      - { tp: 2, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: draft_model, conc-list: [1, 2, 4, 8, 16, 32, 64, 128, 256], srt-recipe: agentic.yaml }
+      - { tp: 2, ep: 2, dp-attn: true, kv-offloading: none, spec-decoding: draft_model, conc-list: [64, 128, 256], srt-recipe: agentic.yaml }
+      - { tp: 4, ep: 1, dp-attn: false, kv-offloading: none, spec-decoding: draft_model, conc-list: [1, 8], srt-recipe: agentic.yaml }
 
 # The golden acceptance curve remains the provisional GLM-5.2-derived curve
 # documented in infx/golden_al_distribution/glm5.3_mtp.yaml.

inferencex-e2e/perf-changelog.yaml

@@ -11499,3 +11499,15 @@
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4509
   scenario-type: [agentic-coding]
   append-only: true
+
+- config-keys:
+    - dsv41flash-fp4-mi355x-atom-agentic-dspark
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Update the image from rocm/atom-dev:nightly_202610121516 to nightly_202610281457 (digest sha256:96a9614276518c2619e88800dec5f276bc76a95691331b4abb2210184e1431af): nightly_202610271459 rebuilt on ATOM 646da00fe and AITER 11a9b4f5c. AITER includes ROCm/aiter#7085, which keeps the MX quant-sort and multi-phase MoE sort in bounds under num_local_tokens (plain TP2 hit an intermittent illegal memory access at concurrency 16 and above on nightly_202610261558), plus #7020 and #7068."
+    - "Switch the KV cache from BF16 to MXFP8 (kv_cache_dtype fp8, ROCm/ATOM#2725), as ATOM's own DSv4.1 AgentX CI runs it. The index cache stays FP4."
+    - "Run two shapes. Plain TP: TP2 at concurrency [1, 2, 4, 8, 16, 32, 64, 128, 256] and TP4 at [1, 8]. DEP2 (DP attention without TBO plus expert parallelism, the DEP2 cell of ROCm/ATOM#2784) at concurrency [64, 128, 256]: dp-load-balance least_tokens, ATOM_DP_SESSION_AFFINITY=1, ATOM_DP_LB_REQ_EQUIV=512, ATOM_ENABLE_PREFILL_DELAYER=1, ATOM_PREFILL_DECODE_INTERVAL=10, GPU_MAX_HW_QUEUES=5."
+    - "Drop the DP-attention points without expert parallelism (TP2 concurrency 32-256, TP4 concurrency 32); TP2 concurrency 32 and 64 move to plain TP."
+    - "Concurrency 1 (TP2 and TP4) follows the C1 recipe of ROCm/ATOM#2659: ATOM_ENABLE_PREFILL_DELAYER=1 with ATOM_PREFILL_DECODE_INTERVAL=1000000, max-num-seqs 2 and decode graphs at [1, 2]. Every other point runs ATOM's default max-num-seqs; the c256 points also capture 256."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3965