From b8b72a1aa35e3026d50fea1e9c68fa1f45f7e4c6 Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Wed, 28 Oct 2026 12:16:51 +0000 Subject: [PATCH 1/4] chore(glm5.2): update GB200 Dynamo+SGLang AgentX image to nightly-dev-20261028-586355a8 Move glm5.2-fp4-gb200-dynamo-sglang-agentic-agg and -disagg to the digest-pinned SGLang nightly 586355a8 and record its bundled sglang and ai-dynamo versions in the recipe identities. Co-Authored-By: Claude Opus 5.5 --- .../glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml | 4 ++-- .../glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml | 4 ++-- inferencex-e2e/configs/nvidia-master.yaml | 4 ++-- 3 files changed, 6 insertions(+), 6 deletions(-) diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml index 395f248c85..e512247391 100644 --- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml +++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml @@ -6,8 +6,8 @@ base: model: {} container: {} frameworks: - dynamo: 1.6.0.dev20261019 - sglang: 0.0.0.dev1+gab580b2e0 + dynamo: 1.7.0.dev20261027 + sglang: 0.0.0.dev1+g586355a83 resources: gpu_type: gb200 gpus_per_node: 4 diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml index 8dd9f2f9b8..26337c0edd 100644 --- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml +++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml @@ -6,8 +6,8 @@ base: model: {} container: {} frameworks: - dynamo: 1.6.0.dev20261019 - sglang: 0.0.0.dev1+gab580b2e0 + dynamo: 1.7.0.dev20261027 + sglang: 0.0.0.dev1+g586355a83 resources: gpu_type: gb200 gpus_per_node: 4 diff --git a/inferencex-e2e/configs/nvidia-master.yaml b/inferencex-e2e/configs/nvidia-master.yaml index dcc8d15576..b9c931ed63 100644 --- a/inferencex-e2e/configs/nvidia-master.yaml +++ b/inferencex-e2e/configs/nvidia-master.yaml @@ -7565,7 +7565,7 @@ glm5.2-fp4-b200-dynamo-sglang-agentic-disagg: dp-attn: true glm5.2-fp4-gb200-dynamo-sglang-agentic-agg: - image: lmsysorg/sglang:nightly-dev-20261020-ab580b2e@sha256:0af411bb320a95bb5e1c58b4c20c9852920a19aa9f12f148f69731b0dca3b576 + image: lmsysorg/sglang:nightly-dev-20261028-586355a8@sha256:b4f09a2ccbcfcc262d49600c3cd7f8581a793ea7d104947257b6a95500953b90 model: nvidia/GLM-5.2-NVFP4 model-prefix: glm5.2 runner: cluster:gb200-nv @@ -7606,7 +7606,7 @@ glm5.2-fp4-gb200-dynamo-sglang-agentic-agg: ep: 1 dp-attn: false glm5.2-fp4-gb200-dynamo-sglang-agentic-disagg: - image: lmsysorg/sglang:nightly-dev-20261020-ab580b2e@sha256:0af411bb320a95bb5e1c58b4c20c9852920a19aa9f12f148f69731b0dca3b576 + image: lmsysorg/sglang:nightly-dev-20261028-586355a8@sha256:b4f09a2ccbcfcc262d49600c3cd7f8581a793ea7d104947257b6a95500953b90 model: nvidia/GLM-5.2-NVFP4 model-prefix: glm5.2 runner: cluster:gb200-nv From 025528300e3a796e9674fc77eccf5427bdfdadda Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Wed, 28 Oct 2026 12:18:13 +0000 Subject: [PATCH 2/4] chore(glm5.2): add perf-changelog entry for GB200 Dynamo+SGLang AgentX image update Co-Authored-By: Claude Opus 5.5 --- inferencex-e2e/perf-changelog.yaml | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml index 907bd21213..1402f3e7c3 100644 --- a/inferencex-e2e/perf-changelog.yaml +++ b/inferencex-e2e/perf-changelog.yaml @@ -11489,3 +11489,11 @@ - "Update the MI355X DeepSeek-V4-Pro-0813 SGLang MoRI disaggregated AgentX serving image from lmsysorg/sglang-rocm:v0.5.22-rocm720-mi35x-20261019 to lmsysorg/sglang-rocm:v0.5.22-rocm10-mi35x-20261026, the latest published MI35x ROCm 10 nightly (digest sha256:edabad83728ab04e3eae39f315eb16261eeb64dbdb051112bd22fc88560b60c4)." - "Preserve all eight concurrency points (1, 16, 32, 48, 192, 256, 384, 512), prefill/decode topology, DSpark settings, UMBP offload, MoRI transfer, and benchmark client image. The workload binder supplies the serving image to every selected recipe variant." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4571 + +- config-keys: + - glm5.2-fp4-gb200-dynamo-sglang-agentic-agg + - glm5.2-fp4-gb200-dynamo-sglang-agentic-disagg + description: + - "Update the SGLang image from nightly-dev-20261020-ab580b2e to nightly-dev-20261028-586355a8." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4611 + scenario-type: [agentic-coding] From 8eda7811b599e6d298549e984f9ac0a6706f500e Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Wed, 28 Oct 2026 16:41:03 +0000 Subject: [PATCH 3/4] fix(glm5.2): pin Dynamo 68c35a1 for GB200 Dynamo+SGLang AgentX readiness Dynamo a789babbe37a (bundled in ai-dynamo 1.7.0.dev20261027) made the frontend /health omit the instances list that srtctl's readiness check reads, so workers never counted as ready. Pin its parent in both recipes. Co-Authored-By: Claude Opus 5.5 --- .../glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml | 8 ++++++-- .../glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml | 8 ++++++-- inferencex-e2e/perf-changelog.yaml | 2 +- 3 files changed, 13 insertions(+), 5 deletions(-) diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml index e512247391..158c016561 100644 --- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml +++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml @@ -6,13 +6,17 @@ base: model: {} container: {} frameworks: - dynamo: 1.7.0.dev20261027 + dynamo: 1.6.0 sglang: 0.0.0.dev1+g586355a83 resources: gpu_type: gb200 gpus_per_node: 4 + # Dynamo a789babbe37a made frontend /health omit instances, which srtctl's + # readiness check needs; pin its parent until srtctl asks for them. dynamo: - install: false + install: true + source: + rev: 68c35a1ce41b035d1a79fff4be97aa1c43de62c3 # Tachometer's per-node exporters slow decode steps and add host memory; # InferenceX does not read its output. observability: diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml index 26337c0edd..0ef5857048 100644 --- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml +++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml @@ -6,13 +6,17 @@ base: model: {} container: {} frameworks: - dynamo: 1.7.0.dev20261027 + dynamo: 1.6.0 sglang: 0.0.0.dev1+g586355a83 resources: gpu_type: gb200 gpus_per_node: 4 + # Dynamo a789babbe37a made frontend /health omit instances, which srtctl's + # readiness check needs; pin its parent until srtctl asks for them. dynamo: - install: false + install: true + source: + rev: 68c35a1ce41b035d1a79fff4be97aa1c43de62c3 # Tachometer's per-node exporters slow decode steps and add host memory; # InferenceX does not read its output. observability: diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml index 1402f3e7c3..c008f620da 100644 --- a/inferencex-e2e/perf-changelog.yaml +++ b/inferencex-e2e/perf-changelog.yaml @@ -11494,6 +11494,6 @@ - glm5.2-fp4-gb200-dynamo-sglang-agentic-agg - glm5.2-fp4-gb200-dynamo-sglang-agentic-disagg description: - - "Update the SGLang image from nightly-dev-20261020-ab580b2e to nightly-dev-20261028-586355a8." + - "Update the SGLang image from nightly-dev-20261020-ab580b2e to nightly-dev-20261028-586355a8 and pin Dynamo 68c35a1." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4611 scenario-type: [agentic-coding] From b9ec4567269028c87de570e3590497751ff0e87a Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Wed, 28 Oct 2026 17:41:52 +0000 Subject: [PATCH 4/4] fix(glm5.2): restore Dynamo admission limits for GB200 Dynamo+SGLang AgentX Dynamo 5df849a06a (#14369) gates each worker at max-running-requests plus a 2x queue and returns 529 beyond it, which aborted AgentX warmup. Set its documented admission env limits to the pre-engine-report defaults. Co-Authored-By: Claude Opus 5.5 --- .../glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml | 5 +++++ .../sglang/gb200-fp4/agentx/disagg-variants.yaml | 10 ++++++++++ 2 files changed, 15 insertions(+) diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml index 158c016561..cde607cf24 100644 --- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml +++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml @@ -44,6 +44,11 @@ base: PIP_BREAK_SYSTEM_PACKAGES: '1' DYN_SKIP_SGLANG_LOG_FORMATTING: '1' SGLANG_TIMEOUT_KEEP_ALIVE: '900' + # Dynamo 5df849a06a (#14369) caps each worker at max-running-requests plus a + # 2x queue and rejects the rest with 529; keep its pre-engine-report limits. + DYN_BACKEND_ADMISSION_ENGINE_REQUEST_LIMIT: '10000' + DYN_BACKEND_ADMISSION_ENGINE_REQUEST_WAITING_LIMIT: '10000' + DYN_BACKEND_ADMISSION_REQUEST_QUEUE_LIMIT: '20000' args: served-model-name: nvidia/GLM-5.2-NVFP4 trust-remote-code: true diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml index 0ef5857048..a4353a6650 100644 --- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml +++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml @@ -59,6 +59,11 @@ base: UCX_CUDA_IPC_ENABLE_MNNVL: 'y' # The default 9,9,10,10,10,10,10,10 split also gives the last stage the LM head and MTP draft layer. SGLANG_PP_LAYER_PARTITION: 10,10,10,10,10,10,10,8 + # Dynamo 5df849a06a (#14369) caps each worker at max-running-requests plus a + # 2x queue and rejects the rest with 529; keep its pre-engine-report limits. + DYN_BACKEND_ADMISSION_ENGINE_REQUEST_LIMIT: '10000' + DYN_BACKEND_ADMISSION_ENGINE_REQUEST_WAITING_LIMIT: '10000' + DYN_BACKEND_ADMISSION_REQUEST_QUEUE_LIMIT: '20000' args: served-model-name: nvidia/GLM-5.2-NVFP4 trust-remote-code: true @@ -111,6 +116,11 @@ base: NVSHMEM_REMOTE_TRANSPORT: none UCX_TLS: cuda_copy,cuda_ipc,sm,self,tcp UCX_CUDA_IPC_ENABLE_MNNVL: 'y' + # Dynamo 5df849a06a (#14369) caps each worker at max-running-requests plus a + # 2x queue and rejects the rest with 529; keep its pre-engine-report limits. + DYN_BACKEND_ADMISSION_ENGINE_REQUEST_LIMIT: '10000' + DYN_BACKEND_ADMISSION_ENGINE_REQUEST_WAITING_LIMIT: '10000' + DYN_BACKEND_ADMISSION_REQUEST_QUEUE_LIMIT: '20000' args: served-model-name: nvidia/GLM-5.2-NVFP4 trust-remote-code: true