From 305a1ccc39750044a6f9c5471998d63eb1550966 Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Sat, 24 Oct 2026 18:38:32 +0000 Subject: [PATCH 1/2] chore(glm5.2): update B200 FP8 SGLang AgentX MTP image to nightly-dev-20261024-5994c559 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 将 GLM-5.2 FP8 B200 SGLang AgentX MTP 镜像更新至 nightly-dev-20261024-5994c559。 Co-Authored-By: Claude Opus 5.5 --- inferencex-e2e/configs/nvidia-master.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/inferencex-e2e/configs/nvidia-master.yaml b/inferencex-e2e/configs/nvidia-master.yaml index 5ef2f719c1..57e4e77692 100644 --- a/inferencex-e2e/configs/nvidia-master.yaml +++ b/inferencex-e2e/configs/nvidia-master.yaml @@ -7335,7 +7335,7 @@ glm5.2-fp4-b200-sglang-agentic-mtp: - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [1, 4, 8, 12, 16], srt-recipe: agentic.yaml } glm5.2-fp8-b200-sglang-agentic-mtp: - image: lmsysorg/sglang:nightly-dev-20261020-ab580b2e@sha256:0af411bb320a95bb5e1c58b4c20c9852920a19aa9f12f148f69731b0dca3b576 + image: lmsysorg/sglang:nightly-dev-20261024-5994c559@sha256:40da9cac78a558237426971c048ab0f4f1a4b13bfb1b6e7a48d4d980e4163c8e model: zai-org/GLM-5.2-FP8 model-prefix: glm5.2 runner: cluster:b200-nscale From 74754a881ac6a77a3c5df6125c287edcbb830636 Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Sat, 24 Oct 2026 19:53:42 +0000 Subject: [PATCH 2/2] fix(glm5.2): keep NCCL graph registration on for B200 FP8 SGLang AgentX MTP MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit SGLang 5994c559 forces NCCL_GRAPH_REGISTER=0 for any multi-rank graph capture; with it the DCP8 c24 GSM8K eval returned only '!' for 68 of 1319 questions. Set it explicitly to 1, which the image honors. SGLang 5994c559 对任何多卡 graph 捕获强制设置 NCCL_GRAPH_REGISTER=0; 在此设置下 DCP8 c24 GSM8K 评测中 1319 道题有 68 道只输出 '!'。 显式设置为 1,镜像会遵循该设置。 Co-Authored-By: Claude Opus 5.5 --- .../srt-slurm-recipes/glm5.2/sglang/b200-fp8-mtp/agentic.yaml | 2 ++ 1 file changed, 2 insertions(+) diff --git a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/glm5.2/sglang/b200-fp8-mtp/agentic.yaml b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/glm5.2/sglang/b200-fp8-mtp/agentic.yaml index 0674301c22..28cffe5d4a 100644 --- a/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/glm5.2/sglang/b200-fp8-mtp/agentic.yaml +++ b/inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/glm5.2/sglang/b200-fp8-mtp/agentic.yaml @@ -44,6 +44,8 @@ base: SGLANG_CACHE_DIR: /tmp/sglang-cache # Outlast AIPerf's pooled keep-alive connections. SGLANG_TIMEOUT_KEEP_ALIVE: '900' + # SGLang otherwise forces 0 for multi-rank graph capture, under which DCP8 GSM8K emitted only '!'. + NCCL_GRAPH_REGISTER: '1' benchmark: env: # Warmup can leave bytes unacknowledged past AIPerf's 30 s default.