From 53483461cd7aace0c8712c2d5cf7ec4a2b3e819a Mon Sep 17 00:00:00 2001 From: Wenyao Gao Date: Sun, 27 Sep 2026 00:54:31 -0700 Subject: [PATCH] fix: disable GLM-5.2 B200 draft MoE FP8 conversion MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Set SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE=0 so the NextN/MTP draft keeps its shipped precision. 设置 SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE=0，使 NextN/MTP draft 保持原始发布精度。 --- .../glm5.2/sglang/b200-fp4/agentx/agg-variants.yaml | 2 +- .../glm5.2/sglang/b200-fp4/agentx/disagg-variants.yaml | 2 +- perf-changelog.yaml | 10 ++++++++++ 3 files changed, 12 insertions(+), 2 deletions(-) diff --git a/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/b200-fp4/agentx/agg-variants.yaml b/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/b200-fp4/agentx/agg-variants.yaml index 8453d5d8a0..baefd1fd34 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/b200-fp4/agentx/agg-variants.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/b200-fp4/agentx/agg-variants.yaml @@ -42,7 +42,7 @@ base: SGLANG_HICACHE_DEBUG_LOG: '1' SGLANG_HICACHE_DEBUG_SAMPLE_RATE: '16384' SGLANG_MOE_NVFP4_DISPATCH: '1' - SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE: '1' + SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE: '0' SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION: '8' PIP_BREAK_SYSTEM_PACKAGES: '1' args: diff --git a/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/b200-fp4/agentx/disagg-variants.yaml b/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/b200-fp4/agentx/disagg-variants.yaml index 02a84e30e1..d1c5719dc5 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/b200-fp4/agentx/disagg-variants.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/b200-fp4/agentx/disagg-variants.yaml @@ -111,7 +111,7 @@ base: SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK: '0' SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '512' SGLANG_MOE_NVFP4_DISPATCH: '1' - SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE: '1' + SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE: '0' SGLANG_DEFAULT_THINKING: '1' SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION: '8' SGLANG_REASONING_EFFORT: max diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 352c61fc0d..17926c7760 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -8986,3 +8986,13 @@ - "Hopper 使用 FA3，Blackwell FA4 fp8 descale 问题不适用，EAGLE3 draft 保持 attention_backend FLASH_ATTN。" - "This change does not alter the EAGLE3 draft model data type. The draft loads unmodified from the published Inferact/MiniMax-M3-EAGLE3-GQA checkpoint via --speculative-config (method=eagle3). kv-cache-dtype fp8 sets KV-cache storage precision, not the draft weights, and no flag overrides or re-quantizes the draft weights." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3503 + +- config-keys: + - glm5.2-fp4-b200-dynamo-sglang-agentic-agg + - glm5.2-fp4-b200-dynamo-sglang-agentic-disagg + scenario-type: + - agentic-coding + description: + - "Disable FP8 conversion of the GLM-5.2 NextN/MTP draft MoE on B200 AgentX so the draft runs at its shipped precision; images, topology, workload and acceptance are unchanged." + - "关闭 B200 GLM-5.2 AgentX 对 NextN/MTP draft MoE 的 FP8 转换，使 draft 保持原始发布精度；镜像、拓扑、工作负载和验收不变。" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3400