From 001e9af8413ed8d323b3cd0dd9e5e3925069e308 Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Tue, 8 Sep 2026 21:31:04 +0000 Subject: [PATCH] config(dsv4-fp4-b300-dynamo-sglang): pin SGLang v0.5.19-cu130 and refresh Dynamo MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Replace the removed nightly lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4 with the digest-pinned release lmsysorg/sglang:v0.5.19-cu130 in the master config and the five srt-slurm recipes it references. Bump the Dynamo source pin to b1c5147f6c9e so the worker imports against the v0.5.19 ServerArgs API, and switch the four DEP recipes from the deprecated SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS / _USE_MXF4_KIND env vars to --enable-w4a4-mxfp4-megamoe so MegaMoE keeps the FP4-activation path. 将已被删除的 nightly 镜像 lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4 替换为按 digest 固定的正式版 lmsysorg/sglang:v0.5.19-cu130，同时更新主配置和 其引用的五个 srt-slurm 配方。将 Dynamo 源码提交固定到 b1c5147f6c9e，以匹配 v0.5.19 的 ServerArgs API；四个 DEP 配方改用 --enable-w4a4-mxfp4-megamoe 取代已弃用的 SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS / _USE_MXF4_KIND 环境变量，保持 MegaMoE 的 FP4 激活路径。 Co-Authored-By: Claude Fable 5.1 --- .../deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml | 10 ++++------ .../deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4.yaml | 4 ++-- .../deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8.yaml | 10 ++++------ .../deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml | 10 ++++------ .../deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8.yaml | 10 ++++------ configs/nvidia-master.yaml | 2 +- 6 files changed, 19 insertions(+), 27 deletions(-) diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml index 42bd30e552..343b1c718a 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml @@ -2,11 +2,11 @@ name: "disagg-b300-1p1d-dep4-dep8" model: path: "deepseek-v4-pro" - container: "lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4" + container: "lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9" precision: "fp4" dynamo: - hash: "41882ae9b07232eed4850fb1daf8c958abb2556a" + hash: "b1c5147f6c9eb036d98e0e25e2a980375f491e78" install: true sbatch_directives: @@ -46,8 +46,6 @@ backend: SGLANG_DSV4_REASONING_EFFORT: "max" SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1" SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "8192" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1" SGLANG_OPT_USE_ONLINE_COMPRESS: "1" NCCL_MNNVL_ENABLE: "1" NCCL_CUMEM_ENABLE: "1" @@ -69,8 +67,6 @@ backend: SGLANG_DSV4_REASONING_EFFORT: "max" SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1" SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "1280" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1" SGLANG_OPT_USE_ONLINE_COMPRESS: "1" NCCL_MNNVL_ENABLE: "1" NCCL_CUMEM_ENABLE: "1" @@ -99,6 +95,7 @@ backend: enable-dp-attention: true moe-a2a-backend: "megamoe" + enable-w4a4-mxfp4-megamoe: true moe-dense-tp-size: 1 disaggregation-mode: "prefill" @@ -118,6 +115,7 @@ backend: stream-interval: 60 moe-a2a-backend: "megamoe" + enable-w4a4-mxfp4-megamoe: true moe-dense-tp-size: 1 disaggregation-mode: "decode" diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4.yaml index e55d8dc2f5..15e91ad869 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4.yaml @@ -2,11 +2,11 @@ name: "disagg-b300-1p1d-tp4-tp4" model: path: "deepseek-v4-pro" - container: "lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4" + container: "lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9" precision: "fp4" dynamo: - hash: "41882ae9b07232eed4850fb1daf8c958abb2556a" + hash: "b1c5147f6c9eb036d98e0e25e2a980375f491e78" install: true sbatch_directives: diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8.yaml index 4c5c5ef4db..fdcc9a0c58 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8.yaml @@ -2,11 +2,11 @@ name: "disagg-b300-2p1d-dep4-dep8" model: path: "deepseek-v4-pro" - container: "lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4" + container: "lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9" precision: "fp4" dynamo: - hash: "41882ae9b07232eed4850fb1daf8c958abb2556a" + hash: "b1c5147f6c9eb036d98e0e25e2a980375f491e78" install: true sbatch_directives: @@ -46,8 +46,6 @@ backend: SGLANG_DSV4_REASONING_EFFORT: "max" SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1" SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "8192" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1" SGLANG_OPT_USE_ONLINE_COMPRESS: "1" NCCL_MNNVL_ENABLE: "1" NCCL_CUMEM_ENABLE: "1" @@ -69,8 +67,6 @@ backend: SGLANG_DSV4_REASONING_EFFORT: "max" SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1" SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "1280" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1" SGLANG_OPT_USE_ONLINE_COMPRESS: "1" NCCL_MNNVL_ENABLE: "1" NCCL_CUMEM_ENABLE: "1" @@ -99,6 +95,7 @@ backend: enable-dp-attention: true moe-a2a-backend: "megamoe" + enable-w4a4-mxfp4-megamoe: true moe-dense-tp-size: 1 disaggregation-mode: "prefill" @@ -118,6 +115,7 @@ backend: stream-interval: 60 moe-a2a-backend: "megamoe" + enable-w4a4-mxfp4-megamoe: true moe-dense-tp-size: 1 disaggregation-mode: "decode" diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml index 8ad309af71..55258735bf 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml @@ -2,11 +2,11 @@ name: "disagg-b300-4p1d-dep4-dep8" model: path: "deepseek-v4-pro" - container: "lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4" + container: "lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9" precision: "fp4" dynamo: - hash: "41882ae9b07232eed4850fb1daf8c958abb2556a" + hash: "b1c5147f6c9eb036d98e0e25e2a980375f491e78" install: true sbatch_directives: @@ -46,8 +46,6 @@ backend: SGLANG_DSV4_REASONING_EFFORT: "max" SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1" SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "8192" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1" SGLANG_OPT_USE_ONLINE_COMPRESS: "1" NCCL_MNNVL_ENABLE: "1" NCCL_CUMEM_ENABLE: "1" @@ -69,8 +67,6 @@ backend: SGLANG_DSV4_REASONING_EFFORT: "max" SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1" SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "1280" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1" SGLANG_OPT_USE_ONLINE_COMPRESS: "1" NCCL_MNNVL_ENABLE: "1" NCCL_CUMEM_ENABLE: "1" @@ -99,6 +95,7 @@ backend: enable-dp-attention: true moe-a2a-backend: "megamoe" + enable-w4a4-mxfp4-megamoe: true moe-dense-tp-size: 1 disaggregation-mode: "prefill" @@ -118,6 +115,7 @@ backend: stream-interval: 60 moe-a2a-backend: "megamoe" + enable-w4a4-mxfp4-megamoe: true moe-dense-tp-size: 1 disaggregation-mode: "decode" diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8.yaml index 6ba7519714..5b118a2a69 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8.yaml @@ -2,11 +2,11 @@ name: "disagg-b300-6p1d-dep4-dep8" model: path: "deepseek-v4-pro" - container: "lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4" + container: "lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9" precision: "fp4" dynamo: - hash: "41882ae9b07232eed4850fb1daf8c958abb2556a" + hash: "b1c5147f6c9eb036d98e0e25e2a980375f491e78" install: true sbatch_directives: @@ -46,8 +46,6 @@ backend: SGLANG_DSV4_REASONING_EFFORT: "max" SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1" SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "8192" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1" SGLANG_OPT_USE_ONLINE_COMPRESS: "1" NCCL_MNNVL_ENABLE: "1" NCCL_CUMEM_ENABLE: "1" @@ -69,8 +67,6 @@ backend: SGLANG_DSV4_REASONING_EFFORT: "max" SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1" SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "1280" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: "1" - SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: "1" SGLANG_OPT_USE_ONLINE_COMPRESS: "1" NCCL_MNNVL_ENABLE: "1" NCCL_CUMEM_ENABLE: "1" @@ -99,6 +95,7 @@ backend: enable-dp-attention: true moe-a2a-backend: "megamoe" + enable-w4a4-mxfp4-megamoe: true moe-dense-tp-size: 1 disaggregation-mode: "prefill" @@ -118,6 +115,7 @@ backend: stream-interval: 60 moe-a2a-backend: "megamoe" + enable-w4a4-mxfp4-megamoe: true moe-dense-tp-size: 1 disaggregation-mode: "decode" diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 0cbe0e8c8f..8c8ce9582c 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -5233,7 +5233,7 @@ dsv4-fp4-b300-dynamo-vllm: # runners/launch_b300-nv.sh. Prefill workers are TP4/DEP4 (two workers per # 8-GPU B300 node); the decode worker is a single DEP8 node. dsv4-fp4-b300-dynamo-sglang: - image: lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4 + image: lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9 model: deepseek-ai/DeepSeek-V4-Pro model-prefix: dsv4 runner: b300