From 98548a11e614556cdbec812a09829d682d051bca Mon Sep 17 00:00:00 2001
From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com>
Date: Sat, 26 Sep 2026 06:51:02 +0000
Subject: [PATCH] feat(qwen3.5): update B200 FP4 SGLang MTP image to
 v0.5.20-cu130
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Update qwen3.5-fp4-b200-sglang-mtp from lmsysorg/sglang:v0.5.14-cu130 to
lmsysorg/sglang:v0.5.20-cu130 (sglang 0.5.20, commit 94602c9c). SGLang v0.5.20
removed the deprecated --cuda-graph-max-bs alias, so the recipe passes the same
values through --cuda-graph-max-bs-decode, the field the alias already set.
Model, topology, MTP settings and all points are unchanged.

将 qwen3.5-fp4-b200-sglang-mtp 的镜像从 lmsysorg/sglang:v0.5.14-cu130 更新为
lmsysorg/sglang:v0.5.20-cu130（sglang 0.5.20，提交 94602c9c）。SGLang v0.5.20
移除了已弃用的 --cuda-graph-max-bs 别名，因此配方改用 --cuda-graph-max-bs-decode
传入相同数值，该别名原本即设置此字段。模型、拓扑、MTP 设置及所有测试点保持不变。

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 .../qwen3.5/sglang/b200-fp4-mtp/8k1k.yaml                 | 8 ++++----
 configs/nvidia-master.yaml                                | 2 +-
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp4-mtp/8k1k.yaml b/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp4-mtp/8k1k.yaml
index 8604043522..34d6a85962 100644
--- a/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp4-mtp/8k1k.yaml
+++ b/benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp4-mtp/8k1k.yaml
@@ -3,7 +3,7 @@ base:
   name: qwen3.5-fp4-b200-sglang-mtp-8k1k
   model:
     path: hf:nvidia/Qwen3.5-397B-A17B-NVFP4-V2
-    container: lmsysorg/sglang:v0.5.14-cu130
+    container: lmsysorg/sglang:v0.5.20-cu130
     precision: fp4
   resources:
     gpu_type: b200
@@ -33,7 +33,7 @@ base:
         mamba-ssm-dtype: bfloat16
         attention-backend: trtllm_mha
         moe-runner-backend: flashinfer_trtllm
-        cuda-graph-max-bs: 4
+        cuda-graph-max-bs-decode: 4
         max-running-requests: 4
         max-prefill-tokens: 16384
         chunked-prefill-size: 16384
@@ -70,7 +70,7 @@ zip_override_tp2_ep1:
     agg:
       gpus: 2
       args:
-        cuda-graph-max-bs: [4, 8, 16, 32, 64]
+        cuda-graph-max-bs-decode: [4, 8, 16, 32, 64]
         max-running-requests: [4, 8, 16, 32, 64]
         scheduler-recv-interval: [10, 30, 30, 30, 30]
         tensor-parallel-size: 2
@@ -82,7 +82,7 @@ zip_override_tp2_ep2:
     agg:
       gpus: 2
       args:
-        cuda-graph-max-bs: [16, 32, 64]
+        cuda-graph-max-bs-decode: [16, 32, 64]
         expert-parallel-size: 2
         linear-attn-backend: triton
         linear-attn-decode-backend: flashinfer
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 30ce89afcd..b91e8641e3 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -1218,7 +1218,7 @@ qwen3.5-fp4-b200-sglang:
         srt-recipe: benchmarks/single_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp4/8k1k.yaml
 
 qwen3.5-fp4-b200-sglang-mtp:
-  image: lmsysorg/sglang:v0.5.14-cu130
+  image: lmsysorg/sglang:v0.5.20-cu130
   model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
   model-prefix: qwen3.5
   runner: cluster:b200-nscale