Switch Qwen3.5 FP4 SGLang configs to NVFP4 V2 / 将 Qwen3.5 FP4 SGLang 配置切换到 NVFP4 V2

configs/nvidia-master.yaml

@@ -1217,7 +1217,7 @@ qwen3.5-fp8-b200-sglang-agentic-mtp:
 
 qwen3.5-fp4-b200-sglang:
   image: lmsysorg/sglang:v0.5.14-cu130
-  model: nvidia/Qwen3.5-397B-A17B-NVFP4
+  model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
   model-prefix: qwen3.5
   runner: b200
   precision: fp4
@@ -1233,7 +1233,7 @@ qwen3.5-fp4-b200-sglang:
 
 qwen3.5-fp4-b200-sglang-mtp:
   image: lmsysorg/sglang:v0.5.14-cu130
-  model: nvidia/Qwen3.5-397B-A17B-NVFP4
+  model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
   model-prefix: qwen3.5
   runner: b200
   precision: fp4
@@ -1295,7 +1295,7 @@ qwen3.5-fp8-b300-sglang:
 
 qwen3.5-fp4-b300-sglang:
   image: lmsysorg/sglang:v0.5.14-cu130
-  model: nvidia/Qwen3.5-397B-A17B-NVFP4
+  model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
   model-prefix: qwen3.5
   runner: b300
   precision: fp4
@@ -1347,7 +1347,7 @@ qwen3.5-fp4-rtx6000pro-sglang-mtp:
 
 qwen3.5-fp4-b300-sglang-mtp:
   image: lmsysorg/sglang:v0.5.14-cu130
-  model: nvidia/Qwen3.5-397B-A17B-NVFP4
+  model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
   model-prefix: qwen3.5
   runner: b300
   precision: fp4

perf-changelog.yaml

@@ -5736,3 +5736,13 @@
     - "Upgrade SGLang from nightly-dev-cu13-20260711-7de33ce8 to nightly-dev-cu13-20260719-99f5a6f4, the latest nightly before the scheduler WAR-barrier regression in SGLang #31687"
     - "Restore the pre-watchdog AgentX concurrency grid and use AIPerf's default 300-second per-trajectory idle-gap cap"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2319
+
+- config-keys:
+    - qwen3.5-fp4-b200-sglang
+    - qwen3.5-fp4-b200-sglang-mtp
+    - qwen3.5-fp4-b300-sglang
+    - qwen3.5-fp4-b300-sglang-mtp
+  description:
+    - "Switch Qwen3.5 FP4 SGLang configs to the nvidia/Qwen3.5-397B-A17B-NVFP4-V2 checkpoint"
+    - "Model pre-staged on all clusters as Qwen3.5-397B-A17B-NVFP4-V2"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2205

runners/launch_b200-dgxc.sh

@@ -37,6 +37,16 @@ elif [[ $MODEL_PREFIX == "qwen3.5" && $PRECISION == "bf16" ]]; then
 elif [[ $MODEL_PREFIX == "qwen3.5" && $PRECISION == "fp8" ]]; then
     export MODEL_PATH="/lustre/fsw/models/Qwen3.5-397B-A17B-FP8"
     export SRT_SLURM_MODEL_PREFIX="qwen3.5-fp8"
+# qwen3.5 fp4 spans two checkpoints, so this must branch on the checkpoint and
+# not on MODEL_PREFIX+PRECISION alone: the sglang keys moved to NVFP4-V2 while
+# qwen3.5-fp4-b200-trt / -trt-mtp still declare plain NVFP4. Both share
+# model-prefix qwen3.5 + precision fp4 + runner b200, and further down this
+# script does `export MODEL="$MODEL_PATH"`, so a single shared branch would
+# serve V2 weights to the TRT configs while publishing them under the old
+# checkpoint name.
+elif [[ $MODEL_PREFIX == "qwen3.5" && $PRECISION == "fp4" && $MODEL == *NVFP4-V2 ]]; then
+    export MODEL_PATH="/scratch/fsw/models/Qwen3.5-397B-A17B-NVFP4-V2"
+    export SRT_SLURM_MODEL_PREFIX="qwen3.5-fp4"
 elif [[ $MODEL_PREFIX == "qwen3.5" && $PRECISION == "fp4" ]]; then
     export MODEL_PATH="/lustre/fsw/models/Qwen3.5-397B-A17B-NVFP4"
     export SRT_SLURM_MODEL_PREFIX="qwen3.5-fp4"

runners/launch_b300-nv.sh

@@ -400,6 +400,7 @@ else
         Qwen3.5-397B-A17B
         Qwen3.5-397B-A17B-FP8
         Qwen3.5-397B-A17B-NVFP4
+        Qwen3.5-397B-A17B-NVFP4-V2
         gpt-oss-120b
     )