[Perf] Skip ll_bf16 router GEMM warmup for non-MoE models (#49659)
Signed-off-by: neweyes <328719365@qq.com>
neweyes committed
Jul 27, 2026 at 20:16 UTC
96fa3f42c95acc75f40a8fddbc3a621c2355a30d
1 file changed
+1
-1
vllm/model_executor/warmup/kernel_warmup.py
+1
-1
@@ -122,7 +122,7 @@ def kernel_warmup(worker: "Worker"):
122
elif has_flashinfer() and current_platform.has_device_capability(90):
123
flashinfer_autotune(worker.model_runner)
124
125
- if current_platform.has_device_capability(90):
125
+ if current_platform.has_device_capability(90) and worker.model_config.is_moe:
126
_warmup_ll_bf16_router_gemm()
127
128
# FlashInfer attention warmup