From 96fa3f42c95acc75f40a8fddbc3a621c2355a30d Mon Sep 17 00:00:00 2001 From: neweyes <328719365@qq.com> Date: Mon, 27 Jul 2026 20:16:42 +0800 Subject: [PATCH] [Perf] Skip ll_bf16 router GEMM warmup for non-MoE models (#49659) Signed-off-by: neweyes <328719365@qq.com> --- vllm/model_executor/warmup/kernel_warmup.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm/model_executor/warmup/kernel_warmup.py b/vllm/model_executor/warmup/kernel_warmup.py index e21ab112e7..fae88eb7f4 100644 --- a/vllm/model_executor/warmup/kernel_warmup.py +++ b/vllm/model_executor/warmup/kernel_warmup.py @@ -122,7 +122,7 @@ def kernel_warmup(worker: "Worker"): elif has_flashinfer() and current_platform.has_device_capability(90): flashinfer_autotune(worker.model_runner) - if current_platform.has_device_capability(90): + if current_platform.has_device_capability(90) and worker.model_config.is_moe: _warmup_ll_bf16_router_gemm() # FlashInfer attention warmup