[Perf] Skip ll_bf16 router GEMM warmup for non-MoE models (#49659)
Signed-off-by: neweyes <328719365@qq.com>
This commit is contained in:
@@ -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
|
||||
|
||||
Reference in New Issue
Block a user