[Bugfix] Handle MLA fallback during FA4 JIT warmup (#49306)
This commit is contained in:
@@ -21,7 +21,11 @@ def fa4_cutedsl_warmup(worker: Worker) -> None:
|
||||
if not vllm_config.model_config.use_mla:
|
||||
return
|
||||
|
||||
backend_cls = get_mla_prefill_backend(vllm_config)
|
||||
try:
|
||||
backend_cls = get_mla_prefill_backend(vllm_config)
|
||||
except ValueError:
|
||||
# fall back to top-k MQA prefill path.
|
||||
return
|
||||
if backend_cls.get_name() != "FLASH_ATTN":
|
||||
return
|
||||
|
||||
|
||||
Reference in New Issue
Block a user