[Bugfix] Handle MLA fallback during FA4 JIT warmup (#49306)

This commit is contained in:
Roberto L. Castro
2026-07-21 13:58:05 +00:00
committed by GitHub
parent c67650f04b
commit 7bb49be4d1
@@ -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