diff --git a/vllm/model_executor/warmup/fa4_cutedsl_warmup.py b/vllm/model_executor/warmup/fa4_cutedsl_warmup.py index cbcca1f38c..7f1042338e 100644 --- a/vllm/model_executor/warmup/fa4_cutedsl_warmup.py +++ b/vllm/model_executor/warmup/fa4_cutedsl_warmup.py @@ -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