[Bugfix] Handle MLA fallback during FA4 JIT warmup (#49306)
Roberto L. Castro committed
Jul 21, 2026 at 15:58 UTC
7bb49be4d1160257f4a27d798d4ccadf5699d292
1 file changed
+5
-1
vllm/model_executor/warmup/fa4_cutedsl_warmup.py
+5
-1
@@ -21,7 +21,11 @@ def fa4_cutedsl_warmup(worker: Worker) -> None:
21
if not vllm_config.model_config.use_mla:
22
return
23
24
- backend_cls = get_mla_prefill_backend(vllm_config)
24
+ try:
25
+ backend_cls = get_mla_prefill_backend(vllm_config)
26
+ except ValueError:
27
+ # fall back to top-k MQA prefill path.
28
+ return
29
if backend_cls.get_name() != "FLASH_ATTN":
30
return
31