From 7bb49be4d1160257f4a27d798d4ccadf5699d292 Mon Sep 17 00:00:00 2001 From: "Roberto L. Castro" <38211239+LopezCastroRoberto@users.noreply.github.com> Date: Tue, 21 Jul 2026 15:58:05 +0200 Subject: [PATCH] [Bugfix] Handle MLA fallback during FA4 JIT warmup (#49306) --- vllm/model_executor/warmup/fa4_cutedsl_warmup.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) 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