[Bugfix] Set enforce_eager automatically for mllama (#12127)

Signed-off-by: Chen Zhang <[email protected]>

Author: heheda12345

examples/offline_inference/vision_language.py

@@ -325,7 +325,6 @@ def run_mllama(question: str, modality: str):
         model=model_name,
         max_model_len=4096,
         max_num_seqs=16,
-        enforce_eager=True,
         disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
     )
 

examples/offline_inference/vision_language_multi_image.py

@@ -186,7 +186,6 @@ def load_mllama(question, image_urls: List[str]) -> ModelRequestData:
         model=model_name,
         max_model_len=4096,
         max_num_seqs=16,
-        enforce_eager=True,
         limit_mm_per_prompt={"image": len(image_urls)},
     )
 

vllm/config.py

@@ -607,10 +607,12 @@ def _verify_cuda_graph(self) -> None:
         self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
                                           self.max_model_len)
 
-        if (self.hf_config.model_type == 'deepseek_v3'
+        MODEL_NOT_SUPPORT_CUDA_GRAPH = ['deepseek_v3', 'mllama']
+        if (self.hf_config.model_type in MODEL_NOT_SUPPORT_CUDA_GRAPH
                 and not self.enforce_eager):
-            logger.warning("CUDA graph is not supported for Deepseek V3 yet, "
-                           "fallback to the eager mode.")
+            logger.warning(
+                "CUDA graph is not supported for %s yet, fallback to the eager "
+                "mode.", self.hf_config.model_type)
             self.enforce_eager = True
 
     def _verify_bnb_config(self) -> None: