set enforce eager

vllm/config.py

@@ -322,7 +322,8 @@ class ModelConfig:
 
         # Set enforce_eager to False if the value is unset.
         if self.enforce_eager is None:
-            self.enforce_eager = False
+            # self.enforce_eager = False
+            self.enforce_eager = True
 
         sliding_window = getattr(self.hf_text_config, "sliding_window", None)
         has_interleaved_attention = (sliding_window is not None) and (

vllm/engine/arg_utils.py

@@ -564,7 +564,8 @@ class EngineArgs:
                             'This should be a JSON string that will be '
                             'parsed into a dictionary.')
         parser.add_argument('--enforce-eager',
-                            action='store_true',
+                            # action='store_true',
+                            default=True,
                             help='Always use eager-mode PyTorch. If False, '
                             'will use eager mode and CUDA graph in hybrid '
                             'for maximal performance and flexibility.')