Fix issue #39191 respect accelerate config to disable torch.dynamo compilation

src/transformers/generation/utils.py

@@ -2140,6 +2140,25 @@ def _tensor_or_none(token, device=None):
         generation_config._pad_token_tensor = pad_token_tensor
         generation_config._decoder_start_token_tensor = decoder_start_token_tensor
 
+
+    def _is_dynamo_compilation_disabled(self):
+        """
+        Check if dynamo compilation should be disabled based on environment variables.
+        Uses standard environment variables to respect user's explicit intention to disable
+        torch.dynamo compilation.
+        """
+        disable_env_vars = [
+            "TORCHDYNAMO_DISABLE"  # Pytorch standard
+        ]
+
+        for env_var in disable_env_vars:
+            value = os.getenv(env_var, "").lower()
+            if value in ("1", "true", "yes", "on"):
+                return True
+
+        return False
+
+
     def _valid_auto_compile_criteria(self, model_kwargs: dict[str, Any], generation_config: GenerationConfig) -> bool:
         """
         Determines whether to trigger auto-compilation of the model's forward pass at generation time.
@@ -2148,6 +2167,10 @@ def _valid_auto_compile_criteria(self, model_kwargs: dict[str, Any], generation_
         if generation_config.disable_compile:
             return False
 
+        # honor accelerate config to disable dynamo
+        if self._is_dynamo_compilation_disabled():
+            return False
+
         # Base logic
         valid_hardware = self.device.type in ["cuda", "xpu"] or bool(
             generation_config.compile_config is not None and generation_config.compile_config._compile_all_devices