[compile] Add fallback path to AOT compile when serialization fails.

Summary:
Fixing issue vllm-project#27348

For dynamo caching, it's possible that the compilation succeeds but
the serialization step fails. In this case, the failure of serialization
step shouldn't block user from getting compilation results correctly.

Therefore we add a handling of the serialization error and only
give warning when model saving fails. When saving fails, VLLM model
runner should be able to just fallback to the old path, and in the
next process, it will fail to load dynamo cache but still fallback
to retracing with dynamo + loading inductor cache, which is the same
behavior to AOT compile turned of off.

This is mostly a short term fix and in the long term we should resolve
the serialization bugs by eliminating pickling of graph modules.

i.e. Once vllm-project#25205 is merged,
we should be able to resolve the issue at a lower level.

Test Plan:

pytest tests/lora/test_quant_model.py

Reviewers:

Subscribers:

Tasks:

Tags:

Signed-off-by: zhxchen17 <zhxchen17@fb.com>

Author: zhxchen17

vllm/compilation/decorators.py

@@ -402,8 +402,17 @@ def patched_inline_call(self_):
                     output = self.aot_compiled_fn(self, *args, **kwargs)
                     assert aot_compilation_path is not None
                     assert cache_dir is not None
-                    os.makedirs(cache_dir, exist_ok=True)
-                    self.aot_compiled_fn.save_compiled_function(aot_compilation_path)
+                    try:
+                        os.makedirs(cache_dir, exist_ok=True)
+                        self.aot_compiled_fn.save_compiled_function(
+                            aot_compilation_path
+                        )
+                    except Exception as e:
+                        logger.warning(
+                            "Cannot save aot compilation to path %s, error: %s",
+                            aot_compilation_path,
+                            str(e),
+                        )
                 else:
                     output = self.compiled_callable(*args, **kwargs)
             return output