defer to aot eager instead of skip frame

eellison · eellison · commit fb3956c29b7b · 2025-05-12T13:37:55.000-07:00
ghstack-source-id: ce98f2d Pull Request resolved: #153409
diff --git a/torch/__init__.py b/torch/__init__.py
@@ -2290,7 +2290,7 @@ def compiled_with_cxx11_abi() -> builtins.bool:
 from torch.utils.dlpack import from_dlpack, to_dlpack
 
 
-def skip_frame_if_max_graphs() -> None:
+def has_hit_max_graphs() -> bool:
     """
     If we have hit a user specified max number of graphs, skip this frame.
     """
@@ -2301,8 +2301,8 @@ def skip_frame_if_max_graphs() -> None:
         return
 
     GraphsCompiledState.increment()
-    if GraphsCompiledState.get_num_graphs() > builtins.int(max_graphs):
-        raise torch._dynamo.exc.SkipFrame(f"Hit max graph limit: {max_graphs}")
+    return GraphsCompiledState.get_num_graphs() > builtins.int(max_graphs)
+        # raise torch._dynamo.exc.SkipFrame(f"Hit max graph limit: {max_graphs}")
 
 
 class _TorchCompileInductorWrapper:
@@ -2377,7 +2377,9 @@ def apply_options(self, options: _Optional[dict[str, _Any]]):
     def __call__(self, model_, inputs_):
         from torch._inductor.compile_fx import compile_fx
 
-        skip_frame_if_max_graphs()
+        if has_hit_max_graphs():
+            return _TorchCompileWrapper("aot_eager", "default", {}, self.dynamic).__call__(model_, inputs_)
+
         return compile_fx(model_, inputs_, config_patches=self.config)
 
     def get_compiler_config(self):
@@ -2423,7 +2425,6 @@ def __eq__(self, other):
         )
 
     def __call__(self, model_, inputs_):
-        skip_frame_if_max_graphs()
         return self.compiler_fn(model_, inputs_, **self.kwargs)
 
     def reset(self):
