[torch.compile] Stop doing unnecessary FakeTensorProp in PiecewiseCompileInterpreter (#34093)
Signed-off-by: Richard Zou <zou3519@gmail.com>
This commit is contained in:
@@ -570,7 +570,9 @@ class PiecewiseCompileInterpreter(torch.fx.Interpreter): # type: ignore[misc]
|
||||
) -> Any:
|
||||
assert isinstance(target, str)
|
||||
|
||||
output = super().call_module(target, args, kwargs)
|
||||
gm = getattr(self.module, target)
|
||||
outputs = gm.graph.output_node().args[0]
|
||||
output = fx.map_arg(outputs, lambda node: node.meta["example_value"])
|
||||
|
||||
if target in self.compile_submod_names:
|
||||
index = self.compile_submod_names.index(target)
|
||||
|
||||
Reference in New Issue
Block a user