Dont clone inputs if using fake tensor (pytorch#88208)

Not sure that this will really reduce memory use but it is an extraneous copy in our stack right now.
Pull Request resolved: pytorch#88208
Approved by: https://github.com/anijain2305

torch/_dynamo/optimizations/analysis.py

@@ -121,10 +121,6 @@ def has_mutation(gm, example_inputs, inputs_only=False):
     true, we only check for mutation of inputs"""
     # TODO - moco gives bad accuracy with Aliasing. gm is getting mutated in a bad way.
 
-    # Clone the inputs such that intermediate tensors (not leaf tensors) with
-    # requires_grad to True are now converted to False to avoid Runtime Error
-    # like "leaf variable that requires grad is inplace modified"
-    example_inputs = clone_inputs(example_inputs)
     if fake_tensors_available and config.fake_tensor_propagation:
         with FakeTensorMode() as fake_mode:
             pass
@@ -134,6 +130,10 @@ def has_mutation(gm, example_inputs, inputs_only=False):
         with fake_mode.restore() if hasattr(fake_mode, "restore") else fake_mode:
             ShapeAliasingAndMutationProp(new_gm).run(*example_inputs)
     else:
+        # Clone the inputs such that intermediate tensors (not leaf tensors) with
+        # requires_grad to True are now converted to False to avoid Runtime Error
+        # like "leaf variable that requires grad is inplace modified"
+        example_inputs = clone_inputs(example_inputs)
         new_gm = copy.deepcopy(gm)
         example_inputs = copy.deepcopy(example_inputs)
         ShapeAliasingAndMutationProp(new_gm).run(*example_inputs)