Relax aten.to restriction (pytorch#142420)

Summary:

if we have a.to(b), and b has a different dtype with a, then it must be a copy. In this case, we do not need to freeze the tensor. Instead, we use torch.ops.aten._assert_tensor_metadata.default to ensure that a must not have the same dtype as b.


We need to remove ` torch.ops.aten._assert_tensor_metadata.default` for executorch and mtia. 

Update to executorch pin to include pytorch/executorch#7277.

Reviewed By: bdhirsh, shruthign

Differential Revision: D66988295

Author: yushangdi

.ci/docker/ci_commit_pins/executorch.txt

@@ -1 +1 @@
-6f638937d64e3396793956d75ee3e14802022745
+d0580ddaf8bee29b863969eadc8fdca457804861

test/export/test_export.py

@@ -5108,6 +5108,29 @@ def forward(self, x):
         for op in ops:
             self.assertIn(op, (torch.ops.aten._to_copy.default,))
 
+    def test_float_conversion_from_int(self):
+        class Module(torch.nn.Module):
+            def forward(self, x):
+                return x.float()
+
+        ep = export(Module(), (torch.tensor(1, dtype=torch.int32),)).run_decompositions(
+            {}
+        )
+        ops = []
+        for node in ep.graph.nodes:
+            if node.op == "call_function":
+                ops.append(node.target)
+        self.assertGreater(len(ops), 0)
+        self.assertIn(torch.ops.aten._to_copy.default, ops)
+        self.assertIn(torch.ops.aten._assert_tensor_metadata.default, ops)
+
+        self.assertEqual(ep.module()(torch.tensor(1, dtype=torch.int32)), 1)
+
+        # Raises error because the input dtype is not the same as the input
+        # tensor when exporting.
+        with self.assertRaisesRegex(RuntimeError, "Tensor dtype mismatch!"):
+            ep.module()(torch.tensor(1, dtype=torch.float32))
+
     def test_device_to_mutation_float(self):
         class Module(torch.nn.Module):
             def forward(self, x):

torch/_subclasses/functional_tensor.py

@@ -535,7 +535,27 @@ def unwrap(x):
                             torch.ops.aten.dropout.default,
                             torch.ops.aten._to_copy.default,
                         ):
-                            torch._freeze_functional_tensor(outs_unwrapped)  # type: ignore[attr-defined]
+
+                            def must_copy():
+                                """
+                                Return True if the output of the op must be copied, not an alias
+                                """
+                                # output dtype is different from input
+                                return (
+                                    func == torch.ops.aten._to_copy.default
+                                    and "dtype" in kwargs
+                                    and kwargs["dtype"] != args_unwrapped[0].dtype
+                                )
+
+                            if must_copy():
+                                # We can further relax to args_unwrapped[0] != kwargs["dtype"], but I don't think
+                                # we have an aten op for that.
+                                torch.ops.aten._assert_tensor_metadata.default(
+                                    torch._from_functional_tensor(args_unwrapped[0]),
+                                    dtype=args_unwrapped[0].dtype,
+                                )
+                            else:
+                                torch._freeze_functional_tensor(outs_unwrapped)  # type: ignore[attr-defined]
                     outs_wrapped = pytree.tree_map_only(
                         torch.Tensor, wrap, outs_unwrapped
                     )