[export] Support preserving submodule callling convention in non-strict export (#117796)

Summary: Title

Test Plan: CI

Reviewed By: zhxchen17

Differential Revision: D52889236

Pull Request resolved: https://github.com/pytorch/pytorch/pull/117796
Approved by: https://github.com/angelayi

test/export/test_unflatten.py

@@ -30,7 +30,7 @@ from torch._export.utils import (
 from torch.export import Constraint, Dim, export
 from torch.fx.experimental.proxy_tensor import make_fx
 from torch.testing import FileCheck
-from torch.testing._internal.common_utils import run_tests, TestCase
+from torch.testing._internal.common_utils import skipIfTorchDynamo, run_tests, TestCase
 from torch.utils._pytree import (
     LeafSpec,
     tree_flatten,
@@ -199,6 +199,7 @@ class TestUnflatten(TestCase):
             id(getattr(unflattened_module.sub_net, "2")),
         )
 
+    @skipIfTorchDynamo("Non strict mode is not meant to run with dynamo")
     def test_unflatten_preserve_signature(self):
         class NestedChild(torch.nn.Module):
             def forward(self, zx, y):
@@ -234,41 +235,43 @@ class TestUnflatten(TestCase):
 
         orig_eager = MyModule()
         inps = torch.rand(2, 3), torch.rand(2, 3)
-        export_module = export(
-            orig_eager,
-            inps,
-            {},
-            preserve_module_call_signature=("foo.nested",),
-        )
-        unflattened = unflatten(export_module)
-        self.compare_outputs(export_module, unflattened, inps)
-        unflattened.foo.nested = NestedChild()
-        self.compare_outputs(export_module, unflattened, inps)
+        for strict in [True, False]:
+            export_module = export(
+                orig_eager,
+                inps,
+                {},
+                preserve_module_call_signature=("foo.nested",),
+                strict=strict
+            )
+            unflattened = unflatten(export_module)
+            self.compare_outputs(export_module, unflattened, inps)
+            unflattened.foo.nested = NestedChild()
+            self.compare_outputs(export_module, unflattened, inps)
 
-        # Test tree spec mismatched input
-        orig_outs = export_module(*inps)
-        new_inps = *inps, torch.rand(2, 3)
-        with self.assertRaisesRegex(
-            TypeError,
-            "There is no flat args adapter sepcified. Are you sure you are calling this with the right arguments?",
-        ):
-            unflattened(new_inps)
+            # Test tree spec mismatched input
+            orig_outs = export_module(*inps)
+            new_inps = *inps, torch.rand(2, 3)
+            with self.assertRaisesRegex(
+                TypeError,
+                "There is no flat args adapter sepcified. Are you sure you are calling this with the right arguments?",
+            ):
+                unflattened(new_inps)
 
-        # With flat args adapter
-        class KeepTwoFlatArgsAdapter(FlatArgsAdapter):
-            def adapt(
-                self,
-                target_spec: TreeSpec,
-                input_spec: TreeSpec,
-                input_args: List[Any],
-            ) -> List[Any]:
-                while len(input_args) > 2:
-                    input_args.pop(-1)
-                return input_args
+            # With flat args adapter
+            class KeepTwoFlatArgsAdapter(FlatArgsAdapter):
+                def adapt(
+                    self,
+                    target_spec: TreeSpec,
+                    input_spec: TreeSpec,
+                    input_args: List[Any],
+                ) -> List[Any]:
+                    while len(input_args) > 2:
+                        input_args.pop(-1)
+                    return input_args
 
-        unflattened = unflatten(export_module, KeepTwoFlatArgsAdapter())
-        new_outs = unflattened(*new_inps)
-        self.assertTrue(torch.allclose(orig_outs, new_outs))
+            unflattened = unflatten(export_module, KeepTwoFlatArgsAdapter())
+            new_outs = unflattened(*new_inps)
+            self.assertTrue(torch.allclose(orig_outs, new_outs))
 
     def test_unflatten_param_list_dict(self):
         class Mod(torch.nn.Module):

torch/_export/wrappers.py

@@ -61,7 +61,9 @@ def _wrap_submodule(mod, path, module_call_specs):
         submodule = getattr(submodule, name)
 
     def update_module_call_signatures(path, in_spec, out_spec):
-        assert path not in module_call_specs
+        if path in module_call_specs:
+            assert module_call_specs[path]["in_spec"] == in_spec
+            assert module_call_specs[path]["out_spec"] == out_spec
         module_call_specs[path] = {"in_spec": in_spec, "out_spec": out_spec}
 
     assert "forward" not in submodule.__dict__

torch/export/_trace.py

@@ -517,12 +517,24 @@ def _export(
     constraints = constraints or []
     kwargs = kwargs or {}
 
+    flat_args, orig_in_spec = pytree.tree_flatten((args, kwargs))
+
     if not strict:
         assert isinstance(f, torch.nn.Module)
-        assert len(preserve_module_call_signature) == 0
         assert len(kwargs) == 0, "keyword arguments NYI"
         out_spec = None
 
+        module_call_specs: Dict[str, Dict[str, pytree.TreeSpec]] = {}
+
+        def strip_root(x):
+            if isinstance(x, str) and x.startswith("_export_root"):
+                stripped = x[len("_export_root") :]
+                return stripped[1:] if stripped.startswith(".") else stripped
+            return x
+
+        def fixup_key(x):
+            return "L__self__" + strip_root(x)
+
         def _tuplify_outputs(aot_export):
             def _aot_export_non_strict(mod, args, **kwargs):
                 class Wrapper(torch.nn.Module):
@@ -537,16 +549,16 @@ def _export(
                         )
                         return tuple(flat_outs)
 
-                gm, sig = aot_export(Wrapper(mod), args, **kwargs)
-
-                def strip_root(x):
-                    if isinstance(x, str) and x.startswith("_export_root"):
-                        stripped = x[len("_export_root") :]
-                        return stripped[1:] if stripped.startswith(".") else stripped
-                    return x
-
-                def fixup_key(x):
-                    return "L__self__" + strip_root(x)
+                wrapped_mod = Wrapper(mod)
+                # Patch export_root to the signatures so that wrapper module correctly populates the
+                # in/out spec
+                new_preserved_call_signatures = [
+                    "_export_root." + i for i in preserve_module_call_signature
+                ]
+                with _wrap_submodules(
+                    wrapped_mod, new_preserved_call_signatures, module_call_specs
+                ):
+                    gm, sig = aot_export(wrapped_mod, args, **kwargs)
 
                 sig.parameters = pytree.tree_map(strip_root, sig.parameters)
                 sig.buffers = pytree.tree_map(strip_root, sig.buffers)
@@ -585,9 +597,39 @@ def _export(
             fake_mode, src_equalities, original_signature, ep_non_strict.gm
         )
         assert out_spec is not None
+
+        gm = ep_non_strict.gm
+
+        module_call_signatures = {
+            strip_root(fqn): ModuleCallSignature(inputs=[], outputs=[], **specs)
+            for fqn, specs in module_call_specs.items()
+        }
+
+        if len(preserve_module_call_signature) > 0:
+            for node in gm.graph.nodes:
+                if node.target == torch.ops.higher_order._export_tracepoint:
+                    if "path" in node.kwargs:
+                        path = strip_root(node.kwargs["path"])
+                        with gm.graph.inserting_before(node):
+                            new_node = gm.graph.create_node(
+                                "call_function",
+                                torch.ops.higher_order._export_tracepoint,
+                                args=node.args,
+                                kwargs={
+                                    "path": path,
+                                    "kind": node.kwargs["kind"],
+                                },
+                            )
+                            node.replace_all_uses_with(new_node)
+                            gm.graph.erase_node(node)
+
+            res = CollectTracepointsPass(module_call_signatures, ep_non_strict.sig)(gm)
+            assert res is not None
+            gm = res.graph_module
+
         return ExportedProgram(
-            root=ep_non_strict.gm,
-            graph=ep_non_strict.gm.graph,
+            root=gm,
+            graph=gm.graph,
             graph_signature=ep_non_strict.sig,
             state_dict=_get_params_buffers(f),
             range_constraints=range_constraints,
@@ -595,9 +637,12 @@ def _export(
                 ModuleCallEntry(
                     "",
                     ModuleCallSignature(
-                        [], [], pytree.tree_flatten((args, {}))[1], out_spec
+                        inputs=[], outputs=[], in_spec=orig_in_spec, out_spec=out_spec
                     ),
                 )
+            ]
+            + [
+                ModuleCallEntry(fqn, sig) for fqn, sig in module_call_signatures.items()
             ],
             example_inputs=(args, kwargs),
             constants=ep_non_strict.constants,
@@ -768,7 +813,6 @@ def _export(
         ),
         len(export_graph_signature.input_specs),
     )
-    flat_args, orig_in_spec = pytree.tree_flatten((args, kwargs))
     range_constraints = _process_constraints(
         gm,
         num_lifted,