[export] Turn off output value from sources for export. (#115442)

Fixes #ISSUE_NUMBER

Pull Request resolved: https://github.com/pytorch/pytorch/pull/115442
Approved by: https://github.com/tugsbayasgalan

test/export/test_export.py

@@ -1562,6 +1562,32 @@ def forward(self, l_x_):
         with self.assertRaisesRegex(RuntimeError, "shape\[0\] is specialized at 4"):
             ep_v2(*test_inp)
 
+    def test_constant_output(self):
+        class ModuleConstant(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.b = torch.randn(3, 2)
+
+            def forward(self):
+                return self.b
+
+        class ModuleNestedConstant(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.bff = torch.randn(3, 2)
+
+            def forward(self, x, y):
+                return {"prediction": (x + y, self.bff)}
+
+        mod = ModuleConstant()
+        ep = torch.export.export(mod, ())
+        self.assertEqual(ep(), mod())
+
+        args = (torch.randn(3, 2), torch.randn(3, 2))
+        mod = ModuleNestedConstant()
+        ep = torch.export.export(mod, args)
+        self.assertEqual(ep(*args), mod(*args))
+
     def test_non_arg_name_dynamic_shapes_api_with_kwarg(self):
         def foo(a, b, kw1, kw2):
             return a.sum() + b.sum() + kw1.sum() - kw2.sum()

torch/_dynamo/codegen.py

@@ -60,14 +60,18 @@ class PyCodegen:
         self.cell_and_freevars = self.tx.cell_and_freevars
         self.new_var = self.tx.output.new_var
         self.mutable_side_effects_from_source = False
+        self.value_from_source: bool = True
 
-    def restore_stack(self, stack_values):
+    def restore_stack(self, stack_values, *, value_from_source=True):
         prior = self.mutable_side_effects_from_source
         self.mutable_side_effects_from_source = True
+        prev = self.value_from_source
+        self.value_from_source &= value_from_source
         try:
             self.foreach(stack_values)
         finally:
             self.mutable_side_effects_from_source = prior
+            self.value_from_source = prev
 
     def graph_output_vars(self):
         return [x.variable for x in self.graph_outputs.values()]
@@ -108,7 +112,7 @@ class PyCodegen:
                 self.top_of_stack = value
                 return
 
-        if value.source is not None and allow_cache:
+        if value.source is not None and allow_cache and self.value_from_source:
             output.extend(value.source.reconstruct(self))
         elif value.is_python_constant() and is_safe_constant(
             value.as_python_constant()

torch/_dynamo/output_graph.py

@@ -911,7 +911,7 @@ class OutputGraph(Checkpointable[OutputGraphState]):
             pass1 = PyCodegen(tx, root, graph_output_var)
             self.side_effects.codegen_hooks(pass1)
             self.side_effects.codegen_save_tempvars(pass1)
-            pass1.restore_stack(stack_values)
+            pass1.restore_stack(stack_values, value_from_source=not tx.export)
             self.side_effects.codegen_update_mutated(pass1)
 
             # one more time now that we have established tempvars
@@ -923,7 +923,7 @@ class OutputGraph(Checkpointable[OutputGraphState]):
             )
             self.side_effects.codegen_hooks(pass2)
             self.side_effects.codegen_save_tempvars(pass2)
-            pass2.restore_stack(stack_values)
+            pass2.restore_stack(stack_values, value_from_source=not tx.export)
             self.side_effects.codegen_update_mutated(pass2)
 
             output = []

torch/_export/passes/lift_constant_tensor_pass.py

@@ -23,14 +23,13 @@ def lift_constant_tensor_pass(gm, graph_signature) -> Dict[str, torch.Tensor]:
     )
     assert fake_mode is not None
 
-    first_user_input_loc, first_user_input = None, None
+    first_user_input_loc, first_user_input = 0, None
-    for i, node in enumerate(gm.graph.nodes):
+    for node in gm.graph.nodes:
         if node.op == "placeholder" and node.name in graph_signature.user_inputs:
             first_user_input = node
-            first_user_input_loc = i
             break
+        first_user_input_loc += 1
 
-    assert first_user_input is not None and first_user_input_loc is not None
     tensor_constants = {}
 
     for node in gm.graph.nodes:

torch/export/_trace.py

@@ -19,10 +19,11 @@ from torch._export.passes.lift_constant_tensor_pass import lift_constant_tensor_
 from torch._export.wrappers import _wrap_submodules
 from torch._functorch.aot_autograd import aot_export_module, GraphSignature
 from torch._guards import detect_fake_mode
-from torch._subclasses.fake_tensor import FakeTensor
+from torch._subclasses.fake_tensor import FakeTensor, FakeTensorMode
 from torch.fx.experimental.symbolic_shapes import (
     ConstraintViolationError,
     GuardOnDataDependentSymNode,
+    ShapeEnv,
 )
 from torch.fx.graph import _PyTreeCodeGen, _PyTreeInfo
 from torch.utils._sympy.value_ranges import ValueRangeError
@@ -57,16 +58,8 @@ DEFAULT_EXPORT_DYNAMO_CONFIG = ExportDynamoConfig()
 
 
 def _convert_input_to_fake(gm, args, kwargs):
-    if (
+    params_buffers = _get_params_buffers(gm)
-        len(args) == 0
-        and len(kwargs) == 0
-        and len(dict(gm.named_parameters())) == 0
-        and len(dict(gm.named_buffers())) == 0
-    ):
-        return [], {}, {}, None
-
     fake_inps: List[torch.Tensor] = []
-    fake_mode = None
     for node in gm.graph.nodes:
         if node.op == "placeholder" and "val" in node.meta:
             fake_val = node.meta["val"]
@@ -75,10 +68,11 @@ def _convert_input_to_fake(gm, args, kwargs):
 
     if detected_fake_mode := detect_fake_mode(fake_inps):
         fake_mode = detected_fake_mode
+    else:
+        fake_mode = FakeTensorMode(shape_env=ShapeEnv())
 
-    assert (
+    if len(args) == 0 and len(kwargs) == 0:
-        fake_mode is not None
+        return (), {}, params_buffers, fake_mode
-    ), "Cannot find fake_mode attatched to the graph's placeholders."
 
     count = 0
 
@@ -94,10 +88,7 @@ def _convert_input_to_fake(gm, args, kwargs):
     fake_params_buffers = pytree.tree_map_only(
         torch.Tensor,
         functools.partial(fake_mode.from_tensor, static_shapes=True),
-        {
+        params_buffers,
-            **dict(gm.named_parameters(remove_duplicate=False)),
-            **dict(gm.named_buffers(remove_duplicate=False)),
-        },
     )
     return fake_args, fake_kwargs, fake_params_buffers, fake_mode
 
@@ -653,8 +644,6 @@ def _export(
     # The unbacked symint symbols are updated in aot_export
     # so we serialize them here instead of inside dynamo
 
-    # dynamo_fake_mode can be None if there's no placeholder in gm_torch_level
-    if dynamo_fake_mode:
     gm.meta["inline_constraints"] = {
         k: v
         for k, v in dynamo_fake_mode.shape_env.runtime_var_to_range.items()
@@ -667,7 +656,7 @@ def _export(
             for i, s in enumerate(export_graph_signature.input_specs)
             if s.kind == InputKind.USER_INPUT
         ),
-        0,
+        len(export_graph_signature.input_specs),
     )
     flat_args, orig_in_spec = pytree.tree_flatten((args, kwargs))
     range_constraints, equality_constraints = _process_constraints(