[dynamo] Mark a vt unspecialized nn module variable source earlier (pytorch#154780)

I am working on providing some skip guard helper functions to allow users to reduce guard overhead. This is a refactor to allow that.

Pull Request resolved: pytorch#154780
Approved by: https://github.com/StrongerXi, https://github.com/jansel

Author: anijain2305

benchmarks/dynamo/ci_expected_accuracy/cpu_inductor_torchbench_inference.csv

@@ -138,7 +138,7 @@ hf_Bert_large,pass,0
 
 
 
-hf_BigBird,pass,18
+hf_BigBird,pass,24
 
 
 

benchmarks/dynamo/ci_expected_accuracy/dynamic_cpu_inductor_torchbench_inference.csv

@@ -122,7 +122,7 @@ hf_Bert_large,pass,0
 
 
 
-hf_BigBird,pass,18
+hf_BigBird,pass,24
 
 
 

test/dynamo/test_modules.py

@@ -1299,6 +1299,7 @@ def test_unsupportedmodule(self):
         self.assertTrue(torch._dynamo.testing.same(r, m(i)))
         self.assertEqual(cnt.op_count, 6)
 
+    @patch.object(torch._dynamo.config, "allow_unspec_int_on_nn_module", True)
     def test_self_mutating1(self):
         m1 = torch.nn.Linear(10, 10)
         m2 = SelfMutatingModule(m1)

test/functorch/test_control_flow.py

@@ -7454,14 +7454,13 @@ def forward(self, a, b):
         self.assertExpectedInline(
             backend.graphs[0].code.strip(),
             """\
-def forward(self, s97 : torch.SymInt, L_a_ : torch.Tensor, L_b_ : torch.Tensor, L_self_num : torch.SymInt):
+def forward(self, s97 : torch.SymInt, L_a_ : torch.Tensor, L_b_ : torch.Tensor):
     l_a_ = L_a_
     l_b_ = L_b_
-    l_self_num = L_self_num
     tensor = torch.tensor([True])
     cond_true_0 = self.cond_true_0
     cond_false_0 = self.cond_false_0
-    cond = torch.ops.higher_order.cond(tensor, cond_true_0, cond_false_0, (l_a_, l_b_, l_self_num, s97));  tensor = cond_true_0 = cond_false_0 = l_a_ = l_b_ = l_self_num = s97 = None
+    cond = torch.ops.higher_order.cond(tensor, cond_true_0, cond_false_0, (l_a_, l_b_, s97));  tensor = cond_true_0 = cond_false_0 = l_a_ = l_b_ = s97 = None
     getitem = cond[0];  cond = None
     return (getitem,)""",  # noqa: B950
         )

torch/_dynamo/utils.py

@@ -2402,6 +2402,10 @@ def is_int_specialization_case(value, source):
             source.guard_source().is_unspecialized_builtin_nn_module()
             and not config.allow_unspec_int_on_nn_module
         )
+        or (
+            source.guard_source().is_unspecialized_nn_module()
+            and not config.allow_unspec_int_on_nn_module
+        )
         or is_from_defaults(source)
         # TODO: Delete this condition when rollout is done.  NB: this
         # condition never evaluates True in open source

torch/_dynamo/variables/builder.py

@@ -115,6 +115,8 @@
     Source,
     SubclassAttrListSource,
     TupleIteratorGetItemSource,
+    UnspecializedBuiltinNNModuleSource,
+    UnspecializedNNModuleSource,
 )
 from ..utils import (
     _extract_tensor_dict,
@@ -434,7 +436,10 @@ def __call__(self, value):
             return cached_vt
 
         vt = self._wrap(value)
-        vt.source = self.source
+
+        if vt.source is None:
+            vt.source = self.source
+
         if (
             self._can_lift_attrs_to_inputs(vt)
             and value not in self.tx.output.side_effects
@@ -1714,7 +1719,6 @@ def wrap_module(self, value: torch.nn.Module):
                 value = value.get_base()
                 self.source = AttrProxySource(self.source)
 
-            self.install_guards(GuardBuilder.TYPE_MATCH)
             if torch._dynamo.config.inline_inbuilt_nn_modules:
                 freezing = is_parameter_freezing()
 
@@ -1749,12 +1753,23 @@ def wrap_module(self, value: torch.nn.Module):
                     # this will get cleaned up once compile ends
                     self.tx.output.nn_modules[self.name] = value
 
-            if value.__module__.startswith(("torch.nn.", "torch.ao.")) or getattr(
-                value.__class__, "_dynamo_marked_static", False
-            ):
-                result = UnspecializedBuiltinNNModuleVariable(value, source=self.source)
+            if (
+                value.__module__.startswith(("torch.nn.modules", "torch.ao."))
+                and not value.__module__.startswith("torch.nn.modules.container")
+            ) or getattr(value.__class__, "_dynamo_marked_static", False):
+                new_source = self.source
+                if config.inline_inbuilt_nn_modules:
+                    # Export corner case - look at test_repros.py test_inlining_cornercase
+                    new_source = UnspecializedBuiltinNNModuleSource(self.source)
+                result = UnspecializedBuiltinNNModuleVariable(value, source=new_source)
+                install_guard(new_source.make_guard(GuardBuilder.TYPE_MATCH))
             else:
-                result = UnspecializedNNModuleVariable(value, source=self.source)
+                new_source = self.source
+                if config.inline_inbuilt_nn_modules:
+                    # Export corner case - look at test_repros.py test_inlining_cornercase
+                    new_source = UnspecializedNNModuleSource(self.source)
+                result = UnspecializedNNModuleVariable(value, source=new_source)
+                install_guard(new_source.make_guard(GuardBuilder.TYPE_MATCH))
 
             if not SideEffects.cls_supports_mutation_side_effects(type(value)):
                 # don't allow STORE_ATTR mutation with custom __setattr__
@@ -2127,6 +2142,10 @@ def wrap_numpy_ndarray(self, value):
         )
         proxy.node.meta["grapharg"] = grapharg
 
+        # TODO - Why do we need to set the source of the np ndarray vt back to
+        # original source. Many tests fails.
+        numpy_ndarray_variable.source = self.source
+
         return numpy_ndarray_variable
 
     def wrap_symint(

torch/_dynamo/variables/higher_order_ops.py

@@ -2658,8 +2658,8 @@ def call_function(
 
 class FlexAttentionBackwardHighOrderVariable(TorchHigherOrderOperatorVariable):
     def proxy_submod(self, tx, arg):
-        assert isinstance(arg.source, DictGetItemSource)
-        submod_name = tx.output.install_subgraph(arg.source.index, arg.value)
+        assert isinstance(arg.source.base, DictGetItemSource)
+        submod_name = tx.output.install_subgraph(arg.source.base.index, arg.value)
         p_submod = make_attr(tx, submod_name)
         set_example_value(p_submod.node, arg.value)
         return p_submod

torch/_dynamo/variables/nn_module.py

@@ -48,7 +48,6 @@
     FSDPNNModuleSource,
     GetItemSource,
     NNModuleSource,
-    UnspecializedBuiltinNNModuleSource,
     UnspecializedNNModuleSource,
 )
 from ..utils import (
@@ -891,8 +890,7 @@ def __init__(self, value, **kwargs) -> None:
         self.nn_module_stack_source = self.source
 
     def _wrap_source(self, attr_source):
-        if not isinstance(attr_source, UnspecializedNNModuleSource):
-            return UnspecializedNNModuleSource(attr_source)
+        # the vt is already wrapped with UnspecializedNNModuleSource
         return attr_source
 
     def get_nn_module_stack_source(self):
@@ -1193,8 +1191,7 @@ class UnspecializedBuiltinNNModuleVariable(UnspecializedNNModuleVariable):
     """
 
     def _wrap_source(self, attr_source):
-        if not isinstance(attr_source, UnspecializedBuiltinNNModuleSource):
-            return UnspecializedBuiltinNNModuleSource(attr_source)
+        # vt is already wrapped with the UnspecializedBuiltinNNModuleSource
         return attr_source