Dynamo x autograd.Function supports setup_context (#124802)

yanboliang · pytorchmergebot · commit ce503c1b4020 · 2024-04-27T04:57:13.000Z
Fixes part of #118397 Pull Request resolved: #124802 Approved by: https://github.com/zou3519
diff --git a/test/dynamo/test_autograd_function.py b/test/dynamo/test_autograd_function.py
@@ -253,11 +253,11 @@ def test_autograd_function_has_graph_break(self):
 
     def test_linear_setup_context(self):
         model = ModuleLinear()
-        opt_model = torch._dynamo.optimize("eager")(model)
+        opt_model = torch._dynamo.optimize("eager", nopython=True)(model)
         input = torch.randn(2, 2, dtype=torch.double, requires_grad=True)
         weight = torch.randn(3, 2, dtype=torch.double, requires_grad=True)
-        optim_result = opt_model(input, weight)
         eager_result = model(input, weight)
+        optim_result = opt_model(input, weight)
         self.assertEqual(optim_result, eager_result)
 
     def test_materialize_grad(self):
diff --git a/torch/_dynamo/trace_rules.py b/torch/_dynamo/trace_rules.py
@@ -3205,6 +3205,7 @@ def _module_dir(m: types.ModuleType):
     "torch._dynamo.comptime",
     "torch._dynamo.polyfill",
     "torch._functorch.vmap",
+    "torch._functorch.autograd_function",
     "torch._library.custom_ops",
     "torch._functorch.eager_transforms",
     "torch._inductor.test_operators",
diff --git a/torch/_dynamo/variables/higher_order_ops.py b/torch/_dynamo/variables/higher_order_ops.py
@@ -1628,13 +1628,12 @@ def bwd(ctx, grad, x):
         fwd_src = AttrSource(self.parent_source, member="forward")
         ctx = AutogradFunctionContextVariable.create(tx, args, kwargs)
         if isinstance(self.fwd_graph, types.FunctionType):
-            fwd_fn = UserFunctionVariable(self.fwd_graph, source=fwd_src)
+            fwd_fn = UserFunctionVariable(self.fwd_graph)
             fwd_args = [ctx, *args]
         elif isinstance(self.fwd_graph, types.MethodType):
             fwd_fn = UserMethodVariable(
                 self.fwd_graph.__func__,
                 UserDefinedClassVariable(self.fwd_graph.__class__),
-                source=fwd_src,
             )
             fwd_args = [fwd_fn.obj, ctx, *args]
         else:
diff --git a/torch/_dynamo/variables/misc.py b/torch/_dynamo/variables/misc.py
@@ -357,14 +357,19 @@ def visit(node):
             and torch.is_grad_enabled()
             and config.capture_autograd_function
         ):
-            # Note - this is the same check used in autograd/function.py, except inverted.
-            # If we want to support functorch transforms here, we will need to enable this.
-            if (
-                self.fn_cls.setup_context
-                != torch.autograd.function._SingleLevelFunction.setup_context
-            ):
-                unimplemented(
-                    "NYI - autograd.Function with custom setup_context method"
+            from torch._functorch.autograd_function import (
+                autograd_function_forward_rewritten,
+            )
+            from torch.autograd.function import _is_setup_context_defined
+
+            forward_fn = self.fn_cls.forward
+
+            is_setup_ctx_defined = _is_setup_context_defined(self.fn_cls.setup_context)
+            if is_setup_ctx_defined:
+                # If setup_context is defined, we generate a new forward function which includes
+                # the original forward and setup_context function, and trace the new forward function.
+                forward_fn = autograd_function_forward_rewritten(
+                    self.fn_cls.forward, self.fn_cls.setup_context
                 )
 
             vjp_fn = self.fn_cls.vjp  # type: ignore[attr-defined]
@@ -383,12 +388,25 @@ def visit(node):
                     tx.import_source(self.fn_cls.__module__), self.fn_cls.__name__
                 )
 
-            return AutogradFunctionApplyVariable(
-                self.fn_cls.forward,
+            val = AutogradFunctionApplyVariable(
+                forward_fn,
                 self.fn_cls.backward,
                 source,
                 source=AttrSource(source, member="apply"),
             ).call_function(tx, args, kwargs)
+            # Inside of AutogradFunctionApplyVariable.call_function, we use sourceless variable wrapping
+            # the forward function, as we don't want to generate guards for new_forward.__closure__
+            # if forward is rewritten by autograd_function_forward_rewritten.
+            # But we still need to generate correct guards for the original forward and setup_context
+            # functions, so we have to add guards manually.
+            if self.source:
+                fwd_src = AttrSource(self.source, "forward")
+                install_guard(fwd_src.make_guard(GuardBuilder.FUNCTION_MATCH))
+                if is_setup_ctx_defined:
+                    setup_ctx_src = AttrSource(self.source, "setup_context")
+                    install_guard(setup_ctx_src.make_guard(GuardBuilder.FUNCTION_MATCH))
+
+            return val
 
         if self.source:
             source = AttrSource(self.source, "forward")
@@ -443,7 +461,32 @@ def call_method(
                 return self.call_apply(tx, args, kwargs)
 
         else:
-            unimplemented(f"Unsupported method: {name}")
+            from .. import trace_rules
+
+            source = AttrSource(self.source, name) if self.source is not None else None
+            try:
+                obj = inspect.getattr_static(self.fn_cls, name)
+            except AttributeError:
+                obj = None
+
+            if isinstance(obj, staticmethod):
+                func = obj.__get__(self.fn_cls)
+                if source is not None:
+                    return (
+                        trace_rules.lookup(func)
+                        .create_with_source(func, source=source)
+                        .call_function(tx, args, kwargs)
+                    )
+                else:
+                    return trace_rules.lookup(func)(func).call_function(
+                        tx, args, kwargs
+                    )
+            elif isinstance(obj, classmethod):
+                return variables.UserMethodVariable(
+                    obj.__func__, self, source=source
+                ).call_function(tx, args, kwargs)
+            else:
+                unimplemented(f"Unsupported method: {name}")
 
 
 @dataclasses.dataclass
diff --git a/torch/_dynamo/variables/user_defined.py b/torch/_dynamo/variables/user_defined.py
@@ -2,6 +2,7 @@
 
 import collections
 import contextlib
+import enum
 import functools
 import importlib
 import inspect
@@ -107,7 +108,7 @@ def can_constant_fold_through(self):
 
     def var_getattr(self, tx, name: str) -> "VariableTracker":
         from .. import trace_rules
-        from . import ConstantVariable
+        from . import ConstantVariable, EnumVariable
         from .builder import VariableBuilder
 
         if name == "__name__":
@@ -144,14 +145,16 @@ def var_getattr(self, tx, name: str) -> "VariableTracker":
         if self.value is collections.OrderedDict and name == "fromkeys":
             return super().var_getattr(tx, name)
 
-        if name in getattr(self.value, "__dict__", {}) or (
+        if ConstantVariable.is_literal(obj):
+            return ConstantVariable.create(obj)
+        elif isinstance(obj, enum.Enum):
+            return EnumVariable(obj)
+        elif name in getattr(self.value, "__dict__", {}) or (
             self.value.__module__.startswith("torch.")
             or self.value.__module__ == "torch"
         ):
             if source:
                 return VariableBuilder(tx, source)(obj)
-        elif ConstantVariable.is_literal(obj):
-            return ConstantVariable.create(obj)
 
         return super().var_getattr(tx, name)
 
diff --git a/torch/_functorch/autograd_function.py b/torch/_functorch/autograd_function.py
@@ -682,6 +682,15 @@ def reductify_leaf(
     return grad_input
 
 
+def autograd_function_forward_rewritten(original_forward, original_setup_context):
+    def new_forward(ctx, *args, **kwargs):
+        output = original_forward(*args, **kwargs)
+        original_setup_context(ctx, args, output)
+        return output
+
+    return new_forward
+
+
 class AutogradFunctionApply(HigherOrderOperator):
     def __init__(self):
         super().__init__("autograd_function_apply")
diff --git a/torch/autograd/function.py b/torch/autograd/function.py
@@ -561,7 +561,7 @@ def bind_default_args(func, *args, **kwargs):
 
             return bound_args.args
 
-        is_setup_ctx_defined = cls.setup_context != _SingleLevelFunction.setup_context
+        is_setup_ctx_defined = _is_setup_context_defined(cls.setup_context)
         if is_setup_ctx_defined:
             args = bind_default_args(cls.forward, *args, **kwargs)
 
@@ -585,6 +585,10 @@ def _compiled_autograd_key(ctx):
         return (ctx._autograd_function_id,)
 
 
+def _is_setup_context_defined(fn):
+    return fn != _SingleLevelFunction.setup_context
+
+
 def once_differentiable(fn):
     @functools.wraps(fn)
     def wrapper(ctx, *args):
