Ensure export joint with descriptors + compile works (#159337)

Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/159337
Approved by: https://github.com/wconstab
ghstack dependencies: #159336

test/functorch/test_aot_joint_with_descriptors.py

@@ -739,6 +739,28 @@ class inner_f(torch.nn.Module):
 
             self.assertEqual(param_count, len(param_nodes))
 
+    def test_export_and_compile(self):
+        class SimpleModule(nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.linear = nn.Linear(3, 2)
+
+            def forward(self, x):
+                return self.linear(x)
+
+        model = SimpleModule()
+        inputs = (torch.randn(4, 3),)
+
+        with ExitStack() as stack:
+            joint_with_descriptors = aot_export_joint_with_descriptors(
+                stack, model, inputs
+            )
+            model_fn = aot_compile_joint_with_descriptors(joint_with_descriptors)
+
+        compiled_fn = torch.compile(fullgraph=True)(model_fn)
+        compiled_fn(*dict(model.named_parameters()).values(), inputs).sum().backward()
+        self.assertIsNotNone(model.linear.weight.grad)
+
 
 if __name__ == "__main__":
     run_tests()

torch/_functorch/_aot_autograd/graph_compile.py

@@ -1762,7 +1762,10 @@ def aot_stage2_autograd(
                     placeholder_list[i] = ph_arg.as_strided(ph_arg.size(), real_stride)
 
             compiled_bw_func = None
-            if num_symints_saved_for_bw > 0:
+            if (
+                num_symints_saved_for_bw > 0
+                or aot_config.force_non_lazy_backward_lowering
+            ):
                 try:
                     # See Note: [Backward graph lazy lowering]
                     with torch._subclasses.fake_tensor.unset_fake_temporarily():
@@ -1775,6 +1778,8 @@ def aot_stage2_autograd(
                     )
                     del bw_module_copy
                 except Exception as e:
+                    if aot_config.force_non_lazy_backward_lowering:
+                        raise
                     exc = e
                     trace_structured(
                         "artifact",

torch/_functorch/_aot_autograd/runtime_wrappers.py

@@ -1949,12 +1949,11 @@ class AOTDispatchAutograd:
             expected_meta = meta.meta
 
         runtime_type = type(x)
-        if torch._dynamo.compiled_autograd.in_compiled_autograd_region:
+        # When we're inside compiled autograd's AOTDispatcher step,
-            # When we're inside compiled autograd's AOTDispatcher step,
+        # regular Tensors look like FunctionalTensors.
-            # regular Tensors look like FunctionalTensors.
+        # Tensor subclasses still look like Tensor subclasses though.
-            # Tensor subclasses still look like Tensor subclasses though.
+        if isinstance(x, torch._subclasses.functional_tensor.FunctionalTensor):
-            if isinstance(x, torch._subclasses.functional_tensor.FunctionalTensor):
+            runtime_type = torch.Tensor
-                runtime_type = torch.Tensor
 
         runtime_meta = None
         runtime_subclass_keys: Sequence[str] = []

torch/_functorch/_aot_autograd/schemas.py

@@ -982,6 +982,7 @@ class AOTConfig:
     # Used only by standalone_compile.
     ignore_shape_env: bool = False
     precompile_backend_id: Optional[str] = None
+    force_non_lazy_backward_lowering: bool = False
 
     def __post_init__(self):
         if self.pre_dispatch:

torch/_functorch/aot_autograd.py

@@ -893,6 +893,8 @@ def prepare_aot_module_simplified(
     boxed_forward_device_index: BoxedDeviceIndex,
     ignore_shape_env: bool,
     flatten: bool,
+    *,
+    force_non_lazy_backward_lowering: bool = False,
 ):
     if not flatten:
         assert kwargs is None
@@ -982,6 +984,7 @@ def prepare_aot_module_simplified(
         cache_info=None,
         ignore_shape_env=ignore_shape_env,
         precompile_backend_id=getattr(mod, "_backend_id", None),
+        force_non_lazy_backward_lowering=force_non_lazy_backward_lowering,
     )
     fake_mode, shape_env = construct_fake_mode(full_args, aot_config)
     # NB: full_args_descs not needed here, fake_flat_args is 1:1 with full_args
@@ -1225,6 +1228,12 @@ def aot_export_joint_with_descriptors(
         None,
         ignore_shape_env,
         flatten=True,
+        # Without this, we will attempt to "compile" the backward lazily
+        # at runtime, but this is pointless because it's just boxed_nop,
+        # it's trivial.  But this will get Inductor confused about scoping
+        # Metric(s) {'is_forward'} have already been set in the current
+        # context.
+        force_non_lazy_backward_lowering=True,
     )
 
     # TODO: Maybe this should be in create_aot_state?  Not sure, that would
@@ -1271,6 +1280,7 @@ def aot_compile_joint_with_descriptors(jd: JointWithDescriptors) -> callable:
 
     # Cribbed from torch/export/pt2_archive/_package.py
     @simple_wraps(compiled_fn)
+    @torch._dynamo.nonstrict_trace  # allow recursive compilation
     def unflattened_compiled_fn(*args, **kwargs):
         flat_inputs = pytree.tree_flatten((args, reorder_kwargs(kwargs, jd.in_spec)))[0]
         # TODO: do I need to filter? I hope not!