Support 'BaseOutput' and subclasses from 'diffusers' in dynamo (pytor…

…ch#111978)

Extending the workarounds for `transformers` `ModelOutput` to cover `diffusers` `BaseOutput`. Together with huggingface/diffusers#5459 it should unblock export for `diffusers` models.

Pull Request resolved: pytorch#111978
Approved by: https://github.com/jansel

test/dynamo/test_base_output.py

@@ -0,0 +1,97 @@
+# Owner(s): ["module: dynamo"]
+import unittest.mock
+
+import torch
+
+import torch._dynamo.test_case
+import torch._dynamo.testing
+from torch._dynamo.testing import same
+
+try:
+    from diffusers.models import unet_2d
+except ImportError:
+    unet_2d = None
+
+
+def maybe_skip(fn):
+    if unet_2d is None:
+        return unittest.skip("requires diffusers")(fn)
+    return fn
+
+
+class TestBaseOutput(torch._dynamo.test_case.TestCase):
+    @maybe_skip
+    def test_create(self):
+        def fn(a):
+            tmp = unet_2d.UNet2DOutput(a + 1)
+            return tmp
+
+        torch._dynamo.testing.standard_test(self, fn=fn, nargs=1, expected_ops=1)
+
+    @maybe_skip
+    def test_assign(self):
+        def fn(a):
+            tmp = unet_2d.UNet2DOutput(a + 1)
+            tmp.sample = a + 2
+            return tmp
+
+        args = [torch.randn(10)]
+        obj1 = fn(*args)
+
+        cnts = torch._dynamo.testing.CompileCounter()
+        opt_fn = torch._dynamo.optimize_assert(cnts)(fn)
+        obj2 = opt_fn(*args)
+        self.assertTrue(same(obj1.sample, obj2.sample))
+        self.assertEqual(cnts.frame_count, 1)
+        self.assertEqual(cnts.op_count, 2)
+
+    def _common(self, fn, op_count):
+        args = [
+            unet_2d.UNet2DOutput(
+                sample=torch.randn(10),
+            )
+        ]
+        obj1 = fn(*args)
+        cnts = torch._dynamo.testing.CompileCounter()
+        opt_fn = torch._dynamo.optimize_assert(cnts)(fn)
+        obj2 = opt_fn(*args)
+        self.assertTrue(same(obj1, obj2))
+        self.assertEqual(cnts.frame_count, 1)
+        self.assertEqual(cnts.op_count, op_count)
+
+    @maybe_skip
+    def test_getattr(self):
+        def fn(obj: unet_2d.UNet2DOutput):
+            x = obj.sample * 10
+            return x
+
+        self._common(fn, 1)
+
+    @maybe_skip
+    def test_getitem(self):
+        def fn(obj: unet_2d.UNet2DOutput):
+            x = obj["sample"] * 10
+            return x
+
+        self._common(fn, 1)
+
+    @maybe_skip
+    def test_tuple(self):
+        def fn(obj: unet_2d.UNet2DOutput):
+            a = obj.to_tuple()
+            return a[0] * 10
+
+        self._common(fn, 1)
+
+    @maybe_skip
+    def test_index(self):
+        def fn(obj: unet_2d.UNet2DOutput):
+            return obj[0] * 10
+
+        self._common(fn, 1)
+
+
+if __name__ == "__main__":
+    from torch._dynamo.test_case import run_tests
+
+    run_tests()

torch/_dynamo/convert_frame.py

@@ -303,7 +303,11 @@ def _convert_frame_assert(
         ):
             return None
         if code.co_name == "<genexpr>" and code.co_filename.endswith(
-            ("transformers/file_utils.py", "transformers/utils/generic.py")
+            (
+                "transformers/file_utils.py",
+                "transformers/utils/generic.py",
+                "diffusers/utils/outputs.py",
+            )
         ):
             # not needed, but cleans up torchbench error stats
             return None

torch/_dynamo/variables/dicts.py

@@ -437,6 +437,30 @@ def unpack_var_sequence(self, tx):
         return [x.add_options(self) for x in self.items]
 
 
+def _is_matching_transformers_cls(cls) -> bool:
+    if not cls.__module__.startswith("transformers."):
+        return False
+
+    try:
+        from transformers.file_utils import ModelOutput
+
+        return issubclass(cls, ModelOutput)
+    except ImportError:
+        return False
+
+
+def _is_matching_diffusers_cls(cls) -> bool:
+    if not cls.__module__.startswith("diffusers."):
+        return False
+
+    try:
+        from diffusers.utils import BaseOutput
+
+        return issubclass(cls, BaseOutput)
+    except ImportError:
+        return False
+
+
 class DataClassVariable(ConstDictVariable):
     """
     This is a bit of a hack to deal with
@@ -452,20 +476,27 @@ class DataClassVariable(ConstDictVariable):
     @staticmethod
     @functools.lru_cache(None)
     def _patch_once():
-        from transformers.file_utils import ModelOutput
+        try:
+            from transformers.file_utils import ModelOutput
 
-        for obj in ModelOutput.__dict__.values():
-            if callable(obj):
-                skip_code(obj.__code__)
+            for obj in ModelOutput.__dict__.values():
+                if callable(obj):
+                    skip_code(obj.__code__)
+        except ImportError:
+            pass
 
-    @staticmethod
-    def is_matching_cls(cls):
         try:
-            from transformers.file_utils import ModelOutput
+            from diffusers.utils import BaseOutput
 
-            return issubclass(cls, ModelOutput)
+            for obj in BaseOutput.__dict__.values():
+                if callable(obj):
+                    skip_code(obj.__code__)
         except ImportError:
-            return False
+            pass
+
+    @staticmethod
+    def is_matching_cls(cls):
+        return _is_matching_transformers_cls(cls) or _is_matching_diffusers_cls(cls)
 
     @classmethod
     def is_matching_object(cls, obj):
@@ -578,26 +609,19 @@ def var_getattr(self, tx, name: str) -> "VariableTracker":
 class CustomizedDictVariable(ConstDictVariable):
     @staticmethod
     def is_matching_cls(cls):
-        try:
-            # True if using default OrderedDict.__init__ and did not implement __post_init__
-            if (
-                issubclass(cls, collections.OrderedDict)
-                and cls.__init__ is collections.OrderedDict.__init__
-                and not hasattr(cls, "__post_init__")
-            ):
-                return True
-            # hack for HF usecase:
-            #   assume dataclass annotation for ModelOutput subclass
-            #   assume self.create is AA to ModelOutput.__post_init__
-            # for non-HF usecase:
-            #   check __module__ string to avoid costy HF import
-            if cls.__module__ != "transformers.modeling_outputs":
-                return False
-            from transformers.file_utils import ModelOutput
-
-            return issubclass(cls, ModelOutput)
-        except ImportError:
-            return False
+        # True if using default OrderedDict.__init__ and did not implement __post_init__
+        if (
+            issubclass(cls, collections.OrderedDict)
+            and cls.__init__ is collections.OrderedDict.__init__
+            and not hasattr(cls, "__post_init__")
+        ):
+            return True
+        # hack for HF usecase:
+        #   assume dataclass annotation for ModelOutput subclass
+        #   assume self.create is AA to ModelOutput.__post_init__
+        # for non-HF usecase:
+        #   check __module__ string to avoid costy HF import
+        return _is_matching_transformers_cls(cls) or _is_matching_diffusers_cls(cls)
 
     @classmethod
     def is_matching_object(cls, obj):