avoid compiling compute_loss (#511)

Summary:
Pull Request resolved: #511

# Context
As per discussion in D48361308, looks like we can remove this since it's only used for torch dynamo but it's not officially supported

# This diff
Separating the change from the above since we also need to clean up some existing UTs which don't make sense anymore. Going to rebase the above change on top of this

Reviewed By: JKSenthil

Differential Revision: D48581289

fbshipit-source-id: 0fc08554edb2ff726ab5d6249a9b1ce045c5e7d5

tests/framework/test_auto_unit.py

@@ -266,6 +266,7 @@ def test_compile_state_dict(self) -> None:
         """
         device = init_from_env()
         my_module = torch.nn.Linear(2, 2, device=device)
+        self.assertIsNone(my_module._compiled_call_impl)
         my_module_state_dict = my_module.state_dict()
         auto_unit = DummyAutoUnit(
             module=my_module,
@@ -283,94 +284,7 @@ def test_compile_state_dict(self) -> None:
             self.assertTrue(
                 torch.allclose(my_module_state_dict[k], compiled_state_dict[k])
             )
-
-    @unittest.skipUnless(
-        condition=COMPILE_AVAIL,
-        reason="This test needs PyTorch 1.13 or greater to run.",
-    )
-    def test_compile_eager(self) -> None:
-        """
-        e2e torch compile test
-        """
-
-        my_module = torch.nn.Linear(2, 2)
-
-        input_dim = 2
-        dataset_len = 16
-        batch_size = 2
-        max_epochs = 1
-
-        auto_unit = DummyAutoUnit(
-            module=my_module,
-            torch_compile_params=TorchCompileParams(backend="eager"),
-        )
-
-        train_dl = generate_random_dataloader(dataset_len, input_dim, batch_size)
-        self.assertFalse(auto_unit._compile_used)
-        train(auto_unit, train_dl, max_epochs=max_epochs)
-        self.assertTrue(auto_unit._compile_used)
-
-    @unittest.skipUnless(
-        condition=COMPILE_AVAIL,
-        reason="This test needs PyTorch 1.13 or greater to run.",
-    )
-    @unittest.skipUnless(
-        condition=cuda_available, reason="This test needs a GPU host to run."
-    )
-    def test_compile_train(self) -> None:
-        """
-        e2e torch compile on train
-        """
-
-        my_module = torch.nn.Linear(2, 2)
-
-        input_dim = 2
-        dataset_len = 16
-        batch_size = 2
-        max_epochs = 1
-
-        auto_unit = DummyAutoUnit(
-            module=my_module,
-            torch_compile_params=TorchCompileParams(backend="inductor"),
-        )
-
-        train_dl = generate_random_dataloader(dataset_len, input_dim, batch_size)
-
-        self.assertFalse(auto_unit._compile_used)
-        train(auto_unit, train_dl, max_epochs=max_epochs)
-        self.assertTrue(auto_unit._compile_used)
-
-    @unittest.skipUnless(
-        condition=COMPILE_AVAIL,
-        reason="This test needs PyTorch 1.13 or greater to run.",
-    )
-    @unittest.skipUnless(
-        condition=cuda_available, reason="This test needs a GPU host to run."
-    )
-    def test_compile_eval(self) -> None:
-        """
-        e2e torch compile on eval
-        """
-
-        my_module = torch.nn.Linear(2, 2)
-
-        input_dim = 2
-        dataset_len = 16
-        batch_size = 2
-
-        auto_unit = DummyAutoUnit(
-            module=my_module,
-            torch_compile_params=TorchCompileParams(backend="inductor"),
-        )
-
-        input_dim = 2
-        dataset_len = 8
-        batch_size = 2
-
-        eval_dl = generate_random_dataloader(dataset_len, input_dim, batch_size)
-        self.assertFalse(auto_unit._compile_used)
-        evaluate(auto_unit, eval_dl)
-        self.assertTrue(auto_unit._compile_used)
+        self.assertIsNotNone(auto_unit.module._compiled_call_impl)
 
     @unittest.skipUnless(
         condition=COMPILE_AVAIL,
@@ -983,16 +897,9 @@ def test_predict_detect_anomaly(self, mock_detect_anomaly) -> None:
 
 # pyre-fixme[11]: Annotation `Batch` is not defined as a type.
 class DummyAutoUnit(AutoUnit[Batch]):
-    # pyre-fixme[3]: Return type must be annotated.
-    # pyre-fixme[2]: Parameter must be annotated.
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self._compile_used = False
 
     # pyre-fixme[3]: Return annotation cannot contain `Any`.
     def compute_loss(self, state: State, data: Batch) -> Tuple[torch.Tensor, Any]:
-        if COMPILE_AVAIL:
-            self._compile_used = torch._dynamo.is_compiling()
         inputs, targets = data
         outputs = self.module(inputs)
         loss = torch.nn.functional.cross_entropy(outputs, targets)

torchtnt/framework/auto_unit.py

@@ -470,11 +470,6 @@ def __init__(
             )
 
         if torch_compile_params:
-            # pyre-ignore
-            self.compute_loss = torch.compile(
-                self.compute_loss,
-                **asdict(torch_compile_params),
-            )
             try:
                 # use in-place compile to avoid altering the state_dict keys
                 module.compile(**asdict(torch_compile_params))