Stop ignoring NotImplementedErrors in cuda CriterionTests. (pytorch#4…

…4381)

Summary:
Pull Request resolved: pytorch#44381

Perhaps this was necessary when the test was originally introduced, but it's difficult to figure out what is actually tested.  And I don't think we actually use NotImplementedErorrs.

Test Plan: Imported from OSS

Reviewed By: mruberry

Differential Revision: D23598646

Pulled By: gchanan

fbshipit-source-id: aa18154bfc4969cca22323e61683a301198823be

torch/testing/_internal/common_nn.py

@@ -5106,47 +5106,45 @@ def convert_dtype(obj, dtype, requires_grad=False):
 
         if not TEST_CUDA or not self.should_test_cuda:
             raise unittest.SkipTest('Excluded from CUDA tests')
-        try:
+
+        cpu_input = self._get_input()
+        cpu_target = self._get_target()
+        cpu_module = self.constructor(*self.constructor_args)
+        gpu_module = self.constructor(*self.constructor_args)
+
+        # Convert input, target and module parameters to dtype
+        if dtype is not None:
+            cpu_input = convert_dtype(cpu_input, dtype, True)
+            # NLLLoss requires target to be LongTensor
+            if not isinstance(cpu_target, torch.LongTensor) and self.convert_target:
+                cpu_target = convert_dtype(cpu_target, dtype)
+            cpu_module.type(dtype)
+            gpu_module.type(dtype)
+
+        # GPU setup
+        gpu_input = to_gpu(cpu_input)
+        gpu_target = to_gpu(cpu_target)
+        gpu_module.cuda()
+
+        # torch.HalfTensor doesn't support most operations, converting back to default
+        if dtype in {torch.half, torch.bfloat16}:
             cpu_input = self._get_input()
             cpu_target = self._get_target()
+            # Loss modules with weights require consistent input/module weight types
             cpu_module = self.constructor(*self.constructor_args)
-            gpu_module = self.constructor(*self.constructor_args)
-
-            # Convert input, target and module parameters to dtype
-            if dtype is not None:
-                cpu_input = convert_dtype(cpu_input, dtype, True)
-                # NLLLoss requires target to be LongTensor
-                if not isinstance(cpu_target, torch.LongTensor) and self.convert_target:
-                    cpu_target = convert_dtype(cpu_target, dtype)
-                cpu_module.type(dtype)
-                gpu_module.type(dtype)
-
-            # GPU setup
-            gpu_input = to_gpu(cpu_input)
-            gpu_target = to_gpu(cpu_target)
-            gpu_module.cuda()
-
-            # torch.HalfTensor doesn't support most operations, converting back to default
-            if dtype in {torch.half, torch.bfloat16}:
-                cpu_input = self._get_input()
-                cpu_target = self._get_target()
-                # Loss modules with weights require consistent input/module weight types
-                cpu_module = self.constructor(*self.constructor_args)
-
-            cpu_output = test_case._forward_criterion(cpu_module, cpu_input, cpu_target, extra_args=extra_args)
-            gpu_output = test_case._forward_criterion(gpu_module, gpu_input, gpu_target, extra_args=extra_args)
-            # dtype can be None, so set precision in this way instead of a precision map
-            # TODO(#38095): Replace assertEqualIgnoreType. See issue #38095
-            test_case.assertEqualIgnoreType(cpu_output, gpu_output,
-                                            atol=1e-1 if dtype in {torch.half, torch.bfloat16} else 4e-4, rtol=0)
 
-            cpu_gradInput = test_case._backward_criterion(cpu_module, cpu_input, cpu_target, extra_args=extra_args)
-            gpu_gradInput = test_case._backward_criterion(gpu_module, gpu_input, gpu_target, extra_args=extra_args)
-            # TODO(#38095): Replace assertEqualIgnoreType. See issue #38095
-            test_case.assertEqualIgnoreType(cpu_gradInput, gpu_gradInput,
-                                            atol=1e-1 if dtype in {torch.half, torch.bfloat16} else 4e-4, rtol=0)
-        except NotImplementedError:
-            pass
+        cpu_output = test_case._forward_criterion(cpu_module, cpu_input, cpu_target, extra_args=extra_args)
+        gpu_output = test_case._forward_criterion(gpu_module, gpu_input, gpu_target, extra_args=extra_args)
+        # dtype can be None, so set precision in this way instead of a precision map
+        # TODO(#38095): Replace assertEqualIgnoreType. See issue #38095
+        test_case.assertEqualIgnoreType(cpu_output, gpu_output,
+                                        atol=1e-1 if dtype in {torch.half, torch.bfloat16} else 4e-4, rtol=0)
+
+        cpu_gradInput = test_case._backward_criterion(cpu_module, cpu_input, cpu_target, extra_args=extra_args)
+        gpu_gradInput = test_case._backward_criterion(gpu_module, gpu_input, gpu_target, extra_args=extra_args)
+        # TODO(#38095): Replace assertEqualIgnoreType. See issue #38095
+        test_case.assertEqualIgnoreType(cpu_gradInput, gpu_gradInput,
+                                        atol=1e-1 if dtype in {torch.half, torch.bfloat16} else 4e-4, rtol=0)
 
     def _get_target(self):
         return self._get_arg('target', False)