[BE] Raise NotImplementedError (#155470)

malfet · pytorchmergebot · commit 5ab6a3fb6fd3 · 2025-06-12T23:19:12.000Z
When op is unimplemented for a specific dtype Which makes more sense, than a RuntimeError Example ```python >>> import torch >>> torch.nn.Hardshrink()(torch.randint(0, 5, (10,))) NotImplementedError: "hardshrink_cpu" not implemented for 'Long' ``` release notes bc-breaking: After this release `NotImplementedError` exception will be raised when ATen operation is called on the combinaiton of input tensor dtypes it has not been implemented for Pull Request resolved: #155470 Approved by: https://github.com/albanD, https://github.com/Skylion007
diff --git a/aten/src/ATen/Dispatch.h b/aten/src/ATen/Dispatch.h
@@ -200,7 +200,7 @@ inline at::ScalarType scalar_type(at::ScalarType s) {
     switch (_st) {                                                          \
       __VA_ARGS__                                                           \
       default:                                                              \
-        TORCH_CHECK(                                                        \
+        TORCH_CHECK_NOT_IMPLEMENTED(                                        \
             false,                                                          \
             '"',                                                            \
             at_dispatch_name,                                               \
diff --git a/test/torch_np/numpy_tests/core/test_getlimits.py b/test/torch_np/numpy_tests/core/test_getlimits.py
@@ -135,9 +135,9 @@ def test_basic(self):
         [
             np.uint8,
             # xfail: unsupported add (uint[16,32,64])
-            subtest(np.uint16, decorators=[xfail]),
-            subtest(np.uint32, decorators=[xfail]),
-            subtest(np.uint64, decorators=[xfail]),
+            subtest(np.uint16, decorators=[] if TEST_WITH_TORCHDYNAMO else [xfail]),
+            subtest(np.uint32, decorators=[] if TEST_WITH_TORCHDYNAMO else [xfail]),
+            subtest(np.uint64, decorators=[] if TEST_WITH_TORCHDYNAMO else [xfail]),
         ],
     )
     def test_unsigned_max(self, T):
