pytorch - 💡(How to fix) Fix DISABLED test_opcheck_opinfo_NumpyTakeCustomOp_cpu_float32 (__main__.TestCustomOpTestingCPU) [2 comments, 1 participants]

Traceback (most recent call last): File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/testing/_internal/optests/generate_tests.py", line 695, in opcheck tester(op, args, kwargs, rtol=rtol, atol=atol) ~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/testing/_internal/optests/generate_tests.py", line 61, in safe_schema_check result = op(*args, **kwargs) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_ops.py", line 871, in call return self._op(*args, **kwargs) ~~~~~~~~^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/testing/_internal/common_utils.py", line 1727, in torch_function r = func(*args, **kwargs) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_ops.py", line 871, in call return self._op(*args, **kwargs) ~~~~~~~~^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_compile.py", line 54, in inner return disable_fn(*args, **kwargs) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_dynamo/eval_frame.py", line 1281, in _fn return fn(*args, **kwargs) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_subclasses/schema_check_mode.py", line 184, in torch_dispatch out = func(*args, **kwargs) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_ops.py", line 871, in call return self._op(*args, **kwargs) ~~~~~~~~^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_library/custom_ops.py", line 375, in backend_impl result = self._backend_fns[device_type](*args, **kwargs) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_compile.py", line 54, in inner return disable_fn(*args, **kwargs) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_dynamo/eval_frame.py", line 1281, in _fn return fn(*args, **kwargs) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_library/custom_ops.py", line 410, in wrapped_fn return fn(*args, **kwargs) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/testing/_internal/custom_op_db.py", line 157, in numpy_take x = to_numpy(x) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/testing/_internal/custom_op_db.py", line 29, in to_numpy return tensor.cpu().numpy() ~~~~~~~~~~~~~~~~~~^^ RuntimeError: Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead.

The above exception was the direct cause of the following exception:

Traceback (most recent call last): File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/testing/_internal/common_device_type.py", line 1175, in test_wrapper return test(*args, **kwargs) File "/__w/pytorch/pytorch/test/test_custom_ops.py", line 444, in test_opcheck_opinfo torch.library.opcheck(op.op, args, kwargs) ~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/library.py", line 1749, in opcheck return optests.opcheck( ~~~~~~~~~~~~~~~^ op, ^^^ ...<5 lines>... atol=atol, ^^^^^^^^^^ ) ^ File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/testing/_internal/optests/generate_tests.py", line 699, in opcheck raise OpCheckError( ...<2 lines>... ) from ex torch.testing._internal.optests.generate_tests.OpCheckError: opcheck(op, ...): test_schema failed with Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead. (scroll up for stack trace)

The above exception was the direct cause of the following exception:

Traceback (most recent call last): File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/testing/_internal/common_utils.py", line 3444, in wrapper method(*args, **kwargs) ~~~~~~^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/testing/_internal/common_device_type.py", line 439, in instantiated_test result = test(self, **param_kwargs) File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/testing/_internal/common_utils.py", line 1821, in wrapper fn(*args, **kwargs) ~~^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/testing/_internal/common_device_type.py", line 1187, in test_wrapper raise e_tracked from e Exception: opcheck(op, ...): test_schema failed with Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead. (scroll up for stack trace)

Caused by sample input at index 0: SampleInput(input=Tensor[size=(3, 5), device="cpu", dtype=torch.float32], args=(Tensor[size=(3, 5), device="cpu", dtype=torch.int64],Tensor[size=(3, 5), device="cpu", dtype=torch.int64],1), kwargs={}, broadcasts_input=False, name='')

To execute this test, run the following from the base repo dir: PYTORCH_OPINFO_SAMPLE_INPUT_INDEX=0 PYTORCH_TEST_WITH_CROSSREF=1 python test/test_custom_ops.py TestCustomOpTestingCPU.test_opcheck_opinfo_NumpyTakeCustomOp_cpu_float32

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0

TL;DR

The most likely fix is to modify the to_numpy function in custom_op_db.py to use tensor.detach().numpy() instead of tensor.cpu().numpy() to resolve the RuntimeError caused by attempting to call numpy() on a Tensor that requires grad.

Guidance

• Review the to_numpy function in custom_op_db.py to ensure it correctly handles Tensors that require gradients.
• Update the to_numpy function to use tensor.detach().numpy() to detach the Tensor from the computation graph before converting it to a NumPy array.
• Verify that the test test_opcheck_opinfo_NumpyTakeCustomOp_cpu_float32 in test_custom_ops.py passes after applying the fix.
• Investigate other potential instances of this issue in the codebase where tensor.cpu().numpy() is used on Tensors that may require gradients.

Example

# custom_op_db.py
def to_numpy(x):
    return x.detach().cpu().numpy()  # Update this line to detach the Tensor

Notes

This fix assumes that detaching the Tensor from the computation graph is the intended behavior. If the Tensor needs to remain part of the computation graph, an alternative solution may be required.

Recommendation

Apply the workaround by updating the to_numpy function to use tensor.detach().numpy() to resolve the immediate issue, and then investigate the codebase for other potential instances of this problem.