With out: OK Traceback (most recent call last): File "/data/users/xmfan/gb200_moe_sol/tests/repro_custom_op_optional_mutate.py", line 25, in <module> result = foo(x) ^^^^^^ File "/home/xmfan/miniconda3/envs/gb200_moe_sol/lib/python3.12/site-packages/torch/_library/custom_ops.py", line 726, in call return self._opoverload(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/xmfan/miniconda3/envs/gb200_moe_sol/lib/python3.12/site-packages/torch/_ops.py", line 871, in call return self._op(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/xmfan/miniconda3/envs/gb200_moe_sol/lib/python3.12/site-packages/torch/_library/autograd.py", line 112, in autograd_impl result = forward_no_grad(*args, Metadata(keyset, keyword_only_args)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/xmfan/miniconda3/envs/gb200_moe_sol/lib/python3.12/site-packages/torch/_library/autograd.py", line 41, in forward_no_grad result = op.redispatch(keyset & _C._after_autograd_keyset, *args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/xmfan/miniconda3/envs/gb200_moe_sol/lib/python3.12/site-packages/torch/_ops.py", line 878, in redispatch return self._handle.redispatch_boxed(keyset, *args, **kwargs) # type: ignore[return-value] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/xmfan/miniconda3/envs/gb200_moe_sol/lib/python3.12/site-packages/torch/_library/custom_ops.py", line 686, in adinplaceorview_impl increment_version(args[idx]) ~~~~^^^^^ IndexError: tuple index out of range

TL;DR

The most likely fix is to modify the foo function to handle the case when out is None without attempting to access it as if it were a tuple element.

Guidance

• Verify that the error occurs when the out parameter is not provided (i.e., defaults to None).
• Check the mutates_args parameter of the @torch.library.custom_op decorator to ensure it correctly handles optional arguments.
• Consider modifying the foo function to explicitly handle the case when out is None, potentially by adding a conditional statement to avoid attempting to access out as if it were a tuple element.
• Review the PyTorch documentation for custom ops to ensure that the implementation aligns with the expected behavior for optional arguments.

Example

@torch.library.custom_op("mylib::foo", mutates_args=("out",))
def foo(x: torch.Tensor, out: torch.Tensor | None = None) -> torch.Tensor:
    if out is not None:
        out.copy_(x)
        return x.new_empty(0)  # don't alias out
    else:
        # Handle the case when out is None without attempting to access it as a tuple element
        return x.clone()

Notes

The provided code snippet and error message suggest that the issue is related to the handling of optional arguments in custom PyTorch ops. However, without further information about the intended behavior of the foo function, it is difficult to provide a more specific solution.

Recommendation

Apply a workaround by modifying the foo function to explicitly handle the case when out is None, as shown in the example above. This should prevent the IndexError from occurring when the out parameter is not provided.