Error logs

TL;DR

The issue can be worked around by re-examining the custom op's behavior when autograd is disabled.

Guidance

• Review the @torch.library.custom_op decorator to ensure it correctly handles the no_grad context.
• Consider modifying the bar function to create a new tensor instead of returning a view into the mutated buffer when autograd is disabled.
• Investigate if there are any existing issues or discussions related to custom ops and autograd in the PyTorch library.
• Verify if the issue persists when using a different device (e.g., CPU) instead of CUDA.

Example

@torch.library.custom_op("mylib::bar", mutates_args=("buf",))
def bar(buf: torch.Tensor, n: int) -> torch.Tensor:
    with torch.no_grad():
        new_buf = buf.clone()
        new_buf[:n].fill_(1.0)
        return new_buf[:n]

Notes

The provided code snippet and issue description suggest that the problem lies in the interaction between custom ops and autograd. However, without more information about the PyTorch version and the specific use case, it's difficult to provide a definitive solution.

Recommendation

Apply workaround: The example code snippet provides a possible workaround by creating a new tensor instead of returning a view into the mutated buffer when autograd is disabled. This approach may help mitigate the issue, but further investigation is needed to determine the root cause and a permanent fix.