Python: torch.cuda.nccl.unique_id()

distributed
1

Hi,

I facing the same issue as in this post:

torch.cuda.nccl.init_rank does not handle “uid” properly, causing runtime error · Issue #38019 · pytorch/pytorch (github.com)

I am running the same example, and see this assert triggering https://github.com/pytorch/pytorch/blob/9f47580ad772c0da7e37a799711563534a51cba5/torch/csrc/cuda/python_nccl.cpp#L114.

Now id_len required is 128 Bytes.

I have torch v2.2.

Any solutions / work around for this?

Thank you.