RuntimeError: Distributed package doesn't have NCCL built in

[manoj21192]

When trying to run example_completion.py file in my windows laptop, I am getting below error:

I am using pytorch 2.0 version with CUDA 11.7 . On typing the command
import torch.distributed as dist

if dist.is_nccl_available():
print("NCCL is available and built into PyTorch.")
else:
print("NCCL is not available in this PyTorch installation.")

I am getting the output "NCCL is not available in this PyTorch installation."
What should I do ?

[GaganHonor]

Based on the error message you provided, it seems that you are encountering a ChildFailedError when trying to run the example_completion.py file in your Windows laptop. This error is related to distributed training in PyTorch.

To fix this issue, you can try the following steps:

Check PyTorch Documentation: Visit the PyTorch documentation page on elastic errors to enable traceback and get more information about the error.

Update PyTorch and CUDA Versions: Make sure you are using compatible versions of PyTorch and CUDA. It is recommended to use the latest stable versions to ensure compatibility and access to the latest

[manoj21192]

Everything checked but unable to resolve. I am using latest version of Pytorch and 11.7 CUDA.

[srinivaskumarramdas]

Getting the same error on my macbook as well.

[GaganHonor]

please try to explain more and share information like system info , and all logs in txt form or link to them

[manoj21192]

As of now, for 7B parameter model, its working on windows by making changes to generator.py file by using torch.distributed.init_process_group("gloo"), instead of "nccl".
Is this methodology fine if I want to use high parameter model in future?

[srinivaskumarramdas]

"gloo" worked on Apple M2 chips mac as well. had some more additional changes covered in this PR: #18 .

[GaganHonor]

Please close this issue if issue is Fixed

[manoj21192]

Use gloo to make it work on Windows