Jetson Orin Nano Compatibility with NCCL for PyTorch

[m.tharani.193]

Hi,

I’ve been working with a Jetson Orin Nano and recently installed PyTorch v1.11.0. While testing the distributed capabilities, I noticed that torch.distributed.is_available() returns True and torch.distributed.is_nccl_available() returns False. This led me to wonder if Jetson Orin Nano supports NCCL for PyTorch.

My attempts to build PyTorch from source using the command python setup.py bdist_wheel have resulted in system hangs and compilation termination errors during the rebuild process.

I’m facing difficulties resolving this issue and would appreciate guidance on how to enable NCCL support for PyTorch on my Jetson Orin Nano. Any advice or insights from the community would be greatly appreciated.

Thanks

[TomNVIDIA]

Moving this topic to the Jetson Orin nano category for visibility.

[dusty_nv]

Hi @m.tharani.193, Jetson is single-GPU architecture and doesn’t support NCCL:

Can Jetson Orin support nccl?

Can Jetson Orin support NCCL for communication among processes on the same chip or among pcie connected different chips? Also, what is the best way to do inter-chip data communication for Jetson Orin platform? NVSci? Or NCCL? What is the. difference between these two approaches?

You can recompile PyTorch with USE_DISTRIBUTED enabled (try mounting additional swap if it hangs), but it will use MPI instead of NCCL.