I have NCCL 2.5 installed on the system, but torch.cuda.nccl.version() shows 2.4.8.
On a single machine with 2 gpus, it works fine.
Hey @nash, NCCL is packaged in PyTorch as a submodule. The current version if 2.4.8. If you would like to use your own version, you can set USE_SYSTEM_NCCL=1.
1 Like