Open XinyiZhang0724 opened 1 year ago
erro: Distributed package doesn't have NCCL built in
I managed to get this working after receiving the same issue. I add this line os.environ["PL_TORCH_DISTRIBUTED_BACKEND"] = "gloo" at the top of the run.py file. Then I removed strategy parameter from line 53 of run.py file strategy=DDPPlugin(find_unused_parameters=False). Seems DDPPlugin doesn't support gloo, please someone correct me if wrong on this.
Hello, I run it on Windows. How can I tansfer "nccl" to "gloo"?