The issue with training on multiple GPUs.

[longfafffa]

Thank you for your great work. When I tried to reproduce your code with multiple GPUs, I encountered the problem: 'Duplicate GPU detected: rank 2 and rank 0 both on CUDA device 52000.' From the code print(f"Process {os.getpid()} CUDA_VISIBLE_DEVICES: {os.getenv('CUDA_VISIBLE_DEVICES')}"), I found that each process is on the same GPU. How can I solve this problem?