reduce.py

import torch
import torch.distributed as dist

def init_process():
    dist.init_process_group(backend="nccl")
    torch.cuda.set_device(dist.get_rank())

def example_reduce():
    tensor = torch.tensor([dist.get_rank()] * 4, dtype=torch.float32).cuda()
    print(f"Before reduce on rank {dist.get_rank()}: {tensor}")
    dist.reduce(tensor, dst=0, op=dist.ReduceOp.SUM)
    print(f"After reduce on rank {dist.get_rank()}: {tensor}")

init_process()
example_reduce()
dist.destroy_process_group()