Create all_reduce.py

all_reduce.py

@@ -0,0 +1,16 @@
+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.all_reduce(tensor, op=dist.ReduceOp.SUM)
+    print(f"After reduce on rank {dist.get_rank()}: {tensor}")
+
+init_process()
+example_reduce()
+dist.destroy_process_group()