ReduceOp¶
指定规约类操作的逐元素操作类型,需要是下述值之一:
ReduceOp.SUM
ReduceOp.MAX
ReduceOp.MIN
ReduceOp.PROD
代码示例¶
import numpy as np
import paddle
from paddle.distributed import ReduceOp
from paddle.distributed import init_parallel_env
paddle.set_device('gpu:%d'%paddle.distributed.ParallelEnv().dev_id)
init_parallel_env()
if paddle.distributed.ParallelEnv().local_rank == 0:
np_data = np.array([[4, 5, 6], [4, 5, 6]])
else:
np_data = np.array([[1, 2, 3], [1, 2, 3]])
data = paddle.to_tensor(np_data)
paddle.distributed.all_reduce(data, op=ReduceOp.SUM)
out = data.numpy()
# [[5, 7, 9], [5, 7, 9]]