import paddle
from paddle.framework import core
from paddle.distributed.fleet import auto
import numpy as np
from paddle.distributed.auto_parallel.static.dist_attribute import (
    DistTensorSpec,
    TensorDistAttr,
)
import paddle.distributed as dist
mesh0 = dist.ProcessMesh([0, 1], dim_names=['x'])
mesh1 = dist.ProcessMesh([2, 3], dim_names=['x'])
paddle.seed(0)
a = paddle.rand([2])
a = dist.shard_tensor(a, mesh0, [dist.Shard(0)])
print(a, flush=1)
print(np.array(a), flush=1)
# on gpu 2 and 3, the result should be:
# Tensor(shape=[2], dtype=float32, place=Place(gpu:2), stop_gradient=True, process_mesh={shape: [2], process_ids: [0,1], dim_names: [x]}, placements=[Shard(dim=0)], GlobalDenseTensor Not initialized)
# []
# However, before this pr, it is:
# Tensor(shape=[2], dtype=float32, place=Place(gpu:2), stop_gradient=True, process_mesh={shape: [2], process_ids: [0,1], dim_names: [x]}, placements=[Shard(dim=0)], GlobalDenseTensor=
       [1., 1.])
# [1. 1.]
# the data [1. 1.] may differ because the memory allocated by numpy is not initialized.