dist2dense_pass fix shape errors in shard randomly sampled data

[jeff41404]

PR Category

Auto Parallel

PR Types

Bug fixes

Description

pcard-86711
After PR #67589 supporting sharding of randomly sampled data, it is necessary to synchronously modify the shape in the attribute of previously operator of pd_op.full_int_array when use paddle.distributed.to_static and pir(FLAGS_enable_pir_api=true). Otherwise, the executor will still create data according to this wrong shape during actual execution, resulting in incorrect reporting of shape in subsequent calculations.

import paddle
import paddle.distributed as dist
process_mesh = dist.ProcessMesh([0, 1], dim_names=['dp'])

noise = paddle.randn(x.shape)  
noise = dist.shard_tensor(noise, process_mesh, [dist.Shard(0)) # same shard as x
y = noise * x # will error when use paddle.distributed.to_static and pir

Printing pir program also shows that the shape in the attribute of operator of pd_op.full_int_array (line1 and line3) is inconsistent with correct shape (line2 and line4)

this PR will fix this issue.

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[JZ-LIANG]

LGTM