[Auto Parallel] fix recompute reentrant:false bugs in auto parallel&dynamic graph

[GITD245]

PR Category

Auto Parallel

PR Types

Bug fixes

Description

动态图自动并行使用recompute并设置reentrant:false时的一些问题：

1. 使用原先的tmp_tensor = core.eager.Tensor会出现backward过程中读取ctx里存储的tensor未初始化的情况（fused_layers.py中）
2. 如果ctx中保存的tensor有grad，使用原先的tmp_tensor会将grad清除
   该pr修复上述两种情况

NOTE: 原先使用以下代码

# TODO(jeff41404): it seems better to use `tmp_tensor = core.eager.Tensor(inner_x)`,
# but other errors will be triggered during the current period, and can be modified after resolution
tmp_tensor = core.eager.Tensor(
    inner_x.dtype,
    inner_x.shape,
    inner_x.name + "cpy",
    core.VarDesc.VarType.DENSE_TENSOR,
    inner_x.persistable,
    inner_x.process_mesh,
    inner_x.placements,
)

是因为早前框架存在某些bug，需通过这种方式绕过，现替换为

tmp_tensor = core.eager.Tensor(inner_x)

修改后已在动半qwen、baichuan、gpt2、llama7b上完成测试，可正常生效且loss对齐
pcard-86802

[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.

[liym27]

LGTM

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 50.00000% with 1 line in your changes missing coverage. Please review.

Please upload report for BASE (develop@726b769). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...on/paddle/distributed/fleet/recompute/recompute.py	50.00%	1 Missing ⚠️

❌ Your patch status has failed because the patch coverage (50.00%) is below the target coverage (90.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             develop   #74075   +/-   ##
==========================================
  Coverage           ?   50.00%           
==========================================
  Files              ?        1           
  Lines              ?        2           
  Branches           ?        0           
==========================================
  Hits               ?        1           
  Misses             ?        1           
  Partials           ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[liym27]

LGTM