[AutoParallel] Support qwen for auto_parallel

[GhostScreaming]

PR types

New features

PR changes

Models

Description

Support qwen for auto_parallel.

1. 在QWen上验证自动并行架构

• 对比动态图半自动并行和动态图手动并行，收敛和精度结果符合预期，涉及以下策略验证
  • dp2mp2pp2
  • dp2mp2pp2 + amp
  • dp2mp2pp2 + 动转静 + amp

2. 本 PR 代码包含最终模型代码、示例脚本、单测。验证结果如下

• dp2mp2pp2 FP32 Loss Diff

- dp2mp2pp2 BF16 Loss Diff

- dp2mp2pp2 动转静 FP32 Loss Diff

- dp2mp2pp2 动转静 BF16 Loss Diff

[paddle-bot]

Thanks for your contribution!

[codecov]

Codecov Report

Attention: Patch coverage is 16.56051% with 393 lines in your changes missing coverage. Please review.

Project coverage is 55.18%. Comparing base (1ffa290) to head (1799dea).
Report is 295 commits behind head on develop.

Files	Patch %	Lines
paddlenlp/transformers/qwen/modeling_3D_auto.py	16.38%	393 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #8312      +/-   ##
===========================================
- Coverage    55.25%   55.18%   -0.07%     
===========================================
  Files          613      614       +1     
  Lines        95626    96326     +700     
===========================================
+ Hits         52837    53161     +324     
- Misses       42789    43165     +376     

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

[liym27]

LGTM

[wawltor]

LGTM