[体验优化] 整合训练的CUDA和Triton算子为 paddlenlp_kernel

[JunnYu]

PR types

New features

PR changes

APIs

Description

新增 paddlenlp_kernel wheel包，整合训练过程中用到的CUDA以及Triton算子
目前支持：

• mamba1 && mamba2 算子
• fast_ln && fused_ln 算子
• fused_linear_cross_entropy 算子
• paddle的反向算子：如 flash attn，flash mask，add，matmul
• inf_cl 算子

编译 cuda 算子

cd csrc
rm -rf build dist *.egg-info  # 清理之前的构建文件和目录
python setup.py build  # 开始编译

打包 wheel 包

完成 CUDA 算子的编译后，接下来打包成 Wheel 包以便安装：

python setup.py bdist_wheel

安装 wheel 包

使用 pip 命令安装刚刚打包好的 Wheel 包：

pip install dist/*.whl

[paddle-bot]

Thanks for your contribution!

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 52.79%. Comparing base (25415fb) to head (a04be9f).
Report is 12 commits behind head on develop.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #9471      +/-   ##
===========================================
- Coverage    53.16%   52.79%   -0.37%     
===========================================
  Files          718      718              
  Lines       113862   112252    -1610     
===========================================
- Hits         60532    59263    -1269     
+ Misses       53330    52989     -341     

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