算子开发流程

[fpzh2011]

搭建开发环境

conda create -y -n temp python=3.12
conda activate temp
# A100: torch==2.10 否则会报驱动错误
pip install torch==2.10 pre-commit
pip uninstall -y triton
RES="--index-url=https://resource.flagos.net/repository/flagos-pypi-hosted/simple \
    --trusted-host=https://resource.flagos.net"
# flagtree 需要时安装最新版本
# https://github.com/flagos-ai/FlagTree/tags
python3 -m pip install flagtree==0.5.0+3.1 $RES
conda install -y numpy sqlalchemy

# FlagGems
git clone git@github.com:siliconflow/FlagGems.git
cd FlagGems
pip install -U scikit-build-core>=0.11 pybind11 ninja cmake
pip install --no-build-isolation -e .
pre-commit install

# set env
export LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH

# exec unit test
pytest tests/test_special_ops.py -m reflection_pad2d

代码目录

• LLM 的算子一般在 src/flag_gems/fused
• 单元测试用例在 tests 目录，例如 tests/test_conv3d.py
• 性能测用例在 benchmark 目录

Git upstream

git clone git@github.com:siliconflow/FlagGems.git
cd FlagGems
git remote add upstream git@github.com:flagos-ai/FlagGems.git
git fetch upstream master
# git remote -v 显示
# origin https://github.com/siliconflow/FlagGems.git
# upstream  https://github.com/flagos-ai/FlagGems.git

Git 开发流程

• 在本仓库提交 feature 分支
  • pre-commit run --all-files
  • pre-commit run isort --files file1 file2
• git push origin my-feature
• 创建 Pull Request
  • 标题如 [SiliconFlow] Op: Reflection pad2d
  • 参考 addcmul, acos
  • 检查 Pull Request 页面各种检查是否有不通过的情况
• Pull Request 合并后，将 upstream/master 更新到 origin/master

Pull Request 配置

base repository: flagos-ai/FlagGems
base branch: master

head repository: siliconflow/FlagGems
compare: my-feature

更新 upstream

git checkout master
git pull upstream master
git push origin master