FlagTree/documents/build_cn.md at main · flagos-ai/FlagTree

English

从源代码安装

从源码构建技巧

自动下载依赖库的速度可能受限于网络环境，编译前可自行下载至缓存目录 ~/.flagtree（可通过环境变量 FLAGTREE_CACHE_DIR 修改），无需自行设置 LLVM_BUILD_DIR 等环境变量。
各后端完整构建命令如下：

💫 ILUVATAR（天数智芯）iluvatar

对应的 Triton 版本为 3.1，基于 x64 平台

1. 构建及运行环境

推荐使用 Ubuntu 20.04

2. 手动下载 FlagTree 依赖库

mkdir -p ~/.flagtree/iluvatar; cd ~/.flagtree/iluvatar
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/iluvatar-llvm18-x86_64_v0.4.0.tar.gz
tar zxvf iluvatar-llvm18-x86_64_v0.4.0.tar.gz
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/iluvatarTritonPlugin-cpython3.10-glibc2.35-glibcxx3.4.30-cxxabi1.3.13-ubuntu-x86_64_v0.4.0.tar.gz
tar zxvf iluvatarTritonPlugin-cpython3.10-glibc2.35-glibcxx3.4.30-cxxabi1.3.13-ubuntu-x86_64_v0.4.0.tar.gz

3. 手动下载 Triton 依赖库

cd ${YOUR_CODE_DIR}/FlagTree
# For Triton 3.1 (x64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.1.x-linux-x64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.1.x-linux-x64.tar.gz

4. 源码构建命令

cd ${YOUR_CODE_DIR}/FlagTree/python
export FLAGTREE_BACKEND=iluvatar
python3 -m pip install . --no-build-isolation -v

💫 KLX xpu

对应的 Triton 版本为 3.0，基于 x64 平台

1. 构建及运行环境

推荐使用镜像（22GB）ubuntu_2004_x86_64_v30.tar
联系 kunlunxin-support@baidu.com 可获取进一步支持

2. 手动下载 FlagTree 依赖库

mkdir -p ~/.flagtree/xpu; cd ~/.flagtree/xpu
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/XTDK-llvm19-ubuntu2004_x86_64_v0.3.0.tar.gz
tar zxvf XTDK-llvm19-ubuntu2004_x86_64_v0.3.0.tar.gz
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/xre-Linux-x86_64_v0.3.0.tar.gz
tar zxvf xre-Linux-x86_64_v0.3.0.tar.gz
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/xpu-liblaunch_shared_so-ubuntu-x64_v0.3.1.tar.gz
tar zxvf xpu-liblaunch_shared_so-ubuntu-x64_v0.3.1.tar.gz

3. 手动下载 Triton 依赖库

cd ${YOUR_CODE_DIR}/FlagTree
# For Triton 3.1 (x64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.1.x-linux-x64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.1.x-linux-x64.tar.gz

4. 源码构建命令

cd ${YOUR_CODE_DIR}/FlagTree/python
export FLAGTREE_BACKEND=xpu
python3 -m pip install . --no-build-isolation -v

💫 Moore Threads（摩尔线程）mthreads

对应的 Triton 版本为 3.1，基于 x64/aarch64 平台

1. 构建及运行环境

推荐使用镜像 Dockerfile-ubuntu22.04-python3.10-mthreads

2. 手动下载 FlagTree 依赖库

mkdir -p ~/.flagtree/mthreads; cd ~/.flagtree/mthreads
# x64
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/mthreads-llvm19-glibc2.35-glibcxx3.4.30-x64_v0.4.0.tar.gz
tar zxvf mthreads-llvm19-glibc2.35-glibcxx3.4.30-x64_v0.4.0.tar.gz
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/mthreadsTritonPlugin-cpython3.10-glibc2.35-glibcxx3.4.30-cxxabi1.3.13-ubuntu-x64_v0.4.1.tar.gz
tar zxvf mthreadsTritonPlugin-cpython3.10-glibc2.35-glibcxx3.4.30-cxxabi1.3.13-ubuntu-x64_v0.4.1.tar.gz
# aarch64
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/mthreads-llvm19-glibc2.35-glibcxx3.4.30-aarch64_v0.4.0.tar.gz
tar zxvf mthreads-llvm19-glibc2.35-glibcxx3.4.30-aarch64_v0.4.0.tar.gz
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/mthreadsTritonPlugin-cpython3.10-glibc2.35-glibcxx3.4.30-cxxabi1.3.13-ubuntu-aarch64_v0.4.0.tar.gz
tar zxvf mthreadsTritonPlugin-cpython3.10-glibc2.35-glibcxx3.4.30-cxxabi1.3.13-ubuntu-aarch64_v0.4.0.tar.gz

3. 手动下载 Triton 依赖库

cd ${YOUR_CODE_DIR}/FlagTree
# For Triton 3.1 (x64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.1.x-linux-x64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.1.x-linux-x64.tar.gz

4. 源码构建命令

cd ${YOUR_CODE_DIR}/FlagTree/python
export FLAGTREE_BACKEND=mthreads
python3 -m pip install . --no-build-isolation -v

💫 ARM China（安谋科技）aipu

对应的 Triton 版本为 3.3，基于 x64/arm64 平台

1. 构建及运行环境

推荐使用 Ubuntu 22.04

2. 手动下载 FlagTree 依赖库

模拟环境中使用 x64 版本的 llvm，在 ARM 开发板上使用 arm64 版本的 llvm

mkdir -p ~/.flagtree/aipu; cd ~/.flagtree/aipu
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/llvm-a66376b0-ubuntu-x64-clang16-lld16_v0.4.0.tar.gz
tar zxvf llvm-a66376b0-ubuntu-x64-clang16-lld16_v0.4.0.tar.gz

3. 手动下载 Triton 依赖库

cd ${YOUR_CODE_DIR}/FlagTree
# For Triton 3.3 (x64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.3.x-linux-x64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.3.x-linux-x64.tar.gz

4. 源码构建命令

cd ${YOUR_CODE_DIR}/FlagTree/python
git checkout -b triton_v3.3.x origin/triton_v3.3.x
export FLAGTREE_BACKEND=aipu
python3 -m pip install . --no-build-isolation -v

💫 Tsingmicro（清微智能）tsingmicro

对应的 Triton 版本为 3.3，基于 x64 平台

1. 构建及运行环境

推荐使用 Ubuntu 20.04

2. 手动下载 FlagTree 依赖库

mkdir -p ~/.flagtree/tsingmicro; cd ~/.flagtree/tsingmicro
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/tsingmicro-llvm21-glibc2.30-glibcxx3.4.28-python3.11-x64_v0.2.0.tar.gz
tar zxvf tsingmicro-llvm21-glibc2.30-glibcxx3.4.28-python3.11-x64_v0.2.0.tar.gz
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/tx8_depends_release_20250814_195126_v0.2.0.tar.gz
tar zxvf tx8_depends_release_20250814_195126_v0.2.0.tar.gz

3. 手动下载 Triton 依赖库

cd ${YOUR_CODE_DIR}/FlagTree
# For Triton 3.3 (x64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.3.x-linux-x64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.3.x-linux-x64.tar.gz

4. 源码构建命令

cd ${YOUR_CODE_DIR}/FlagTree/python
git checkout -b triton_v3.3.x origin/triton_v3.3.x
export TX8_DEPS_ROOT=~/.flagtree/tsingmicro/tx8_deps
export FLAGTREE_BACKEND=tsingmicro
python3 -m pip install . --no-build-isolation -v

💫 Huawei Ascend（华为昇腾）ascend

对应的 Triton 版本为 3.2，基于 aarch64 平台

1. 构建及运行环境

推荐使用镜像 Dockerfile-ubuntu22.04-python3.11-ascend
或者使用镜像（5.4GB）https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/docker_image_cann-8.2.rc1.alpha003-a3-ubuntu22.04-py3.11-flagtree.tar.gz
上述步骤完成后，需要重新安装 Cann 相关的工具包：在 https://www.hiascend.com/developer/download/community/result?module=cann 注册账号后下载对应平台的 cann-toolkit、cann-ops

# cann-toolkit
chmod +x Ascend-cann-toolkit_8.5.0_linux-aarch64.run
./Ascend-cann-toolkit_8.5.0_linux-aarch64.run --install
# cann-ops for 910B (A2)
chmod +x Ascend-cann-910b-ops_8.5.0_linux-aarch64.run
./Ascend-cann-910b-ops_8.5.0_linux-aarch64.run --install
# cann-ops for 910C (A3)
chmod +x Ascend-cann-A3-ops_8.5.0_linux-aarch64.run
./Ascend-cann-A3-ops_8.5.0_linux-aarch64.run --install

2. 手动下载 FlagTree 依赖库

mkdir -p ~/.flagtree/ascend; cd ~/.flagtree/ascend
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/llvm-a66376b0-ubuntu-aarch64-python311-compat_v0.3.0.tar.gz
tar zxvf llvm-a66376b0-ubuntu-aarch64-python311-compat_v0.3.0.tar.gz

3. 手动下载 Triton 依赖库

cd ${YOUR_CODE_DIR}/FlagTree
# For Triton 3.2 (aarch64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.2.x-linux-aarch64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.2.x-linux-aarch64.tar.gz

4. 源码构建命令

cd ${YOUR_CODE_DIR}/FlagTree/python
git checkout -b triton_v3.2.x origin/triton_v3.2.x
export FLAGTREE_BACKEND=ascend
python3 -m pip install . --no-build-isolation -v

💫 HYGON（海光信息）hcu

Based on Triton 3.0, x64
对应的 Triton 版本为 3.0，基于 x64 平台

1. 构建及运行环境

推荐使用镜像 Dockerfile-ubuntu22.04-python3.10-hcu

2. 手动下载 FlagTree 依赖库

mkdir -p ~/.flagtree/hcu; cd ~/.flagtree/hcu
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/hcu-llvm20-df0864e-glibc2.35-glibcxx3.4.30-ubuntu-x86_64_v0.3.0.tar.gz
tar zxvf hcu-llvm20-df0864e-glibc2.35-glibcxx3.4.30-ubuntu-x86_64_v0.3.0.tar.gz

3. 手动下载 Triton 依赖库

cd ${YOUR_CODE_DIR}/FlagTree
# For Triton 3.1 (x64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.1.x-linux-x64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.1.x-linux-x64.tar.gz

4. 源码构建命令

cd ${YOUR_CODE_DIR}/FlagTree/python
export FLAGTREE_BACKEND=hcu
python3 -m pip install . --no-build-isolation -v

💫 Enflame（燧原）enflame

对应的 Triton 版本为 3.3，基于 x64 平台

1. 构建及运行环境

推荐使用镜像（2.4GB）https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/enflame-flagtree-0.3.1.tar.gz

2. 手动下载 FlagTree 依赖库

mkdir -p ~/.flagtree/enflame; cd ~/.flagtree/enflame
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/enflame-llvm21-d752c5b-gcc9-x64_v0.3.0.tar.gz
tar zxvf enflame-llvm21-d752c5b-gcc9-x64_v0.3.0.tar.gz

3. 手动下载 Triton 依赖库

cd ${YOUR_CODE_DIR}/FlagTree
# For Triton 3.3 (x64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.3.x-linux-x64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.3.x-linux-x64.tar.gz

4. 源码构建命令

cd ${YOUR_CODE_DIR}/FlagTree/python
export FLAGTREE_BACKEND=enflame
python3 -m pip install . --no-build-isolation -v

💫 Sunrise（曦望芯科）sunrise

对应的 Triton 版本为 3.4，基于 x64 平台

1. 构建及运行环境

推荐使用 Ubuntu 22.04

2. 手动下载 FlagTree 依赖库

mkdir -p ~/.flagtree/sunrise; cd ~/.flagtree/sunrise
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/sunrise-llvm21-glibc2.39-glibcxx3.4.33-x86_64_v0.4.0.tar.gz
tar zxvf sunrise-llvm21-glibc2.39-glibcxx3.4.33-x86_64_v0.4.0.tar.gz
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/sunriseTritonPlugin-cpython3.10-glibc2.39-glibcxx3.4.33-x86_64_v0.4.0.tar.gz
tar zxvf sunriseTritonPlugin-cpython3.10-glibc2.39-glibcxx3.4.33-x86_64_v0.4.0.tar.gz

3. 手动下载 Triton 依赖库

cd ${YOUR_CODE_DIR}/FlagTree
# For Triton 3.4 (x64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.4.x-linux-x64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.4.x-linux-x64.tar.gz

4. 源码构建命令

cd ${YOUR_CODE_DIR}/FlagTree
export TRITON_BUILD_WITH_CLANG_LLD=1
export TRITON_OFFLINE_BUILD=1
export TRITON_BUILD_PROTON=OFF
export FLAGTREE_BACKEND=sunrise
python3 -m pip install . --no-build-isolation -v

💫 NVIDIA & AMD nvidia & amd

对应的 Triton 版本为 3.1/3.2/3.3/3.4/3.5，基于 x64/arm64 平台

1. 构建及运行环境

推荐使用镜像（12GB）https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/docker_image_nvidia_pytorch_25.05-py3.tar.gz

2. 手动下载 LLVM 依赖包

cd ${YOUR_LLVM_DOWNLOAD_DIR}
# For Triton 3.1
wget https://oaitriton.blob.core.windows.net/public/llvm-builds/llvm-10dc3a8e-ubuntu-x64.tar.gz
tar zxvf llvm-10dc3a8e-ubuntu-x64.tar.gz
export LLVM_SYSPATH=${YOUR_LLVM_DOWNLOAD_DIR}/llvm-10dc3a8e-ubuntu-x64
# For Triton 3.2
wget https://oaitriton.blob.core.windows.net/public/llvm-builds/llvm-86b69c31-ubuntu-x64.tar.gz
tar zxvf llvm-86b69c31-ubuntu-x64.tar.gz
export LLVM_SYSPATH=${YOUR_LLVM_DOWNLOAD_DIR}/llvm-86b69c31-ubuntu-x64
# For Triton 3.3
wget https://oaitriton.blob.core.windows.net/public/llvm-builds/llvm-a66376b0-ubuntu-x64.tar.gz
tar zxvf llvm-a66376b0-ubuntu-x64.tar.gz
export LLVM_SYSPATH=${YOUR_LLVM_DOWNLOAD_DIR}/llvm-a66376b0-ubuntu-x64
# For Triton 3.4
wget https://oaitriton.blob.core.windows.net/public/llvm-builds/llvm-8957e64a-ubuntu-x64.tar.gz
tar zxvf llvm-8957e64a-ubuntu-x64.tar.gz
export LLVM_SYSPATH=${YOUR_LLVM_DOWNLOAD_DIR}/llvm-8957e64a-ubuntu-x64
# For Triton 3.5
wget https://oaitriton.blob.core.windows.net/public/llvm-builds/llvm-7d5de303-ubuntu-x64.tar.gz
tar zxvf llvm-7d5de303-ubuntu-x64.tar.gz
export LLVM_SYSPATH=${YOUR_LLVM_DOWNLOAD_DIR}/llvm-7d5de303-ubuntu-x64
#
export LLVM_INCLUDE_DIRS=$LLVM_SYSPATH/include
export LLVM_LIBRARY_DIR=$LLVM_SYSPATH/lib

3. 手动下载 Triton 依赖库

详见离线构建支持：预下载依赖包。

4. 源码构建命令

cd ${YOUR_CODE_DIR}/FlagTree
cd python  # For Triton 3.1, 3.2, 3.3, you need to enter the python directory to build
git checkout main                                   # For Triton 3.1
git checkout -b triton_v3.2.x origin/triton_v3.2.x  # For Triton 3.2
git checkout -b triton_v3.3.x origin/triton_v3.3.x  # For Triton 3.3
git checkout -b triton_v3.4.x origin/triton_v3.4.x  # For Triton 3.4
git checkout -b triton_v3.5.x origin/triton_v3.5.x  # For Triton 3.5
unset FLAGTREE_BACKEND
python3 -m pip install . --no-build-isolation -v
# If you need to build other backends afterward, you should clear LLVM-related environment variables
unset LLVM_SYSPATH LLVM_INCLUDE_DIRS LLVM_LIBRARY_DIR

离线构建支持

上文介绍了构建时 FlagTree 各后端可手动下载依赖包以避免受限于网络环境。Triton 构建时原本就带有一些依赖包，因此我们提供预下载包，可以手动安装至环境中，避免在构建时卡在自动下载阶段。

cd ${YOUR_CODE_DIR}/FlagTree
# For Triton 3.1 (x64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.1.x-linux-x64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.1.x-linux-x64.tar.gz
# For Triton 3.2 (x64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.2.x-linux-x64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.2.x-linux-x64.tar.gz
# For Triton 3.2 (aarch64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.2.x-linux-aarch64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.2.x-linux-aarch64.tar.gz
# For Triton 3.3 (x64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.3.x-linux-x64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.3.x-linux-x64.tar.gz
# For Triton 3.4 (x64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.4.x-linux-x64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.4.x-linux-x64.tar.gz
# For Triton 3.5 (x64)
wget https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/build-deps-triton_3.5.x-linux-x64.tar.gz
sh python/scripts/unpack_triton_build_deps.sh ./build-deps-triton_3.5.x-linux-x64.tar.gz

执行完上述脚本后，原有的 ~/.triton 目录将被重命名，新的 ~/.triton 目录会被创建并存放预下载包。

Q&A

Q: 安装完成后，运行时报错：version GLIBC or GLIBCXX not found

A: 查询环境中的 libc.so.6、libstdc++.so.6.0.30 支持的 GLIBC / GLIBCXX 版本：

strings /lib/x86_64-linux-gnu/libc.so.6 |grep GLIBC
strings /usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.30 | grep GLIBCXX

若支持该版本 GLIBC / GLIBCXX，那么也可尝试：

export LD_PRELOAD="/lib/x86_64-linux-gnu/libc.so.6"  # If GLIBC cannot be found
export LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.30"  # If GLIBCXX cannot be found
export LD_PRELOAD="/lib/x86_64-linux-gnu/libc.so.6 \
    /usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.30"  # If neither GLIBC nor GLIBCXX can be found

FilesExpand file tree

build_cn.md

Latest commit

History

build_cn.md

File metadata and controls

从源代码安装

从源码构建技巧

💫 ILUVATAR（天数智芯）iluvatar

1. 构建及运行环境

2. 手动下载 FlagTree 依赖库

3. 手动下载 Triton 依赖库

4. 源码构建命令

💫 KLX xpu

1. 构建及运行环境

2. 手动下载 FlagTree 依赖库

3. 手动下载 Triton 依赖库

4. 源码构建命令

💫 Moore Threads（摩尔线程）mthreads

1. 构建及运行环境

2. 手动下载 FlagTree 依赖库

3. 手动下载 Triton 依赖库

4. 源码构建命令

💫 ARM China（安谋科技）aipu

1. 构建及运行环境

2. 手动下载 FlagTree 依赖库

3. 手动下载 Triton 依赖库

4. 源码构建命令

💫 Tsingmicro（清微智能）tsingmicro

1. 构建及运行环境

2. 手动下载 FlagTree 依赖库

3. 手动下载 Triton 依赖库

4. 源码构建命令

💫 Huawei Ascend（华为昇腾）ascend

1. 构建及运行环境

2. 手动下载 FlagTree 依赖库

3. 手动下载 Triton 依赖库

4. 源码构建命令

💫 HYGON（海光信息）hcu

1. 构建及运行环境

2. 手动下载 FlagTree 依赖库

3. 手动下载 Triton 依赖库

4. 源码构建命令

💫 Enflame（燧原）enflame

1. 构建及运行环境

2. 手动下载 FlagTree 依赖库

3. 手动下载 Triton 依赖库

4. 源码构建命令

💫 Sunrise（曦望芯科）sunrise

1. 构建及运行环境

2. 手动下载 FlagTree 依赖库

3. 手动下载 Triton 依赖库

4. 源码构建命令

💫 NVIDIA & AMD nvidia & amd

1. 构建及运行环境

2. 手动下载 LLVM 依赖包

3. 手动下载 Triton 依赖库

4. 源码构建命令

离线构建支持

Q&A

Q: 安装完成后，运行时报错：version GLIBC or GLIBCXX not found