Skip to content

Build Triton From Scratch

Jump to bottom
Stonepia edited this page Aug 7, 2023 · 10 revisions

This BKC is for the full building process for Triton. It offers a guide to setup and run Triton models on xpu platform from scratch.

1. Overview:

For CUDA Platform:

  • LLVM
  • PyTorch
  • Triton

For XPU Platform: In addition to CUDA's requirements, you also need

  • Intel® oneAPI Base Toolkit / driver (See Appendix for more detail)
  • Intel® Extension for PyTorch*

For all the cases, CUDA/XPU Shares the same building repo as well as the building steps. Thus, unless explicitly pointed out, the build process should fit both CUDA/XPU.

This BKC will be arranged into Env Setting, Build Process, TroubleShooting, Test Process, and Appendix. If you encounter any build errors, please refer to the troubleshooting section first.

2. Env Setting

We recommend using a conda environment for setup. You could skip this part if it is familiar to you. The Python version supports 3.x, you could use any version you like.

$conda create -n triton-env python=3.10
$conda activate triton-env

For proxy settings, please refer to the appendix.

It is recommended that you set the proxy if needed.

export https_proxy=your_proxy
export http_proxy=your_proxy

2.1. XPU Specific env setting

You need to specify Intel's GPU toolchain. Normally they are in the Intel® oneAPI Base Toolkit. You could refer to the Installation Guide for the latest updated version.

A sample env setting for oneAPI would be like this:

# Sample env setting
source ~/intel/oneapi/compiler/latest/env/vars.sh
source ~/intel/oneapi/mkl/latest/env/vars.sh
source ~/intel/oneapi/dnnl/latest/env/vars.sh

export MKL_DPCPP_ROOT=${HOME}/intel/oneapi/mkl/latest
export LD_LIBRARY_PATH=${MKL_DPCPP_ROOT}/lib:${MKL_DPCPP_ROOT}/lib64:${MKL_DPCPP_ROOT}/lib/intel64:${LD_LIBRARY_PATH}
export LIBRARY_PATH=${MKL_DPCPP_ROOT}/lib:${MKL_DPCPP_ROOT}/lib64:${MKL_DPCPP_ROOT}/lib/intel64:${LIBRARY_PATH}

# The AOT_DEVLIST should be set according to your device.
# export USE_AOT_DEVLIST='ats-m150'
# export USE_AOT_DEVLIST='pvc'
source ~/intel/oneapi/tbb/latest/env/vars.sh

# Helps to build quicker
export BUILD_SEPARATE_OPS=ON

You could store it in a file like env.sh, then call the source env.sh to activate it next time. Note that please source oneAPI after PyTorch is built, there are known bugs when using oneAPI's compiler to build PyTorch.

3. Build Process

It is required to build everything from the source.

3.1. LLVM

Temporary on triton_debug

git clone -b triton_debug https://github.com/chengjunlu/llvm/

Before build, make sure the related dependencies are installed in your system. See LLVM-requirements for detail. One thing worth noticing is that zlib is required for Triton, thus you should install it first.

sudo apt-get install zlib1g-dev
# Note: Choose the branch you need!
cd llvm
mkdir build && cd build
cmake ../llvm -G Ninja  -DLLVM_ENABLE_PROJECTS="mlir" -DCMAKE_BUILD_TYPE=Release -DLLVM_USE_LINKER=gold  -DLLVM_TARGETS_TO_BUILD="X86;NVPTX;AMDGPU"
ninja all

This should build LLVM and all its related targets. You could check those files under build folder.

In case there are build errors, please refer to troubleshooting page.

Triton build will use the LLVM_SYSPATH for specific LLVM dir. If you choose your own LLVM, you should do the following:

export LLVM_SYSPATH={abs_path_to}/llvm/build/

3.1.1. Other optional steps

In a later process, if you need something like LLVM_LIBRARY_DIR, you could export it using something like:

# Not required for triton, example only
export LLVM_LIBRARY_DIR=.../llvm/build/lib

For some reason, if you wish to use mlir-opt or other binaries, it is recommended to install the LLVM to another directory, you could take the following steps:

# Not required for triton, example only
# Your local install folder, could be anything you like
mkdir /home/user/local/llvm-install-folder
# In llvm/build
cmake -DCMAKE_INSTALL_PREFIX=/home/user/local/llvm-install-folder -P cmake_install.cmake

Export the path:

# Not required for Triton, example only
export PATH=/home/gta/tongsu/llvm/build/bin:${PATH}

3.2. PyTorch

Note that PyTorch should be built from the source with all torch patches applied and build with _GLIBCXX_USE_CXX11_ABI=1 enabled:

First clone PyTorch with v2.0.1 tag and intel-extension-for-pytorch repo with xpu-master branch:

git clone -b v2.0.1 https://github.com/pytorch/pytorch/
git clone -b xpu-master https://github.com/intel/intel-extension-for-pytorch

Important : Apply the torch patches:

cd pytorch
git apply ../intel-extension-for-pytorch/torch_patches/*.patch

Important Be sure to build PyTorch with _GLIBCXX_USE_CXX11_ABI=1 enabled:

export _GLIBCXX_USE_CXX11_ABI=1
git submodule sync
git submodule update --init --recursive --jobs 0

conda install cmake ninja mkl mkl-include
pip install -r requirements.txt

Then build PyTorch. Note that please make sure the oneAPI is not sourced at this step. Just use gcc to build PyTorch instead of icpx (The compiler in oneAPI).

export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
python setup.py develop

Test if PyTorch is installed without error:

cd ..
python -c "import torch;print(torch.__version__)"

3.3. Intel® Extension for PyTorch*

cd intel-extension-for-pytorch
git submodule sync
git submodule update --init --recursive --jobs 0
pip install -r requirements.txt

First source env.sh mentioned earlier:

source env.sh
python setup.py bdist_wheel
pip install dist/*.whl

Test if intel-extension-for-pytorch is installed without error:

cd ..
python -c "import torch;import intel_extension_for_pytorch as ipex;print(ipex.__version__)"

3.4. Triton

git clone https://github.com/openai/triton triton

Now set the env flags so that Triton build uses LLVM in the environment instead of re-download it.

# Formerly build LLVM path
export LLVM_SYSPATH={abs_path_to}/llvm/build/

Install pybind11 if it is missing:

pip install pybind11

Build Triton

$cd triton
$git submodule sync
$git submodule update --init --recursive --jobs 0
# check to latest third_party backend if needed
$cd third_party/intel_xpu_backend
$git checkout main && git pull
# cd back to python folder
$cd ../../python
$TRITON_CODEGEN_INTEL_XPU_BACKEND=1 python setup.py develop

4. Troubleshooting

See page on

https://github.com/intel/intel-xpu-backend-for-triton/wiki/Possible-Build-Bugs

5. Triton Test Process

For the XPU test, Please refer to the doc in the docs folder.

6. Appendix

6.1. ToolChain

For driver and Toolchains, you could go to this page for the latest alignment.

https://wiki.ith.intel.com/display/PyTorchdGPU/Tool+Chain+and+BKC+alignment

6.2. Driver & System Install

For the Installation guide about the system, you could refer to:

https://dgpu-docs.intel.com/driver/installation.html

Clone this wiki locally