| Official Documentation | Operator Development User Journey | Community Meetings | About Ascend |
- [2026.07.31] Triton-Ascend 3.2.2 official release is now available
- [2026.04.30] Triton-Ascend 3.2.1 official release is now available
- [2026.01.20] Triton-Ascend 3.2.0 official release is now available
More latest news
- [2025.11.14] Triton-Ascend 3.2.0rc4 pre-release is now available:
- Extended the tt.fp_to_fp interface to add FP8 type conversion support
- Added the scatter_ub_to_out interface to support efficient data scatter operations from UB to GM - [2025.09.30] Improved Scan/Sort Triton Python APIs, supporting non-contiguous memory access, and completed adaptation of key Triton operators in vLLM and sglang open-source repositories
- [2025.09.19] Supported Triton-Ascend nightly package extraction
- [2025.08.15] Improved Atomic-class Triton Python API support, completed adaptation of key Triton operators in the Flaggems open-source repository, and provided reference examples for high-performance implementations of simple operators such as Matmul
- [2025.06.30] Supported 85% of Triton Python APIs, supporting contiguous memory access, covering basic usage scenarios
- [2025.05.20] Triton-Ascend is open-sourced, Gitcode repository is alive!
Supported operating systems: linux (aarch64/x86_64)
Supported Ascend products: Atlas A2 products\Atlas A3 products\Ascend 950PR&950DT products
Minimum hardware configuration: single card with 32GB memory (recommended)
Determine and install the Python, CANN, and TorchNPU software versions. This step must be completed before both package installation and source code compilation installation.
-
Python version selection: py3.9-py3.13 are all supported.
-
CANN version selection: You can visit the Ascend community website and follow the community software installation guide to complete the CANN installation and configuration. It is recommended to download and install version 9.1.0.
-
TorchNPU version selection: The currently bundled TorchNPU version is 2.7.1.post8.
If you need to access Ascend NPU computing resources for development or testing, please visit the HiDevLab - Online Development page on the HiDevLab platform to apply for free access.
# Taking the installation of triton-ascend 3.2.2 as an example
pip install triton-ascend --extra-index-url=https://mirrors.huaweicloud.com/ascend/repos/pypiMore source installation
apt update
# Supported clang versions: clang-15, clang-17, clang-20.1.8
apt install zlib1g-dev clang-15 lld-15
apt install ccache # optional
update-alternatives --install /usr/bin/clang clang /usr/bin/clang-15 100
update-alternatives --install /usr/bin/clang++ clang++ /usr/bin/clang++-15 100
pip install ninja cmake wheel pybind11 # build-time dependenciesgit clone https://github.com/triton-lang/triton-ascend.git && cd triton-ascend
git checkout main
pip install -e .# If you need to customize the LLVM build process, you can execute this step first before compiling Triton-Ascend
# Check out the specified version of LLVM source code and apply patches
git clone --no-checkout https://github.com/llvm/llvm-project.git
cd llvm-project
git checkout f6ded0be897e2878612dd903f7e8bb85448269e5
wget https://raw.githubusercontent.com/triton-lang/triton-ascend/main/third_party/ascend/patch/llvm_patch_f6ded0b.patch
git apply llvm_patch_f6ded0b.patch
export LLVM_INSTALL_PREFIX=/path/to/llvm-install
# Build a custom LLVM version
cd {PATH_TO}/llvm_project
mkdir build
cd build
cmake ../llvm \
-G Ninja \
-DCMAKE_C_COMPILER=/usr/bin/clang-15 \
-DCMAKE_CXX_COMPILER=/usr/bin/clang++-15 \
-DCMAKE_LINKER=/usr/bin/lld-15 \
-DCMAKE_BUILD_TYPE=Release \
-DLLVM_ENABLE_ASSERTIONS=ON \
-DLLVM_ENABLE_PROJECTS="mlir;llvm;lld" \
-DLLVM_TARGETS_TO_BUILD="host;NVPTX;AMDGPU" \
-DLLVM_ENABLE_LLD=ON \
-DCMAKE_INSTALL_PREFIX=${LLVM_INSTALL_PREFIX}
ninja install
# Compile Triton-Ascend
git clone https://github.com/triton-lang/triton-ascend.git && cd triton-ascend
LLVM_SYSPATH=${LLVM_INSTALL_PREFIX} \
TRITON_BUILD_WITH_CCACHE=true \
TRITON_BUILD_WITH_CLANG_LLD=true \
TRITON_BUILD_PROTON=OFF \
TRITON_WHEEL_NAME="triton-ascend" \
TRITON_APPEND_CMAKE_ARGS="-DTRITON_BUILD_UT=OFF" \
python3 setup.py installMore Docker image usage
-
We provide a Dockerfile to help you install the Docker environment image. The build process uses the
quay.io/ascend/cannpre-built image as the base image, skipping the CANN installation step and significantly speeding up the build. -
You need to specify the
CANN_BASE_IMAGEparameter via--build-argto select the appropriate CANN base image for your machine. Available CANN base image tags can be found at quay.io/ascend/cann. -
You can check the NPU model on your system using the npu-smi command.
git clone https://github.com/triton-lang/triton-ascend.git && cd triton-ascend
docker build \
--build-arg CANN_BASE_IMAGE=quay.io/ascend/cann:9.1.0-a3-ubuntu22.04-py3.12 \
-t triton-ascend-image:latest -f ./docker/Dockerfile .- To start a container from this image, you can refer to the following command:
docker run -u 0 -dit --shm-size=512g --name=triton-ascend_container --net=host --privileged \
--security-opt seccomp=unconfined \
--device=/dev/davinci0 \
--device=/dev/davinci1 \
--device=/dev/davinci2 \
--device=/dev/davinci3 \
--device=/dev/davinci4 \
--device=/dev/davinci5 \
--device=/dev/davinci6 \
--device=/dev/davinci7 \
--device=/dev/davinci_manager \
--device=/dev/devmm_svm \
--device=/dev/hisi_hdc \
-v /usr/local/dcmi:/usr/local/dcmi \
-v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-v /usr/local/sbin/npu-smi:/usr/local/sbin/npu-smi \
-v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
-v /home:/home \
-v /etc/ascend_install.info:/etc/ascend_install.info \
triton-ascend-image:latest \
/bin/bash
# Enter the container
docker exec -u root -it triton-ascend_container /bin/bash-
Welcome to participate in Triton-Ascend development and code contribution. For details, please refer to the Contribution Guide
-
Please report any bugs you encounter via Issue.