Triton Operator Development Environment Configuration

Core Principles

Prerequisite: Obtain the Latest Matching Requirements (MANDATORY)

• Python version requirements
• torch / torch_npu version requirements
• triton-ascend version requirements
• CANN version requirements
• Version correspondence of each component

Environment Check and Configuration Process

1. CANN Environment Configuration (Step 1)

1. Execute

   npu-smi info

   to check if the driver is loaded successfully
2. Execute

   which bisheng

   to check if the CANN environment is loaded and the npuir compiler is obtained, a path should be output
3. If there is no output, try loading the CANN environment:
   • Priority:

     source /usr/local/Ascend/cann/set_env.sh

   • Alternative:

     source /usr/local/Ascend/ascend-toolkit/set_env.sh

4. Perform step 1 and step 2 checks again. If unsuccessful, remind the user to check and resolve the CANN environment configuration, and suggest downloading and installing CANN from https://www.hiascend.com/cann/download

2. Python Version Check

1. Check the current Python path:

   which python3

2. If failed, try executing:

   export PATH="/usr/bin:$PATH"

3. Perform step 1 check again. If unsuccessful, remind the user to install Python, preferring the miniconda method below
4. If Python3 exists, check the Python version:

   python3 --version

   , which must comply with the requirements in the official quick_start.md documentation

3. Python Environment Installation (On-Demand)

1. If the user needs to install a Python environment, first execute

   conda init bash

   to check if a conda environment exists. If yes, jump to step 5; if no, execute step 2 to install
2. Execute

   uname -m

   to confirm the current system architecture If the system architecture is aarch64, execute:

   wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-aarch64.sh

   If the system architecture is x86_64, execute:

   wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh

3. Execute the installation script:

   bash Miniconda3-latest-Linux-x86_64.sh

   or

   bash Miniconda3-latest-Linux-aarch64.sh

   , activate the miniconda environment after installation as prompted
4. Check if miniconda is installed successfully:

   conda init bash

5. Create a Python environment:

   conda create -n triton python=<official required version>

6. Activate the Python environment:

   conda activate triton

4. torch Configuration

1. Execute:

   pip list | grep "torch"

2. Check the torch / torch_npu versions, which must comply with the requirements in the official documentation
3. If versions do not meet requirements, install according to the documentation:
   • First install torch:

     pip install torch==<official required version>

   • If encountering the error

     ERROR: No matching distribution found for torch==2.7.1+cpu

     , try:
     bash

     pip install torch==2.7.1+cpu --index-url https://download.pytorch.org/whl/cpu

   • Then install torch_npu:

     pip install torch_npu==<official required version>

4. Check if the torch environment is configured successfully:

   python3 -c "import torch; print(torch.__version__)"

5. Run a simple torch code to check if the NPU device is loaded successfully:

   python3 -c "import torch; a = torch.randn(2, 3); print(a)"

   , which should output results similar to:

   tensor([[2.86, 1.0406, 1.5811], [0.8329, 1.0024, 1.3639]])

5. triton-ascend Configuration

1. Execute:

   pip list | grep "triton"

2. Community Triton and Triton-Ascend cannot coexist:
   • If native triton is installed, uninstall it first:

     pip uninstall triton

   • If triton-ascend is also installed, uninstall it first, then reinstall
3. Install the latest triton-ascend package:

   pip install triton-ascend

4. For installing the nightly version, refer to: https://triton-ascend.readthedocs.io/zh-cn/latest/installation_guide.html#nightly-build

Environment Validation (MANDATORY)

scripts/01-vector-add.py

python3 <skill-dir>/scripts/01-vector-add.py

tensor([0.8329, 1.0024, 1.3639,  ..., 1.0796, 1.0406, 1.5811], device='npu:0')
tensor([0.8329, 1.0024, 1.3639,  ..., 1.0796, 1.0406, 1.5811], device='npu:0')
The maximum difference between torch and triton is 0.0

Troubleshooting

pip install torch==2.7.1+cpu --index-url https://download.pytorch.org/whl/cpu

pip uninstall triton

apt-get install g++

sudo apt install zlib1g-dev clang-15 lld-15

Anti-Pattern List (NEVER)

• ❌ Skip reading the official online documentation to obtain the latest matching requirements
• ❌ Skip CANN environment check and directly configure Python
• ❌ Use Python / torch versions that do not meet official requirements
• ❌ Install both native triton and triton-ascend at the same time
• ❌ Start development without environment validation
• ❌ Execute steps in different terminal sessions (must be in the same terminal)

Checklist

• Have you read the official online documentation to obtain the latest matching requirements?
• Is the CANN environment loaded successfully (output from

  npu-smi info

  and

  which bisheng

  )?
• Does the Python version meet official requirements?
• Do the torch / torch_npu versions meet official requirements?
• Has triton-ascend been installed (native triton has been uninstalled)?
• Has the environment validation passed (

  01-vector-add.py

  runs successfully)?

Notes

• If there are still problems after following the above process, you can check the official online documentation for new matching requirements: https://triton-ascend.readthedocs.io/zh-cn/latest/quick_start.html