FAQ
【免费下载链接】asnumpy哈尔滨工业大学计算学部苏统华、王甜甜老师团队联合华为CANN团队开发的华为昇腾NPU原生Numpy仓库项目地址: https://gitcode.com/cann/asnumpy
Back to README
Frequently asked questions about installing and using AsNumpy.
How do I check if CANN is correctly installed?
cat /usr/local/Ascend/ascend-toolkit/latest/version.cfgWhat should I do if I encounter a compilation error?
- Confirm CMake version >= 3.26:
cmake --version - Confirm GCC version >= 11.2:
gcc --version - Ensure the CANN environment variable is set:
export ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest - Try a clean rebuild:
pip install -e . --no-build-isolation
How compatible is AsNumpy with NumPy?
AsNumpy is designed to be API-compatible with NumPy, but operator coverage is not yet complete. The current version (v0.2.0) covers the most common math, logic, sorting, and array-creation APIs.
The roadmap target is to cover thetop 100 most-used NumPy APIsby v1.0. See the README roadmap for the full plan.
Why is AsNumpy slower than NumPy for small arrays?
For small tensors (e.g., 500×500), NPU kernel launch overhead dominates the measured time. NPU acceleration becomes significant starting around 1000×1000 (float32). See Benchmarks for detailed data.
Do I need to manually initialize or finalize the NPU?
No. AsNumpy handles device initialization automatically onimport asnumpyand releases the device on program exit. You only need to callap.set_device(n)if you want to select a specific NPU other than device 0.
【免费下载链接】asnumpy哈尔滨工业大学计算学部苏统华、王甜甜老师团队联合华为CANN团队开发的华为昇腾NPU原生Numpy仓库项目地址: https://gitcode.com/cann/asnumpy
创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考
平台agentheroes：构建AI协作应用的开源框架)
