English | 中文
OnyxCore(ˈɑnɪks kɔr) Onyx → Overlapping NPU(黑曜石的层叠纹理=>计算与通信的时空咬合) Core → Kernel Acceleration(核心算力引擎)
Supported Hardware Models: Atlas A3 Series Products Platform: aarch64/x86 Supporting Software
- Driver firmware Ascend HDK 25.0.RC1.1, CANN Community Edition 8.2.RC1.alpha001 and later versions (refer to the "CANN Software Installation Guide" to install the CANN development kit package, as well as the supporting firmware and drivers)
- Before installing CANN software, you need to install the relevant dependency list
- Python >= 3.9
- PyTorch >= 2.5.1, torch-npu >= 2.5.1-7.0.0
- Prepare the CANN environment variables (modify according to the installation path)
source /usr/local/Ascend/ascend-toolkit/set_env.sh- Build the project
Before executing the engineering build script build.sh, modify
_ASCEND_INSTALL_PATHon line 7 of build.sh according to the CANN installation path.
# Building Project
bash build.sh
# Link to the deep_ep_cpp.*.so file based on your settings
ln -s build/lib.linux-aarch64-cpython-39/deep_ep/deep_ep_cpp.cpython-39-aarch64-linux-gnu.so
# Run test cases
bash tests/run_test.sh- Run the installation script to install the
.whlfile into your Python environment
bash install.shInstall required environment variables:
ASCEND_HOME_PATH: CANN installation path
- Execute the environment variables for CANN (modify according to the installation path)
source /usr/local/Ascend/ascend-toolkit/set_env.sh- In the Python project, import
deep_ep.
- If installing the
.whlfile results in the inability to importdeep_epin the project, check whether it is correctly installed in thesite-packagesdirectory of the current Python environment; View installation path:
pip show deep-ep
- If after installing the
.whl, you encounter an issue wheredeep_ep_cppis not found, you need to create a symbolic link of thedeep_ep_cpp*.sofiles from thesite-packages/deep_epdirectory to thesite-packagesdirectory; Execute the following command in thesite-packagesdirectory:
ln -s deep_ep/deep_ep_cpp*.so