Star0
0
代码
代码
Issues
Pull Requests
Actions
项目讨论
Wiki
分析
Star0
0
ascend-robotascend-robot[docs] The links, structure and expression of "overview" have been modified.
17b775ff创建于 17 天前历史提交

TorchNPU

English | 中文

The plugin is a deep learning adaptation framework based on Atlas, enabling Atlas NPUs to support the PyTorch framework and providing users of the PyTorch framework with the powerful computing power of Atlas AI processors.

Quick Reference

Supported Tags and Dockerfile Usage

Tag Specification

Tags follow the format:

<TorchNPU_version>-<CANN_version>-<chip>-<os>-<python_version>
Field value Description
TorchNPU Version 2.12.0 For details, see the version notes in the readme.
CANN_version 9.1.0 For details, see the version notes in the readme.
Chip Specific example values can be found in the CANN mirror chip model identifier
OS ubuntu22.04 / openeuler24.03 OS distribution used for the base image
Python Version py3.12 Major Python version pre-installed in the image

TorchNPU timely adaptation version (pt2.13)

  • 2.13.0rc1-cann9.1.0-310p-ubuntu22.04-py3.12
  • 2.13.0rc1-cann9.1.0-310p-openeuler24.03-py3.12
  • 2.13.0rc1-cann9.1.0-910b-ubuntu22.04-py3.12
  • 2.13.0rc1-cann9.1.0-910b-openeuler24.03-py3.12
  • 2.13.0rc1-cann9.1.0-a3-ubuntu22.04-py3.12
  • 2.13.0rc1-cann9.1.0-a3-openeuler24.03-py3.12
  • 2.13.0rc1-cann9.1.0-950-ubuntu22.04-py3.12
  • 2.13.0rc1-cann9.1.0-950-openeuler24.03-py3.12

Tag list for version 26.1.1

As shown below are all the image tags of the latest 26.1.1 version released by TorchNPU. For all the historical versions, please refer to Supported Tags

  • 2.12.0.post2-cann9.1.0-310p-ubuntu22.04-py3.12
  • 2.12.0.post2-cann9.1.0-310p-openeuler24.03-py3.12
  • 2.12.0.post2-cann9.1.0-910b-ubuntu22.04-py3.12
  • 2.12.0.post2-cann9.1.0-910b-openeuler24.03-py3.12
  • 2.12.0.post2-cann9.1.0-a3-ubuntu22.04-py3.12
  • 2.12.0.post2-cann9.1.0-a3-openeuler24.03-py3.12
  • 2.12.0.post2-cann9.1.0-950-ubuntu22.04-py3.12
  • 2.12.0.post2-cann9.1.0-950-openeuler24.03-py3.12
  • 2.11.0.post2-cann9.1.0-310p-ubuntu22.04-py3.12
  • 2.11.0.post2-cann9.1.0-310p-openeuler24.03-py3.12
  • 2.11.0.post2-cann9.1.0-910b-ubuntu22.04-py3.12
  • 2.11.0.post2-cann9.1.0-910b-openeuler24.03-py3.12
  • 2.11.0.post2-cann9.1.0-a3-ubuntu22.04-py3.12
  • 2.11.0.post2-cann9.1.0-a3-openeuler24.03-py3.12
  • 2.11.0.post2-cann9.1.0-950-ubuntu22.04-py3.12
  • 2.11.0.post2-cann9.1.0-950-openeuler24.03-py3.12
  • 2.10.0.post6-cann9.1.0-310p-ubuntu22.04-py3.12
  • 2.10.0.post6-cann9.1.0-310p-openeuler24.03-py3.12
  • 2.10.0.post6-cann9.1.0-910b-ubuntu22.04-py3.12
  • 2.10.0.post6-cann9.1.0-910b-openeuler24.03-py3.12
  • 2.10.0.post6-cann9.1.0-a3-ubuntu22.04-py3.12
  • 2.10.0.post6-cann9.1.0-a3-openeuler24.03-py3.12
  • 2.10.0.post6-cann9.1.0-950-ubuntu22.04-py3.12
  • 2.10.0.post6-cann9.1.0-950-openeuler24.03-py3.12
  • 2.9.0.post8-cann9.1.0-310p-ubuntu22.04-py3.12
  • 2.9.0.post8-cann9.1.0-310p-openeuler24.03-py3.12
  • 2.9.0.post8-cann9.1.0-910b-ubuntu22.04-py3.12
  • 2.9.0.post8-cann9.1.0-910b-openeuler24.03-py3.12
  • 2.9.0.post8-cann9.1.0-a3-ubuntu22.04-py3.12
  • 2.9.0.post8-cann9.1.0-a3-openeuler24.03-py3.12
  • 2.9.0.post8-cann9.1.0-950-ubuntu22.04-py3.12
  • 2.9.0.post8-cann9.1.0-950-openeuler24.03-py3.12
  • 2.7.1.post10-cann9.1.0-310p-ubuntu22.04-py3.12
  • 2.7.1.post10-cann9.1.0-310p-openeuler24.03-py3.12
  • 2.7.1.post10-cann9.1.0-910b-ubuntu22.04-py3.12
  • 2.7.1.post10-cann9.1.0-910b-openeuler24.03-py3.12
  • 2.7.1.post10-cann9.1.0-a3-ubuntu22.04-py3.12
  • 2.7.1.post10-cann9.1.0-a3-openeuler24.03-py3.12
  • 2.7.1.post10-cann9.1.0-950-ubuntu22.04-py3.12
  • 2.7.1.post10-cann9.1.0-950-openeuler24.03-py3.12

Dockerfile build parameters

See dockerfile:dockerfile

parameters Description Required Reference Source Value
TORCH_VERSION Full TorchNPU version number Yes TorchNPU repo releases 2.12.0.post2
CHIP_ARCH chip architecture identifier Yes CANN image tag rules Specific example values can be found in the CANN mirror
OS Base image operating system Yes CANN image tag rules ubuntu / openeuler
OS_VERSION Operating system version Yes CANN image tag rules 22.04 / 24.03
PY_VERSION Python version pre-installed in base image Yes CANN image tag rules 3.11
CANN_VERSION CANN toolkit version Yes CANN base image repo 9.1.0
ARCH Host hardware architecture Yes Environment hardware arm / x86
PY_TAG Python package ABI tag (cp + version number) Yes Strictly match PY_VERSION cp311
TORCH_NPU_RELEASE_VERSION Official TorchNPU release tag (including PyTorch version) Yes TorchNPU repo releases v26.1.1-pytorch2.12.0.post2
TORCH_NPU_PATCH_TAG TorchNPU version number in the release package name Yes TorchNPU repo releases 2.12.0.post2
MANYLINUX_VER PyPI package compatible system version No torch official wheel spec manylinux_2_28
PIP_MIRROR_URL pip installation source URL (Tsinghua mirror by default) No PyPI mirror sources https://pypi.tuna.tsinghua.edu.cn/simple

Parameter Sources

  1. Image tags, operating systems, and version information:CANN Base Image Repository tag

  2. TORCH_NPU_RELEASE_VERSION and TORCH_NPU_PATCH_TAG parameter values are from TorchNPU Official Release Versions. Example with a wheel download URL:

https://gitcode.com/Ascend/pytorch/releases/download/v26.1.1-pytorch2.12.0/torch_npu-2.12.0.post2-cp310-cp310-manylinux_2_28_aarch64.whl

  • TORCH_NPU_RELEASE_VERSION is the part between download/ and /torch_npu-, e.g. v26.1.1-pytorch2.12.0.
  • TORCH_NPU_PATCH_TAG is the part between torch_npu- and -cp310, e.g. 2.12.0.post2.

Quick Start

Build TorchNPU Image

Taking the construction of the 2.12.0.post2-a3-ubuntu22.04-py3.12 image as an example:

docker build \
  --build-arg CHIP_ARCH=a3 \
  --build-arg OS=ubuntu \
  --build-arg OS_VERSION=22.04 \
  --build-arg PY_VERSION=3.11 \
  --build-arg CANN_VERSION=9.1.0 \
  --build-arg ARCH=arm \
  --build-arg PY_TAG=cp311 \
  --build-arg TORCH_NPU_RELEASE_TAG=v26.1.1-pytorch2.12.0 \
  --build-arg TORCH_NPU_PATCH_TAG=2.12.0.post2 \
  -t image_name:tag \
  -f Dockerfile .

Note: If your build environment requires a proxy, pass proxy variables via --build-arg, for example:

docker build \
  --build-arg HTTP_PROXY=http://proxy.example.com:8080 \
  --build-arg HTTPS_PROXY=http://proxy.example.com:8080 \
  --build-arg NO_PROXY=localhost,127.0.0.1 \
  ... \
  -f Dockerfile .

Replace the proxy address and port with your actual environment values.

Run TorchNPU Container

docker run \
    --name pta_container \
    --device /dev/davinci1 \
    --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/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
    -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -it ascend/pta:tag bash

Running CANN containers on 950 series aarch64 architecture products

docker run \
    --name pta_container \
    --device /dev/davinci1 \
    --device /dev/davinci_manager \
    --device /dev/hisi_hdc \
    --device /dev/ummu \
    --device /dev/uburma \
    -v /usr/local/dcmi:/usr/local/dcmi \
    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
    -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
    -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -it ascend/pta:tag bash

Secondary Development

# Use TorchNPU image as base image and add user software
FROM quay.io/ascend/torch-npu:2.12.0.post2-cann9.0.0-910b-ubuntu22.04-py3.12 

RUN apt update -y && \
    apt install gcc ...

...

License/Disclaimer

See the license information for TorchNPU included in these images.

Like all container images, pre-installed software packages (Python, system libraries, etc.) may be subject to their own licenses.

The released Atlas software images are community versions; they are not intended for commercial use and serve solely as references for production practices.

Liability disclaimers are displayed in the image startup information and on the Atlas image platform.