ascendai/mindspeed-core repository overview

⁠MindSpeed Core Docker Image Overview

⁠Quick Reference

⁠Key Field Description of Image Tags

Recommended tag template:

{mindspeed_version}-cann{cann_version}-torch_npu{TorchNPU_version}-{npu_type}-{os}-py{python_version}

The NPU type in the image tag and CANN base image name must be in lowercase: a3, 910b, or 950. The complete --base-image is passed through as-is, so its tag must exactly match the published CANN image name.

Examples:

• v26.1.0_core_r0.12.1-cann9.1.0-torch_npu2.7.1.post8-910b-ubuntu22.04-py3.11
• v26.1.0_core_r0.12.1-cann9.1.0-torch_npu2.7.1.post8-a3-openeuler24.03-py3.11

⁠Latest CANN 9.1.0

All images that support the latest CANN 9.1.0 release are listed below.

⁠Build Parameters

It is recommended to use docker/build.sh as the build entry point. The script supports selecting the CANN base image by operating system, NPU type, Python tag, CANN base image version, and target architecture.

If http_proxy, https_proxy, HTTP_PROXY, HTTPS_PROXY, NO_PROXY, or no_proxy is set on the host, the script automatically forwards it as a Docker build argument. These values are not persisted in the final image.

⁠Quick Start

Default:

cd docker
bash build.sh

Build the A3 + openEuler + aarch64 + CANN 9.1.0 base image:

cd docker
bash build.sh -t a3 -o openeuler24.03 -a aarch64

Build using the full base image name. The script will try to automatically identify the CANN version, NPU type, operating system, and Python version from the image tag:

cd docker
bash build.sh \
  --arch aarch64 \
  --base-image swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-910b-openeuler24.03-py3.11

Download the image:

docker pull swr.cn-south-1.myhuaweicloud.com/ascendhub/mindspeed-core:v26.1.0_core_r0.12.1-cann9.1.0-torch_npu2.7.1.post8-a3-openeuler24.03-py3.11

Run image:

Before copying the startup command below, replace the two paths {path-to-data} and {path-to-weights} in the parameters with the actual host paths for data and model weights. Otherwise, the container will fail to mount the host paths correctly after startup.

docker run -it -d \
  --name mindspeed-core \
  --privileged \
  --network host \
  --ipc=host \
  -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
  -v /usr/local/dcmi:/usr/local/dcmi \
  -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
  -v /etc/ascend_install.info:/etc/ascend_install.info \
  -v {path-to-data}:/data \
  -v {path-to-weights}:/weights \
  swr.cn-south-1.myhuaweicloud.com/ascendhub/mindspeed-core:v26.1.0_core_r0.12.1-cann9.1.0-torch_npu2.7.1.post8-a3-openeuler24.03-py3.11 \
  bin/bash

Enter the started container:

docker exec -it mindspeed-core /bin/bash

⁠Compatibility Notes

• The current version uses a unified Dockerfile plus build script structure, supporting configurable CANN base image selection.
• The default base image uses CANN 9.1.0, 910B, openEuler 24.03, and Python 3.11.
• You can switch the operating system, NPU type (a3, 910b, or 950), target architecture, or CANN base image version via docker/build.sh.
• MindSpeed is cloned to /MindSpeed, and Megatron-LM is cloned to /Megatron-LM.
• The image installs PyTorch, TorchNPU, MindSpeed Core, Megatron-LM, and the Python dependencies in requirements.txt, then restores and verifies the configured PyTorch, TorchNPU, and NumPy versions.

⁠License

MindSpeed is released under the Apache License 2.0. For details, see the LICENSE⁠ file.

As with all Docker images, these images may also contain other software that is subject to other licenses (such as Bash from the base distribution, and any direct or indirect dependencies of the primary software being included).

For any use of pre-built images, it is the image user's responsibility to ensure that any use of this image complies with the relevant licenses of all software contained within it.

⁠Disclaimer

The released Ascend software images are community versions and are not intended for commercial accountability. They are provider solely as references for production practices.