ascendai/mindiesd

•Updated 28 days ago

Image

464

ascendai/mindiesd repository overview

⁠MindIE SD

⁠Quick Reference

Item	Value
Image	`mindiesd`
Tags	`v3.1.0-cann9.1.0-torch_npu2.10.0.post4-910b-ubuntu22.04-py3.12`
	`v3.1.0-cann9.1.0-torch_npu2.10.0.post4-a3-ubuntu22.04-py3.12`
	`v3.1.0-cann9.1.0-torch_npu2.10.0.post4-950-ubuntu22.04-py3.12`
Base Images	Atlas 800I A2 inference server: `quay.io/ascend/vllm-omni:v0.28.0`
	Atlas 800I A3 SuperPoD Server: `quay.io/ascend/vllm-omni:v0.28.0-a3`
	950PR&950DT products: `quay.io/ascend/vllm-omni:v0.28.0-a5`
Architecture	`linux/amd64` (x86_64)
OS	Ubuntu 22.04
Python	3.12
CANN	9.1.0
TorchNPU	2.10.0.post4
License	Mulan PSL v2

This image is maintained by the MindIE community⁠.

Get help:

⁠Image Overview

This image combines vLLM-Omni and MindIE SD (Mind Inference Engine Stable Diffusion) into a single container, enabling both multi-modal LLM inference and Stable Diffusion image generation on Atlas NPUs.

It is built on top of the quay.io/ascend/vllm-omni base image (which includes CANN 9.1.0, torch, TorchNPU, vllm, and vllm_ascend). Three variants are provided for different NPU series:

Atlas 800I A2 inference server: based on quay.io/ascend/vllm-omni:v0.28.0, for Atlas 800I A2 inference server
Atlas 800I A3 SuperPoD Server: based on quay.io/ascend/vllm-omni:v0.28.0-a3, for Atlas 800I A3 SuperPoD Server
950PR&950DT products: based on quay.io/ascend/vllm-omni:v0.28.0-a5, for 950PR&950DT products

All three base tags publish linux/amd64 only.

All three variants add the following Atlas tuning and debugging tools:

Component	Version	Description
mindiesd	3.1.0	MindIE Stable Diffusion inference engine
msprobe	0.1.4	Precision debugging tool
msmodelslim	8.2.1	Model compression and quantization tool
msprof-analyze	26.0.0	MindStudio Profiler analysis tool
msprof	(bundled with CANN)	NPU profiling tool

⁠Image Tags & Dockerfile Path

⁠Tag Naming Convention

{version}-{cann version}-{torch_npu version}-{supported product}-{os}-{python version}

Series	Example Tag	Base Image
Atlas 800I A2 inference server	`v3.1.0-cann9.1.0-torch_npu2.10.0.post4-910b-ubuntu22.04-py3.12`	`quay.io/ascend/vllm-omni:v0.28.0`
Atlas 800I A3 SuperPoD Server	`v3.1.0-cann9.1.0-torch_npu2.10.0.post4-a3-ubuntu22.04-py3.12`	`quay.io/ascend/vllm-omni:v0.28.0-a3`
950PR&950DT products	`v3.1.0-cann9.1.0-torch_npu2.10.0.post4-950-ubuntu22.04-py3.12`	`quay.io/ascend/vllm-omni:v0.28.0-a5`

⁠v3.1.0 Version Dockerfile Directory

Each series has a dedicated Dockerfile, archived in the docker/omni directory of the MindIE SD source repository:

Series	Example Tag	Dockerfile
Atlas 800I A2 inference server	`v3.1.0-cann9.1.0-torch_npu2.10.0.post4-910b-ubuntu22.04-py3.12`	Dockerfile⁠
Atlas 800I A3 SuperPoD Server	`v3.1.0-cann9.1.0-torch_npu2.10.0.post4-a3-ubuntu22.04-py3.12`	Dockerfile⁠
950PR&950DT products	`v3.1.0-cann9.1.0-torch_npu2.10.0.post4-950-ubuntu22.04-py3.12`	Dockerfile⁠

Released image tags, including v3.0.0, are listed in Supported Tags⁠.

⁠Quick Start

⁠Pull Base Image

Browse all available tags at quay.io/ascend/vllm-omni⁠.

Atlas 800I A2 inference server:

docker pull quay.io/ascend/vllm-omni:v0.28.0

Atlas 800I A3 SuperPoD Server:

docker pull quay.io/ascend/vllm-omni:v0.28.0-a3

950PR&950DT products:

docker pull quay.io/ascend/vllm-omni:v0.28.0-a5

Tip: You can also use podman pull in place of docker pull if you prefer Podman as your container runtime.

⁠Run the Container

Atlas 800I A2 inference server:

docker run -it --rm --name=mindiesd \
    --privileged \
    --shm-size=1g \
    --device /dev/davinci0 \
    --device /dev/davinci1 \
    --device /dev/davinci2 \
    --device /dev/davinci3 \
    --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 \
    -v /root/.cache:/root/.cache \
    mindiesd:v3.1.0-cann9.1.0-torch_npu2.10.0.post4-910b-ubuntu22.04-py3.12 \
    bash

Atlas 800I A3 SuperPoD Server:

docker run -it --rm --name=mindiesd \
    --privileged \
    --shm-size=1g \
    --device /dev/davinci0 \
    --device /dev/davinci1 \
    --device /dev/davinci2 \
    --device /dev/davinci3 \
    --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 \
    -v /root/.cache:/root/.cache \
    mindiesd:v3.1.0-cann9.1.0-torch_npu2.10.0.post4-a3-ubuntu22.04-py3.12 \
    bash

950PR&950DT products:

docker run -it --rm --name=mindiesd \
    --privileged \
    --shm-size=1g \
    --device /dev/davinci0 \
    --device /dev/davinci1 \
    --device /dev/davinci2 \
    --device /dev/davinci3 \
    --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 \
    -v /root/.cache:/root/.cache \
    mindiesd:v3.1.0-cann9.1.0-torch_npu2.10.0.post4-950-ubuntu22.04-py3.12 \
    bash

Note: The --privileged flag and device mappings are required for NPU access. The host must mount driver libraries (/usr/local/Ascend/driver/lib64), driver version info, DCMI, npu-smi, and Atlas install info into the container.

⁠Build Locally

Clone the MindIE SD repository and build from the docker/omni directory:

Atlas 800I A2 inference server:

git clone https://gitcode.com/Ascend/MindIE-SD.git
cd MindIE-SD
git checkout dev
cd docker/omni

docker build --platform linux/amd64 -t mindiesd:v3.1.0-cann9.1.0-torch_npu2.10.0.post4-910b-ubuntu22.04-py3.12 \
    -f Dockerfile.a2.ubuntu .

Atlas 800I A3 SuperPoD Server:

git clone https://gitcode.com/Ascend/MindIE-SD.git
cd MindIE-SD
git checkout dev
cd docker/omni

docker build --platform linux/amd64 -t mindiesd:v3.1.0-cann9.1.0-torch_npu2.10.0.post4-a3-ubuntu22.04-py3.12 \
    -f Dockerfile.a3.ubuntu .

950PR&950DT products:

git clone https://gitcode.com/Ascend/MindIE-SD.git
cd MindIE-SD
git checkout dev
cd docker/omni

docker build --platform linux/amd64 -t mindiesd:v3.1.0-cann9.1.0-torch_npu2.10.0.post4-950-ubuntu22.04-py3.12 \
    -f Dockerfile.a5.ubuntu .

⁠Customize (Secondary Development)

To add your own dependencies or application code, create a new Dockerfile based on this image:

FROM mindiesd:v3.1.0-cann9.1.0-torch_npu2.10.0.post4-910b-ubuntu22.04-py3.12

# Add your custom packages
RUN pip install --no-cache-dir your-package

# Copy your application
COPY ./your-app /workspace/your-app
WORKDIR /workspace/your-app

⁠Compatibility Changes

Refer to the MindIE SD documentation⁠ for the latest release notes and compatibility information.

⁠License & Disclaimer

This image is licensed under the Mulan Permissive Software License, Version 2 (Mulan PSL v2). See the LICENSE⁠ file for the full text.

By pulling and using this container image, you accept the terms and conditions of the Huawei Container License Agreement. A copy of the license is available at: https://www.hiascend.com/en/legal/ascendhub-download⁠

You agree and undertake that when using Huawei or third-party software in this image, you will comply with the license agreement of the corresponding Huawei or third-party software.

Tag summary

Recent tags

Content type

Image

Digest

sha256:92f09c1cb…

Size

7.2 GB

Last updated

28 days ago

docker pull ascendai/mindiesd:v3.1.0-A5-ubuntu22.04-py3.11