Help Center/ ModelArts/ Image Management/ ModelArts Preset Images
Updated on 2026-06-05 GMT+08:00
View PDF
Share

ModelArts Preset Images

When using ModelArts for machine learning development, you often need to configure the environment based on different frameworks and CUDA versions, which may increase configuration complexity and cause usage issues. To address these challenges, ModelArts provides preset images based on different frameworks and CUDA versions. You can directly select a preset image provided by ModelArts when creating a notebook instance, training a model, or performing real-time inference. By using preset images, you can quickly start the development environment, reduce configuration time, and focus on model development and training, thereby improving development efficiency.

Constraints

Container images based on the HCE system have the following restrictions:

The Docker version of the resource pool cluster must meet the following requirements: EulerOS 2.9 ≥ h59, EulerOS 2.10 ≥ h53, and EulerOS 2.11 ≥ h2. Otherwise, system compatibility problems may occur.

Log in to the resource pool cluster node to query the Docker version.

docker --version

Naming Conventions for Preset Images

Preset images in ModelArts comply with certain naming conventions. You can learn the basic information about an image based on its name. Generally, an image name contains the following fixed fields. You are advised to use unified naming conventions when adding custom images.

Table 1 Naming conventions for preset images

Framework

Example Name of a Preset Image

Image Name Interpretation

VeOmni

veomni_v0.1.6-pytorch_2.7.1-cann_8.3.rc1-py_3.11-hce_2.0.2509-aarch64-snt9b
  • veomni_v0.1.6: indicates that the Veomni framework version is 0.1.6.
  • pytorch_2.7.1: indicates that the PyTorch framework version is 2.7.1.
  • cann_8.3.rc1: indicates that the version of the Compute Architecture for Neural Networks (CANN) software stack is 8.3 Release Candidate 1.
  • py_3.11: indicates that the Python version is 3.11.
  • hce_2.0.2509: indicates that the Huawei Cloud EulerOS version is 2.0.2509.
  • aarch64: indicates that the image is compiled for the 64-bit Arm architecture.
  • snt9b: indicates that the image is adapted to the Ascend snt9b chip.

MindSpeed-LLM

mindspeed_llm_2.2.0-pytorch_2.7.1-cann_8.2.rc2-py_3.11-hce_2.0.2509-aarch64-snt9b
  • mindspeed_llm_2.2.0: indicates that the MindSpeed Large Language Model (LLM) framework version is 2.2.0.
  • pytorch_2.7.1: indicates that the PyTorch framework version is 2.7.1.
  • cann_8.2.rc2: indicates that the version of the CANN software stack is 8.2 Release Candidate 2.
  • py_3.11: indicates that the Python version is 3.11.
  • hce_2.0.2509: indicates that the Huawei Cloud EulerOS version is 2.0.2509.
  • aarch64: indicates that the image is compiled for the 64-bit Arm architecture.
  • snt9b: indicates that the image is adapted to the Ascend snt9b chip.

VeRL

verl_0.7.0-pytorch_2.7.1-cann_8.3.rc1-py_3.11-hce_2.0.2509-aarch64-snt9b
  • verl_0.7.0: indicates that the VeRL framework version is 0.7.0.
  • pytorch_2.7.1: indicates that the PyTorch framework version is 2.7.1.
  • cann_8.3.rc1: indicates that the version of the CANN software stack is 8.3 Release Candidate 1.
  • py_3.11: indicates that the Python version is 3.11.
  • hce_2.0.2509: indicates that the Huawei Cloud EulerOS version is 2.0.2509.
  • aarch64: indicates that the image is compiled for the 64-bit Arm architecture.
  • snt9b: indicates that the image is adapted to the Ascend snt9b chip.

AReaL

areal_0.5.1-pytorch_2.7.1-cann_8.3.rc1-py_3.11-hce_2.0.2509-aarch64-snt9b
  • areal_0.5.1: indicates that the AReaL framework version is 0.5.1.
  • pytorch_2.7.1: indicates that the PyTorch framework version is 2.7.1.
  • cann_8.3.rc1: indicates that the version of the CANN software stack is 8.3 Release Candidate 1.
  • py_3.11: indicates that the Python version is 3.11.
  • hce_2.0.2509: indicates that the Huawei Cloud EulerOS version is 2.0.2509.
  • aarch64: indicates that the image is compiled for the 64-bit Arm architecture.
  • snt9b: indicates that the image is adapted to the Ascend snt9b chip.

MindSpore

mindspore_2.7.1-cann_8.3.rc1-py_3.11-euler_2.10.11-aarch64-snt9b
  • mindspore_2.7.1: indicates that the MindSpore framework version is 2.7.1.
  • cann_8.3.rc1: indicates that the version of the CANN software stack is 8.3 Release Candidate 1.
  • py_3.11: indicates that the Python version is 3.11.
  • euler_2.10.11: indicates that the EulerOS version is 2.10.11.
  • aarch64: indicates that the image is compiled for the 64-bit Arm architecture.
  • snt9b: indicates that the image is adapted to the Ascend snt9b chip.

TensorFlow

tensorflow_2.1.0-cuda_10.1-py_3.7-ubuntu_18.04-x86_64
  • tensorflow_2.1.0: indicates that the TensorFlow framework version is 2.1.0.
  • cuda_10.1: indicates that the CUDA version is 10.1.
  • py_3.7: indicates that the Python version is 3.7.
  • ubuntu_18.04: indicates that the OS is Ubuntu 18.04.
  • x86_64: indicates that the image is compiled for x86_64.

PyTorch

pytorch_2.7.1-cann_8.3.rc1-py_3.11-hce_2.0.2509-aarch64-snt9b
  • pytorch_2.7.1: indicates that the PyTorch framework version is 2.7.1.
  • cann_8.3.rc1: indicates that the version of the CANN software stack is 8.3 Release Candidate 1.
  • py_3.11: indicates that the Python version is 3.11.
  • hce_2.0.2509: indicates that the Huawei Cloud EulerOS version is 2.0.2509.
  • aarch64: indicates that the image is compiled for the 64-bit Arm architecture.
  • snt9b: indicates that the image is adapted to the Ascend snt9b chip.

Capabilities of Preset Images

ModelArts provides official images based on different machine learning frameworks. The following describes the official images based on mainstream frameworks such as VeOmni, MindSpeed-LLM, VeRL, AReaL, and Conda.

VeOmni

VeOmni is an open-source full-modal distributed training framework developed by the seed team and built on PyTorch. With models as the core, the framework decouples distributed parallel logic from the model computation process, supports flexible combination of multiple parallel policies (such as FSDP, SP, and EP), and can be efficiently extended to training scenarios of ultra-long sequences and large-scale MoE models. VeOmni provides lightweight, full-modal APIs to streamline access to multimodal codecs. It incorporates system-level optimizations, including dynamic batch processing and efficient operators, to boost both training efficiency and stability.

Highlights

  • Decoupling between model computation and parallelism: Based on PyTorch's native distributed capabilities, the framework decouples parallelism logic from model computation, supporting flexible combinations of Fully Sharded Data Parallel (FSDP), Sequence Parallelism (SP), and Expert Parallelism (EP).
  • Full-modal access: Lightweight full-modal APIs are provided to simplify the access process of multimodal (such as visual and voice) codecs and support complex multimodal model training.
  • Ascend operator-level optimization: The framework is deeply optimized for PyTorch NPU (torch_npu 2.7.1), integrating dynamic batching and high-performance fused operators to enhance computing throughput across full-modal scenarios.
  • Ultra-large-scale scalability: The framework enables efficient training of ultra-long sequences and is optimized for large-scale MoE architectures to ensure training stability.

MindSpeed-LLM

MindSpeed-LLM is a distributed training suite for LLMs built on the Ascend computing ecosystem. It deeply integrates the MindSpeed acceleration library and Megatron-LM core architecture to provide end-to-end large model training solutions for partners. This image integrates distributed pre-training, instruction fine-tuning, and full-process toolchain (data preprocessing, weight conversion, operator acceleration, and baseline evaluation) to implement an out-of-the-box Huawei Ascend NPU development environment.

Core Component Version

The following core components have been pre-installed and adapted to the environment in this image:

Highlights

  • Ascend native acceleration: MindSpeed is deeply integrated to optimize the instruction set for torch_npu and support efficient mixed precision training and operator fusion.
  • Distributed training: Based on the Megatron-LM core, multi-dimensional parallelism policies such as tensor parallelism (TP) and pipeline parallelism (PP) are supported.
  • Full-stack toolchain: It provides built-in weight conversion tools (supporting format conversion of mainstream open-source models to NPU-compatible formats), distributed data preprocessing, and Ray-based resource scheduling.
  • Ecosystem compatibility: Mainstream LLM architectures are supported, and the PEFT and Transformers libraries are integrated to facilitate the migration of open-source community models.

VeRL

VeRL is an easy-to-use and high-performance open-source framework in the reinforcement learning field. With its modular algorithm library and native distributed training features, VeRL has become a preferred tool for implementing reinforcement learning solutions in scientific research and industry. This image deeply integrates the VeRL core framework, covering the entire process of reinforcement learning algorithm R&D, policy training, and simulation verification.

Application Scenarios

  • Reinforcement learning algorithm development and experiment: This image can be used to quickly design and verify reinforcement learning algorithms in Notebook.
  • Training and tuning: A single device or multiple devices are used to perform large-scale reinforcement learning training in a training job.
  • Education and training: A stable and consistent lab environment is provided to facilitate course teaching and skill training.

AReaL

As one of the mainstream open-source frameworks in the reinforcement learning field, AReaL is widely favored due to its fully asynchronous training and inference and active community ecosystem. This image integrates AReaL and its ecosystem components (vLLM and vLLM-Ascend) and provides multiple version combinations to meet the requirements of different development, training, and deployment scenarios.

ModelArts provides flexible combinations of AReaL, Python, and CANN versions (applicable to Huawei Ascend NPUs). You can select a proper image version to implement an out-of-the-box reinforcement learning environment.

Highlights

  • Pre-installed dependencies
    • Common scientific computing libraries: NumPy and Pandas
    • Inference frameworks: vLLM and vLLM-Ascend
    • Model deployment tool: TensorBoard
  • Out-of-the-box availability: No additional environment configuration is required. The best practice script is directly started to automatically start the cluster and training.

Application Scenarios

This framework applies to most scenarios of reinforcement learning.

MindSpore

MindSpore is a deep learning framework in all scenarios, aiming to achieve easy development, efficient execution, and all-scenario unified deployment. MindSpore preset images have MindSpore pre-installed and provide multiple combinations of Python, CANN, EulerOS, and Ubuntu versions. They support Ascend accelerator card scenarios.

TensorFlow

TensorFlow is an end-to-end platform that enables researchers to advance cutting-edge technologies in the field of machine learning and allows developers to easily build and deploy machine learning-driven applications. TensorFlow preset images have TensorFlow pre-installed and provide multiple combinations of Python, CUDA, CANN, EulerOS, and Ubuntu versions. They support GPU and Ascend accelerator card scenarios.

PyTorch

As one of the mainstream open-source frameworks in the deep learning field, PyTorch is widely favored for its dynamic graph mechanism, Python-first design, and active community ecosystem. PyTorch preset images integrate PyTorch and its ecosystem components (Torchvision and Torchaudio) and provide multiple version combinations to meet the requirements of different development, training, and deployment scenarios.

ModelArts provides flexible combinations of PyTorch, Python, CUDA (for NVIDIA GPUs), and CANN (for Huawei Ascend NPUs), and is compatible with multiple OSs, such as Ubuntu and EulerOS. You can select a proper image version based on your hardware platform (NVIDIA GPU or Huawei Ascend NPU) and project requirements to implement an out-of-the-box deep learning environment.

Highlights
  • Multi-hardware support
    • NVIDIA GPU: supports CUDA acceleration and provides optimization libraries such as cuDNN and NCCL.
    • Huawei Ascend NPU: integrates the torch_npu plug-in to support mixed precision training and distributed training.
  • Pre-installed dependencies
    • Common scientific computing libraries: NumPy, SciPy, and Pandas
    • Visual processing libraries: OpenCV and Pillow
    • Model deployment tools: ONNX Runtime and TensorBoard
    • Development tools: JupyterLab, IPython, tqdm, and more
  • Out-of-the-box availability: No additional environment configuration is required. You can directly start notebook instances, real-time services, or training jobs.

Application Scenarios

  • AI development and experiment: This image can be used to quickly design, debug, and visualize model prototypes in Notebook.
  • Training and tuning: A single device or multiple devices are used to perform large-scale data training in a training job.
  • Deployment and servitization: You can deploy a trained model as a RESTful API through a real-time service.
  • Education and training: A stable and consistent lab environment is provided to facilitate course teaching and skill training.

Core Images

The following lists the core preset images based on the PyTorch, MindSpore, and TensorFlow frameworks. The images available in each region may vary.

Table 2 PyTorch preset images

Image

Supported Chip

Applicable Scope

Created

Description

Image Address

pytorch_2.7.1-cann_8.5.2-py_3.12-hce_2.0.2512-aarch64-snt9b23

Ascend Snt9b2x (such as Snt9b23)

Notebook, training, and inference deployment

2026/04/29

Updated CANN to 8.5.2, Python to 3.12, and Huawei Cloud EulerOS to 2.0.2512.

Viewing the image Address

pytorch_2.7.1-cann_8.5.2-py_3.12-hce_2.0.2512-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/04/29

Updated CANN to 8.5.2, Python to 3.12, and Huawei Cloud EulerOS to 2.0.2512.

pytorch_2.7.1-cann_8.5.2-py_3.12-euler_2.10.11-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/04/29

Updated CANN to 8.5.2, Python to 3.12, and EulerOS to 2.10.11.

pytorch_2.6.0-cann_8.2.rc1-py_3.11-euler_2.10.11-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated PyTorch to 2.6.0, CANN to 8.2.RC1, and the driver to Ascend HDK 25.2.0.

pytorch_2.1.0-cann_8.1.rc1-py_3.10-euler_2.10.11-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated CANN to 8.1.RC1.B150 and the driver to Ascend HDK 25.0.RC1.

pytorch_2.1.0-cann_8.0.rc3-py_3.9-euler_2.10.10-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated CANN to 8.0.rc3 and the driver to Ascend HDK 24.1.RC3.

pytorch_2.1.0-cann_8.0.rc2-py_3.9-euler_2.10.7-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated CANN to 8.0.rc2 and the driver to Ascend HDK 24.1.RC2.

pytorch_2.1.0-cann_8.0.rc1-py_3.9-euler_2.10.7-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated CANN to 8.0.rc1.

pytorch_2.1.0-cann_8.0.0-py_3.10-euler_2.10.11-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated CANN to 8.0.0.beta1 and the driver to Ascend HDK 24.1.0.

pytorch_1.11.0-cann_8.0.rc2-py_3.9-euler_2.10.7-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated CANN to 8.0.rc2 and the driver to Ascend HDK 24.1.RC2.

pytorch_1.11.0-cann_8.0.rc1-py_3.9-euler_2.10.7-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated CANN to 8.0.rc1.

pytorch_2.7.1-cann_8.3.rc1-py_3.11-euler2.10.11-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2025-12-08

Upgraded Python to 3.11, upgraded third-party software, and more.

pytorch_1.11.0-cann_6.3.2-py_3.7-euler_2.10.7-aarch64-d910b

Ascend snt9b

Notebook, training, and inference deployment

2023/09/14

Updated CANN to 6.3.2 and Python to 3.7.

pytorch_2.7.1-cann_8.3.rc1-py_3.11-hce_2.0.2509-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2025-11-11

Upgraded Python to 3.11 and OS to HCE 2.0.2509, and updated AReaL.
Table 3 MindSpore preset images

Preset Image

Supported Chip

Applicable Scope

Created

Description

Image Address

mindspore_2.7.2-cann_8.5.2-py_3.11-euler_2.10.11-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/04/29

Updated CANN to 8.5.2, Python to 3.11, and EulerOS to 2.10.11.

Viewing the image Address

mindspore_2.7.2-cann_8.5.2-py_3.11-hce_2.0.2512-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/04/29

Updated CANN to 8.5.2, Python to 3.11, and Huawei Cloud EulerOS to 2.0.2512.

mindspore_ascend:mindspore_2.7.2-cann_8.5.2-py_3.11-hce_2.0.2512-aarch64-snt9b23

Ascend Snt9b2x (such as Snt9b23)

Notebook, training, and inference deployment

2026/04/29

Updated CANN to 8.5.2, Python to 3.11, and Huawei Cloud EulerOS to 2.0.2512.

mindspore_2.7.0rc1-cann_8.2.rc1-py_3.11-euler_2.10.11-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated MindSpore to 2.7.0rc1, CANN to 8.2.RC1, and the driver to Ascend HDK 25.2.0.

mindspore_2.6.0rc1-cann_8.1.rc1-py_3.10-euler_2.10.11-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated MindSpore to mindspore_2.6.0rc1, CANN to 8.1.RC1.B150, and the driver to Ascend HDK 25.0.RC1.

mindspore_2.4.10-cann_8.0.0-py_3.10-euler_2.10.11-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated MindSpore to 2.4.10, CANN to 8.0.0.beta1, and the driver to Ascend HDK 24.1.0.

mindspore_2.4.0-cann_8.0.rc3-py_3.9-euler_2.10.10-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated CANN to 8.0.rc3 and the driver to Ascend HDK 24.1.RC3.

mindspore_2.3.0-cann_8.0.rc2-py_3.9-euler_2.10.7-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated MindSpore to 2.3.0, CANN to 8.0.rc2, and the driver to Ascend HDK 24.1.RC2.

mindspore_2.3.0-cann_8.0.rc1-py_3.9-euler_2.10.7-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2026/02/04

Updated MindSpore to 2.3.0-rc4 and CANN to 8.0.rc1.

Taken ma-cau1.1.6 and ma-cau-adapter1.1.3 offline.

mindspore_2.1.0-cann_6.3.2-py_3.7-euler_2.10.7-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2023/11/17

Updated CANN to 6.3.2 and Python to 3.7.

mindspore_2.2.0-cann_7.0.1-py_3.9-euler_2.10.7-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2023/08/11

Updated CANN to 7.0.1 and Python to 3.9.

mindspore_2.7.1-cann_8.3.rc1-py_3.11-euler_2.10.11-aarch64-snt9b

Ascend snt9b

Notebook, training, and inference deployment

2025-12-05

Upgraded Python to 3.11, upgraded third-party software, and more.
Table 4 TensorFlow preset images

Preset Image

Supported Chip

Applicable Scope

Updated

Description

Image Address

tensorflow_2.1.0-cuda_10.1-py_3.7-ubuntu_18.04-x86_64

GPU (CUDA 10.1)

Notebook, training, and inference deployment

2022-09-26

Updated CUDA to 10.1 and Python to 3.7.

Viewing the image Address

Viewing the Preset Image Version List/Image Address

  1. Log in to the ModelArts console. In the navigation pane on the left, choose Asset Management > Image Management.
  2. On the page that appears, view the image name, organization, and total number of versions.
  3. Click the image name to view the preset image version list, including the image version, status, resource type, image size, and SWR address.

Viewing Details About an Image Component

You can use either of the following methods to view details about an image component:

  • Method 1: View the details in a container.
    1. Run the following command to start the container: 
      docker run -it image_name bash

      Replace image_name with the actual image path. For details about how to obtain the image name, see Viewing the Preset Image Version List/Image Address.

    2. After accessing the container, run the following command to view the version information: 
      pip list
  • Method 2: View the details in a notebook instance.
    • Use a preset image to create a notebook instance and access the instance in JupyterLab mode. For details, see Using JupyterLab to Develop and Debug Code Online.
    • Open a Terminal interface, and then run the following command to view the version information: 
      pip list
Figure 1 Viewing image component details