Help Center/ ModelArts/ Best Practices/ LLM Training/ Adapting Mainstream Open-Source Models to AscendFactory NPU Training Based on Lite Server/ Version Software Description and Requirements
Updated on 2025-11-04 GMT+08:00
View PDF
Share

Version Software Description and Requirements

Version Differences

Use this guide for AI Compute Service version 6.5.906 or newer. The latest version is 6.5.907. You are advised to use the latest software package and image.

Table 1 Version differences

Version

Description

6.5.907

Compared with 6.5.906, 6.5.907 supports the following new features:

  1. VeRL reinforcement learning framework: The Qwen2.5VL, Qwen3, Qwen2.5 series models are added and support PPO, DAPO, and GRPO.
  2. Llama-Factory framework: The Qwen2.5-14B, Qwen2.5-VL-7B, and Qwen2.5-72B models are added and support DPO reinforcement learning.
  3. MindSpeed-MM framework: The Qwen2.5VL-7B and Qwen2.5VL-3B models are added and support pre-training and fine-tuning.

Compared with 6.5.906, 6.5.907 does not support the following feature changes:

  1. Llama-Factory framework: Internvl2.5-8B, Internvl2.5-38B, and Internvl2.5-78B do not support full-parameter and LoRA fine-tuning.

6.5.906

Compared with 6.5.905, 6.5.906 has the following feature changes:

  1. The MindSpeed_RL reinforcement learning framework enables GRPO algorithm training for the Qwen2.5 model series.
  2. The VeRL reinforcement learning framework works with both the Qwen3-8B LLM and the Qwen2.5-VL multimodal model series.

Base Image Versions

The following tables list the base image addresses and their versions for this tutorial.

Table 2 Base image addresses

Usage

Address

Version

Base Snt9b image

CN Southwest-Guiyang1:

swr.cn-southwest-2.myhuaweicloud.com/atelier/pytorch_ascend:pytorch_2.5.1-cann_8.2.rc1-py_3.11-hce_2.0.2503-aarch64-snt9b-20250729103313-3a25129

CN-Hong Kong:

swr.ap-southeast-1.myhuaweicloud.com/atelier/pytorch_ascend:pytorch_2.5.1-cann_8.2.rc1-py_3.11-hce_2.0.2503-aarch64-snt9b-20250729103313-3a25129

6.5.907

Base Snt9b23 image

CN Southwest-Guiyang1:

swr.cn-southwest-2.myhuaweicloud.com/atelier/pytorch_ascend:pytorch_2.5.1-cann_8.2.rc1-py_3.11-hce_2.0.2503-aarch64-snt9b23-20250729103313-3a25129

CN-Hong Kong:

swr.ap-southeast-1.myhuaweicloud.com/atelier/pytorch_ascend:pytorch_2.5.1-cann_8.2.rc1-py_3.11-hce_2.0.2503-aarch64-snt9b23-20250729103313-3a25129

6.5.907
Table 3 Base image addresses (dedicated for DeepSeek)

Usage

Address

Version

Base Snt9b image (dedicated for DeepSeek)

CN Southwest-Guiyang1:

swr.cn-southwest-2.myhuaweicloud.com/atelier/pytorch_2_1_ascend:pytorch_2.1.0-cann_8.2.rc1-py_3.11-hce_2.0.2503-aarch64-snt9b-20250729103313-3a25129

CN-Hong Kong

swr.ap-southeast-1.myhuaweicloud.com/atelier/pytorch_2_1_ascend:pytorch_2.1.0-cann_8.2.rc1-py_3.11-hce_2.0.2503-aarch64-snt9b-20250729103313-3a25129

6.5.907

Base Snt9b23 image (dedicated for DeepSeek)

CN Southwest-Guiyang1:

swr.cn-southwest-2.myhuaweicloud.com/atelier/pytorch_2_1_ascend:pytorch_2.1.0-cann_8.2.rc1-py_3.11-hce_2.0.2503-aarch64-snt9b23-20250729103313-3a25129

CN-Hong Kong:

swr.ap-southeast-1.myhuaweicloud.com/atelier/pytorch_2_1_ascend:pytorch_2.1.0-cann_8.2.rc1-py_3.11-hce_2.0.2503-aarch64-snt9b23-20250729103313-3a25129

6.5.907
Table 4 Model image versions

Server Model

Model

Version

Snt9b

CANN

8.2.RC1

Driver

25.2.1

PyTorch

2.5.1

2.1.0 (for DeepSeek)

Snt9b23

CANN

8.2.RC1

Driver

25.2.1

PyTorch

2.5.1

2.1.0 (for DeepSeek)

Software Package Obtaining

The following table lists the software version and dependency package required by this solution and how to obtain them.

Table 5 Software versions and download URL

AI Compute Service Version

Software

Description

How to Obtain

6.5.907

AscendCloud-6.5.907-timestamp.zip

The training code used in this tutorial is included.

1. Go to Support-E.

2. Find ModelArts 6.5.907.1.

NOTE:

If the software information does not appear when opening the download link, you lack access permissions. Contact your company's Huawei technical support for assistance with downloading.

Software Package Structure

Key training files in the AscendCloud-LLM code package: 
|——AscendCloud-LLM
  |──llm_train                    # Model training code package
    |──AscendFactory   
      |──examples/                # Directory of the config file
      |──data.tgz                 # Compressed package of sample data
      |──third-party/             # Patch package
      |──src/acs_train_solution/  # Training package
      |──install.sh                # Environment initialization script
      |──Dockerfile               
      |──scripts_install          # Installation script packages of each framework
      |──dependences.yaml         # YAML file of the open-source community code version