Updated on 2026-09-03 GMT+08:00

Overview

Description

ModelArts allows you to fine-tune popular open-source models with the MindSpeed-LLM framework.

MindSpeed-LLM

MindSpeed-LLM is a distributed training suite for LLMs based on the Ascend ecosystem. It aims to provide an E2E LLM training solution for Huawei Ascend chip ecosystem partners, including distributed pre-training, distributed instruction fine-tuning, and the corresponding development toolchain, such as data preprocessing, weight transformation, online inference, and baseline evaluation.

Supported LLMs

**Table 1** Supported LLMs
Model	Training Scenario	Megatron	Training Framework	Minimum PUs
Qwen3-8B	SFT	√	MindSpeed-LLM	8x Snt9b2
Qwen3-30B-A3B	SFT	√	MindSpeed-LLM	16 x Snt9b2
Qwen3-32B	SFT	√	MindSpeed-LLM	16 x Snt9b2

Constraints

Snt9b2 is used for training.

Resources and Files Required

**Table 2** Resources and data files required for training
No.	Item	Mandatory	Description
1	ModelArts compute resources	Yes	You are advised to use Snt9b resources in a ModelArts dedicated resource pool. The pool must be created in advance.
2	OBS bucket	Yes	ModelArts requires you to create an OBS bucket for storing the model weight files, training data, training scripts, and training artifacts.
3	Model weight files	Yes	Model files for training. You must prepare them in advance. This document provides the download addresses.
4	Training datasets	Yes	Datasets for training. You must prepare them in advance. This document provides the download addresses.
5	Training scripts	Yes	Dataset preprocessing scripts and training job startup scripts. You must prepare them in advance. This document provides script examples, which can be directly used.
6	Training environment image	No	The training images are already set up on the platform. You do not need to prepare them. Just select an image when you create a training job. The image contains the software required for training, such as the CANN package, driver, PyTorch, and training framework code package.

Procedure

**Table 3** Procedure description
Phase	Step	Description
Preparations	Configuring Cloud Resources	Prepare the dedicated resource pool and OBS bucket required for training.
	Preparing Model Weights and Training Data	Prepare the model weight files, training datasets, and data processing script files, and upload them to the OBS bucket.
	Preparing the Training Scripts	Prepare the script files required for training in advance and upload them to the OBS bucket.
Training	Executing a Training Job	Create a training job. During the creation, select a preset training environment image, select the resource specifications required for training, and set input and output parameters.
Viewing the training results	Viewing the Training Output	View the training logs and performance results.
Viewing the training results	Viewing the Training Output	Obtain the trained model from the output path set in the training job.

Parent topic: Training the Qwen3 Series Models Adapted to the MindSpeed-LLM Framework on ModelArts