Overview

Features

ModelArts is a one-stop AI development and training/inference platform provided by Huawei Cloud, offering a full-lifecycle toolchain that includes algorithm development, model training, deployment, and resource management.

Model training is the core module of ModelArts, supporting you in creating training tasks through GUI operations and APIs. You can use the training module to rapidly iterate on model parameters, validate different combinations of datasets and hyperparameters, and train models that meet your requirements.

The process of developing an AI model is referred to as modeling, which typically involves two stages:

Development stage: Prepare and configure the environment, and debug code so that it can begin deep learning training. You are advised to enable JupyterLab or remote SSH when creating a debug job in ModelArts. For details, see Creating a Debug Training Job.
Experiment stage: Adjust datasets and hyperparameters through multiple rounds of experimentation to train an ideal model. This is recommended to be completed in ModelArts Training Jobs. For details, see Creating a Training Job and Distributed Model Training.

These two processes can transition into one another. Once code is stabilized in the development stage, you move into the experimentation stage to iterate the model by tuning hyperparameters; if performance optimization is needed during the experimentation stage, you can return to the development stage to re-optimize the code.

Figure 1 Model development workflow
Click to enlarge

Advantages

One-stop model development: ModelArts provides an end-to-end AI development workflow from data preparation to model training and deployment, helping you complete full-lifecycle model development with ease.
High-performance training resource support: ModelArts is optimized for Huawei's latest supernode servers based on Ascend chips. Powered by Huawei Collective Communication Library (HCCL) interconnection technology, it delivers ultra-high training bandwidth and powerful compute, enabling you to efficiently complete ultra-large-scale distributed model training tasks.
Compatibility with mainstream training frameworks: ModelArts provides built-in images for mainstream training frameworks, including PyTorch, TensorFlow, MindSpore, MindSpeed, Llama-Factory, and VeOmni, while also supporting custom images, allowing you to adapt flexibly to different scenarios.
High availability solutions for stable job execution: ModelArts offers robust resumable training (checkpointing), a comprehensive training fault detection mechanism, and multiple fast recovery options (at the job, pod, and process levels). These features significantly improve training resource utilization and ensure a high success rate for training jobs.
Observability and operations: Supports integration with TensorBoard, MindStudio Insight, and AOM monitoring metrics, making the entire training process fully visible.

New and Old Consoles

To improve user experience, ModelArts has launched a brand-new console with a more modern interface and more comprehensive functions. It is strongly recommended that you switch to the new console. This section describes the differences between the new and old consoles, how to switch between them, and how to choose the right console for your needs.

Differences Between New and Old Consoles

There are differences between the new and old versions of the ModelArts console in terms of feature entry points, creation methods, and feature support. For detailed differences, see Table 1.

**Table 1** Feature comparison between new and old consoles
Comparison Item	New Console	Old Console
Training job entry point	Redirects to the new page by default.	Requires manual switching. For details, see 2.
Navigation path	Model Build > Model Training	Model Training > Training Jobs
Training experiment management	Not supported	Supported
Quick start/One-click training	Supported	Not supported
Creation method	Fine-tuning job/Custom job	Algorithm management/Common framework/Custom
Publish to assets	Supported	Not supported
High availability & observability configuration	Supported	Not supported

Console Switching Methods
- Switching to the new console: Click New Version in the lower left or upper right corner of the old console Overview page. The switching method is shown in Figure 2.
  Figure 2 Switching to the new console
- Switching to the old console: Click Old Version in the lower left corner of the new console Overview page. The switching method is shown in Figure 3.
  Figure 3 Switching to the old console

Decision Guide for Console Selection

You can decide whether to use the new or old console based on your scenario as described in Table 2.

**Table 2** Decision guide for console selection
Your Scenario	Recommended Version	Reason & Recommendation
You are a new user and want to learn and experience model training.	New console (strongly recommended)	Offers a Getting Started section, one-click training cases, and clear guidance for development/experimentation stages. Provides a better user experience, faster feature iterations, and easier onboarding.
You need to fine-tune LLMs.	New console	Supports the latest features, such as one-click training and fine-tuning jobs, allowing you to complete LLM fine-tuning directly on the console.
You have running old training jobs.	Use old console for viewing/managing existing jobs; use new console for creating new tasks	You can continue monitoring old jobs, but migrating new tasks to the new console offers better tool support.
You rely on specific legacy features.	Old console (temporary use)	Features will gradually be migrated to the new console in the future; adapting as early as possible is recommended.

The entire ModelArts Model Training documentation is based on the new console workflows. If you are still using the old console, see Old Console Features.
The new console page is being rolled out gradually. If you cannot switch between the new and old consoles, the feature may not yet be online in your current region/site. Refer to your actual site interface.

Workflow

You can complete a full training task by following the process shown in Figure 4. Alternatively, you can select appropriate sections based on your training scenario as described in Training Scenario Selection to complete your tasks.

Figure 4 ModelArts model training workflow
Click to enlarge

Table 3 provides an outline of the entire model training process. Refer to the corresponding sections based on your actual training scenario.

**Table 3** Full model training workflow
Task	Subtask	Description
Preparations	Prepare training resources.	Before starting training, ensure that compute resources are in place.
	Prepare training data.	Training data includes training datasets and can also consist of prediction models. Prepare training data before creating a training job. If the data is ready for use without further processing, upload it directly to an OBS bucket. Specify the OBS path as the input parameter when creating the training job. If the training dataset is unlabeled or requires data preprocessing, import the data to the ModelArts data preparation module first. When creating a training job, select a custom dataset from the data asset module for the input parameter location.
	Prepare a training image.	ModelArts provides preset images for various mainstream training frameworks and supports custom images for flexible options. Common preset frameworks provided by ModelArts can be used directly. For details, see ModelArts Preset Images. If preset frameworks do not meet your training requirements, you can build a custom image for training.
	Prepare training code.	Training code consists of the boot file or command and training dependency packages. You can use training scripts provided by preset frameworks or execute tasks using custom training code.
Creating a Training Job	Use basic training features.	Create training jobs via the console visual interface. For details, see Creating a Custom Training Job. Create training jobs by calling an API. For details, see Creating a Training Job Using an API.
Creating a Training Job	Use advanced training features.	ModelArts supports the following advanced training features: Incremental Model Training Distributed Model Training Model Training Optimization and Acceleration Fault Tolerance and Recovery
Managing Model Training Jobs	View training job details.	During or after a training job, you can view parameter settings, job events, and other details on the training job details page. For details, see Viewing Training Jobs and Details.
Managing Model Training Jobs	View training job logs.	Training logs record the execution process and exception information. You can use these logs to locate issues occurred during job execution.

Training Scenario Selection

ModelArts provides a variety of training scenarios, ranging from quick-start scenarios that complete training with a single click to custom configurations for massive distributed training tasks. Whether you are an AI enthusiast with no prior experience or a seasoned LLM expert, you can find the exact training scenario you need.

This section provides an overview of training scenarios tailored to different user requirements, helping you quickly locate the relevant information.

**Table 4** Training scenario selection
User Type	Focus	Reference
Beginner: Interested in model training with no prior experience, looking to understand the LLM training process by creating a training task.	How to quickly run a training task end-to-end.	Creating a One-Click Training Job
Intermediate: Able to independently complete model training tasks and use preset training frameworks.	How to create a training task on ModelArts. How to migrate local training tasks to ModelArts.	Creating a Custom Training Job Creating a Fine-Tuning Training Job Example: Migrating Local PyTorch Training Code to ModelArts Creating a Debug Training Job Managing Model Training Jobs
Expert: Experienced model training expert with in-depth knowledge of ModelArts, looking to execute ultra-large-scale training tasks.	How to build custom images for model training. Executing ultra-large-scale training jobs using distributed training. Automating training workflows using APIs. Optimizing training performance.	Creating a Training Job Using an API Distributed Model Training Model Training Optimization and Acceleration Fault Tolerance and Recovery

Next topic: Preparations