ModelArts Best Practices

Adapting Mainstream Open-Source Models to AscendFactory NPU Training

Pre-training, SFT full-parameter fine-tuning training, and LoRA fine-tuning training

Describes the training process of mainstream open-source models, such as DeepSeek, Llama, Qwen3, and Qwen-VL series, based on ModelArts. The AscendFactory framework and AI Compute Service NPUs are used for training.

Adapting Mainstream Open-Source Models for NPU Inference on Ascend-vLLM with PyTorch

Inference deployment, inference performance test, inference accuracy test, and inference model quantization

Describes the inference deployment process of mainstream open-source models, such as Llama, Qwen3, and Qwen-VL series, based on ModelArts. The Ascend-vLLM framework and AI Compute Service NPUs are used for inference.

Adapting Stable Diffusion for NPU Inference with Diffusers/ComfyUI and Lite Server (6.5.907)

SD1.5, SDXL, SD3.5, and Hunyuan model inference

Describes the inference process of mainstream image generation models based on ModelArts Lite Server. AI Compute Service NPUs are used for inference.

After the inference service is started, it can be used in image generation scenarios.

Stable Diffusion XL Inference Guide Based on ModelArts Notebook (6.5.907)

SDXL model inference

Describes the inference process of mainstream image generation models based on ModelArts Standard Notebook. AI Compute Service NPUs are used for inference.

After the inference service is started, it can be used in image generation scenarios.

Inference Guide for Wan Series Video Generation Models Adapted to NPU via Lite Server

Wan series model inference

Describes the inference process of Wan series models based on ModelArts Lite Server. The PyTorch framework and AI Compute Service NPUs are used for inference.

ModelArts Standard Permission Management

IAM permission configuration and management

Permission assignment for IAM users

Assign specific ModelArts operation permissions to the IAM users under a Huawei Cloud account. This prevents exceptions from occurring due to permissions when the IAM users access ModelArts.

Building a Handwritten Digit Recognition Model with ModelArts Standard

PyTorch

Algorithm customization

Handwritten digit recognition

Use your customized algorithm to train a handwritten digit recognition model and deploy the model for prediction.

Migrating a Third-Party Inference Framework to a Custom Inference Engine

Third-party frameworks

Inference deployment

-

ModelArts allows the deployment of third-party inference frameworks. This section describes how to migrate TF Serving and Triton to a custom inference engine.