MindSpeed
This introduction helps developers understand MindSpeed and use it for model development and tuning.
- MindSpeed Overview
Get a quick overview of MindSpeed's core features, capabilities, and key components.
- Release Notes
Learn about the version information for each component of MindSpeed.
- Quick Start
Get started with LLM pre-training and fine-tuning using MindSpeed. For details, see Quick Start for LLM Training, Quick Start for Multimodal Model Training, and Quick Start for Distributed LLM Pre-training.
Large Language Model (LLM) Suite
Powered by the Ascend ecosystem, MindSpeed LLM is a distributed training framework designed to provide Ascend ecosystem partners with end-to-end LLM training solutions. It includes distributed pre-training, distributed instruction fine-tuning, and a dedicated development toolchain.
- Software Installation
Learn how to install MindSpeed LLM.
- Distributed Pre-training
This section uses the Qwen3-8B model as an example to demonstrate multi-sample pre-training methods.
- Distributed Fine-tuning
Learn how to perform instruction fine-tuning using single-sample data formats.
- Performance Profiling
Learn how to profile and analyze performance data on Ascend hardware.
Multimodal Model (MM) Suite
MindSpeed MM is an Ascend-based multimodal suite built for large-scale distributed training. Supporting both multimodal generation and understanding, it provides end-to-end multimodal training solutions for Huawei Ascend processors. It features mainstream foundation models, data engineering capabilities, distributed training acceleration, and supports pre-training, fine-tuning, online inference, and reinforcement learning.
Large Model Training Acceleration Library
MindSpeed Core is a dedicated large model acceleration library for Ascend devices. It enables seamless migration of large model workloads to the Ascend platform and supports Ascend-proprietary algorithms, ensuring an out-of-the-box experience.
- Software Installation
Learn how to install MindSpeed Core.
- Feature List
Review the release status of MindSpeed Core features, PyTorch framework support, and performance optimization case studies for sequence parallelism.
- Custom Operators
Learn about the release status of custom operators and their compatibility with PyTorch.
Reinforcement Learning (RL) Suite
MindSpeed RL is an Ascend-based reinforcement learning acceleration framework. It delivers end-to-end RL training and inference solutions for Ascend ecosystem partners, featuring core acceleration capabilities such as co-located or disaggregated training/inference on ultra-large Ascend clusters, asynchronous multi-model pipeline scheduling, and heterogeneous sharded communication.