MindSpeed

MindSpeed is a high-performance acceleration library engineered specifically for the Ascend platform. It features four core components: MindSpeed Core (large model training acceleration library), MindSpeed LLM, MindSpeed MM, and MindSpeed RL suites. Designed to streamline your workflow, MindSpeed empowers developers to easily master the suite and build, optimize, and fine-tune models with maximum efficiency.

MindSpeed Key Features

  • MindSpeed LLM: Delivers major capability enhancements. The Megatron training backend now supports Seed-OSS and GLM5 model training, adds asynchronous weight saving, and extends compatibility to the Ascend 950 series products.
  • MindSpeed MM: Enables asynchronous activation offloading for Qwen3.5 and Kimi K2.5, and supports chunk loss and chunk micro-batch size (MBS).
  • MindSpeed Core: Adds multi-precision training, quantization, adaptation for new models and features, and operator-level optimizations.

MindSpeed Documentation

Table 1 Release and installation

Release and Installation

MindSpeed Overview

Release Notes

Quick Installation

Ascend Product Overview

Table 4 Large model training acceleration library

MindSpeed Core

Introduction

Software Installation

Quick Start

Features

Feature Development

Model Migration

Performance Tuning

Table 5 Reinforcement learning (RL) suite

MindSpeed RL

MindSpeed RL

Table 6 References

References

Security Statement

Glossary