MindSpeed

This introduction helps developers understand MindSpeed and use it for model development and tuning.

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.

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.

References

  • Security Statement

    Best practices and recommended security hardening measures for MindSpeed components.