MindSpeed
Product Stack
Service System
MindSpeed for
Training Acceleration
Foundation Model Suite
MindSpeed LLM
MindSpeed MM
MindSpeed RL
Acceleration Module
MindSpeed Core
Atlas Hardware
MindSpeed LLM
Ascend hardware-oriented LLM suite offering usable open‑source models that support pre-training, full-parameter fine-tuning, parameter‑efficient fine‑tuning along with comprehensive toolchains for data preprocessing, weight conversion, and diverse evaluation methods, enhancing the development and deployment of LLMs on Ascend hardware.
Development Scenarios
LLM
Ready-to-use mainstream LLMs, supporting data preprocessing, training, fine-tuning, and preference alignment.
Multimodal Model
Preconfigured multimodal understanding and generation models, offering diverse multimodal data processing functions and large-scale distributed training capabilities and enabling flexible model customization.
Reinforcement Learning (RL) Model
E2E training and inference solution focused on RL training scenarios, enabling flexible cluster deployment, out-of-the-box usability, and improved distributed training and inference efficiency through optimization technologies.
Key Features
Distributed Sequence Training
MindSpeed multi-dimensional optimization for training of million-token sequences
Parallel Optimization
Three context parallel algorithms, MoE parallelism
Computing Optimization
FlashAttention
Memory Optimization
ALiBi, Reset Attention Mask compression
Communication Optimization
P2P streaming and concurrency
