Paper 2606.13392
MiniMax Sparse Attention
01 In brief
Summary
MiniMax Sparse Attention (MSA) is a blockwise sparse attention mechanism built on Grouped Query Attention (GQA) to address the quadratic cost of softmax attention in ultra-long contexts.
A lightweight Index Branch scores key-value blocks and selects a Top-k subset per GQA group, while the Main Branch computes exact block-sparse attention over selected blocks.
MSA is co-designed with GPU kernels, including exp-free Top-k selection and KV-outer sparse attention, to achieve practical speedups.
On a 109B-parameter multimodal model, MSA matches GQA performance while reducing per-token attention compute by 28.4x at 1M context, with 14.2x prefill and 7.6x decoding speedups on H800.
Training uses a KL alignment loss with gradient detach, indexer warmup, and a forced local block.
Experiments show MSA-PT (from scratch) and MSA-CPT (continued pretraining) remain competitive with full attention across benchmarks, with long-context extension results close to baseline.
The inference kernel is open-sourced, and a production model is publicly released.
02 From the paper
Abstract
Ultra-long-context capability is becoming indispensable for frontier LLMs: agentic workflows, repository-scale code reasoning, and persistent memory all require the model to jointly attend over hundreds of thousands to millions of tokens, yet the quadratic cost of softmax attention makes this untenable at deployment scale. We introduce MiniMax Sparse Attention (MSA), a blockwise sparse attention built upon Grouped Query Attention (GQA). A lightweight Index Branch scores key-value blocks and independently selects a Top-k subset for each GQA group, enabling group-specific sparse retrieval while maintaining efficient block-level execution; the Main Branch then performs exact block-sparse attention over only the selected blocks. Designed around a principle of simplicity and scalability, MSA is deliberately streamlined, making it straightforward to deploy efficiently across a broad range of GPUs. To translate sparsity into practical speedups, we co-design MSA with a GPU execution path that uses exp-free Top-k selection and KV-outer sparse attention to improve tensor-core utilization under block-granular access. On a 109B-parameter model with native multimodal training, MSA performs on par with GQA while reducing per-token attention compute by 28.4x at 1M context. Paired with our co-designed kernel, MSA achieves 14.2x prefill and 7.6x decoding wall-clock speedups on H800. Our inference kernel is available at: https://github.com/MiniMax-AI/MSA. A production-grade natively multimodal model powered by MSA has been publicly released at: https://huggingface.co/MiniMaxAI/MiniMax-M3.