mit-han-lab/radial-attention
A sparse attention mechanism with O(nlogn) complexity that accelerates diffusion-based video generation models while preserving output quality.
Not currently ranked — collecting fresh signals.
star history
Radial Attention implements a sparse attention pattern designed to reduce the computational overhead of transformer attention in video generation models. It achieves near-linear complexity scaling, enabling single-GPU video generation in 33-90 seconds on H100/4090 GPUs. The project integrates with multiple video generation frameworks including Wan2.1, HunyuanVideo, and Mochi-1, and is compatible with techniques like SageAttention and LoRA for further optimization.
Frequently asked
- What is mit-han-lab/radial-attention?
- A sparse attention mechanism with O(nlogn) complexity that accelerates diffusion-based video generation models while preserving output quality.
- Is radial-attention open source?
- Yes — mit-han-lab/radial-attention is open source, released under the Apache-2.0 license.
- What language is radial-attention written in?
- mit-han-lab/radial-attention is primarily written in Python.
- How popular is radial-attention?
- mit-han-lab/radial-attention has 604 stars on GitHub.
- Where can I find radial-attention?
- mit-han-lab/radial-attention is on GitHub at https://github.com/mit-han-lab/radial-attention.