MoonMath.ai builds the performance layer for Physical AI.
We are a small team of mathematicians & engineers building production-grade acceleration for the next wave of AI systems via low level algorithms and system engineering.
MoonLite is an inference acceleration toolkit designed for large generative models.
- LiteAttention: Transforming Video Diffusion with Temporal Sparse Attention
- LiteLinear: Replaces standard FFN layers with a decomposed module.
- BackLite: Wraps Flash Attention 3 and uses attention sparsity to speed up the backward pass via gradient approximation.
- LiteRunner: Runner for generative models with local and W&B tracking.
Website: https://moonmath.ai
Blog: https://moonmath.ai/posts
GitHub Organization: https://github.com/moonmath-ai
LinkedIn: https://www.linkedin.com/company/moonmath-ai
YouTube: https://www.youtube.com/@MoonMath_ai