Dao-AILab/quack
A Quirky Assortment of CuTe Kernels
What it solves
QuACK provides a collection of high-performance GPU kernels for common deep learning operations, allowing developers to achieve "speed-of-light" performance for memory-bound operations directly from Python using the CuTe-DSL.
How it works
The project utilizes the CuTe-DSL to implement a variety of optimized kernels. It supports modern NVIDIA GPUs (H100, B200/B300, and RTX 50 series) and provides interfaces for both standard Python and JAX.
Who it’s for
It is designed for developers and researchers working with high-end NVIDIA GPUs who need highly optimized implementations of fundamental AI operations like normalization and matrix multiplication.
Highlights
- Optimized kernels for RMSNorm, Softmax, Cross Entropy, and LayerNorm (both forward and backward passes).
- Specialized GEMM (General Matrix Multiply) kernels for Hopper and Blackwell architectures.
- Optional JAX bindings for seamless integration with the JAX ecosystem.
- Support for the latest CUDA toolkits (12.9+) and cutting-edge GPU hardware.
Related
- Project
- Project
- Project
- Project
- Project