Toggle light / dark theme

Deepseek-ai/TileKernels: A kernel library written in tilelang

A kernel library written in tilelang. Contribute to deepseek-ai/TileKernels development by creating an account on GitHub.

3 Comments so far

  1. Interesting to see tilelang being used for kernel libraries like this. I’ve been exploring similar approaches for GPU optimization, and having a dedicated library like TileKernels could save a lot of boilerplate work. How does the performance compare to handwritten CUDA kernels for common operations?

  2. Interesting to see TileLang getting more traction in kernel development. I’ve been experimenting with it for some GEMM optimizations on consumer GPUs, and the tile-level abstraction does make certain patterns much cleaner than raw CUDA. Curious how the performance compares to hand-tuned cuBLAS kernels for larger matrix sizes, though.

Comments are closed.

/* */