Triton-Puzzles

Why we included this project

Triton sits between high-level PyTorch-like code and raw CUDA, and it powers a lot of modern deep learning. Most tutorials expect you to already understand CUDA's memory model, but this puzzle set starts from the basics and uses Triton's interpreter, so you can work through everything in a Colab notebook without a GPU. You begin with simple tensor ops and work up to Flash Attention and quantized networks, which is a solid way to get a feel for low-level performance thinking. That makes it useful for anyone writing custom kernels or trying to read performance-sensitive code, and also one of the gentler introductions to accelerator programming around.