Triton Puzzles
Write Triton GPU kernels in a notebook, from vector add to FlashAttention
Triton Puzzles is a notebook of 12 exercises for learning Triton, the Python-based language for writing GPU kernels. Each puzzle states a small math task and gives you a kernel stub to fill in using block loads, stores and masks. The puzzles start with adding a constant to a vector and build up to softmax, a simple FlashAttention, 2D convolution, matrix multiply and quantized matrix multiply.
Free · No login · Browser · Advanced
About Triton Puzzles
It trains how GPU code splits work into blocks and how memory is loaded and stored, which is the hard part of fast kernels. It suits people who know NumPy or PyTorch. A visualizer shows which memory cells each block reads and writes.
It is free and open source. It runs in Google Colab or a local Jupyter setup, and it does not need a GPU because it uses a Triton interpreter. Colab needs a Google account.
Made this game? Claim it →
Tips and discussion
No tips yet.
Sign in to share a tip.