GPU-Puzzles

Why we included this project

Most people who train models never write a line of GPU code, and this interactive notebook is one of the quickest ways to fix that: the puzzles get progressively harder, and you fill in the body of a CUDA kernel, starting with a simple vector map and working up to the shared-memory and reduction algorithms that power most of today's deep learning. Because the code runs through Numba, your Python-like functions get compiled into actual CUDA kernels, which keeps the entry barrier low without hiding what the hardware is doing. Each puzzle checks your kernel against a reference implementation, so you find out immediately whether your indexing is right. The best way to use it is as a hands-on tutorial: open it in Colab, turn on GPU mode, and work through the exercises instead of reading about them. It also makes a solid first exercise for anyone onboarding onto GPU work, and pairs naturally with Sasha Rush's companion Tensor Puzzles for the PyTorch side.