tile-ai/tilelang-puzzles

Learning TileLang with 10 puzzles!

What it solves

Writing high-performance deep learning kernels is hard. TileLang is a domain-specific language that helps, but learning it from scratch is steep. These puzzles give you a guided path: you start with trivial examples and gradually build up to modern kernels like GEMM (matrix multiplication) and FlashAttention, so you understand the design principles of TileLang by doing.

How it works

You install TileLang and its dependencies, then run a script to check your setup is correct. The project ships 10 puzzles, each a standalone Python script in the puzzles/ directory. Each puzzle asks you to implement a kernel; reference solutions are provided in the ans/ directory so you can compare your work. You just run the puzzle script and the answer script side by side.

Who it’s for

Programmers who know the basics of deep learning and want to write fast GPU kernels, particularly those wanting to learn TileLang specifically. It’s also useful for anyone who likes learning by solving incremental challenges.

Highlights

  • 10 progressive puzzles, from simple copy operations to GEMM and FlashAttention
  • Reference implementations included for every puzzle
  • Inspired by Triton Puzzles, Triton Puzzles Lite, and LeetGPU
  • Includes an environment-check script to verify your TileLang and GPU setup

Related

  • Project
    bentoml/BentoDiffusion

    A collection of example projects for self-hosting and deploying various Stable Diffusion family models using the BentoML framework.

  • Project
    TensorOpsAI/LLMstudio

    A prompt engineering platform that provides a unified proxy for multiple LLM providers and local models, featuring a playground UI and monitoring tools.

  • Project
    LeiWang1999/ZYNQ-NVDLA

    ZYNQ‑NVDLA is an open‑source project that ports NVIDIA’s Deep Learning Accelerator (NVDLA) to Xilinx Zynq FPGA boards. It provides RTL, Vivado IP, Linux drivers, pre‑built ARM64 libraries, and a Tengine‑Lite backend so you can run quantised neural‑network models (e.g., ResNet‑18, YOLOX‑nano) on the FPGA‑based accelerator with low latency.

  • Project
    pingmike2/freebuff2api-wokers

    A single-file API proxy that turns freebuff's free models into an OpenAI-compatible API, with a web panel for managing multiple accounts, quotas, and session rotation.