InfiniTensor/ninetoothed

A domain-specific language (DSL) based on Triton but providing higher-level abstractions.

What it solves

NineToothed simplifies the creation of high-performance GPU kernels. It removes the complexity typically associated with writing low-level GPU code by providing a domain-specific language (DSL) that makes the process more accessible.

How it works

It uses a paradigm called tensor-oriented meta-programming (TOM) and an "arrange-and-apply" approach. Developers define three distinct components: the arrangement (how data is tiled and expanded), the application (the actual computation logic), and the tensors themselves. These are then integrated using ninetoothed.make to generate a Triton-based kernel.

Who it’s for

Developers and researchers who need to write custom, high-performance GPU kernels but want a higher-level abstraction than raw Triton or CUDA code.

Highlights

  • Triton-based DSL: Leverages Triton for GPU performance.
  • Arrange-and-Apply Paradigm: Separates data layout and computation logic for better clarity.
  • Tensor-Oriented Meta-programming: Simplifies the mapping of high-level tensor operations to GPU hardware.

Related

  • Project
    triton-lang/triton

    Triton is an open‑source language and compiler for writing custom GPU kernels for deep‑learning primitives. It provides a Python‑like DSL (`@triton.jit`) that compiles via an MLIR‑based backend to efficient NVIDIA and AMD GPU code, with auto‑tuning, an interpreter for CPU debugging, and extensible compiler passes.

  • Project
    pytorch/helion

    Helion is a Python‑embedded DSL that lets you write GPU kernels like regular PyTorch code. It compiles the inner loops to Triton, automatically chooses tiling, grid sizes, and memory‑access strategies, and runs an extensive autotuner to find the fastest configuration. You can freeze the best config for production, and experimental back‑ends (TileIR, CuTe) broaden hardware support.

  • Project
    NVIDIA/tilus

    Tilus is a tile-level GPU kernel programming language that enables high-performance, low-precision computation with explicit control over shared memory and registers.

  • Project
    tile-ai/tilelang

    A concise domain-specific language and compiler built on TVM for developing high-performance GPU/CPU/NPU kernels like GEMM and FlashAttention.

  • Project
    ByteDance-Seed/Triton-distributed

    Triton‑distributed is a ByteDance‑Seed open‑source compiler that extends the Triton language with primitives for overlapping GPU compute and communication. It enables high‑performance distributed kernels (GEMM, MoE, flash‑decode, AllToAll) on NVIDIA and AMD GPUs, offering speed‑ups comparable to hand‑tuned libraries. Install via source or a pre‑built pip wheel, write kernels with `@triton_dist.jit`, and use the provided low‑level communication APIs. The project is actively updated (latest news Sep 2026) and released under MIT.