jax-ml/jax-triton

jax-triton contains integrations between JAX and OpenAI Triton

What it solves

It provides a way to integrate Triton kernels into JAX programs, allowing developers to write high-performance custom GPU kernels using Triton's Python-based language and execute them seamlessly within JAX's functional ecosystem, including inside jax.jit-compiled functions.

How it works

The library introduces the triton_call function, which acts as a bridge. It allows JAX arrays to be passed as arguments to Triton kernels. It also supports in-place mutations of JAX Ref objects (created via jax.new_ref), enabling kernels to modify data without allocating new output arrays.

Who it’s for

Developers and researchers who need the performance of custom Triton kernels but want to maintain the use of JAX for their overall model architecture and high-level orchestration.

Highlights

  • Support for calling Triton kernels from JAX arrays.
  • Compatibility with jax.jit for efficient execution.
  • Support for in-out parameters via JAX Ref for in-place mutation.
  • Integration with the Gluon dialect.

Related

  • Dispatch
    OpenAI Triton 1.0 Release

    OpenAI has released Triton 1.0, an open-source Python-like language and compiler that allows researchers to write highly efficient GPU kernels without deep CUDA expertise.

  • Project
    triton-lang/triton

    A language and compiler for writing highly efficient custom deep-learning primitives with higher productivity than CUDA and greater flexibility than other DSLs.

  • 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.

  • Project
    triton-inference-server/client

    Client libraries and examples for C++, Python, and Java that simplify communication with Triton Inference Server for model inferencing and server management.

  • Project
    BobMcDear/attorch

    A Triton-based subset of PyTorch's nn module that provides a readable and hackable collection of neural network layers for developing custom deep learning operations.