triton-lang/triton

📈 Trending★ 19,210MLIRAudience · researcherComplexity · 4/5ActiveLicenseSetup · hard

Mindmap

mindmap
  root((Triton))
    What it does
      GPU kernel language
      Higher-level than CUDA
      Compiles to machine code
    Tech stack
      MLIR compiler
      LLVM backend
      Python syntax
    Use cases
      Custom model layers
      Performance optimization
      GPU computation
    Audience
      ML researchers
      Deep learning engineers
      Performance specialists
    Integration
      PyTorch torch.compile
      AI ecosystem

mindmap root((Triton)) What it does GPU kernel language Higher-level than CUDA Compiles to machine code Tech stack MLIR compiler LLVM backend Python syntax Use cases Custom model layers Performance optimization GPU computation Audience ML researchers Deep learning engineers Performance specialists Integration PyTorch torch.compile AI ecosystem

Things people build with this

USE CASE 1

Write custom GPU kernels for bottleneck layers in neural networks without learning CUDA.

USE CASE 2

Optimize matrix operations and attention mechanisms for faster model inference.

USE CASE 3

Implement specialized mathematical operations that existing libraries don't provide.

Tech stack

PythonMLIRLLVMCUDAGPU

Getting it running

Difficulty · hard Time to first run · 1h+

Requires CUDA toolkit, LLVM/MLIR build infrastructure, and GPU hardware to test compiled operations.

Use freely for any purpose including commercial, as long as you keep the copyright notice.

In plain English

Triton is a programming language and compiler for writing highly efficient custom operations for deep learning, particularly the kind that run on GPUs. When training or running AI models, much of the heavy computation happens in custom mathematical kernels (small, highly optimized programs that run on GPU hardware). Writing these in CUDA (NVIDIA's low-level GPU programming language) requires deep hardware expertise. Triton aims to offer a higher-level, more productive alternative while still producing fast code, described as offering higher productivity than CUDA but higher flexibility than other specialized languages. Triton uses MLIR (a compiler infrastructure framework) and LLVM internally to transform Python-like kernel code into GPU machine code. It is tightly integrated with the AI/ML ecosystem and is a key component powering PyTorch's compiled execution path (torch.compile). You would use Triton if you are a machine learning researcher or engineer who needs to write custom GPU kernels for performance-critical model components, but wants to work at a higher level of abstraction than raw CUDA. It installs via pip for CPython 3.10 through 3.14.

Copy-paste prompts

Prompt 1

Show me a Triton kernel that implements a fused matrix multiply and activation function for a transformer layer.

Prompt 2

How do I install Triton and write my first GPU kernel to speed up a custom PyTorch operation?

Prompt 3

Convert a CUDA kernel I wrote into Triton code and explain what's simpler about the Triton version.

Prompt 4

Create a Triton kernel that performs element-wise operations on GPU tensors and integrates with torch.compile.

Open on GitHub → Explain another repo

Generated 2026-05-18 · Model: sonnet-4-6 · Verify against the repo before relying on details.