triton-lang/triton

Health v2 · maintenance only

Adoption not part of the score

20017 stars · 3141 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Triton is an open-source language and compiler for writing highly efficient custom deep-learning GPU kernels with Python-level productivity. It compiles tiled neural network computations via LLVM/MLIR, offering more flexibility than typical DSLs and higher productivity than raw CUDA.

Use cases

• write custom GPU kernels for deep learning
• implement fused attention or matmul kernels without CUDA
• optimize neural network primitives for NVIDIA GPUs
• prototype high-performance ML kernels in Python
• compile tiled tensor computations to GPU code

When to choose

• you need custom GPU kernels beyond what PyTorch ops provide
• you want CUDA-level performance with Python-level productivity
• you are building a compiler or runtime that generates GPU code
• you need portable kernel code across GPU architectures

When to avoid

• you only need standard neural network layers already covered by PyTorch or JAX
• you target CPUs exclusively without GPU hardware
• you need fine-grained low-level control over warp scheduling that CUDA exposes
• you want a stable, frozen API for production embedding without tracking upstream changes

Facets

library · maturity active

compiler programming-language gpu-computing machine-learning deep-learning compilers gpu-computing machine-learning deep-learning programming-languages python cpp dsl cuda-alternative kernel-authoring mlir llvm compiler-infrastructure linux gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="triton-lang/triton")