Triton OpenAI Triton
More info
- Creator
- Philippe Tillet (Harvard / OpenAI)
- Released
- Type
- Python-like language for GPU kernels
- License
- MIT
- Source
- github.com/triton-lang/triton
Python DSL for writing CUDA kernels — Tillet at OpenAI. Powers FlashAttention, PyTorch 2's torch.compile, and most modern LLM training.
Pitch: write kernels at Python level (without managing shared memory, banks, warps by hand), the Triton compiler generates competitive PTX. Tri Dao (FlashAttention) wrote both Cutlass and Triton versions — Triton won on readability. AMD ROCm support landed in 2024. Not to be confused with Nvidia Triton Inference Server (totally different thing).