Philippe Tillet, H. T. Kung, and David Cox presented Triton: An Intermediate Language and Compiler for Tiled Neural Network Computations at MAPL '19, introducing a tile-based IR that automates shared-memory management and thread scheduling for dense and fused neural kernels.
Triton
Also known as: OpenAI Triton · Triton Language
Plain English
A Python-like GPU programming language and compiler created by OpenAI that lets ML engineers write fast custom GPU kernels without dropping down to complex C++ CUDA code.
Technical Overview
Triton replaces CUDA's thread-level SIMT programming model with block/tile-level abstractions, automating shared-memory coalescing, SRAM tiling, and tensor-core scheduling in MLIR/LLVM. It serves as the default GPU codegen backend for PyTorch torch.compile (TorchInductor) and powers custom attention, FP8 GEMM, and PagedAttention kernels across vLLM and SGLang.
Timeline (3)
2019
- Paper
2021
2023
- Ecosystem
PyTorch 2.0 adopted Triton as the primary GPU backend for TorchInductor and
torch.compile, making Triton-generated fused kernels the default acceleration path across training and inference workloads.
See also (bkz.) (10)
Open full graphTriton