Back to Gap Map Infrastructure / Core ML frameworks & libraries

Triton

open source / Overall score: 5.0(leading)

Triton is a Python-embedded language and compiler for writing efficient custom GPU kernels, aiming for higher productivity than CUDA with more flexibility than other domain-specific languages. It compiles through MLIR and targets NVIDIA and AMD GPUs, with CPU support in progress. Created by Philippe Tillet at OpenAI, it is now developed under the triton-lang organization and serves as the kernel backend for PyTorch's inductor.

Openness

5 high confidence

5.0

license: MIT(OSI)
source: public
core features withheld: no

The LICENSE file is the standard MIT text, and the full source is public. The README describes the whole library with no paid, enterprise or hosted tier beside it, so the core is ungated.

https://cdn.jsdelivr.net/gh/triton-lang/triton@main/LICENSE recorded 2026-09-29
The LICENSE file carries the MIT permission grant: "Permission is hereby granted, free of charge, to any person obtaining" the software, with "Copyright 2020-2022 OpenAI" beside Philippe Tillet.
https://cdn.jsdelivr.net/gh/triton-lang/triton@main/README.md recorded 2026-09-29
README of the development repository: "The aim of Triton is to provide an open-source environment to write fast code at higher productivity than CUDA", installable with "pip install triton" or from source. Searched the body for enterprise, paid, commercial, license key and hosted: none.
https://github.com/triton-lang/triton/blob/main/LICENSE recorded 2026-08-13
LICENSE file is the verbatim MIT License text, Copyright 2018-2020 Philippe Tillet / 2020-2022 OpenAI
https://pypi.org/pypi/triton/json recorded 2026-09-29
PyPI package metadata for triton declares license "MIT" and summary "A language and compiler for custom Deep Learning operations".

Adoption

5 high confidence

5.0

Adoption is measured by PyPI downloads of the triton package, most of which arrive as a PyTorch dependency rather than a direct install. The wheel is still Triton's own distribution channel, so no adjustment is made for that indirect route.

https://pypistats.org/api/packages/triton/recent recorded 2026-08-13
last_month downloads = 69,538,627 for triton

Capability

5 high confidence

5.0

Triton is the kernel-authoring layer that PyTorch's inductor compiler uses to generate code for NVIDIA and AMD GPUs. It targets both vendors' hardware, while CPU support is still under development.

https://cdn.jsdelivr.net/gh/triton-lang/triton@main/README.md recorded 2026-09-29
README compatibility section lists "NVIDIA GPUs (Compute Capability 8.0+)" and "AMD GPUs (ROCm 6.2+)", with CPUs "Under development"; it describes "a language and compiler for writing highly efficient custom Deep-Learning primitives".
https://github.com/triton-lang/triton recorded 2026-08-13
README still describes the GPU kernel language and MLIR-based compiler, with public nightly wheels
https://pytorch.org/blog/pytorch-2-0-release/ recorded 2026-09-29
PyTorch 2.0 release post: "For NVIDIA and AMD GPUs, it uses OpenAI Triton as a key building block."
https://triton-lang.org/main/index.html recorded 2026-09-29
Documentation home: Triton "is a language and compiler for parallel programming" providing "a Python-based programming environment for productively writing custom DNN compute kernels".

Verified 2026-09-24