Triton is a language and compiler designed for writing highly efficient custom deep-learning primitives with greater productivity than CUDA and more flexibility than existing domain-specific languages. The project is developed as an open-source environment that aims to bridge the gap between low-level GPU programming and high-level abstraction, with foundations described in the MAPL2019 publication "Triton: An Intermediate Language and Compiler for Tiled Neural Network Computations."
The repository is written primarily in MLIR and serves as the development hub for both the Triton language and its compiler infrastructure. The project maintains comprehensive documentation at triton-lang.org and provides multiple installation options, including binary wheels for CPython versions 3.10 through 3.14 available via pip, as well as source-based installation with support for custom LLVM builds. The build system is flexible, allowing developers to configure compilation with environment variables such as TRITON_BUILD_WITH_CLANG_LLD for faster builds using clang and lld, TRITON_BUILD_WITH_CCACHE for ccache integration, and MAX_JOBS to manage memory constraints during compilation.
The development environment provides extensive debugging and optimization capabilities through configuration knobs documented in python/triton/knobs.py. Developers can enable MLIR IR dumps with MLIR_ENABLE_DUMP, inspect LLVM IR transformations with LLVM_IR_ENABLE_DUMP, generate MLIR reproducers for failing compilation stages, and use the Triton interpreter via TRITON_INTERPRET for GPU-free testing with Python breakpoint support. Additional tools include LLVM address sanitizer integration for memory safety on AMD backends and fine-grained debug output control through TRITON_LLVM_DEBUG_ONLY.
The project maintains active community engagement through the Triton Developer Conference, with the third iteration held in October 2025 at Microsoft's Silicon Valley Campus. Conference materials including recordings and slides are publicly available, alongside documentation from previous 2024 and 2023 conferences. The repository also references third-party educational resources such as Triton Puzzles, which can be executed using the Triton interpreter without requiring GPU hardware, making the language accessible for learning and experimentation.
