triton-lang/triton is the development repository for Triton, a language and compiler for writing efficient custom deep learning primitive operations. As the repository description puts it, the Triton language and compiler are developed together in one place. It began on August 30, 2014, and the latest public release tag is v3.8.0.
The README states that Triton aims to provide an open-source environment that is more productive than writing CUDA code directly and more flexible than existing DSLs (domain-specific languages). In other words, it assumes that the user writes the needed operation directly in code, rather than calling a fixed set of high-level operations. The implementation language is MLIR and the license is MIT. The official documentation covers installation and tutorials, and the README also explains how to build from source and run the tests. The star count is 20424.
However, no performance figures compared with CUDA are given. Nor can what changed in v3.8.0 be confirmed from public information. The 20424 stars are only a signal of recognition and do not substitute for performance or actual adoption. Of the evaluation items, only momentum is scored, at 4.2, while the freshness, technology and quality scores are empty. In other words, the claim that it can produce faster code stops at the README's statement of intent. Teams that already have concrete custom operations to write may find it worth checking, but without performance evidence it is not yet a default choice.