Meta (Engineering at Meta) · GPU Kernel Extension

TLX (Triton Low-Level Extensions)

Meta describes TLX, an extension to the Triton GPU programming language that gives engineers explicit control over warp-group and cluster-level execution while keeping Triton's blocked programming model. It targets Hopper and Blackwell features such as warp specialisation, TMA and TMEM, and is used on NVIDIA and AMD GPUs. The post says TLX has improved performance in RecSys training, RecSys inference and large-scale LLM pretraining, but the excerpt gives no benchmark figures. The page text does not state the licence.