Meta open-sources TLX, a Triton extension for low-level GPU control

Meta has published TLX (Triton Low-Level Extensions), an open-source extension to the Triton kernel language that gives engineers explicit control over how work is scheduled across warp groups and thread-block clusters on modern GPUs. It targets capabilities such as warp specialization, asynchronous memory movement, tensor memory and cluster-level execution, which Triton's compiler-managed model previously left implicit. The design raises the warp group to the unit of orchestration, an approach Meta calls Multiple Instructions, Multiple Warps. Meta also reports that Triton kernel count in its stack grew about 3.8x over a 10-month period, versus roughly 2.0x for CUDA.