Meta Details TLX Extension for Triton GPU Kernels

Meta detailed TLX in an engineering blog post, describing an extension that adds explicit GPU execution controls to Triton. TLX lets developers coordinate warp groups, memory movement, synchronization and cluster execution while retaining Triton's Python programming model and autotuning tools.

TLX supports hardware features including asynchronous tensor operations, tensor memory and cooperation between multiple thread blocks. It currently targets NVIDIA Hopper and Blackwell GPUs, as well as AMD Mi300 and Mi350x GPUs.

Meta says TLX runs in production for recommendation system training, recommendation system inference and large language model pretraining. Its Generative Ads Model training stack reached 20% to 25% model FLOPs use on several thousand NVIDIA Blackwell GPUs, while a recommendation inference implementation reduced compute intensive latency by as much as two thirds for co-designed models.

The company is also using a TLX grouped matrix multiplication kernel for mixture of experts models across a large GPU fleet. Meta plans to support more GPU architectures and its MTIA accelerators, add automatic warp group orchestration and make TLX a backend for torch.compile.