facebook/KernelLLM

KernelLLM is an 8 billion parameter large language model developed by Meta, based on Llama 3.1 Instruct, specifically fine-tuned for authoring GPU kernels using Triton. It translates PyTorch modules into efficient Triton kernel implementations, aiming to democratize GPU programming. The model demonstrates competitive performance against much larger models on kernel generation tasks, achieving a score of 20.2 (pass@1) on KernelBench-Triton. Its primary strength lies in automating the generation of high-performance Triton kernels from PyTorch code.