AI agent ports Karpathy's CUDA GPT-2 trainer to SYCL, verified by hidden checks
A developer built Muse Port, a tool in which Meta's Muse model plans and ports CUDA repositories to SYCL, an open standard that runs on NVIDIA, AMD, Intel and CPU hardware. Porting is accepted only after hidden checks, unseen by the model, confirm the output matches the original. The developer applied it to Andrej Karpathy's llm.c, a GPT-2 trainer written in raw CUDA, and the resulting port passed the project's own test on CPU, AMD and NVIDIA hardware. The initial port ran 36 times slower than CUDA, but after tuning on real GPUs it reached roughly 1.1 to 1.5 times CUDA's step time on NVIDIA hardware, and it also ran on AMD Instinct GPUs. The work was submitted to Meta's Global AI Developer Hackathon, with code and a demo published online.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.


Discussion (0)
Log in to join the discussion and vote.
Log in