torch.compile is cool but LLM compile: takes your .py repo as string and outputs a brand new, custom, from scratch, minimal code repository directly running your network in highly optimized CUDA

→ View original post on X — @karpathy,