tinygrad

Built by tiny corp, the company George Hotz started around it, tinygrad tries to be a full deep learning stack that one person can read top to bottom. On the surface sits a PyTorch-style Tensor with autograd, nn layers, optimizers and dataset helpers. Underneath, a small compiler records operations lazily, fuses chains of them into one generated kernel, and can run a BEAM search over kernel variants; raising the DEBUG environment variable prints the code it emits, and TinyJit captures kernel graphs once and replays them. Devices include NVIDIA via CUDA or its own NV runtime, AMD, Apple Metal, OpenCL, WebGPU, Qualcomm and CPU, and a new accelerator only has to implement a short list of primitive ops. It ships in real products: comma.ai's openpilot runs its driving model with it on a Snapdragon 845 GPU, and tiny corp sells tinybox machines built for it. Install from source with pip; the project is pre-1.0, so expect API churn. The MIT license puts no limits on commercial use. A good match for researchers, compiler hackers and hardware vendors who want to see how tensor code becomes GPU programs.