tinygrad

tinygrad, maintained by the tiny corp, is a deep learning framework that sits between PyTorch and Karpathy's micrograd, kept intentionally small and hackable. It draws ergonomics from PyTorch, functional transforms from JAX, and scheduling ideas from TVM, using lazy evaluation to fuse operations into single kernels. It supports several GPU backends and can run real models while staying easy to read. Released under the MIT license.