Why I built it

To strip away the magic of PyTorch's loss.backward() and understand what an autograd engine really does. minute-grad is small enough to read in an afternoon — under a thousand lines — and demonstrates reverse-mode autodiff on a real tensor type with broadcasting.

What it does

Tensors track their parent operations as a DAG. Each op records a backward function that, given an upstream gradient, computes gradients for its inputs. A topological traversal from the loss tensor walks those backward functions in order, accumulating gradients into leaf nodes.

It supports enough ops to train a small MLP on MNIST end-to-end on CPU.