bendygrad

A Bend 2 port of tinygrad's front-end concepts: tensor shapes and views, buffers, a lazy operation graph, reverse-mode automatic differentiation, linear layers, and stochastic gradient descent. A small neural-network demo learns XOR.

The implementation adapts tensor operations and gradient accumulation to Bend's functional and affine programming model. LAWS.bend, PROOF.bend, and focused tests accompany the tensor, shape, and differentiation code. This is an exploratory front-half port, not a drop-in replacement for the complete tinygrad stack.