Automatic Differentiation With Vjp And Jvp

via “automatic-differentiation-with-vjp-and-jvp”

Apple's ML framework for Apple Silicon — NumPy-like API, unified memory, LLM support.

Unique: Implements both VJP and JVP as composable transforms that build gradient computation graphs mirroring the forward graph. Unlike frameworks that hard-code backprop rules per operation, MLX uses a transform system where each primitive defines its VJP/JVP, enabling extensibility. Gradients are first-class transforms, not special-cased.

vs others: More flexible than PyTorch's fixed backprop because VJP/JVP are composable transforms; more efficient than TensorFlow's tape-based autodiff for complex control flow because it builds explicit gradient graphs.