model.jl implements the core Flux API for forward and backward passes.

The compiler folder implements the @net macro and some dataflow manipulation (like loop unrolling). The mxnet and tensorflow folders in backend each describe how to run @net code on those backends. These are the most involved parts of Flux where the magic-y stuff happens, but everything else is pretty straightforward Julia code.

layers is Flux's "standard library" of model parts; layers, activations, cost functions. Most are implemented using @net. control.jl implements Chain and others in pure Julia. shims.jl includes some things like convolutions which don't have "real" implementations yet, but can compile to a backend regardless.

dims implements the mechanisms for typing batch (and other) dimensions. This is basically standalone.

data.jl and utils.jl implement misc utilities.