This post shows where the underling methods for a linear layer are defined. Convolutions dispatch to different implementations from here (i.e. the CPU kernel, native CUDA kernel, cuDNN, etc.).