Making small PyTorch ops faster on a laptop GPU
A power daemon that capped my GPU memory at 810 MHz, prefetching the next layer into L2, and a decode attention kernel that runs in one launch. Interactive charts, including the cases where none of it helped.