Fitting training and inference into the memory and bandwidth actually available.
Colour is the family; a dashed line is the second member of it.
Trains a small model against the large model's full output distribution rather than the hard labels. The relative probabilities of the wrong answers… · 2015
Keeps only a subset of activations and recomputes the rest during the backward pass, trading arithmetic for memory at roughly the square root of the… · 2016
Stores and multiplies in sixteen bits while keeping a master copy of the weights and the accumulations in thirty-two, with a loss scale to keep small… · 2017
Routes each token to a few of many expert subnetworks, so parameter count grows without the arithmetic growing with it. The routing is the hard part,… · 2017
Freezes the pretrained weights and learns a low-rank update beside them. Fine-tuning then costs a fraction of the memory, and the adapters can be swa… · 2021
Represents weights, and sometimes activations, as low-bit integers with a scale. Most of the difficulty is the handful of outlier channels whose rang… · 2022
A small draft model proposes several tokens and the large model verifies them in one batched pass, keeping the longest correct prefix. Exact: the out… · 2022
Stores the keys and values already computed for the prefix so each generated token attends without recomputing them. Turns quadratic regeneration int…