sparsegpt

Definition

A one-shot pruning method for very large language models: it removes a large fraction of the weights and updates the remaining ones to compensate, layer by layer, with no retraining and no gradient steps. The contribution is that this stays tractable at hundred-billion-parameter scale, where retraining-based pruning is simply out of reach. It shares its solver lineage with GPTQ and composes with quantization.