SLoPe: Double-Pruned Sparse Plus Lazy Low-Rank Adapter Pretraining of LLMs

TL;DR

SLoPe introduces a novel method combining double-pruned sparse structures and lazy low-rank adapters to enhance accuracy, speed, and memory efficiency in large language model pretraining and inference.

Contribution

It presents a new pretraining approach that adds low-rank adapters late in training and employs double-pruned backward passes for efficiency, improving sparse LLM performance.

Findings

Accelerates training by up to 1.25x and inference by 1.54x on large models.

Reduces memory usage during training and inference by up to 0.63x and 0.61x.

Improves accuracy of sparse LLMs without significant overheads.

Abstract

We propose SLoPe, a Double-Pruned Sparse Plus Lazy Low-rank Adapter Pretraining method for LLMs that improves the accuracy of sparse LLMs while accelerating their pretraining and inference and reducing their memory footprint. Sparse pretraining of LLMs reduces the accuracy of the model, to overcome this, prior work uses dense models during fine-tuning. SLoPe improves the accuracy of sparsely pretrained models by adding low-rank adapters in the final 1% iterations of pretraining without adding significant overheads to the model pretraining and inference. In addition, SLoPe uses a double-pruned backward pass formulation that prunes the transposed weight matrix using N:M sparsity structures to enable an accelerated sparse backward pass. SLoPe accelerates the training and inference of models with billions of parameters up to $1.25\times$ and $1.54\times$ respectively (OPT-33B and OPT-66B) while reducing their memory usage by up to $0.63\times$ and $0.61\times$ for training and inference respectively.