The rapid proliferation of large language models (LLMs) in natural language processing (NLP) has created a critical need for techniques that enable efficient deployment on memory-constrained devices without compromising performance. We present a method to prune LLMs that selectively prunes model blocks based on an importance score and replaces them with a low-parameter replacement strategy. Specifically, we propose a principled metric to replace each pruned block using a weight-sharing mechanism that leverages unpruned counterparts from the model and block-specific low-rank adapters. Furthermore, we facilitate the learning of these replacement blocks with output feature normalization and an adapter initialization scheme built on low-rank SVD reconstructions. Empirical evaluations demonstrate substantial performance gains over existing methods, achieving state-of-the-art performance on 5/6 benchmarks for a compression rate of 30% and 6/6 benchmarks for a compression rate of 40%. We also demonstrate that our approach can extend smaller models, boosting performance on 6/6 benchmarks using only ~0.3% tokens of extended training with minimal additional parameter costs.

Speculative decoding has emerged as a prominent alternative to autoregressive decoding for expediting inference in large language models (LLMs). However, prevailing assumptions often focus solely on latency reduction, neglecting the computational expenses. In this paper, we present Speculate Less, validate More (SLiM), a speculative decoding enhancement to reduce the speculation set while validating more effective tokens. SLiM is designed to mitigate LLMs’ computation costs associated with the token verification by introducing hypothesis reduction based on a fast posterior estimation. It consistently surpasses counterparts lacking cost reduction across a spectrum from CPU to GPU. Our evaluation with diverse conversational datasets shows that SLiM can achieve a substantial 70% reduction in FLOPs while generating more effective predictions on top of prior arts.

Traditional language models operate autoregressively, i.e., they predict one token at a time. Rapid explosion in model sizes has resulted in high inference times. In this work, we propose DynaMo, a suite of multi-token prediction language models that reduce net inference times. Our models *dynamically* predict multiple tokens based on their confidence in the predicted joint probability distribution. We propose a lightweighttechnique to train these models, leveraging the weights of traditional autoregressive counterparts. Moreover, we propose novel ways to enhance the estimated joint probability to improve text generation quality, namely co-occurrence weighted masking and adaptive thresholding. We also propose systematic qualitative and quantitative methods to rigorously test the quality of generated text for non-autoregressive generation. One of the models in our suite, DynaMo-7.3B-T3, achieves same-quality generated text as the baseline (Pythia-6.9B) while achieving 2.57× speed-up with only 5.87% and 2.67% parameter and training time overheads, respectively.