Attention Is All You Need But You Don't Need All Of It For Inference of Large Language Models

The inference demand for LLMs has skyrocketed in recent months, and serving models with low latencies remains challenging due to the quadratic input length complexity of the attention layers. In this work, we investigate the effect of dropping MLP and attention layers at inference time on the performance of Llama-v2 models. We find that dropping dreeper attention layers only marginally decreases performance but leads to the best speedups alongside dropping entire layers. For example, removing 33\% of attention layers in a 13B Llama2 model results in a 1.8\% drop in average performance over the OpenLLM benchmark. We also observe that skipping layers except the latter layers reduces performances for more layers skipped, except for skipping the attention layers.

Long-run Behaviour of Multi-fidelity Bayesian Optimisation

Multi-fidelity Bayesian Optimisation (MFBO) has been shown to generally converge faster than single-fidelity Bayesian Optimisation (SFBO) (Poloczek et al. (2017)). Inspired by recent benchmark papers, we are investigating the long-run behaviour of MFBO, based on observations in the literature that it might under-perform in certain scenarios (Mikkola et al. (2023), Eggensperger et al. (2021)). An under-performance of MBFO in the long-run could significantly undermine its application to many research tasks, especially when we are not able to identify when the under-performance begins. We create a simple benchmark study, showcase empirical results and discuss scenarios and possible reasons of under-performance.