Large language models (LLMs) exhibit remarkable reasoning abilities, allowing them to generalize across a wide range of downstream tasks, such as commonsense reasoning or instruction following. However, as LLMs scale, inference costs become increasingly prohibitive, accumulating significantly over their life cycle. This poses the question: Can we compress pre-trained LLMs to meet diverse size and latency requirements? We leverage Neural Architecture Search (NAS) to compress LLMs by pruning structural components, such as attention heads, neurons, and layers, aiming to achieve a Pareto-optimal balance between performance and efficiency. While NAS already achieved promising results on small language models in previous work, in this paper we propose various extensions that allow us to scale to LLMs. Compared to structural pruning baselines, we show that NAS improves performance up to 3.4% on MMLU with an on-device latency speedup.

The increasing size of language models necessitates a thorough analysis across multiple dimensions to assess trade-offs among crucial hardware metrics such as latency, energy consumption, GPU memory usage, and performance. Identifying optimal model configurations under specific hardware constraints is becoming essential but remains challenging due to the computational load of exhaustive training and evaluation on multiple devices. To address this, we introduce HW-GPT-Bench, a hardware-aware benchmark that utilizes surrogate predictions to approximate various hardware metrics across 13 devices of architectures in the GPT-2 family, with architectures containing up to 774M parameters. Our surrogates, via calibrated predictions and reliable uncertainty estimates, faithfully model the heteroscedastic noise inherent in the energy and latency measurements. To estimate perplexity, we employ weight-sharing techniques from Neural Architecture Search (NAS), inheriting pretrained weights from the largest GPT-2 model. Finally, we demonstrate the utility of HW-GPT-Bench by simulating optimization trajectories of various multi-objective optimization algorithms in just a few seconds.

Pareto front profiling in multi-objective optimization (MOO), i.e. finding a diverse set of Pareto optimal solutions, is challenging, especially with expensive objectives like neural network training. Typically, in MOO neural architecture search (NAS), we aim to balance performance and hardware metrics across devices. Prior NAS approaches simplify this task by incorporating hardware constraints into the objective function, but profiling the Pareto front necessitates a search for each constraint. In this work, we propose a novel NAS algorithm that encodes user preferences for the trade-off between performance and hardware metrics, and yields representative and diverse architectures across multiple devices in just one search run. To this end, we parameterize the joint architectural distribution across devices and multiple objectives via a hypernetwork that can be conditioned on hardware features and preference vectors, enabling zero-shot transferability to new devices. Extensive experiments with up to 19 hardware devices and 3 objectives showcase the effectiveness and scalability of our method. Finally, we show that, without additional costs, our method outperforms existing MOO NAS methods across qualitatively different search spaces and datasets, including MobileNetV3 on ImageNet-1k and a Transformer space on machine translation.

Weight sharing is a fundamental concept in neural architecture search (NAS), enabling gradient-based methods to explore cell-based architecture spaces significantly faster than traditional blackbox approaches. In parallel, weight entanglement has emerged as a technique for intricate parameter sharing among architectures within macro-level search spaces. Since weight-entanglement poses compatibility challenges for gradient-based NAS methods, these two paradigms have largely developed independently in parallel sub-communities. This paper aims to bridge the gap between these sub-communities by proposing a novel scheme to adapt gradient-based methods for weight-entangled spaces. This enables us to conduct an in-depth comparative assessment and analysis of the performance of gradient-based NAS in weight-entangled search spaces. Our findings reveal that this integration of weight-entanglement and gradient-based NAS brings forth the various benefits of gradient-based methods (enhanced performance, improved supernet training properties and superior any-time performance), while preserving the memory efficiency of weight-entangled spaces. The code for our work is openly accessible here. The concept of weight-sharing in Neural Architecture Search (NAS) arose from the need to improve the efficiency of conventional blackbox NAS algorithms, which demand significant computational resources to evaluate individual architectures. Here, weight-sharing (WS) refers to the paradigm by which we represent the search space with a single large supernet, also known as the one-shot model, that subsumes all the candidate architectures in that space. Every edge of this supernet holds all the possible operations that can be assigned to that edge. Gradient-based NAS algorithms (or optimizers), such as DARTS (Liu et al., 2019), GDAS (Dong and Yang, 2019) and DrNAS (Chen et al., 2021b), assign an architectural parameter to every choice of operation on a given edge of the supernet.

Face recognition systems are widely deployed in safety-critical applications, including law enforcement, yet they exhibit bias across a range of socio-demographic dimensions, such as gender and race. Conventional wisdom dictates that model biases arise from biased training data. As a consequence, previous works on bias mitigation largely focused on pre-processing the training data, adding penalties to prevent bias from effecting the model during training, or post-processing predictions to debias them, yet these approaches have shown limited success on hard problems such as face recognition. In our work, we discover that biases are actually inherent to neural network architectures themselves. Following this reframing, we conduct the first neural architecture search for fairness, jointly with a search for hyperparameters. Our search outputs a suite of models which Pareto-dominate all other high-performance architectures and existing bias mitigation methods in terms of accuracy and fairness, often by large margins, on the two most widely used datasets for face identification, CelebA and VGGFace2. Furthermore, these models generalize to other datasets and sensitive attributes. We release our code, models and raw data files at https://github.com/dooleys/FR-NAS.