Large Language Model Inference Acceleration: A Comprehensive Hardware Perspective

Large Language Models (LLMs) have demonstrated remarkable capabilities across various fields, from natural language understanding to text generation. Compared to non-generative LLMs like BERT and DeBERTa, generative LLMs like GPT series and Llama series are currently the main focus due to their superior algorithmic performance. The advancements in generative LLMs are closely intertwined with the development of hardware capabilities. Various hardware platforms exhibit distinct hardware characteristics, which can help improve LLM inference performance. Therefore, this paper comprehensively surveys efficient generative LLM inference on different hardware platforms. First, we provide an overview of the algorithm architecture of mainstream generative LLMs and delve into the inference process. Then, we summarize different optimization methods for different platforms such as CPU, GPU, FPGA, ASIC, and PIM/NDP, and provide inference results for generative LLMs. Furthermore, we perform a qualitative and quantitative comparison of inference performance with batch sizes 1 and 8 on different hardware platforms by considering hardware power consumption, absolute inference speed (tokens/s), and energy efficiency (tokens/J). We compare the performance of the same optimization methods across different hardware platforms, the performance across different hardware platforms, and the performance of different methods on the same hardware platform. This provides a systematic and comprehensive summary of existing inference acceleration work by integrating software optimization methods and hardware platforms, which can point to the future trends and potential developments of generative LLMs and hardware technology for edge-side scenarios.

MARCA: Mamba Accelerator with ReConfigurable Architecture

We propose a Mamba accelerator with reconfigurable architecture, MARCA.We propose three novel approaches in this paper. (1) Reduction alternative PE array architecture for both linear and element-wise operations. For linear operations, the reduction tree connected to PE arrays is enabled and executes the reduction operation. For element-wise operations, the reduction tree is disabled and the output bypasses. (2) Reusable nonlinear function unit based on the reconfigurable PE. We decompose the exponential function into element-wise operations and a shift operation by a fast biased exponential algorithm, and the activation function (SiLU) into a range detection and element-wise operations by a piecewise approximation algorithm. Thus, the reconfigurable PEs are reused to execute nonlinear functions with negligible accuracy loss.(3) Intra-operation and inter-operation buffer management strategy. We propose intra-operation buffer management strategy to maximize input data sharing for linear operations within operations, and inter-operation strategy for element-wise operations between operations. We conduct extensive experiments on Mamba model families with different sizes.MARCA achieves up to 463.22$\times$/11.66$\times$ speedup and up to 9761.42$\times$/242.52$\times$ energy efficiency compared to Intel Xeon 8358P CPU and NVIDIA Tesla A100 GPU implementations, respectively.

Pareto-Optimal Learning from Preferences with Hidden Context

Ensuring AI models align with human values is essential for their safety and functionality. Reinforcement learning from human feedback (RLHF) uses human preferences to achieve this alignment. However, preferences sourced from diverse populations can result in point estimates of human values that may be sub-optimal or unfair to specific groups. We propose Pareto Optimal Preference Learning (POPL), which frames discrepant group preferences as objectives with potential trade-offs, aiming for policies that are Pareto-optimal on the preference dataset. POPL utilizes Lexicase selection, an iterative process to select diverse and Pareto-optimal solutions. Our empirical evaluations demonstrate that POPL surpasses baseline methods in learning sets of reward functions, effectively catering to distinct groups without access to group numbers or membership labels. Furthermore, we illustrate that POPL can serve as a foundation for techniques optimizing specific notions of group fairness, ensuring inclusive and equitable AI model alignment.

DALex: Lexicase-like Selection via Diverse Aggregation

Lexicase selection has been shown to provide advantages over other selection algorithms in several areas of evolutionary computation and machine learning. In its standard form, lexicase selection filters a population or other collection based on randomly ordered training cases that are considered one at a time. This iterated filtering process can be time-consuming, particularly in settings with large numbers of training cases. In this paper, we propose a new method that is nearly equivalent to lexicase selection in terms of the individuals that it selects, but which does so significantly more quickly. The new method, called DALex (for Diversely Aggregated Lexicase), selects the best individual with respect to a weighted sum of training case errors, where the weights are randomly sampled. This allows us to formulate the core computation required for selection as matrix multiplication instead of recursive loops of comparisons, which in turn allows us to take advantage of optimized and parallel algorithms designed for matrix multiplication for speedup. Furthermore, we show that we can interpolate between the behavior of lexicase selection and its "relaxed" variants, such as epsilon or batch lexicase selection, by adjusting a single hyperparameter, named "particularity pressure," which represents the importance granted to each individual training case. Results on program synthesis, deep learning, symbolic regression, and learning classifier systems demonstrate that DALex achieves significant speedups over lexicase selection and its relaxed variants while maintaining almost identical problem-solving performance. Under a fixed computational budget, these savings free up resources that can be directed towards increasing population size or the number of generations, enabling the potential for solving more difficult problems.

Optimizing Neural Networks with Gradient Lexicase Selection

One potential drawback of using aggregated performance measurement in machine learning is that models may learn to accept higher errors on some training cases as compromises for lower errors on others, with the lower errors actually being instances of overfitting. This can lead to both stagnation at local optima and poor generalization. Lexicase selection is an uncompromising method developed in evolutionary computation, which selects models on the basis of sequences of individual training case errors instead of using aggregated metrics such as loss and accuracy. In this paper, we investigate how lexicase selection, in its general form, can be integrated into the context of deep learning to enhance generalization. We propose Gradient Lexicase Selection, an optimization framework that combines gradient descent and lexicase selection in an evolutionary fashion. Our experimental results demonstrate that the proposed method improves the generalization performance of various widely-used deep neural network architectures across three image classification benchmarks. Additionally, qualitative analysis suggests that our method assists networks in learning more diverse representations. Our source code is available on GitHub: https://github.com/ld-ing/gradient-lexicase.

Objectives Are All You Need: Solving Deceptive Problems Without Explicit Diversity Maintenance

Navigating deceptive domains has often been a challenge in machine learning due to search algorithms getting stuck at sub-optimal local optima. Many algorithms have been proposed to navigate these domains by explicitly maintaining diversity or equivalently promoting exploration, such as Novelty Search or other so-called Quality Diversity algorithms. In this paper, we present an approach with promise to solve deceptive domains without explicit diversity maintenance by optimizing a potentially large set of defined objectives. These objectives can be extracted directly from the environment by sub-aggregating the raw performance of individuals in a variety of ways. We use lexicase selection to optimize for these objectives as it has been shown to implicitly maintain population diversity. We compare this technique with a varying number of objectives to a commonly used quality diversity algorithm, MAP-Elites, on a set of discrete optimization as well as reinforcement learning domains with varying degrees of deception. We find that decomposing objectives into many objectives and optimizing them outperforms MAP-Elites on the deceptive domains that we explore. Furthermore, we find that this technique results in competitive performance on the diversity-focused metrics of QD-Score and Coverage, without explicitly optimizing for these things. Our ablation study shows that this technique is robust to different subaggregation techniques. However, when it comes to non-deceptive, or ``illumination" domains, quality diversity techniques generally outperform our objective-based framework with respect to exploration (but not exploitation), hinting at potential directions for future work.

Ever Evolving Evaluator (EV3): Towards Flexible and Reliable Meta-Optimization for Knowledge Distillation

We introduce EV3, a novel meta-optimization framework designed to efficiently train scalable machine learning models through an intuitive explore-assess-adapt protocol. In each iteration of EV3, we explore various model parameter updates, assess them using pertinent evaluation methods, and adapt the model based on the optimal updates and previous progress history. EV3 offers substantial flexibility without imposing stringent constraints like differentiability on the key objectives relevant to the tasks of interest. Moreover, this protocol welcomes updates with biased gradients and allows for the use of a diversity of losses and optimizers. Additionally, in scenarios with multiple objectives, it can be used to dynamically prioritize tasks. With inspiration drawn from evolutionary algorithms, meta-learning, and neural architecture search, we investigate an application of EV3 to knowledge distillation. Our experimental results illustrate EV3's capability to safely explore model spaces, while hinting at its potential applicability across numerous domains due to its inherent flexibility and adaptability.

Quality Diversity through Human Feedback

Reinforcement learning from human feedback (RLHF) has exhibited the potential to enhance the performance of foundation models for qualitative tasks. Despite its promise, its efficacy is often restricted when conceptualized merely as a mechanism to maximize learned reward models of averaged human preferences, especially in areas such as image generation which demand diverse model responses. Meanwhile, quality diversity (QD) algorithms, dedicated to seeking diverse, high-quality solutions, are often constrained by the dependency on manually defined diversity metrics. Interestingly, such limitations of RLHF and QD can be overcome by blending insights from both. This paper introduces Quality Diversity through Human Feedback (QDHF), which employs human feedback for inferring diversity metrics, expanding the applicability of QD algorithms. Empirical results reveal that QDHF outperforms existing QD methods regarding automatic diversity discovery, and matches the search capabilities of QD with human-constructed metrics. Notably, when deployed for a latent space illumination task, QDHF markedly enhances the diversity of images generated by a Diffusion model. The study concludes with an in-depth analysis of QDHF's sample efficiency and the quality of its derived diversity metrics, emphasizing its promise for enhancing exploration and diversity in optimization for complex, open-ended tasks.

CLERA: A Unified Model for Joint Cognitive Load and Eye Region Analysis in the Wild

Non-intrusive, real-time analysis of the dynamics of the eye region allows us to monitor humans' visual attention allocation and estimate their mental state during the performance of real-world tasks, which can potentially benefit a wide range of human-computer interaction (HCI) applications. While commercial eye-tracking devices have been frequently employed, the difficulty of customizing these devices places unnecessary constraints on the exploration of more efficient, end-to-end models of eye dynamics. In this work, we propose CLERA, a unified model for Cognitive Load and Eye Region Analysis, which achieves precise keypoint detection and spatiotemporal tracking in a joint-learning framework. Our method demonstrates significant efficiency and outperforms prior work on tasks including cognitive load estimation, eye landmark detection, and blink estimation. We also introduce a large-scale dataset of 30k human faces with joint pupil, eye-openness, and landmark annotation, which aims to support future HCI research on human factors and eye-related analysis.

Particularity

We describe a design principle for adaptive systems under which adaptation is driven by particular challenges that the environment poses, as opposed to average or otherwise aggregated measures of performance over many challenges. We trace the development of this "particularity" approach from the use of lexicase selection in genetic programming to "particularist" approaches to other forms of machine learning and to the design of adaptive systems more generally.