The Effect of Scheduling and Preemption on the Efficiency of LLM Inference Serving

The growing usage of Large Language Models (LLMs) highlights the demands and challenges in scalable LLM inference systems, affecting deployment and development processes. On the deployment side, there is a lack of comprehensive analysis on the conditions under which a particular scheduler performs better or worse, with performance varying substantially across different schedulers, hardware, models, and workloads. Manually testing each configuration on GPUs can be prohibitively expensive. On the development side, unpredictable performance and unknown upper limits can lead to inconclusive trial-and-error processes, consuming resources on ideas that end up ineffective. To address these challenges, we introduce INFERMAX, an analytical framework that uses inference cost models to compare various schedulers, including an optimal scheduler formulated as a constraint satisfaction problem (CSP) to establish an upper bound on performance. Our framework offers in-depth analysis and raises essential questions, challenging assumptions and exploring opportunities for more efficient scheduling. Notably, our findings indicate that preempting requests can reduce GPU costs by 30% compared to avoiding preemptions at all. We believe our methods and insights will facilitate the cost-effective deployment and development of scalable, efficient inference systems and pave the way for cost-based scheduling.

Beyond Autoregression: Fast LLMs via Self-Distillation Through Time

Autoregressive (AR) Large Language Models (LLMs) have demonstrated significant success across numerous tasks. However, the AR modeling paradigm presents certain limitations; for instance, contemporary autoregressive LLMs are trained to generate one token at a time, which can result in noticeable latency. Recent advances have indicated that search and repeated sampling can enhance performance in various applications, such as theorem proving, code generation, and alignment, by utilizing greater computational resources during inference. In this study, we demonstrate that diffusion language models are capable of generating at least 32 tokens simultaneously, while exceeding the performance of AR models in text quality and on the LAMBADA natural language understanding benchmark. This outcome is achieved through a novel distillation method for discrete diffusion models, which reduces the number of inference steps by a factor of 32-64. Practically, our models, even without caching, can generate tokens at a rate that is up to 8 times faster than AR models employing KV caching, and we anticipate further improvements with the inclusion of caching. Moreover, we demonstrate the efficacy of our approach for diffusion language models with up to 860M parameters.

Unpacking SDXL Turbo: Interpreting Text-to-Image Models with Sparse Autoencoders

Sparse autoencoders (SAEs) have become a core ingredient in the reverse engineering of large-language models (LLMs). For LLMs, they have been shown to decompose intermediate representations that often are not interpretable directly into sparse sums of interpretable features, facilitating better control and subsequent analysis. However, similar analyses and approaches have been lacking for text-to-image models. We investigated the possibility of using SAEs to learn interpretable features for a few-step text-to-image diffusion models, such as SDXL Turbo. To this end, we train SAEs on the updates performed by transformer blocks within SDXL Turbo's denoising U-net. We find that their learned features are interpretable, causally influence the generation process, and reveal specialization among the blocks. In particular, we find one block that deals mainly with image composition, one that is mainly responsible for adding local details, and one for color, illumination, and style. Therefore, our work is an important first step towards better understanding the internals of generative text-to-image models like SDXL Turbo and showcases the potential of features learned by SAEs for the visual domain. Code is available at https://github.com/surkovv/sdxl-unbox

SIKeD: Self-guided Iterative Knowledge Distillation for mathematical reasoning

Large Language Models (LLMs) can transfer their reasoning skills to smaller models by teaching them to generate the intermediate reasoning process required to solve multistep reasoning tasks. While LLMs can accurately solve reasoning tasks through a variety of strategies, even without fine-tuning, smaller models are not expressive enough to fit the LLMs distribution on all strategies when distilled and tend to prioritize one strategy over the others. This reliance on one strategy poses a challenge for smaller models when attempting to solve reasoning tasks that may be difficult with their preferred strategy. To address this, we propose a distillation method SIKeD (Self-guided Iterative Knowledge Distillation for mathematical reasoning), where the LLM teaches the smaller model to approach a task using different strategies and the smaller model uses its self-generated on-policy outputs to choose the most suitable strategy for the given task. The training continues in a self-guided iterative manner, where for each training iteration, a decision is made on how to combine the LLM data with the self-generated outputs. Unlike traditional distillation methods, SIKeD allows the smaller model to learn which strategy is suitable for a given task while continuously learning to solve a task using different strategies. Our experiments on various mathematical reasoning datasets show that SIKeD significantly outperforms traditional distillation techniques across smaller models of different sizes. Our code is available at: https://github.com/kumar-shridhar/SIKeD

The Role of Deep Learning Regularizations on Actors in Offline RL

Deep learning regularization techniques, such as \emph{dropout}, \emph{layer normalization}, or \emph{weight decay}, are widely adopted in the construction of modern artificial neural networks, often resulting in more robust training processes and improved generalization capabilities. However, in the domain of \emph{Reinforcement Learning} (RL), the application of these techniques has been limited, usually applied to value function estimators \citep{hiraoka2021dropout, smith2022walk}, and may result in detrimental effects. This issue is even more pronounced in offline RL settings, which bear greater similarity to supervised learning but have received less attention. Recent work in continuous offline RL has demonstrated that while we can build sufficiently powerful critic networks, the generalization of actor networks remains a bottleneck. In this study, we empirically show that applying standard regularization techniques to actor networks in offline RL actor-critic algorithms yields improvements of 6\% on average across two algorithms and three different continuous D4RL domains.

In Search for Architectures and Loss Functions in Multi-Objective Reinforcement Learning

Multi-objective reinforcement learning (MORL) is essential for addressing the intricacies of real-world RL problems, which often require trade-offs between multiple utility functions. However, MORL is challenging due to unstable learning dynamics with deep learning-based function approximators. The research path most taken has been to explore different value-based loss functions for MORL to overcome this issue. Our work empirically explores model-free policy learning loss functions and the impact of different architectural choices. We introduce two different approaches: Multi-objective Proximal Policy Optimization (MOPPO), which extends PPO to MORL, and Multi-objective Advantage Actor Critic (MOA2C), which acts as a simple baseline in our ablations. Our proposed approach is straightforward to implement, requiring only small modifications at the level of function approximator. We conduct comprehensive evaluations on the MORL Deep Sea Treasure, Minecart, and Reacher environments and show that MOPPO effectively captures the Pareto front. Our extensive ablation studies and empirical analyses reveal the impact of different architectural choices, underscoring the robustness and versatility of MOPPO compared to popular MORL approaches like Pareto Conditioned Networks (PCN) and Envelope Q-learning in terms of MORL metrics, including hypervolume and expected utility.

Investigating Low-Rank Training in Transformer Language Models: Efficiency and Scaling Analysis

State-of-the-art LLMs often rely on scale with high computational costs, which has sparked a research agenda to reduce parameter counts and costs without significantly impacting performance. Our study focuses on Transformer-based LLMs, specifically applying low-rank parametrization to the computationally intensive feedforward networks (FFNs), which are less studied than attention blocks. In contrast to previous works, (i) we explore low-rank parametrization at scale, up to 1.3B parameters; (ii) within Transformer language models rather than convolutional architectures; and (iii) starting from training from scratch. Experiments on the large RefinedWeb dataset show that low-rank parametrization is both efficient (e.g., 2.6$\times$ FFN speed-up with 32\% parameters) and effective during training. Interestingly, these structured FFNs exhibit steeper scaling curves than the original models. Motivated by this finding, we develop the wide and structured networks surpassing the current medium-sized and large-sized Transformer in perplexity and throughput performance.

HiPPO-Prophecy: State-Space Models can Provably Learn Dynamical Systems in Context

This work explores the in-context learning capabilities of State Space Models (SSMs) and presents, to the best of our knowledge, the first theoretical explanation of a possible underlying mechanism. We introduce a novel weight construction for SSMs, enabling them to predict the next state of any dynamical system after observing previous states without parameter fine-tuning. This is accomplished by extending the HiPPO framework to demonstrate that continuous SSMs can approximate the derivative of any input signal. Specifically, we find an explicit weight construction for continuous SSMs and provide an asymptotic error bound on the derivative approximation. The discretization of this continuous SSM subsequently yields a discrete SSM that predicts the next state. Finally, we demonstrate the effectiveness of our parameterization empirically. This work should be an initial step toward understanding how sequence models based on SSMs learn in context.

Self-Recognition in Language Models

A rapidly growing number of applications rely on a small set of closed-source language models (LMs). This dependency might introduce novel security risks if LMs develop self-recognition capabilities. Inspired by human identity verification methods, we propose a novel approach for assessing self-recognition in LMs using model-generated "security questions". Our test can be externally administered to keep track of frontier models as it does not require access to internal model parameters or output probabilities. We use our test to examine self-recognition in ten of the most capable open- and closed-source LMs currently publicly available. Our extensive experiments found no empirical evidence of general or consistent self-recognition in any examined LM. Instead, our results suggest that given a set of alternatives, LMs seek to pick the "best" answer, regardless of its origin. Moreover, we find indications that preferences about which models produce the best answers are consistent across LMs. We additionally uncover novel insights on position bias considerations for LMs in multiple-choice settings.

Building on Efficient Foundations: Effectively Training LLMs with Structured Feedforward Layers

State-of-the-art results in large language models (LLMs) often rely on scale, which becomes computationally expensive. This has sparked a research agenda to reduce these models' parameter count and computational costs without significantly impacting their performance. Our study focuses on transformer-based LLMs, specifically targeting the computationally intensive feedforward networks (FFN), which are less studied than attention blocks. We consider three candidate linear layer approximations in the FFN by combining efficient low-rank and block-diagonal matrices. In contrast to many previous works that examined these approximations, our study i) explores these structures from the training-from-scratch perspective, ii) scales up to 1.3B parameters, and iii) is conducted within recent Transformer-based LLMs rather than convolutional architectures. We first demonstrate they can lead to actual computational gains in various scenarios, including online decoding when using a pre-merge technique. Additionally, we propose a novel training regime, called \textit{self-guided training}, aimed at improving the poor training dynamics that these approximations exhibit when used from initialization. Experiments on the large RefinedWeb dataset show that our methods are both efficient and effective for training and inference. Interestingly, these structured FFNs exhibit steeper scaling curves than the original models. Further applying self-guided training to the structured matrices with 32\% FFN parameters and 2.5$\times$ speed-up enables only a 0.4 perplexity increase under the same training FLOPs. Finally, we develop the wide and structured networks surpassing the current medium-sized and large-sized Transformer in perplexity and throughput performance. Our code is available at \url{https://github.com/CLAIRE-Labo/StructuredFFN/tree/main}.