Efficient deployment of large language models, particularly Mixture of Experts (MoE) models, on resource-constrained platforms presents significant challenges in terms of computational efficiency and memory utilization. The MoE architecture, renowned for its ability to increase model capacity without a proportional increase in inference cost, greatly reduces the token generation latency compared with dense models. However, the large model size makes MoE models inaccessible to individuals without high-end GPUs. In this paper, we propose a high-throughput MoE batch inference system, MoE-Lightning, that significantly outperforms past work. MoE-Lightning introduces a novel CPU-GPU-I/O pipelining schedule, CGOPipe, with paged weights to achieve high resource utilization, and a performance model, HRM, based on a Hierarchical Roofline Model we introduce to help find policies with higher throughput than existing systems. MoE-Lightning can achieve up to (10.3x) higher throughput than state-of-the-art offloading-enabled LLM inference systems for Mixtral 8x7B on a single T4 GPU (16GB). When the theoretical system throughput is bounded by the GPU memory, MoE-Lightning can reach the throughput upper bound with 2-3x less CPU memory, significantly increasing resource utilization. MoE-Lightning also supports efficient batch inference for much larger MoEs (e.g., Mixtral 8x22B and DBRX) on multiple low-cost GPUs (e.g., 2--4 T4s).

Evaluating aligned large language models' (LLMs) ability to recognize and reject unsafe user requests is crucial for safe, policy-compliant deployments. Existing evaluation efforts, however, face three limitations that we address with **SORRY-Bench**, our proposed benchmark. **First**, existing methods often use coarse-grained taxonomies of unsafe topics, and are over-representing some fine-grained topics. For example, among the ten existing datasets that we evaluated, tests for refusals of self-harm instructions are over 3x less represented than tests for fraudulent activities. SORRY-Bench improves on this by using a fine-grained taxonomy of 44 potentially unsafe topics, and 440 class-balanced unsafe instructions, compiled through human-in-the-loop methods. **Second**, evaluations often overlook the linguistic formatting of prompts, like different languages, dialects, and more --- which are only implicitly considered in many evaluations. We supplement SORRY-bench with 20 diverse linguistic augmentations to systematically examine these effects. **Third**, existing evaluations rely on large LLMs (e.g., GPT-4) for evaluation, which can be computationally expensive. We investigate design choices for creating a fast, accurate automated safety evaluator. By collecting 7K+ human annotations and conducting a meta-evaluation of diverse LLM-as-a-judge designs, we show that fine-tuned 7B LLMs can achieve accuracy comparable to GPT-4 scale LLMs, with lower computational cost. Putting these together, we evaluate over 50 proprietary and open-weight LLMs on SORRY-Bench, analyzing their distinctive safety refusal behaviors. We hope our effort provides a building block for systematic evaluations of LLMs' safety refusal capabilities, in a balanced, granular, and efficient manner. Benchmark demo, data, code, and models are available through [https://sorry-bench.github.io](https://sorry-bench.github.io).

Studying how people interact with large language models (LLMs) in real-world scenarios is increasingly important due to their widespread use in various applications. In this paper, we introduce LMSYS-Chat-1M, a large-scale dataset containing one million real-world conversations with 25 state-of-the-art LLMs. This dataset is collected from 210K unique IP addresses in the wild on our Vicuna demo and Chatbot Arena website. We offer an overview of the dataset's content, including its curation process, basic statistics, and topic distribution, highlighting its diversity, originality, and scale. We demonstrate its versatility through four use cases: developing content moderation models that perform similarly to GPT-4, building a safety benchmark, training instruction-following models that perform similarly to Vicuna, and creating challenging benchmark questions. We believe that this dataset will serve as a valuable resource for understanding and advancing LLM capabilities. The dataset is publicly available at https://huggingface.co/datasets/lmsys/lmsys-chat-1m.

Large Language Models (LLMs) have unlocked new capabilities and applications; however, evaluating the alignment with human preferences still poses significant challenges. To address this issue, we introduce Chatbot Arena, an open platform for evaluating LLMs based on human preferences. Our methodology employs a pairwise comparison approach and leverages input from a diverse user base through crowdsourcing. The platform has been operational for several months, amassing over 240K votes. This paper describes the platform, analyzes the data we have collected so far, and explains the tried-and-true statistical methods we are using for efficient and accurate evaluation and ranking of models. We confirm that the crowdsourced questions are sufficiently diverse and discriminating and that the crowd-sourced human votes are in good agreement with those of expert raters. These analyses collectively establish a robust foundation for the credibility of Chatbot Arena. Because of its unique value and openness, Chatbot Arena has emerged as one of the most referenced LLM leaderboards, widely cited by leading LLM developers and companies. The platform is publicly available at https://chat.lmsys.org.

Large language models (LLMs) are increasingly used for complex tasks that require multiple generation calls, advanced prompting techniques, control flow, and structured inputs/outputs. However, efficient systems are lacking for programming and executing these applications. We introduce SGLang, a system for efficient execution of complex language model programs. SGLang consists of a frontend language and a runtime. The frontend simplifies programming with primitives for generation and parallelism control. The runtime accelerates execution with novel optimizations like RadixAttention for KV cache reuse and compressed finite state machines for faster structured output decoding. Experiments show that SGLang achieves up to $6.4\times$ higher throughput compared to state-of-the-art inference systems on various large language and multi-modal models on tasks including agent control, logical reasoning, few-shot learning benchmarks, JSON decoding, retrieval-augmented generation pipelines, and multi-turn chat. The code is publicly available at https://github.com/sgl-project/sglang.

The high computational and memory requirements of large language model (LLM) inference make it feasible only with multiple high-end accelerators. Motivated by the emerging demand for latency-insensitive tasks with batched processing, this paper initiates the study of high-throughput LLM inference using limited resources, such as a single commodity GPU. We present FlexGen, a high-throughput generation engine for running LLMs with limited GPU memory. FlexGen can be flexibly configured under various hardware resource constraints by aggregating memory and computation from the GPU, CPU, and disk. By solving a linear programming problem, it searches for efficient patterns to store and access tensors. FlexGen further compresses the weights and the attention cache to 4 bits with negligible accuracy loss. These techniques enable FlexGen to have a larger space of batch size choices and thus significantly increase maximum throughput. As a result, when running OPT-175B on a single 16GB GPU, FlexGen achieves significantly higher throughput compared to state-of-the-art offloading systems, reaching a generation throughput of 1 token/s for the first time with an effective batch size of 144. On the HELM benchmark, FlexGen can benchmark a 30B model with a 16GB GPU on 7 representative sub-scenarios in 21 hours. The code is available at https://github.com/FMInference/FlexGen.

Large Language Models (LLMs), despite their recent impressive accomplishments, are notably cost-prohibitive to deploy, particularly for applications involving long-content generation, such as dialogue systems and story writing. Often, a large amount of transient state information, referred to as the $\mathsf{KV}$ $\mathsf{cache}$, is stored in GPU memory in addition to model parameters, scaling linearly with the sequence length and batch size. In this paper, we introduce a novel approach for implementing the $\mathsf{KV}$ $\mathsf{cache}$ which significantly reduces its memory footprint. Our approach is based on the noteworthy observation that a small portion of tokens contributes most of the value when computing attention scores. We call these tokens Heavy Hitters ($\mathsf{H_2}$). Through a comprehensive investigation, we find that ($i$) the emergence of $\mathsf{H_2}$ is natural and strongly correlates with the frequent co-occurrence of tokens in the text, and ($ii$) removing them results in significant performance degradation. Based on these insights, we propose Heavy Hitter Oracle ($\mathsf{H_2O}$), a $\mathsf{KV}$ $\mathsf{cache}$ eviction policy that dynamically retains a balance of recent and $\mathsf{H_2}$ tokens. We formulate the $\mathsf{KV}$ $\mathsf{cache}$ eviction as a dynamic submodular problem and prove (under mild assumptions) a theoretical guarantee for our novel eviction algorithm which could help guide future work. We validate the accuracy of our algorithm with OPT, LLaMA, and GPT-NeoX across a wide range of tasks. Our implementation of $\mathsf{H_2O}$ with 20\% heavy hitters improves the throughput over three leading inference systems DeepSpeed Zero-Inference, Hugging Face Accelerate, and FlexGen by up to $29\times$, $29\times$, and $3\times$ on OPT-6.7B and OPT-30B. With the same batch size, $\mathsf{H_2O}$ can reduce the latency by up to $1.9\times$.

Evaluating large language model (LLM) based chat assistants is challenging due to their broad capabilities and the inadequacy of existing benchmarks in measuring human preferences.To address this, we explore using strong LLMs as judges to evaluate these models on more open-ended questions.We examine the usage and limitations of LLM-as-a-judge, including position, verbosity, and self-enhancement biases, as well as limited reasoning ability, and propose solutions to mitigate some of them.We then verify the agreement between LLM judges and human preferences by introducing two benchmarks: MT-bench, a multi-turn question set; and Chatbot Arena, a crowdsourced battle platform.Our results reveal that strong LLM judges like GPT-4 can match both controlled and crowdsourced human preferences well, achieving over 80\% agreement, the same level of agreement between humans.Hence, LLM-as-a-judge is a scalable and explainable way to approximate human preferences, which are otherwise very expensive to obtain.Additionally, we show our benchmark and traditional benchmarks complement each other by evaluating several variants of LLaMA and Vicuna.The MT-bench questions, 3K expert votes, and 30K conversations with human preferences are publicly available at https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.

Large Language Models (LLMs) and other large foundation models have achieved impressive results, but their size exacerbates existing resource consumption and latency challenges. In particular, the large-scale deployment of these models is hindered by the significant resource requirements during inference. In this paper, we study two approaches for mitigating these challenges: employing a cache to store previous queries and learning a model selector to choose from an ensemble of models for query processing. Theoretically, we provide an optimal algorithm for jointly optimizing both approaches to reduce the inference cost in both offline and online tabular settings. By combining a caching algorithm, namely Greedy Dual Size with Frequency (GDSF) or Least Expected Cost (LEC), with a model selector, we achieve optimal rates in both offline and online settings. Empirically, simulations show that our caching and model selection algorithm greatly improves over the baselines, with up to $50\times$ improvement over the baseline when the ratio between the maximum cost and minimum cost is $100$. Experiments on real datasets show a $4.3\times$ improvement in FLOPs over the baseline when the ratio for FLOPs is $10$, and a $1.8\times$ improvement in latency when the ratio for average latency is $1.85$.

High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. However, existing systems struggle because the key-value cache (KV cache) memory for each request is huge and grows and shrinks dynamically. When managed inefficiently, this memory can be significantly wasted by fragmentation and redundant duplication, limiting the batch size. To address this problem, we propose PagedAttention, an attention algorithm inspired by the classical virtual memory and paging techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce memory usage. Our evaluations show that vLLM improves the throughput of popular LLMs by 2--4× with the same level of latency compared to the state-of-the-art systems, such as FasterTransformer and Orca. The improvement is more pronounced with longer sequences, larger models, and more complex decoding algorithms. vLLM's source code is publicly available at https://github.com/vllm-project/vllm.

Abstract cvc5 is the latest SMT solver in the cooperating validity checker series and builds on the successful code base of CVC4. This paper serves as a comprehensive system description of cvc5 ’s architectural design and highlights the major features and components introduced since CVC4 1.8. We evaluate cvc5 ’s performance on all benchmarks in SMT-LIB and provide a comparison against CVC4 and Z3.

Abstract In this paper, we discuss exact and parameterized algorithms for the problem of finding a read-once refutation (ROR) in an unsatisfiable Horn constraint system (HCS). Recall that a linear constraint system $\mathbf {A \cdot x \ge b}$ is said to be an HCS if each entry in $\textbf {A}$ belongs to the set $\{0,1,-1\}$ and at most one entry in each row of $\textbf {A}$ is positive. In this paper, we examine the importance of constraints in which more variables have negative coefficients than positive coefficients. In particular, we study the impact of the proportion of these ‘net-negative’ constraints has on the difficulty of finding RORs. There exist several algorithms for checking whether an HCS is feasible. To the best of our knowledge, these algorithms are not certifying, i.e. they do not provide a certificate of infeasibility. Our work is concerned with providing a specialized class of certificates called ‘read-once refutations’. In an ROR, each constraint defining the HCS may be used at most once in the derivation of a refutation. The problem of checking if an HCS has an ROR has been shown to be NP-hard. We analyse the HCS ROR problem from three different algorithmic perspectives, viz., parameterized algorithms, exact exponential algorithms and approximation algorithms. In particular, we show that the HCS ROR problem is fixed-parameter tractable (FPT) with respect to the number of constraints in the system that have more variables with negative coefficient than variables with positive coefficient. Additionally, we show that the HCS ROR problem becomes easy when this parameter is both small and large. We also derive an algorithm that runs in time $O(1.66^{m})$, where $m$ is the number of constraints in the HCS. On the lower-bound side, we derive a lower bound on the algorithmic resources needed for this problem using the Exponential Time Hypothesis. We also establish that the HCS ROR problem does not have a polynomial kernel when the number of constraints with three or more variables in a refutation is used as a parameter. Finally, we show that the problem of approximating the length of the shortest ROR in an HCS is NPO PB-complete1.

Abstract We make two contributions to the study of polite combination in satisfiability modulo theories. The first is a separation between politeness and strong politeness, by presenting a polite theory that is not strongly polite. This result shows that proving strong politeness (which is often harder than proving politeness) is sometimes needed in order to use polite combination. The second contribution is an optimization to the polite combination method, obtained by borrowing from the Nelson-Oppen method. The Nelson-Oppen method is based on guessing arrangements over shared variables. In contrast, polite combination requires an arrangement overallvariables of the shared sorts. We show that when using polite combination, if the other theory is stably infinite with respect to a shared sort, only the shared variables of that sort need be considered in arrangements, as in the Nelson-Oppen method. The time required to reason about arrangements is exponential in the worst case, so reducing the number of variables considered has the potential to improve performance significantly. We show preliminary evidence for this by demonstrating a speed-up on a smart contract verification benchmark.

Algebraic datatypes, and among them lists and trees, have attracted a lot of interest in automated reasoning and Satisfiability Modulo Theories (SMT). Since its latest stable version, the SMT-LIB standard defines a theory of algebraic datatypes, which is currently supported by several mainstream SMT solvers. In this paper, we study this particular theory of datatypes and prove that it is strongly polite, showing also how it can be combined with other arbitrary disjoint theories using polite combination. Our results cover both inductive and finite datatypes, as well as their union. The combination method uses a new, simple, and natural notion of additivity, that enables deducing strong politeness from (weak) politeness.