Speculative decoding accelerates Large Language Model (LLM) inference by employing a small speculative model (SSM) to generate multiple candidate tokens and verify them using the LLM in parallel. This technique has been widely integrated into LLM inference serving systems. However, inference requests typically exhibit uncertain execution time, which poses a significant challenge of efficiently scheduling requests in these systems. Existing work estimates execution time based solely on predicted output length, which could be inaccurate because execution time depends on both output length and token acceptance rate of verification by the LLM. In this paper, we propose a semi-clairvoyant request scheduling algorithm called Least-Attained/Perceived-Service for Speculative Decoding (LAPS-SD). Given a number of inference requests, LAPS-SD can effectively minimize average inference latency by adaptively scheduling requests according to their features during decoding. When the token acceptance rate is dynamic and execution time is difficult to estimate, LAPS-SD maintains multiple priority queues and allows request execution preemption across different queues. Once the token acceptance rate becomes stable, LAPS-SD can accurately estimate the execution time and schedule requests accordingly. Extensive experiments show that LAPS-SD reduces inference latency by approximately 39\% compared to state-of-the-art scheduling methods.

As large language models (LLMs) have shown great success in many tasks, they are used in various applications. While a lot of works have focused on the efficiency of single-LLM application (e.g., offloading, request scheduling, parallelism strategy selection), multi-LLM applications receive less attention, particularly in offline inference scenarios. In this work, we aim to improve the offline end-to-end inference efficiency of multi-LLM applications in the single-node multi-GPU environment. The problem involves two key decisions: (1) determining which LLMs to run concurrently each time (we may not run all the models at the same time), and (2) selecting a parallelism strategy to use for each LLM. This problem is NP-hard. Naive solutions may not work well because the running time for a model to complete a set of requests depends on the request workload and the selected parallelism strategy, and they lack an accurate model of the running time. As the LLM output lengths are unknown before running, to estimate the model running time, we propose a sampling-then-simulation method which first estimates the output lengths by sampling from an empirical cumulative function we obtained from a large dataset in advance, and then simulates the LLM inference process accordingly. Based on the simulation, we estimate the per-iteration latencys to get the total latency. A greedy method is proposed to optimize the scheduling of the LLMs in the application across the GPUs. We then propose a framework SamuLLM which contains two phases: planning, which calls the greedy method for an application and running, which runs the application and dynamically adjust the model scheduling based on the runtime information. Experiments on 3 applications and a mixed application show that SamuLLM can achieve 1.0-2.4$\times$ end-to-end speedups compared to the competitors.

In Large Language Model (LLM) serving, the KV-cache (KVC) bottleneck causes high tail Time-to-First-Token (TTFT) and Time-Between-Tokens (TBT), impairing user experience, particularly in time-sensitive applications. However, satisfying both TTFT and TBT service-level objectives (SLOs) is challenging. To address this, we propose a system, named CacheOPT for mitigating KV Cache competition, based on key insights from our measurements, incorporating novel components. First, it estimates a request's output length, bounding the deviation with a high specified probability, adjusted based on the request arrival rate. Second, it allocates the estimated KVC demand to a request, and reuses other requests' allocated KVC to avoid preemptions while reducing waiting time. Third, it proactively allocates KVC before instead of at the time a request exhausts its allocation and reserves KVC globally to prevent preemptions. Fourth, it chooses a request that has long TBT SLO, long job remaining time and short preemption time to preempt. Fifth, it selects the shortest-latency strategy between swapping and recomputation for preemptions. Experiments show that CacheOPT achieves up to 3.29$\times$ and 2.83$\times$ lower tail TBT and tail TTFT, 47\% and 53\% higher TTFT and TBT SLO attainments, and supports up to 1.58$\times$ higher request arrival rate than the state-of-the-art methods.

Large Language Models have revolutionized natural language processing, yet serving them efficiently in data centers remains challenging due to mixed workloads comprising latency-sensitive (LS) and best-effort (BE) jobs. Existing inference systems employ iteration-level first-come-first-served scheduling, causing head-of-line blocking when BE jobs delay LS jobs. We introduce QLLM, a novel inference system designed for Mixture of Experts (MoE) models, featuring a fine-grained, priority-aware preemptive scheduler. QLLM enables expert-level preemption, deferring BE job execution while minimizing LS time-to-first-token (TTFT). Our approach removes iteration-level scheduling constraints, enabling the scheduler to preempt jobs at any layer based on priority. Evaluations on an Nvidia A100 GPU show that QLLM significantly improves performance. It reduces LS TTFT by an average of $65.5\times$ and meets the SLO at up to $7$ requests/sec, whereas the baseline fails to do so under the tested workload. Additionally, it cuts LS turnaround time by up to $12.8\times$ without impacting throughput. QLLM is modular, extensible, and seamlessly integrates with Hugging Face MoE models.

Artificial intelligence (AI) model creators commonly attach restrictive terms of use to both their models and their outputs. These terms typically prohibit activities ranging from creating competing AI models to spreading disinformation. Often taken at face value, these terms are positioned by companies as key enforceable tools for preventing misuse, particularly in policy dialogs. But are these terms truly meaningful? There are myriad examples where these broad terms are regularly and repeatedly violated. Yet except for some account suspensions on platforms, no model creator has actually tried to enforce these terms with monetary penalties or injunctive relief. This is likely for good reason: we think that the legal enforceability of these licenses is questionable. This Article systematically assesses of the enforceability of AI model terms of use and offers three contributions. First, we pinpoint a key problem: the artifacts that they protect, namely model weights and model outputs, are largely not copyrightable, making it unclear whether there is even anything to be licensed. Second, we examine the problems this creates for other enforcement. Recent doctrinal trends in copyright preemption may further undermine state-law claims, while other legal frameworks like the DMCA and CFAA offer limited recourse. Anti-competitive provisions likely fare even worse than responsible use provisions. Third, we provide recommendations to policymakers. There are compelling reasons for many provisions to be unenforceable: they chill good faith research, constrain competition, and create quasi-copyright ownership where none should exist. There are, of course, downsides: model creators have fewer tools to prevent harmful misuse. But we think the better approach is for statutory provisions, not private fiat, to distinguish between good and bad uses of AI, restricting the latter.

Recent years have witnessed an explosive growth of AI models. The high cost of hosting AI services on GPUs and their demanding service requirements, make it timely and challenging to lower service costs and guarantee service quality. While spot instances have long been offered with a large discount, spot preemptions have discouraged users from using them to host model replicas when serving AI models. To address this, we introduce SkyServe, a system that efficiently serves AI models over a mixture of spot and on-demand replicas across regions and clouds. SkyServe intelligently spreads spot replicas across different failure domains (e.g., regions or clouds) to improve availability and reduce correlated preemptions, overprovisions cheap spot replicas than required as a safeguard against possible preemptions, and dynamically falls back to on-demand replicas when spot replicas become unavailable. We compare SkyServe with both research and production systems on real AI workloads: SkyServe reduces cost by up to 44% while achieving high resource availability compared to using on-demand replicas. Additionally, SkyServe improves P50, P90, and P99 latency by up to 2.6x, 3.1x, 2.7x compared to other research and production systems.

Please provide some additional clarification on what is to be compared here – in particular, I also stumbled over why (as stated in several corresponding theorems) one ends up with competitive ratios „at most min{ robustness-ratio, consistency-ratio }". At first glance, I had thought that if an algorithm does well for good predictions but the robustness bound is bad, that the latter should dominate. As this is apparently not the case, I ask for a brief explanation/clarification, preferably already in the introduction, as to why the minimum of the two values yields a bound on the competitive ratio.

Efficient scheduling is crucial for interactive Large Language Model (LLM) applications, where low request completion time directly impacts user engagement. Size-based scheduling algorithms like Shortest Remaining Process Time (SRPT) aim to reduce average request completion time by leveraging known or estimated request sizes and allowing preemption by incoming jobs with shorter service times. However, two main challenges arise when applying size-based scheduling to LLM systems. First, accurately predicting output lengths from prompts is challenging and often resource-intensive, making it impractical for many systems. As a result, the state-of-the-art LLM systems default to first-come, first-served scheduling, which can lead to head-of-line blocking and reduced system efficiency. Second, preemption introduces extra memory overhead to LLM systems as they must maintain intermediate states for unfinished (preempted) requests. In this paper, we propose TRAIL, a method to obtain output predictions from the target LLM itself. After generating each output token, we recycle the embedding of its internal structure as input for a lightweight classifier that predicts the remaining length for each running request. Using these predictions, we propose a prediction-based SRPT variant with limited preemption designed to account for memory overhead in LLM systems. This variant allows preemption early in request execution when memory consumption is low but restricts preemption as requests approach completion to optimize resource utilization. On the theoretical side, we derive a closed-form formula for this SRPT variant in an M/G/1 queue model, which demonstrates its potential value. In our system, we implement this preemption policy alongside our embedding-based prediction method.

Although various aspects of soft-constraint based norms have been explored, it is still challenging to understand preemption. Preemption is a situation where higher-level norms override lower-level norms when new information emerges. To address this, we propose a derivation state argumentation framework (DSA-framework). DSA-framework incorporates derivation states to explain how preemption arises based on evolving situational knowledge. Based on DSA-framework, we present an argumentative approach for explaining preemption. We formally prove that, under local optimality, DSA-framework can provide explanations why one consequence is obligatory or forbidden by soft-constraint based norms represented as logical constraint hierarchies.

Inference serving for large language models (LLMs) is the key to unleashing their potential in people's daily lives. However, efficient LLM serving remains challenging today because the requests are inherently heterogeneous and unpredictable in terms of resource and latency requirements, as a result of the diverse applications and the dynamic execution nature of LLMs. Existing systems are fundamentally limited in handling these characteristics and cause problems such as severe queuing delays, poor tail latencies, and SLO violations. We introduce Llumnix, an LLM serving system that reacts to such heterogeneous and unpredictable requests by runtime rescheduling across multiple model instances. Similar to context switching across CPU cores in modern operating systems, Llumnix reschedules requests to improve load balancing and isolation, mitigate resource fragmentation, and differentiate request priorities and SLOs. Llumnix implements the rescheduling with an efficient and scalable live migration mechanism for requests and their in-memory states, and exploits it in a dynamic scheduling policy that unifies the multiple rescheduling scenarios elegantly. Our evaluations show that Llumnix improves tail latencies by an order of magnitude, accelerates high-priority requests by up to 1.5x, and delivers up to 36% cost savings while achieving similar tail latencies, compared against state-of-the-art LLM serving systems. Llumnix is publicly available at https://github.com/AlibabaPAI/llumnix.