Jaeseong Lee, Seung-won Hwang, Aurick Qiao, Daniel F Campos, Zhewei Yao, Yuxiong He. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

LLM inference for enterprise applications, such as summarization, RAG, and code-generation, typically observe much longer prompt than generations, leading to high prefill cost and response latency.We present SwiftKV, a novel model transformation and distillation procedure targeted at reducing the prefill compute (in FLOPs) of prompt tokens while preserving high generation quality.First, SwiftKV prefills later layers' KV cache using an earlier layer's output, allowing prompt tokens to skip those later layers.Second, SwiftKV employs a lightweight knowledge-preserving distillation procedure that can adapt existing LLMs with minimal accuracy impact.Third, SwiftKV can naturally incorporate KV cache compression to improve inference performance in low-memory scenarios.Our comprehensive experiments show that SwiftKV can effectively reduce prefill computation by 25-50% across several LLM families while incurring minimum quality degradation.In the end-to-end inference serving, SwiftKV realizes up to 2× higher aggregate throughput and 60% lower time per output token.It can achieve a staggering 560 TFlops/GPU of normalized inference throughput, which translates

Test-time reasoning algorithms such as chain-of-thought, self-consistency, and MCTS enhance LLM problem-solving but can wastefully generate many tokens without improving accuracy. At the same time, we observe that these algorithms exhibit answer stabilization: their intermediate solutions often cease to change after a certain point, and further investment of compute does not change their final answer. To quantify this phenomenon, we introduce Certaindex, an algorithm-agnostic metric measuring this evolving stability, signaling when further computation is unlikely to alter the final result. Certaindex is lightweight, can accelerate reasoning program inference via early exit, and further enables dynamic token allocation, gang scheduling, and many opportunities when integrated with real-world LLM serving systems. To quantify real-world benefits, we built Certaindex as a scheduler into Dynasor, our reasoning-aware LLM serving system, and demonstrate up to 50\% compute savings and 3.3$\times$ higher throughput in real workloads with no accuracy drop. Our code is available at https://github.com/hao-ai-lab/Dynasor.git

Speculative decoding is widely adopted to reduce latency in large language model (LLM) inference by leveraging smaller draft models capable of handling diverse user tasks. However, emerging AI applications, such as LLM-based agents, present unique workload characteristics: instead of diverse independent requests, agentic frameworks typically submit repetitive inference requests, such as multi-agent pipelines performing similar subtasks or self-refinement loops iteratively enhancing outputs. These workloads result in long and highly predictable sequences, which current speculative decoding methods do not effectively exploit. To address this gap, we introduce \emph{SuffixDecoding}, a novel method that utilizes efficient suffix trees to cache long token sequences from prompts and previous outputs. By adaptively speculating more tokens when acceptance likelihood is high and fewer when it is low, SuffixDecoding effectively exploits opportunities for longer speculations while conserving computation when those opportunities are limited. Evaluations on agentic benchmarks, including SWE-Bench and Text-to-SQL, demonstrate that SuffixDecoding achieves speedups of up to 3.9$\times$, outperforming state-of-the-art methods -- 2.2$\times$ faster than model-based approaches like EAGLE-2/3 and 1.6$\times$ faster than model-free approaches such as Token Recycling. SuffixDecoding is open-sourced.

Abstract KVCache, by storing key-value pairs for reuse, has been crucial for enhancing inference efficiency for large language models (LLMs). However, the increasing memory demands of KVCache, especially with recent trends of longer input sequences, present a major challenge. In this work, we propose an innovative token-adaptive low-rank approximation strategy for KVCache compression. By applying varying ranks based on token significance, our method compresses KVCache efficiently while retaining critical information. Moreover, we introduce a lazy approximation technique, which approximates lazily only when needed, alongside a reconstruction-free design to bypass costly recalculations. Combined with multi-level quantization, this method reduces KVCache size by 9.1× on the Llama-3.1-8B model, with minimal performance degradation on complex tasks such as GSM8K. Moreover, our custom attention implementation shows up to 2× latency reduction compared to the conventional method in long context scenarios. The code is publicly available.

In Large Language Model (LLM) inference, the output length of an LLM request is typically regarded as not known a priori. Consequently, most LLM serving systems employ a simple First-come-first-serve (FCFS) scheduling strategy, leading to Head-Of-Line (HOL) blocking and reduced throughput and service quality. In this paper, we reexamine this assumption -- we show that, although predicting the exact generation length of each request is infeasible, it is possible to predict the relative ranks of output lengths in a batch of requests, using learning to rank. The ranking information offers valuable guidance for scheduling requests. Building on this insight, we develop a novel scheduler for LLM inference and serving that can approximate the shortest-job-first (SJF) schedule better than existing approaches. We integrate this scheduler with the state-of-the-art LLM serving system and show significant performance improvement in several important applications: 2.8x lower latency in chatbot serving and 6.5x higher throughput in synthetic data generation. Our code is available at https://github.com/hao-ai-lab/vllm-ltr.git

The Sia scheduler efficiently assigns heterogeneous deep learning (DL) cluster resources to elastic resource-adaptive jobs. Although some recent schedulers address one aspect or another (e.g., heterogeneity or resource-adaptivity), none addresses all and most scale poorly to large clusters and/or heavy workloads even without the full complexity of the combined scheduling problem. Sia introduces a new scheduling formulation that can scale to the search-space sizes and intentionally match jobs and their configurations to GPU types and counts, while adapting to changes in cluster load and job mix over time. Sia also introduces a low-profiling-overhead approach to bootstrapping (for each new job) throughput models used to evaluate possible resource assignments, and it is the first cluster scheduler to support elastic scaling of hybrid parallel jobs.

Machine learning (ML) training algorithms often possess an inherent self-correcting behavior due to their iterative- convergent nature. Recent systems exploit this property to achieve adaptability and efficiency in unreliable computing environments by relaxing the consistency of execution and allowing calculation errors to be self-corrected during training. However, the behavior of such systems are only well understood for specific types of calculation errors, such as those caused by staleness, reduced precision, or asynchronicity, and for specific algorithms, such as stochastic gradient descent. In this paper, we develop a general framework to quantify the effects of calculation errors on iterative-convergent algorithms. We then use this framework to derive a worst-case upper bound on the cost of arbitrary perturbations to model parameters during training and to design new strategies for checkpoint-based fault tolerance. Our system, SCAR, can reduce the cost of partial failures by 78%{–}95% when compared with traditional checkpoint-based fault tolerance across a variety of ML models and training algorithms, providing near-optimal performance in recovering from failures.

At the core of Machine Learning (ML) analytics is often an expert-suggested model, whose parameters are refined by iteratively processing a training dataset until convergence. The completion time (i.e. convergence time) and quality of the learned model not only depends on the rate at which the refinements are generated but also the quality of each refinement. While data-parallel ML applications often employ a loose consistency model when updating shared model parameters to maximize parallelism, the accumulated error may seriously impact the quality of refinements and thus delay completion time, a problem that usually gets worse with scale. Although more immediate propagation of updates reduces the accumulated error, this strategy is limited by physical network bandwidth. Additionally, the performance of the widely used stochastic gradient descent (SGD) algorithm is sensitive to step size. Simply increasing communication often fails to bring improvement without tuning step size accordingly and tedious hand tuning is usually needed to achieve optimal performance.

The idea of multi-pivot quicksort has recently received the attention of researchers after Vladimir Yaroslavskiy proposed a dual pivot quicksort algorithm that, contrary to prior intuition, outperforms standard quicksort by a a significant margin under the Java JVM [10]. More recently, this algorithm has been analysed in terms of comparisons and swaps by Wild and Nebel [9]. Our contributions to the topic are as follows. First, we perform the previous experiments using a native C implementation thus removing potential extraneous effects of the JVM. Second, we provide analyses on cache behavior of these algorithms. We then provide strong evidence that cache behavior is causing most of the performance differences in these algorithms. Additionally, we build upon prior work in multi-pivot quicksort and propose a 3-pivot variant that performs very well in theory and practice. We show that it makes fewer comparisons and has better cache behavior than the dual pivot quicksort in the expected case. We validate this with experimental results, showing a 7–8% performance improvement in our tests.