HELIOS: Adaptive Model And Early-Exit Selection for Efficient LLM Inference Serving

• URL: http://arxiv.org/abs/2504.10724v2
• Date: Fri, 31 Oct 2025 16:06:03 GMT
• Title: HELIOS: Adaptive Model And Early-Exit Selection for Efficient LLM Inference Serving
• Authors: Avinash Kumar, Shashank Nag, Jason Clemons, Lizy John, Poulami Das,
• Abstract summary: Early-Exit Large Language Models (EE-LLMs) enable high throughput inference by allowing tokens to exit early at intermediate layers.<n>Existing EE-LLM frameworks rely on a single model and therefore, their token generation latencies are bottlenecked.<n>We propose $textitHELIOS, a framework that improves both token generation latency and batch sizes.
• Score: 2.7433801927536074
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Early-Exit Large Language Models (EE-LLMs) enable high throughput inference by allowing tokens to exit early at intermediate layers. However, their throughput is limited by the computational and memory savings. Existing EE-LLM frameworks rely on a single model and therefore, their token generation latencies are bottlenecked by tokens that do not exit early and traverse additional layers. Moreover, early exits are only known at runtime and depend on the request. Therefore, these frameworks load the weights of all model layers even though large portions remain unused when tokens exit early. The lack of memory savings limit us from scaling the batch sizes. We propose $\textit{HELIOS}$, a framework that improves both token generation latency and batch sizes to enable high-throughput in EE-LLMs. HELIOS exploits two insights. $\textit{First}$, early exits are often complimentary across models, tokens that do not exit early on one model often take an early-exit on another. HELIOS employs multiple models and dynamically switches between them to collectively maximize the number of tokens that exit early, and minimize token generation latencies. $\textit{Second}$, even when a predicted token does not exit early due to poor confidence, it often remains unchanged even after additional layer traversal. HELIOS greedily allows such tokens to exit early and only loads the weights of the most likely to be used layers, yielding memory savings which is then re-purposed to increase batch sizes. HELIOS employs real-time profiling to accurately identify the early-exit distributions, and adaptively switches between models by tracking tokens in real-time to minimize the performance degradation caused by greedy model loading and exiting. Our evaluations show that HELIOS achieves $1.48\times$ higher throughput and $15.14\times$ larger batch size compared to existing EE-LLM frameworks.

Related papers

• Continuous Autoregressive Language Models [56.49239051750678]
  We introduce Continuous Autoregressive Language Models (CALM)<n>CALM uses a high-fidelity autoencoder to compress a chunk of K tokens into a single continuous vector.<n>We develop a comprehensive likelihood-free framework that enables robust training, evaluation, and controllable sampling.
  arXiv  Detail & Related papers  (2025-10-31T17:58:11Z)
• Fast Inference via Hierarchical Speculative Decoding [65.40448210801763]
  We introduce Hierarchical Speculative Decoding (HSD), an algorithm that stacks draft models into a hierarchy, where each model proposes tokens, and the next larger model verifies them in a single forward pass.<n>HSD gives up to 1.2x speed-up over the best single-draft baseline.
  arXiv  Detail & Related papers  (2025-10-22T15:56:19Z)
• Pipeline Parallelism is All You Need for Optimized Early-Exit Based Self-Speculative Decoding [73.67253077506672]
  Large language models (LLMs) deliver impressive generation quality, but incur very high inference cost.<n>Early-exit based self-speculative decoding (EESD) has emerged to mitigate this cost.<n>We propose Pipeline-Parallel Self-Speculative Decoding (PPSD) that fully pipelines the draft and verification work.
  arXiv  Detail & Related papers  (2025-09-19T04:51:41Z)
• SlimInfer: Accelerating Long-Context LLM Inference via Dynamic Token Pruning [3.502168555273189]
  SlimInfer aims to accelerate inference by directly pruning less critical prompt tokens during the forward pass.<n>We show that SlimInfer can achieve up to $mathbf2.53times$ time-to-first-token (TTFT) speedup and $mathbf1.88times$ end-to-end latency reduction for LLaMA3.1-8B-Instruct.
  arXiv  Detail & Related papers  (2025-08-08T16:42:38Z)
• ELIS: Efficient LLM Iterative Scheduling System with Response Length Predictor [5.097511974401423]
  ELIS is a serving system for Large Language Models (LLMs) featuring an Iterative Shortest Remaining Time First (ISRTF) scheduler.<n>ISRTF scheduler efficiently manages inference tasks with the shortest remaining time.
  arXiv  Detail & Related papers  (2025-05-14T04:50:00Z)
• DEL: Context-Aware Dynamic Exit Layer for Efficient Self-Speculative Decoding [7.204881999658682]
  We introduce DEL, a plug-and-play method that adaptively selects the exit layer and speculation length during inference.<n>Del achieves overall speedups of $2.16times$$sim$$2.62times$ over vanilla auto-regressive decoding.
  arXiv  Detail & Related papers  (2025-04-08T01:12:59Z)
• Priority-Aware Preemptive Scheduling for Mixed-Priority Workloads in MoE Inference [4.7730970530715835]
  Large Language Models have revolutionized natural language processing, yet serving them efficiently in data centers remains challenging.<n>We introduce QLLM, a novel inference system designed for Mixture of Experts (MoE) models.<n>QLLM enables expert-level preemption, deferring BE job execution while minimizing LS time-to-first-token (TTFT)
  arXiv  Detail & Related papers  (2025-03-12T11:56:01Z)
• Smart Routing: Cost-Effective Multi-LLM Serving for Multi-Core AIOS [31.60019342381251]
  Existing scheduling frameworks mainly target at latency optimization.<n>This paper proposes an efficient capability-cost coordinated scheduling framework, ECCOS, for multi-LLM serving.
  arXiv  Detail & Related papers  (2025-02-27T22:35:31Z)
• MixLLM: Dynamic Routing in Mixed Large Language Models [57.309520357563215]
  Large Language Models (LLMs) exhibit potential artificial generic intelligence recently, however, their usage is costly with high response latency.<n>We develop MixLLM, a dynamic contextual-bandit-based routing system for query-LLM assignment.
  arXiv  Detail & Related papers  (2025-02-09T02:26:15Z)
• Leveraging Online Olympiad-Level Math Problems for LLMs Training and Contamination-Resistant Evaluation [55.21013307734612]
  AoPS-Instruct is a dataset of more than 600,000 high-quality QA pairs.<n>LiveAoPSBench is an evolving evaluation set with timestamps, derived from the latest forum data.<n>Our work presents a scalable approach to creating and maintaining large-scale, high-quality datasets for advanced math reasoning.
  arXiv  Detail & Related papers  (2025-01-24T06:39:38Z)
• An Early FIRST Reproduction and Improvements to Single-Token Decoding for Fast Listwise Reranking [50.81324768683995]
  FIRST is a novel approach that integrates a learning-to-rank objective and leveraging the logits of only the first generated token. We extend the evaluation of FIRST to the TREC Deep Learning datasets (DL19-22), validating its robustness across diverse domains. Our experiments confirm that fast reranking with single-token logits does not compromise out-of-domain reranking quality.
  arXiv  Detail & Related papers  (2024-11-08T12:08:17Z)
• SuffixDecoding: Extreme Speculative Decoding for Emerging AI Applications [13.948608558319307]
  Speculative decoding is widely adopted to reduce latency in large language model (LLM) inference.<n>Agentic frameworks submit repetitive inference requests, which result in long and highly predictable computation.<n>We introduce emphSuffixDecoding, a novel method that utilizes efficient suffix trees to cache long token sequences.
  arXiv  Detail & Related papers  (2024-11-07T18:49:33Z)
• ALISE: Accelerating Large Language Model Serving with Speculative Scheduling [7.367068885621016]
  Large Language Models (LLMs) represent a revolutionary advancement in the contemporary landscape of artificial general intelligence (AGI) In this paper, we propose a new efficient LLM inference serving framework, named ALISE. We show that ALISE improves the throughput of inference serving by up to 1.8x and 2.1x under the same latency constraint on the Alpaca and ShareGPT datasets, respectively.
  arXiv  Detail & Related papers  (2024-10-31T00:58:11Z)
• Fast Inference for Augmented Large Language Models [14.195265302357148]
  Augmented Large Language Models (LLMs) enhance the capabilities of standalone LLMs by integrating external data sources through API calls. Traditional size-based scheduling algorithms, such as Shortest Job First (SJF), become less effective at minimizing completion times. We propose LAMPS, a novel LLM inference framework for augmented LLMs.
  arXiv  Detail & Related papers  (2024-10-23T19:53:30Z)
• Don't Stop Me Now: Embedding Based Scheduling for LLMs [22.099820814682513]
  Size-based scheduling algorithms like Shortest Remaining Process Time (SRPT) aim to reduce average request completion time. We propose a prediction-based SRPT variant with limited preemption designed to account for memory overhead in LLM systems.
  arXiv  Detail & Related papers  (2024-10-01T19:51:07Z)
• SELF-GUIDE: Better Task-Specific Instruction Following via Self-Synthetic Finetuning [70.21358720599821]
  Large language models (LLMs) hold the promise of solving diverse tasks when provided with appropriate natural language prompts. We propose SELF-GUIDE, a multi-stage mechanism in which we synthesize task-specific input-output pairs from the student LLM. We report an absolute improvement of approximately 15% for classification tasks and 18% for generation tasks in the benchmark's metrics.
  arXiv  Detail & Related papers  (2024-07-16T04:41:58Z)
• DAISY: Data Adaptive Self-Supervised Early Exit for Speech Representation Models [55.608981341747246]
  We introduce Data Adaptive Self-Supervised Early Exit (DAISY), an approach that decides when to exit based on the self-supervised loss. Our analysis on the adaptivity of DAISY shows that the model exits early (using fewer layers) on clean data while exits late (using more layers) on noisy data.
  arXiv  Detail & Related papers  (2024-06-08T12:58:13Z)
• SMART: Automatically Scaling Down Language Models with Accuracy Guarantees for Reduced Processing Fees [21.801053526411415]
  Large Language Models (LLMs) have significantly boosted performance in natural language processing (NLP) tasks. The deployment of high-performance LLMs incurs substantial costs, primarily due to the increased number of parameters aimed at enhancing model performance. We introduce SMART, a novel framework designed to minimize the inference costs of NLP tasks while ensuring sufficient result quality.
  arXiv  Detail & Related papers  (2024-03-11T17:45:47Z)
• HiRE: High Recall Approximate Top-$k$ Estimation for Efficient LLM Inference [68.59839755875252]
  HiRE comprises of two novel components: (i) a compression scheme to cheaply predict top-$k$ rows/columns with high recall, followed by full computation restricted to the predicted subset, and (ii) DA-TOP-$k$: an efficient multi-device approximate top-$k$ operator. We demonstrate that on a one billion parameter model, HiRE applied to both the softmax as well as feedforward layers, achieves almost matching pretraining and downstream accuracy, and speeds up inference latency by $1.47times$ on a single TPUv5e device.
  arXiv  Detail & Related papers  (2024-02-14T18:04:36Z)
• Cheaply Evaluating Inference Efficiency Metrics for Autoregressive Transformer APIs [66.30706841821123]
  Large language models (LLMs) power many state-of-the-art systems in natural language processing. LLMs are extremely computationally expensive, even at inference time. We propose a new metric for comparing inference efficiency across models.
  arXiv  Detail & Related papers  (2023-05-03T21:51:42Z)
• Confident Adaptive Language Modeling [95.45272377648773]
  CALM is a framework for dynamically allocating different amounts of compute per input and generation timestep. We demonstrate the efficacy of our framework in reducing compute -- potential speedup of up to $times 3$ -- while provably maintaining high performance.
  arXiv  Detail & Related papers  (2022-07-14T17:00:19Z)
• Token Dropping for Efficient BERT Pretraining [33.63507016806947]
  We develop a simple but effective "token dropping" method to accelerate the pretraining of transformer models. We leverage the already built-in masked language modeling (MLM) loss to identify unimportant tokens with practically no computational overhead. This simple approach reduces the pretraining cost of BERT by 25% while achieving similar overall fine-tuning performance on standard downstream tasks.
  arXiv  Detail & Related papers  (2022-03-24T17:50:46Z)
• Magic Pyramid: Accelerating Inference with Early Exiting and Token Pruning [19.93342734884434]
  We propose a novel idea, Magic Pyramid (MP), to reduce both width-wise and depth-wise computation via token pruning and early exiting for Transformer-based models. MP is capable of achieving an average of 8.06x speedup on two popular text classification tasks, regardless of the sizes of the inputs.
  arXiv  Detail & Related papers  (2021-10-30T11:07:43Z)
• Accelerating BERT Inference for Sequence Labeling via Early-Exit [65.7292767360083]
  We extend the recent successful early-exit mechanism to accelerate the inference of PTMs for sequence labeling tasks. We also propose a token-level early-exit mechanism that allows partial tokens to exit early at different layers. Our approach can save up to 66%-75% inference cost with minimal performance degradation.
  arXiv  Detail & Related papers  (2021-05-28T14:39:26Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.