LLM Inference Efficiency

Large language models (LLMs) have demonstrated strong capabilities in handling long-context tasks, but processing such long contexts remains challenging due to the substantial memory requirements and inference latency. In this work, we discover that certain attention heads exhibit sequential consistency in their attention patterns, which can be persistently identified using a coefficient-of-variation-based algorithm. Inspired by this observation, we propose CateKV, a hybrid KV cache method that retains only critical token information for consistent heads, thereby reducing KV cache size and computational overhead, while preserving the majority of KV pairs in adaptive heads to ensure high accuracy. We show the unique characteristics of our algorithm and its extension with existing acceleration methods. Comprehensive evaluations on long-context benchmarks show that, while maintaining accuracy comparable to full attention, CateKV reduces memory usage by up to $2.72\times$ and accelerates decoding by $2.18\times$ in single-sample inputs, and boosts throughput by $3.96\times$ in batch scenarios.

Long-context LLM applications such as document summarization and multi-turn agents require generation from prefixes spanning tens of thousands of tokens, making decoding latency a major bottleneck. Speculative decoding (SD) reduces latency without changing model outputs, but its speedup depends on both accepted draft tokens and draft-step latency: Lightweight drafts are fast but lack the capacity to capture long-range dependencies, whereas strong independent drafts recover acceptance but incur growing KV-access cost at long prefixes. We introduce memory-augmented drafting for long-context SD, equipping a strong independent draft with compressed draft-side KV memory: A lightweight adaptor constructs and incrementally updates this memory to retain distant information and exact recent context. The target verifier retains its full KV cache and applies the standard accept/reject rule, preserving SD's lossless guarantee. Experiments on Llama~3.1-8B and 70B targets at prefix lengths up to 32K show that our method reduces draft-side memory by over 70%. It achieves speedups of up to 2.08x and 3.33x , respectively, over autoregressive decoding.

Speculative decoding accelerates large language model inference by using a draft model to generate candidate tokens, which are verified by the target model in a single forward pass. Verification proceeds sequentially and discards every position from the first rejection onward, yet existing draft training relies on token-level imitation of the target with a fixed per-position weighting that reflects neither property. We introduce Verification-Aware Training (VAT), a plug-in framework that simulates verification at every training step and turns the resulting accept and reject patterns into supervision. VAT consists of two components: (i) a verification head, a lightweight jointly trained binary classifier that supervises the draft model on whether each position survives sequential verification; (ii) verification-adaptive weighting, which replaces the fixed weighting schedule by keeping full weight up to each sample's first rejection point and re-anchoring the decay to start there. VAT modifies only the training objective, so it can be layered on top of existing methods without changing the draft architecture, the target model, or the inference procedure. Applied to EAGLE-3 and DFlash on Qwen3-4B, Qwen3-8B, and LLaMA-3.1-8B, VAT improves average acceptance length by up to 11.4% and wall-clock speedup by up to 8.7%, with consistent gains across math, code, and chat benchmarks. Code will be available at https://github.com/naver-ai/vat

Single-GPU deployment of 70B-parameter language models on an NVIDIA GPU is constrained by device memory, long-context throughput, and engineering integration cost. We cast single-GPU inference as a budget-aware design problem over these three axes and study how pruning, quantization, and KV-cache compression interact under realistic execution. Controlled ablations show that layer-wise pruning makes weight quantization more robust. KV-cache sparsification complements INT8 KV quantization by reducing memory without hurting decoding speed, while static vector quantizers often conflict with dynamic caching. Guided by these coupling results and explicit budget tracking, we assembled a practical pipeline and compressed a 70B model to about 33 GB, sustained about 57 tokens/s on 10k token prompts on a single A40, and kept absolute accuracy within 5% on common and reasoning benchmarks. We contribute design rules and a reproducible evaluation protocol that jointly report quality, memory, and end-to-end speed, and we provide a foundation for automated pipeline search under realistic single-GPU constraints.

KV-cache compression reduces LLM inference memory by evicting context tokens, but when the evicted tokens contain answer-bearing evidence, the model may hallucinate instead of recognizing that the compressed context is insufficient. We address this failure from a behavioral perspective: to our knowledge, this is the first work to formulate compression-aware abstention as a learning problem, in which a model learns to answer when supporting evidence survives compression and abstain when it does not. We construct supervision from compressor survival masks and tight answer-bearing spans, labeling examples as Confident when evidence survives and Abstain when it is removed. A 10.1M-parameter LoRA adapter trained on ~2.6K MuSiQue 2-hop QA examples reduces base-model hallucinations by 97% under prompt-style truncation while preserving correct answering on evidence-retaining examples. Unlike prompt-only abstention baselines, which over-abstain on many answerable high-retention examples, the trained adapter learns a conditional policy. We also evaluate the method under actual compressed-cache decoding, where multi-compressor training yields a 6-22x relative lift over the unaided base on evidence-retaining examples. Controlled-deletion experiments show that the learned behavior is driven by evidence content rather than input length alone.

Speculative decoding accelerates autoregressive language model inference by having a lightweight draft model propose multiple candidate tokens, which are then verified in parallel by a larger target model. However, after the first rejection, standard prefix-based verification discards the remaining draft suffix, so the computation spent generating and verifying those positions does not contribute to decoding progress. Focusing on DFlash, we show that rejected positions in a rejected suffix may still align with the target continuation, indicating that the draft model can retain useful semantic and structural information despite local token-level errors. Motivated by this observation and inspired by conditional diffusion, we introduce~\textbf{ReTrace}, a rejected-trajectory conditioning method that conditions each draft block on the rejected suffix from the previous round rather than generating it from fresh mask placeholders alone. ReTrace retains the hidden representations of the rejected suffixes, aligns them with the next draft block, refines them using target-aware correction signals from the same verification pass, and admits them into the drafter's input embeddings through gated residual fusion. Because rejected tokens are never committed and target-side verification remains unchanged, ReTrace preserves the lossless property of speculative decoding without requiring an additional model forward pass. Experiments with Qwen3 models across mathematical reasoning, code generation, and open-ended dialogue demonstrate that ReTrace consistently improves average acceptance length and end-to-end decoding speed over its DFlash backbone. By introducing cross-round conditioning without modifying within-round proposal generation, ReTrace is largely orthogonal to existing drafting improvements and might be combined with them for further gains.

Quantization has been widely adopted in LLM training and inference to reduce cost and improve efficiency. However, low-bit quantization of the \emph{attention} module often introduces large errors at very low bit-widths, causing performance degradation. Existing methods mainly rely on smoothing techniques to handle outliers, while we propose a hybrid quantization design to better balance accuracy and efficiency. Specifically, we propose \textbf{HyQuant}, an efficient hybrid quantization framework for LLM attention. HyQuant quantizes most attention states into low-bit formats while retaining a small set of vertical-line tokens and local-window states in high precision. These accuracy-critical regions are selected using lightweight vertical-line-aware attention-pattern signals, reducing quantization error with limited overhead. In the Prefill stage, HyQuant uses a hybrid-precision quantized attention operator that preserves vertical-line tokens and a local sliding window in full precision while quantizing the remaining context. In the Decode stage, HyQuant applies the same principle to KV-cache compression and fuses KV dequantization with attention computation to improve memory and hardware efficiency. Across diverse tasks, models, and datasets, HyQuant maintains nearly lossless accuracy with an extremely simple design, demonstrating the efficiency and practical feasibility of hybrid quantization for LLM attention. Code is available at: https://github.com/jerrysfls/HyQuant .

Long-context inference is bottlenecked by the memory footprint of the key-value (KV) cache, especially for small models under tight resource budgets. Existing KV cache eviction methods score tokens using the model's attention distribution or, in attention-free variants, each key's distance from a global reference point. Using a controlled leave-one-out probe, we find that attention magnitude is unrelated to a token's causal contribution to the answer (Spearman $ρ=-0.004$), challenging the premise behind dominant eviction methods. We introduce TwinKV, a training-free, attention-free redundancy signal that detects whether a token's key has a near-duplicate elsewhere in context. Rather than replacing existing policies, TwinKV acts as a composable repair pass: given a policy's fixed retained set, it identifies evicted tokens with no surviving duplicate (\emph{orphans}) and retained tokens whose information is duplicated elsewhere (\emph{redundant donors}), then swaps them while preserving the original budget and scoring rule. We compose TwinKV with four recent eviction policies across LongBench, LooGLE, RULER, and a short-context MMLU-Pro no-harm control at compression ratios ${0.3,0.5,0.7}$. On Qwen3-4B, TwinKV improves a majority of configurations for two policies, is near-even for a third, and helps only a minority for a fourth adaptive baseline already near a performance ceiling; gains across the three non-ceiling policies are smallest at the loosest ratio. On RULER with Llama-3.2-1B, however, that fourth policy improves in every evaluated cell because its Alone score leaves substantial room to improve. More broadly, Llama-3.2-1B shows a smaller average LongBench gain but a higher fraction of improved cells on LongBench and LooGLE than Qwen3-4B, plus a clean RULER win. We also identify few-shot classification exemplars as a task structure where TwinKV does not help on either model.

Key-value (KV) caching is essential for efficient autoregressive large language model (LLM) inference, but the cache grows linearly with context length, increasing storage and decoding costs. KV cache compression mitigates this cost by retaining only a subset of cached tokens. This challenge is particularly important for multi-step LLM agents, where a query expands into trajectories of reasoning, tool interactions, and retrieved observations. Existing pruning methods typically treat the cache as a flat token stream and rank tokens by recency or attention saliency. This creates a mismatch between the unit of compression and the unit of reasoning: token-level pruning removes individual entries, whereas useful information in multi-step agents is often organized into reasoning steps with uneven and delayed importance. Consequently, an early observation or intermediate decision may receive little recent attention yet remain essential for later evidence synthesis. We term this failure mode Reasoning Continuity Disruption.These observations motivate KV cache compression that jointly considers token- and reasoning-step-level information. StepKV addresses this goal by treating reasoning steps as first-class retention units. It associates cache entries with their generating steps, estimates step utility from trajectory-derived signals, and combines this utility with token-level saliency. The resulting scores globally rank prunable tokens, from which StepKV retains the top-scoring entries under a target budget. StepKV thus provides a step-centric perspective for agent KV cache compression. Across multi-hop QA and long-horizon web reasoning tasks, StepKV sustains accuracy under low KV budgets where token-level baselines degrade sharply, offering a more robust efficiency-accuracy trade-off for multi-step agent inference.

Recent years have witnessed remarkable achievements of Large Language Models (LLMs) in multiple domains, while the excessive resource requirements of LLMs hinder the deployment on resource-constrained devices. Although model quantization stands out as an effective approach, conventional quantization approaches typically incur severe performance degradation due to uniform bit-width or simple heuristic sensitivity evaluation. In this paper, we propose a novel Fisher information-based Adaptive Mixed Precision Weight Quantization approach, i.e., FAMPWQ, which performs layer-adaptive weight quantization for effective LLM inference on commodity GPUs. First, we propose a system model with a novel Fisher information metric to measure the layer-wise sensitivity to quantization. Second, we propose a reinforcement learning-based bit-width allocator in FAMPWQ, which generates an adaptive bit-width allocation strategy based on the Fisher information sensitivity metric. Extensive experiments on 7 models and 5 benchmarks demonstrate that FAMPWQ significantly outperforms 7 baseline approaches in terms of PPL (up to 3.39 smaller), accuracy (up to 6.87% higher), and LLM-as-a-judge comparison (up to 76% win rate).

Speculative decoding accelerates language-model inference by drafting future tokens the target model verifies in parallel. A diffusion-style drafter such as DFlash drafts an entire block in one forward pass. It is trained on the per-position marginals rather than on the joint distribution over the block, so the tokens it emits are individually plausible yet jointly incoherent. We introduce LiLiCorr, a Lightweight Likelihood-based model that Correlates the per-position marginals such a drafter produces. It keeps the top-K tokens at each position and processes them jointly, emitting an in and an out vector for each. Two candidates at consecutive positions match when the earlier out vector aligns, in cosine similarity, with the later in vector. Training scores the correct pairings highest and pushes competing ones down, so coherent blocks outscore incoherent ones. The joint distribution over the block, exponential in its length, is never materialized. One lightweight network pass produces all the vectors, the pairwise scores follow as batched matrix operations, leaving only a cheap greedy walk sequential. We co-train the DFlash drafter with LiLiCorr, so it proposes candidates that correlate into longer accepted sequences. Over the vanilla DFlash drafter it builds on, LiLiCorr accepts more and serves faster at all 72 settings we test: nine benchmarks at two target sizes under greedy and temperature-one decoding, plus a throughput sweep over six concurrencies, two input lengths and three output-entropy tiers. It raises acceptance length by 7 to 19%, while its single-pass scoring head costs only about 3% of the per-block latency. Against three concurrently developed methods that also restore coherence at draft time, all equally optimized on a common stack, LiLiCorr holds the highest throughput in 63 of those settings, ties within a measured noise floor in 6, and trails in only 3.

Transformer-based language models achieve strong performance but incur substantial inference cost due to repeated high-dimensional matrix multiplications. We propose Reduced Matrix Multiplication (RMM), a training-free, input-adaptive inference method that reduces Transformer matrix products by selecting informative slices along their contraction dimensions, without modifying model weights. Under a simple retention-ratio control, RMM provides a smooth and predictable accuracy-efficiency trade-off. Across language models ranging from 1B to 70B parameters, we find that reduction tolerance depends on the model family, task, component, and retention ratio, although it often improves with model scale. Under moderate reduction, RMM remains robust across the evaluated discriminative, autoregressive generation, and long-context settings. We further show that the same principle extends to multimodal vision-language inference. Mechanistic ablations reveal a structural asymmetry within Transformers: attention-side computations are substantially more reducible than MLP components. Finally, wall-clock benchmarks with custom kernels on an NVIDIA A100 show that these computational savings can translate into practical runtime gains, especially at longer sequence lengths. Together, these results position RMM as a scalable direction for input-adaptive inference-time optimization.

Large Language Models (LLMs) have achieved remarkable success in natural language understanding and generation, but their deployment is constrained by high computational demands. Deploying smaller LLMs directly on the edge can circumvent this, but with degraded accuracy. Deploying smaller cloud-based big LLMs preserves performance, but at the cost of expensive per-token computation. We present a distributed inference framework, \our{}, that integrates speculative decoding (SD) across edge and cloud. A compact draft model deployed on the edge generates candidate tokens rapidly, and a large verifier model on the cloud validates these tokens in parallel. Accepted tokens are retained, while only rejections trigger verifier correction, substantially reducing the number of cloud queries. Our plug-and-play design shifts the bulk of computation to the edge, significantly lowers inference time and cloud cost, and preserves the accuracy of the big model without any retraining requirement. Our approach demonstrates a practical path toward scalable, cost-efficient, and accurate deployment of LLMs in real-world environments. Experimental results across multiple Natural Language Processing tasks using SpecBench and CNN/Dailymail datasets demonstrate that \our{} reduces the cloud model calls by $76\%$ with zero loss in accuracy as compared to the full model.

The rapid growth of LLM inference is shifting sustainability concerns from one-time training to continuous serving, where infrastructure decisions shape energy use, carbon emissions, water consumption, and service quality. Yet operators often need to compare deployment alternatives before large-scale infrastructure is built, making direct measurement costly, slow, and sometimes infeasible. We present InFactPlanner, a trace-driven decision-support framework for what-if analysis of sustainable AI data center deployment for LLM inference across single and geo-distributed sites. InFactPlanner combines query traces, hardware-model profiles, candidate site configurations, PUE/WUE parameters, renewable generation models, and time-varying grid carbon intensity to estimate power, energy, carbon emissions, water use, latency, and server utilization. The framework abstracts low-level serving effects into configurable hardware-model profiles, enabling rapid comparison of site selection, capacity placement, hardware, model, renewable integration, and routing choices. We validate the energy accounting pipeline by reproducing reference LLM inference energy estimates with less than 10% deviation, evaluate scalability across multiple data centers and server counts, and demonstrate scenario-driven decision analyses for hardware selection, renewable placement, geographic deployment, and carbon-aware routing. Our results show that sustainability-optimal choices can differ from latency-optimal ones, and that the carbon value of deployment depends strongly on the local grid mix.

Should you replace your text-embedding pipeline with a large language model? We answer this with a controlled, cost-aware comparison of ten LLMs across six families and 26 embedding models (118M to 14B parameters) on 37 tasks spanning classification, semantic textual similarity (STS), clustering, pair classification, and retrieval. In aggregate the two paradigms are effectively tied: the best LLM (Gemini 3.1 Pro, 77.6) and the best embedding model (77.2) differ by 0.4 points. Their strengths differ by task: LLMs lead on reasoning-heavy retrieval, embedding models lead on classification, and the two match on clustering, STS, and pair classification. Reaching that parity is expensive. An LLM costs up to 1,431x more than an embedding model of comparable quality (USD 154 vs. USD 0.11 per benchmark pass), and the open LLMs tested process tokens 2.5 to 736x more slowly on the same GPU. Reasoning tokens account for 28 to 81% of LLM inference cost; lower reasoning budgets preserve or improve retrieval quality for most models in our ablation. The Pareto frontier contains the leading embedding models and one LLM, Gemini 3.1 Pro. These results support a division of labour: use embedding models for similarity, classification, and clustering, and reserve LLMs for reasoning-intensive retrieval. Our code, datasets, and results are publicly available at https://github.com/embeddings-benchmark/embedders-dilemma.

Few-shot in-context learning (ICL) with multi-modal large language models (MLLMs) enables task adaptation without parameter updates, but its performance is highly sensitive to the quality and coverage of the selected demonstrations. While unlabeled multi-modal data is abundant, it remains elusive how to exploit them for ICL. We propose MAG (MAnifold-Guided semi-supervised in-context demonstra- tion selection), an efficient framework that leverages unlabeled data to improve multi-modal ICL. MAG formulates demonstration selection as a semi-supervised propagation problem on a multi-modal graph and adopts a two-stage strategy: (i) relevance score propagation identifies a compact set of high-impact unlabeled samples for pseudo-labeling, reducing MLLM inference cost; (ii) multi-modal relevance is used to select the final demonstrations. We show that textual represen- tations are more effective for relevance propagation, while both visual and textual modalities are crucial for high-quality demonstration selection. Experiments on eight multi-modal benchmarks demonstrate that MAG consistently outperforms strong baselines in label-scarce regimes, achieving significant gains with a limited pseudo-labeling budget.

ReRound (Reconstructive Rounding) is a post-training quantization method that addresses the midpoint ambiguity inherent in standard round-to-nearest (RTN) schemes when quantizing weights near the centers of quantization intervals. Starting from a pretrained LLM, ReRound trains a conditional diffusion model to produce continuous reconstructions of low-bit weights for the LLM. These reconstructed weights act as a guidance signal to disambiguate the rounding direction of weights located close to interval midpoints. To integrate this reconstruction-guided rounding with conventional RTN, ReRound introduces a tolerance metric measuring how far the quantized weight (not the final quantized integer) is away from the midpoint: quantized weights within a tolerance region around midpoints are quantized using diffusion-based reconstructions, whereas weights closer to quantization boundaries are quantized with RTN. By sweeping the tolerance parameter, ReRound generates multiple candidate quantized integer weight matrices and selects the de-quantized weight matrix candidate whose leading singular values most closely match those of the original full-precision weights. This selected candidate determines the tolerance parameter ReRound uses. ReRound is particularly effective for smaller LLMs. Across a range of such models, it consistently outperforms standard RTN for 3-bit and 4-bit weight quantization. ReRound achieves superior accuracy compared to an extensive set of calibration-free methods, remains competitive with calibration-dependent approaches, and operates entirely offline, introducing no additional overhead during low-bit inference. The ReRound strategy represents a new approach for low-bit quantization. The method applies to AI models beyond LLMs. This paper focuses on its applications to small LLMs.

Speculative decoding accelerates autoregressive large language model (LLM) inference by using a lightweight draft model to speculate multiple tokens, reducing expensive target model decoding steps. Its effectiveness depends heavily on draft selection, motivating adaptive methods that exploit variation across inputs and generation stages. On memory-constrained edge devices, however, these methods often fail to improve end-to-end throughput due to the overhead of switching between draft models. We identify a key limitation in this setting: the mismatch between draft selection and draft availability under tight memory budgets. To address this challenge, we present MemSpec, a prediction-guided, memory-aware runtime for adaptive speculative decoding on edge devices. MemSpec decouples draft selection from execution through proactive resident working-set management. A lightweight predictor estimates draft effectiveness from prompt and generation context, while a memory-aware scheduler reduces reactive model loading overhead. Experiments on a Jetson Orin Nano show that MemSpec improves steady-state generation throughput by 40.7% on average over state-of-the-art bandit-based adaptive methods while closely approaching the oracle upper bound.

A main promise of looped language models (LMs) is depth-adaptive inference. By iterating a block of shared layers a variable number of times, the model can use less compute for "easy" tokens and more for "hard" ones. However, this adaptivity breaks standard batching: tokens in the same batch now require a different number of loops, so there is no unified forward pass, making efficient inference difficult. Standard inference frameworks like vLLM schedule on the token level and cannot handle this because tokens need to be removed from the batch within the forward pass. Loop-level scheduling has been proposed as a solution, but never implemented end to end. The key challenge is that looped architectures also contain non-looped boundary stages (e.g., token embedding and LM head) that must be scheduled at different frequencies than the loop. We introduce continuous depth batching (CDB), which schedules at the granularity of individual loop iterations. CDB handles boundary stages and loop steps in separate priority queues, makes exit decisions one step ahead, and overlaps all scheduling work with GPU computation. On Ouro 1.4B and Huginn 3.5B, CDB can realize up to $99\%$ of the theoretical maximum speed-up from adaptive-depth, translating to $1.5$-$1.9\times$ higher offline throughput and $45$-$90\%$ lower normalized latency under dynamic serving load.

The key-value (KV) cache is the primary throughput optimization in modern large language model (LLM) inference, enabling prefix reuse across requests. In multi-tenant deployments this cache is shared across tenants, creating a timing side channel: an adversarial tenant can reconstruct another tenant's private prompt by probing cache-hit latency. Three published attacks exploit it -- PROMPTPEEK, EarlyBird and InputSnatch -- reaching up to 100% attack success rate against unprotected vLLM and SGLang, with rates varying by cache architecture and prompt structure. We present KVGov, a governance layer addressing all three attack families' prefix-cache paths under one mechanism. A per-principal salt sigma_p = HMAC_K(secret, principal_id) seeds the block-hash chain, making cache keys cryptographically disjoint across principals. An ablation (N=1000 trials, seed 2026, deterministic judges) isolates this salt as the necessary and sufficient component. KVGov adds ORIGAMI, a Stackelberg water-filling audit scheduler that reduces adversary expected utility by 12.6% at realistic tenant heterogeneity (Gini 0.63), and an evolutionary stability analysis giving a 31.6% adversary-prevalence tipping point below which global caching remains stable. On real hardware (Qwen2.5-7B-Instruct, vLLM 0.26.0, NVIDIA A100) we measure a gate-verified cold/cached TTFT ratio of 0.22, confirming the channel is exploitable at production scale; the defense itself is evaluated in simulation calibrated to those measurements. We replicate the channel on an independent stack (llama.cpp on Apple Metal, ratio 0.093). Finally, isolation and cache efficiency need not conflict: identifying information resides only where prompts diverge, so injecting the salt at that boundary rather than the chain root retains an estimated 93% of the prefix-cache benefit with no cross-principal signal.

Recent advances in retrieval-augmented generation (RAG) and large language models (LLMs) enable researchers to integrate AI into scientific workflows. However, using proprietary commercial AI systems raises concerns about transparency, reproducibility and privacy, which are essential for scientific practices. To this end, AquiLLM was developed as an open-source modular RAG-LLM framework using open-weight models, designed to support research groups in capturing tacit knowledge. In this work, we present a series of architectural improvements and feature enhancements to AquiLLM, including local embedding and reranking, multimodal capabilities, OpenAI-compatible inference interfaces, user interface improvements, semantic and episodic memory capabilities, and skills support. These enhancements were informed by discussions with domain experts, including astrophysicists and environmental researchers, and represent a step toward AI systems more closely aligned with scientific research practices.

Transformer-based large language models (LLMs) achieve strong performance across many tasks, but their Key-Value (KV) cache grows linearly with sequence length, creating a severe memory bottleneck for long-context inference. Existing heuristic eviction methods (e.g., H$_2$O and SnapKV) rely on static attention or positional signals that often fail to capture a token's future predictive influence. We propose DistillCache, a reinforcement learning framework that formulates KV-cache eviction as a sequential decision problem. DistillCache learns a lightweight policy network using rich internal model signals (attention statistics, value norms, entropy, and position) and trains it with REINFORCE via a per-step KL-divergence reward to preserve the full-cache output distribution. On a 7B-parameter instruction-tuned Transformer (Mistral-7B-Instruct-v0.3), DistillCache retains 94.2% of full-cache accuracy on LongBench at a 25% cache budget, outperforming both strong heuristic baselines (H$_2$O, SnapKV) by up to 2.7 absolute points and, under our re-implementations, concurrent RL-based methods (ForesightKV, RLKV) by up to 1.4 points on long-context tasks. On reasoning benchmarks, DistillCache is competitive with the best concurrent method and surpasses it under aggressive compression. It also delivers up to 2.1x full-cache throughput while maintaining competitive practical efficiency. These results highlight the effectiveness of learned, distribution-aware policies for memory-efficient long-context LLM inference.

The carbon footprint of any deployed Large Language Model (LLM) accumulates during inference, where repeated use of the model substantially exceeds the one-time cost of fine-tuning. Yet most efficiency interventions target either pre-training scale or post-hoc compression. We ask whether folding a calibrated, differentiable energy surrogate into the fine-tuning objective can produce inference behavior that gains task accuracy at zero or near-zero carbon cost, a break-even configuration. We propose a joint loss mechanism with a per-model carbon-emission parameter, a linear surrogate over parameter norm, FLOP proxy, and a memory proxy, fit from on-hardware energy profiling. We fine-tune three architecturally distinct families: Gemma-2 2B, Llama-3.1 8B, and Qwen-2.5 14B, and evaluate inference F1 and CO$_2$ emissions on three MMLU subjects: abstract algebra, philosophy, and formal logic. We discover from several outcomes that the carbon term behaves as either harmful interference or beneficial regularization depending on the task structure. We position calibrated carbon-aware fine-tuning as a lightweight, drop-in regularizer with a non-empty but model and task-dependent break-even region. This is an ongoing work, and we will release our codebase soon.

Large Language Models are deployed to multiple types of environments, from internet browsers to edge devices, and WebGPU serves as a modern cross-platform standard. The engines for browser-based LLM inference have proliferated, yet the overhead of WebGPU per-operation dispatch remains poorly characterized. In this work, we introduce a sequential-dispatch measurement method and show that naive single-operation measurements overestimate per-dispatch cost by conflating dispatch with synchronization. Using our method, we measure the per-dispatch cost and show that it is independent of data type used. We show that the dispatch overhead, not kernel quality, is the bottleneck at batch size 1, and isolate the dispatch count as the cause. Therefore, we conclude that at batch size 1, the effective approach to LLM inference optimization in WebGPU is reducing dispatch count. Our findings point to dispatch amortization, in the inference engines and in the WebGPU specification, as a path to practical browser-based inference.

Speculative decoding accelerates large language model inference by drafting multiple tokens for parallel verification, with efficiency critically determined by the speculative length selected at each decoding round. Existing dynamic speculation methods select the speculation length by estimating how many tokens will be accepted, which is reasonable for autoregressive drafters that generates tokens sequentially. The recent wave of diffusion-based drafters, however, generates candidate blocks in parallel at substantially lower drafting cost, shifting the key question from how many tokens to generate to how many generated tokens are worth verifying. We therefore reformulate dynamic speculative-length selection as expected-speedup optimization and derive a marginal criterion that extends the speculative sequence only when its acceptance gain outweighs the additional verification cost. Building on this criterion, we develop \textit{LibraSpec}, a training-free and plug-and-play algorithm that iteratively determines the speculative length using drafter confidence scores. Theoretically, we prove that LibraSpec monotonically converges toward the optimal speculative length. Experiments across six target models, three diffusion-based speculative decoding methods, and math, coding, and chat benchmarks show consistent improvements under both greedy and sampling settings, achieving a further $0.5\sim1.5\times$ improvement over baselines and up to $8.49\times$ speedup over autoregressive decoding.

Long-context LLM inference is bottlenecked by KV cache memory, yet distributing a limited cache budget across layers remains challenging. Existing methods rely on proxies such as layer depth, attention statistics, or representation change. These proxies do not measure how perturbations at each layer propagate to the output and may therefore cause sensitive layers to be underallocated while tolerant layers are overallocated. To address this issue, we propose RippleKV, which allocates cache across layers by estimating how perturbations to each layer's value cache affect the final predictive distribution. RippleKV independently injects norm-adaptive perturbations into each layer's value cache and measures the induced KL divergence at the model output over a small calibration set. Averaging these responses yields a sensitivity profile specific to the model that need not vary monotonically with depth. RippleKV then converts the sensitivity profile into layer budget multipliers by normalizing the sensitivity scores and applying an exponential mapping. A ratio parameter controls the allocation disparity between sensitive and tolerant layers, while a final normalization preserves the KV cache budget. Experiments on LongBench demonstrate that RippleKV achieves the highest average performance among the evaluated KV cache compression methods under matched cache budgets.

Multi-agent systems (MAS) built on Large Language Models (LLMs) are proliferating rapidly, but their heterogeneous execution traces provide no common basis for evaluation across methods. Outcome-only benchmarks discard collaborations, whereas LLM-as-Judge evaluation requires additional, model-dependent inference and can vary with the LLM and rubric. We introduce a generalizable evaluation framework that maps native MAS traces into a shared space of unified collaboration graphs, enabling different methods to be evaluated under the same representation, reference set, and metric panel. Candidate graphs are compared with a query-specific reference forest. Each forest is a benchmark-provided collection of verified-success graphs: it records diverse ways in which representative MAS methods can complete the task, rather than prescribing a unique optimal process. Instantiating the framework as ForestBench, we filter $844$ collaboration-necessary queries from seven public datasets, precompute ten successful target-conditioned reference graphs per query, and evaluate six representative MAS frameworks. Controlled backbone, reference-construction, and perturbation studies test the stability and scope of evaluation. Once the benchmark forests are built, ForestBench scores a trace in milliseconds without further LLM inference, providing a reusable structural basis for comparing diverse MAS collaboration traces.

As LLM inference shifts to multi-tenant GPU clusters, co-batching improves throughput but obscures per-tenant usage and limits control. Enabling fractional sharing of the inference engine requires a real-time, per-request attribution primitive that is accurate and light enough to run inside the scheduling loop. We present LLMVisor, a roofline-guided latency attribution model that captures the memory-bound and compute-bound phases via a concise piecewise-linear form over features proportional to FLOPs and memory I/O traffic. LLMVisor decomposes batch latency into additive, per-request shares and runs efficiently at microsecond scale. We evaluate LLMVisor across Llama 3.1-8B and Qwen 2.5-14B/32B on A100/H100 GPUs under varying tensor parallelism and workload mixes. Compared to a token-count baseline, LLMVisor attains near-perfect R-squared and reduces relative error by up to 2.5x and 3.3x at p90 and p99, respectively, for prefill, and by up to 3.5x and 4.4x for decode, despite batching variability and sequence divergence.

Long-context LLM inference is bottlenecked by quadratic attention computation and growing KV-cache costs. Existing sparse attention and KV-compression methods typically decide which tokens or heads to preserve from runtime attention scores, observation windows, calibration prompts, or learned gates, making head diagnosis input-dependent and costly to deploy. We propose Autonomy-of-Heads (AoH), a data-free method that identifies retrieval and streaming heads from the spectral geometry of query-key projections. AoH defines the kernel attention operator $M_h = W_K^{h\top}W_Q^h$ and uses its effective-rank as a weight-space measure of head function: concentrated spectra indicate a small number of dominant query-key matching directions and are associated with retrieval heads, whereas diffuse spectra indicate the absence of a dominant global matching direction and are associated with streaming heads. We further derive an efficient $d_\text{head}$-dimensional computation that avoids constructing the full $d_\text{model}\times d_\text{model}$ matrix. We conducted extensive experiments across models demonstrating that at 50% sparsity, AoH retains 96.5% of Full Attention performance on average while reducing prefill and decode latency by up to 41.4% and 66.0%, respectively, and KV-cache memory by 50.0% at 256K tokens.

Weight quantization for large-language-model inference must balance adaptive reconstruction levels with representations regular enough for efficient GPU execution. Uniform integers constrain each group to a linear grid. Low-bit floating-point formats use a fixed exponent-mantissa structure, while learned codebooks gain flexibility at the cost of irregular decoding and additional metadata. We introduce CubicQuant, a parametric non-uniform scalar format that preserves a dense integer code stream while adapting reconstruction levels within each weight group. A monotonic cubic curve, specified by two shape parameters and one scale, maps uniformly spaced magnitude codes to non-uniform levels. The family spans 1-8-bit weight payloads, contains symmetric uniform integer quantization as an exact special case, and has effective width B + 64/G bits per weight for payload width B and group size G. We derive population distortion under Uniform, Gaussian, and Laplace distributions, formulate continuous and Dynamic-A8-carrier-aware fitting objectives, and describe direct packed-weight GPU execution. For finite groups of G=128 with 15,360 samples per distribution, W4 CubicQuant reduced reconstruction RMSE relative to optimally clipped four-bit uniform integer quantization by 3.90% on Uniform, 13.49% on Gaussian, and 28.14% on Laplace samples. Relative to the best enumerated four-bit finite floating-point format, the reductions were 3.90%, 9.44%, and 6.27%. Preliminary H200 kernel measurements show a workload-dependent crossover: model-dtype execution is faster for narrow GEMV, while Dynamic A8 becomes favorable as row count grows. The results establish the format's representational promise and direct executability; downstream model quality and cross-device end-to-end performance remain open evaluation questions.

The rapid scaling of Large Language Models (LLMs) has significantly increased computational cost, energy consumption, and inference latency, making accurate estimation essential for sustainable artificial intelligence deployment and hardware-aware design. In this work, we introduce Hybrid Modeling for Energy and Latency of LLMs (HYMELL), a hybrid three-level framework for estimating LLM inference latency and energy by combining analytical modeling with machine learning (ML). HYMELL models LLM execution through a three-level hierarchy: analytical estimation of primitive operations, ML prediction of higher-level components, and an end-to-end model that captures system-level overheads across both prefill and decode phases. The framework supports diverse architectures, including dense and mixture-of-experts (MoE) feed-forward networks (FFNs), as well as multi-head attention (MHA) and grouped-query attention (GQA) mechanisms. Evaluated on an NVIDIA H100 graphics processing unit (GPU), HYMELL achieves high predictive accuracy; notably, for LLaMA 3 8B, it attains less than 5% error for both prefill and decode phases. By predicting execution costs directly from architectural parameters, it enables fast, hardware-free design space exploration and energy-efficient optimization.

Diffusion language models (dLLMs) offer a promising alternative to autoregressive models by accelerating inference through parallel decoding. Recent dLLMs commonly use blockwise semi-autoregressive decoding, generating blocks autoregressively while denoising tokens within each active block in parallel. However, despite KV caching, each denoising step still attends to all previous blocks, repeatedly incurring prefix-attention cost. Motivated by this bottleneck, we ask whether dLLM inference can be further accelerated by linearizing attention over previous blocks. We introduce block-hybrid attention, which retains exact softmax attention within the active denoising block while applying linear attention over previous blocks. We show that this hybrid attention can be retrofitted into a pretrained dLLM with minimal post-training: LLaDA-Hybrid replaces 6 of the 20 attention layers in LLaDA~2.1, a 16B open-source dLLM, largely following LoLCAT (Zhang et al, 2024). The conversion takes only approximately 60 hours while preserving benchmark performance: 72.0% vs. 75.6% on HumanEval, 63.0% vs. 57.7% on MBPP+, and 86.7% vs. 88.3% on CMATH. With a Triton implementation, LLaDA-Hybrid achieves up to $1.7\times$ higher decoding throughput and supports more concurrent requests before exhausting memory, showing that pretrained dLLMs can be efficiently linearized for faster inference. Our code is available at: https://github.com/Diuven/LLaDA-Hybrid.

The reasoning and agentic capabilities of large language models have expanded the range of applications they support, from short interactive exchanges to long, compute-heavy requests. LLM serving platforms today define response-latency service-level objectives, even though requests within the same service can differ by orders of magnitude in input length, generation length, execution cost, and the availability of reusable KV-cache state. As a result, requests governed by the same service level objective have different urgency: after accounting for the time required to execute them, some have substantial latency headroom while others have almost none. We define this headroom---the difference between a request's service level objective and its predicted remaining service time---as its per-request latency budget. We present Cascade, an LLM serving system that estimates and continuously updates this budget from request characteristics, KV-cache state, and current system load. Unlike prior SLO-aware schedulers that use deadlines to govern request ordering alone, Cascade uses a single per-request budget to jointly coordinate request scheduling and KV-cache management across the memory hierarchy. Its scheduler prioritizes requests with little remaining budget, while its memory manager uses the same budget to decide whether non-resident KV state should be restored or prefetched from a deeper tier, retained in HBM, or recomputed. By directing queueing and data-movement overhead toward requests that can absorb it, Cascade improves SLO-satisfied goodput while preserving fairness across heterogeneous request classes. On production traces across three large language models, Cascade improves goodput by up to2.4x and reduces SLO violations by 40% relative to the default vLLM first-come, first-served scheduler.

Large Language Models (LLMs) such as ChatGPT and Claude are widely used for information retrieval and problem-solving. Recent work has focused on improving scheduling algorithms to boost throughput while maintaining low latency. However, these approaches often assume Poisson request arrivals with constant rates - an assumption that fails to reflect the inherently bursty and dynamic nature of real-world traffic. We propose a lightweight extension to the state-of-the-art WAIT algorithm [1], which adapts to time-varying arrival rates without prior traffic knowledge. The proposed algorithm performs online estimation of request intensity based on observed interarrival times. Using Markov Modulated Poisson Process (MMPP)-based synthetic workloads with diverse request types, we conduct a simulation-based evaluation demonstrating that the proposed method achieves higher throughput than Sarathi-Serve [2], ORCA [3], and vLLM [4] in the evaluated low arrival-rate shift scenarios while maintaining comparable latency.

Edge inference is a promising paradigm to provide large language model (LLM) inference services in next-generation mobile networks. LLM inference mainly relies on two approaches: Autoregressive decoding (AD) generates output tokens sequentially, resulting in long latency; Speculative decoding (SD) accelerates inference by using a small language model (SLM) to generate multiple draft tokens for LLM verification, but incurs extra memory costs. Due to this latency-memory tradeoff, neither approach alone can efficiently serve users with heterogeneous demands under limited edge computing resources. To address this challenge, we propose a hybrid autoregressive-speculative inference (BALANCE) framework for edge LLM inference. In BALANCE, an edge server hosts both an SLM and an LLM, assigns each user to AD or SD, and performs the two modes simultaneously. To maximize the number of served users, we formulate a task throughput maximization problem to jointly determine user scheduling and computing resource allocation between AD and SD under user latency requirements and server memory constraints. Since the problem is NP-hard, we develop a polynomial-time algorithm that transforms the original problem into two sub-problems and obtains a sub-optimal solution with a constant approximation guarantee. Experiments demonstrate that BALANCE consistently outperforms conventional AD and SD and significantly improves task throughput.

Speculative decoding accelerates large language models' inference by using a lightweight drafter to propose multiple future tokens and a target model to verify them. While recent block and diffusion-style drafters can predict several positions in a single pass, their training and sampling procedures are typically optimized for greedy decoding or assume that positions in the draft block are conditionally independent. This assumption becomes brittle in non-greedy speculative decoding, where the target distribution is deliberately stochastic and multiple continuations become plausible. We study this mismatch for block diffusion drafters and show that the accepted draft length degrades as the entropy of the target sampling distribution increases. We propose a dependent block drafter based on a low-rank latent mixture over token positions, complemented by an acceptance-oriented training objective that directly targets the expected verified length. Experiments with Qwen3-4B and Qwen3-8B on GSM8K, MT-Bench, HumanEval, and creative-writing benchmarks show that our approach, namely DBLast, consistently improves accepted length over independent block sampling, especially in higher-entropy decoding regimes.

Autoregressive large language model inference is increasingly constrained by the memory footprint of the Key-Value (KV) cache. A dominant line of work reduces this footprint by evicting tokens that appear unimportant under attention-derived scores. However, such policies make an implicit irreversible decision: once a token is evicted, it cannot become useful again. We show that this assumption is brittle during decoding. Token and window importance drift as generated queries evolve, causing standard eviction policies to permanently discard states that later receive substantial attention under the full-cache model. To characterize this behaviour, we introduce Future Missed Mass and Global LIR, two diagnostics that measure future attention assigned to discarded states and the reactivation of historically inactive regions. We propose QEvict, a three-tier KV-cache management scheme that replaces binary retain-or-delete eviction with recoverable eviction. QEvict maintains high-confidence windows in full precision, stores intermediate windows in a quantized recoverable tier, and deletes only the lowest-confidence windows. During decoding, cumulative attention scores update window importance and when a quantized window becomes important again, it is dequantized and promoted to the full-precision. Under a fixed memory budget, this design preserves broader historical context while retaining exact full precision for the most important regions. Across long-context understanding, retrieval, and reasoning benchmarks, QEvict consistently improves over representative eviction and quantization baselines, reducing missed attention and improving information retention

On-device deployment of Large Language Models (LLMs) has become essential for personalized edge applications. A primary bottleneck is external memory access (EMA) in feed-forward network (FFN) layers. Speculative decoding and mixture-of-experts (MoE) are promising solutions. Speculative decoding reduces the number of decoding stages by generating multiple tokens per stage, and MoE minimizes per-stage cost through sparse expert activation. However, there is an incompatibility when combining these two techniques. We propose EdgeXpert, a software-hardware co-designed LLM accelerator that resolves this incompatibility. In the prefill stage, the prompt-wise expert reuse reformulates routing as prompt-level expert reuse rather than independent per-token expert selection. It identifies important tokens using a lightweight encoder, constructs a shared expert set from them, and routes less important tokens with a reduced expert budget to lower expert EMA. In the decode stage, depth-aware expert coalescing exploits the contextual similarity and mutual exclusivity of same-depth candidate tokens. Rather than loading the union of all required channels, EdgeXpert loads only salient channels and applies computational calibration to recover accuracy without additional memory access. Synthesized in Samsung 28nm technology at 800 MHz, EdgeXpert achieves up to 56.3% latency reduction and 44.1% energy reduction compared to prior works, while maintaining near-baseline accuracy.

Benchmark scores are reported as properties of a model, yet the inference framework used to produce them, such as HuggingFace, vLLM, or Ollama, are considered non-influential and their names and versions are almost never disclosed. In this work we investigate how much this choice can influence the model output. In a fully-crossed study (three instruction-tuned models x five inference frameworks x six benchmarks x four generation modes) we investigate how different tools (wrappers/backend) influence benchmark scores and how their score changes is influenced by generation hyper-parameters. We find backend to be a non-negligible factor where even under greedy, sampling-noise-free decoding, changing the backend can significantly alter models performance and this effect is structural and strongly model-dependent. Decomposing the variance according to generation mode reveal that considerable portion of the variability (roughly 39%) a practitioner sees out-of-the-box can stem from the backend, while the remaining stems from sampling noise and each framework's default generation parameters, both of which are avoidable by disclosing and matching the generation configuration. These divergences are more pronounced on factual than on social-bias benchmarks. Overall, benchmark numbers are not backend-agnostic therefore, we recommend disclosing the backend, its version, and the full generation configuration, also using deterministic decoding for cross-backend comparison.

Long-context inference with large language models is constrained by the linear growth of the key-value cache to sequence length. While pruning offers mitigation, prevailing methods determine query-specific token importance that cannot be reused across unseen queries. In contrast, we introduce TaskPress, a framework for task-guided, query-agnostic KV cache eviction. Instead of optimizing the cache for a single query, TaskPress constructs a reusable memory representation conditioned on a high-level task guide. The guide functions as a meta-query during prefill to filter irrelevant tokens before downstream queries are issued. In addition, TaskPress leverages quantization scale factors as a zero-cost signal for detecting influential representation outliers, providing an efficient proxy for token importance. Experiments on conducted on various tasks with long context input demonstrate that TaskPress efficiently creates a compact, reusable cache across diverse queries.

The attention score with rotary position embeddings (RoPE) decomposes exactly into a sum over its 2D-rotation frequency pairs, and each pair's wavelength limits how far it can discriminate position. Aligned with this structure, we propose the per-RoPE-wavelength distance window: it prunes the query--key inner-product terms beyond a wavelength-proportional distance. Unlike a sliding window, every key remains reachable, at least through the low-frequency pairs. The reduction rate is input-independent, with a closed form logarithmic in the sequence length $N$, in contrast to dynamic-sparse methods like MInference. Such token-level selection is orthogonal to our frequency-level pruning. The window can therefore be applied on top of those methods. On Qwen2.5-0.5B and Llama-3.2-3B, the window prunes 37--48% of the query--key inner-product terms within each model's native context length. Relative to full attention, the top-1 match rate stays at 96--98% and the mean output-distribution KL at the $10^{-3}$-nat level on LongBench-v2 contexts. We examine absolute scores on long-context benchmarks such as RULER, OpenAI-MRCR, LongCodeQA, and $\infty$Bench: they are broadly preserved. We implement the window as a slice of the query--key contraction axis, leaving the online-softmax recurrences untouched, and port it with minimal diffs into the released FlashAttention-4 prefill and FlashInfer decode. On RTX PRO 6000 with Llama, both ports outpace stock with gains growing with context length, up to $1.29\times$ at 128K. End to end on Qwen2.5-7B-1M, with 57% of the inner-product terms pruned, the speedup reaches $1.31\times$ at a 1M-token context.

The key-value (KV) cache is the primary memory bottleneck in long-context LLM inference. Existing approaches attack it from opposite ends: eviction methods permanently discard tokens, degrading performance whenever a discarded token later proves essential, while quantization methods retain all tokens at low precision but offer limited compression. We propose AnchorKV, a compression scheme that shrinks the cache by $20\times$ without discarding a single token. AnchorKV represents the cache using a small set of anchors stored exactly, expresses every other token through its most similar anchor, and refines only those whose approximation most affects the model's output. AnchorKV consistently preserves accuracy across models and datasets, retaining 99% of the full-cache score at the 70B scale, while keeping the entire context at a fraction of its cost.

Kernel generation for hardware accelerators such as GPUs and NPUs has become a proving ground for large language models (LLMs), and state-of-the-art systems raise correctness through pipelines that couple LLMs with agentic reinforcement learning and evolutionary search. Such pipelines generate, compile, and execute large numbers of candidate kernels, discarding most of them and forgoing the opportunity to distill failures into reusable knowledge. Many discarded candidates are near-miss operators that compile and run but fail numerical validation; each embodies genuine domain knowledge and a nontrivial investment in LLM inference, cross-compilation, and hardware execution. We argue for a paradigm shift: rather than regenerate, debug. Debugging is far more constrained than generating from scratch: the search space is small and feedback is dense. We present a domain-specific debug agent that addresses three core challenges in autonomous repair: mitigating knowledge scarcity through retrieved patterns and diagnostic instrumentation, ensuring integrity through anti-cheat detection and full-coverage evaluation, and controlling cost via convergence guards and bounded iteration. Debugging serves two complementary roles: it extends the capability frontier by recovering operators that repeated regeneration fails to produce, and it lowers cost per deliverable operator. Debug Pass@1 achieves 66.7% versus Regenerate Avg Pass@1's 25.9% and Regenerate Pass@3's 40.7%, while consuming 92.8% fewer tokens per success than three-trial regeneration. Component ablations show that the knowledge base drives recovery, while integrity gates reject 12.5-33.3% of the successes the workflow itself accepted.

Long-running assistants and agents consume interaction streams that eventually outgrow the context. Existing context retention, summarization, and retrieval preserve access to selected history, but do not provide a persistent state over the full lifecycle when working context changes. We formulate this missing inference capability as \emph{state continuity under context turnover}: carrying computation forward through a fixed-capacity memory state whose lifetime is independent of the active context. We introduce an intrinsic memory method, \textbf{LiveMem}, which augments a pretrained full-attention LLM with a memory state that preserves the historical information over the whole lifecycle while the main attention path retains a bounded KV window. Context turnover and memory state maintaining, memory-oriented post-training, and state-aware serving jointly make this memory state load bearing after its originating tokens are released. Our experiments show that LiveMem achieves leading overall performance among evaluated systems and other intrinsic memory methods. Experiments on LongMemEval show that LiveMem is able to answer the question based on the memory state, even when the supporting evidence has been removed from the current context, and evidence-distance analysis shows that useful information persists beyond the active window. LiveMem thus establishes state continuity as a distinct and complementary abstraction for continual LLM inference.

Speculative decoding accelerates LLM inference only when drafted continuations survive target-model verification. Semi-autoregressive drafters such as DSpark predict an entire token block with one backbone forward and refine it with a lightweight Markov head. However, DSpark decodes this block as a single chain, so an early mismatch invalidates the remaining suffix and limits the benefit of large draft blocks. We show that the conditional structure already learned by DSpark can support multiple parent-consistent continuations without retraining or additional backbone passes. We introduce Parent-Conditioned Drafting Tree (PCTree), which uses the pretrained Markov head to score alternative children separately for each concrete parent and allocates a fixed verification budget to the most probable paths. This converts DSpark's linear draft into a tree while preserving its one-pass parallel backbone. Across Qwen3-{4B,8B,14B} and nine benchmarks, at $B{=}7$, measured speedup gains over autoregressive (AR) decoding, relative to matched DSpark, range from $3.1\%$ to $29.5\%$. On Qwen3-4B GSM8K at $B{=}16$, PCTree increases mean acceptance length from $9.41$ to $11.16$ and three-run mean AR speedup from $6.14{\times}$ to $6.60{\times}$. These show that parent-conditioned branching can turn conditional capacity already present in a semi-autoregressive drafter into end-to-end inference gains through an inference-only change.

This paper investigates collaborative mobile edge computing (MEC) servers for large language model (LLM) inference under soft deadline constraints. In this system, to improve the quality of service, computations are expected to be completed within their deadlines. However, due to dependencies among tasks or subtasks, any missed deadline can lead to catastrophic consequences for the entire request. In this context, this work proposes an extended deadline mechanism with constrained flexibility. The main challenges lie in handling large-scale computations under strict latency constraints while limiting the number of allowable deadline extensions, especially in the presence of task dependencies within each request. To tackle these challenges, we develop a transformer-enhanced proximal policy optimization (PPO) framework that enables efficient collaboration among MEC servers. The proposed approach aims to maximize the number of tasks completed within their deadlines while minimizing the use of deadline extensions. By capturing temporal dependencies and cross-server interactions, the transformer improves decision-making for task migration. Simulation results demonstrate that the proposed method significantly outperforms conventional PPO and heuristic-based approaches in terms of task completion rate and overall system efficiency.

Large language model (LLM) inference has evolved from an offline workload into a continuously operated software service, yet root-cause analysis remains difficult because a single request spans the inference engine, Python/C++ backend, host CUDA APIs, GPU kernels, and distributed communication. Existing profilers expose raw timelines, while log-based diagnosis often misses cross-layer execution semantics and request-level structure. We present TELLER, a non-intrusive Trace- and Log-aware LLM inference Root-cause analysis framework. TELLER first collects NVTX/CUPTI traces and service logs without modifying model binaries, then reconstructs per-request call-chain trees and aligns log lines with the corresponding execution steps. We introduce a dependency-aware causal-context slice that preserves parent-child structure, temporal order, and communication relations, and a Trace Pair Encoding (TPE) tokenizer that compresses such slices into compact structural token sequences with parent, depth, and duration attributes. On top of these representations, TELLER combines numeric candidate localization with a multimodal root-cause model that jointly predicts abnormal steps, localizes suspicious operators, and generates natural-language explanations. Experiments on multi-node GPU inference workloads show a clear compression-accuracy trade-off: a moderate TPE vocabulary reduces per-step trace length by more than 80% while achieving the best overall performance on both horizontal (cross-node communication) and vertical (within-node execution stack) views, whereas more aggressive compression substantially degrades diagnosis quality. Further analyses under low-fault priors, strengthened baselines, modality ablations, explanation-quality checks, and tracing overhead show that TELLER provides a practical triage and evidence-localization substrate for LLM inference RCA.

The key-value (KV) cache has become a major memory and bandwidth bottleneck in long-context large language model inference, making ultra-low-bit quantization increasingly important. However, existing rotation-based INT2 methods optimize cache statistics or proxy errors before the complete attention readout, even though the model is ultimately affected by the error propagated through attention and the output projection $W_O$. To address this mismatch, we propose \textit{OptR}, an output-aware rotation method that minimizes post-$W_O$ attention-output error. OptR decomposes the post-$W_O$ attention-output error into key- and value-induced terms and learns per-head orthogonal corrections through the full INT2 quantization and attention path. OptR further applies an attention-equivalent key reparameterization to reduce large channel-wise offsets without changing the softmax distribution. Across three models and five reasoning and coding benchmarks, OptR consistently improves both QuaRot and OSCAR and strengthens long-context retrieval, while preserving the paged KV-cache format with negligible inference overhead.

Hybrid-attention large language models combine full attention with recurrent linear attention to reduce long-context inference costs, yet their autoregressive decoding remains memory-bound. Tree speculative decoding offers an attractive acceleration path, but existing tree-speculation systems are designed around the key--value caches of full-attention models. On hybrid models, they traverse recurrent layers branch by branch and materialize a full state for every proposal node, causing verification latency and transient memory to scale poorly with tree and batch sizes. We present Bole, a kernel--runtime co-design that enables efficient tree speculation for hybrid-attention LLMs. Bole transforms the linear-attention recurrence into a tree-structured closed form and realizes it with a resource-efficient GPU kernel, verifying all proposal nodes in parallel and accelerating linear-attention tree verification by 3.4--7.7$\times$. It losslessly encodes speculative state updates as token-level factors and reconstructs only the state selected after sampling, reducing transient state memory by 82--99$\times$ and freeing GPU capacity for KV caches. Its integration into SGLang, a widely deployed production LLM serving engine, couples efficient state management with a batch-wide verification budget calibrated to the complete hybrid forward. Across four models, two GPU platforms, and diverse datasets, Bole delivers up to $4.72\times$ the offline decode throughput of autoregressive decoding and up to $2.03\times$ that of the strongest tree-speculative baseline. Under online agent workloads, it reduces TTFT and TPOT by up to $67.6$ and $49.9$, respectively, over the strongest tree-speculative baseline.

Large Language Models (LLMs) increasingly rely on sparsity to reduce inference cost, but most prior work targets a single sparsity source-either weight or activation-and optimizes for batched multi-user inference. Dual-sparsity, which combines unstructured weight pruning with runtime activation sparsity, offers a compelling tradeoff among model size, accuracy, and latency for single-user decoding, but formulates as a Sparse Matrix-Sparse Vector (spMspV) workload that existing GPU kernels handle poorly. We propose Celty, a co-designed sparse format, GPU kernel, and SIMT microarchitecture for efficient spMspV in LLM inference. At the kernel level, Celty introduces a Run-Length Compressed CSC (RLC-CSC) format that enables vectorized loading of compressed weight columns and exploits both sparsity sources to skip unnecessary memory accesses, with shared memory used for scattered partial-product accumulation. At the microarchitecture level, the Celty Sparse SIMT Core integrates a pipelined RLC decoder to eliminate software-level index reconstruction and repurposes local register files for conflict-free accumulation-operating directly on the same RLC-CSC format without data layout changes. The Celty GPU kernel achieves up to 2.8x speedup over cuBLAS and 2.4x over Flash-LLM. With the Sparse SIMT Core, speedups reach up to 5.3x over cuBLAS at 70% dual-sparsity.

LLM agents accumulate long trajectories of reasoning steps, tool calls, and environment feedback, making the KV cache a major inference bottleneck. KV cache compaction can reduce this cost, but most prior methods assume a static context where future queries are known or can be approximated offline. Agents instead require online compaction: new information must be compressed before future relevance is known, using proxy queries cheap enough for the inference path. We study online compaction across token eviction (TE) and attention matching (AM), adapting both to compact agent turns and comparing cheap proxy sources such as boundary, repeat-prefill, and delayed future-generation queries. Experiments on BrowseComp-Plus and WideSearch show that immediate compaction often hurts performance, whereas delaying compaction to use the agent's future queries recovers much of the gap. Moreover, TE is often more robust than AM under imperfect proxies. Across models at different scales, TE preserves most of the accuracy while reducing KV cache by 80%, and can improve throughput over the no compaction baseline. These results position proxy-query selection as a core design choice for practical online KV compaction.

Speculative decoding mitigates the latency of sequential generation in autoregressive Large Language Models (LLMs) by interleaving draft generation with target verification. However, existing parallel drafting backends often suffer from rapid accuracy degradation over long horizons, leading to high rejection rates during verification and suboptimal wall-clock speedups. We observe that drafting errors are not uniformly distributed but typically stem from localized high-uncertainty tokens that destabilize downstream generation trajectories. Motivated by this token error pattern, we propose CURE, a budget-aware dynamic repair tree designed to repair errors at uncertainty focal points without incurring prohibitive tree-verification overheads. Specifically, our method uses predictive confidence margins to dynamically locate candidate error tokens within a block-parallel draft, expands bounded repair paths only at these fragile nodes, and employs a novel repair resynchronization mechanism to realign draft states post-verification. Evaluations on code-generation benchmarks (HumanEval, MBPP, and LiveCodeBench-lite) and mathematical reasoning benchmark (GSM8K) demonstrate that CURE increases the average accepted length by 4.2-7.5% over parallel baselines without repair, translating to an end-to-end speedup of $2.66-3.49\times$ over target-only decoding. Furthermore, we provide a plug-and-play repair module compatible with standard parallel drafting frameworks. We also characterize the trade-off between draft compute and verification efficiency.

As large language models serve more requests, cumulative inference cost is becoming increasingly important relative to one-time training cost. The two inference phases stress hardware differently: prompt prefill is parallel and typically compute-bound, whereas autoregressive decode is sequential and often memory-bandwidth-bound. Conventional width or depth scaling increases both costs together because every added layer is evaluated in both phases. We ask whether additional learned computation can instead be allocated to continuation prediction while preserving the prompt-wide primary computation and a single persistent key-value (KV) cache. We introduce the Dual-Flow Transformer. Its primary flow is a complete causal language model that processes the prompt and writes the KV cache. The auxiliary flow is omitted during prompt processing and activated only from the final prompt position onward, adding continuation-prediction computation without writing persistent state or influencing the primary flow. The two flows share major attention, MLP, and output matrices, while using separate token embeddings and lightweight coupling. Sharing weights and the primary cache also creates opportunities to reuse loaded weights and cached keys and values during grouped execution. Across matched-token comparisons, Dual-Flow achieves lower validation loss across architectures and data configurations. In MoE models, the separation makes primary and auxiliary expert fan-outs independent controls over prompt cost, continuation cost, and predictive quality. We study two regimes: increasing decode computation at fixed prefill expert computation, and reallocating a fixed decode expert budget between the two flows. These experiments expose a prefill-decode-quality trade-off and demonstrate the potential of phase-specific expert allocation.

Recent byte-level large language models (LLMs) have made tokenizer-free modeling increasingly competitive by grouping bytes into dynamically sized patches. However, existing byte-patch architectures still apply the same dense feed-forward computation to every patch. This uniform computation cannot adapt model capacity to variations in patch semantics and granularity. We address this limitation with EntropyMoE, a Mixture-of-Experts (MoE) architecture designed for dynamic byte patches. EntropyMoE replaces the dense feed-forward modules in the global patch Transformer with Top-K expert layers. Each dynamic patch serves as the basic unit of expert routing, and its byte coverage determines its contribution to workload accounting. The router selects experts directly from patch entropy, using the same granularity signal that underlies dynamic patch construction to organize sparse computation. Patch entropy and length jointly define the feature space for regulating expert specialization. Experiments show that EntropyMoE achieves the lowest held-out bits-per-byte among matched dense and sparse baselines while maintaining comparable downstream accuracy. These results establish patch entropy as an effective routing coordinate for sparse conditional computation and extend Mixture-of-Experts modeling beyond tokenizer-based representations.

Deploying large language models in realistic server environments poses challenges, as the system needs to provide high-quality responses with low latency. Quantization is a common approach to reduce the memory footprint and improve inference efficiency, yet its impact on latency and throughput is rarely evaluated under controlled, orchestration-level workloads. In this work we study the quantization trade-offs of two translation model families, EuroLLM \citep{martins2025eurollm} and Hy-MT2 \citep{zheng2026hy} across five models ranging from 1.7B to 22B for efficient deployment on a single A100 or H100 GPU. We demonstrate that combining a document-chunking strategy with W4A8 or W8A8 quantization improves the latency-throughput Pareto-curve under a wide range of workloads. Furthermore, since standard machine translation (MT) benchmarks rely on isolated sentences and fail to capture long-context dynamics, we introduce a document-level evaluation from WMT24++ to assess how text chunking strategies affect translation quality under quantization. Our results reveal that standard segment-level evaluation can fail to predict the interaction between quantization and long-context document translation. While Hy-MT2 remains robust under quantization, EuroLLM shows strong sensitivity and translation quality collapses rapidly for all considered quantization formats. Overall, our experiments show that the trade-off between inference efficiency and translation quality depends not only on the quantization format, but also on the choice of text chunking strategy.

AI coding agents like GitHub Copilot, Claude Code, and Codex interleave multi-step LLM inference with tool execution, creating a workload different from chatbots. We present the first production-scale characterization of this workload using sampled GitHub Copilot traces from June 2026, comprising 3.2M users, 13M sessions, 761M LLM calls, and 95T tokens. Our analysis reveals distinctive workload properties with important systems implications. For example, agentic coding sessions consist of sparse user-initiated turns, each unfolding into an autonomous agent loop of LLM calls almost always coupled with tool execution. This structure yields KV cache hit rates averaging 90% within a turn, but falling to 55% across turn boundaries and drastically invalidated after events like model switches or context compaction. Diverse workflows and user behaviors are observed with variable and long-tailed token consumption, time span, and tool calls. We highlight the difference between quick agentic turnaround times and the minutes-long user idle periods at turn boundaries, and design a lightweight idle-time predictor that captures 86-90% of total idle time, enabling proactive decisions for efficient resource orchestration. These findings challenge assumptions underlying current LLM-serving systems and provide an empirical foundation for agent-native infrastructure.

Pruning is a promising approach for improving the efficiency of LLMs. Existing static structured pruning methods are hardware-friendly and can deliver practical throughput gains, but their input-agnostic computation allocation often causes substantial accuracy degradation under aggressive sparsity. Recent dynamic sparsity methods improve quality retention by adapting computation to individual inputs, yet they remain largely limited to coarse-grained structural decisions and their practical acceleration under real-world inference scenarios remains challenging. To address these challenges, we present WIDE, the first end-to-end differentiable token-level dynamic width pruning framework designed for both prefill and decode scenarios. WIDE enables fine-grained computation allocation by allowing each token to dynamically select attention-head groups and FFN-channel groups, extending dynamic pruning beyond layer-level decisions to neuron-block-level granularity. Through a two-stage training pipeline, WIDE learns effective token-wise sparse execution patterns and achieves substantially better quality retention than existing approaches. To make such fine-grained dynamic pruning practical, we further propose a pruning--kernel co-design framework that decomposes dynamic sparsity acceleration into mask reordering, hardware-agnostic block-level skipping, and hardware-dependent intra-block skipping, enabling efficient execution across different granularities. At 50% sparsity, WIDE provides 55.1% performance boost when compared to the state-of-the-art dynamic depth pruning under calibration-only settings. Under prefill and decoding inference workloads, WIDE achieves close-to-theoretical kernel-level speedups of up to 1.98x for prefill and 4.95x for decoding, as well as 1.68x and 1.55x end-to-end acceleration. Our code is available at https://github.com/EIT-NLP/LLM-Pruning/tree/main/WIDE.

Transformer depth is not used uniformly: lower and middle layers build semantic representations, while upper layers increasingly specialize them for prediction. We turn this division of labor into CoMem (Comprehension Memory), which writes each context chunk only through an intermediate layer, retrieves a fixed number of cached residual states, and recomputes the query-conditioned upper layers over the resulting pack. For a fixed retrieval budget, model-side read compute and memory are independent of stored-context length. We evaluate a continued-trained Qwen3-8B base LM under a unified chat-template-free protocol. The backbone is frozen; the flagship trains only a rank-32 self-distillation LoRA on plain PG19, and we report an adapter-free arm separately. CoMem reaches 97.05 on RULER and 38.27 on LoCoMo versus 34.59 for full-context KV-Direct; the dialogue-memory advantage survives conversation-cluster resampling and an independent judge. Results on additional long-context and long-document tasks expose both the benefits of bounded retrieval and its in-window compression tax. Controlled depth sweeps show that deeper caching lowers per-query recomputation but incurs a fidelity loss that self-distillation substantially repairs. In a separate adapter-free efficiency control on an NVIDIA H20 at 128k, CoMem uses 18.26 GB rather than 89.36 GB and achieves a 7.83x prefill speedup. These results show that long-context memory can be organized along the layer axis, not only the token axis.

Multimodal foundation models are reshaping edge-cloud visual intelligence from task-specific feature pipelines into token-based interfaces, where edge devices encode visual inputs into tokens for a general-purpose cloud MLLM. However, dense visual-token sequences increase cloud-side inference costs. Existing pruning methods mainly target centralized inference: vision-driven methods can operate before cloud execution but are typically query-agnostic, whereas query-guided methods often rely on internal states of the target MLLM and cannot determine token relevance before transmission. Compact guidance models offer an alternative, but existing designs may require costly attention aggregation or auxiliary generation. We propose LAST, a training-free framework for query-dependent visual token pruning in edge-cloud collaborative MLLM inference. LAST uses a compact edge-side VLM as a guidance proxy and derives a lightweight importance signal from the last query token's attention to visual tokens. Under causal attention, the last query token can attend to the full visual sequence and the entire query context, enabling query-aware pruning without cloud-model access, autoregressive generation, or costly aggregation over multiple query positions. LAST then retains a diverse set of query-relevant visual tokens under a fixed token budget. We evaluate LAST on 11 multimodal benchmarks under multiple token budgets against pruning methods with different guidance strategies. Experiments show that LAST consistently achieves the strongest performance, preserving 95.4% of the full-token accuracy while retaining only 12.5% of the visual tokens, with low edge-side selection overhead and reduced cloud-side computation.

Speculative decoding alleviates the memory-bandwidth bottleneck in large language model inference, but its acceleration is jointly constrained by drafting overhead, token acceptance, and speculation length. We present a unified efficiency analysis showing that extending the speculation horizon can reduce rather than improve speedup when the marginal acceptance probability falls below the relative drafting cost. Guided by this analysis, we introduce SparseSpec-L, a training-free self-speculative decoding framework for long-context inference. SparseSpec-L generates lightweight drafts directly from the target model using a dynamically sparsified and recallable KV cache. It recycles per-head attention statistics produced during full-context verification as a no-extra-forward importance signal, allowing critical historical tokens to be recalled without permanently discarding the dense KV cache. An online entropy-based controller further selects the speculation length according to expected step-wise efficiency. Experiments across multiple long-context tasks and model scales show consistent end-to-end acceleration, with up to speedup over autoregressive decoding while preserving the target model's output distribution.