LLM Inference Efficiency

We present SemantiClean, a modular framework for extracting structured semantic signals from e-commerce session data and driving pluggable inference targets including purchase intent, customer segmentation, and product affinity through a shared element library. Unlike conventional end-to-end predictors that optimise solely for accuracy, SemantiClean prioritises auditability, structural governance, and sigma=0 reproducibility, explicitly trading marginal predictive gains for element-level transparency and defensible decision trails. Built upon the Online Shoppers Purchasing Intention (OSPI) dataset, the framework organises twenty-four behavioural elements into a four-layer architecture (Functional, Interaction, Systemic, Contextual) and enforces signal quality through three anti-inflation mechanisms: RedundancyGroup contribution caps, TieredPenaltyCalculator bias penalties, and AdaptiveConstraintMode cold-start protection.This report introduces the LLM-Integrated Semantic Inference Engine, a fully implemented two-phase LLM-driven inference architecture that leverages complete element metadata at inference time. All quantitative results reported herein are produced by this engine. Deterministic engine outputs remain fully reproducible (sigma=0); LLM-dependent results (E8, E10) are subject to controlled output variability under fixed provider/model/temperature settings. The gender inference target remains non-functional in the current implementation and is excluded from all quantitative results.

Efficient inference for on-device Large Language Models (LLMs) remains challenging due to limited hardware resources and the high cost of the prefill stage, which processes the full input context to construct Key-Value (KV) caches. We present SparKV, an adaptive KV loading framework that combines cloud-based KV streaming with on-device computation. SparKV models the cost of individual KV chunks and decides whether each chunk should be streamed or computed locally, while overlapping the two execution paths to reduce latency. To handle fluctuations in wireless connectivity and edge resource availability, SparKV further refines offline-generated schedules at runtime to rebalance communication and computation costs. Experiments across diverse datasets, LLMs, and edge devices show that SparKV reduces Time-to-First-Token by 1.3$x-5.1x with negligible impact on response quality, while lowering per-request energy consumption by 1.5x to 3.3x, demonstrating its robustness and practicality for real-world on-device deployment.

Compressing large language models (LLMs) for deployment on commodity GPUs remains challenging: conventional scalar quantization is limited to fixed bit-widths (e.g., 8/4/3-bit), offers only a few discrete compression points, and typically requires calibration data. We present FASQ (Flexible Accelerated Subspace Quantization), a calibration-free framework that applies product quantization to LLM weight matrices. By tuning two parameters, sub-vector size and codebook cardinality, FASQ exposes a continuous design space spanning 27-49% of the original FP16 model size, filling compression gaps that fixed-bit schemes cannot reach. On Meta-Llama-3-8B, FASQ surpasses 4-bit GPTQ and AWQ in accuracy (67.1-67.7 avg.) at 37-42% model size, with consistent results on Qwen3-8B and Qwen3.5-9B-Base. To make product quantization practical at inference time, we design custom CUDA kernels: a LUT-free direct-compute GEMV for decode and an output-stationary double-buffered LUT GEMM for prefill, both with split-K parallelism. On an RTX~3090, FASQ achieves 45.2 tok/s decode at effective 4-bit (2.56x memory reduction) and 51.8 tok/s at effective 3-bit (2.80x), both surpassing FP16 tensor-core performance (43.9 tok/s) and delivering 1.6 to 1.8x the throughput of AWQ, 2.5 to 2.5x of GPTQ, and 4.3 to 5x of RTN. FASQ is the only compressed method that accelerates decode beyond FP16, offering calibration-free compression, continuous size-quality trade-offs, and real-time inference on a single consumer GPU.

Deploying large language models to heterogeneous hardware is often constrained by memory, not compute. We introduce MCAP (Monte Carlo Activation Profiling), a load-time per-layer importance estimator that enables dynamic precision and memory placement decisions on the target device. MCAP produces a lightweight per-layer signal that drives both precision dispatch (W4A8 vs. W4A16) and residency tier (GPU, RAM, SSD), allowing a single set of weights to operate across diverse memory budgets. Our system, NVE, achieves 1.5-1.8x higher decode throughput than llama-cpp Q4_0 on NVIDIA T4 and enables models to run in memory regimes previously infeasible without modifying weights.

Cloud-hosted LLM inference for autonomous driving adds round-trip delay and depends on stable connectivity, while purely local edge models struggle under occlusion. We present SwarmDrive, a semantic Vehicle-to-Vehicle (V2V) coordination framework in which nearby vehicles run local Small Language Models (SLMs), share compact intent distributions only when uncertainty is high, and fuse them through event-triggered consensus. We evaluate SwarmDrive in a 5-seed executable study built around one occluded intersection case, combining matched operating-point comparisons with robustness sweeps. In that setting, SwarmDrive under its 6G communication setting ("Swarm 6G") raises success from 68.9% to 94.1% over a single local SLM while reducing latency from a 510 ms cloud reference to 151.4 ms. However, an increased number of participating vehicles leads to higher communication overhead and packet loss. SwarmDrive also evaluates the impact of swarm-size, packet-loss, and entropy-threshold sweeps and shows that the cooperative gain holds across ablations and is best balanced near an active swarm size of 4 vehicles and an entropy trigger threshold of 0.65 in the current prototype. These results show that semantic edge cooperation can work under tight latency constraints in the targeted intersection case, but they are not a deployment-grade validation of a real 6G stack.

Long-context inference in Large Language Models (LLMs) is bottlenecked by the linear growth of Key-Value (KV) cache memory. Existing KV cache compression paradigms are fundamentally limited by heuristics: heuristic budgeting relies on statistical priors rather than task objectives, causing resource misallocation, while heuristic selection relies on coupled query-key interactions or static inductive biases (e.g., attention sinks). To address this limitation, we introduce LKV (Learned KV Eviction), which formulates KV compression as an end-to-end differentiable optimization problem. LKV integrates LKV-H to learn task-optimized global budgets, and LKV-T to derive intrinsic KV importance without materializing attention matrices. This design bypasses heuristic proxies, strictly aligning compression with task objectives. Extensive evaluations demonstrate that LKV achieves state-of-the-art performance on both LongBench and RULER benchmarks at high compression rates. In particular, on LongBench, LKV achieves near-lossless performance with only 15% KV cache retention. Crucially, our analysis identifies learned budgeting as the dominant driver of fidelity, demonstrating that data-driven allocation is essential to overcome the limitations of hand-crafted heuristics.

Sparse attention can reduce the cost of long-context inference, but most variants introduce new architectural components. We introduce Simplified Sparse Attention (SSA), a simpler approach to sparse attention that requires no architectural changes. Concretely, we first perform continued pretraining on sequences interleaved with gist tokens. We optimize the standard next-token loss as usual, but the gist tokens use an attention mask to restrict what parts of the context the language model can attend to; this teaches the model to pack each chunk's important information into the gist tokens. At inference time, SSA scores chunks via attention between the current query and the small set of gist tokens, selectively unfolding the top-k chunks by reintroducing their corresponding raw tokens. Since the query is scored only against the gist tokens, we avoid the memory-bandwidth cost associated with naive scoring against the full KV cache, without requiring the auxiliary KV cache approach used by sparse attention methods. On LongBench, SSA consistently outperforms compression and inference-time sparse-attention baselines under the same compression ratio. More strikingly, in retrieval-augmented generation, SSA can even outperform full attention after continued pretraining by over 5.7 points. We attribute this to the ability of SSA's selective unfolding, which concentrates attention on the query-relevant chunks and effectively filters out noise. SSA further extends to a hierarchical gist-of-gist variant (H-SSA) that achieves log-linear decoding complexity while maintaining or improving accuracy at high compression ratios up to 32x. The code is available at https://github.com/yuzhenmao/simplified-sparse-attention/.

Large language models are increasingly deployed on CPU-only platforms where memory bandwidth is the primary bottleneck for autoregressive generation. Weight quantization to four bits or below reduces memory pressure, yet existing systems still dequantize weights and perform floating-point multiplications, limiting the achievable gains. Ternary weights in {-1, 0, +1} provide a more efficient alternative, replacing multiplications with conditional additions, subtractions, or no-ops. While Fairy2i shows that ternary LLMs can match FP16 quality, its runtime does not exploit this structure. We present FairyFuse, an inference system that enables multiplication-free execution on commodity CPUs by fusing the eight real-valued sub-GEMVs of each widely-linear layer into a single AVX-512 loop using masked additions and subtractions, with zero floating-point multiplications. Roofline analysis shows that 16x weight compression shifts memory-bound GEMV toward the compute regime on bandwidth-limited CPUs, yielding a 29.6x kernel speedup while offering little benefit on GPUs. End-to-end, FairyFuse achieves 32.4 tokens per second on a single Intel Xeon 8558P, outperforming llama.cpp Q4_K_M by 1.24x with near-lossless quality (WikiText-2 perplexity 5.52 vs. 5.47 FP16; downstream accuracy 66.0%).

The quadratic computational complexity of the standard attention mechanism constitutes a fundamental bottleneck for large language models in long-context inference. While existing KV cache compression methods alleviate memory pressure, they often sacrifice generation quality and fail to address the high overhead of floating-point arithmetic. This paper introduces DASH-KV, an innovative acceleration framework that reformulates attention as approximate nearest-neighbor search via asymmetric deep hashing. Under this paradigm, we design an asymmetric encoding architecture that differentially maps queries and keys to account for their distinctions in precision and reuse characteristics. To balance efficiency and accuracy, we further introduce a dynamic mixed-precision mechanism that adaptively retains full-precision computation for critical tokens. Extensive experiments on LongBench demonstrate that DASH-KV significantly outperforms state-of-the-art baseline methods while matching the performance of full attention, all while reducing inference complexity from O(N^2) to linear O(N).

Inference-time LLM alignment methods, particularly activation steering, offer an alternative to fine-tuning by directly modifying activations during generation. Existing methods, however, often rely on non-anticipative interventions that ignore how perturbations propagate through transformer layers and lack online error feedback, resulting in suboptimal, open-loop control. To address this, we show empirically that, despite the nonlinear structure of transformer blocks, layer-wise dynamics across multiple LLM architectures and scales are well-approximated by locally-linear models. Exploiting this property, we model LLM inference as a linear time-varying dynamical system and adapt the classical linear quadratic regulator to compute feedback controllers using layer-wise Jacobians, steering activations toward desired semantic setpoints in closed-loop with minimal computational overhead and no offline training. We also derive theoretical bounds on setpoint tracking error, enabling formal guarantees on steering performance. Using a novel adaptive semantic feature setpoint signal, our method yields robust, fine-grained behavior control across models, scales, and tasks, including state-of-the-art modulation of toxicity, truthfulness, refusal, and arbitrary concepts, surpassing baseline steering methods. Our code is available at: https://github.com/trustworthyrobotics/lqr-activation-steering

Diffusion Large Language Models (dLLMs) have emerged as a promising alternative to autoregressive generation by enabling parallel token prediction. However, practical dLLM decoding still suffers from high inference latency, which limits deployment. In this work, we observe that a substantial part of this inefficiency comes from recurring redundancy in the decoding process, including spatial redundancy caused by confidence clusters and positional ambiguity, and temporal redundancy caused by repeatedly remasking predictions that have already stabilized. Motivated by these patterns, we propose $R^{2}$-dLLM, a unified framework for reducing decoding redundancy from both inference and training perspectives. At inference time, we introduce training-free decoding rules that aggregate local confidence and token predictions, and finalize temporally stable tokens to avoid redundant decoding steps. We further propose a redundancy-aware supervised fine-tuning pipeline that aligns the model with efficient decoding trajectories and reduces reliance on manually tuned thresholds. Experiments demonstrate that $R^{2}$-dLLM consistently reduces the number of decoding steps by up to 88% compared to existing decoding strategies, while maintaining competitive generation quality across different models and tasks. These results validate that decoding redundancy is a central bottleneck in dLLMs, and that explicitly reducing it yields substantial practical efficiency gains. Our code and models are available at https://github.com/GATECH-EIC/R2-dLLM.

Apple Neural Engine (ANE) is a dedicated neural processing unit (NPU) present in every Apple Silicon chip. Mixture-of-Experts (MoE) LLMs improve inference efficiency via sparse activation but are challenging for NPUs in three ways: expert routing is unpredictable and introduces dynamic tensor shapes that conflict with the shape-specific constraints of NPUs; several irregular operators, e.g., top-k, scatter/gather, etc., are not NPU-friendly; and launching many small expert kernels incurs substantial dispatch and synchronization overhead. NPUs are designed to offload AI compute from CPU and GPU; our goal is to enable such offloading for MoE inference, particularly during prefill, where long-context workloads consume substantial system resources. This paper presents NPUMoE, a runtime inference engine that accelerates MoE execution on Apple Silicon by offloading dense, static computation to NPU, while preserving a CPU/GPU fallback path for dynamic operations. NPUMoE uses offline calibration to estimate expert capacity and popularity that drives three key techniques: (1) Static tiers for expert capacity to address dynamic expert routing (2) Grouped expert execution to mitigate NPU concurrency limits (3) Load-aware expert compute graph residency to reduce CPU-NPU synchronization overhead. Experiments on Apple M-series devices using three representative MoE LLMs and four long-context workloads show that NPUMoE consistently outperforms baselines, reducing latency by 1.32x-5.55x, improving energy efficiency by 1.81x-7.37x, and reducing CPU-cycle usage by 1.78x-5.54x through effective NPU offloading.

Large Language Models (LLMs) have demonstrated exceptional performance across diverse domains but are increasingly constrained by high inference latency. Early Exit has emerged as a promising solution to accelerate inference by dynamically bypassing redundant layers. However, in decoder-only architectures, the efficiency of Early Exit is severely bottlenecked by the KV Cache Absence problem, where skipped layers fail to provide the necessary historical states for subsequent tokens. Existing solutions, such as recomputation or masking, either introduce significant latency overhead or incur severe precision loss, failing to bridge the gap between theoretical layer reduction and practical wall-clock speedup. In this paper, we propose River-LLM, a training-free framework that enables seamless token-level Early Exit. River-LLM introduces a lightweight KV-Shared Exit River that allows the backbone's missing KV cache to be naturally generated and preserved during the exit process, eliminating the need for costly recovery operations. Furthermore, we utilize state transition similarity within decoder blocks to predict cumulative KV errors and guide precise exit decisions. Extensive experiments on mathematical reasoning and code generation tasks demonstrate that River-LLM achieves 1.53 to 2.16 times of practical speedup while maintaining high generation quality.

Processing-in-Memory (PIM) architectures offer a promising solution to the memory bottlenecks in data-intensive machine learning, yet often overlook the growing challenge of activation memory footprint. Conventional PIM approaches struggle with massive KV cache sizes generated in long-context scenarios by Transformer-based models, frequently exceeding PIM's limited memory capacity, while techniques like sparse attention can conflict with PIM's need for data locality. Existing PIM approaches and quantization methods are often insufficient or poorly suited for leveraging the unique characteristics of activations. This work identifies an opportunity for PIM-specialized activation quantization to enhance bandwidth and compute efficiency. We explore clustering-based vector quantization approaches, which align well with activation characteristics and PIM's internal bandwidth capabilities. Building on this, we introduce AQPIM, a novel PIM-aware activation quantization framework based on Product Quantization (PQ), optimizing it for modern Large Language Models (LLMs). By performing quantization directly within memory, AQPIM leverages PIM's high internal bandwidth and enables direct computation on compressed data, significantly reducing both memory footprint and computational overhead for attention computation. AQPIM addresses PQ's accuracy challenges by introducing several algorithmic optimizations. Evaluations demonstrate that AQPIM achieves significant performance improvements, drastically reducing of GPU-CPU communication that can account for 90$\sim$98.5% of decoding latency, together with 3.4$\times$ speedup over a SOTA PIM approach.

Decentralized LLM inference networks need lightweight, reference-free quality evaluation for Proof of Quality (PoQ). We present PoQ-Judge, a framework that trains dedicated judge models to score query-output pairs without ground-truth references. We study three architectures across the quality-cost tradeoff: a TextCNN judge, a MiniLM cross-encoder, and a DeBERTa judge. Using two-stage training on UltraFeedback plus GPT-labeled in-domain data, the best model reaches 0.747 Pearson correlation with the ground-truth proxy on a held-out test set, outperforming reference-based evaluators from prior work. As a reference-free component in composite scoring, it achieves 0.645 Pearson correlation, matching the best single reference-based evaluator while removing the need for reference answers. We also show that online calibration identifies semantic quality as the dominant dimension and that cascade evaluation reduces cost by 72.7 percent with only modest quality loss. Results are much stronger on QA than summarization, pointing to proxy quality as the main remaining limitation.

Existing works on large language model (LLM) decomposition mainly focus on improving performance on downstream tasks, but they ignore the poor parallel inference performance when trying to scale up the model size. To mitigate this important performance issue, this paper introduces DeInfer, a high-performance inference system dedicated to parallel inference of decomposed LLMs. It consists of multiple optimizations to maximize performance and be compatible with state-of-the-art optimization techniques. Extensive experiments are carried out to evaluate DeInfer's performance, where the results demonstrate its superiority, suggesting it can greatly facilitate the parallel inference of decomposed LLMs.

While distributed device-edge speculative decoding enhances resource utilization across heterogeneous nodes, its performance is often bottlenecked by conventional token-level verification strategies. Such rigid alignment leads to excessive rejections, significantly diminishing the accepted sequence length and increasing interaction rounds under fluctuating wireless conditions. In this paper, we propose WISV (Wireless-Informed Semantic Verification), a novel distributed speculative decoding framework that goes beyond strict token-level matching via a channel-aware semantic acceptance policy. WISV integrates a lightweight decision head into the edge-side target LLM to dynamically evaluate speculative tokens by synthesizing high-dimensional hidden representations with instantaneous channel state information (CSI). To optimize the trade-off between verification fidelity and communication overhead, we further design two tailored communication protocols: full-hidden upload and mismatch-first selective-hidden upload. Extensive simulations using a 1B drafter and an 8B target model demonstrate that WISV achieves up to a 60.8% increase in accepted length, a 37.3% reduction in interaction rounds, and a 31.4% improvement in end-to-end latency compared to vanilla speculative decoding across tested settings, while maintaining a negligible task accuracy drop (<1%). Finally, we validate WISV on a hardware testbed comprising an NVIDIA Jetson AGX Orin and an A40-equipped server, confirming its real-world efficacy in accelerating edge-deployed LLM inference.

KV cache memory is the dominant bottleneck for long-context LLM inference. Existing compression methods each act on a single axis of the four-dimensional KV tensor -- token eviction (sequence), quantization (precision), low-rank projection (head dimension), or cross-layer sharing -- but apply the same recipe to every layer. We show that this homogeneity leaves accuracy on the table: different layers respond very differently to each compression operation, and the optimal per-layer mix of eviction and quantization is far from uniform. We propose MoE-nD, a mixture-of-experts framework that routes each layer to its own (eviction-ratio, K-bits, V-bits) tuple under a global memory budget. An offline-calibrated greedy solver chooses the routing that minimizes predicted quality loss; at inference time, per-layer heterogeneous eviction and quantization are applied jointly through a single attention patch. On a 4-task subset of LongBench-v1 (16k inputs, n=50 per task, adapted reasoning-model protocol; see section Experiments), MoE-nD's hetero variant matches our uncompressed 1.9GB baseline at 14x compression (136MB) while every other compressed baseline we tested (1d, 2d_uniform, 2d) at comparable or smaller memory stays under 8/100. The gains hold on AIME reasoning benchmarks (+6 to +27 pts over the strongest per-layer-quantization baseline across eight configurations). Two null results -- MATH-500 and LongBench's TREC -- share a principled cause (short inputs, solver picks keep=1.0 on most layers), cleanly characterizing when per-layer eviction routing has headroom to help.

Serialization formats designed for document interchange impose structural overhead that becomes prohibitive when large language models consume operational data at scale. A modest dataset of 1,000 IoT sensor readings serialized as JSON requires approximately 80,000 tokens - the majority spent on repeated field names, nested braces, and structural punctuation rather than semantic content. We present ONTO (Object Notation for Token Optimization), a columnar notation that declares field names once per entity and arranges values in pipe-delimited rows with indentation-based hierarchy. This schema-once, data-many design eliminates per-record key repetition while preserving human readability and nested structure support. Evaluation across three synthetic operational datasets demonstrates 46-51% token reduction versus JSON, with stable scaling from 100 to 1,000 records. Controlled inference benchmarks on Qwen2.5-7B show corresponding 5-10% latency improvement. Comprehension validation confirms no material degradation in LLM task accuracy across lookup, counting, extraction, and aggregation operations when format context is provided. Ablation analysis reveals that key repetition accounts for the majority of JSON overhead, with indentation costs in nested structures explaining the 4-percentage-point gap between flat and hierarchical data. ONTO occupies a previously unfilled position in the serialization landscape: columnar efficiency with hierarchical structure, optimized for LLM context windows rather than document interchange. Code and specification are available at https://github.com/harsh-aranga/onto.

We present Open-TQ-Metal, the first implementation of fused compressed-domain attention on Apple Silicon, enabling 128K-context inference for Llama 3.1 70B on a single 64GB consumer Mac -- a configuration impossible with all existing inference frameworks. Open-TQ-Metal quantizes the KV cache to int4 on the fly and computes attention directly on the compressed representation via custom Metal compute shaders, eliminating all intermediate dequantization matrices. Across 330 experiments spanning two model families (Gemma 4 31B and Llama 3.1 70B), the fused sdpa_int4 kernel achieves 48x attention speedup at 128K context over the dequantize-then-attend baseline, reduces KV cache memory from 40 GB to 12.5 GB (3.2x compression), and maintains identical top-1 token predictions to FP16 inference. We further provide the first cross-architecture analysis of KV cache quantization methods, revealing that the attention scale factor -- not model size -- determines whether angular quantization schemes like PolarQuant succeed or fail, with Gemma 4's attn_scale=1.0 amplifying directional error 25-100x more than Llama's standard 1/sqrt(d) scaling.

In long-context decoding for LLMs and LMMs, attention becomes increasingly memory-bound because each decoding step must load a large amount of KV-cache data from GPU memory. Existing acceleration strategies often trade efficiency for accuracy by relying on heuristic pruning that may discard useful information. At a deeper level, they also tend to indiscriminately preserve all high-scoring tokens, treat early tokens as indispensable anchors, or rely on heuristic head routing, reflecting an insufficient mechanistic understanding of the attention sink phenomenon. In this paper, we show that the attention sink phenomenon corresponds to a stable, reachable, and error-controllable fixed point constructed during training. Based on this insight, we propose SinkRouter, a training-free selective routing framework that detects the sink signal and skips computations that would otherwise produce near-zero output. To translate this mechanism into real-world acceleration, we develop a hardware-aware Triton kernel with block-level branching and Split-K parallelism. We conduct extensive evaluations on a diverse suite of long-context benchmarks, including LongBench, InfiniteBench, CVBench, MileBench, and MMVP, using both text-only and multimodal backbones such as Llama-3.1-8B, Llama-3.1-70B, Yi-9B-200K, LLaVA-1.5-7B, and LLaVA-1.5-13B. Across these settings, SinkRouter consistently improves decoding efficiency while maintaining competitive accuracy, and reaches 2.03x speedup with a 512K context.

Multimodal large language models (MLLMs) have recently demonstrated strong capabilities in understanding and generating responses from diverse visual inputs, including high-resolution images and long video sequences. As these models scale to richer visual representations, inference increasingly relies on storing large numbers of vision tokens in the key-value (KV) cache, making memory consumption a central bottleneck. Existing methods address this issue by identifying redundancy in vision tokens and compressing the cache, but such compression is typically applied only after all inputs are processed, resulting in high peak memory usage during the prefill stage. In this work, we show that MLLMs exhibit inherent structural regularities and representational redundancy that can be exploited to control memory growth throughout inference. Based on this insight, we propose a sequential input-compression mechanism that enforces a fixed memory budget by performing structure-aware key-value cache compression during the prefill process. This approach substantially reduces peak memory usage while maintaining generative performance with only minimal degradation, enabling more practical and memory-efficient multimodal inference.

Speculative decoding accelerates Large Language Model inference via a draft-then-verify paradigm, yet the output projection layer becomes a bottleneck as vocabulary sizes scale. While existing static pruning methods effectively reduce this overhead, they suffer from precipitous drops in acceptance rate in specialized domains or topic-switching scenarios due to their inability to capture dynamic distribution shifts. To address this, we introduce EvoSpec, a framework that enables real-time evolution of the draft model through dynamic vocabulary and parameter adaptation. Unlike static or purely retrieval-based approaches, EvoSpec employs a context-aware mechanism that retrieves critical long-tail tokens via efficient semantic and statistical indexing. Furthermore, we propose a lightweight online alignment strategy utilizing curriculum learning to continually minimize the distributional gap between the draft and target models. Extensive evaluations across specialized domains (coding, law, and medicine) confirm that EvoSpec overcomes the limitations of static baselines. On EAGLE-3, it achieves a 1.13x speedup in these settings over the state-of-the-art static baseline FR-Spec, with 27% lower memory overhead than standard online adaptation.

Speculative decoding (SD) accelerates language model inference by drafting tokens from a cheap proposal model and verifying them against an expensive target model via rejection sampling. Because rejection truncates the draft block at the first error, throughput degrades when draft and target diverge. Rather than rejecting draft tokens outright, we propose to reweight them. To this end, we introduce sequential Monte Carlo speculative decoding (SMC-SD), which replaces token-level rejection with importance-weighted resampling over a population of draft particles. SMC-SD is a principled approximate inference scheme that trades exactness for additional speed, while preserving theoretical bounds on its per-step approximation error. Because LLM inference is memory bandwidth-bound, the arithmetic needed to draft particles and to score them in parallel comes nearly for free -- SMC-SD uses idle compute to turn verification into a vectorized, fixed-size operation with no rollback. Empirically, SMC-SD achieves 2.36x speed-up over speculative decoding and a 5.2x speed-up over autoregressive decoding, while remaining within 3% of the target model's accuracy on reasoning, instruction-following, and coding benchmarks.

Large Language Model (LLM) deployment is increasingly shifting to cost-efficient accelerators like Google's Tensor Processing Units (TPUs), prioritizing both performance and total cost of ownership (TCO). However, existing LLM inference kernels and serving systems remain largely GPU-centric, and there is no well-established approach for efficiently mapping LLM workloads onto TPU architectures--particularly under the dynamic and ragged execution patterns common in modern serving. In this paper, we present Ragged Paged Attention (RPA), a high-performance and flexible attention kernel for TPUs, implemented using Pallas and Mosaic. RPA addresses these challenges through three key techniques: (1) fine-grained tiling to enable efficient dynamic slicing over ragged memory, (2) a custom software pipeline that fuses KV cache updates with attention computation, and (3) a distribution-aware compilation strategy that generates specialized kernels for decode, prefill, and mixed workloads. Evaluated on Llama 3 8B on TPU7x, RPA achieves up to 86% memory bandwidth utilization (MBU) in decode and 73% model FLOPs utilization (MFU) in prefill. Integrated as the primary TPU backend in vLLM and SGLang, RPA provides a production-grade foundation for efficient TPU inference and offers practical insights into kernel design.

Speculative decoding (SD) accelerates large language model inference by allowing a lightweight draft model to propose outputs that a stronger target model verifies. However, its token-centric nature allows erroneous steps to propagate. Prior approaches mitigate this using external reward models, but incur additional latency, computational overhead, and limit generalizability. We propose SpecGuard, a verification-aware speculative decoding framework that performs step-level verification using only model-internal signals. At each step, SpecGuard samples multiple draft candidates and selects the most consistent step, which is then validated using an ensemble of two lightweight model-internal signals: (i) an attention-based grounding score that measures attribution to the input and previously accepted steps, and (ii) a log-probability-based score that captures token-level confidence. These signals jointly determine whether a step is accepted or recomputed using the target, allocating compute selectively. Experiments across a range of reasoning benchmarks show that SpecGuard improves accuracy by 3.6% while reducing latency by ~11%, outperforming both SD and reward-guided SD.

KV caching is a ubiquitous optimization in autoregressive transformer inference, long presumed to be numerically equivalent to cache-free computation. This assumption fails under standard FP16 precision: cache-ON and cache-OFF execution paths employ different floating-point accumulation orderings which, due to FP16 non-associativity, produce a deterministic divergence in decoded token sequences. Across three open-weight models (LLaMA-2-7B, Mistral-7B-v0.3, Gemma-2-2B) evaluated on GSM8K, we observe a 100% token divergence rate across all sampling strategies, including greedy decoding, which rules out sampling randomness as a cause, and also with cache-ON yielding higher accuracy in 8 of 9 conditions, where the accuracy difference serves as an indicator that the divergence direction is systematic rather than random. Controlled FP32 falsification reduces divergence by eight orders of magnitude, eliminates token flips, and drops the flip rate to exactly 0.0%, confirming FP16 non-associativity as the sole causal driver. Layer-wise drift profiling reveals architecturally predictable propagation patterns: models using Grouped-Query Attention exhibit sharp divergence at the first layer, while Gemma's larger head dimension and sliding window attention produce uniform accumulation across all layers. Finally, activation patching of the entire residual stream fails to recover the cache-free trajectory, localizing the causal variable to the stateful KV cache. These findings establish that FP16 KV cache inference is fundamentally non-equivalent to recomputation and provide a mechanistic framework for understanding numerical instability in modern LLM inference systems.

LLM-based coding agents can generate functionally correct GPU kernels, yet their performance remains far below hand-optimized libraries on critical computations such as matrix multiplication, attention, and Mixture-of-Experts (MoE). Peak GPU performance requires coordinated reasoning over tightly coupled optimizations, including tiling, shared-memory staging, software pipelining, and instruction scheduling, while existing agents rely on sparse pass/fail feedback, leaving them unable to diagnose global constraint violations. We present Argus, an agentic framework that addresses this through data-flow invariants: compile-time specifications encoding how data must be choreographed throughout kernel execution. Argus introduces a tile-based, Pythonic DSL exposing hardware instructions and compiler policies while hiding low-level representations. The DSL provides tag functions to propagate symbolic annotations through data and control flow, and tag assertions to enforce relational constraints at use sites. When violations occur, the compiler returns concrete counterexamples identifying the thread, data element, and program point, enabling dense, structured feedback for targeted fixes. Invariants are verified at compile time via abstract interpretation over a layout algebra and SMT solving, with zero runtime overhead. An in-context reinforcement learning planner learns to select optimizations and synthesize effective invariants, supported by a curated knowledge base of GPU optimization techniques. We evaluate Argus on the AMD MI300X GPU across GEMM, flash attention, and MoE kernels accounting for over 90% of GPU time in LLM inference. Generated kernels achieve 99-104% of state-of-the-art hand-optimized assembly throughput and are 2-1543x faster than existing agentic systems. Argus further generalizes to 200 KernelBench tasks, solving 100% of Level 1 and 90% of Level 2 problems.

Open-weight Small Language Models(SLMs) can provide faster local inference at lower financial cost, but may not achieve the same performance level as commercial Large Language Models (LLMs) that are orders of magnitudes larger. Consequently, many of the latest applications of LLMs, such as software engineering agents, tend to be evaluated on larger models only, leaving the issue of improving the cost-benefit trade-off of such applications neglected. This paper proposes Atropos, a predictive early-termination analysis and hotswap technique that aims to improve the cost-benefit trade-off for LLM-based agents that use self-consistency. The core component of ATROPOS is a predictive model based on structural properties of LLM inferences: after merging multiple agentic inference paths into a graph representation, ATROPOS uses Graph Convolutional Network (GCN) to predict whether an ongoing inference will eventually succeed or not. If an agentic task instance running on the source LLM is predicted to fail, ATROPOS subsequently performs hotswapping, i.e., migrating the on-going inference context onto the more capable target LLM: this is feasible because LLM contexts are stateless. An empirical evaluation of ATROPOS using three recent LLM-based agents shows that ATROPOS can predict early termination of eventually failing inferences with the accuracy of 0.85 at the midpoint of the inference. Hotswapping LLMs for such inferences can convert up to 27.57% of them to be successful. Consequently, ATROPOS achieves 74.35% of the performance of closed LLMs with as low as only 23.9% of the cost.

Flow matching retains the generation quality of diffusion models while enabling substantially faster inference, making it a compelling paradigm for generative modeling. However, when applied to language modeling, it exhibits fundamental limitations in representing complex latent distributions with irregular geometries, such as anisotropy and multimodality. To address these challenges, we propose a mixture-of-experts flow matching (MoE-FM) framework, which captures complex global transport geometries in latent space by decomposing them into locally specialized vector fields. Building on MoE-FM, we develop a non-autoregressive (NAR) language modeling approach, named YAN, instantiated with both Transformer and Mamba architectures. Across multiple downstream tasks, YAN achieves generation quality on par with both autoregressive (AR) and diffusion-based NAR language models, while requiring as few as three sampling steps. This yields a $40\times$ speedup over AR baselines and up to a $10^3\times$ speedup over diffusion language models, demonstrating substantial efficiency advantages for language modeling.

Autoregressive decoding in Large Language Models (LLMs) generates one token per step, causing high inference latency. Speculative decoding (SD) mitigates this through a guess-and-verify strategy, but existing training-free variants face trade-offs: retrieval-based drafts break when no exact match exists, while logits-based drafts lack structural guidance. We propose $\textbf{RACER}$ ($\textbf{R}$etrieval-$\textbf{A}$ugmented $\textbf{C}$ont$\textbf{e}$xtual $\textbf{R}$apid Speculative Decoding), a lightweight and training-free method that integrates retrieved exact patterns with logit-driven future cues. This unification supplies both reliable anchors and flexible extrapolation, yielding richer speculative drafts. Experiments on Spec-Bench, HumanEval, and MGSM-ZH demonstrate that RACER consistently accelerates inference, achieving more than $2\times$ speedup over autoregressive decoding, and outperforms prior training-free methods, offering a scalable, plug-and-play solution for efficient LLM decoding. Our source code is available at $\href{https://github.com/hkr04/RACER}{https://github.com/hkr04/RACER}$.

An OS kernel that runs LLM inference internally can read the model's own next-token logit distribution before any text is generated, and act on it as a governance primitive. I present ProbeLogits, a kernel-level operation that performs a single forward pass and reads specific token logits to classify an agent's action as safe or dangerous, with zero learned parameters. Because the probe reads a logit from the same base model the agent already runs, it removes the second model a fine-tuned guard requires: the marginal cost of a safety check becomes a single logit read. I evaluate ProbeLogits on three base models (Qwen2.5-7B, Llama-3-8B, Mistral-7B) across three external benchmarks (HarmBench, XSTest, ToxicChat). On HarmBench non-copyright, all three reach a 97-99% block rate. On ToxicChat (n=1,000), ProbeLogits attains F1 parity-or-better against Llama Guard 3: Qwen2.5-7B Safe/Dangerous reaches F1=0.812 (+13.7 pp, bootstrap 95% CIs disjoint), Llama-3 matches within CI (+0.4 pp), and Mistral exceeds by +4.4 pp. Classification is a measured 2.4-3.4x faster than Llama Guard 3 (332-556 ms vs. 851-1,142 ms), because it reads a single logit position instead of generating tokens. A calibration strength alpha acts as a deployment-time policy knob rather than a learned hyperparameter, trading recall for precision per operation class. I implement ProbeLogits within Anima OS, a bare-metal x86-64 kernel written in ~285,000 lines of Rust. Because agent actions must pass through kernel-mediated host functions, enforcement operates below the WASM sandbox boundary, making it substantially harder to circumvent than application-layer classifiers.

Disaggregation maps parts of an AI workload to different types of GPUs, offering a path to utilize modern heterogeneous GPU clusters. However, existing solutions operate at a coarse granularity and are tightly coupled to specific model architectures, leaving much room for performance improvement. This paper presents Tessera, the first kernel disaggregation system to improve performance and cost efficiency on heterogeneous GPUs for large model inference. Our key insight is that kernels within a single application exhibit diverse resource demands, making them the most suitable granularity for aligning computation with hardware capabilities. Tessera integrates offline analysis with online adaptation by extracting precise inter-kernel dependencies from PTX to ensure correctness, overlapping communication with computation through a pipelined execution model, and employing workload-aware scheduling with lightweight runtime adaptation. Extensive evaluations across five heterogeneous GPUs and four model architectures, scaling up to 16 GPUs, show that Tessera improves serving throughput and cost efficiency by up to 2.3x and 1.6x, respectively, compared to existing disaggregation methods, while generalizing to model architectures where prior approaches do not apply. Surprisingly, a heterogeneous GPU pair under Tessera can even exceed the throughput of two homogeneous high-end GPUs at a lower cost.

Efficient context compression is critical for retrieval-augmented question answering in resource-constrained settings, where long retrieved contexts increase latency, memory use, and LLM reader cost. We propose a lightweight encoder-only framework for query-driven sentence pruning that preserves answer-critical evidence while aggressively reducing irrelevant context. Our method learns marginal contribution scores for sentences using counterfactual training signals and optimizes a contrastive ranking objective that separates critical evidence from noncritical context. Our approach scores all sentences from a single full-context encoding, enabling fast inference with low computational overhead. Experiments show that it maintains accuracy comparable to the strongest baseline while using 3.7$\times$ less peak memory and achieving nearly 3$\times$ lower compression latency, demonstrating an effective quality--efficiency trade-off for practical resource-constrained deployment.

Device-side Large Language Models (LLMs) have witnessed explosive growth, offering higher privacy and availability compared to cloud-side LLMs. During LLM inference, both model weights and user data are valuable, and attackers may even compromise the OS kernel to steal them. ARM TrustZone is the de facto hardware-based isolation technology on mobile devices, used to protect sensitive applications from a compromised OS. However, protecting LLM inference with TrustZone incurs significant overhead due to its inflexible isolation of memory and the NPU. To address these challenges, this paper introduces FlexServe, a fast and secure LLM serving system for mobile devices. It first introduces a Flexible Resource Isolation mechanism to construct Flexible Secure Memory (Flex-Mem) and Flexible Secure NPU (Flex-NPU). Both memory pages and the NPU can be efficiently switched between unprotected and protected modes. Based on these mechanisms, FlexServe designs a fast and secure LLM inference framework within TrustZone's secure world. The LLM-Aware Memory Management and Secure Inference Pipeline are introduced to accelerate inference. A Multi-Model Scheduler is proposed to optimize multi-model workflows. We implement a prototype of FlexServe and compare it with two TrustZone-based strawman designs. The results show that FlexServe achieves an average $10.05\times$ speedup in Time to First Token (TTFT) compared to the strawman, and an average $2.44\times$ TTFT speedup compared to an optimized strawman with pipeline and secure NPU enabled. For multi-model agent workflows, the end-to-end speedup is up to $24.30\times$ and $4.05\times$ compared to the strawman and optimized strawman, respectively.

Speculative decoding (SD) accelerates large language model inference by using a smaller draft model to propose draft tokens that are subsequently verified by a larger target model. However, the performance of standard SD is often limited by the strictly sequential execution of these drafting and verification stages. To address this, this paper proposes MineDraft, a batch parallel speculative decoding (PSD) framework designed to effectively hide drafting latency by overlapping it with verification. Our theoretical analysis shows that PSD is substantially more efficient than standard SD. MineDraft realizes the PSD through a novel batch-parallel design that maintains two batches of requests, overlapping drafting for one batch with verification for the other. Our experimental results show significant improvements of \alg{} in both throughput (up to 75%) and end-to-end latency (up to 39%) over standard SD. Furthermore, we have implemented MineDraft as a plugin for vLLM, demonstrating its practicality for production-ready inference systems.

The rapid growth of large language models (LLMs) with diverse capabilities, costs, and domains has created a critical need for intelligent model selection at inference time. While smaller models suffice for routine queries, complex tasks demand more capable models. However, static model deployment does not account for the complexity and domain of incoming queries, leading to suboptimal performance and increased costs. Dynamic routing systems that adaptively select models based on query characteristics have emerged as a solution to this challenge. This survey provides a systematic analysis of multi-LLM routing and cascading approaches, focusing on systems that route queries across a pool of independently trained LLMs at inference time. We cover diverse routing paradigms, including query difficulty, human preferences, clustering, uncertainty quantification, reinforcement learning, multimodality, and cascading. For each paradigm, we analyze representative methods and examine key trade-offs. Beyond taxonomy, we introduce a conceptual framework that characterizes routing systems along three dimensions: when decisions are made, what information is used, and how they are computed. This perspective highlights that practical systems are often compositional, integrating multiple paradigms under operational constraints. Our analysis demonstrates that effective multi-LLM routing requires balancing competing objectives. Choosing the optimal routing strategy depends on deployment and computational constraints. Well-designed routing systems can outperform even the most powerful individual models by strategically leveraging specialized capabilities across models while maximizing efficiency gains. Meanwhile, open challenges remain in developing and evaluating routing mechanisms that generalize across diverse architectures, modalities, and applications.

Sparse Mixture-of-Experts (SMoE) architectures are increasingly used to scale large language models efficiently, delivering strong accuracy under fixed compute budgets. However, SMoE models often suffer from severe load imbalance across experts, where a small subset of experts receives most tokens while others are underutilized. Prior work has focused mainly on training-time solutions such as routing regularization or auxiliary losses, leaving inference-time behavior, which is critical for deployment, less explored. We present a systematic analysis of expert routing during inference and identify three findings: (i) load imbalance persists and worsens with larger batch sizes, (ii) selection frequency does not reliably reflect expert importance, and (iii) overall expert workload and importance can be estimated using a small calibration set. These insights motivate inference-time mechanisms that rebalance workloads without retraining or router modification. We propose Replicate-and-Quantize (R&Q), a training-free and near-lossless framework for dynamic workload rebalancing. In each layer, heavy-hitter experts are replicated to increase parallel capacity, while less critical experts and replicas are quantized to remain within the original memory budget. We also introduce a Load-Imbalance Score (LIS) to measure routing skew by comparing heavy-hitter load to an equal allocation baseline. Experiments across representative SMoE models and benchmarks show up to 1.4x reduction in imbalance with accuracy maintained within +/-0.6%, enabling more predictable and efficient inference.

Large language models(LLMs) are now used to power complex multi-turn agentic workflows. Existing systems run agentic inference by loosely assembling isolated components: an LLM inference engine (e.g., vLLM) and a tool orchestrator (e.g., Kubernetes). Although agentic workflows involve multiple LLM and tool requests, these systems schedule and allocate resources separately on a per-request basis, without end-to-end knowledge of the workflow. This leads to sub-optimal management of KV cache and tool execution environments. To address the challenges, we propose ThunderAgent, a fast, simple, and program-aware agentic inference system. We first abstract agentic workflows as LLM Programs, enabling a unified view of heterogeneous resources, including KV caches, system states, and external tool assets such as disk memory and network ports. Built upon this abstraction, ThunderAgent introduces a program-aware scheduler and a tool resource manager designed to maximize KV cache hit rates, mitigate memory imbalances, and enable asynchronous environment preparation. Evaluations across coding, routing, and scientific discovery agents demonstrate that ThunderAgent achieves 1.5-3.6x throughput improvements in serving, 1.8-3.9x in RL rollout, and up to 4.2x disk memory savings compared to state-of-the-art inference systems. To facilitate reproducibility and support future development, we open-source the system implementations of the whole ThunderAgent at: https://github.com/Agentic-Kinetics/ThunderAgent.

The growing size of Large Language Models (LLMs) makes efficient inference challenging, primarily due to the memory demands of the autoregressive Key-Value (KV) cache. Existing eviction or compression methods reduce cost but rely on heuristics, such as recency or past attention scores, which serve only as indirect proxies for a token's future utility and introduce computational overhead. We reframe KV cache eviction as a reinforcement learning (RL) problem: learning to rank tokens by their predicted usefulness for future decoding. To this end, we introduce KV Policy (KVP), a framework of lightweight per-head RL agents trained on pre-computed generation traces using only key and value vectors. Each agent learns a specialized eviction policy guided by a holistic reward, derived from future utility, that evaluates the quality of the ranking across all cache budgets, requiring no modifications to the underlying LLM or additional inference. Evaluated across two model families on the long-context benchmark RULER (up to 128K tokens) and the multi-turn dialogue benchmark OASST2-4k, KVP significantly outperforms strong baselines. Zero-shot tests on standard downstream tasks (BoolQ, LongBench passage retrieval, GovReport) further show that KVP generalizes beyond its training distribution and to considerably longer sequence lengths. These results demonstrate that learning to predict future token utility is a powerful and scalable paradigm for adaptive KV cache management.

Large Language Models (LLMs) achieve strong performance across many tasks but suffer from high inference latency due to autoregressive decoding. The issue is exacerbated in Large Reasoning Models (LRMs), which generate lengthy chains of thought. While speculative decoding accelerates inference by drafting and verifying multiple tokens in parallel, existing methods operate at the token level and ignore semantic equivalence (i.e., different token sequences expressing the same meaning), leading to inefficient rejections. We propose SemanticSpec, a semantic-aware speculative decoding framework that verifies entire semantic sequences instead of tokens. SemanticSpec introduces a semantic probability estimation mechanism that probes the model's internal hidden states to assess the likelihood of generating sequences with specific meanings. Experiments on four benchmarks show that SemanticSpec achieves up to 2.7x speedup on DeepSeekR1-32B and 2.1x on QwQ-32B, consistently outperforming token-level and sequence-level baselines in both efficiency and effectiveness.

The performance of large language models (LLMs) on verifiable tasks is usually measured by pass@k, the probability of answering a question correctly at least once in k trials. At a fixed budget, a more suitable metric is coverage@cost, the average number of unique questions answered as a function of the total number of attempts. We connect the two metrics and show that the empirically-observed power-law behavior in pass@k leads to a sublinear growth of the coverage@cost (diminishing returns). To solve this problem, we propose Reset-and-Discard (ReD), a query method of LLMs that increases coverage@cost for a given budget, regardless of the pass@k form. Moreover, given a pass@k, we can quantitatively predict the savings in the total number of attempts using ReD. If pass@k is not available for the model, ReD can infer its power-law exponent. Experiments on three LLMs across coding (HumanEval), math (GSM8K), and reasoning (MMLU-Pro) benchmarks demonstrate that ReD substantially reduces the required attempts, tokens, and USD cost to reach a desired coverage, while also offering an efficient way to measure inference power-laws. ReD's advantage is maintained for imperfect verifiers and outperforms the tested allocation baselines.

Diffusion language models (DLMs) generate text through iterative denoising, but inference requires full-sequence attention at every iteration, resulting in substantial redundant computation on masked tokens. Block-wise diffusion can reduce this cost, yet it typically relies on retraining and constrained update orders, limiting its direct applicability to pretrained DLMs. Our token-level analysis reveals pronounced structural locality in DLM inference. Decoding is driven by a small set of prefix-localized active tokens; the influence of distant undecoded context diminishes rapidly, and decoded tokens exhibit stage-wise temporal stability, enabling reuse of intermediate representations except for a brief post-decode transient. Motivated by these observations, we propose \textbf{\placeholder}\footnote{The source code is available at https://github.com/vhicrgit/Window-Diffusion.}, a window-based token pruning and caching method for inference. We maintain a local computation window that slides rightward as denoising progresses, and partition undecoded tokens into: (i) \textit{active tokens} that are computed online, (ii) \textit{buffer tokens} whose KV states are cached and periodically refreshed, and (iii) \textit{far-field tokens} that are pruned outside the window. Computation is restricted to active and buffer tokens within the window, while far-field tokens are omitted at each stage. Experiments on LLaDA and Dream show that, under matched compute budgets, our method achieves up to $99\times$ inference speedup while largely preserving generation performance.

Diffusion Large Language Models (dLLMs) offer a compelling paradigm for natural language generation, leveraging parallel decoding and bidirectional attention to achieve superior global coherence compared to autoregressive models. While recent works have accelerated inference via KV cache reuse or heuristic decoding, they overlook the intrinsic inefficiencies within the block-wise diffusion process. Specifically, they suffer from spatial redundancy by modeling informative-sparse suffix regions uniformly and temporal inefficiency by applying fixed denoising schedules across all the decoding process. To address this, we propose Streaming-dLLM, a training-free framework that streamlines inference across both spatial and temporal dimensions. Spatially, we introduce attenuation guided suffix modeling to approximate the full context by pruning redundant mask tokens. Temporally, we employ a dynamic confidence aware strategy with an early exit mechanism, allowing the model to skip unnecessary iterations for converged tokens. Extensive experiments show that Streaming-dLLM achieves up to 68.2X speedup while maintaining generation quality, highlighting its effectiveness in diffusion decoding. The code is available at https://github.com/xiaoshideta/Streaming-dLLM.

The emergence of fine-grained numerical formats like NVFP4 presents new opportunities for efficient Large Language Model (LLM) inference. However, it is difficult to adapt existing Post-Training Quantization (PTQ) strategies to these formats: rotation-based methods compromise fine-grained block isolation; smoothing techniques struggle with significant 4-bit quantization errors; and mixed-precision approaches often conflict with hardware constraints on unified-precision computation. To address these challenges, we propose ARCQuant, a framework that boosts NVFP4 performance via Augmented Residual Channels. Distinct from methods that compromise block isolation or hardware uniformity, ARCQuant maintains a strictly unified NVFP4 format by augmenting the activation matrix with quantized residual channels. This design integrates the error compensation process directly into the matrix reduction dimension, enabling the use of standard, highly optimized GEMM kernels with minimal overhead. Theoretical analysis confirms that the worst-case error bound of our dual-stage NVFP4 quantization is comparable to that of standard 8-bit formats such as MXFP8. Extensive experiments on LLaMA and Qwen models demonstrate that ARCQuant achieves state-of-the-art accuracy, comparable to full-precision baselines in perplexity and downstream tasks. Furthermore, deployment on RTX 5090 and RTX PRO 6000 GPUs confirms practical benefits, achieving up to 3x speedup over FP16. Our code is available at https://github.com/actypedef/ARCQuant.

Ambiguity-bearing inputs reach downstream systems through interfaces that favor a single resolved response before later context arrives. Even when alternatives are externalized, their representation and relative activation depend on the update rule. We address this state-maintenance problem within Non-Resolution Reasoning (NRR) by specifying a typed external text-to-state interface and explicit state-update contract. A mapping (phi: T -> S) constructs typed (v,c,w,m) records; declared operators make record carry-forward inspectable, while a record-weight entropy criterion separately tests normalized-weight concentration. The contract covers dampening, calibration, deferred resolution, contradiction-preserving integration, and temporal persistence. Across 580 constructed states/pairs, an executable suite performs 2,740 operator-state measurements. Tested non-violating transitions and calibration/identity checks show 0% record-weight entropy violations; a uniform-subtraction comparison violates the criterion at 1.7%, 6.1%, and 17.8% as subtraction increases. A separate 68-input construction audit finds multiple positive-weight typed records (mean record-weight entropy H = 1.087 bits) under the reported rule-based and archived LLM-assisted procedures, including a Japanese marker-set instantiation. The entropy criterion does not by itself certify record identity or cardinality, semantic adequacy, or end-to-end behavioral improvement. The repository provides deterministic reruns, archived prompt/output artifacts, transcript audit, and a fixed 18-set LLM-case sanity rerun. Phi therefore turns retained-state maintenance into a typed, executable, and falsifiable interface: record carry-forward is inspectable in declared operators, and normalized-weight concentration is separately testable before downstream commitment.

Attention is the dominant source of latency during long-context LLM inference, an increasingly popular workload with reasoning models and RAG. We propose Kascade, a training-free sparse attention method that leverages known observations such as 1) post-softmax attention is intrinsically sparse, and 2) the identity of high-weight keys is stable across nearby layers. Kascade computes exact Top-k indices in a small set of anchor layers, then reuses those indices in intermediate reuse layers. The anchor layers are selected algorithmically, via a dynamic-programming objective that maximizes cross-layer similarity over a development set, allowing easy deployment across models. The method incorporates efficient implementation constraints (e.g. tile-level operations), across both prefill and decode attention. The Top-k selection and reuse in Kascade is head-aware and we show in our experiments that this is critical for high accuracy. Kascade achieves up to 4.1x speedup in decode attention and 2.2x speedup in prefill attention over FlashAttention-3 baseline on H100 GPUs while closely matching dense attention accuracy on long-context benchmarks such as LongBench and AIME-24.

Many LLM applications require only narrow capabilities, yet standard post-training quantization (PTQ) methods allocate precision without considering the target task. This can waste bits on layers that are less relevant to the task signal while over-compressing layers that are critical for downstream behavior. We propose Task-Aware Quantization (TAQ), a training-free, weight-only mixed-precision PTQ framework that uses a small set of unlabeled task calibration prompts to allocate higher precision to task-relevant transformer layers under a fixed bit budget. TAQ estimates layer importance from hidden representations and output sensitivity, and we instantiate it with three scoring rules: TAQ-IS, based on activation information and stability; TAQ-KL, based on output-distribution sensitivity under a quantization-noise proxy; and TAQ-O, a label-informed oracle diagnostic for analyzing layer sensitivity. Across several benchmarks, TAQ outperforms task-agnostic baselines such in most settings, with especially strong gains in the accuracy--memory ratio. We further validate that these gains translate to real deployment behavior through hardware throughput and latency measurements, and analyze calibration robustness and residual-stream error propagation. Overall, TAQ turns mixed-precision PTQ from a model-centric compression step into a task-conditioned precision-allocation problem. A reference implementation is available at \href{https://anonymous.4open.science/r/TAQ-9217/README.md}{\includegraphics[height=1em]{imgs/github-mark.png}}.

The substantial inference costs of attention in transformers motivated the development of efficient sequence mixers: namely sparse and sliding window attention, convolutions and linear attention. Although these approaches result in impressive reductions in inference costs, they often trade-off with quality, specifically in-context recall. Apriori fixing this quality-cost tradeoff at training time means being suboptimal from the get-go: some downstream applications might fundamentally require more memory for in-context recall, while other tasks may require lower latency and memory. We propose a conceptually simple meta-sequence mixer with inference-cost controllability: the Compress & Attend Transformer (CAT). CAT decodes chunks of tokens by attending to compressed chunks of the sequence so far. Both compression and decoding can use any existing sequence mixer. Decoding from the compressed sequence yields compute and memory savings, with chunk size setting the operating point on the quality-cost trade-off. Importantly, training CAT across multiple chunk sizes at once unlocks test-time control of this trade-off without any retraining, all in a single model. Instantiated with the most basic choice, dense attention as the mixer, CAT surprisingly suffices to match 10 popular and diverse efficient models (linear, hybrids, sparse) on real-world long-context recall at comparable inference costs, all from a single trained model. CAT further performs competitively on long-context understanding benchmarks while providing 1.4-3.7x higher generation throughput than a dense transformer. Code is at: https://github.com/rajesh-lab/cat-transformer

Deploying large language models (LLMs) is challenging due to their massive parameters and high computational costs. Ultra low-bit quantization can significantly reduce storage and accelerate inference, but extreme compression (i.e., mean bit-width <= 2) often leads to severe performance degradation. To address this, we propose Squeeze10-LLM, effectively "squeezing" 16-bit LLMs' weights by 10 times. Specifically, Squeeze10-LLM is a staged mixed-precision post-training quantization (PTQ) framework and achieves an average of 1.6 bits per weight by quantizing 80% of the weights to 1 bit and 20% to 4 bits. We introduce Squeeze10LLM with two key innovations: Post-Binarization Activation Robustness (PBAR) and Full Information Activation Supervision (FIAS). PBAR is a refined weight significance metric that accounts for the impact of quantization on activations, improving accuracy in low-bit settings. FIAS is a strategy that preserves full activation information during quantization to mitigate cumulative error propagation across layers. Experiments on LLaMA and LLaMA2 show that Squeeze10-LLM achieves state-of-the-art performance for sub-2bit weight-only quantization, improving average accuracy from 43% to 56% on six zero-shot classification tasks--a significant boost over existing PTQ methods.

Efficiently scaling Large Language Models (LLMs) necessitates exploring alternatives to dominant autoregressive (AR) methods, with Masked Diffusion Models (MDMs) emerging as candidates. However, comparing AR (typically decoder-only) and MDM (often encoder-only) paradigms is confounded by differing architectures, obscuring true algorithmic and efficiency trade-offs. This research decouples these factors by evaluating MDMs within a decoder-only framework to: (1) Equitably compare MDM (as Any-Order AR) and standard AR paradigms through discrepancies on orders. (2) Investigate MDM architectural impacts on computational efficiency. We show decoder-only MDMs, despite a larger modeling space, can achieve significant inference speedups ($\sim25\times$) and comparable perplexity with techniques like temperature annealing, offering a path to reduced inference compute. This work provides insights for developing more computationally efficient foundation models by disentangling core modeling choices from architectural influences. Code is available at https://github.com/scxue/AO-GPT-MDM.

Large language models (LLMs) require substantial compute, and thus energy, at inference time. While quantizing weights and activations is effective at improving efficiency, naive quantization of LLMs can significantly degrade performance due to large magnitude outliers. This paper describes FPTQuant, which introduces three novel, lightweight, and expressive function-preserving transforms (FPTs) to facilitate quantization of transformers: (1) a mergeable pre-RoPE transform for queries and keys, (2) a mergeable transform for values, and (3) a cheap, dynamic per-token scaling transform. By leveraging the equivariances and independencies inherent to canonical transformer operation, we designed these FPTs to maintain the model's function while shaping the intermediate activation distributions to be more quantization friendly. FPTQuant requires no custom kernels and adds virtually no overhead during inference. The FPTs are trained both locally to reduce outliers, and end-to-end such that the outputs of the quantized and full-precision models match. FPTQuant enables static INT4 quantization with minimal overhead and shows SOTA speed-up of up to 3.9X over FP. Empirically, FPTQuant has an excellent accuracy-speed trade-off -- it is performing on par or exceeding most prior work and only shows slightly lower accuracy compared to a method that is up to 29% slower.

Large Language Model (LLM) inference is typically memory-intensive, especially when processing large batch sizes and long sequences, due to the large size of key-value (KV) cache. Vector Quantization (VQ) is recently adopted to alleviate this issue, but we find that the existing approach is susceptible to distribution shift due to its reliance on calibration datasets. To address this limitation, we introduce NSNQuant, a calibration-free Vector Quantization (VQ) technique designed for low-bit compression of the KV cache. By applying a three-step transformation-1) a token-wise normalization (Normalize), 2) a channel-wise centering (Shift), and 3) a second token-wise normalization (Normalize)-with Hadamard transform, NSNQuant effectively aligns the token distribution with the standard normal distribution. This alignment enables robust, calibration-free vector quantization using a single reusable codebook. Extensive experiments show that NSNQuant consistently outperforms prior methods in both 1-bit and 2-bit settings, offering strong generalization and up to 3$\times$ throughput gain over full-precision baselines. Code is available at https://github.com/DHdroid/NSNQuant.

On-device inference offers privacy, offline use, and instant response, but consumer hardware restricts large language models (LLMs) to low throughput and capability. To overcome this challenge, we present prima.cpp, a distributed on-device inference system that runs 30-70B LLMs on consumer home clusters with mixed CPUs/GPUs, insufficient RAM/VRAM, slow disks, Wi-Fi links, and heterogeneous OSs. We introduce pipelined-ring parallelism (PRP) to overlap disk I/O with compute and communication, and address the prefetch-release conflict in mmap-based offloading. We further propose Halda, a heterogeneity-aware scheduler that co-optimizes per-device CPU/GPU workloads and device selection under RAM/VRAM constraints. On four consumer home devices, a 70B model reaches 674 ms/token TPOT with <6% memory pressure, and a 32B model with speculative decoding achieves 26 tokens/s. Compared with llama.cpp, exo, and dllama, our proposed prima.cpp achieves 5-17 times lower TPOT, supports fine-grained model sizes from 8B to 70B, ensures broader cross-OS and quantization compatibility, and remains OOM-free, while also being Wi-Fi tolerant, privacy-preserving, and hardware-independent. The code is available at https://github.com/OpenCPIL/prima.cpp.

To reduce memory consumption during LLM inference, a handful of methods have been proposed for KV cache pruning. While these techniques can accomplish lossless memory reduction on many datasets, they often hinge on an under-emphasized condition: an input/domain-specific threshold for KV cache budget needs to be pre-determined to achieve the optimal performance. However, such input-sensitive design may be considerably limited in real-world scenarios, as open-domain inputs span diverse domains, lengths and difficulty levels, without clear boundaries for threshold selection. As a result, the dependence of such input-sensitive threshold can be a fundamental limitation that causes large degradation on arbitrary inputs. In this work, we propose a new objective that lifts the threshold constraints for robust KV compression, advocating for "threshold-free" methods that adaptively adjust budget allocation while preserving full-cache performance. We then propose a novel method, ReFreeKV, serving as the first instantiation of this objective. Extensive experiments across 13 datasets with diverse context lengths, task types, and model sizes demonstrate its efficacy and efficiency. Our code is publicly released at https://github.com/Patrick-Ni/ReFreeKV.

Long-context understanding remains challenging for LLMs due to limited context windows. This paper introduces Long Input Fine-Tuning (LIFT), a framework that improves the long-context performance of arbitrary short-context LLMs by dynamically adapting their parameters to each long input. Instead of endlessly extending context windows to fit longer inputs in context, LIFT stores and absorbs the input in parameters. By fine-tuning long inputs into parameters, LIFT enables short-context LLMs to answer questions even when required information is absent from the inference context, avoiding the quadratic input-length complexity of standard long-context models. Rather than simple continued pretraining on new long contexts, LIFT uses carefully designed LLM-generated synthetic tasks to enhance comprehension beyond memorization. To offset fine-tuning overhead, we design a highly optimized pipeline that reduces Time to First Token (TTFT) to under 10 seconds for 8k context. We further analyze LIFT's strengths and limitations, discuss large-scale deployment feasibility, and highlight future research directions. Implementation is open-sourced at https://github.com/MuLabPKU/LIFT.

LLMs are increasingly used world-wide from daily tasks to agentic systems and data analytics, requiring significant GPU resources. While LLM inference systems are capable of serving millions of requests from multiple users, they often lack theoretical models to determine whether they achieve the performance upper bounds of underlying hardware resources. Beyond online workload serving, merely analyzing existing systems-or developing yet another one-is both GPU-intensive and labor-intensive. This paper provides a comprehensive survey of LLM inference systems, focusing on their cache management policies and availability. We then show that simulations can be an effective tool to save GPU hours in the development and analysis phase of inference systems, revealing useful insights for developing better inference techniques, unlike how existing studies used simulations to find the best parameters inside a given system. Finally, we provide theoretical tools to estimate the optimal performance and formulate new ideas. Based on the theoretical analysis, especially on the cache management in LLM inference, we propose a simple yet effective cache replacement policy that can be easily plugged into existing preemptive schedulers and systems. We show that such a simple policy inspired from database systems can substantially save GPU hours in actual inference systems on online workloads. We share our experience submitting a journal paper to a database venue in November 2025 for anyone considering a similar path.

Retrieval-Augmented Language Modeling (RALM) by integrating large language models (LLM) with relevant documents from an external corpus is a proven method for enabling the LLM to generate information beyond the scope of its pre-training corpus. Previous work utilizing retrieved content by simply prepending it to the input poses a high runtime issue, which degrades the inference efficiency of the LLMs because they fail to use the Key-Value (KV) cache efficiently. In this paper, we propose FlashBack, a modular RALM designed to improve the inference efficiency of RALM with appending context pattern while maintaining decent performance after fine-tuning by Low-Rank Adaption. FlashBack appends retrieved documents at the end of the context for efficiently utilizing the KV cache instead of prepending them. And we introduce Marking Token as two special prompt tokens for marking the boundary of the appending context during fine-tuning. Our experiments on testing generation quality show that FlashBack can remain decent generation quality in perplexity. And the inference speed of FlashBack is up to $4\times$ faster than the prepending counterpart on a 7B LLM (Llama 2) in the runtime test. Via bypassing unnecessary re-computation, it demonstrates an advancement by achieving significantly faster inference speed, and this heightened efficiency will substantially reduce inferential cost.

Long-context inference with large language models (LLMs) is costly: self-attention during prefill scales quadratically with sequence length, and the key-value (KV) cache grows with the number of processed tokens. Larger context windows also do not ensure reliable evidence use. Context compression reduces this cost, but many soft-compression methods use LLMs as compressors and rely on compact memory tokens both to preserve information and to condition the decoder. We propose SeDeM, a selective decompression framework that decouples compact memory storage from decoder conditioning. SeDeM stores context as compact hidden-state memory blocks, selects query-relevant blocks, and decompresses only the selected blocks for decoder conditioning. Thus, the decoder avoids both full-context processing and direct generation from highly compressed memory slots. On four long-context QA benchmarks, SeDeM achieves higher QA scores than the compression baselines in our main comparison in both 1B and 3B same-backbone settings, and with the 3B backbone exceeds full-context fine-tuning on three datasets. SeDeM also provides favorable quality--efficiency trade-offs, achieving 1.74--2.46$\times$ lower online time-to-first-token and 1.08--1.10$\times$ higher autoregressive decoding throughput relative to ICAE while maintaining strong answer quality.

In this study, we identify depth-dependent prefix redundancy in final-readout LLM embedding models, notably across representative backbones including Qwen3-Embedding and Qwen3-VL-Embedding. We find that removing prefix states is substantially more damaging in shallow layers than at greater depth, showing that prefix states become increasingly compressible as the prefix and readout states propagate through the network. To this end, we introduce FastE, a training-free, plug-and-play method. FastE uses a shared fixed threshold on batch-mean readout-prefix alignment as a lightweight online heuristic for selecting when compression occurs, and ranks prefix states by the attention scores they receive from the readout position to determine which states are retained in subsequent layers. Our evaluations demonstrate FastE's ability to substantially reduce computational costs: on NarrativeQA with Qwen3-Embedding-0.6B, it reduces decoder-backbone FLOPs by 40.11% while retaining 99.53% of Full Forward nDCG@10. Across five text embedding benchmarks, two backbone scales, and three cross-modal retrieval tasks, the quality-efficiency trade-off is directly customizable through the maximum removal ratio without retraining. We believe FastE offers practical value for scalable embedding generation in retrieval, indexing, clustering, and multimodal representation systems.