Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning

Large language models achieve superior performance on tasks that require extended reasoning, but long chains of thought make the KV cache a severe memory bottleneck. Existing KV cache compression methods share one paradigm: score each cached token by some estimate of how much it will matter later, and keep the top-scoring ones. We show that the selection signal contributes almost nothing. Random Attention keeps the prompt and evicts uniformly at random within each attention head, computing no score at all; across four models and six reasoning tasks it matches the strongest prior evictor while serving 32-43% higher throughput than it in vLLM deployment. Controlled experiments explain this by showing that 1) the prompt is the fragile part of the cache, and most of the gap between selectors is just whether their selection signal happened to keep it; 2) the reasoning trace protects itself against eviction with redundancy at two levels, in the text (the model restates what it still needs as it works) and across attention heads (each keeps its own copy of the trace), so once the prompt is safe, a random draw retains enough copies of what the model still needs, and no score is required to pick them. Our code is publicly available at https://github.com/SalesforceAIResearch/Random-Attention.

LeanStream: A Speculate-and-Refine Streaming Framework for Efficient on-Device LLM Inference

On-device LLM inference is attractive for privacy and responsiveness, but remains challenging on mobile and embedded devices because model weights far exceed available DRAM. Prior systems exploit activation sparsity and offload weights to SSD or flash storage, but face a fundamental systems trade-off: accurate sparse execution decisions require the latest context, whereas efficient computation-I/O overlap requires early prediction. As a result, existing designs either serialize execution or incur redundant weight fetches, extra computation, and large cache overheads. We present LeanStream, a streaming speculate-and-refine framework for efficient on-device LLM inference. LeanStream progressively refines computation, loading, and cache-retention priorities using partial GPU results, enabling fine-grained overlap between GPU execution and storage I/O. We implement LeanStream on both mobile and embedded platforms. Compared with prior on-device LLM inference systems, LeanStream reduces memory usage by 4.8$\times$ to 7.5$\times$ at the best throughput achieved by prior work, while further improving token generation throughput by 1.6$\times$ to 2.1$\times$.

CRISP: Cliff-awaRe Input-adaptive Sparse Prefilling with Structural-Mass-Motivated Routing

The attention prefilling phase of long-context LLM inference scales quadratically, making self-attention a severe computational bottleneck. Traditional sparse attention methods mitigate this through fixed patterns or offline profiling, but lack the flexibility to adapt to input-dependent attention structure. Recent dynamic methods address this by routing heads to sparse patterns in real-time, but rely on indirect routing proxies with overhead and budget allocation mechanisms that overlook the post-softmax mass hierarchy. We present CRISP (Cliff-awaRe Input-adaptive Sparse Prefilling), which identifies and addresses two structural challenges in this dynamic routing paradigm. First, we show that the routing decision can be read directly off the structure of the proxy attention map. We replace the Jensen-Shannon Divergence (JSD) routing with C_struct, a structural proxy that measures mass at Vertical-Slash compatible positions and reproduces JSD's routing decisions while eliminating both the pooled matmul and subsequent KL divergence overhead. Second, we formalize the post-softmax mass cliff and demonstrate theoretically that strictly cumulative coverage thresholds accumulate O(n) background noise at long contexts. CRISP navigates this via a sink-aware threshold grounded in the noise floor. Empirically, across InfiniteBench, RULER and LongBench on two model families, CRISP is the strongest sparse method overall and matches or exceeds exact dense attention on retrieval-heavy benchmarks, recovering up to +28.0 pp on retrieval tasks over baselines and achieving up to a 5.30x attention speedup at 512k tokens, driven primarily by our O(n) noise elimination during selection while preserving structural integrity.

Residual Sparsification via Output Importance for Compressing Mixture-of-Experts LLMs

Mixture-of-experts (MoE) architectures scale large language models efficiently, but they demand massive GPU memory. To cope with such demand, models are commonly compressed to reduce their memory footprint. Residual sparsification is a representative compression technique that decomposes each projection matrix of an expert into a shared base matrix and per-expert residual matrix, and then compresses the residuals. Existing sparsification methods compress each residual matrix independently by minimizing its compression error, thereby minimizing the error of each projection matrix. However, our analysis shows that this objective is misaligned with preserving model accuracy after compression. In an expert, the final output is produced through computations coupled across multiple projections and hidden representations. Therefore, even small errors in individual matrices can propagate through hidden representations and projection interactions, leading to large expert output errors and accuracy degradation. To address this misalignment, we propose PARSER, a new residual sparsification method that shifts the compression objective from minimizing isolated matrix errors to preserving the expert output error. PARSER achieves this by introducing output importance, which measures the actual contribution to the expert output error. Our experiments show that, compared with existing methods, PARSER narrows the accuracy gap to the uncompressed model by 1.41$\times$ on Qwen and 1.44$\times$ on DeepSeek, while achieving the same peak memory reduction.

HBQ: Hierarchical Scaling Block Quantization with Hardware-Efficiency-Aware Design for Accurate LLM Inference

Block Quantization (BQ) is a promising approach for efficient deployment of large language models (LLMs), enabling low-precision computation with controlled accuracy degradation. Compared to scalar weight-only quantization (WoQ), BQ quantizes both weight and activation, offering higher hardware efficiency and end-to-end inference on a unified datapath, but its design space, spanning bit-width, block size, scaling, and numeric formats, remains underexplored. We provide hardware/benchmark results through design space exploration (DSE). We find that increasing block size improves hardware efficiency by amortizing dequantization and accumulation costs, but degrades accuracy. This trade-off limits conventional BQ methods. Motivated by this insight, we propose Hierarchical Block Quantization (HBQ). Unlike prior methods [1], [2], which use small blocks and conventional Power-of-Two (PoT) or integer-based scaling, HBQ uses large blocks to maximize efficiency and introduces low-overhead significand (SIG) scaling for second-level quantization. By allocating quantization levels effectively and accounting for distinct activation and weight distributions, SIG scaling compensates for large-block errors more effectively than prior PoT and INT schemes. HBQ-A (accurate) achieves W4A16-level accuracy using only W4A5 while requiring less silicon area than NVFP4. HBQ-E (efficient) further reduces hardware cost by 17% while maintaining higher accuracy than all existing BQ methods. We implemented a 28nm ASIC accelerator applying HBQ to weights, activations, and KV cache, and integrated a novel partial-sum BQ scheme to further reduce EMA energy. Compared to state-of-the-art WoQ, HBQ delivers $2.3\times$/$4.6\times$ higher area/energy efficiency at the same accuracy level; $1.6$--$3.3\times$ system energy reduction and $1.5$--$3.0\times$ speedup over prior BQ methods while providing best accuracy.

Budget-Aware Compression Pipeline for Single-GPU LLM Inference: Methods, Trade-offs, and Coupling Effects

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.

OCGQuant: Outlier-Companion Grouping for NVFP4 Quantization

NVFP4 is an efficient microscaling format for low-bit inference, but activation outliers can still degrade quantization accuracy within NVFP4 blocks. Within each quantization block, large activations can dominate the block scale, increasing the quantization error of the remaining values sharing the same scale. Existing post-training quantization (PTQ) methods mitigate outlier errors through strategies such as mixed precision, rotation, or residual compensation, but these approaches are either not specifically tailored to NVFP4 or introduce additional computation. In this work, we revisit NVFP4 from a channel-grouping perspective and define the reducible error incurred by remaining block values under the scale set by the block maximum as Collateral Quantization Error. Based on this insight, we propose OCGQuant, a post-training quantization method centered on Outlier-Companion Grouping (OCG), which adaptively pairs outlier channels with low-magnitude companion channels to improve NVFP4 activation block composition. Experiments on Llama3 and Qwen3 show that OCGQuant achieves the lowest WikiText-2 perplexity and highest average downstream accuracy among evaluated PTQ methods, while maintaining prefill speedup close to RTN and matching its peak decoding memory. Code is available at https://github.com/Eshamont/OCGQuant.

A rigor-matched audit of periodic-step layer skipping for efficient llm inference: conflayers versus swift, with a supplemental analysis of trained routing alternatives

Layer-skipping methods for efficient LLM inference decide, at some granularity, which transformer layers to execute for a given input. We present a rigor-matched, three-seed audit of two periodic-step, search-based methods that make this decision online at inference time and re-evaluate it every few generation steps: a confidence-gated early-exit baseline (ConfLayers) and genuine self-speculative decoding (SWIFT, Xia et al. 2024), together with vanilla autoregressive decoding, across two model scales (Qwen2.5-0.5B and Qwen2.5-1.5B) and two tasks (GSM8K reasoning and CNN/DailyMail summarization). SWIFT is the strongest method on accuracy in three of four cells; ConfLayers is dominated everywhere, with particularly large deficits on GSM8K at 1.5B. Once online-search overhead is separated from pure inference cost, SWIFT's true inference speed is faster than ConfLayers's in all four cells (5-21%), reversing the naive wall-clock ranking in three of them. ConfLayers's search overhead is small and stable (1-2% of cost), while SWIFT's is larger and more variable (up to 28.7%). We additionally examine two trained-routing methods, LayerRoute (Sikdar, 2026) and LayerDrop (Fan et al. 2020), as a supplemental analysis because they operate at coarser decision granularities. Under a verified protocol with genuine per-input gating, a genuine full-model baseline, and genuine inference-time compute skipping, both show modest speedups (1.08-1.33x) but accuracy well below the periodic-step methods, including a near-total collapse for LayerRoute on GSM8K at 1.5B (0.003 mean exact-match across three seeds). We release the full audit protocol as a template for rigor-matched efficiency comparisons.

H-Scale: Hessian-Guided Scale Refinement for NVFP4 Sub-Byte LLM Inference

The NVIDIA Blackwell architecture, with native support for the ultra-fine-grained NVFP4 format, opens new opportunities for accelerating large language model (LLM) inference. NVFP4's micro-block design, such as a group size of 16, offers strong representational flexibility for capturing local weight distributions and isolating outliers, but it also introduces a large and highly sensitive space of per-group scaling factors. Existing post-training quantization (PTQ) methods primarily focus on refining quantized weight values, leaving this scale-selection step underexplored. To address this gap, we propose \textbf{H-Scale}, a lightweight post-processing method for NVFP4 per-group scale refinement. Instead of minimizing plain weight reconstruction error, H-Scale selects hardware-valid group scales using a diagonal second-order proxy derived from calibration activations, thereby targeting layer output perturbation more directly. It is designed as a drop-in replacement for RTN-style scale selection in diverse NVFP4 pipelines, requires only modest offline calibration, and introduces strictly zero overhead at inference time. Under a fixed evaluation protocol, experiments on mainstream LLMs show that H-Scale generally improves a broad range of NVFP4 baselines and brings several variants closer to the BF16 reference.

HyQuant: Hybrid-Precision Quantization for LLM Attention

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 .

Understanding the Energy Scaling of Large Language Model Inference Across Context Lengths and Attention Architectures

The growing adoption of large language models (LLMs) has raised increasing concerns about the energy consumption and environmental impact of inference. This paper presents a systematic empirical study of decode-phase energy consumption across representative open-source LLMs employing Multi-Head Attention (MHA), Grouped Query Attention (GQA), and Grouped Query Attention with Sliding Window Attention (SWA) to characterize how attention architecture influences decode-phase energy consumption under varying inference workloads. We evaluate four models across different context lengths, batch sizes, and generation workloads while measuring GPU energy using NVIDIA hardware counters. We examine the effects of context length, attention mechanism, Key-Value (KV) cache growth, and batching on decode-phase energy consumption. Results show that attention mechanism is the primary factor governing how decode energy scales with context length. MHA models exhibit substantially steeper energy growth than GQA models, whereas GQA with SWA maintains nearly constant energy consumption. We further show that model size primarily determines absolute energy consumption, while batching reduces both energy per generated token and request latency by up to 87%. These findings provide practical guidance for selecting energy-efficient LLM architectures and inference configurations.

SHIFT-LLM: Distribution Shift Correction in Depth-Pruned LLMs

Depth pruning removes entire Transformer blocks to reduce the inference cost of large language models, but disrupts the hidden-state distributions expected by downstream layers, leading to significant accuracy loss. We introduce SHIFT-LLM, a training-free post-pruning correction framework that inserts a Linear Residual Adapter (LRA) at each pruning site. Each LRA preserves the identity pathway of the original residual block and adds a lightweight affine residual correction. This correction is calibrated via closed-form least-squares regression on a small held-out set, without gradient computation, to approximate the missing residual update produced by the pruned block. Together with the preserved identity pathway, the resulting LRA output approximates the hidden state produced by the original block, thereby mitigating the distributional mismatch introduced by layer removal while avoiding the expensive attention and feed-forward computations of the removed blocks. The resulting LRAs support low-rank factorization and exact merging across consecutive pruned layers for additional compression, and combine naturally with parameter-efficient fine-tuning for further recovery beyond fine-tuning the pruned model alone. Experiments on five model families, six layer-selection criteria, and seven zero-shot benchmarks show that SHIFT-LLM consistently recovers accuracy lost to depth pruning across most configurations, achieving gains up to +15.7 points on Llama-3.1-8B-Instruct while requiring only a few hundred calibration samples and no gradient computation.

FLINT: Efficiently Leveraging High Bandwidth Flash for Capacity-Scalable LLM Inference Acceleration

LLM inference is increasingly constrained by accelerator memory capacity rather than compute throughput. This constraint is especially acute in single-accelerator and small-node inference systems, where limited on-package memory capacity restricts the size of deployable models. HBF is an emerging 3D-stacked NAND flash technology that provides multi-terabyte near-accelerator capacity, making it a promising capacity tier for storing LLM weights. However, existing HBF-based proposals face three adoption challenges: they (1) rely on coarse-grained static prefetching for LLM weights aiming to hide the microsecond-level read latency of the NAND flash device while maximizing HBF's read throughput, (2) expose NAND flash management tasks (e.g., refresh operations) to the accelerator-visible critical inference path, and (3) miss optimization opportunities to specialize and optimize the flash-management mechanisms to the workload behavior. Our goal is to design an efficient HBF substrate that integrates HBF as a memory-capacity tier alongside HBM while addressing these three challenges. To this end, we propose FLINT, a workload-driven HBF substrate for capacity-scalable LLM inference. FLINT introduces three mechanisms: (1) a hardware burst-buffer controller that dynamically coalesces and pipelines HBF reads aiming to utilize existing NAND flash buffers while sustaining high HBF bandwidth, (2) a phantom-plane refresh mechanism, which removes refresh from the critical inference path by moving refresh-related NAND flash operations outside the read foreground back via low-cost resource duplication, and (3) a read-only FTL, which replaces SSD-class support for arbitrary writes with a compact table that translates logical weight bursts to physical HBF locations.

Hydra: Phase-Aware Workload Characterization of LLM Inference across Edge SoC Generations, Backends, and Quantization Levels

Edge LLM deployment is shaped by more than model size and precision: inference backend, hardware platform, memory traffic, and power management all affect latency and efficiency. We present Hydra, a common-schema, phase-aware workload characterization framework for LLM inference on edge SoCs. Hydra instruments HuggingFace Transformers and llama.cpp with a shared per-prompt timing schema and fuses those records with hardware telemetry, enabling a multi-dimensional characterization of performance, system-resource utilization, and efficiency across prefill and decode phases. Using Hydra, we evaluate three consecutive edge System-on-Chip (SoC) generations (AGX Xavier, AGX Orin, and AGX Thor), 13 instruction-tuned LLMs from seven families, five execution formats, and consider input/output-length sensitivity. The resulting artifact contains roughly 107K per-prompt records and is publicly released with Hydra. Our analysis shows that aggregate latency alone hides key deployment effects: backend structure changes where latency is introduced, quantization reduces memory traffic and energy but does not predict power monotonically, and SoC generation changes how utilization and efficiency should be interpreted. By connecting phase-level timing with system-resource utilization and efficiency metrics, Hydra enables reproducible, phase-aware characterization of edge LLM inference. Hydra's source code and the collected per-prompt trace corpus are available open-source at: https://github.com/amirtaherin/hydra

ResiSpec: Enhancing Multi-Candidate Speculative Sampling via Residual Distribution Shaping

The efficiency of Large Language Model (LLM) serving is fundamentally limited by the sequential nature of autoregressive decoding. Speculative Decoding (SD) mitigates this by using a lightweight draft model to speculate future tokens, which are then validated by the LLM in a single parallel forward pass. To further boost efficiency, multi-candidate schemes propose diverse candidate sets to increase the likelihood of token acceptance. However, we show that these schemes are bottlenecked by Residual Drift: a phenomenon where the rejection of initial candidates causes the residual target distribution to diverge from the draft model's predictions. This shift renders subsequent candidates ineffective and forces the system into expensive resampling. To resolve this, we propose ResiSpec, a framework that strategically reforms the proposal distribution during verification to anchor the residual target mass within the draft model's high-confidence regions. By mathematically re-aligning the verification process without compromising output exactness, ResiSpec prevents candidate obsolescence and achieves up to 1.92$\times$ speedup over state-of-the-art multi-candidate methods. Code is available at https://github.com/Czzzk/Resispec.

Deterministic LLM Inference Across GPU Kernels: Power-of-Two INT8 Quantization Scales and the Limits of Tolerance-Based Conformance

Conformance suites for quantized GEMM kernels ask whether two implementations agree within a tolerance. We measure what such a suite can detect. Injecting nine faults into a reference INT8 pipeline over 8,232 layer--fault--regime cells of Qwen3-1.7B, we find that every one of five epilogue faults -- scale precision, double rounding, multiplication order, output truncation, fused ordering -- moves the output by at most a single bfloat16 spacing, and by exactly one whenever it moves it at all, across 5,880 cells. A tolerance of one spacing is therefore blind to the entire class by construction: four of the five faults are detected by no check in the suite, and the fifth only under power-of-two scales. Faults that violate the accumulator's exactness preconditions, or that break operand sharing, are detected without exception, and a null fault never fires. What a tolerance-based suite of this shape establishes is therefore narrower than interchangeability: that the preconditions hold, that operands are shared, and that differences stay within one spacing. The power-of-two constraint that exposes the one detected fault is also deployable. Requantizing every weight scale to its nearest power of two makes CUTLASS and Triton agree bitwise at every linear layer (196/196 and 252/252, against 8/196 and 10/252 under the checkpoints' own scales) and yields byte-identical generated token sequences at 1.7B, 8B and 14B (8/8 prompts, against 0/8 at all three). Observed perplexity point estimates are +0.32%, -0.28% and +0.48%; the 90% intervals cover zero at the two smaller sizes but not at 14B, reaching +0.71% and +0.76%. A previously reported +157% perplexity for this intervention was an artifact of a probe that rewrote scales without requantizing the weights; separating the effects attributes 99.8% of it to the resulting weight--scale mismatch rather than to the power-of-two constraint itself.

FAMPWQ: Fisher Information-based Adaptive Mixed Precision Weight Quantization for Effective LLM 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).

NeuroPrefetcher: Storage-Aware Sparse LLM Inference via Delta Prefetching

Deploying large language models on edge devices is increasingly limited by a widening gap between model size and available memory. Existing approaches such as quantization, smaller models, and offloading can raise the effective memory limit, but they still assume that the model can be compressed or partitioned to fit within some budget. We target the harder model-exceeds-memory setting, in which the model remains larger than resident memory throughout execution and storage becomes an active source of weights on the critical path. We observe that MLP activity during autoregressive decoding has strong temporal locality: approximately 82-85% of active neurons persist from one token to the next. This means that most sparse weights needed for the current token are already resident, and only the newly needed rows must be fetched from storage. We present NeuroPrefetcher, a storage-backed LLM inference system that exploits this property through predictive delta prefetching. After layer 0, a single GPU-resident predictor, occupying 2.86% of base model parameters, predicts sparse activity for all downstream MLP layers in one forward pass. The runtime compares these predictions against resident GPU buffers and issues application-scheduled NVMe reads only for incoming delta rows, replacing reactive operating-system demand paging with explicit, model-aware weight movement. On real unified-memory edge hardware, NeuroPrefetcher achieves 7.9-12.0x speedup over llama.cpp across constrained memory budgets.

COEC: Calibrated Orthogonal-Equivalence Compensation for Structured Pruning of Large Language Models

Structured pruning reduces the size and inference cost of large language models (LLMs) by removing weight columns, but the resulting output error can degrade accuracy. Existing training-free compensation methods use an additive bias or a single orthogonal rotation on the output side of the retained weight. These corrections leave its input singular frame unchanged and therefore limit how the retained weight can adapt after column removal. We propose COEC (Calibrated Orthogonal-Equivalence Compensation), a training-free compensation framework that applies alternating left and right orthogonal rotations to the retained weight. The right rotation is optimized on a reduced Stiefel manifold, while singular values are rescaled using generalized cross-validation to select the regularization strength for each layer. COEC further tempers the calibration Gram matrix to reduce the dominance of high-energy activation directions and introduces an alignment penalty that preserves the geometric relation between adjacent attention projections.All components use second-order statistics from a small calibration set and require neither backpropagation through the LLM nor retraining of the model parameters. COEC is independent of the column pruning criterion and can be applied to multiple structured pruning methods. Experiments on the Llama-3, Llama-3.1, and Qwen2.5 model families across multiple structured sparsity levels show that COEC improves perplexity on every model and zero-shot accuracy in most settings over existing compensation methods, with larger gains at higher sparsity. These results show that post-pruning compensation can recover part of the performance lost to column removal.

Which Eviction Policy Should an LLM Cache Use? A Systematic Study Across Workloads, Capacities, and Encoders

Semantic caches reuse an LLM response when the incoming query embedding lies near a cached query, but proposed eviction policies have rarely been compared under one protocol. Using CLEVER, we evaluate FIFO, LRU, LFU, ARC, GDSF, a single-pass streaming adaptation of SISO, and a semantic-redundancy policy across three ordered, deduplicated query corpora, three cache capacities, and two encoders. No evaluated policy improves on LFU by more than 0.041 percentage points in any of the eighteen settings. Replacement is not irrelevant: FIFO and streaming SISO trail LFU by as much as 8.67 and 8.55 points, respectively, at tight capacity. We explain the missing upside with a conditional packing result. Under exact lookup and insert-on-miss, a newly inserted entry cannot have a resident neighbor within the hit radius, so a geometry-aware eviction rule receives little new redundancy signal. A separate audit exposes a larger problem with the evaluated operating point. At MiniLM's median nearest-neighbor threshold, only 2.1-3.9% of sampled LMSYS and QQP hits are judged answer-substitutable, reducing raw hit rates of 51-60% to quality-adjusted rates of 1.1-2.2%. The cross-encoder study further shows that thresholds do not transfer between embedding models. LFU is the strongest simple default in this protocol; deployment decisions should first establish answer validity and then test sub-point policy differences with exact search.

Flama: a Python framework for development and deployment of production-ready APIs, machine learning, and LLM services

We present Flama, an open-source Python framework for developing and deploying production-ready web APIs, machine learning services, and large-language-model (LLM) applications. Built on the Asynchronous Server Gateway Interface (ASGI), Flama offers a type-driven, async-first programming model that unifies REST API development, predictive model serving, and generative AI inference in one architecture. It is organised around seven subsystems: a component-based dependency injection system resolving handler parameters from type annotations at startup; a pluggable schema layer supporting Pydantic, Marshmallow and Typesystem behind a single adapter; an automatic CRUD generator turning a SQLAlchemy table and a schema class into REST endpoints backed by the Repository and Unit of Work patterns; a portable binary format (.flm) packaging models from scikit-learn, TensorFlow, PyTorch and Hugging Face Transformers with their metadata for zero-code deployment; a multi-backend LLM server running vLLM (Linux/CUDA) or MLX (Apple Silicon) and exposing four wire protocols (OpenAI, Anthropic, Ollama, and a native streaming dialect) through a shared codec; a Rust-accelerated core compiled via Maturin for routing, JSON encoding, compression and parsing; and a Model Context Protocol module turning any application into an MCP server over JSON-RPC 2.0. Built-in capabilities include JWT authentication, two pagination strategies, background tasks in threads or processes, WebSocket endpoints, Server-Sent Event and NDJSON streaming, OpenAPI 3.2.0 generation from handler signatures, and a command-line interface for running applications and for serving, packaging and inspecting models. We describe the architecture, present the programming model through worked examples, and compare Flama with existing frameworks, model serving platforms and LLM inference engines.

TileMix: Tile-Centric Mixed-Precision Attention for LLM Inference Acceleration

Long-context prefill in large language models (LLMs) incurs substantial computation and memory traffic because dense self-attention computes quadratic query-key scores. Existing methods either use a uniform low-precision path or select token interactions, leaving spatial precision routing over hardware-aligned score tiles outside fused dense attention. We introduce TileMix, a tile-centric precision-routing kernel that makes numerical precision an executable spatial decision over score-tile groups within fused dense attention. TileMix partitions the attention matrix into hardware-aligned score tiles, packs routing decisions into compact bitmasks, and dispatches each tile group through FP16 or INT8 score computation while both paths update a shared online-softmax state. Scalable precision grouping lets each routing bit govern multiple adjacent key tiles, preserving hardware-aligned compute tiles and compact metadata at long contexts. By routing all legal tile groups, TileMix preserves dense token connectivity, requires no training, and supports grouped-query attention, variable-length batches, and INT8 key/value caches. Across LongEval, LV-Eval, and A100 prefill benchmarks on LLaMA, Qwen, and Vicuna, TileMix recovers long-context quality lost under uniform INT8 and improves prefill throughput over FP16, yielding a controllable accuracy-efficiency frontier across model families. The implementation is available at https://github.com/HanzhiZhang-Ulrica/TileMix.

Pallas: A Proactive KV Cache Migration Framework for LLM Inference in AI-RAN

AI-RAN brings large language model (LLM) serving close to mobile users, but cellular handover can separate an active request from its inference state: the user attaches to a target base station (gNB) while the large and growing key-value (KV) cache remains at the source. Retaining inference at the source preserves service continuity but persistently increases inter-token latency (ITL), whereas recovering the state at the target restores serving locality but requires KV-cache transfer, recomputation, or a combination of both only after handover, directly prolonging service interruption time (SIT). This work presents Pallas, a \textit{proactive} KV-cache migration framework that prepares the inference state at the predicted target before handover, in parallel with ongoing source-side inference and token delivery. At the preparation trigger, Pallas partitions the token sequence into a stable historical prefix and an evolving suffix. The target reconstructs the prefix through local prefill, while the source streams the KV blocks generated for the suffix. At handover, the target assembles both portions into an up-to-date KV cache and resumes decoding locally, leaving only unfinished preparation to contribute to SIT. An online scheduler selects the \textit{prefetching window}, which determines how early preparation begins before handover, based on mobility predictions and runtime telemetry. Across three LLMs and $100$--$500~\mathrm{Mbps}$ inter-gNB links, our vLLM-based prototype reduces average SIT by factors of $2.28$--$89.68$ over target-side recovery approaches and lowers average ITL by $16.0\%$--$50.0\%$ compared with source-side forwarding.

Global Simulation-Guided Dynamic Operator Scheduling for Efficient Multi-Tenant Model Serving

Container-granularity scheduling leaves abundant short-lived idle slices within containers unexploited. Reallocating containers is too heavyweight to utilize such fine-grained opportunities under SLA constraints, and operator-level scheduling requires reasoning about dependencies, memory safety, and cluster-wide execution dynamics in real time. In this paper, we present SliceScheduler, a dynamic operator-level scheduling system for multi-tenant model serving. The key idea is to expose cluster-wide operator execution state and enable what-if reasoning over scheduling decisions. SliceScheduler consists of four key components. First, we introduce the Global Mapping Graph (GMG), a unified abstraction that captures operator dependencies, tensor shapes, resource mappings, and execution states, providing a real-time, cluster-wide view with explicit resource semantics. Second, we build a global simulator on top of GMG to predict operator-level execution and memory evolution under candidate placements. Third, we design an incremental, simulation-based scheduling module that selects placements to exploit fragmented idle slices while avoiding memory violations and preserving SLA. Finally, we develop an operator executor that materializes scheduling decisions on GPUs and coordinates computation and cross-accelerator transfers. We implement SliceScheduler as a PyTorch backend and evaluate it using production trace replay. Experimental results show that SliceScheduler improves token throughput by 1.10--2.29$\times$ compared to existing approaches, while maintaining SLA violations within 9\%. SliceScheduler demonstrates that operator-level scheduling is a practical and effective approach to improving GPU utilization for multi-tenant LLM serving.

FluxBin: Flexible LUT-based Ultra-low-bit LLM Inference by Algorithm-Kernel Synergy

While binary quantization theoretically promises extreme compression and acceleration for Large Language Models (LLMs), existing research often overlooks the necessity of specialized hardware kernels, thus failing to unleash the full acceleration potential due to persistent reliance on expensive floating-point arithmetic or runtime dequantization overheads. To bridge this gap, we propose FluxBin (\textbf{F}lexible \textbf{L}UT-based \textbf{U}ltra-low-bit e\textbf{X}ecution with \textbf{Bin}ary bases), an algorithm-kernel co-design that synergizes post-training quantization with a highly optimized CUDA kernel. Algorithmically, we introduce Decoupled Row-Column Binary Decomposition to enhance representational capacity while maintaining hardware efficiency, complemented by a Hessian-guided saliency-aware hybrid bases that preserve critical information. At the kernel level, we implement a Lookup Table Building Approach with Scale Fusion to reduce floating-point arithmetic, featuring a Virtual Columnar Mapping that transforms irregular, sparse, and salient matrices into dense execution. Extensive evaluations demonstrate FluxBin achieves up to $5.92\times$ speedup and $10.19\times$ energy savings across diverse model architectures, delivering comparable accuracy to heavily fine-tuned methods. This effectively enables the deployment of 70B-scale models on one single A100 GPU with a $4\times$ memory reduction. Code is available at https://github.com/nicyyyy/FluxBin.

S2-MoE: Enabling Efficient Self-Speculative Decoding for Mixture-of-Experts on Edge Devices

Deploying large language models (LLMs) for inference on edge devices is challenging due to severe memory and bandwidth constraints. While speculative decoding and Mixture-of-Experts (MoE) have been proposed to improve inference efficiency, naively combining them often incurs excessive verification overhead and poor expert reuse, limiting their effectiveness in memory-bound edge settings. In this work, we propose S2-MoE, an efficient self-speculative decoding framework for MoE inference on edge devices. S2-MoE reduces redundant verification through routing-aware adaptive speculative expansion, improves verification efficiency with reuse-aware expert gating, and aligns draft and target execution via shared context. Implemented in llama$.$cpp, S2-MoE achieves up to $5.3\times$ speedup (about $2.0\times$ on average) over standard autoregressive decoding across diverse MoE models and datasets on edge devices. Code is available at https://github.com/angerybob/S2-MoE.

The Integer Alibi: Localizing Cross-Kernel Divergence in INT8-Quantized LLM Inference

Two GPU kernels implementing the same scaled INT8 GEMM interface are usually treated as interchangeable. We test that assumption: holding the checkpoint, prompts, hardware, inference engine, decoding, and quantization configuration fixed, we swap only the INT8 linear kernel (CUTLASS versus Triton) inside vLLM. At 1.7B each arm reproduces itself bit-for-bit across cold restarts, yet the arms agree on no sequence in any end-to-end comparison we ran (0/8, 0/16, and 0/64). What makes this more than a benchmark discrepancy is an integer alibi: for shared INT8 operands under a verified no-overflow bound, the INT32 dot product is exact and order-independent, so the accumulator cannot be the source of any difference. Feeding both kernels identical operands from every linear layer of Qwen3-1.7B and 8B (196 and 252 layers), we find bit-identical outputs under power-of-two scales, confirming a pinned prediction list 196/196 and 252/252 (pre-registered at 1.7B, pinned but not blind at 8B), and observed differences of at most one bfloat16 spacing under the checkpoints' real scales. This localizes the divergence to scale application and output rounding after the exact accumulator. Applied as a probe checkpoint, the same intervention restores end-to-end bitwise agreement (8/8 and 16/16 sequences). Cross-implementation FP8 GEMM shows a different signature: both the prevalence and the magnitude of differences grow with reduction depth, while the INT8 fraction stays at parts per million and within one spacing over a 64x range of K. Teacher-forced replay ties layers to tokens: flips concentrate at small logit margins, which predict flip risk with ROC-AUC 0.94 on 16,384 positions. We will release the pre-registration, per-layer predictions, manifests with kernel-selection evidence, and a conformance procedure that turns these controls into a concrete check for kernel interchangeability.

Reduced Matrix Multiplication: Input-Adaptive Matrix-Product Reduction for LLM Inference

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.

InFactPlanner: Planning Sustainable Geo-Distributed LLM Data Centers

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.

MemSpec: Memory-Aware Runtime for Adaptive Draft Scheduling in Speculative Decoding on Edge Devices

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.