The rapid scaling of large language models (LLMs) has unlocked remarkable capabilities across various domains. However, deploying LLMs on mobile devices remains challenging because of their high computational and memory requirements. Low-bit quantization (e.g., 4-bit weights, 8-bit activations) can reduce model size while preserving accuracy, but its practical efficacy on mobile GPUs has been limited so far. In this paper, we identify that the performance gap stems primarily from frequent data type conversions between quantized and floating-point values, and dynamic and inefficient memory access patterns caused by the growing Key-Value (KV) cache during autoregressive decoding. We introduce BitLLM, a mobile inference engine that addresses type conversion overhead and suboptimal memory accesses in quantized LLMs. Our approach features an adaptive layout and packing strategy for the KV cache that ensures efficient memory access throughout LLM decoding. This is coupled with an integer-native execution pipeline that defers and overlaps type conversion outside the compute hot path, keeping the critical MatMul stages in the integer domain. Implemented on off-the-shelf mobile GPUs, BitLLM outperforms leading engines, i.e., LLAMA.CPP, MNN-LLM, LiteRT-LM, and MLC-LLM, achieving average speedups of 1.78$\times$, 1.73$\times$, 1.58$\times$, and 1.47$\times$, respectively, across various models and quantization levels. Under long-context generation, BitLLM reduces average energy consumption (per token) by up to 88.9%.
On-policy distillation (OPD) trains a student on its own generations by matching the teacher's token probabilities. These probabilities entangle two kinds of information: which tokens the teacher prefers, and how confident it is. The latter reflects the teacher's own capacity and calibration, which a smaller student may be unable to reproduce. Because the teacher temperature changes these values without changing the ordering, OPD is also sensitive to the temperature choice. We ask whether the teacher's ordering of candidate tokens is enough, and propose Order-Only On-Policy Distillation (O$^{3}$PD). At each student-generated prefix, O$^{3}$PD trains the student to reproduce the order of the teacher's top-$K$ tokens, without fitting their probability values. The objective is therefore invariant to the teacher temperature and needs no temperature tuning. Across six teacher--student pairs, O$^{3}$PD outperforms OPD on the average over HMMT25, AIME24, and AIME25 in every pair, by up to 2.9 points, with gains of up to 5.6 points on AIME25. Other ranking losses yield similar gains, so the benefit comes from ordinal supervision itself. Analysis shows that O$^{3}$PD departs further from the teacher's probabilities than OPD, yet reproduces its ranking more faithfully and yields a stronger student. The ranking, not the probability values, thus carries the knowledge that on-policy distillation needs to transfer.
Gaurav Kolhe, Cansu Demirkiran, Shubhi Tyagi et al. · cs.LG
Compression 96%Inference speed 90%
Abstract
Memory capacity and bandwidth are major bottlenecks in LLM inference, motivating weight compression beyond conventional 4-bit quantization. Vector quantization (VQ) enables approximately 2-bit weight storage but often relies on codebooks incompatible with native FP4 computation. We introduce MXCode (MX-Native Codebook Quantization), a framework bridging ultra-low-bit VQ compression with MicroXcaling (MX)-native FP4 computation. MXCode proposes three strategies: PostScale, which quantizes learned codebooks to MXFP4; PreScale, which extracts MX scales before codebook learning; and PreScale+, which further compresses these scales using VQ. These strategies explore tradeoffs among quantization quality, scale overhead, and codebook granularity. All reconstruct MX-native W4A4 operands without materializing intermediate high-precision weights. We evaluate MXCode on Qwen3 (235B), DeepSeek-R1 (671B), and GLM-5.2 (753B) using the Microsoft MX Emulation Library. MXCode achieves approximately 2 bits per weight with competitive, workload-dependent model quality. On GB300 hardware, experiments with Qwen3 (235B) show approximately 50% lower model-load memory than Native FP4 while maintaining near-Native-FP4 serving throughput. Under memory-constrained prefix reuse, P90 TTFT improves by up to 94.5%. These results demonstrate software-based reconstruction on existing FP4 hardware and inform the development of future MX-native inference systems.
Hongwei Ren, Jincheng Ren, Yuchuan Zhou et al. · cs.LG
Compression 96%Inference speed 82%
Abstract
We introduce GateTrim, an adaptive-rank post-training quantization method for large language models that learns matrix-specific ranks through gate optimization and threshold-based rank selection. GateTrim standardizes the low-rank factors of a quantized low-rank--residual representation, learns a separate gate vector for each target weight matrix, and uses a shared threshold to automatically determine its retained rank. The resulting ranks can differ across matrices, and the compact quantized representations are subsequently reconstructed. We evaluate GateTrim under W4A8 on Qwen3-4B/8B/14B, Llama-3.2-3B-Instruct, and Llama-3.1-8B-Instruct. Across all five models, we observe a counterintuitive finding: increasing the threshold from $\tau=0.3$ to $\tau=0.7$ reduces the mean retained rank by 46.60%--67.89% while simultaneously lowering WikiText-2 perplexity by 0.59%--5.17%. Threshold 0.7 achieves the lowest mean retained rank and final perplexity among the three tested thresholds, 0.3, 0.5, and 0.7, and reduces measured target-export size by 1.50%--4.25% relative to 0.3. GateTrim uses an $8\times512$-token fitting pool and a separate development pool of the same size. Its Qwen3-4B/8B/14B perplexities at $\tau=0.7$ are numerically below TwinQuant's published W4A8 values, reported with a $128\times2048$-token calibration pool, in a cross-study comparison. On Qwen3-8B, the threshold-0.7 configuration already has the lowest perplexity before reconstruction under both tested initialization seeds; reconstruction improves it further without increasing retained rank or export size.
Inference speed 96%Compression 88%Serving systems 77%
Abstract
Long-horizon reasoning and agentic workloads strain KV-cache capacity and decode attention throughput. We present Jet-QSparse, a training-free method that combines compact NVFP4 cache storage with native FP4 query-key computation on Blackwell tensor cores. A dense scan scores the full key history and selects 128-token blocks for sparse FP16 value aggregation with FP32 accumulation, preserving higher precision in the output-forming step while enabling contiguous tile loads. Calibrated key centering and at most two BF16 pages per sequence and layer address quantization error; FP8 prefill uses a temporary arena while the persistent cache is packed directly from BF16 activations. Across four hybrid LLMs, a budget of 128 blocks per KV head shows no detected accuracy loss on the evaluated mathematical reasoning, coding, and agentic benchmarks, although some smaller budgets incur losses. With 64 blocks, Qwen3.8-27B achieves a $1.94\times$ end-to-end decode speedup over dense BF16 at 128K context and batch size 16 on one GB300, and the measured Qwen KV pools provide $2.58$-$2.76\times$ BF16's token capacity.
Baran Atalar, Selly Beyene, Xutong Liu et al. · cs.LG, cs.CL, cs.DC
Inference speed 96%
Abstract
Semantic caching lowers the cost of serving large language models (LLMs) by reusing stored responses for similar requests. In multi-turn conversations, whether a cached response is acceptable depends on both conversations' histories, which embedding similarity of requests and histories predicts poorly (AUC 0.62--0.74 on five workloads). We therefore learn the \emph{reuse risk}, the probability that a cached response is unacceptable, from judge feedback, and choose a budgeted set of conversation states to cache by weighing the generation cost a reuse avoids against a price for unacceptable reuse. We propose CUCB-SC-MT, which learns a cache from logged conversations with a suboptimality guarantee, and CLCB-LS-MT, which learns online and has $\tilde O(\sqrt N)$ regret over $N$ requests when conversation states recur. On math-tutoring (SocraTeach) and movie-ticketing (Taskmaster-3) conversations, both achieve lower loss than prior semantic caching approaches: ContextCache, multi-turn variants of ElastiCache and vCache, and single-turn semantic caching at every cache budget. With 1,000 cache entries, CLCB-LS-MT avoids 35--41\% of the generation cost through acceptable reuses, against at most 22--25\% for the baselines, and lowers estimated serving latency by 19--35\%.
Naoto Sugiura, Yuta Tokusashi, Hiroki Matsutani · cs.DC
Inference speed 96%Serving systems 47%
Abstract
As large language model inference environments expand toward consumer devices, it remains challenging to efficiently run models that exceed GPU memory capacity in single-GPU settings. The latency of transferring weights that do not fit in GPU memory accounts for about 98\% of inference time and remains a major obstacle. In this paper, we focus on the fact that during the decode phase, layer computation time on the CPU can be shorter than weight transfer time, presenting an opportunity to hide weight transfers by utilizing this computation time. Therefore, we propose a method to hide transfer latency by executing certain layers on the CPU during the decode phase while asynchronously transferring the next layer's weights to the GPU in the background of that computation. To realize this scheduling, we formulate the placement of CPU/GPU computation and weight transfers as a latency minimization problem and perform optimization using an approximation model with relaxed temporal constraints. Evaluation results confirm that in a consumer desktop environment, our approach accelerates OPT-30B inference by 3.5x compared to HuggingFace Accelerate and 6.3x compared to DeepSpeed Zero-Inference.
Kunjun Li, Wenhao Chai, Yihao Liang et al. · cs.LG
Inference speed 96%Serving systems 34%
Abstract
Speculative decoding accelerates memory-bound LLM decoding with a lightweight drafter, and parallel drafters extend this speedup by drafting a whole block in one forward pass. However, it remains unclear whether their released systems (stock) attain the full speedup that parallel drafting permits. We identify three challenges. First, with one request, each step accepts few tokens but reads every target weight. Second, with batching, per-request verification work roughly halves the speedup over autoregressive (AR) decoding. Third, speculative decoding is deployment-specific, and no universal optimization fits every serving regime, drafter, target model and GPU. To address these challenges, we present SpecPilot, an autonomous agentic loop that discovers algorithms and systems optimizations for each deployment. In each round, a cost model locates the largest gaps between the current system and its ceiling. Parallel agents explore directions and integrate the successful ones. Token forcing checks every output token against stock to detect reward hacking. A structured memory carries validated mechanisms to new rounds and deployments. SpecPilot discovered a general lossless 12-bit weight format, tree verification for hybrid recurrent targets, two-level verification and three batching mechanisms. We evaluate it across four target models, three drafters, two GPU platforms and five benchmarks. Without training or quality loss, it raises the speedup over AR from stock's 4.46$\times$ to 10.87$\times$ with one request, and from 1.94$\times$ to 3.93$\times$ with batching.
Key-value (KV) caches impose substantial memory pressure on large language model serving systems, requiring cache eviction when capacity is insufficient. Existing mechanisms typically evict KV from the ends of cached sequences to preserve prefix sharing across requests. However, agent workloads repeatedly reuse long histories across successive turns. Recomputing the KV cache for tokens near the end of a sequence is substantially more expensive than for earlier tokens because their attention computation involves longer histories. To address this problem, we propose preserving shared prefixes and evicting the KV cache from the beginning of the remaining history while retaining the cached suffix. We build Penelope around this strategy by co-designing the eviction policy, KV recomputation, and unified GPU-CPU cache management. Returning requests reuse the KV cache retained in GPU or CPU memory and recompute only missing segments while preserving reuse of the cached suffix. We implement Penelope in SGLang and evaluate it on three Qwen3 model sizes using real agent traces. Compared with SGLang, Penelope reduces mean time to first token by up to 60% and job completion time by up to 37% with a GPU-only cache, and by up to 76% and 60% with a two-tier GPU-CPU cache.
Smartphone applications increasingly require large language models to process long contexts, whereas high prefill latency and memory demands constrain local inference. Block-sparse attention reduces computation and memory traffic, but its irregular selections conflict with mobile NPUs' preference for large, regular tiles. Efficient execution must exploit cross-query reuse while avoiding redundant computation and preserving selections needed for accuracy. In this paper, we present TileWeave, which partitions sparse attention by this cross-query reuse: highly shared key blocks run in coarse NPU tiles, while low-reuse blocks run on the GPU only for the query blocks that selected them. In addition, we design an NPU-native pooled-key selector that keeps selection inexpensive, and a runtime that publishes GPU partials in the order the NPU consumes them and merges them in its existing output store. On RULER, HeteroSparse matches XAttention's accuracy with Qwen3-1.7B and improves it with Llama-3.2-3B-Instruct; its mean LongBench scores match XAttention as well. On Snapdragon SM8750 and SM8850 phones, it achieves 1.35--3.23$\times$ faster full-model prefill than XAttention over 1k--16k tokens and up to 1.58$\times$ faster than dense attention.
Xinwei Qiang, Ziyi Xu, Zaifeng Pan et al. · cs.DC, cs.LG
Inference speed 95%Serving systems 83%
Abstract
Parallel speculative drafters use a lightweight serial head to recover token dependencies within a block generated in parallel. Each head step updates only a small state, yet projects that state through the same vocabulary-sized matrix. Step-wise execution repeatedly transfers this large, immutable projection, making the head a meaningful source of latency despite its low arithmetic cost. We present Weights That Stay, an execution design that keeps projection shards on chip for the entire serial draft walk. A persistent megakernel couples weight residency with GPU-side token selection and state advancement, reusing resident weights while exchanging only compact state between steps. The design supports full or partial residency and accommodates both Markov and recurrent heads on H100 and B200. Integrated into SGLang, Weights That Stay improves throughput for Markov drafting on Qwen3-4B and Qwen3-8B, with gains of up to 12.3%. At the head level, the DSpark walk is up to 11.6$\times$ faster than the fastest measured SGLang reference path. These results show that aligning weight residency with the full draft walk translates cross-step reuse into end-to-end serving gains.
Rongfeng Guo, Zhuang Zhuang, Ji Dai et al. · cs.LG, cs.AI
Compression 95%
Abstract
Generative recommendation retrieves items by autoregressively generating semantic IDs, typically via beam search. On-policy self-distillation (OPSD) enhances training supervision by transferring privileged self-teacher knowledge at student-generated states. However, beam-search retention depends on competition among prefixes: visited states differ in importance, and initially reliable decisions may approach the pruning boundary as training evolves. Moreover, conventional distribution matching couples target-token confidence with preferences among competing continuations, potentially weakening relational supervision when the self-teacher strongly favors the target. To address these challenges, we propose Search-Aware On-Policy Self-Distillation (SA-OPSD), a framework that transfers privileged self-teacher knowledge to the states visited during the student's own beam search. Specifically, we introduce Evolving-Boundary Sampling, which prioritizes search-critical states by considering both current pruning-boundary proximity and initial-reference retention-margin decline under a controlled distillation budget. Furthermore, we develop Confidence-Decoupled Branch Distillation, which separates target-probability calibration from informative sibling relations to better preserve the teacher's preferences among competing branches. Experiments on Industrial, Office, and Instruments demonstrate strong performance for SA-OPSD.
Inference speed 95%Compression 93%Serving systems 35%
Abstract
Mixture-of-Experts (MoE) architecture has been widely adopted in large language models (LLMs) since it reduces the computational cost through sparse expert activation. However, the massive parameter scale of MoE models hinders their deployment on resource-constrained edge devices. Although existing mixed-precision quantization reduces the model size, they are inherently static, fixing the bit-width of each expert across decoding steps. Such static schemes yield suboptimal performance as they overlook the dynamic variation in expert importance during decoding. Furthermore, they fail to adapt to the dynamic latency constraint driven by the fluctuating system utilization. To address these challenges, we propose MatMoE, a dynamic mixed-precision quantization framework for efficient MoE inference. MatMoE features: (1) a Matryoshka MoE PTQ algorithm that enables experts to operate flexibly at multiple precisions within a stringent memory budget; (2) a floor-and-compensation quantization method that eliminates redundant memory accesses in the Matryoshka representation; and (3) a dynamic bit-width allocation strategy that determines the precision for each expert based on their routing scores to satisfy the dynamic latency requirement. We implement our MatMoE system atop llama.cpp for both CPU and GPU platforms. Extensive evaluations across three modern MoE models and challenging benchmarks demonstrate that MatMoE achieves $4.0\% \sim 18.3\%$ higher accuracy than state-of-the-art static mixed-precision method. Compared to llama.cpp, MatMoE delivers $1.3\times \sim 3.7\times$ speedup on Raspberry Pi 5, and $1.2\times \sim 1.3\times$ speedup on NVIDIA RTX 4090 and Jetson AGX Orin.
Reverse and forward Kullback--Leibler (KL) divergences provide complementary supervision in on-policy distillation: reverse KL refines the student's active token rankings, and forward KL allocates mass to teacher-supported candidates. At student-generated prefixes, we form a four-observation matrix by treating the student and teacher as both candidate sources and evaluators. The diagonal self-observations measure uncertainty, while source-centered cross-observations recover the two KL objectives. We establish an exact identity equating the Fisher inverse-metric coupling of their logit gradients with the sum of these contrasts, yielding a common local descent direction for distinct positive distributions. Observation-coordinated on-policy distillation (OC-OPD) uses observations to set preferences, closed-form bargaining to coordinate gradients, and distributional discrepancy to scale the update. Across three heterogeneous student settings (non-thinking generation, thinking generation, and base-model initialization), OC-OPD achieves strong gains in mathematical reasoning and transfers effectively to out-of-distribution scientific question answering and code generation. Our code is publicly available at: https://github.com/RuoranXu/OC-OPD.
On-policy distillation (OPD) trains a student language model on its own trajectories using token-level supervision from a stronger teacher. We show that standard sampled reverse-KL OPD systematically suppresses updates to high-confidence tokens, such that their weak gradient contribution does not necessarily imply weak or uninformative teacher supervision. We introduce Bernoulli On-Policy Distillation (Bernoulli OPD), which directly matches the teacher and student probabilities for that token. This yields a bounded probability space correction that preserves teacher-student disagreement while avoiding the additional confidence-dependent suppression induced by vanilla OPD. Fixed-rollout analyses show that Bernoulli OPD substantially restores optimization signal to high-confidence tokens while remaining sensitive to teacher-student disagreement. We further analyze the resulting model behavior through confidence distributions, answer calibration, policy movement, and training dynamics, showing that Bernoulli OPD's gains cannot be explained simply by longer reasoning, larger optimization steps, or more uniform imitation of the teacher. Code is available at https://github.com/duterscmy/Bernoulli-OPD.
Kevin Xu, Tatsuya Kubo, Daichi Tokuda et al. · cs.AR
Inference speed 95%Serving systems 37%
Abstract
Autoregressive decoding of large language models (LLMs) on edge devices is dominated by memory-bound general matrix--vector multiplication (GeMV) operations. Processing-using-DRAM (PUD) is a promising computing paradigm that can alleviate this bottleneck by reducing processor--memory data transfers and enabling massively parallel computation within DRAM (e.g., across 65,536 columns). However, weight matrices in individual LLM layers (e.g., a QKV matrix in Llama3 8B with 6,144 rows) are typically too small to fully utilize parallelism in PUD. To address this challenge, we propose an algorithm--system co-design for PUD-based LLM inference. At the algorithm level, we enable a subsequent projection to be packed with the current projection by reformulating it as a computation on the current residual state plus a correction for the residual update. Layer redundancy allows this correction to be approximated using fewer bit-serial steps. At the system level, we tailor data mapping and execution to improve utilization of either columns within DRAM subarrays (SIMD) or independently controlled banks (MIMD). Overall, we achieves up to $1.41\times$ end-to-end decoding speedup over the PUD baseline, corresponding to a $2.45\times$ speedup over processor-based execution.
CPU/GPU hybrid inference enables Mixture-of-Experts (MoE) models to run on memory-constrained PCs by keeping all expert weights in host memory and executing activated experts across both processors. Existing hybrid systems, however, assume exclusive resource access and determine an upfront CPU/GPU partition, so CPU contention from colocated applications can severely imbalance the two execution paths. This paper presents Apto, a self-adaptive hybrid inference system that applies dynamic scheduling to expert execution. The CPU and GPU executors claim experts from opposite ends of a dual-ended queue as they progress, allowing the CPU/GPU work split to adapt to their execution progress without predicting contention. Apto further uses affinity-aware scheduling to expose each executor's preferred experts at its queue end and cancellable tail racing to mitigate final-expert stragglers. Compared with state-of-the-art GPU-only and hybrid baselines, Apto achieves average prefill and decode speedups of 2.96$\times$ and 1.52$\times$ under standalone execution, and 5.97$\times$ and 3.45$\times$ under concurrent CPU workloads. It also accelerates real-world coding-agent applications by 2.02$\times$ on average.
Philipp Nazari, Neehal Tumma, Daniela Rus et al. · cs.LG, cs.AI, cs.CL
Compression 94%Inference speed 79%
Abstract
Linear recurrences are becoming an increasingly common building block in large language models, replacing the growing KV cache of softmax attention with a fixed-size recurrent state that enables efficient decoding. However, the quantization stack has largely been developed around softmax Transformers, whose numerical properties and inference state differ from those of recurrent models. Our goal is to understand these differences and study the design space of quantization methods for recurrent models. In our analysis, we differentiate between two targets. First, extra-state quantization concerns weights and activations outside the recurrent state, whose distributions are shaped indirectly by the sequence mixer during training. In controlled 1.3B parameter models, we find that decayed recurrences develop weaker residual outlier channels than softmax attention and are therefore easier to quantize without incoherence processing. Second, intra-state quantization concerns the recurrent state itself during decode, where repeated updates create a different numerical problem from KV cache quantization. We instantiate our findings as LoReS, a state quantization pipeline combining value-axis scale grouping to match output channel magnitude, decay-based mixed-precision to protect slowly decaying heads or channels, and stochastic rounding to limit error accumulation under repeated requantization. Across Qwen-3.5 9B, Nemotron-H 8B, and Kimi-Linear 48B-A3B, LoReS supports robust 8-bit state quantization and lower effective precision on shorter decode horizons. We implement LoReS in vLLM and demonstrate end-to-end decode speedups.
Yutong Lin, Yingdi Shan, Shaoyuan Chen et al. · cs.AR
Inference speed 94%Serving systems 73%Compression 37%
Abstract
Large Mixture-of-Experts (MoE) models are promising for local inference because sparse routing allows the routed-expert pool to reside on SSD while DRAM caches recently used experts. However, under practical DRAM cache capacities, substantial SSD I/O remains on the critical path despite I/O-computation overlap. Token-level dynamic expert precision offers a natural way to reduce this data demand by allocating bits according to expert importance and quantization sensitivity. Nevertheless, treating each bit-width of the same expert as an independent object prevents cross-precision cache reuse, causing overlapping data to be fetched repeatedly. The key insight is that precision should instead denote the refinement depth of one shared expert state. Based on this insight, we present PagedMoE, which stores each expert once using a prefix-compatible tree codebook and bitplane layout. Lower-precision representations are valid prefixes of higher-precision ones, allowing cached data to be reused or extended by fetching only missing bitplanes. PagedMoE selects the required prefix length before I/O using current routing weights and offline expert-specific error profiles, and directly executes cached 1-8-bit prefixes without materializing full-precision experts. At an average routed-expert precision of 2.5 bits, PagedMoE achieves quality close to GPTQ-INT4 on Qwen3.5-397B-A17B and to MXFP4 on most evaluated GLM5.2 tasks. With a 24 GiB expert cache, it reaches 8.89 tokens/s on Qwen3.5, 1.33x the throughput of the same dynamic requests using independent precision objects and 1.56x that of uniform 4-bit execution.
Kyle Lemoi, Jonathan Yong, Shubhankar Tripathy et al. · cs.LG, cs.AI
Compression 93%Inference speed 78%
Abstract
Depth pruning removes entire transformer blocks and therefore depends on a reliable criterion for identifying which blocks can be removed. Block Influence (BI) ranks blocks by how minimally they rotate the hidden state. However, local redundancy does not necessarily imply downstream irrelevance. We address this limitation using the Jacobian Lens (J-Lens) neural artifact, a map transporting perturbations at a layer to the final residual stream. At a common sparsity level, pruning scores applying J-Lens transportation to $BI$, what we call $J-BI$, were found to improve post-pruning accuracy by up to 19 points over $BI$ while reducing damage to perplexity for Qwen3-8B, Llama-3.1-8B-IT, and Gemma-3-12B-IT. We also perform a single-block ablation study and use J-Lens probing to explain why the proposed scores improved performance.
Leonidas Tsigkounakis, Abraham Jaeger Mountain, Yilang Zhang et al. · cs.LG
Compression 93%Inference speed 77%
Abstract
Post-training quantization (PTQ) reduces the memory burden of large model deployment by casting the high precision trained weights to fewer bits. While this promotes memory-efficient inference, it also incurs quantization error that impairs model performance. To compensate for quantization error, the present contribution decomposes the original trained weights into a quantized matrix, a low-rank component, and a sparse matrix that accounts for outlier rows in the weights. Due to analytical intractability and the discrete structure of the decomposition, a simulated annealing search is employed to explore candidate decompositions while avoiding poor local minima. Numerical tests demonstrate the effectiveness of this approach across a broad suite of large language models, including Llama 3.2 1B, Llama 3.2 3B, Llama 3.1 8B, and Qwen2.5 7B, evaluated on natural language generation benchmarks using 4-bit NF4 and FP4 quantizers. The novel method can markedly outperform direct quantization while incurring minimal memory overhead.
Yueyang Pan, Qianchao Zhu, Yuliang Liu et al. · cs.DC, cs.LG
Inference speed 93%Serving systems 37%
Abstract
Position-independent caching reuses chunk KV caches across positions and requests. Independent encoding, however, omits cross-chunk interactions and can degrade accuracy. In our study, neither PIC training nor rule-based recomputation, alone or combined, consistently recovers full-prefill accuracy. Oracle-guided recomputation on PIC-trained models improves accuracy at the same budget, motivating better selection. We present CacheWeave, which adapts the full backbone to PIC, then freezes it and trains a model-native block selector on the adapted model's oracles. The selector uses compact cached chunk features and question hidden states without extracting backbone attention scores. It selects contiguous token blocks aligned with KV blocks to support efficient recomputation and limit copy-on-write amplification. Its runtime, CacheWeave-Serve, supports private and shared KV caches, keeps selection and recomputation on the GPU, and stages sliding-window KV for hybrid attention models. Across three models and four tasks, CacheWeave outperforms CacheBlend and ProphetKV by 7.2 and 3.5 percentage points on average at a 5% recomputation budget, with a mean relative accuracy loss of 1.6% to full prefill. With 32-token blocks, CacheWeave-Serve cuts single-request TTFT by up to 16.1x over full prefill and 1.7x over ProphetKV, and raises SLO goodput under load by up to 7.8x.
Full-duplex speech models execute dense model inference for every audio frame throughout a live interaction. This uniform framewise schedule creates sustained GPU load across dialogue states whose computation has strong, reusable temporal structure. ReFrame formulates full-duplex inference as a silence-aware computation problem. Mutual Silence Speculation (MSS) operates when both participants are silent: it evaluates multi-frame Temporal continuations, verifies text and hidden-state stability, and applies batched Mimi audio-token correction. Listening Frame Folding (LFF) operates while the user speaks and the agent listens: it processes already-arrived user frames through parallel Temporal execution and frame-batched audio-token generation while accepting only the verified prefix. Frame-Budget Reinvestment (FBR) assigns the released frame intervals to resumable external retrieval and language-model workers, then injects timely references through Moshi's conditioning interface. We instantiate ReFrame on five full-duplex models spanning Moshi-family, phase-structured, and flattened autoregressive generation. Across the complete 1,225-conversation FDBench, ReFrame reduces measured GPU execution time by 18.80-31.01% and retains 96.92-100% non-PAD text agreement on five different models. On HaluEval tests, FBR increases the number of correct generated spoken answers from 86 to 205 under the same auxiliary-service budget. ReFrame turns recurring dialogue structure into lower inference cost and higher end-to-end service capability.
Youhe Jiang, Pavle Glu\v{s}ac, Binhang Yuan et al. · cs.DC
Inference speed 93%Serving systems 77%
Abstract
Long-context large language model (LLM) serving places substantial demands on compute, memory capacity, and memory bandwidth. Economical accelerators can offer these resources at lower cost, but their limited per-device capacity requires pooling across devices, making communication overhead critical. We study cluster efficiency, measured in tokens per dollar and per joule, and show how low-latency, high-bandwidth interconnects enable economical accelerators to retain their cost advantage. Guided by a characterization of coding-agent workloads and accelerator scaling, we realize this design on the Tenstorrent Galaxy, a cluster of Blackhole chips. The serving stack combines integrated communication engines, pipelined ring transfers, communication-computation overlap, and device-side decode replay to reduce collective overhead. Across three dense models in BF16 and 8-bit configurations with 8K-256K-token contexts, the Galaxy achieves 1.07-2.24$\times$ the tokens per dollar of an eight-GPU RTX PRO 6000 Blackwell server under a three-year acquisition-and-electricity cost model. Emulating TCP-class communication on the same chips reduces throughput by 23.5-67.4% and eliminates the cost advantage in every evaluated configuration. These results demonstrate how fast interconnects translate the per-dollar resource advantages of economical accelerators into cost-efficient long-context LLM serving.
Speculative decoding accelerates LLM inference with a drafter whose proposed tokens are verified by the target model in parallel. Under lossless speculative sampling, a draft token is accepted with probability one minus the total variation (TV) distance between the target and draft distributions, yet drafters are often trained with cross-entropy (CE) or KL. We first show that under matched and well-tuned training, TV matches or surpasses several published draft-training objectives when fine-tuning, and a fixed TV and KL mixture does the same when training from scratch. We identify two remaining problems, TV's vanishing gradients on undercovered tokens and the prefix products that starve later positions in acceptance-length objectives. To address both, we derive TVP, a loss family based on expected acceptance length, whose power parameter interpolates between coverage and acceptance oriented objectives. Annealing the power from zero to one gives the highest or tied-highest acceptance length across three drafter architectures and three target models, including longer drafts and block verification. TVP improves acceptance length over CE by 9.8% when fine-tuning, and over the public DFlash2 training recipes by 6.2 to 12.9%.
Existing serving systems for large language models batch requests for many different LoRA adapters on one GPU. However, these systems reject adapters from other parameter-efficient fine-tuning (PEFT) methods such as DoRA, (IA), VeRA and OFT. One solution to this is to convert these adapters to LoRA, but this induces a significant memory and accuracy cost. Instead, we show that, under some modifications, a batched LoRA kernel can serve all of these methods exactly and efficiently. We achieve this by writing each method in one normal form, a LoRA computation with six slots for an adapter's own vectors and matrices. Some slots can merge into an adapter's weights at load time; the others remain live and induce a cost. We give ways to efficiently pay this cost and predict it from a zero-parameter byte model. We package this system, Scute, and build it on top of vLLM. Scute serves all five methods, coming as checkpoints from the Hugging Face PEFT library, in one batch, and it matches that library with a mean error of at most 1.5% of each adapter's effect. Under a burst of requests on an RTX A6000, a mix of these methods takes 5-17% more decode time than the same number of LoRA adapters. Furthermore, using Scute, the same GPU can serve 64 to 1024 (IA) and VeRA adapters at 1.5x to at least 2.5x the request rate of their LoRA conversions.
Jianchang Su, Dawei Xiang, Wei Zhang · cs.LG, cs.AI
Inference speed 92%Serving systems 67%
Abstract
Speculative decoding speeds up LLM serving by drafting tokens that the target model verifies in one pass, and its gain grows with the drafts the target accepts. Structured outputs should be an ideal case, because much of a record follows from relations among its fields: an itinerary's days follow from its cities' stays and flights. Serving engines, however, draft such text by proposing it from the prompt or a learned head, and their grammar masks enforce only syntax, so the model must satisfy the relations, which it violates: on NATURAL PLAN's trip problems, Qwen3-8B writes a valid itinerary for 10%, and retries that name the violated relation raise this to 14%. We present FactorSpec, which serves exact relational constraints as speculative drafts. FactorSpec compiles a JSON Schema and its relations into a factor graph with an exact feasibility check, drafts every token the constraint determines as certain, verifies each request at a width chosen for it, and computes the masks in worker processes outside the scheduler while the GPU runs. In SGLang, its drafts yield up to 10.5 tokens per verify step, up to 1.9x those of the fastest n-gram drafter, and it decodes up to 1.43x faster than speculators tuned per load, at a grammar mask's cost; it also leads in vLLM. Because every record is also valid, it serves up to 13.9x the valid records per GPU-second of syntax-only baselines.
Michael Helcig, Earl Killian, Dan Alistarh · cs.LG
Compression 92%Inference speed 45%
Abstract
Block-scaled lookup-table (LUT) formats, which represent weights as indices into a small value set (the grid), have become standard for low-bit LLM quantization, ranging from NF4 and BOF4 to the 3-bit LUT instruction in NVIDIA Rubin. While these formats leave the grid definition flexible, existing implementations rely on fixed grids derived either from numerical data types (INT, FP) or assumed weight distributions (NF, BOF). We demonstrate that no single fixed grid is universally optimal; the best fixed grid varies with bit-width: BOFS at 2 bits, NF at 4 and 5 bits, and INT at 6 bits. We introduce SHq, a closed-form family of grids parameterized by two critical degrees of freedom: a companding strength {\alpha} that interpolates level spacing from uniform to floating-point-like, and an offset c that introduces asymmetry. This offset allows the sign bit of the FP8 block scale, otherwise unused, to select between the grid and its mirror image, up to doubling the number of distinct values available per block. SHq modifies only the lookup table contents and serves as a drop-in replacement for the grid in any such format. Across five models ranging from 3B to 35B parameters, including a mixture-of-experts architecture, evaluated at 2-6 bits using RTN and GPTQ, SHq achieves the lowest KL divergence in every setting at 3 and 4 bits and in most settings at 2, 5 and 6 bits. It outperforms the best fixed grid by up to 34% and remains superior under mixed-precision configurations. At 4.5 bits per weight, the SHq format reduces the KL divergence of NVFP4 by 41-67% while improving downstream accuracy. Grid families incorporating both spacing and offset parameters perform comparably, whereas those lacking an offset exhibit clearly higher error, confirming that these two degrees of freedom drive the observed gains. We therefore propose SHq as the standard grid for block-scaled lookup-table formats.
Dancheng Liu, Markus M\"uller, Rupak Vignesh Swaminathan et al. · cs.LG, cs.AI, cs.SD
Inference speed 91%
Abstract
Speech language models encode audio at a much higher rate than text, causing the KV cache to dominate the memory, bandwidth, and decoding cost of long-form generation. KV retrieval can reduce per-step access by attending to only a query-dependent subset of the cache, but existing methods are designed primarily for text and do not account for the structure of audio attention. We first show that retrieval errors may compound autoregressively into generation drift. We then analyze full audio attention during long-form generation and identify three properties relevant to retrieval: unequal attention concentration across query heads, non-contiguous audio support, and temporal continuity across decoding steps. Guided by these observations, we propose $C^3$(Cluster-Compete-Compete), which organizes audio KVs in a hierarchical key-space index, lets query heads sharing a KV head compete for a common retrieval budget, and adaptively reallocates retrieval resolution through paired split-merge updates. Across long-form speech recognition, translation, and summarization on four SpeechLLMs, $C^3$ achieves the strongest overall performance among retrieval methods under comparable final-KV access budgets, with particularly large gains in the tight-budget regime, while improving long-context decoding efficiency and accelerator-memory scaling.
Mixture-of-experts (MoE) models expand model capacity by activating only a subset of experts for each token. On accelerators with static execution plans, expert capacities and placement are fixed before routing reveals each expert's load. Uniform worst-case capacities waste work on padding; smaller capacities tied to expert identities can overflow. We present RankLane, which ranks each device's experts by current load and assigns them to fixed-capacity lanes. Runtime indices select experts and token rows while tier populations, shapes and device placement remain fixed. An aggregation graph preserves device ownership and reduces each token's local contributions on one core. Routing calibration and measured execution costs guide tier selection; overflowing chunks rerun at full capacity. We evaluate full-model prefill of Qwen3-30B-A3B and Gemma 4 26B-A4B with MXFP6 weights on four Cloud AI 100 Ultra SoCs, using 128-2048-token chunks and including overflow detection and in-program recovery support. Against a dense baseline, RankLane achieves end-to-end prefill speedups of 1.27-4.18x on Qwen3 and 1.93-3.46x on Gemma.
Multi-LoRA serving allows many task-specific adapters to share a common base model, but scaling to large adapter fleets introduces substantial storage, loading, and execution overheads. Post-hoc compression can reduce the memory footprint of trained LoRA adapters, yet smaller ranks do not automatically translate into faster serving. For instance, padding, factor access, and generic kernels can obscure the retained low-rank structure, while independently recompressing thousands of modules also incurs significant preparation overhead. We present LoRAForge, a system connecting batched factor construction to rank-specialized execution through an explicit retained-rank representation. It supports two allocation policies: uniform retention ratios (LoRAForge-U) and spectral-energy budget allocation (LoRAForge-S). Both share FlashTSQR factor construction and the same serving backend. FlashTSQR exploits parallelism across modules and row tiles, accelerating factor construction by up to $2.28\times$ over matched library QR. In a fixed-work Qwen control, the shared runtime adds $1.17$--$1.22\times$ throughput across both compressed policies, with negligible gains on original adapters. Across 85 trained open-weighted Qwen and Llama adapters, the spectral policy's floored quality configuration retains about 7.7% of adapter parameters with fleet-average score drops of at most 1.2 points on a 0--100 scale. The same rank-floored $1/12$ spectral exports achieve substantial token-throughput gains on both AMD and NVIDIA GPUs.
Shinyoung Kang, Junhyung Ahn, Yujun Kim et al. · cs.LG
Inference speed 90%Compression 85%Serving systems 64%
Abstract
On hardware whose fastest matrix datapath is INT8$\times$INT8, such as the Neural Accelerators of Apple's M5 GPUs, 4-bit weights must become INT8 operands. Existing MXFP4 kernels already map FP4 (E2M1) values to integers by doubling them, but apply every 32-element block scale inside the reduction. We show when no per-block work is needed: a row of MXFP4 weights is exactly one INT8 vector times one power-of-two scale if its block exponents span at most three, so a linear layer becomes a single INT8 GEMM with one rescale per output. Enforcing the condition adds at most 1/16 of a row's largest rounding error to the blocks it changes, and real weights already satisfy it: 97.9-99.97% of rows across seven models, including 99.93% of the 95M rows of a 125B-parameter mixture-of-experts (MoE) model, Qwen3.8-Flash-Next. The resulting spread-constrained MXFP4 files stay standard MXFP4. With integer instead of power-of-two block multipliers, an integer-exact format (16-wide blocks, multipliers 1..10, 4.26 bits) runs as one exact INT8 GEMM and is more accurate than NVFP4 (4.5 bits) with its standard scale rule on five small models, within 0.06 perplexity of NVFP4 with scale search. On Qwen3.8-Flash-Next's routed experts with INT8 activations it is as close to the FP8 model as NVFP4 with scale search (KL divergence 0.0557 vs. 0.0553) and closer than MLX's MXFP4 and affine INT4; with every matrix quantized it solves 94.8% of GSM8K vs. 91.0% for MLX's MXFP4. Quantizing only the 74k routed expert matrices (96% of the parameters) to either of our formats with INT8 activations costs at most 0.2% perplexity. A complete W4A8 system on the M5 produces outputs bit-identical to the quantized reference; prefill is 1.34-1.74x faster than MLX on a dense model, 1.10-1.57x on two MoE models up to 35B parameters and 1.22-1.29x on Qwen3.8-Flash-Next, where shared non-GEMM work bounds the gain.
Hamidreza Alikhani, Nazmul Takbir, Matin Khajavi et al. · cs.DC
Inference speed 90%Serving systems 53%
Abstract
LLM-based applications often employ agentic workflows and retrieval-augmented generation (RAG), which build long prompts from shared context. When prefixes are shared across requests, serving engines reuse the key-value (KV) cache to skip parts of the prefill computation. Reuse is typically graded: the uncached work of requests in a queue can differ by up to 100x. In this setting, the order of scheduling determines the time-to-first-token (TTFT) for the queued requests. KV reuse-first scheduling, which prioritizes the requests with the longest KV cache hit, gives the lowest median TTFT but starves expensive requests under load spikes; first-come-first-served (FCFS) scheduling bounds the tail but raises the median TTFT. We present Fairway, a prefill scheduler built on adaptive reuse prioritization: it favors KV cache reuse while queues are short, and shifts toward fair allocation as requests wait longer relative to their own prefill time. Fairway employs a virtual dual lane model: a KV reuse-first Fast lane and a Fair lane atop the serving engine's waiting queue that enables speeding up of cheap reuse requests without starving expensive ones. This model automatically adapts to the workload without any per-workload tuning. Fairway lowers the tail TTFT at no cost to the median in transient overload. Across workloads and load levels, Fairway matches or beats the best baseline, with a p99 TTFT up to 8.6x lower than prior prefix-aware schedulers. During load spikes, it keeps the tail of FCFS at a 6x lower median TTFT.
Reasoning models answer better when they think longer, but a thinking trace keeps going well after its answer has formed. We measure this slack in Qwen3.5-9B and remove most of it with concise self-training: two rounds of on-policy rejection-sampling fine-tuning that keep the shortest correct samples, plus a small share of answer-synthesis targets, giving Concise-9B. On 8,983 held-out problems (competition and hard mathematics, SuperGPQA, a five-area confirmation set and MMLU-Pro), Concise-9B generates 0.66x the tokens of Qwen3.5-9B at a 32K thinking budget, with accuracy equivalent within +/-1 point (-0.20 [-0.77, +0.36] points). Forcing each trace on the hard and confirmation sets to answer every 2,048 thinking tokens, we find that 54-64% of Qwen3.5-9B's thinking tokens come after the point from which every forced answer equals the final one (41-48% for Concise-9B). Stopping there would keep every answer at 0.35-0.44x the base model's tokens, but finding that point takes hindsight: stopping when three consecutive forced answers agree costs the base model 2.4-9.2 points. Concise self-training removes slack without that cost: 86% [83, 90] of the thinking tokens it saves come after the answer has formed. It has a limit: with traces continued to 80K tokens, Concise-9B solves 25 fewer of 1,276 hard problems (-25 [-49, -1]), mostly in competition mathematics. The measurement also yields a cheap commit signal for adaptive sampling that matched two-sample escalation within +/-1 point at 0.67x its tokens in a pre-registered test. Results on Qwen3.5-4B, on which the recipe was developed, and a replication on Olmo-3-7B-Think suggest that the saving depends on how verbose the base model is.