生成模型与LLM推理优化

Long-horizon large language model (LLM) agents require context assembly: the runtime must decide what to include in each prompt, in what order, and when to compact history under a hard context-window budget and a byte-sensitive prompt cache. In production agentic systems, this logic is scattered across prompt builders, ad hoc compaction routines, cache-break workarounds, and per-provider shims. We argue that context assembly is structurally isomorphic to query execution in a relational database: both execute under a hard budget, exploit a tiered cache, and leverage statistics. We adopt this discipline in ContextPipe: a five-phase pipeline (Plan Bind Optimize Execute Feedback) backed by a structured data-source catalog, a deterministic cache-aware optimizer, and an EXPLAIN ANALYZE trace. We show that context in ContextPipe is auditable, replayable, and failure-isolated. A preliminary evaluation using the SWE-bench Pro Qutebrowser subset shows that, compared with the append-only context construction policy, ContextPipe reduces total token volume by 31%, LLM calls by 23%, and response time by 9%, at the cost of a lower KV cache-hit ratio.

Density Functional Theory (DFT) is indispensable for materials science and drug discovery, yet the exchange--correlation (XC) evaluation remains a major bottleneck due to its cubic scaling. Although linear-scaling methods exploit electronic nearsightedness to reduce asymptotic complexity, they produce irregular sparse workloads that hide implicit sparsity and prevent efficient use of modern AI accelerators. We present MakoXC, a modular matrix-aligned XC evaluation engine that rearchitects nearsightedness-induced sparsity into regular, accelerator-friendly computations. MakoXC co-designs three key techniques: (1) Matrix-Aligned Cells reorganize nearsightedness-induced interactions into dense, accelerator-aligned data clusters; (2) Sparsity-Guided Activation translates deeper implicit sparsity into numerically correct structured execution for practical linear scaling; and (3) Kernel-Fused Pipeline consolidates fragmented workloads into a unified, compute-intensive execution path that fully unleashes accelerator throughput. Extensive evaluations show that MakoXC achieves average speedups of 67.8$\times$ speedup over standard XC evaluation and 4.7$\times$ over state-of-the-art linear-scaling methods. When integrated into a production-grade commercial DFT package, MakoXC scales XC evaluation to ubiquitin (1,231 atoms, def2-SVP) on 64 GPUs, enabling the end-to-end DFT calculation to complete in under five minutes. By restructuring XC evaluation into a unified, structured computation, MakoXC demonstrates how scientific workloads can achieve genuine low complexity while maximizing parallel efficiency on AI accelerators.

Multi-agent combinatorial optimization problems are notoriously challenging due to their NP-hard nature. Recent parallel autoregressive neural solvers improve inference efficiency by allowing agents to make decisions simultaneously, but their performance often degrades on large-scale instances. This is largely attributable to weak modeling of local geometric structures and the fact that conflicting task selections are handled only after action generation. To address these limitations, we propose GeoPAR, a geometry-guided parallel autoregressive reinforcement learning framework for scalable multi-agent combinatorial optimization. GeoPAR integrates three key components: (1) a projection-window sparse geometry mechanism that builds lightweight local candidate neighborhoods through multi-directional projections, (2) sparse edge-biased attention that injects these geometric relations into node representations, and (3) cache-guided conflict-aware assignment that reuses the geometric cache during decoding to suppress duplicate selections of exclusive tasks. Experiments on heterogeneous vehicle routing and open multi-depot pickup-and-delivery problems show that GeoPAR improves large-scale zero-shot generalization while substantially reducing rollout steps and maintaining efficient inference.

A multi-model language service must route each request while preserving workload-level budgets for compute, latency, memory, or monetary cost. Two features make this problem materially harder than static model selection. Prompt representations are high dimensional, so only a small subset of embedding directions may predict the incremental value of a model, and both the request mix and the model frontier drift after launches, fine-tunes, quantization changes, and system updates. We formulate nonstationary sparse contextual routing with multiple knapsack constraints and an optional shadow-audit stream that evaluates a small fraction of prompts on several models. We propose Drift-Aware Sparse Routing (DRS). The policy estimates reward and resource use from a rolling audit window, routes using pessimistic reward and optimistic cost estimates, updates resource shadow prices online, and applies a hard meter before commitment. The analysis separates control from statistics. On any event with uniform prediction radii $\{β_t\}$, regret against a paced dynamic fluid benchmark is bounded by the sum of the radii, a capacity-buffer term, and an $O(\sqrt{T})$ pacing term. Under a sparse linear model and bounded drift $V_T$, rolling estimation gives \[ \widetilde O\left( T\sqrt{\frac{s}{ρW}}+WV_T+\sqrt{T} \right), \] where $s$ is sparsity, $ρ$ is the audit rate, and $W$ is the window length. Optimizing $W$ yields the usual stationary $O(\sqrt{sT/ρ})$ rate when $V_T=0$ and a $O(T^{2/3}(s/ρ)^{1/3}V_T^{1/3})$ adaptation term under drift.

Edge-AI model selection is commonly driven by one isolated metric - accuracy, latency, memory, energy, or safety, even though a deployable language model must balance all five. Our work focuses on answering the question whether na- tively trained small language models (SLMs) or large language models (LLMs) compressed through post-training quantization offer the more sustainable edge- deployment trade-off. We introduce a reproducible Holistic Sustainability Score (HSS) organized around the triple bottom line: an economic pillar for capability and systems efficiency, an environmental pillar for operational GPU energy and a social pillar for harmful-prompt robustness. Five BF16 SLMs and five LLMs under different quantization approaches - BF16, INT8, NF4 4-bit, GPTQ 4-bit, and GGUF Q4 produce 30 measured configurations. Capability is assessed on five zero-shot benchmarks; efficiency uses latency, throughput, peak VRAM and energy; and safety is approximated by attack success rate on five harmful prompts. Qwen3-30B-A3B/GGUF Q4 ranks first in the combined pool (93.38), followed by Mistral-Small-24B/GGUF Q4 (92.40), while Phi-4-mini/BF16 is the highest- ranked SLM in that pool (89.49). Thus, the hypothesis that native SLMs must be the most sustainable edge choice is not supported universally; optimized quantized LLMs can win overall, while SLMs remain competitive through lower resource demand. Quantization is a systems-level choice rather than a monotonic precision- efficiency trade-off and HSS remains relative to its comparison pool and proxy definitions.

Post-training quantization (PTQ) is widely used to reduce the cost of serving large language models (LLMs), but its accuracy cost is uneven and is often tuned per model. We study where quantization damage occurs and how to allocate a small additional precision budget. Using causal mixed-precision intervention as ground truth (raise each layer to 8-bit in turn and measure the accuracy it recovers) across 9 open-weight models in 4 architecture families, we test 3 intuitive hypotheses: that quantization damage lives in task circuits, where the model computes, or in weight statistics. None of them predicts which layers benefit from restored precision. Recovery is instead diffuse: for 8 of 9 models, recovering 75% of the gap takes roughly half the layers; the lone exception, Qwen3-8B, is sharply concentrated. At a matched precision budget, spending it globally on finer quantization granularity beats locally repairing the most recoverable layers for all 8 group-128-compatible models (all but OpenLLaMA, whose width rules out group-128), by 21-52 points, including the concentrated Qwen3-8B. We report 2 secondary findings: the residual is budget-limited (8-bit is near-lossless in our evaluation across RTN, GPTQ, and AWQ), and the location of peak recovery correlates with architecture within a family, though not across families. Within this budget setting, global granularity is a better default than selectively protecting critical layers. More broadly, cheap signals that correlate with quantization damage do not necessarily identify where restoring precision improves accuracy; this must be tested with causal intervention.

The heavy-tailed distribution of output lengths in Large Language Model (LLM) serving poses major challenges for resource provisioning and cluster scheduling. Although output-length prediction can mitigate these issues, existing approaches have key drawbacks: external proxy models add substantial latency and often have limited fidelity, whereas internal state-based methods are efficient but rely on shallow probes of current model states. We identify a structural connection between speculative decoding (SD) and length prediction: latent representations produced by the draft decoder in advanced frameworks (e.g., EAGLE-3) encode signals that are predictive of generation length. Building on this insight, we introduce OUTLETS (Output-Length Prediction from Speculative Decoding Backbones), which repurposes the speculative backbone as a trajectory-aware length predictor. When its draft representations are already computed for speculative decoding, OUTLETS adds only a lightweight regression head and achieves lower MAE than the evaluated methods. Under saturated disaggregated serving, OUTLETS predictions enable standard scheduling policies to prioritize shorter requests and distribute requests more evenly across decoding instances, reducing short-request P99 latency by 34.8%.

On-device mobile Large Language Model (LLM) inference is gaining significant attention. However, mobile devices operate in highly dynamic multitasking environments where users frequently switch between applications. This creates memory pressure, forcing LLM memory (model weights and KV cache) to be evicted by the operating system. When a new inference request arrives, the inference system must restore the evicted memory through slow storage reads or recompute the entire KV cache, severely degrading responsiveness. To address this, we present mzCache, an on-device LLM inference system with specialized memory management for multitasking environments. Under unpredictable memory pressure, mzCache elastically evicts LLM memory and leverages the unified memory of mobile SoCs to enable zero-wait inference on the GPU with concurrent CPU-side restoration. mzCache realizes this through restoration-oriented memory management: LLM memory is partitioned into fine-grained shared buffers to enable partial eviction and restoration with concurrent cross-processor access, while hybrid swap and backward-out eviction policies ensure low-latency restoration from any eviction state. Implemented on llama.cpp and deployed as an Android application, mzCache achieves 2.1-5.5$\times$ reduction in Time-to-First-Token compared to storage-backed partial offload and demonstrates its effectiveness in real multitasking scenarios.

Looped Transformers increase effective depth by iterating a shared block of layers, but most evaluations compare at fixed model size, conflating architectural advantage with extra FLOPs. We study looping on Mixture-of-Experts Transformers while closely matching per-token FLOPs, total non-embedding parameters, and KV cache. Through a series of ablations, we arrive at a recipe we call SMELT (Sparse MoE Transformer, middle layers Loop Twice), which loops the middle half of layers twice while matching the unlooped Baseline on all three budgets. We scale SMELT across four sizes up to 54B non-embedding parameters and fit a separate Chinchilla-style scaling law for each architecture. SMELT's loss drops faster with compute, saving 6.8--18.0% of training FLOPs on the compute-optimal frontier. The advantage transfers to downstream benchmarks beyond what validation loss predicts, is largest on Code, and grows with sample length and the number of in-context examples. Mechanistic analysis shows that the second visit reduces the attention sink and redirects mass toward content-relevant tokens, an inductive bias that may underlie the observed performance gains. These results show that looping can improve Transformers even under budget matching, offering a practical recipe that turns depth reuse into measurable gains.

LLM inference has become an essential service, yet it imposes unprecedented demands on memory bandwidth, computational density, and communication efficiency. While IMC is a promising solution to the memory wall issue, the heterogeneous data dynamicity of LLM requires complementary resources to handle intermediate data generated during run-time. Furthermore, the massive number of parameters in LLM necessitates scale-up architectures where on-chip data movement is often the primary performance bottleneck. This article presents a hardware-software co-design framework that unifies distributed compute, memory, and communication into a seamless processing-communication fabric. On the hardware side, we propose a scalable architecture, named LEAP, that integrates IMC PE, NMC PE, and INC. This allows each hardware layer to execute specialized tasks: IMC for static weights, NMC for dynamic data, and INC for partial result reduction. On the software side, we introduce a partitioning, mapping, and scheduling framework optimized for key metrics in LLM serving, including throughput and latency. To address the distinct computational intensities of the prefill and decode phases, we present a prefill-decode disaggregation approach that dynamically reconfigures PE organizations to maximize resource utilization. Compared to commercial GPU platforms, the proposed architecture provides a throughput and an energy efficiency improvement of $\geq{}1.52\times$ and $24.91\times$, respectively.

Sharing context between LLMs in a multi-model system requires the receiving model to prefill the shared prefix because KV caches are model-specific. Recent closed-form cross-model KV transfer, hereafter Full-Head Mapping, avoids this replay by fitting a training-free affine mapper from source to target caches. However, its full-head design maps each target KV head from every source KV head in the selected layers, making transfer quality sensitive to architectural differences and causing mapper storage and application cost to grow with layer support. To this end, we introduce CacheBridge, which co-designs architecture-indexed mapper support, attention-aligned calibration, and bounded mapper construction while retaining a closed-form affine interface for online deployment. CacheBridge restricts each target head to a matched source head, weights reconstruction errors by causal attention sensitivity, and uses a fused GPU kernel to construct weighted sufficient statistics without materializing full observation tensors. Across three transfer directions, CacheBridge recovers the two Ministral 3 transfer directions where Full-Head Mapping loses substantial accuracy while preserving 99.83% mean target retention on Qwen3. On Qwen3 $14\mathrm{B}\to32\mathrm{B}$, it reduces mapper storage by $8\times$, accelerates application by up to $3.0\times$, matches \fullhead with one tenth of the calibration data, and reduces 500-sequence construction from 92.63 to 8.63 seconds ($10.7\times$).

Rollout inference often dominates the wall-clock time of large-scale reinforcement learning (RL). In agentic RL, each trajectory alternates between model generation and environment interaction over multiple turns. Asynchronous trajectories consequently introduce new prefill (P) work while other trajectories remain in decode (D), making P/D coexistence a persistent property of the rollout rather than a one-time prompt-ingestion event. On shared accelerators, persistent P/D coexistence can make prefill interfere with latency-sensitive decode and prolong rollout completion. P/D disaggregation avoids this co-location but requires separate device pools and KV-cache transfers. In-place multiplexing retains shared devices and KV state, but existing designs lack the communication isolation needed for large MoE deployments that combine attention TP/DP with distributed expert execution. In practical implementations, P and D can issue intersecting collectives in inconsistent cross-rank orders; DeepEP's P and D paths also share mutable protocol state. We present AInfer-PD, which extends in-place P/D multiplexing to distributed MoE rollouts. AInfer-PD coordinates P/D collective order across ranks and gives the two DeepEP paths independent communication state, making crossed ADP/ATP and DeepEP paths safe for concurrent P/D execution. The design retains shared model weights and KV storage while coordinating P and D on the same devices. Across repeated single-node prefill-intensive workloads, AInfer-PD reduces fixed-workload rollout completion time by 7.1-22.5% relative to the same AInfer engine with P/D multiplexing disabled and by 24.8-32.9% relative to SGLang. On two nodes, the reductions are 18.0-35.3% and 18.3-31.8%, respectively. In a same-engine ablation, fine-grained boundaries reduce completion time by a further 8.6-19.8% over whole-epoch asynchronous enqueue.

Despite their strong multimodal understanding ability, multimodal large language models (MLLMs) incur substantial computational overhead when processing long visual token sequences. To reduce inference costs, recent studies have explored visual token pruning through vision-centric or text-guided strategies. However, these methods often overlook high-norm outlier tokens, i.e., tokens with abnormally large feature norms, leading to suboptimal pruning decisions. In this work, we show that such high-norm outlier tokens are highly redundant in both feature and spatial dimensions, yet are often mistakenly preserved as informative cues by existing methods. Motivated by this observation, we propose SinkPruner, a training-free visual token pruning framework for efficient MLLM inference. SinkPruner follows a coarse-to-fine design with two key modules: a visual sanitizer that filters high-norm redundancies and alleviates attention sink and attention dispersion, and a text-guided pruner that further retains tokens semantically aligned with the text query. Extensive experiments on twelve image-language and four video-language benchmarks demonstrate the effectiveness, efficiency, and generalizability of our framework. Notably, SinkPruner preserves 96.5% (91.8%) of the original performance of LLaVA-1.5 (Qwen2.5-VL) under an 89% token reduction. Experiments further indicate that our visual sanitizer exhibits promising transferability in enhancing the performance of existing pruning methods. Our code is available at https://github.com/LaVi-Lab/SinkPruner.

A prominent paradigm in inference-time alignment employs lightweight supervisors to steer Large Language Models (LLMs). Through empirical analysis, we identify a structural mismatch in this paradigm: weak supervisors exhibit pervasive high entropy across the vast majority of tokens, yet prevailing dense intervention approaches mandate supervision at every decoding step. This leads to frequent low-confidence interventions that can disrupt valid base-model reasoning and incur substantial utility costs. To resolve this, we propose TUSA (Trust-based Uncertainty Sparse Alignment). Moving away from continuous oversight, TUSA reframes alignment as a dynamic arbitration process, introducing an uncertainty-aware arbiter that authorizes intervention only when two conditions are met: the supervisor is confident and the token is semantically salient. This mechanism effectively filters out uncertainty-driven noise and redundant supervision. Extensive experiments across multiple models and benchmarks show that TUSA consistently improves both safety alignment and general helpfulness. By bypassing approximately 50% of alignment steps, it not only enhances safety preference by up to 15.6%, but also boosts general preference rates by up to 12.0% compared to the dense baseline, demonstrating that selective, high-precision alignment can outperform continuous supervision.

Large Reasoning Models (LRMs) impose substantial energy costs during deployment, yet current compression methods apply uniform quantization across all components, risking damage to critical reasoning circuits. We present a reasoning-aware compression framework that benchmarks quantization conditions across five reasoning benchmarks, GSM8K, FOLIO, MATH-500, ProofWriter, and MuSiQue, with hardware-level GPU energy measurement; profiles per-module INT4 vulnerability across all 196-224 (layer, projection) pairs via a perturbation sweep on a held-out calibration split, then selectively restores the most sensitive circuits to FP16. Three findings emerge. First, INT4 quantization can increase energy by extending reasoning chains; a 25% power reduction becomes a net energy increase on GSM8K. Second, vulnerability is task-dependent: attention projections are more critical for mathematical reasoning, and sensitivity patterns differ by architecture in logical inference. Third, selective compression achieves Pareto-optimal points inaccessible to uniform methods: R1-Qwen-7B Top-10% on ProofWriter gains +12 pp over FP16 at -9.7% energy, validated on held-out data across five reasoning benchmarks.

We present Turing-20B-A2B, a 20B-parameter Mixture-of-Experts language model that activates approximately 2B parameters per token, designed for long-context and latency-sensitive physical AI applications. The model adopts Quantile Routing in a dynamic top-k configuration, enabling token-adaptive expert allocation while maintaining balanced expert utilization and a controlled average compute budget. During deployment, we further apply capacity-constrained routing to prompt prefill for more regular and efficient expert execution, while retaining dropless routing during pretraining. Turing-20B-A2B also employs a hybrid attention architecture that combines Lightning Attention with a small number of full-attention layers for efficient long-context modeling. The model is pretrained with a progressive three-stage curriculum and extended to a native context length of 128K through continued pretraining, with further inference-time extension to 512K using YaRN. Despite its compact active-parameter budget, Turing-20B-A2B achieves, at the base-model stage, overall general capability exceeding Qwen3-8B Base and approaching Qwen3.5-9B Base, while maintaining strong long-context performance and favorable prefill-latency scaling. These results demonstrate an effective balance among model capability, long-context scalability, and practical inference efficiency.

Multimodal models often build on architectures designed for generative vision-language modeling, typically combining separately pretrained vision encoders with causal language models. Visual document retrievers such as ColPali repurpose these models as encoders, carrying over the parameter and compute overhead of a VLM for a non-generative task. We introduce NeoMME, a family of 260M and 800M-parameter Multimodal and Multilingual bidirectional Encoders that process multilingual text and raw image patches in a single bidirectional Transformer encoder. Both models are pretrained from scratch with a masked discrete-diffusion text objective, conditioned on visible image patches for multimodal examples. Both support a 16,384-token context, enough to encode up to two standard 4K UHD images. To demonstrate its downstream capabilities, we fine-tune NeoMME with jointly trained dense and late-interaction heads. On the ViDoRe v3 benchmark, the resulting NeoMME-Retriever 260M outperforms all evaluated models strictly below 800M parameters with 0.523 nDCG@10, while NeoMME-Retriever 800M reaches 0.556. At a matched 2048x2048 image input size on an NVIDIA L40S, NeoMME-260M encodes pages with about 2x the throughput of ColModernVBERT. Hierarchical token pooling and asymmetric quantization compress late-interaction multimodal document embeddings by 255x while preserving over 95% of baseline nDCG@10. We contribute NeoMME to Hugging Face Transformers and release the pretrained backbone and retrieval-compatible checkpoints under Apache 2.0 at https://hf.co/collections/Hcompany/neomme.

Efficient evaluation changes the protocol used to support claims about model behavior, yet it is rarely tested whether those claims remain stable after the evaluation itself is made cheaper. We stress-test conclusion robustness in responsible-AI benchmarking by evaluating three dense and mixture-of-experts models on BBQ and BBQ-V under seven conditions spanning batching, quantization, benchmark reduction, and their combinations. Rather than treating preserved aggregate accuracy as sufficient, we compare accuracy, bias severity and prevalence, reasoning quality, subgroup behavior, subset-membership stability, runtime, and measured GPU energy against a full-benchmark BF16 baseline. Larger batching keeps accuracy within 0.35 percentage points of baseline and produces comparatively small subgroup changes, while reducing energy in five of six model--dataset settings. INT8 largely preserves quality but uses 1.79--4.26$\times$ baseline energy. INT4 causes larger, model- and context-dependent changes. Reduced benchmarks provide the most consistent savings, but very small subsets are substantially more sensitive to which items are retained. Efficient evaluation should therefore be treated as a measurement intervention whose validity must be checked across the conclusions the benchmark is intended to support. Our project website is https://vectorinstitute.github.io/sustainable-rai-evaluation/ and the code is available at https://github.com/VectorInstitute/sustainable-rai-evaluation.

Speculative decoding speeds up generation with an efficient draft model (drafter) that proposes tokens for a target model to verify in one pass, preserving the target's output distribution. High-acceptance block-diffusion drafters such as DFlash and DFlare fill an entire block in one parallel pass. In many cycles, the target accepts the whole block, so the drafter exhausts its trained block horizon before verification fails. We call this unrealized acceptance stranded speed-up. A mean committed length, per prompt or per cycle, hides it, whereas the acceptance histogram exposes it as a spike in the ceiling bin, the fraction of cycles that accept the entire block. We recommend the histogram as a preflight check before spending training compute. Naively widening the block at inference does not recover the speed-up, because once the block outgrows its training size, the drafter's bidirectional attention shifts its distribution even at early positions and erodes front-of-block verification. Instead, we post-train the drafter on a longer block with a short curriculum that emphasizes the newly exposed positions, a method we call DBloom. Expanding the pretrained DFlash and DFlare drafters from block size 16 to 24 across Qwen3-8B and Qwen3-4B targets raises the per-prompt committed length on the high-ceiling benchmarks by a median of +0.8 tokens (up to +1.1). Once continuation fine-tuning precedes expansion, the increase reaches 1.37 tokens. The same expansion also lifts committed length on all seven benchmarks for Gemma-4-12B-IT, a different model family, by a median of +0.41 tokens (Arm A), and the full continuation-then-expand pipeline (Arm B) adds +0.29 to +0.98 tokens over the same B16 drafter. In a prompt-matched comparison against JetSpec, a contemporary tree-based drafter not used in our design, DBloom commits more tokens on every benchmark at tree budgets up to 64 nodes.

We describe the architecture and ablations of Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125B parameters, 6B activated per token, and additional 51B parameters of n-gram embedding tables held off the accelerator. On fourteen pre-training benchmarks the model leads the 397B-A17B predecessor on eight and trails it on the rest by at most 2.6 points, at 1/3 the activated parameters, 1/3 the training tokens, and roughly 1/9 the training FLOPs. Token mixing uses a layer-wise hybrid of Gated DeltaNet (GDN) and global attention, with one full-attention layer in every four; at continued-pretraining time those full-attention layers are replaced by Qwen Sparse Attention (QSA), which scores context at micro-block granularity with a compressed lightweight indexer. The residual stream is widened to four branches and read through an elementwise gate, a design we call the Gated Residual (GR). Capacity is added outside the backbone by a single n-gram embedding layer whose tables are prefetched from host memory. We evaluate every candidate change along three axes: loss together with downstream benchmarks; the cost of the change in training, prefill and decode; and its effect on the optimal hyperparameters and training stability. Loss and downstream accuracy do not always move together: enlarging the n-gram vocabulary lowers loss monotonically while downstream accuracy saturates. The architecture and the Muon optimizer together shift the optimal learning rate and batch size upwards, render batch-size warmup unnecessary, and substantially improve stability under stress tests. Loss, benchmarks, efficiency and stability form one design problem. Solved jointly, they yield a recipe that is simultaneously more efficient, more capable and more stable.

Large language model (LLM) services increasingly process heterogeneous queries with diverse latency, accuracy, and resource requirements. While edge deployment reduces response time, the heterogeneity of devices and the diversity of model families, parameter scales, and quantization levels make efficient LLM query orchestration challenging. This paper introduces DRLM, a Deep Reinforcement Learning-based LLM query orchestration framework in edge environments. DRLM integrates two lightweight predictors: (i) a class-conditioned quality estimator that maps queries to semantic categories and infers model performance, and (ii) a feature-driven latency predictor that estimates inference time across model-device configurations. These predictions, combined with system state, feed a factorized Proximal Policy Optimization (PPO) agent that performs state-aware orchestration decisions. To enable data-driven orchestration, we construct a large-scale benchmarking dataset with 223 835 measurements spanning 1258 queries, 6 query classes, 8 model families (32 deployed instances), 5 quantization levels, and heterogeneous edge devices. Evaluation on a 64-node edge cluster and comparison with three baselines and two state-of-the-art methods show that DRLM reduces inference latency by up to 51% and queuing delay by up to 67 %, while incurring at most 8% accuracy loss. It improves latency under increasing workloads up to 61.4%, demonstrating robust and stable orchestration.

LLM agents that cache recovery suggestions from API errors can skip re-derivation in later episodes, spending fewer tokens and fewer model calls on constraints they have already learned. Server-side data drift turns those cached fixes into silent failures, and the usual remedy, re-deriving on every episode, gives the savings back. We introduce invalidation contracts, a protocol layer that attaches version stamps and cacheability hints to every recovery suggestion so the client can evict stale entries without trial and error, and keep the rest. The contract decomposes realized savings into two independent factors: validity, the fraction of cached suggestions that remain correct after a drift event, and compliance, the fraction the planner applies on the first attempt. Validity depends only on the protocol and is vendor-independent. Compliance depends on the planner model: identical wire bytes yield 100% first-try compliance on Claude Haiku 4.5 and 11% or below on Claude Sonnet 5, which exhibits input-schema conservatism, refusing fixes that add fields the original request did not contain. We evaluate across seven models, three serving paths, two domains, and approximately 9,400 episodes. Row-level invalidation raises compliance by 0 to 66.7 percentage points across the seven models, 55.6 to 66.7 on three, and recovers 29-33% of baseline token cost on four of seven models, while table-level invalidation destroys co-located entries and drops post-drift first-try rates to 0% on five of seven. Eviction precision is 1.00 at row granularity on every model under the row-level oracle of Section 4.1. The contract adds 15% to response payload. Version-stamp validity is deterministic by construction and produced identical results across every model and serving path, with zero contract failures in the entire evaluation.

Speculative decoding accelerates generation without changing its output, yet on vision-language models (VLMs) it has been caught in a self-defeating cycle. The drafter stays autoregressive, so it must stay small. A small drafter cannot afford the image at every step, so vision is compressed, pruned, or hidden. A drafter cut off from the image is then least reliable exactly where the image makes text predictable. We present GLANCE, the first one-pass block drafter that is lossless on an unmodified VLM target, and it breaks the cycle at both ends. A block-diffusion head reads the target's already-fused vision-language state, so vision costs the drafter nothing, and fills a whole block in one forward pass, so depth costs no sequential steps. A wide candidate tree is verified in one target pass, and every audited prompt reproduces greedy decoding exactly. Grounded workloads reward this most, entering a verbatim-copy regime whose long runs cost an autoregressive drafter a pass for every token and a block drafter one in total. Under one engine and one round budget, GLANCE decodes up to 2.93x faster than autoregression, from one draft pass a round where the production EAGLE3-VL head takes eight, and accepts 2.7x longer blocks than an EAGLE-3 head trained on the same corpus. One law organizes these results. Accepted length is set by the target's next-token entropy, with a fitted slope that steepens with grounding across all five tasks. The law transfers across targets and modalities and names its own boundary, since free-running text still favors a chain. Our code is available at https://github.com/js-lee-AI/GLANCE.

ReRAM-based in-memory computing (IMC) architectures are widely regarded as a promising approach to alleviating the computational bottleneck of conventional architectures. Since ReRAM crossbars perform matrix-vector multiplication (MVM) in the analog domain, their computational energy consumption is highly dependent on weight and activation distributions. However, most existing ReRAM accelerators focus primarily on weight optimization while paying limited attention to the impact of activations on computational energy consumption, leaving the energy-saving potential of activation sparsity largely underexploited. In this paper, we propose a voltage-aware ReRAM-based accelerator (VARA), along with its accompanying design methodology. Specifically, we first introduce a voltage-aware training (VAT) algorithm that incorporates a preset threshold into the activation function to steer the activation distribution toward zero values, thereby enhancing activation sparsity. Building upon this, we further propose a co-zero activation reordering (CAR) scheme for crossbar-level computation skipping. CAR clusters activation dimensions based on their co-zero correlations and consistently reorders both the activation matrix and its corresponding weights. This process consolidates scattered zero activations into contiguous zero-valued regions to maximize the benefits of crossbar-level computation skipping. Extensive experimental results demonstrate that, with only marginal accuracy loss, VARA reduces the average total system energy consumption by 60.12% and improves the average system energy efficiency by 2.68$\times$ compared to the baseline, outperforming existing state-of-the-art accelerators for sparse-activation optimization.

The growing scale of scientific data makes lossy compression essential for reducing data volume under controllable error. Transformation-based compressors using multilevel decomposition, such as MGARD, achieve strong compression ratios but map poorly to GPU architectures. We propose BlockMGARD, an adaptive, Region-of-Interest (ROI)-supported GPU lossy compressor, with four contributions: (1) an In-cache Block decomposition leveraging GPU on-chip memory and constant lookup tables to accelerate decomposition; (2) a hybrid hierarchy combining In-cache Block and global decomposition to balance speed and compression ratio; (3) an end-to-end pipeline with fine-grained ROI error control for feature preservation; and (4) an evaluation against state-of-the-art methods on five real-world datasets. Compared to MGARD-X, BlockMGARD achieves up to 4.2x and 9.1x higher compression and decompression throughput, and up to 8.63x higher compression ratio than uniform-tolerance baselines under ROI-aware error control. Across four GPUs, BlockMGARD achieves near-ideal linear scaling and up to 1.58x I/O cost reduction over MGARD-X.

Weight-only post-training quantization (PTQ) can alleviate the computational burden of serving large language models (LLMs) at scale. However, existing PTQ methods often fail to generalize across models and suffer severe accuracy loss below 2 bits. Many leverage unstructured sparsity to mitigate this loss, but at the cost of regularity and GPU-friendly execution. We present QTEA, a sub-2-bit PTQ framework that quantizes weights into ternary values and uses salient weights as residual error compensators. To maintain hardware efficiency, residuals are assigned to selected columns with semi-structured $1:4$ sparsity within the salient columns. We further add column-wise rescale refinement to GPTQ-style column-by-column quantization, alternately updating per-column scales and ternary assignments to reduce reconstruction error. We also identify order-dependent error propagation in GPTQ and introduce error decay to attenuate late-stage error accumulation. On Qwen3-14B, QTEA compresses all weights to an effective 1.7 bits per weight while improving average accuracy over the strongest ternary PTQ baseline by 16.7%. It also achieves 1.40$\times$ and 2.61$\times$ lower perplexity on WikiText and C4 respectively. This trend holds on Llama3-8B, where QTEA obtains a 6.6% accuracy gain and 1.34$\times$ / 1.95$\times$ lower perplexity on the same datasets. Finally, we develop a lookup-table based kernel that achieves 7.2$\times$ faster per-token generation over an FP16 baseline. Code is available at https://github.com/Intelligent-Microsystems-Lab/QTEA.

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.

Mixture-of-Experts (MoE) models enable efficient scaling of large language model (LLM) inference but suffer from substantial data-movement overhead when deployed on neural processing unit (NPU)-based systems. Near-Data Processing (NDP) provides a promising way to mitigate this bottleneck via cooperative NPU-NDP execution. However, existing NPU-NDP MoE systems do not fully account for hardware heterogeneity, dynamic expert-level concurrency, and temporal expert reuse during batched inference. This paper presents DynaNDE, a dynamic near-data expert scheduling framework that exploits NPU-NDP collaboration to accelerate batched MoE inference. DynaNDE introduces an analytical performance model that captures hardware heterogeneity, data-movement costs, and communication-computation overlap in cooperative NPU-NDP execution. Guided by this model, DynaNDE determines per-layer expert scheduling across the NPU and NDP while accounting for expert-level concurrency. DynaNDE also incorporates a reuse-aware runtime that avoids redundant parameter movement when experts reside in NPU memory. Experimental results show that DynaNDE achieves substantial throughput improvements over the state-of-the-art NPU-NDP MoE serving framework, with average speedups of 2.6$\times$ and 2.2$\times$ for the prefill and decoding stages, respectively.

The development of long-context Large Language Models (LLMs) is constrained by the memory bandwidth bottleneck and quadratic complexity of the attention mechanism during decoding. To overcome the inherent trade-offs between the memory overhead of metadata-based metrics and the computational inefficiency of adaptive selection strategies, we present Faster Flash Decoding (FFD), a novel hardware-algorithm co-design framework designed to break the memory wall in long-context decoding. FFD integrates the selector and computer into a fully fused kernel, replacing external metadata indices with content-aware scanning via low-bit quantization. Furthermore, we introduce the top-delta strategy, which dynamically filters blocks to achieve distribution-adaptive sparsity without global synchronization. Offering a training-free and plug-and-play solution, FFD also enables the reuse of scanning results for computation, achieving up to 11.6x kernel-level speedup and scaling to 256K context length, with 2.37x end-to-end throughput improvement. Empirical validation on RULER and LongBench confirms that FFD maintains model accuracy while delivering high-ratio sparsity, with code available at https://github.com/qluoluo/faster-flash-decoding

A central question in modern machine learning is how much a trained model can be compressed without changing its behavior, to reduce the memory, compute and energy required to deploy it. To study this, we quantify functional degeneracy through the behavioral recovery rank, defined as the number of leading behavioral-Hessian eigendirections required to recover a trained model's performance. Using the behavioral recovery rank as a geometric benchmark for compression, we find that structural and magnitude pruning retain more degrees of freedom, even after the task is saturated. This gap suggests that functional redundancy is distributed across parameter directions and is not exposed by individual weights or neurons.