生成模型与LLM推理优化

Large Language Model (LLM) agents increasingly solve long-horizon tasks through multi-turn interactions with users and external tools. In these settings, relevant task information often unfolds over time rather than being fully specified at the initial prompt. Service agents make this challenge especially concrete: users may clarify or revise their goals, while tool responses provide information needed for subsequent decisions. Thus, a final reward alone cannot indicate which actions contributed to resolving the task. Recent methods rely on comparative evidence from other trajectories or resampled continuations, or on separately constructed step-level learning signals, to refine credit. However, a completed rollout already records how information and errors flow between agent actions. We introduce Influence-Aware Policy Optimization (IAPO), which represents each rollout as a typed influence-dependency graph over trainable agent actions, with user and tool observations serving as evidence. IAPO converts support-use and failed-use structure into routing weights that redistribute the same trajectory-level advantage. Experiments with Qwen3-4B and Qwen3-8B demonstrate superior performance over multi-turn reinforcement learning (RL) baselines across three service-agent benchmarks: ${τ^2}$-Bench, UserBench, and AgentChangeBench. BFCL-v4 Multi-Turn further shows that these gains do not compromise multi-turn function-calling performance. This work advances the understanding of credit assignment in multi-turn user interactions and provides a principled approach to training service agents from sparse outcome feedback.

Enterprise AI agents in production often need to be bounded, stateful, observable, and governable rather than fully autonomous. We present PinSieve, a production case study in a large-scale content-quality pipeline. Its deployed component is a selective vision-language-model (VLM) Serving Agent that operates only on the grey-zone slice left unresolved by lightweight upstream models, exposes a scalar routing score online, and preserves controlled human escalation. On this slice, the deployed system filters 2.05x more non-actionable items than the previous production module while slightly reducing estimated miss rate; after promotion, it improves review productivity by 25.7%, reduces normalized operating cost by 16.2%, and moves signal delivery from next-day to same-day. We then study maintenance through a governed memory flywheel under selective feedback, where escalated items are reviewed by default and auto-passed items are labeled mainly through audit sampling. Feedback Memory records routing traces, observation paths, audit propensities, and replay metadata for evaluation and debugging. The Data Curation Agent uses a bounded proposal-verifier loop over representative, uncertainty, recency, and fresh-review replay, with positive-rate and score-bin guardrails before batch acceptance. In chained monthly refresh over six months of production data, this design reduces average FNR@50% from 17.73% under representative random replay to 13.29%. A Reasoning Review Agent audits teacher-generated rationales and supports keep/repair/drop decisions. Production claims are attributed only to the deployed Serving Agent; replay and rationale-review results are offline or sampled-governance evidence. The same serving-agent recipe has been adopted to several additional internal signals, suggesting transferability beyond one task.

pigz is a widely deployed parallel gzip utility, but its process-global mutable state means that it was not designed as a reentrant, directly embeddable library. pigzpp is a from-scratch C++23 rewrite that turns the design into a thread-safe library with one accelerated DEFLATE core exposed to C++, Python, WebAssembly, Go, and Rust. As application-level conveniences built on that same core, it also provides multi-entry ZIP archives and a fast PNG codec. Its portable zlib-ng backend preserves gzip/zlib's ratio; its Intel ISA-L backend is an x86-64-only fast path that produces about 10% larger output at level 6 on our text corpus. In that configuration, zlib-ng reaches 2.1 times the CLI throughput of pigz and 16 times that of Python's in-memory gzip; ISA-L reaches 8.5 times and 50 times, respectively. Outputs remain standards-compliant and cross-decode with gzip, pigz, and unzip. The result is a self-contained, multi-platform compression stack and a case study in AI-assisted modernization under automated compatibility tests.

Joint text-to-video-audio generation produces synchronized visual and acoustic content, but the long sampling trajectories and heterogeneous multimodal computation of large models make inference prohibitively expensive. We present TurboT2VA, a distillation and inference framework for accelerating a 19B-parameter joint video-audio model. Large-scale T2VA distillation is challenged by modality-imbalanced optimization, the difficulty of continuous-time consistency training at scale, and the quality--diversity trade-off. TurboT2VA addresses these issues with per-modality normalization and a progressive curriculum comprising discrete consistency warm-up, continuous consistency refinement, and joint consistency--distribution matching. The curriculum first establishes a stable, diverse generation trajectory and only then introduces distribution-level refinement. On LTX-2, four-step distillation reduces generator latency from 50.52s to 2.51s at the standard evaluation resolution of 512$\times$768, achieving a 20.1$\times$ speedup while maintaining strong visual quality, audio fidelity, diversity, and video-audio synchronization. We further develop an architecture-aware inference stack that combines guarded W8A8 and fused operators, padded-text compaction, and modality-aware sparse attention while preserving dense cross-modal and text-conditioning paths. Under the high-resolution deployment setting at 1024$\times$1792, the complete stack reduces generator latency from 318.74s to 5.83s on one NVIDIA H20, achieving a 54.67$\times$ generator-only speedup. Inference code and generation demos are available at https://github.com/thu-ml/TurboDiffusion/tree/main/turbot2va.

Quantization correction methods usually optimize weights, quantization parameters, or reconstruction objectives, while the underlying parameter subspaces responsible for effective correction remain unclear. In this work, we study quantization correction from a parameter subspace perspective and reveal that correction capability is highly non-uniform across parameter groups. By decomposing trainable parameters into backbone weights, normalization-affine parameters, and quantization parameters, we show that the low-dimensional normalization-affine subspace provides a highly efficient correction direction under matched budgets. Based on this finding, we propose SandwichQuant, a two-stage normalization-affine correction framework that performs adaptation before and after quantization. The pre-stage improves quantization robustness, while the post-stage compensates residual errors after the quantized graph is fixed. Extensive experiments on vision models and large language models demonstrate consistent improvements under various low-bit quantization settings, validating the effectiveness of subspace-aligned correction.

Transformer-based object trackers are renowned for their strong performance, yet dense token processing often leads to prohibitive computational cost, limiting real-time deployment on edge devices. While recent works explore token pruning to reduce computation, they often stop short of an end-to-end sparse pipeline, as early-layer token scores can be noisy without a motion prior, and many trackers ultimately fall back to dense reshaping to feed the dense prediction head that partially negates the savings. We introduce Motion-aware Sparse Tracker (MaST), a sparse tracking framework that makes sparsity effective from tokens to boxes. First, MaST injects a lightweight motion prior to refine cross-attention-based importance scores, enabling earlier and more stable token reduction in the search region. Second, we introduce a natively sparse prediction head that operates directly on the retained unstructured tokens with a score-first, regress-once design, eliminating dense padding/reshaping and reducing redundant computation. Extensive experiments on multiple benchmarks demonstrate that MaST establishes new state of the art among lightweight trackers, where MaST-tiny attains 63.8 AUC on LaSOT and 80.1 SUC on TrackingNet, surpassing the prior best AsymTrack-S by +1.0 AUC and +2.2 SUC while running at 152 FPS on Jetson Nano, nearly twice as fast as AsymTrack-S at 88 FPS. Code is available at https://github.com/TsingWei/MaST.

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

System-level simulation is an essential tool for exploring the rapidly expanding design space of LLM serving systems, where real deployments remain costly and often infeasible. However, modern LLM serving now evolves faster than human-driven simulator development can track, and emerging workloads and mechanisms, from agentic workflows to disaggregated serving, no longer fit the monolithic simulation pipeline that existing simulators assume. Each new mechanism therefore demands an invasive rewrite, leaving a widening development gap between deployed serving systems and the simulators that model them. To close this gap, we present Simthesizer, a framework that realizes agent-driven simulator development. Simthesizer introduces a composable simulator infrastructure that uniformly expresses the complete serving workflow, including the control decisions that coordinate it, and realizes it as a unified dynamic graph in Simthesizer simulator. Synthesizer agent, a harnessed coding agent, then lowers natural-language feature requests onto this abstraction under simulator-specific guardrails and fidelity validation, evolving one shared simulator instead of building a new one for every feature. Under the same coding agent and harnesses, extensions built on Simthesizer follow a vLLM-based real system with 2.51% average throughput error, versus 6.03% for extensions built on existing simulators. On identical workloads, Simthesizer also simulates up to 284.96x and 23.19x faster than two state-of-the-art simulators, LLMServingSim2.0 and Vidur, respectively.

Post-training quantization lowers the memory footprint of Large Language Models (LLMs) and speeds up inference, which is why it is now common for on-device deployment. Most of what we know about its effects, however, comes from English benchmarks. It is not clear whether the same holds for morphologically complex, low-resource languages such as Bangla, and this gap is what we address here. We evaluate three model families---Qwen-2.5-7B, LLaMA-3.1-8B, and GPT-OSS-20B---in full precision and in three quantized formats (GPTQ-Int8, GPTQ-Q8, GGUF-W8A16) across five Bangla natural language understanding benchmarks (Bangla MMLU, CommonsenseQA-BN, OpenBookQA-BN, PIQA-BN, and BoolQ-BN), using zero-shot evaluation through lm-evaluation-harness. To our knowledge this is the first controlled comparison of quantization formats on Bangla NLU. The three families do not respond the same way: GPT-OSS loses up to 57.35% accuracy on reasoning-heavy tasks under GGUF-W8A16, while Qwen and LLaMA hold steady under GPTQ, and in a few cases the quantized version edges out the full-precision one. BoolQ-BN, a comprehension task, stays stable across all three families regardless of format. Taken together, these results suggest quantization can work well for Bangla deployment, but the choice of architecture and quantization method matters more than the bit width alone. We discuss what this means for practitioners choosing a model to run on constrained hardware.

Large language model (LLM)-based agent applications often incur high response time. Speculative decoding is a promising solution to improve the inference efficiency of LLM agents without impacting generation quality. However, state-of-the-art speculative decoding algorithms exhibit substantial speed degradation under large batch sizes, limiting their effectiveness to deploy in real-world agent applications. In this work, we first present a systematic analysis of speculative decoding for LLM agents and identify two dominant factors of speedup degradation: high rejection rate of speculative tokens, and under-utilization of dynamic token budgets.B ased on these observations, we propose AgentSpec, a speculative decoding algorithm that addresses the limitations of existing methods for LLM agents. AgentSpec incorporates structure-isolated drafting that constrains speculation to semantically coherent segments of the agent workflow, reducing the drafts of irrelevant semantic paths and achieving an extremely low rejection rate. Moreover, AgentSpec adopts redundancy-aware budget allocation that exploits agent-level information to better utilize the dynamically-free token budget during the agent inference. We implement and evaluate AgentSpec on five different workloads and four different models from four different LLM families in vLLM. Our results demonstrate the superiority of AgentSpec over state-of-the-arts.

When an LLM serving deployment runs out of KVcache room, there are two well-established ways out. Tensor parallelism shards the weights and the KV cache across two, four, or eight devices, buying memory headroom at the price of an all-reduce on every layer and a hardware bill that grows with the device count. The algorithms community shrinks the cache in place, with KV quantisation and eviction keeping a single GPU and spending a little quality instead. Compression papers report memory ratios, parallel-scaling papers report throughput curves, and almost nobody puts the two on the same cost axis. We place tensor-parallel configurations (degree 1 to 8) and KV-compressed configurations (16/8/4-bit, keep-ratios down to 0.25) on one costnormalised axis, cost per million tokens against latency, using a profiled simulator calibrated on A100, A40, and H100 hardware, and we go looking for the cost-equivalence crossover. We do not find one. Across two models (Llama-2 at 7B and 70B), three GPU types, and every level of memory relief we could construct, compression is cheaper by 1.20x to 2.00x. A 7B model on an 80 GB device cannot exhaust its KV budget within its own context window, and the boundary that decides between the strategies is model size relative to device memory, at roughly 36B parameters for an 80 GB card. Below that wall, compression dominates and extra GPUs are largely wasted spend; above it, tensor parallelism stops being a choice and becomes an entry ticket: Llama-2-70B is infeasible on one A100 at any KV setting, because the binding resource is weights, which KV compression does not touch. Tensor parallelism is the only lever that improves latency (compression makes per-token latency worse, by 8 to 93%, through batching contention), while compression is the only lever that multiplies capacity per dollar (16.5x, against 1.21x for an eightfold spend on GPUs).

While Vision Large Language Models (VLLMs) have achieved remarkable success in multimodal reasoning, their long-context inference remains prohibitively expensive due to the massive computation and memory overhead of visual Key-Value (KV) caches. Existing KV compression methods often apply uniform pruning across visual tokens and layers, leading to substantial information loss and degraded performance.To address this challenge, we propose VisCache, a plug-and-play framework for coarse-to-fine Visual KV Cache pruning without training, which consists of two synergistic stages. First, a lightweight VLM filters temporal redundancy by selectively forwarding semantically informative keyframes. Second, we introduce {PruneKV}, a surgical KV compression algorithm tailored to the attention dynamics of VLLMs. Unlike rigid pruning strategies, PruneKV adopts a parabolic layer-wise budget allocation together with an asymmetric update mechanism that selectively prunes keys while fusing values, thereby preserving critical contextual information. Extensive experiments demonstrate that VisCache substantially improves inference efficiency, achieving up to {2.35$\times$ speedup} and significant memory reduction while maintaining competitive performance with only {19--28%} KV cache retention. VisCache consistently outperforms existing baselines, establishing a new Pareto frontier between efficiency and performance for long-context VLLM inference. Code is available at https://github.com/Wlklk/VisCache

Diffusion Transformers achieve high-fidelity image and video generation, but their iterative sampling remains expensive, for each denoising step requires large matrix operations. Existing cache-based acceleration reduces redundant computation yet increases the VRAM footprint by storing intermediate states, which can directly constrain inference batch size. In this work, we propose a training-free acceleration method that performs stepwise forecasting for DiT sampling using a Barycentric Extrapolator. By leveraging barycentric extrapolation, our predictor is numerically stable and alleviates oscillatory artifacts analogous to the Runge phenomenon during forward forecasting. Across extensive experiments on both image and video generation, our approach provides a favorable trade-off between memory usage and perceptual quality, while delivering up to 3.30x end-to-end sampling speedup compared with baseline DiT inference.

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

Mixture-of-Experts (MoE) models scale capacity for strong quality while keeping per-token compute bounded through sparse expert activation. Yet low-latency MoE serving is increasingly challenging, because it spans two inference phases with fundamentally different bottlenecks: prefill is dominated by token-wise expert computation, whereas decode is constrained by memory traffic from the batch-wise activated expert set. However, existing training-free acceleration methods optimize only a single resource proxy, either the experts each token executes or the experts a batch activates, and either discard the excluded experts' contribution or leave it only implicitly approximated. In this paper, we propose ExFold, a unified training-free expert-folding framework for jointly accelerating MoE prefill and decode. ExFold casts both prefill and decode as one budgeted output-approximation problem: execute only a phase-specific constrained expert set while projecting the contribution of budget-excluded experts onto retained experts using calibrated scalar projectors. Motivated by the observation that many expert outputs are directionally aligned but differ in magnitude, ExFold calibrates a pairwise scalar-projector matrix on unlabeled data and uses it at inference time to fold excluded expert contributions into retained experts. Under this view, prefill acceleration becomes token-level Top-K folding, and decode acceleration becomes batch-level expert-pool folding. The two phases differ only in how retained experts are selected, while excluded contributions are recovered by one shared folding mechanism. We implement ExFold as a plug-and-play plugin in vLLM, with a lightweight expert-folding CUDA kernel, delivering up to 1.41x TTFT and 2.45x TPOT speedups while retaining about 99% of the original average quality.

Everything a language model sees is tokens. The serving stack knows what each span is — user input, tool output, instructions — but the model must keep track of that itself, and it can lose track or be confused: text can be written to read like anything. Prompt injection is a natural exploit of this phenomenon. By scrambling the model's understanding of span identity, an attacker can induce unwanted and potentially dangerous actions. Adding a non-textual channel to the model's input — a way to communicate span identity beyond text — mitigates this class of attack. We thus introduce a general steering technique called Semantic Overlays: small learned adapters applied at chosen prefill positions to a frozen model's residual stream. Laying an overlay over a span creates an out-of-band annotation channel that cannot be replicated by tokens. Unlike steering vectors, Semantic Overlays are trained, adaptable, and selectively applied. An overlay can encode complex semantics that reshape how the model perceives the marked span: asked to copy a code snippet under an overlay asserting that it is in a different programming language than it is, the model rewrites the snippet, faithfully, in the asserted language. Overlays are also composable, allow for transparent reading of underlying content, and can carry complex payloads — including imperatives that the model will follow. An overlay which marks a span as "non-executable" defends against the broad class of prompt injections that add instructions in untrusted context. We report strong results on prompt injection benchmarks: SEP separation rises from 24.3% to 96.5% with utility unchanged (our scoring rule; we also correct a defect in the published grader), TensorTrust attack success rate falls from 34.8% to 6.6%, and all four PIArena attack families drop to 0% compliance, all while marked spans stay readable (92.5% exact copy rate).

Single-token autoregressive decode on CPUs is bound by memory bandwidth, not arithmetic: a modern CPU sustains roughly 1 TFLOP/s of compute but only about 50 GB/s from main memory, and each generated token must stream every active weight once. This report argues that the most effective response is to co-design the model architecture and the inference runtime together. It presents cflow, a CPU-first streaming engine, alongside a family of pipeline-native transformer architectures whose inter-layer dependency graphs are constructed to permit a vertical, stage-major execution schedule. cflow stores weights as L2-sized tiles in compute-consumption order, reads only the top-k experts of each mixture-of-experts layer, fuses projections, and executes a delay-aware schedule from per-model dependency parameters. Across five architectures trained on TinyStories, one (arch2_4_combined) achieves a 2.00x reduction in critical-path weight bandwidth (9.00 to 4.50 MB/token) within 0.24 perplexity of the best candidate, and the tile layout incurs 7.29x fewer L1-data read misses than a row-major baseline. On a 30.9-billion-parameter pipeline-native MoE, cflow decodes at 5.94 tokens/s (tok/s) on a 32-vCPU Ice Lake server, ahead of llama.cpp (4.75) and the vLLM CPU backend (1.65) on comparably sized dense models. Realizing the expert-delay window as asynchronous I/O overlap on a disk-resident expert tier yields a further net win of up to 1.68x, matching the overlap model within 1%. Measurement refutes one of the eight design claims and leaves a second inconclusive; both are reported in full, with the conditions under which they would hold.

Recent Vision-Language Models encode high-resolution images into long visual token sequences, incurring prohibitive prefill costs. To compress them, existing methods score each visual token by averaging text-to-visual attention uniformly across all heads, which assumes every head matches the query. However, our empirical analysis shows that misaligned heads dominate the average, amplifying background tokens and drowning out fine-grained cues. To address this, we propose PAQ (Prompt-Grounded Attention Quality), a metric quantifying how well each head aligns the prompt with image regions. Built on PAQ, our pruning proceeds in three stages. Given a target FLOPs budget, we first partition the transformer layers into groups and allocate a visual token budget to each. Within each group, we then aggregate per-head attention maps via PAQ-weighted softmax into a group-level matrix. Finally, we score visual tokens by this matrix's magnitude and retain the allocated budget per group. By weighting heads with PAQ, our method scores tokens by attention signals that more faithfully reflect prompt relevance, rather than diluting them through uniform averaging. Across 18 benchmarks, our method delivers state-of-the-art trade-offs. Specifically, on LLaVA-1.5-7B (9 tasks), retaining only 5.6% tokens preserves 99.1% of the original performance, surpassing the strongest baseline AutoPrune by 4.2 points. Code is available in https://github.com/baokou-fw2/HAP.

We introduce Giga-Embeddings, a family of text embedding models designed to combine strong retrieval quality with efficient serving. Its largest member is a sparse 10B-parameter Mixture-of-Experts encoder with approximately 1.8B active parameters per token. Across English, Russian, multilingual, and code MTEB benchmarks, this model achieves the strongest aggregate performance within the family on all four evaluated suites. In our vLLM benchmark with 1024-token inputs, it processes 114.5k tokens per second, providing 25 percent higher throughput than the dense 3B model and 1.56-2.65x the throughput of the evaluated external systems. The family also includes a dense 3B encoder and a distilled 480M encoder for tighter compute and memory budgets. We train the compact model using a dimension-agnostic objective that aligns teacher and student similarity distributions. The resulting 480M model scores 70.98 on Russian MTEB, surpassing FRIDA while using 42 percent fewer parameters. We release all three model checkpoints.

Masked diffusion language models (dLLMs) can in principle generate text faster than autoregressive (AR) models, since they denoise many tokens at once. Recent systems have begun building serving infrastructure for dLLMs, but none first measure how these models behave under real, concurrent serving load. Serving systems built without this grounding risk carrying over assumptions from AR serving that may not hold for dLLMs. We characterize dLLM serving to close this gap, using LLaDA-8B-Instruct with a D2F (Discrete Diffusion Forcing) LoRA adapter on a single NVIDIA H200 GPU, evaluated on GSM8K and HumanEval. We report three findings. First, request difficulty, the number of denoising steps a request needs, is discrete rather than continuous: requests fall into 11 fixed step-count levels (178 + 29k), and no signal we test predicts the level before generation starts (best R2 = 0.150). Second, benchmarks with short generation budgets below 320 tokens understate serving variance, since requests are cut off before the latency spread appears. Third, only 24% of single-request wall-clock time is GPU computation; the rest is CPU-side dispatch overhead. Batching mainly helps by amortizing this overhead: sharing one forward pass per denoising step improves throughput by 16.0x at batch size 16 over a per-request-dispatch baseline. We also argue structurally that output quality should not degrade with batch size, stating three assumptions this rests on; we measure 74 to 76% GSM8K accuracy at single-request scale. Finally, we derive a batch-timeout rule for fixed-fill synchronized batching under Poisson arrivals. Together, these results show that serving diffusion language models needs parallelism at the level of each denoising step, which differs from AR serving in how admission and eviction interact with an already shared forward pass.

Storage-backed inference is easy to overclaim: process RSS excludes charged page cache, process-local device readings exclude board-wide use, and successful generation does not establish correct asynchronous reuse. We present a falsifiable certificate separating representation, semantic demand, scheduler requests, and traffic while naming resource authorities, exactness horizons, and tested reuse transitions. At one Qwen3-Next identity, the stock router selects all 48 x 512 managed layer-expert objects during 32K prefill. Their duplicate-free, overlap-free canonical union gives a 43.59375 GiB semantic-demand lower bound, exceeding the declared 34 GiB full-residency envelope and the 11 GiB host-hard plus physical 24 GiB-device envelope. LRU64 execution stays within its host-hard/GPU-audited contract. Against one prespecified zero-cache oracle, all 64 token IDs, every byte of 64 complete 151,936-float logit rows, 3,408 route events, response bytes, and recorded consumer and destination identities are exact. Recurrent-state and upstream-runtime equality are excluded. In a matched source campaign, the buffered path completes exactly but reaches the 11 GiB host ceiling and records 33,481 memory.max events. The blocking one-window direct path and complete eight-window asynchronous component are exact with positive margin and zero limit events. Across six counterbalanced pairs, the complete asynchronous component takes 32.3% of the blocking direct path's wall time at identical physical source bytes per output; the comparison jointly changes queue depth, overlap, and lifecycle implementation. Separately, fourteen prespecified control/fault cells pass across later Qwen3-Next and Gemma 4 binaries, supporting fail-closed behavior only for the named transitions. The principal experiment is one fixed model, workload, runtime, device, and 64-output horizon.

Long-context inference in large language models (LLMs) is increasingly limited by the memory required for the key-value (KV) cache. KV cache compression addresses this problem by reducing the storage cost of previous tokens. Among existing approaches, low-rank compression is particularly attractive because it represents every token in reduced dimensions. Previous low-rank methods typically derive fixed projection spaces from model weights, construct fixed spaces from calibration activations, or construct a shared basis over a broad cache region. Such representations may not capture detailed but important information. We partition each per-head KV cache into fixed-length logical pages and observe substantial low-rank structure within individual pages. Based on this observation, we propose PuzzleKV, a training- and calibration-free method that treats each completed page as an independent compression unit. PuzzleKV decomposes pages within each layer and KV head, computes attention directly over dense and factorized pages, and incrementally compresses newly eligible pages during autoregressive decoding. Experiments across models, context lengths, and benchmarks demonstrate the effectiveness of PuzzleKV under matched storage budgets. At approximately 60% of the original KV cache storage, PuzzleKV achieves more than 96% of Full KV performance across both evaluated models and all benchmark settings, with substantial gains over Global SVD on RULER and competitive performance on LongBench. To achieve a more aggressive compression ratio, PuzzleKV can be further combined with quantization while retaining more than 93% of Full KV performance using only 18.7% of the original storage.

The key-value (KV) cache is a primary capacity and bandwidth bottleneck in long-context LLM serving. We present Minima-KV, a retention-preserving hierarchy for mixed-format paged attention. Recent and protected Anchor pages remain in FP8, while older non-anchor pages move to packed TQ3; every live-request page remains addressable. Format-specific kernels compute partial attention states and combine them through a globally normalized online-softmax merge, enabling direct heterogeneous decode without a cache-sized dense shadow. Across separate, configuration-bound Qwen3.6-27B profiles on a single 96-GB NVIDIA RTX PRO 6000 Blackwell GPU, deployment accounting reports 18.3 KiB of attention KV per live token, corresponding to 3.50x compression relative to BF16 and 1.75x relative to FP8. A materializing quality profile matches its dense control on 16K RULER needle-in-a-haystack tasks. On the same 503-question LongBench v2 set, measured deltas are -0.80, -0.60, and -0.40 percentage points at 16K, 32K, and 64K. A separate single-pair direct-decode canary with two 59,008-token requests measures 3.625x active-KV compression and 0.9821x throughput relative to its control, routes all 16 full-attention layers without fallback, and retains no dense shadow. These results establish a practical mixed-format path for compressing long-context state without evicting live-request KV pages.

An LLM serving engine sizes its key-value (KV) cache once, at startup, permanently setting aside a reserve for the worst-case prefill activation. During decode-dominant phases that reserve sits idle, yet it cannot be handed to the KV pool because it is exactly the memory a large prefill needs. We ask whether this reserve is reclaimable, and build a mechanism to test it. Our elastic KV cache lends the reserve to the KV pool during decode and returns it before prefill, driven by the scheduler's one-step-ahead view of the next batch. It is pure userspace on the CUDA virtual-memory path: two physical handles mapped into one contiguous virtual range per layer, so the attention kernel is unchanged and no driver patch is required. It decommits in a few milliseconds and recommits in tens of milliseconds, works with CUDA graphs and prefix caching, and never triggers an out-of-memory event. A static commit of the same memory is unsafe, crashing on prefill bursts, which makes the dynamic toggle necessary. Having built the mechanism, we test the premise it rests on and report an honest negative result. It only pays off if a small prefill chunk size badly hurts prefill latency. In a controlled experiment injecting long prompts into a live decode load, that penalty is small (median time-to-first-token differs by about 1% between chunk sizes of 8192 and 32768 tokens), because prefill is compute bound and decode consumes only about one token per sequence per step. Simply lowering max_num_batched_tokens recovers more KV than the controller does, at nearly equal latency. The reserve also dilutes under tensor parallelism, from 16% of KV at TP1 to 2.7% at TP4. We state precisely when reclaiming the reserve could still help, and release the mechanism as a reusable userspace elastic-VMM allocator.

Standard decoding rules for autoregressive language models promote diversity by rescaling the full next-token distribution or truncating its low-probability tail. These strategies overlook a common regime of open-ended generation in which several continuations are plausible but too much probability mass remains concentrated on the most generic choice. We introduce XTC (Exclude Top Choices), a lightweight head-aware decoding operator that targets this regime directly. XTC identifies tokens whose probabilities exceed an absolute plausibility threshold $τ$: when at least two qualify, it removes the dominant eligible choices with probability $ρ$ and retains only the weakest plausible alternative before renormalization. Across 60 experiments on Gemma 3 27B Q4, Gemma 3 12B Q6, and DeepSeek R1 14B Q6, with scaling validation on Llama 3.3 70B Q4, XTC improves the diversity-repetition Pareto frontier. On creative generation, Distinct-2 increases by 11--15% and repeat trigrams decrease by 27--47% across the four models. Combined with temperature scaling, gains reach 38% in Distinct-2 and 71% in repeat-trigram reduction over baseline. A blinded Amazon Mechanical Turk study with 150 Master raters yields a 62.3% creativity preference for XTC ($p<10^{-4}$) without reduced fluency, while a GPT-4o control judge reproduces the Anthropic-judge direction on every measure. On IFEval with Llama 3.3 70B Q4, XTC preserves prompt-level strict accuracy within 1.7 percentage points of baseline while recovering most of the diversity gain; a temperature setting matched on Distinct-2 reduces IFEval by 8.8 points. The effect is additive with temperature and repetition penalties, robust across quantization levels and model families, and consistent across twelve prompt genres. XTC has been adopted by llama.cpp, ExLlamaV2, and text-generation-webui.

Embodied Reasoning constitutes a fundamental capability of embodied intelligence, serving as the basis for autonomous perception, reasoning, and interaction within physical environments. Recent studies have shifted the paradigm of embodied reasoning from static perception toward dynamic exploration, where agents acquire task-relevant information through interactions with the environment. However, existing active reasoning approaches generally generate exploration trajectories incrementally without long-horizon planning. Even recently emerged test-time scaling frameworks often resort to myopic, single-step lookaheads, which struggle to resolve the delayed feedback inherent in complex, occluded spatial environments. To address this limitation, we propose ParallelWorld, a multi-horizon test-time scaling framework for embodied reasoning. Instead of greedy, single-step trials, ParallelWorld empowers agents to simulate and evaluate multi-step future trajectories in parallel before committing to an action. Specifically, we introduce a verifier-guided tree-search paradigm. Starting from the current state, ParallelWorld branches into multiple parallel trajectories and rolls them out continuously across a multi-step horizon. At each simulation step, a verifier agent evaluates the intermediate state transitions, dynamically pruning unpromising branches and prioritizing paths with the highest information gain. Once the multi-step prospective simulation is complete, the agent synthesizes the long-horizon outcomes to commit to the optimal action sequence. Finally, an answer agent performs reasoning over the selected trajectory to produce the final reasoning. Extensive experiments on ESI-Bench demonstrate that ParallelWorld consistently improves active perception and reasoning performance.

Hybrid CPU architectures have entered the mainstream desktop computing with the announcement of Intel's Alderlake architecture. Such a transition to heterogeneous CPU architecture has various performance and power implications on existing parallel workloads. In this paper we study the effects and impact of hybrid core and cache architecture on the performance of highly parallel HPC workloads. We also illustrate interesting thread scaling behavior for parallel workloads and describes the reason for such behavior both qualitatively and quantitatively. We also explore the impact of hybrid cache architecture on parallel shared data HPC applications. Finally, we illustrate that 1) parallel applications with work imbalance (i.e., threads in application perform different amount of work) scale better across hybrid cores when thread affinity is disabled and 2) hybrid cache architecture has very little impact on parallel shared data applications except for some workloads with locks. This work lays the foundation for our future work which focuses on extending this work to model parallel workloads and hybrid CPU architectures to improve their performance in terms of execution time, memory usage and power consumption.

While recent advancements in TinyML have significantly reduced the computational complexity of on-device vision pipelines, image acquisition remains a dominant contributor to system-level energy consumption and memory footprint. In vision-enabled IoT platforms, the image sensor consumes energy comparable to the inference engine, thereby offsetting algorithmic efficiency gains. Consequently, current designs face a fundamental trade-off: continuous and always-on sensing incurs prohibitive energy consumption, whereas aggressive duty cycling increases latency and risks missing transient events. This work presents an energy-proportional, context-aware vision IoT node that addresses this challenge through a heterogeneous multimodal dual-camera architecture. Detection and recognition are decoupled by combining an event-based imager operating asynchronously in an energy-efficient always-on wake-on-motion mode together with an RGB imager. Deployed on a low-power microcontroller, a novel TinyissimoYOLOv12 is introduced for efficient and accurate object detection. By activating the high-power image acquisition and processing stages only upon sparse visual triggers, the proposed architecture improves efficiency and latency, eliminating redundant sensing while maintaining continuous monitoring coverage. Experimental results demonstrate an energy consumption of only 222$μ$Wh. Upon a motion trigger, the system completes a full sense-to-report cycle-RGB acquisition, object detection across 80 classes, and LoRa telemetry-with a total energy consumption of 28.7mJ. The network achieves up to 32.3% mAP with a model size of 1 million parameters. At a 1% daily activity ratio, the platform achieves a three-month operational lifetime with a 1.85Wh battery, enabling always-on visual monitoring in a place-and-forget scenario through autonomous edge intelligence.

Text-driven human motion synthesis has made substantial development with two core modules of motion representation and generative architecture. For representation, Vector Quantization (VQ)-based methods compress motion data into discrete tokens while latent-based models operate directly in continuous space. However, both of these representations exhibit significant limitations. VQ-based methods suffer from inherent information loss, which compromises the quality, diversity, and generalization of generated motions, while continuous representation on holistic whole-body motion hinders part-level flexibility. For architecture, diffusion and autoregressive diffusion models have demonstrated their superiority, yet the fine-grained controllability over individual body parts is also limited. Thus, we propose a unified spatiotemporally decoupled framework named DeMoDiff, which jointly redesigns representation and architecture. To enhance representation extraction capabilities and offer greater part-level controllability, we present a spatial-temporal VAE that encodes each body joint rather than compressing the whole-body motion into a single latent space. Then, we incorporate spatial-temporal masking and attention mechanisms into an autoregressive diffusion generator, achieving both generative capability and controllable editability. Extensive experiments on the HumanML3D and KIT-ML datasets demonstrate that our model achieves state-of-the-art reconstruction performance and compelling motion generation results. Moreover, our framework demonstrates strong temporal and spatial editing capabilities, further validating its effectiveness. Our project page: https://rex0191.github.io/DeMoDiff/

Processing-in-Memory (PIM) promises to reduce data movement overhead by executing computation in or near memory, but its realized application speedup remains highly design-dependent. Non-offloadable host execution, host-PIM transfers, limited PIM capacity, and device programming latency can limit end-to-end speedup, making fast early-stage design-space exploration (DSE) essential. However, existing PIM evaluation methods remain limited: circuit- and device-level tools cannot capture these end-to-end PIM performance factors, while cycle-accurate simulation is too slow for iterative DSE. To address this gap, we present VIPER, a unified, lightweight, and architecture-aware performance evaluation framework for PIM DSE. VIPER profiles host execution once and combines the measured host behavior with a PIM-aware analytical engine that sweeps PIM-side parameters across candidate designs. It supports both Processing Near Memory (PNM) and Processing Using Memory (PUM) under task-offloading and data-triggered execution by capturing host-PIM transfer, array access, in-memory computation, device programming latency, and capacity-induced partitioning, providing rapid architecture-aware performance estimates for iterative DSE without repeated cycle-accurate simulation. We validate VIPER against a commercial UPMEM system and more than 400 cycle-accurate gem5 configurations. VIPER predicts the UPMEM offloading decision and break-even region a priori, and, with a refined transfer model, captures the measured peak-and-rolloff behavior with 12% mean speedup error across the DPU sweep (6% up to the 256-DPU peak). Against gem5, VIPER achieves less than 10% error while reducing evaluation time from hours to under one minute. Case studies of UPMEM, ReRAM/FeFET crossbars, and IMCRYPTO show that architecture-aware DSE reveals key performance trade-offs that device-level evaluation misses.