生成模型与LLM推理优化

Diffusion Transformer (DiT) has emerged as a powerful model architecture for generating high-quality images and videos. In the case of video DiT, 3D Spatio-Temporal Attention increases token length in proportion to the number of frames, sharply increasing computational cost. Token reduction methods mitigate this cost by exploiting spatial redundancy, but existing approaches rely on inaccurate similarity estimates and lightweight matching algorithms, resulting in poor matching quality and only marginal acceleration. To overcome these limitations, we propose ORBIS, an SW-HW co-designed accelerator for video DiT. ORBIS leverages the output activation from the previous timestep to obtain more accurate inter-token similarity, substantially improving matching quality and enabling a higher token reduction ratio. We further introduce a Distribution-Aware Token Matching (DATM) algorithm that captures global token distribution and explicitly minimizes token-pair loss for additional gains. To fully hide DATM latency, we design specialized, deeply pipelined hardware and minimize its hardware cost through quantization, occupying only 2.4% of total area with negligible accuracy loss. Extensive experiments show that ORBIS achieves about 2x higher token reduction ratio than the state-of-the-art approach, AsymRnR, while delivering up to 4.5x speedup and 79.3% energy reduction compared to an NVIDIA A100 GPU.

Modern serving systems for Mixture-of-Experts (MoE) models adopt hybrid data-expert parallelism: expert parallelism (EP) shards experts across GPUs to scale capacity, while data parallelism (DP) replicates attention layers across instances to process independent requests. Existing systems bind each request's attention, MoE communication, and KV cache to a single instance. Because attention latency scales with KV cache size while MoE communication latency scales with batch size, this binding cannot balance both simultaneously, producing EP stragglers; it also fragments KV memory across instances, inflating tail latency under long contexts. While existing context parallelism (CP) mitigates these constraints, its uniform parallelism degree incurs prohibitive communication and attention-side overheads. We present \work, which decouples MoE communication from KV cache placement and achieves dual balance through dynamic context parallelism (DCP). DCP assigns each request a context-parallel degree sized to its KV footprint: long requests distribute attention across multiple instances; short requests remain local. This dynamic parallelism effectively liquefies the KV cache across the cluster, balancing both the per-instance KV cache occupancy and batch sizes without unnecessary load-balancing costs. To bridge DCP with static execution, \work introduces an ahead-of-time (AOT) graph engine paired with a custom routing-based communication backend. Experimental results show that \work maintains up to $1.88\times$--$3.27\times$ higher request rates under strict time-per-output-token (TPOT) service level objectives (SLOs). Furthermore, \work significantly mitigates stragglers, reducing P99 tail latency by up to $1.79\times$--$2.12\times$.

Autoregressive (AR) visual generation has achieved remarkable performance but suffers from high memory usage and low throughput, as it requires caching previously generated visual tokens. Recent research has shown that retaining only a few lines of cache tokens can maintain high-quality images while significantly reducing memory usage and improving throughput. However, these methods allocate a fixed budget to each attention head, overlooking the heterogeneity among attention heads, leading to suboptimal memory allocation. In this paper, we observe that attention heads across different layers exhibit diverse attention patterns, where some heads focus on local neighborhoods while others capture broader contextual dependencies. Based on this insight, we propose a novel head-aware key-value (KV) cache compression framework for autoregressive image generation, called HeadKV, which assigns smaller budgets to locality-biased heads and larger budgets to heads with broader attention. A key challenge lies in identifying the type of each attention head to guide cache compression. We further observe that, within the same layer, each head exhibits consistent attention patterns across token positions, i.e., a head's behavior for early tokens remains consistent with that for later tokens. This insight suggests that head types can be identified during the early stage and reused for KV compression throughout generation. Its advantage is that it requires no additional training or dataset-level statistics and generalizes seamlessly across different inputs. Moreover, we design a Stratified Token Eviction strategy to effectively preserve long-range information. Extensive experiments demonstrate its effectiveness across multiple autoregressive image generation models.

Diffusion Transformers (DiTs) have revolutionized high-fidelity video generation, yet their $\mathcal{O}(L^2)$ attention complexity poses a formidable bottleneck for long-sequence synthesis. While recent sparse-linear attention hybrids aim to mitigate this, their performance severely degrades at extreme sparsity due to the "RoPE Dilemma": standard linear attention fails to preserve the orthogonal relative-position structure of 3D Rotary Position Embeddings (RoPE), neutralizing vital distance awareness. To address this, we propose RoPeSLR, a 3D RoPE-guided Sparse-LowRank attention framework. We establish that under empirically validated assumptions, the DiT attention manifold admits a decoupling into a high-frequency semantic spike set (bounded by $\mathcal{O}(L^{3/2})$ sparsity) and an extreme low-rank ($\mathcal{O}(d_h \log L)$) background continuum. Guided by this structural prior, RoPeSLR eschews standard linear attention for a head-wise low-rank parameterization equipped with a learnable 3D Absolute Positional Embedding (PE) injection, seamlessly synthesizing long-range relative distance decay. By guaranteeing sub-quadratic sparsity and sub-linear rank growth, RoPeSLR is exceptionally suited for scaling to ultra-long video inference. Extensive evaluations validate this scalable superiority: at 90% sparsity, RoPeSLR achieves up to $10\times$ fewer FLOPs on Wan2.1-1.3B and delivers a $2.26\times$ end-to-end inference speedup on the ultra-long 100K+ token sequences of HunyuanVideo-13B, all while maintaining near-lossless generation fidelity (less than 1.3% average VBench degradation).

Computer-use agents (CUA) automate tasks specified with natural language such as "order the cheapest item from Taco Bell" by generating sequences of calls to tools such as click, type, and scroll on a browser. Current implementations follow a sequential fetch-screenshot-execute loop where each iteration requires an LLM call, resulting in high latency and frequent errors from incorrect tool use. We present agent just-in-time (JIT) compilation, an alternative that compiles task descriptions directly into executable code that is free to include LLM calls, tool calls, and parallelization. Our approach comprises three components: (1) JIT-Planner, which generates multiple code plans, validates each against tool specifications, and selects the minimum-cost candidate; (2) JIT-Scheduler, which explores parallelization strategies via Monte Carlo cost estimation from learned latency distributions; and (3) an invariant-enforcing tool protocol specifying precondition and postcondition state requirements that reduce the rate of generating plans with incorrect tool use. Across 5 web applications, JIT-Planner achieves $10.4\times$ speedup and $+28%$ accuracy over Browser-Use, while JIT-Scheduler achieves $2.4\times$ speedup and $+9%$ accuracy over OpenAI CUA.

Running language models in the browser presents a unique opportunity to build efficient, private, and portable AI applications, but requires contending with constrained memory availability and heterogeneous hardware targets. To realize this opportunity, we present Llamas on the Web (LlamaWeb), a WebGPU backend for llama$.$cpp that enables memory-efficient and performance-portable LLM inference across a wide range of model weight formats in the browser. Our design significantly reduces memory overhead through static memory planning and efficient model loading, addresses cross-device variability through a tunable kernel library, and introduces templated GPU kernels that support performant implementations of numerous quantization formats, enabling broad model support and extensibility to new formats. We evaluate LlamaWeb on 16 devices from 8 vendors, collecting data from 10 language models and four model weight formats. We compare LlamaWeb against existing browser-based LLM frameworks and find that LlamaWeb requires 29-33% less memory across several combinations of device, browser, and operating system. We also evaluate LlamaWeb's performance against these frameworks and find that it increases decode throughput by 45-69% across four GPUs from separate vendors. In addition, we compare LlamaWeb's performance against other llama$.$cpp backends, where it is competitive with and even beats vendor-specific backend performance on some devices.

Discrete diffusion models excel at visual synthesis but rely on slow, iterative decoding. Existing single-step distillation methods attempt to bypass this bottleneck, either by training auxiliary score networks that effectively double compute, or by introducing specialized parameterizations and multi-stage pipelines that fragment optimization. In this paper, we introduce Fixed-Point Distillation (FPD), an end-to-end framework that constructs local correction targets by partially corrupting the student's one-step draft and refining it with a single teacher step. To compute the training objective in a semantically meaningful space, we lift discrete tokens into continuous features and apply a multi-bandwidth drift loss that iteratively accumulates these corrections. To backpropagate through the discrete bottleneck, we employ a straight-through estimator that feeds exact hard-sampled tokens to the teacher and decoder during the forward pass, ensuring that training and inference operate on the same codebook manifold, while routing continuous gradients back to the student logits. This fully differentiable pathway additionally accommodates an optional unconditional adversarial objective to enhance perceptual realism. Evaluations on both class- and text-conditional generation validate the effectiveness of our framework. FPD achieves competitive visual fidelity and structural alignment within a single inference step, narrowing the gap to multi-step teachers while outperforming existing discrete distillation baselines.

Large language model (LLM) inference has become a dominant workload in modern data centers, driving significant GPU utilization and energy consumption. While prior systems optimize throughput and latency by batching, scheduling, and parallelism, they largely treat GPU power as a static constraint rather than a controllable resource. In this paper, we present a power-aware runtime for LLM serving, PALS, that treats GPU power caps as a first-class control knob and jointly optimizes them with software parameters such as batch size. The system combines lightweight offline power-performance models with a feedback-driven controller to select configurations that satisfy throughput targets while maximizing energy efficiency. We implement PALS within an existing LLM serving framework, vLLM, demonstrating that it requires no model retraining or API changes. Across multi-GPU systems and both dense and mixture-of-experts (MoE) models, PALS improves energy efficiency by up to 26.3%, reduces QoS violations by 4x to 7x under power constraints, and tracks dynamic power budgets. These results highlight the potential of integrating power control directly into LLM inference runtimes, enabling energy-proportional and grid-interactive AI systems.

Inference in diffusion large language models (dLLMs) is computationally expensive, as full self-attention must be repeatedly executed at each step of the denoising process without KV cache. Recent sparse attention methods for dLLMs mitigate this cost via block-sparse computation, which is applied only in later iterations when model performance is less sensitive to coarse-grained sparse approximation, but yields limited improvements in computational efficiency and acceleration. This motivates a finer-grained sparsification strategy that can be applied from earlier iterations and leverages reusable sparsity patterns, enabling further efficiency gains. In this work, we introduce PulseCol, a periodically refreshed column-sparse attention method for accelerating diffusion language models. PulseCol replaces coarse block-level sparsity with a finer-grained column-sparse structure, allowing important attention interactions to be retained more precisely while exposing greater sparsity. Built on this column-level formulation, PulseCol further identifies sparse patterns at the early denoising step and reuses them across subsequent iterations, refreshing them only at a small number of intermediate steps to track the evolution of sparse attention patterns during denoising. Experiments show that PulseCol achieves higher sparsity and greater practical speedup than prior sparse attention methods for dLLMs, while maintaining model quality. Enabled by optimized GPU kernels for column-sparse attention, PulseCol delivers up to 1.95$\times$ end-to-end speedup over FlashAttention across several context lengths.

The key-value (KV) cache dominates memory bandwidth and footprint in long-context autoregressive inference. Recent rotation-preconditioned codecs (TurboQuant, PolarQuant) show that a structured random rotation followed by a per-coordinate scalar quantizer matched to an analytically tractable marginal is a near-optimal recipe for KV compression. OCTOPUS advances this paradigm through joint quantization of rotated coordinate triplets. Each triplet's direction is mapped to a square via an octahedral parameterization, and the two resulting coordinates and the triplet norm are Lloyd-Max quantized against implementation-matched marginals. Optimizing the per-triplet squared error gives a strictly non-uniform bit allocation depending only on the total dimensionality of the keys. We find the finite-dimensional quality optimum with sweeps to be constant on every real decoder we test. The codec is data-oblivious, online, and deterministic given a seed. Across text, video, and audio, OCTOPUS matches or beats every prior rotation codec at every reported bit width and metric, with a lead that grows as bits drop for extreme compression. Furthermore, a fused Triton implementation reconstructs keys on the fly without materializing the uncompressed key, so the codec adds no decode-time bandwidth or latency over the existing dequantization. Project Page: https://octopus-quant.github.io/

Although existing video editing methods are generally feasible, they often require many costly iterations and still struggle to deliver high-quality yet satisfying editing results. We attribute this limitation to the prevalent data-to-data paradigm, which is less compatible with modern generative models than noise-to-data generation. To address this gap, we revisit video editing from a noise-to-data perspective and propose Streaming-Generation-based Video Editing (StreamGVE), which preserves few-step sampling while seamlessly injecting source-video conditions. Built on pre-trained streaming generation models, StreamGVE introduces dual-branch fast sampling with a self-attention bridge and cross-attention grounding/boosting to satisfy both sampling and conditioning requirements. We further propose source-oriented guidance to improve target-generation quality, and a visual prompting strategy to enhance editing flexibility and practicality. The method is effective, robust, and generalizable across different models. Extensive experiments on diverse video editing tasks show that StreamGVE consistently outperforms existing approaches, even in few-step settings with minimal time cost.

Emerging cyber-physical systems increasingly require low-latency inference from streaming sensor data while maintaining models that reflect complex and evolving physical processes. In many domains, however, model updates depend on high-fidelity simulations and training executed on remote high-performance computing (HPC) systems under batch scheduling. This creates a fundamental mismatch between the responsiveness required at the edge and the cost, throughput, and availability of simulation-driven model updates. We present RBF (Reverse Backfill), a hybrid edge-HPC learning and inference architecture that integrates low-latency edge inference with asynchronous, simulation-driven model improvement. RBF targets simulation-bounded settings in which model updates are constrained by simulation throughput and HPC scheduling delays, and reinterprets HPC backfilling by using opportunistic computation to improve model accuracy rather than system utilization. RBF decouples inference from simulation and training by deploying lightweight surrogate models at the edge while incorporating improved models asynchronously as they become available. The architecture supports pluggable surrogate models and orchestrates computation across heterogeneous infrastructure spanning edge devices, private 5G, cloud, and HPC resources. We instantiate RBF using a real-world digital agriculture deployment that couples edge sensing with computational fluid dynamics (CFD) simulations to infer airflow patterns in a large agricultural screenhouse. Our evaluation characterizes end-to-end system behavior under realistic constraints, quantifying simulation latency, training cost, inference throughput, and the impact of delayed model updates on prediction accuracy. Results demonstrate that RBF enables continuous, low-latency inference while improving model fidelity over time despite delayed and irregular model updates.

Large language models (LLMs) are increasingly deployed on mobile devices, where Neural Processing Units (NPUs) necessitate fully static quantization for optimal inference efficiency. However, existing post-training quantization (PTQ) methods predominantly rely on dynamic activation quantization, rendering them incompatible with NPU hardware constraints. To bridge the gap between high-fidelity PTQ and NPU-constrained inference, we propose Quant.npu, a integer-only fully static quantization framework. It incorporates learnable quantization parameters and rotation matrices, enabling low-bit activation-weight quantization without runtime quantization parameters re-computation. Crucially, we identify that initialization and selective optimization of quantization parameters is pivotal for optimization stability, as improper initialization and naive joint optimization induce gradient instability that disrupts the optimization of rotation matrices. To address this, we propose a rotation-and-bit-width-aware initialization tailored to diverse activation profiles and a distribution-aware selective optimization (two-stage quantization pipeline) tailored to rotated and unrotated tensors. Furthermore, we introduce a sensitivity-guided adaptive mixed-precision scheme to balance accuracy with inference efficiency. Extensive experiments on real-world mobile NPUs demonstrate that Quant.npu achieves comparable accuracy to state-of-the-art methods, while reducing inference latency by up to 15.1%.

MXFP4 arithmetic can dramatically accelerate reinforcement learning (RL) post-training of large language models (LLMs), yet the quantization error introduces severe accuracy degradation. Existing work treats the quantization error as a monolithic noise term, missing the distinct mechanisms upon interpreting how quantization error damages training. We prove an exact three-way decomposition of quantization error and show how each component dominates a distinct RL training pathway. Our theoretical and empirical analysis decomposes the MXFP4 quantization error into three additive components: "scale bias" from power-of-two rounding, "deadzone truncation" from zeroing small values, and "grid noise" from rounding to the nearest 4-bit grid. Each component dominates a distinct RL failure mode: scale bias accumulates multiplicatively through the backward pass, affecting gradient accuracy; deadzone truncation degrades rollout quality; and grid noise raises the policy's entropy. We combine corrections that are RL failure mode-targeted but not component-exclusive: Macro-block scaling to reduce scale bias, Outlier Fallback recovers deadzone entries, but also partially reduces scale bias induced error, and Adaptive Quantization Noise (AQN) for controlling the policy entropy. On Qwen2.5-3B dense and Qwen3-30B-A3B-Base mixture-of-experts model, the targeted corrections recover BF16 accuracy to within 0.7% and 3.0% respectively.

Recent advancements in large language models (LLMs) have shown their remarkable capacities in many NLP tasks. However, their substantial size often presents challenges for deployment. This necessitates efficient techniques for model compression, with quantization emerging as a prominent solution. Despite its benefits, the exact impact of quantization (from 2- to 6-bit) on the performance and accuracy of LLMs remains an active area of research. This paper investigates the performance of eight LLMs at various quantization levels, focusing on tasks such as MMLU-Pro for knowledge processing and reasoning, CRUXEval for code comprehension, and MuSR for reading comprehension. Our results show a consistent trend where higher precision (e.g., 8-bit Q8_0) yields improved performance, albeit with diminishing returns. Aggressive quantization (e.g., 2-bit Q2_K) usually retains acceptable accuracy, though some models show a substantial loss in performance. Our findings indicate that while lower bit precision generally reduces performance, the impact varies across models and tasks. Larger models show greater resilience to aggressive quantization, but can still undergo significant drops at lower precision levels. Mid-sized models in the 7-9 billion parameter range strike an optimal balance between efficiency and resource usage. Such results provide insights into the trade-offs between model size, quantization, and performance.

Speculative decoding accelerates memory-bound LLM inference without quality degradation by using a fast drafter to propose multiple candidate tokens and the target model to verify them in parallel. However, conventional sequential speculative decoding suffers from mutual waiting between drafting and verification, and repeated exchange of intermediate states further increases memory access overhead. Parallel speculative decoding addresses this limitation by performing drafting and verification within a single target forward pass, allowing future drafts to be prepared while current candidates are being verified. Although effective at small batch sizes, existing parallel speculative decoding methods either require costly continual pretraining with quality degradation or suffer from low acceptance rates. More importantly, this paradigm inherently suffers from uncertainty in both the bonus token and the accepted length, leading to draft verification mismatch and causing throughput gains to collapse at large batch sizes. To address these limitations, we introduce FlexDraft, a lossless speculative decoding framework that flexibly adapts to varying batch sizes through three key designs. (1) Attention Tuning enables block diffusion drafting by tuning only the attention projectors of the final few layers on mask tokens, while keeping the autoregressive path frozen to preserve the target distribution and produce high quality drafts with minimal trainable parameters. (2) Bonus-guided Calibration uses a lightweight MLP conditioned on the resolved bonus token to calibrate draft logits, mitigating draft verification mismatch caused by bonus token uncertainty. (3) Flex Decoding dynamically switches between parallel draft and verify at small batch sizes and sequential draft then verify at large batch sizes, and adjusts verification length based on draft confidence to eliminate redundant computation.

Speculative decoding (SD) accelerates large language model inference by leveraging a draft-then-verify paradigm. To maximize the acceptance rate, recent methods construct expansive draft trees, which unfortunately incur severe VRAM bandwidth and computational overheads that bottleneck end-to-end speedups. While dynamic-depth pruning can reduce this latency by removing marginal branches, it also discards potentially valid candidates, preventing the acceptance rate from reaching the upper bound of dense trees. In this paper, we identify a critical opportunity in resource allocation: the transition from dense to pruned drafting frees up significant computational budget. To break this Pareto tradeoff, we introduce Graft, a compensation framework that couples pruning and retrieval as mutually reinforcing operations. Pruning supplies sufficient budget for retrieval, while retrieval compensates for pruning-induced coverage loss and recovers accepted length. By employing a sequential `prune-then-graft' mechanism, Graft attaches highly predictive retrieved tokens into positions opened by pruning, filling the topological gaps with near-zero overhead. Graft is entirely training-free and lossless. Comprehensive evaluations show that Graft establishes a new Pareto frontier across practical deployment settings, including short-context generation, long-context generation, and large-scale models. On short-context benchmarks, it achieves up to 5.41$\times$ speedup and improves average speedup over EAGLE-3 by up to 21.8% on the large-scale Qwen3-235B. We also provide a preliminary exploration of applying Graft to the DFlash-style block drafting paradigm, offering initial evidence and insights for extending grafting beyond autoregressive draft trees.

Modern deployments of Large Language Models (LLMs) increasingly require serving multiple models with diverse architectures, sizes, and specialization on shared, heterogeneous hardware. This setting introduces new challenges for resource allocation, dispatching, and scheduling, particularly under GPU memory constraints where partial CPU-GPU offloading and preemption become necessary. While existing systems primarily optimize throughput for a single model, comparatively little work addresses multi-model scheduling under these conditions. In this paper, we present an empirical study of how different LLMs behave across hardware platforms, focusing on the performance implications of layer offloading and preemption. We show that offloading leads to strongly non-linear and model-dependent degradation in decode throughput, with smaller models exhibiting sharper sensitivity to reduced GPU residency. We further demonstrate that preemption incurs substantial overhead, largely dominated by model state reload rather than key-value cache transfer, and that this cost varies significantly across models and hardware platforms. Additionally, we highlight the role of sequence length and interconnect bandwidth in amplifying data movement and execution inefficiencies. Based on these findings, we identify a set of key features that future schedulers must consider, including model-specific offloading sensitivity, workload characteristics, and the cost structure of preemption and data transfer. These insights provide guidance for the design of next-generation LLM serving systems capable of efficiently managing heterogeneous, multi-model workloads with hybrid CPU-GPU execution.

Vision-Language Models suffer severe KV cache pressure at inference, as a single image often encodes into thousands of tokens. Most existing methods exploit token sparsity through token pruning, but permanently discarding visual content causes substantial degradation on fine-grained perception tasks. This motivates a complementary axis, feature sparsity: under a fixed KV cache budget, compressing the channel dimension preserves more visual tokens at the same memory cost. Prior Key channel pruning methods, however, face a structural trade-off: token-wise channel pruning is expressive but unstructured and slow, while head-wise approach is hardware-friendly but less robust. We resolve this with RotateK, a rotation-based structured Key channel pruning framework. RotateK applies an online PCA-based rotation that aligns token-dependent channel importance into a shared low-dimensional subspace, enabling accurate pruning under lightweight head-wise masks; a fused Triton attention kernel operates directly on sparse-channel Keys for efficient decoding. Experiments on two representative VLM backbones show that RotateK consistently outperforms prior Key channel pruning in both accuracy and decoding latency, while joint token-channel pruning improves over token-only baselines at matched KV cache budgets.

Low-bit post-training quantization (PTQ) is a pivotal technique for deploying Vision-Language Models (VLMs) on resource-constrained devices. However, existing PTQ methods often degrade VLMs' accuracy due to the heterogeneous activation distributions of text and vision modalities during quantization. We find that this cross-modal heterogeneity is distributed unevenly across channels: a small subset of channels contains most modality-specific outliers, and these outliers typically reside in different channels for each modality. Motivated by this, we propose SplitQ, a channel-Splitting-driven post-training Quantization framework. At its core, SplitQ introduces a novel Modality-specific Outlier Channel Decoupling (MOCD) module that effectively isolates salient modality-specific outlier channels with minimal overhead. To further address the remaining cross-modal distribution discrepancies, we design an Adaptive Cross-Modal Calibration (ACC) module that employs dual lightweight learnable branches to dynamically mitigate modality-induced quantization errors. Extensive experiments on popular VLMs demonstrate that SplitQ significantly outperforms existing approaches across 6 popular multi-modal datasets under all evaluated quantization settings, including W4A8, W4A4, W3A3, and W3A2. Notably, SplitQ preserves 93.5% of FP16 performance under the challenging W3A3 setting (69.5 vs. 74.3), pushing the efficiency frontier for deploying advanced VLMs. Our code is available at https://github.com/EMVision-NK/SplitQ

The rapid advancement toward long-context reasoning and multi-modal intelligence has made the memory footprint of the Key-Value (KV) cache a dominant memory bottleneck for efficient deployment. While the established per-channel quantization effectively accommodates intrinsic channel-wise outliers in Key tensors, its efficacy diminishes under extreme compression. In this work, we revisit the inherent limitations of the per-channel quantization paradigm from both empirical and theoretical perspectives. Our analysis identifies Token Norm Imbalance (TNI) as the primary bottleneck to quantization fidelity. We demonstrate that TNI systematically amplifies errors when shared quantization parameters are required to span token groups exhibiting substantial norm disparities. Instead of relying on intricate quantization pipelines (e.g., TurboQuant), we propose OScaR (Omni-Scaled Canalized Rotation), an accurate and lightweight KV cache compression framework for X-LLMs (i.e., text-only, multi-modal, and omni-modal LLMs). Advancing the per-channel paradigm, OScaR employs Canalized Rotation followed by Omni-Token Scaling to mitigate TNI-induced sequence-dimensional variance both effectively and efficiently, further supported by our optimized system design and CUDA kernels. Extensive evaluations across X-LLMs show that OScaR consistently outperforms existing methods and achieves near-lossless performance under INT2 quantization, establishing it as a robust, low-complexity, and universal framework that defines a new Pareto front. Compared with the BF16 FlashDecoding-v2 baseline, our OScaR implementation achieves a notable up to 3.0x speedup in decoding, reduces memory footprint by 5.3x, and increases throughput by 4.1x. The code for OScaR is publicly available at https://github.com/ZunhaiSu/OScaR-KV-Quant.

The rapid proliferation of AI-generated visual media has created an urgent need for efficient, trustworthy deepfake detection systems. However, existing deep learning-based detection methods rely on computationally intensive and energy-demanding inference algorithms, limiting their scalability. Here, we present a hybrid digital-analog deepfake video detection framework that combines a lightweight digital front-end with a spatially multiplexed optical decoding back-end for massively parallel analog inference through a programmable spatial light modulator. By simultaneously processing 15 or more video streams within a single optical propagation pass, the system enables high-throughput and accurate video-level authenticity prediction at reduced computational cost compared with purely digital methods. We validated this hybrid deepfake video processor using different datasets spanning classical face-swapping, real-world deepfake recordings, and fully AI-generated videos. Using a spatially multiplexed experimental set-up operating in the visible spectrum, we achieved average deepfake detection accuracy, sensitivity and specificity of 97.79%, 99.86% and 95.72%, respectively, on the Celeb-DF video dataset with 15 videos tested in parallel in a single optical pass per inference. The multiplexed optical decoder also demonstrates resilience against various types of video degradation, noise, compression, experimental misalignments and black-box adversarial attacks. Our results show that integrating optical computation into AI inference enables simultaneous gains in throughput, energy efficiency, and adversarial robustness - three properties that are difficult to achieve together in purely digital systems.

Recent advances in Video Large Language Models (Video-LLMs) have greatly expanded multimodal reasoning capabilities. However, the massive number of visual tokens extracted from long video sequences incurs prohibitive computational costs, limiting their deployment in real-world scenarios. Existing training-free token compression methods select tokens based on attention magnitude as a proxy for semantic importance, but often overlook positional bias and rely only on short-term temporal locality, leading to redundant spatio-temporal coverage and inefficient token usage. We present DynaTok, a training-free, temporally adaptive and bias-aware token compression framework that allocates token budgets across both temporal and spatial dimensions. Through a lightweight exponential moving average (EMA) memory, the Temporal Budget Allocation (TBA) module dynamically assigns fewer tokens to redundant frames and more to novel frames, capturing long-term temporal variation. The Spatial Budget Allocation (SBA) module complements this by selecting spatially diverse and semantically important features using activation-based attention maps, while leveraging a spatial memory to reduce redundancy from previously selected regions and mitigate positional bias. DynaTok integrates seamlessly with existing Video-LLMs such as LLaVA-OneVision and LLaVA-Video without retraining, and effectively preserves semantic coverage under aggressive compression. Experiments on four representative VideoQA benchmarks-MVBench, LongVideoBench, MLVU, and VideoMME-show that DynaTok retains over 95% of baseline accuracy even with a 90% token reduction, surpassing recent training-free approaches. These results demonstrate that DynaTok provides a principled foundation for efficient and robust video reasoning, paving the way toward real-time streaming video understanding with future Video-LLMs.

Diffusion Language Models (DLMs) enable globally coherent, bidirectional, and controllable text generation, offering advantages over traditional autoregressive LLMs, while scaling to ultra-long sequences remains costly. Many existing block-sparse attention methods select blocks by fixed sampling patterns over the high-resolution attention space, such as tail regions or anti-diagonal stripes. Such prior-driven sampling can miss salient tokens and introduce instability under distribution shifts. In this paper, we propose the Block Approximate Sparse Attention framework (BA-Att) with block-wise pre-downsampled operation, which identifies informative regions within a compact downsampled space, avoiding reliance on brittle positional priors. To analyze its theoretical behavior, we define an oracle post-downsample attention map and formalize the approximation error between pre- and post-downsample schemes. Based on this insight, we introduce a lightweight norm-sorting module and a covariance-compensated correction that approximates full covariance using diagonal QK variances, reducing computational complexity. Extensive experiments show that our operator achieves up to 6.95x acceleration over FlashAttention in attention computation, and maintains near full-attention performance at 50% sparsity across language models, multimodal language models, and video generation models, demonstrating strong efficiency and generalization.

Large language model (LLM) agents increasingly operate over long and recurring external contexts, like document corpora and code repositories. Across invocations, existing approaches preserve either the agent's trajectory, passive access to raw material, or task-level strategies. None of them preserves what we argue is most needed for repeated same-context workloads: reusable orientation knowledge (e.g., what the context contains, how it is organized, and which entities, constants, and schemas have historically been useful) about the recurring context itself. We introduce PEEK, a system that caches and maintains this orientation knowledge as a context map: a small, constant-sized artifact in the agent's prompt that gives it a persistent peek into the external context. The map is maintained by a programmable cache policy with three modules: a Distiller that extracts transferable knowledge from inference-time signals, a Cartographer that translates it into structured edits, and a priority-based Evictor that enforces a fixed token budget. On long-context reasoning and information aggregation, PEEK improves over strong baselines by 6.3-34.0% while using 93-145 fewer iterations and incurring 1.7-5.8x lower cost than the state-of-the-art prompt-learning framework, ACE. On context learning, PEEK improves solving rate and rubric accuracy by 6.0-14.0% and 7.8-12.1%, respectively, at 1.4x lower cost than ACE. These gains generalize across LMs and agent architectures, including OpenAI Codex, a production-grade coding agent. Together, these results show that a context map helps long-context LLM agents interact with recurring external contexts more accurately and efficiently.

Linear attention has recently gained significant attention for long-context inference due to its constant decoding cost with respect to context length. However, existing serving systems typically serve linear attention by recurrently computing and updating a large linear attention state in every decoding step. Since the state is much larger than the per-token key and value, recurrent decoding incurs substantial memory access and becomes inefficient for serving linear attention. In this paper, we propose KVBuffer, an IO-aware serving mechanism for linear attention. By buffering recent keys and values, KVBuffer enables serving systems to compute linear attention outputs in more flexible and memory-efficient ways. For decoding, KVBuffer enables chunkwise computation, which reduces average memory access and decoding latency by deferring state updates and applying them in batch. For speculative decoding, KVBuffer verifies draft tokens in parallel and avoids storing temporary states. For short contexts, KVBuffer computes attention outputs directly from buffered keys and values, without creating or updating the linear attention state. We implement KVBuffer in SGLang for Qwen3-Next. Our evaluations show that KVBuffer can reduce linear attention decoding latency by up to 45.17% and increase the maximum number of serving requests by 5x for speculative decoding when verifying four draft tokens.

We present LongLive-2.0, an NVFP4-based parallel infrastructure throughout the full training and inference workflow of long video generation, addressing speed and memory bottlenecks. For training, we introduce sequence-parallel autoregressive (AR) training, instantiated as Balanced SP, which co-designs the efficient teacher-forcing layout with SP execution by pairing clean-history and noisy-target temporal chunks on each rank, enabling a natural teacher-forcing mask with SP-aware chunked VAE encoding. Combined with NVFP4 precision, it reduces GPU memory cost and accelerates GEMM computation during training, the proportion of which increases as video length grows. Moreover, we show that a high-quality infrastructure and dataset enable a remarkably clean training pipeline. Unlike existing Self-Forcing series methods that rely on ODE initialization and subsequent distribution matching distillation (DMD), LongLive-2.0 directly tunes a diffusion model into a long, multi-shot, interactive auto-regressive (AR) diffusion model. It can be further converted to real-time generation (4 to 2 denoising steps) with standalone LoRA weights. For inference on Blackwell GPUs, we enable W4A4 NVFP4 inference, quantize KV cache into NVFP4 for memory savings, and boost end-to-end throughput with asynchronous streaming VAE decoding. On non-Blackwell GPU architectures, we deploy SP inference to match the speed on Blackwell GPUs, while the quantized KV cache can lower inter-GPU communication of SP. Experiments show up to 2.15x speedup in training, and 1.84x in inference. LongLive-2.0-5B achieves 45.7 FPS inference while attaining strong performance on benchmarks. To our knowledge, LongLive-2.0 is the first NVFP4 training and inference system for long video generation.

Vision-language model (VLM) agents increasingly rely on memory-augmented reinforcement learning to reuse experience across long-horizon tasks, yet most existing frameworks store memory as text and depend on proprietary teacher models to summarize or refine it. This design is poorly matched to spatial decision making: geometric priors are compressed into lossy language, and sparse interaction is often supervised through delayed textual feedback rather than dense visually grounded signals. We argue that reusable experience for VLM agents should remain visually grounded. Based on this insight, we propose AtlasVA, a teacher-free visual skill memory framework that organizes memory into three complementary layers: spatial heatmaps, visual exemplars, and symbolic text skills. AtlasVA further evolves danger and affinity atlases directly from trajectory statistics and lightweight grid heuristics, and reuses these self-evolving atlases as potential-based shaping rewards for reinforcement learning. This unifies perception, memory, and optimization without external LLM supervision. Experiments on \textsc{Sokoban}, \textsc{FrozenLake}, 3D embodied navigation, and 3D robotic manipulation benchmarks show that AtlasVA consistently outperforms text-centric memory baselines and competitive VLM agents, with especially strong gains on spatially intensive tasks. Homepage: https://wangpan-ustc.github.io/AtlasvaWeb

We present a real-time gaze tracking system that directly acquires task-relevant latent features using a fully passive optical encoder. Instead of forming and processing full-resolution images, our approach leverages a microlens array with a co-designed binary chromium mask to perform spatially multiplexed optical encoding, producing a compact set of measurements sufficient for gaze estimation. By integrating sensing and feature extraction in the optical domain, the proposed system eliminates the need for high-bandwidth image readout and substantially reduces computational overhead. The encoded measurements are captured by a 4 x 4 phototransistor array and mapped to gaze direction using a lightweight neural network. Our proof-of-concept prototype enables an end-to-end sensing-to-inference latency of 3.4 ms, outperforming published research systems. We demonstrate the effectiveness of our approach on both simulated and real-world data, achieving competitive gaze estimation accuracy while significantly improving latency and energy efficiency compared to conventional camera-based pipelines. This work highlights the potential of task-driven optical sensing for ultra-low-latency, computationally efficient human-computer interaction systems.

As autonomous agentic systems scale across regulated critical infrastructures, the lack of mechanistic, hardware-rooted enforcement for high-frequency policy updates presents a fundamental safety gap. We introduce Ethical Hyper-Velocity (EHV), a novel architectural framework for the formal verification of AI governance policies at runtime. Unlike retrospective auditing frameworks (ISO/IEC 42001, NIST AI RMF) which introduce 14-30 day latencies, EHV relocates the Policy Enforcement Point (PEP) into the inference pipeline via a Governance-Aware Just-In-Time (JIT) Compiler. By integrating Conflict-free Replicated Data Types (CRDTs) for policy synchronization and Epoch-based Attestation Caching within Trusted Execution Environments (TEEs), EHV achieves Sub-millisecond Formal Determinism (SMFD). We demonstrate via TLA+ formal verification that non-compliant agentic actions are computationally unreachable within the system's bounded operating state space. We prove that O(1) runtime enforcement can eliminate the traditional trade-off between deployment velocity and governance integrity, reducing Governance Latency from O(days) to O(1).