生成模型与LLM推理优化

At each decoding step a language model attends over the key-value (KV) cache of every earlier token, so at long context the attention call is bounded by memory bandwidth. Sparse attention reads only a subset of keys chosen by a cheap score estimate, and most methods give the unread tokens zero weight. The output then draws on only a small fraction of the KV cache, and accuracy drops at small budgets, most on tasks that aggregate information across the context. An inverted-file product-quantization (IVF-PQ) index over the cached keys computes an approximate score for every indexed token in order to rank them; after ranking, those scores approximate the attention logits of the tokens left out. PQ-HSA (hybrid sparse-approximate attention) attends the selected tokens with their original keys and values, and the unselected tokens, the background, enter the same softmax through those scores, summed per inverted list and multiplied by the list's mean value. At 128K and a 1-2% retrieval budget, PQ-HSA is more accurate than Quest and SnapKV on Llama-3.1-8B and Qwen3-30B-A3B and stays close to full attention in macro accuracy; with the same selector, the background term raises macro accuracy on the 8B model from 0.71 to 0.83. In the same 128K setting, inside vLLM on one NVIDIA H20, the decode attention call runs 1.6x faster than the FlashAttention-3 kernel; the speedup grows with context length, and a cost model fitted on 8B to 30B models gives the context length at which it begins. A vLLM plugin runs PQ-HSA on two engine versions without changes to the engine source; code is available at https://github.com/KunmingSHAO/pqhsa_release.

Hybrid LLMs interleave full-attention layers with linear-attention layers to reduce long-context inference cost, but this structure complicates prefix caching. Full-attention KV caches are token-addressable, whereas linear-attention layers maintain recurrent states that cannot be rolled back to arbitrary prefix boundaries. Existing systems materialize recurrent-state checkpoints, restricting prefix reuse to checkpoint-aligned positions. We present SuffixReplay, the first prefix caching system that lets hybrid LLMs reuse cached prefixes at every cache-supported page boundary without materializing recurrent-state checkpoints. Our key insight is to just let linear states forget the distant past. Modern linear-attention mechanisms use recurrent decay and gating to attenuate the influence of old inputs. Therefore, instead of checkpointing every prefix boundary, SuffixReplay approximates the state at a matched boundary by replaying only a recent suffix of the layer's input hidden states, which we retain as anchors. At the algorithmic level, SuffixReplay combines layer-wise and token-wise anchor sparsity with a bounded replay budget to control storage, computation, and quality. At the system level, it uses an independently managed anchor sidecar and a pipelined replay path to overlap anchor movement and state reconstruction with the native serving pipeline. We evaluate SuffixReplay on three hybrid LLMs: OLMo-Hybrid-7B, Qwen3.5-4B, and Qwen3.6-27B-FP8. Across these models, SuffixReplay retains 91.4-100% of full-prefill quality on average across LongBench and RULER, while using only 0.36-0.51x the amortized per-token storage of SGLang's default 8192-token checkpoint cache. Integrated into SGLang, SuffixReplay reduces median TTFT by 15-70% on branching workloads, sustains 2.3-4.3x SGLang's throughput when the working set exceeds HBM, and matches SGLang on high-hit continuation traffic.

Speculative decoding reduces sequential Target model calls by verifying multiple tokens from the Draft model in parallel. Yet KV Cache growth limits long-context serving under constrained GPU memory. Offloading KV to CPU memory relieves this pressure. However, existing offloading schemes restore the full KV history before attention and fail to fully exploit the benefits of KV sharing across queries within a verification round. Existing parallel speculative decoding methods also overlook idle GPU compute capacity while memory-bandwidth-bound Target verification waits for historical KV. We present SpecStream, a speculative decoding system that begins verification without waiting for the full KV history to be restored and exploits compute bubbles during KV transfers for concurrent drafting on the same GPUs. It offloads only Target-committed history, keeping candidate rollback local to the GPUs. Each streamed KV chunk serves all queries in the round, with online softmax preserving full attention. Target-priority scheduling controls Draft concurrent execution under resource limits to constrain interference with Target verification. Experiments show that SpecStream maintains task quality close to SGLang speculative decoding while supporting more concurrent requests under limited GPU memory. Across different datasets, it achieves average throughput speedups of 1.41$\times$ and 1.32$\times$ over the offloading baseline for Qwen3 and InternLM2.5, respectively. Compared with parallel speculative decoding on separate Target and Draft GPUs, SpecStream improves output throughput per GPU by an average of 55.4%.

We circumvent the Sanford-Hsu-Telgarsky (SHT) single-layer induction obstruction within a linear, one-step, causal, bilinear, symplectically consistent design class on the post-RoPE substrate, by lifting attention onto a symplectic phase space, mirroring Hairer's lift of Stormer-Verlet. The lift exits the premise of the SHT counting argument rather than the bound itself. The reframing exhibits the obstruction as a filter-order gap: a one-layer bilinear score realises a $z$-transform of joint order $(0,0)$, whereas the induction discriminator requires key-side order $\geq 1$. Applying the symplectic upper shear $M_γ:(q,p)\mapsto(q+γp,p)$ to the post-RoPE query and key streams closes it. We prove this lift is unique within the factorised subclass, exactly symplectic at operator level, and requires post-RoPE placement; and in an explicit $T_4$-only Gaussian reduction we derive a closed-form two-branch induction phase transition, held out at $r=0.9876$ with zero fitted parameters. That law is an analytically solvable limit, not a robust prediction: restoring the $T_3$ channel moves $γ_c$ at $d_k=64$ from 1.030 to 0.569 and removes the crossover. Deployability follows by exact derivation: the KV cache is unchanged, prefix reuse and speculative decoding are preserved, overhead is $6d$ FLOPs per token per layer, INT8 headroom grows by at most $\log_2(1+2γ)$ bits, fused kernels are unmodified, and no parameters are added. At 91.3M parameters a supercritical sweep locates an emergence band: induction forms 3/3 seeds at $γ=0.80$ in a mean of 717 steps, against 2/3 seeds and 2700 steps at $γ=0$. Adverse results are reported as directly: a key-only half-lift reaches 0.949 against 0.811 for the symmetric operator, so if induction accuracy is the objective, the half-lift is the better construction. Forty-one notebooks and result files ship as ancillary material.

Unified multimodal models combine understanding, generation, and editing within a single network, offering a promising foundation for versatile multimodal applications. However, growing multimodal contexts make KV cache storage and access increasingly costly. Existing KV cache compression methods are typically tailored to specific tasks and single-modality caches, while overlooking changes in cache importance across tasks and timesteps. However, in unified multimodal models, each task involves multiple KV cache types, and both their composition and dynamics differ across tasks. As a result, a single compression policy overlooks task- and type-specific requirements, leading to the loss of critical information and degraded quality across tasks. Based on these findings, we propose UniCache, a training-free framework for task- and type-aware KV cache compression. UniCache identifies the cache segments activated by each task and assigns suitable compression policies through offline calibration. It coordinates their parallel execution under a shared storage budget through attention-guided allocation and task-aware temporal scheduling. Experiments show that UniCache achieves $5\times$ KV cache compression for understanding and editing and $2.5\times$ for generation with negligible quality loss, while increasing throughput by up to $1.78\times$ in long-context settings, significantly improving the practicality of scaling unified multimodal models to longer context.

Vision-language models are increasingly used to understand long videos and continuous streams. However, dense visual tokens accumulate with video duration, making long-context inference prohibitively expensive. Existing training-free visual-token selection methods reduce this cost by retaining informative tokens, but may lose coherent event evidence and fail to distinguish detailed recent observations from long-range history. We propose KeyRec, a training-free framework for constructing bounded visual memory. During query-agnostic writing, KeyRec preserves fine-grained recent observations in a visual cache and organizes historical evidence into a structured event bank. Candidate events are proposed according to their novelty relative to previously stored events and maintained through an online add--merge--evict update. When a question arrives, a text-only router adaptively allocates a fixed readout budget between recent and event memory, without reprocessing historical frames. KeyRec operates on model-facing visual embeddings and supports both modular encoder--projector VLMs and the encoder- and projector-free NEO-ov architecture. Across four streaming and long-video benchmarks and three VLM backbones, KeyRec achieves the best compressed performance in 13 of 15 settings using only 10% of the dense decoder-facing visual-token budget. It outperforms the strongest compressed baseline by 2.21--18.37 points on real-time questions, achieves the best compressed result in five of six long-video settings, and performs best in every NEO-ov 2B setting.

Vision Language Action (VLA) models condition actions directly on current visual and language context, without an explicit account of how the scene evolves under candidate actions. World Action Models (WAM) attempt to address this limitation by predicting future states, but existing designs keep prediction and policy learning architecturally separate, connecting them only through the predicted output, whether through pixel space video generation or a latent forecasting module trained independently of the policy. We present Devol-ONE, a Mixture of Transformers architecture that unifies vision language understanding, latent world dynamics prediction, and action generation within a single autoregressive framework. Instead of encoding vision language tokens once and feeding them to the action expert, Devol-ONE runs autoregressive prediction jointly across a vision language stream and a V-JEPA pretrained dynamics stream, attending to the vision language key-value cache at every layer to forecast future latent states under language guidance. The action expert is in turn shaped continuously by semantic reasoning and predicted physical dynamics rather than by a fixed representation computed in advance. Extensive experiments are conducted on LIBERO, LIBERO-PLUS, RoboTwin2.0 along with real-world evaluation on Flexiv single-arm and dual-arm setups. Ablation studies show the effectiveness of dynamic stream prediction and layer-wise unified attention to validate our model architectural coherency.

Large language model (LLM) agents serving different users often solve related tasks, yet separate user histories can leave reusable experience inaccessible to other agents. Pooling memories expands access but risks transferring preferences that conflict with the receiving user's requirements. We introduce ShareMem, a memory architecture that shares reusable experience while grounding its application in the receiving user's own preferences. Shared experiences indicate how to act and which preferences to consult; the receiving user's memory supplies their concrete values. Two-stage consolidation refines experience locally before integrating accepted edits into a shared pool. During execution, scope-first retrieval jointly selects local and shared experiences under a common entry budget, while a user-bound channel supports initial and agent-initiated preference retrieval. We evaluate ShareMem across web navigation (Mind2Web), online personalized interaction (VitaBench~2.0), and multi-session coding (MemoryCode) with four backbone models. It improves step success, average task success, and dialogue-macro coding scores, respectively, over matched user-local memory across all four models. Ablations favor two-stage consolidation for smaller shared pools, lower induction token usage, and better downstream performance, and support complementarity between experience guidance and active preference retrieval. Further analyses show that sharing helps most when relevant local experience is scarce, while source quality and cross-user preference interference limit useful transfer.

Machine unlearning and model pruning are increasingly coupled in the real world. Models must support unlearning requests, e.g., for safety concerns, while also meeting requirements in latency and memory budget. Until recently, existing works have studied each aspect as an independent problem, e.g., running unlearning and pruning sequentially. In this work, we show that unlearning and pruning are naturally aligned and should be solved jointly to be made aware of each other. Intuitively, parameters that encode information of the unlearned samples are natural pruning targets, as unlearning and pruning both call for the "deletion" of such parameters. We propose PruneForget, a method that uses the unlearn set as a guide for pruning, so that unlearning and pruning mutually benefit each other. Extensive experiments on image classifiers and generative models show that PruneForget removes the influence of the unlearned samples while producing a more compact model with reduced inference cost. It achieves a negligible performance gap relative to an oracle that retrains from scratch for unlearning and then prunes.

Running GUI agents on edge devices can keep sensitive screens and interaction histories local, but the computational cost of inference at every action step makes deployment challenging. Existing GUI agent systems either perform full vision-language model (VLM) inference at each action step or reuse coarse-grained knowledge matched to prior tasks. However, dynamic mobile environments and user tasks make it difficult to fully utilize prior task executions without additional fine-tuning or task-specific offline exploration. To address this challenge, we present PastForward, a system that accelerates GUI agents through validated, fine-grained reuse of computational experience accumulated during ordinary task execution. During decoding, PastForward retrieves prior output sequences as device-adaptive multi-token proposals and verifies them in a single VLM forward pass. Across action steps, it uses prior GUI transitions to begin next-step inference while the device executes the current action, retains the early computation only when the predicted screen matches the observed screen, and carries reusable KV states forward. We evaluate PastForward on AndroidWorld workloads derived from real mobile usage patterns using multiple VLM backbones across server and edge platforms. On device, PastForward achieves action-step latency speedups of 1.63-2.36$\times$ while maintaining task success rates.

External memory lets language-model agents answer questions about histories too long for the answer model's context window. Updates create a harder problem than retrieving a recent fact: changing one relation can redirect a multi-hop question to records about an entity absent from the question. We study this update-dependent evidence selection problem and introduce the Contract Memory Compiler (CMC). Before seeing a question, CMC uses a language model to identify relations in the history and record where each one was stated. It applies later updates to determine the current relations, follows them from entities named in the question, and passes the corresponding original records to the answer model in one call. Thus the current state determines which evidence is read, rather than merely refreshing values in a previously selected context. To the best of our knowledge, CMC achieves state-of-the-art multi-hop accuracy on FactConsolidation, reaching 78.25% overall and 61.0% at 262K. With the extracted relations and answer model held fixed, selecting evidence before resolving updates reduces multi-hop accuracy to 21.50%. We also introduce MQuAKE-MemStream, a derived dataset of ordered memory streams built from MQuAKE-Remastered counterfactual cases.

While Large Language Model (LLM)-based agents demonstrate strong capabilities in long-horizon tasks by interleaving reasoning with external environment interactions, the continuous accumulation of context rapidly creates a critical memory bottleneck. Existing memory compression methods rely on task-specific optimization or external auxiliary models, introducing significant computational overhead. Furthermore, the resulting compressed representations tend to lose structured relationships, leading to information dilution, attention collapse, and degraded decision consistency. To address these limitations, we propose Auto-Regressive State Machine (ARSM), a lightweight training-free framework that enables in-situ reasoning compression through structured state evolution. ARSM introduces two key components: (i) a trajectory abstraction mechanism that reorganizes interaction histories into compact Hypothesis-Action-Result (HAR) micro-chains; (ii) a dynamic state machine that regulates hierarchical memory through atomic operations and a compression-control parameter. These components are unified within an auto-regressive, self-compressive generation space, where each model output jointly performs external action execution and internal state updates. We evaluate ARSM on Webshop, Multi-Objective Multi-Hop QA, and SWE-Bench Lite datasets. Experimental results show that ARSM maintains the task performance while simultaneously reducing token consumption, offering a practical, cost-effective route toward scalable autonomous agents for long-horizon tasks.

Long-horizon LLM agents must convert accumulated experience into durable memory, deciding what to keep, compress, abstract into reusable skills and rules, or forget. We report a four-phase research program on this consolidation problem whose central finding is a shift in what is measured: from how much an agent remembers, to whether its consolidation decisions are any good, to whether those decisions can be trusted. Phase 1 learns episodic boundaries from agent traces by downstream utility; an honest near-miss (oracle correlation 0.691 vs a 0.70 bar) whose lasting output is a three-gate anti-leakage protocol. Phase 2 learns when to promote experience and to which abstraction level under a token budget, achieving a verified +22.7% task-success improvement with 7x compression, but exposing a degenerate-forgetting failure and a distribution-shift failure mode we name lambda-prevalence coupling. Phase 3 introduces ConsolidationBench, an oracle-by-construction benchmark that scores consolidation decisions against a known optimum on three non-circular axes; production retrieval systems retain information yet score zero on cross-level transfer. Phase 4 introduces governed consolidation: the decision wrapped in poison-resistance, reversibility, and auditability guarantees with a quality gate. Governance is statistically distinct from the quality score ($r^2 = 0.43$; partial $r = 0.27$; identical-quality policies differ threefold in governance), so the contribution survives independently of the metric's external validity. On that question we report a resolved negative: after a graded-reuse redesign removed a structural ceiling, a two-benchmark study with 2,532 real answer cells finds the quality score does not predict real transfer accuracy (pooled Spearman $ρ= -0.24$, n = 12, CI spanning zero). An adversarial self-critique pass cleared the final claim set with zero surviving overclaims.

Agent memory grows as agents read inputs, reason, and call tools. Longer histories increase inference cost and eventually exceed the context window. LLM-based summarization reduces this history but adds latency and provides no explicit bound on information loss. We propose LAM, a Lossy Agent Memory system with three components: a deterministic deduplication rule with a substitution bound on retrieval scores - a bound on score perturbation, not a certificate of unchanged ranking; a memory manager that preserves the cached prefix and overlaps compaction with inference; and a performance model that estimates compaction costs before deployment. On 600 agent trajectories, LAM removes 22.47% of observation tokens while retaining 99.984% of the measured gold-patch evidence. At a fixed deletion set, the performance model predicts a 71.4x-91.6x end-to-end speedup from removing records before prefill instead of deleting them from a prefilled context. That benefit comes from the schedule rather than the rule and applies to any prefix-preserving test.

Quantization-Aware Pre-Training (QAPT) can increase the inference efficiency of DNNs, but a problematic behaviour known as rounding boundary weight oscillation can introduce detrimental noise into the training process and significantly reduce convergence speed. While existing methods can reduce this detrimental noise, they either introduce additional hyperparameters or memory overhead, or cannot consistently improve model accuracy. In this work, we propose optimization with $C$onstrained $E$mpirical $W$eight dis$T$ribution (CEWT), the first hyperparameter-free memory-overhead-free oscillation suppression method that consistently improves QAPT performance: an optimizer post-update step that projects weights to the nearest point in weight space whose empirical distribution (histogram) matches a zero-mean Gaussian. Our key insight is many quantizers are designed with the implicit assumption that the to-be-quantized data are permutations of samples from a zero-mean Gaussian, and this assumption is not true during QAPT. By enforcing the zero-mean Gaussian prior as a hard constraint, CEWT can suppress this detrimental noise. Empirical results on various combinations of SOTA quantizers and hypersphere optimizers suggest, that with a geomean increase of 4% in training time, CEWT can consistently reduce the pre-training perplexity (by an average of 2.5 and up to 21 points) of low-precision (down to 1-bit activations and weights and up to 610M parameters) LLaMA/GPT models without introducing any hyperparameters or storage overhead. Code is available at https://github.com/1733116199/cewt

Diffusion Transformers have become the dominant architecture for video generation. Their substantial computational cost motivates scaling inference across multi-GPU servers, yet efficient scaling remains challenging on commodity GPUs connected via PCIe, whose bandwidth is limited. Although sparse attention substantially reduces computation, its implications for communication remain underexplored. This paper argues that attention sparsity should be treated as a communication primitive. We present SparSP, an efficient sparse sequence parallel communication system that co-designs token distribution, communication routing, and asynchronous execution for sparse video diffusion models. First, Dependency-Aware Placement distributes sequence blocks according to diffusion models' sparse attention patterns. Second, Demand-Directed KV Routing transfers KV blocks directly to requesting GPUs without intermediate relays. Third, a Decoupled Transfer Runtime separates communication from GPU computation to reduce resource contention and maximize effective bandwidth. Our evaluation shows that SparSP improves attention performance by 1.38- 1.5$\times$, achieves an average 1.17$\times$ (up to 1.69$\times$) end-to-end speedup across three representative servers and three video diffusion models, and reduces communication volume by 12.54-23.05%. Moreover, we achieve an average 1.53-1.76$\times$ bandwidth improvement over NCCL.

Deploying a language model under different computing and latency budgets calls for compact models with different quality-cost trade-offs. Towards this end, elastic spectral state space models provide ordered and temporally decomposed channels that can be truncated, but they are associated with linear time-invariant filters that cannot selectively preserve relevant past information or forget irrelevant information as the context evolves. To address this, we introduce the Elastic Selective Spectral Hybrid (ESSH), which realizes each Hankel spectral channel as an independent recurrent unit using fitted damped rotation modes. It also features an input-dependent decay and write/read gates that make temporal retention and state update input-dependent while preserving channel-wise truncation and a structured recurrence for efficient execution. ESSH combines these selective spectral mixers with sliding-window attention and jointly trains multiple capacities by reducing spectral-channel count and feed-forward width at different rates through a two-rate capacity map with full-model distillation. The resulting models support chunked parallel training and fused recurrent decoding while avoiding computation for discarded channels. At full capacity, ESSH achieves language-modeling quality comparable to similarly sized independently trained models, while smaller exports exhibit a smooth quality-cost trade-off. We validate the effectiveness of the proposed framework using language understanding, retrieval, cross-domain text, and DNA experiments, by assessing quality retention and the trade-off against independently trained and elastic baselines. At the 1.53B model configuration, fused batch-one decoding takes 1.37 ms per token on a B300, providing a 2.14-2.80x speedup over the tested Mamba-2 and Mamba-3 implementations and 3.03x over Transformer++ at matched parameter counts.

Secure Transformer inference protects sensitive inputs but incurs substantial cryptographic overhead, with nonlinear operations such as Softmax and GeLU becoming major bottlenecks. Existing compression methods reduce nonlinear complexity, sequence-dependent computation, or model structure through separately defined compression variables. Under aggressive compression, however, these independently optimized perturbations can accumulate: at a matched compression level, stacking representative approximation, token-pruning, and model-pruning methods reduces ViT-S accuracy from 80.20% to 76.41%. We introduce DegreeSpar, which formulates secure Transformer compression as structured sparsification over nonlinear polynomial degrees. Polynomial degree directly controls the cost of secure nonlinear evaluation, while computation-aligned zero-degree structures expose token-level and model-dimension computation as removable within the same optimization space. DegreeSpar further incorporates approximation-aware training for low-degree Softmax and GeLU, enabling aggressive degree reduction and creating the optimization headroom required for structured computation removal. Across vision and language Transformers, DegreeSpar consistently improves the accuracy-latency trade-off across model scales, tasks, and sequence lengths, achieving speedups from 2.29x to 6.63x over the corresponding baselines. Under the same network setting, DegreeSpar achieves 92.68% accuracy on BERT/SST-2 in 110.55 s, compared with 92.66% in 167.26 s for CipherPrune, the closest prior hybrid secure-inference approach. These results establish structured polynomial degree as an effective shared optimization space for secure Transformer compression.

The fast Fourier transform (FFT) underlies radar, imaging and scientific computing. A classic rule for fast GPU FFTs is to compute the largest block that fits on chip and compose larger transforms from such blocks. We test this rule on Apple's M6 chip, whose GPU performs 29 arithmetic operations in the time it reads one byte from main memory, and which adds matrix units to both its GPU and CPU. Comparing our kernels with Apple's MPSGraph and vDSP libraries and MLX, we find that data movement, not arithmetic, sets FFT speed. Large batches run at or near the main-memory bandwidth limit in every GPU library, and benchmarks that keep data in cache overstate real throughput by up to $3.7\times$. The on-chip rule still predicts where speed collapses: MPSGraph and MLX lose half or more of their speed once a transform outgrows the GPU's 32\,KiB local memory. Keeping such transforms in registers avoids an extra trip through memory: $2.2\times$ faster than MPSGraph, and $4.4\times$ with half-precision storage. The GPU's matrix units do not help, because recasting the FFT as matrix products adds as much arithmetic as they save. The CPU's matrix unit, 47 times faster than its vector units, does: our kernel for it beats vDSP by up to $5.3\times$. End to end, a $4096\times4096$ synthetic aperture radar image takes 8.1\,ms on the GPU (6.0\,ms with half-precision intermediates), $14$--$19\times$ faster than a CPU reference. Running the CPU's matrix unit alongside the GPU gains nothing: both share one memory bandwidth.

Training language models with ternary weights is commonly judged by loss and downstream accuracy, which record only a modest cost relative to full precision. We show that these metrics can conceal a much larger failure. We instead measure knowledge capacity, the factual bits stored per parameter, on synthetic biographies with known information content. We train GPT-2-style models from scratch with 2.5M to 50M parameters at five precisions. Under the standard cosine schedule, ternary models retain as little as 6% of an identically trained fp16 model's capacity. The deficit widens with model size while perplexity rises by only 1.4 to 1.6 times. Measured throughout training, these models first acquire capacity and then lose most of it. We identify this knowledge collapse as a learning-rate dwell instability. Held near $1$--$2\times10^{-4}$ with no decay, a pre-collapse model collapses within a few hundred exposures, and returning to a safe rate does not restore capacity. We then locate the collapse in the output head. It happens at a value the model can never predict. The weights there grow unchecked, while every other prediction the model makes is unchanged. We find that making the value predictable removes the collapse, regardless of which attribute carries it. A warmup-stable-decay schedule with a 10% cooldown increases ternary capacity by 2.7 times at 25M and 4.3 times at 50M. Gains are larger at lower precision. Cutting the output head's learning rate prevents failure when training from scratch. The post-training quantization methods we tested recover no measurable capacity below 4 bits. Our findings suggest judging low-precision training by retained capacity during training rather than final loss.

Visual token reduction is an effective way to accelerate multimodal large language models (MLLMs), but performance deteriorates rapidly under extremely low token budgets. Existing work has explored both visual-token selection and training-based adaptation to reduced visual inputs. We take a step further by asking how a heavily compressed MLLM should learn from the states induced by its own generations. This setting naturally calls for on-policy self-distillation: a heavily compressed model is supervised on the states induced by its own generations, while its full-token counterpart serves as an information-rich teacher. Based on this insight, we propose LT-OPD, a training framework for extreme visual-token reduction. The student rolls out responses with only a small fraction of visual tokens, and a frozen full-token copy of the same MLLM provides distributional supervision along these student-generated trajectories. To stabilize on-policy learning when visual evidence is severely limited, we further introduce a budget-level curriculum that progressively decreases the token budget during training. Across nine benchmarks on Qwen3.5-4B, LT-OPD raises average retained performance under 5% visual-token retention from 68.6% to 82.3%, outperforming training-free, training-based, and reinforcement-learning baselines at the same budget. The gains transfer consistently to Qwen3.5-9B, GLM-4.6V-9B, and LLaVA-OV-1.5-4B. LT-OPD also reduces KV-cache usage by 85.2% and prefill FLOPs by 85.4% without additional inference overhead, demonstrating that on-policy learning can substantially recover capabilities lost to extreme visual-token reduction.

Matrix multiplications dominate the inference cost of modern transformer-based vision models, yet existing efficiency techniques such as post-training quantization and mixed-precision inference are largely limited to the small set of fixed-width formats (INT4, INT8, BF16, and FP16) supported by conventional accelerators. We revisit stochastic computing (SC) as a way to lift this constraint: viewed as a dense adaptive quantizer, SC controls precision by bit-stream length L rather than a fixed datapath, while each multiplication reduces to a single AND/XNOR gate. We build a GPU library that emulates SC matrix multiplication at scale, exposes stream lengths as first-class kernel arguments, and evaluates SC end-to-end on image classification, object detection and instance segmentation, class-conditional image generation, and visual world-model planning. On top of this substrate, we develop a dynamic per-row mixed-precision policy that assigns stream length per token or group at matched average budget, requires no retraining, and uses the same SC hardware across schedules. Across tasks, SC remains competitive with fixed-format INT quantization at matched bit budgets, while per-row mixed precision helps maintain accuracy at lower average stream lengths. These results provide software-level feasibility evidence that SC can serve as a dense-precision substrate for fine-grained mixed-precision inference on modern vision transformers.

Hybrid attention models have emerged as a crucial architecture for Large Language Models (LLMs) (e.g., the Qwen3.5 and Kimi series). Their memory and computational efficiency make them highly attractive for on-device inference, forming a promising synergy with edge Neural Processing Units (NPUs). However, naive execution of these hybrid models on edge NPUs fails to deliver these benefits, often bottlenecking the prefill stage due to severe memory-system inefficiencies and architectural mismatches within the linear attention (LA) layers. We present HA-NPU, the first system to enable efficient hybrid attention LLM inference on edge NPUs without modifying the underlying algorithms. HA-NPU enhances execution efficiency by reorganizing the dataflow of the LA components across three levels: (1) At the core level, it partitions workloads by the head dimension and fuses dependent operators, eliminating cross-core global memory accesses; (2) At the operator level, it reorders execution to consume intermediate tensors immediately, drastically minimizing local-buffer pressure; (3) At the tensor level, it employs dataflow-aware layout planning to minimize transformation overhead between matrix and vector processing units. Compared to competitive baselines, HA-NPU achieves up to 35.95$\times$ LA kernel speedup and 36.14$\times$ energy reduction, delivering up to 2.03$\times$ faster end-to-end request latency. The source code will be made publicly available at https://github.com/yinyuanzhang/HA-NPU

We present VSA2, a frontier trainable sparse attention for video DiTs. VSA2 includes a variety of new architectural features and training procedures that we apply across all stages of the DiT development cycle, including pretraining, RL, and inference, to produce a DiT with comparable or better quality than a full attention counterpart. Architecturally, VSA2 introduces a fine-grained router that improves the precision of identifying critical tokens and supports dynamic computation by allowing each query to attend to a variable number of key-value pairs. In training, we identify a Hard-to-Easy Curriculum, where models trained under high sparsity and later evaluated with lower sparsity during inference not only generalize effectively, but also outperform models trained with full attention in motion quality. VSA2 is also flexible: it can replace full attention during the middle of progressive low-to-high resolution pretraining, rebasing early-stage full-attention checkpoints. Experiments show that VSA2 reduces attention computation by half over VSA with lower loss. On 720p videos, it accelerates attention by 8.9x and end-to-end generation by 4.62x compared to the FlashAttention-3 baseline, while achieving comparable or better video quality.

Video generation has broad applications in content creation and entertainment. Diffusion transformers have advanced the quality of generated videos, but attention over spatiotemporal tokens becomes increasingly expensive as video resolution and duration increase. We present DraftAttention2, a training-free framework that uses the low-resolution draft attention map to jointly select attention blocks and assign their numerical precision. Specifically, spatial 2D average- and max-pooled queries and keys capture complementary regional statistics to estimate block importance, and a shared ranking assigns higher precision to important blocks, lower precision to less important retained blocks, and skips the rest under configurable budgets. Our analysis separates sparsification error from attention-weighted quantization error, establishing when recovering skipped interactions with low-bit computation tightens the output-error bound. This analysis motivates retaining more interactions at low precision while reserving higher precision for blocks with larger attention mass. To translate these fine-grained assignments into practical speedups, we further develop fused operand preparation and a single attention kernel with precision-specific phases, sharing data movement, softmax statistics, and output accumulation across precisions. Experiments demonstrate that our method achieves a superior quality-efficiency trade-off over existing efficient video generation methods. Notably, its advantage is particularly pronounced for few-step video diffusion, where jointly combining sparsity with 4- and 8-bit mixed-precision computation substantially improves generation quality while retaining significant acceleration. Code is available at https://github.com/anemoi-project/anemoi

LLM-based agents execute multi-turn workflows with interleaved model inference and tool calls, making efficient serving increasingly important. However, evaluating serving optimizations is challenging because identical tasks can produce different execution trajectories. Changes in generated tokens can alter subsequent prompts, tool calls, and reasoning turns, making it difficult to distinguish system improvements from workload variation. Greedy decoding does not guarantee identical outputs, while replaying only sequence lengths loses token information that affects prefix caching and mixture-of-experts (MoE) routing. To address these problems, we propose AgentReplay, a configurable trace record-and-replay framework for agent serving. AgentReplay records input/output tokens, expert selections, request dependencies, and tool durations. During replay, it forces the recorded output tokens while performing normal autoregressive computation, with optional controls for MoE expert selection and tool delays. This allows different systems to execute the same recorded workload while retaining their own batching, scheduling, and parallelization decisions. We further separate trajectory generation from performance evaluation, enabling compatible smaller models to replay long-horizon traces collected with more capable models. Our experiments show that token-wise replay in AgentReplay effectively eliminates workload variation that greedy decoding and length-wise replay cannot avoid, enabling fairer performance comparisons across serving configurations.

Few-step autoregressive video diffusion generates a long video by splitting the video into temporal chunks and generating chunk-by-chunk, each through a short sequence of denoising stages. To memorize chunks that are already generated, previous methods reconstruct a clean or less-noisy key--value (KV) cache by additional forwards to build the cache without advancing an output latent. However, every denoising forward itself already computes the in-flight KV of the current chunk. We introduce FlashForward, which directly reuses this cache to avoid the heavy cache-update-only model forwards. After the current chunk completes one denoising stage, its stage-specific cache is already available for the next chunk. Assigning one GPU to each stage therefore lets different chunks occupy different stages concurrently. This early availability has a quality cost: the resulting stage-matched history is noisy, causing appearance and motion drift among chunks. To complement it, FlashForward produces sparse auxiliary clean anchor latents before the corresponding region is generated so the generation trajectories can be stabilized by this two-sided conditioning. The two memories operate at different temporal scales: sparse clean anchor KV supplies coarse, long-range two-sided structural guidance, while dense stage-matched history preserves fine, recent evolution. With up to four GPUs, FlashForward runs $1.16$--$1.69\times$ faster than HiAR and $1.42$--$2.92\times$ faster than Self-Forcing for 16 FPS videos of 20 seconds or longer across 1.3B and 14B backbone scales at 480p and 720p. On VBench, for the 1.3B model at 480p, it achieves higher scores and remains stable at longer durations, demonstrating that FlashForward generates high-quality and temporally consistent videos across durations of 20s, 35s and 65s at a much faster generation speed.

Distilling bidirectional multi-step video diffusion transformers into few-step causal models has become a common approach for streaming video generation. While these few-step students are significantly faster than the teachers they are distilled from, they remain slow for real-time generation. In this work we present UnStep, a training-free wrapper that accelerates few-step causal video models at inference by running them with fewer diffusion transformer (DiT) steps than during distillation and limiting the temporal window retained in the attention KV cache. We propose two inference-only mechanisms to recover quality lost by step reduction and attention windowing: renoising the generated latent frames to a near-clean level and reusing the existing clean-cache pass to refine them, and applying truncated SVD to the DiT attention value and output projections. We also accelerate inference with a quality-preserving runtime stack for the DiT and VAE decoder, including more efficient attention calls and KV indexing, fused Triton RoPE with cached coefficients, and VAE decoding with optimized memory layout, precision, and convolution kernels. By reducing computation and optimizing the runtime stack, UnStep sets a new throughput regime for causal video diffusion, by running substantially faster than current methods, reaching 50 FPS on a single H100 without quality loss, and 77 FPS on GB200, all without retraining.

Recent multi-agent LLM systems increasingly combine heterogeneous models for specialized agent roles. However, text-based communication requires each receiver to prefill shared context already processed by the sender. Reusing the sender's key-value (KV) cache avoids this redundancy, but prefill-free transfer across model families must handle differences in tokenization, model depth, and KV representations. To address these issues, we propose HeteroFold, a prefill-free cross-family KV cache transfer method that keeps both the sender and receiver frozen. HeteroFold aligns model structures, maps the sender cache into the receiver space, and calibrates it to preserve receiver behavior. Across six transfer directions, HeteroFold achieves the best cache-transfer performance on all four long-context benchmarks and most short-context settings. It also matches text-based communication on the multi-agent benchmark. At 32K context length, Llama-3.1-8B$\rightarrow$Ministral-3-14B transfer is $10.7\times$ faster than Native Prefill and $1.18$--$1.47\times$ faster than the state-of-the-art prefill-free baselines, Dense Latent and KV Ridge. These results show that HeteroFold enables efficient cross-family KV reuse without receiver prefill.

Video diffusion transformers produce high-quality videos, yet iterative denoising incurs substantial inference latency, limiting interactive and large-scale serving. Most existing acceleration methods focus on individual requests, thereby restricting efficiency gains to redundancy within a single generation trajectory. Recent cross-request reuse offers an additional source of savings, but existing approaches often infer reusability from coarse semantic similarity. This conflates semantic relatedness with generation-level computational compatibility, so aggressive reuse may accept incompatible historical computation while conservative reuse leaves substantial acceleration unrealized. We present Carnator, a cross-request acceleration framework that addresses this challenge by extracting and using generation-native compatibility evidence directly from the model's evolving internal states. Specifically, Carnator performs a lightweight early probe to construct an Early Signature from internal diffusion states, assessing reuse validity through risk-aware compatibility decisions. The same evidence characterizes reuse scope by localizing target-specific computation and guiding joint reuse of historical latent trajectories and sparse attention connectivity. Across three text-to-video backbones, Carnator consistently achieves higher cache-hit end-to-end acceleration than the evaluated cross-request baselines despite more selective cache acceptance, reaching up to 2.17$\times$ speedup while maintaining competitive generation quality.