生成模型与LLM推理优化

We present Soofi S 30B-A3B, a sovereign, open-source Mixture-of-Experts (MoE) hybrid Mamba Transformer foundation model for German and English. Its hybrid design activates only 3B of 30B parameters per token and keeps the inference cache near-constant as context grows, giving it a decisive throughput advantage over dense models for long-context, high-concurrency deployment. Pretrained on roughly 27 trillion tokens with deliberately up-weighted German, Soofi S matches dense 14 to 27B models on aggregate English and German benchmarks while achieving the best code aggregates in both languages among 17 open base models, and outperforms every European sovereign baseline in our comparison, including ones far larger in active parameters. Among fully open models, Soofi S obtains the highest English and German evaluation scores, ahead of Olmo 3 32B and Apertus 70B. Soofi S was built end-to-end on the German Industrial AI Cloud, a sovereign HPC scale AI infrastructure operated by Deutsche Telekom in Munich. Soofi S will be released under highly permissive, open-access terms: weights, selected intermediate checkpoints, full per-source data accounting, hyperparameters, and training and evaluation code. Where source licenses permit, data-construction artifacts are released under permissive licenses; commercially licensed sources are documented with aggregate statistics and exact mixture accounting.

Although Video Large Language Models (Video LLMs) have shown strong performance in video understanding, their efficiency is still limited by the large number of visual tokens. Existing video token compression methods typically rely on frame-wise saliency or heuristic token merging, which can over-focus on locally salient regions and produce ambiguous fused features. To address these issues, we propose GeoTrace, a training-free spatiotemporal token compression framework that decomposes video evidence into exact skeleton tokens and traceable residual event tokens. Specifically, Contextual Farthest-Point Anchoring (CFPA) preserves salient, context-consistent, and high-coverage skeleton tokens, while Trajectory-Constrained Residual Condensation (TCRC) compresses residual tokens through one-to-one temporal trajectories and constrained near-manifold condensation, producing traceable event tokens with reduced ambiguity. We evaluate GeoTrace on four Video LLMs across four video understanding benchmarks, and the results demonstrate its effectiveness and generalization across different model architectures and scenarios. On LLaVA-OneVision, with only 10% visual tokens retained, GeoTrace achieves a \(12.99\times\) TFLOPs reduction while preserving 99.1% of the vanilla performance. Overall, GeoTrace offers a compact and traceable token representation for efficient and robust Video LLM inference. Code is available at \href{https://github.com/guohuan-xie/GeoTrace.git}{Code}.

In large-scale computing systems, jobs often demand heterogeneous server allocations: large jobs that occupy a substantial fraction of the servers are of high importance and are thus latency-sensitive, while small jobs fill in the remaining capacity to maintain throughput. To model this dynamic, we introduce the half-batch multiserver-job (MSJ) framework, a queueing model in which large jobs arrive according to a Poisson process and require all servers simultaneously, while small jobs, each needing only one server, are always available. We prove that, in the half-batch MSJ model, the Pareto frontier for large-job mean response time and small-job throughput admits a simple and exact characterization. It is generated by a family of convoy policies, under which the system serves small jobs until $k$ large jobs have arrived and then switches to serving large jobs, together with convex combinations of neighboring convoy policies. Our result is fully general and non-asymptotic, holding for every stable arrival rate $λ$, every number of servers $n$, and every large-job size distribution $S$.

Additive activation steering (injecting a scaled residual-stream direction during generation) is calibrated almost entirely in single-turn chat, yet the models it targets are increasingly deployed as tool-using ReAct agents. We present the first systematic chat-to-agent transfer study of additive steering, coupling behavioral measurement with a representation read-out in a matched-information design: the same items rendered as plain chat or as a ReAct tool-use episode, with matched-norm random-direction controls and the transcript re-encoded every turn to exclude KV-cache contamination. Transfer is real but rescaled, and the right description is a dissociation: the injected direction reaches the late layers at near-full strength in every setting and model tested (install-site agent-over-chat ratios 0.83-1.16 across three families), while the behavioral coupling is reset per model and context. On Qwen2.5-7B a refusal bypass vector amplifies in the agent (T = 1.45, CI [1.20, 1.78], N = 300); across a powered uniform-protocol distribution the coupling spans amplification (Gemma-2-9B T = 2.00) to attenuation (Yi-1.5-9B T = 0.43, CI [0.29, 0.60]), with no universal constant and a single clean attenuator against a universal sign. Directional ablation of the same axis does not amplify (T = 0.93, CI including 1) while additive injection amplifies (T = 1.50), a 20.1-point gain difference (CI [13.4, 26.8]) that identifies an additive-specific mechanism. Two pre-registered instruments converge to localize the rescaling to the ReAct format scaffold, before any tool observation, rather than to the observation boundary where a dilution account would predict it. The safety implication is immediate and unpredictable: agentic deployment amplifies steering-based refusal bypass by up to 2.00x on some models while others attenuate, so a deployment cannot assume a given model is safe under additive steering.

Vision-Language Models (VLMs) have achieved success using homogeneous Transformers to process multimedia data. Recent studies show that heterogeneous structures interleaving efficient mechanisms, like linear attention, improve both performance and inference latency over homogeneous designs. However, these efforts rely on handcrafted static mixing patterns, which are sub-optimal and difficult to adapt to specific hardware. To bridge this gap, we propose Multi-Objective Search for Adaptive Inter-layer Composition (MOSAIC), a hardware-aware search method that automatically transforms homogeneous models into optimized heterogeneous architectures. MOSAIC integrates diverse efficiency mechanisms--including linear, sparse, and low-rank operators--into a unified search space. By formulating the selection as a multi-objective Mixed Integer Programming (MIP) problem, our method identifies optimal configurations that maximize downstream performance under strict hardware latency constraints. To mitigate performance degradation from structural transitions, we introduce a two-stage parameter recovery process: global off-policy distillation to stabilize internal representations, followed by a dual-teacher on-policy distillation leveraging a 235B oracle for knowledge expansion and the original 4B teacher for distributional stability. We validate MOSAIC through MOSAIC-4B, derived from Qwen3-VL-4B-Instruct. Results demonstrate that MOSAIC-4B matches the baseline's performance across multiple benchmarks while requiring less than 2% of the original training cost. Furthermore, it substantially improves inference efficiency, achieving 1.76x prefilling and 2.54x decoding speedups.

Block sparse attention is a hardware friendly way to alleviate the key-value (KV) cache read bottleneck in large language models (LLMs). However, it is not prevalent among leading open-weight LLMs, which rely instead on dense attention or fine-grained selection, thereby motivating our analysis. We study DeepSeek's Native Sparse Attention (NSA) as a representative method, whose three-branch design lets us isolate block selection, the most challenging and consequential stage. We formalize selection and reduce it to ranking blocks by a single quantity, the attention mass: the sum of a block's attention scores. We show that if selection retrieves the blocks with the largest attention mass, block sparse attention can match the quality of dense attention. However, computing the exact attention mass requires reading every key, so the problem of block selection ultimately reduces to approximating this mass from a compact summary instead of the full keys. Via a cumulant expansion, we show why existing methods falter: their selection strategies attempt to estimate the attention mass, but are confined to a first-order approximation. Therefore, we propose COBS (Cumulant Order Block Sparse Attention), an attention method that builds on NSA, incorporating a novel selector that stores a compressed second-order statistic per block. On the 32k RULER long-context retrieval benchmark, COBS raises the NSA baseline's mean score from 0.2999 to 0.8195, approaching dense attention at 0.9040 and closing about 86% of the gap, while using only 1.21x the KV cache read traffic of the NSA baseline and 15.15x less read traffic than dense. The same model preserves short-context behavior and attains lower position-wise negative log-likelihood (NLL) than dense attention in our comparison.

Process Reward Models (PRMs) have been proven to be highly effective in guiding test-time scaling (TTS) methods, which significantly boost the capabilities of LLM-based multi-agent systems. However, existing PRMs are text-based: they re-encode the entire trajectory text from scratch. In long multi-agent rollouts, the scoring cost, growing quadratically with respect to sequence length L, creates a severe computational bottleneck, severely limiting PRMs' application in long-context scenarios. To resolve this, we introduce KV-PRM, a highly efficient process reward model that eliminates the heavy text re-encoding by directly reading the KV cache produced naturally during the LLM's generation phase. By processing a single "verify token" against the pre-existing KV cache, KV-PRM reduces the scoring cost from O(L^2) to O(L). We formally prove that the KV cache contains strictly greater information capacity than text, and is more efficient for downstream reward modeling. Empirically, across the MATH, GSM8K, and AIME benchmarks, KV-PRM matches or strictly outperforms text-PRMs under various TTS methods such as Beam Search, MCTS, and Weighted Voting, with up to a 5,000x reduction in scoring FLOPs, a 37x reduction in latency, and a 34x reduction in per-sequence memory footprint compared to text-based PRMs.

The growing adoption of large language model-based agents within operating system workflows has increased the importance of energy-efficient inference on laptop-class systems-on-chip (SoCs). While cloud offloading remains common, it introduces reliability and privacy concerns that are particularly problematic for agentic workloads. Recent laptop SoCs, therefore, incorporate neural processing engines (NPUs) optimized for energy efficiency; however, effectively mapping attention mechanisms onto NPUs remains challenging due to architectural diversity and explicit data-movement programming models. In this work, we present STEEL, the first open-source implementation of FlashAttention targeting XDNA-like NPUs. STEEL introduces a dataflow formulation of prefill attention, enabling efficient exploitation of spatial parallelism and on-chip memory. Furthermore, STEEL addresses the load imbalance induced by the causal mask by leveraging a sparsity-aware pipeline placement onto the NPU array, reducing synchronization overhead and improving utilization. We evaluate STEEL on the AMD Ryzen AI 9 HX 370 SoC and compare its performance against optimized CPU and GPU implementations. Experimental results show that STEEL reduces energy consumption by an average of 9.17x and 1.75x relative to CPU and GPU baselines, respectively. On XDNA 1, STEEL achieves an average 9.6x latency reduction over the prior state of the art, and delivers a 22.8x speedup on average compared to a layer-by-layer attention implementation on XDNA 2.

Vision-Language Models (VLMs) are the perceptual backbone of embodied AI, but their energy footprint on edge hardware remains poorly understood. Existing efficiency efforts focus predominantly on reducing visual tokens, implicitly treating visual processing as the dominant energy cost. We overturn this implicit assumption through the first systematic energy profiling of on-device VLM inference, spanning five models across three architecture families, four input resolutions, and two hardware platforms (NVIDIA RTX 3070 and Jetson Orin NX). Our analysis yields three findings. First, average inference power is a model-intrinsic constant, invariant to input resolution, image complexity, and prompt type, with less than 5% variation across all conditions. This means that all energy variation across inputs must arise from variation in inference time, not from variation in power draw. Second, each output token costs 11 to 39x more wall-clock time than each input token due to the compute-bound and memory-bound asymmetry between prefill and decode, making output token count the dominant driver of both latency and energy. Third, image complexity, measured by the number of objects in an image, induces up to 4.1x energy differences at identical resolution. This variation arises not from increased visual processing cost, but from differences in output length. These findings expose a fundamental limitation of visual token pruning: even removing all visual tokens saves at most 10% of total energy for fixed-token models. Across models spanning 1 billion to 8 billion parameters, controlling output length saves up to 97% of total energy, with the energy dominance of decoding growing stronger at larger model scale. In short, the true energy bottleneck in edge VLM inference is not what the model sees, but how much it says.

Large Language Models are reshaping how software is developed and maintained. They are typically deployed in production using inference engines such as vLLM, which can efficiently serve pre-trained, highly configurable models. While prior work has focused on model architectures and hardware acceleration, the impact of inference engine configuration on energy consumption, performance, and output quality remains poorly understood. In this paper, we present a large-scale controlled study of three selected vLLM configuration options: attention kernel type, prefix caching, and chunked prefill. We evaluate all combinations of these configurations across 5 open-weight LLMs and 5 diverse inference tasks, totaling $9,000$ runs and $93,600$ measures. We analyze energy consumption, latency, and accuracy, and examine both main effects and interaction effects between configuration options and tasks. Our results show that the studied configuration options significantly impact energy and performance, mainly driven by attention type and prefix caching, while chunked prefill has a limited effect under the default vLLM serving configuration and evaluated workloads. These effects are highly model- and workload-dependent, and no configuration is universally optimal. We further show that model choice dominates global trade-offs, while configuration tuning provides local improvements along the Pareto frontier. Unexpectedly, inference options can also affect model accuracy.

A transformer can be built from operators that are legible by construction — bounded, named units that read as fuzzy set operations rather than dense activations — but legibility must be pressed for during training, and the pressure has a failure mode. A crispness penalty meant to sharpen a bounded operator into a decisive detector instead collapses it into a dead constant. An identity, E[v(1-v)] = mu(1-mu) - var, shows why — the penalty is a variance-minimizer blind to the difference between a live detector and a constant — and names the fix: a per-channel variance floor, the target legibility metric written as a loss, which recovers both legibility and quality. A learned per-unit fraction then retires the hand-set reserved-GELU partition of prior work: given the choice the model keeps no unit as pure GELU and routes 87% of its load-bearing computation through crisp operators. The result is the most legible transformer we have built — 78% of its feed-forward operands and 50% of its attention value channels are crisp-and-contextual detectors, and per-head legibility rises from 18% in shallow layers to 78% in deep ones. Read in the correct rotated per-layer frame, these units separate a clean detection (what a unit responds to) from a harder naming (what its output decodes to); and because the objective makes each unit crisp and sparse, edits to them are far more local — 50-184x in the deep layers where the edit sites concentrate — and can target explicit conjunctions a single neuron cannot express. Finally, a between-unit decorrelation pressure exposes a legibility dial: it trades a circuit's reuse for independence at no quality cost, turning concepts into single, surgically editable units and a prediction into a short explanation read off a handful of named operations. Quality holds at parity with a conventional baseline throughout.

Efficient serving of diffusion large language models (dLLMs) is hindered by convergence heterogeneity: when batching multiple requests, different sequences converge at different rates, causing faster requests to stall behind slower stragglers and introducing compute bubbles and tail latency. We present BlockServe, a continuous batching framework that integrates block-grained scheduling — immediately evicting completed requests at block boundaries — with mixed-state execution that extends dual cache and parallel decoding to heterogeneous batches via gather-scatter indexing. Furthermore, a compute-aware admission controller expands effective batch capacity through token-budgeted refill. On Dream and LLaDA across five benchmarks, BlockServe achieves 1.9--10.6$\times$ throughput over Fast-dLLM with comparable generation quality, establishing block-grained scheduling as a foundation for high-throughput offline dLLM inference.

As large language models (LLMs) scale, their memory and computation demands have grown substantially, making weight-only quantization a widely adopted technique for reducing model size with minimal accuracy loss. However, on current GPUs, CUDA-core-based dequantization introduces substantial instruction overhead, on-chip traffic, and pipeline stalls, making it a major bottleneck for high-throughput, cloud-scale LLM serving. To address these limitations, we propose StreamDQ, a lightweight architectural enhancement that enables on-the-fly dequantization in the memory subsystem for high-throughput, large-batch LLM inference. StreamDQ integrates compact DeQuantization Blocks (DQBs) into the base die of high-bandwidth memory (HBM) and performs inline dequantization on standard memory loads. A lightweight sideband tag on each memory read request selects the dequantization mode while preserving conventional load semantics. By relocating dequantization to the memory side, StreamDQ eliminates GPU-side CUDA-core-based dequantization, thereby reducing on-chip traffic on the GPU and avoiding extra HBM write-back and reload of dequantized weights at large batch sizes. Our evaluation shows that StreamDQ achieves up to 7.08$\times$ speedup and 90.23% lower energy for mixed-precision GEMM, with only 0.127\,mm$^2$ area and 0.355\,W power overhead per DQB in a 12\,nm CMOS process. For end-to-end LLM inference, StreamDQ reduces latency by up to 54.68% and improves decode throughput by up to 2.20$\times$.

The rise of reasoning models and agentic systems has made LLM token-generation latency a key bottleneck. Unlike chatbots, whose latency gains saturate at human reading speed, these systems generate intermediate reasoning tokens not consumed by humans. Thus, per-token latency directly determines end-to-end response time. Low-latency inference uses minimal batching, making token generation bandwidth-bound. Tensor Parallelism addresses this by sharding model weights across GPUs and loading them in parallel. However, scaling to more GPUs introduces All-Reduce overheads that grow with GPU count. Removing All-Reduce improves token throughput by 43% for Llama-3.1-8B on 8 H200 GPUs. We propose Synchronization-Free All-Reduce (SiFAR), which reduces synchronization overhead during low-latency inference. Existing oneshot and twoshot algorithms incur overheads from barriers before and after communication. First, we find that the bottom barrier in oneshot enforces a WAW dependency and eliminate it by co-designing communication and model execution to enable dual buffering. However, oneshot scales poorly with GPU count. Twoshot performs better at higher TP degrees but incurs an unavoidable bottom barrier. To overcome this, we leverage in-switch reduction in modern switches. We propose redundant pull, where each GPU reduces the full All-Reduce payload at the switch. This improves oneshot scalability while retaining its no-bottom-barrier advantage. Finally, to reduce top-barrier overhead, we observe that each decode step issues multiple All-Reduce operations, keeping GPUs tightly synchronized after the first. We therefore propose speculative reduction, which initiates data transfer before the top barrier and ensures correctness via lightweight validation. SiFAR reduces All-Reduce latency by up to 52% and improves end-to-end throughput by 18.6% for Llama-3.1-8B and 13.1% for Qwen3.5-397B-17B at TP=8.

Efficient inference in Large Language Models (LLMs) requires deciding where computation can be reduced while preserving model quality. We study this problem through multilayer perceptron (MLP) activation sparsification and token-level conditional routing. We first propose Sensitivity-Aware Thresholding for Sparsity (SATS), a threshold calibration method to choose layerwise gate thresholds using a local MLP output sensitivity proxy rather than calibrating thresholds directly from activation percentiles. While SATS retains the existing mechanism of sparsifying MLP activations by thresholding gate activations, it replaces percentile-based calibration with a sensitivity-aware selection rule. We then introduce a lightweight token routing framework that dynamically selects between a base path and a modified path on a per-token basis, rather than applying the modified computation uniformly to all tokens. We evaluate both methods on multiple recent open-weight LLMs. Our results show that SATS improves over the threshold-based sparsification baseline at matched actual sparsity and that token routing yields a more favorable quality-throughput trade-off than static activation modification baselines. Overall, our results suggest that improved threshold calibration and token routing can improve the quality-throughput trade-off in LLMs.

Persistent AI agents extend large language models (LLMs) beyond single-turn interaction into long-lived software systems. Unlike traditional chat assistants, unsafe content in these agents can propagate through persistent state, reusable skills, and tool-mediated interactions, creating a substantially larger semantic attack surface. We observe that most security-critical interactions in such agents are transmitted through natural-language token flows, including memory updates, tool arguments, retrieved files, and inter-component communications. This observation enables a new security formulation: unsafe behavior can be intercepted as risky semantic flows before reaching privileged runtime sinks. Based on this insight, we propose TokenWall, a runtime defense framework that acts as a semantic firewall over agent token flows. TokenWall performs boundary-aware semantic auditing over these flows, constructing structured source-sink audit records, applying lightweight local inspection before execution, and selectively escalating ambiguous high-risk cases to stronger arbitration modules. Unlike prior approaches that rely on sparse auditing or remote large-model oversight, TokenWall enables full-coverage pre-execution mediation while reducing remote arbitration and latency. Experiments on CIK-Bench show that TokenWall reduces attack success rate to 12.5% while maintaining a 97.4% benign executable pass rate without human confirmation. TokenWall further introduces only 0.69 seconds of additional latency on benign cases, demonstrating that semantic runtime containment can achieve a practical security-utility trade-off for persistent AI agents.

Recent unified multimodal models show a single architecture can jointly perform vision/language understanding and image generation/editing. However, they repeatedly feed all historical visual and textual inputs into a shared context window, limiting long-horizon multimodal dialogue due to visual token explosion and unreliable cross-turn referencing. We propose a Cognitive-structured Multimodal Agent that externalizes visual information into an Episodic Visual Memory and selectively reactivates relevant episodes during reasoning. The agent consists of a Perceptual Abstraction Engine for structured visual abstraction, a Cognitive Retrieval Engine for cross-turn memory retrieval, and a Multimodal Executive Controller for autonomous task inference and action planning. To address the lack of turn-level retrieval supervision in existing datasets, we develop a Unified Scenario Engine that programmatically generates structured multi-turn conversations with fine-grained retrieval annotations, enabling reinforcement learning to optimize abstraction and retrieval policies. We also construct a long-horizon visual-dialogue benchmark stratified by difficulty to evaluate episodic visual recall. Our 8B agent achieves 91.4% retrieval accuracy over 20-turn sessions, surpassing 32B baselines by +8.2% while nearly halving per-turn inference time (23.1s -> 12.7s). We further present the Cognitive-structured Multimodal Agent Harness (CMA-Harness), a tool-augmented deployment of the same cognitive structure integrating persistent multimodal memory, web access, image generation/editing/composition tools, and OpenAI-compatible serving. Structured memory and modular decision-making offer a more scalable, efficient paradigm for long-horizon multimodal agents than monolithic parameter scaling. Code: https://github.com/caseclose/cma-harness ; Project page: https://caseclose.github.io/cma-harness/

Event-based lip reading has recently emerged as a promising direction for visual speech recognition, benefiting from the high temporal resolution and motion sensitivity of event cameras. However, existing methods typically perform spatial compression before sufficient temporal modeling, which may suppress sparse and localized motion trajectories that are crucial for distinguishing similar lip movements. Moreover, most current approaches optimize temporal representations mainly at the word-classification level, leaving the underlying articulatory structure weakly constrained. To address these limitations, we propose a temporally enhanced framework for event-based lip reading. First, we introduce Trajectory-Aware Differential Aggregation (TDA), which performs local temporal modeling at each spatial location before adaptive spatial aggregation. Second, we propose Viseme-Guided Aggregation (VGA), a unified temporal module composed of a CTC decoder and a viseme-guided gated aggregation branch, which injects viseme-aware sequence supervision and improves final temporal aggregation for word recognition. Third, we incorporate an EMA teacher--student training strategy to enhance robustness under strong event perturbations. Experiments on the DVS-Lip benchmark verify the effectiveness of the proposed design, and extensive ablation studies further validate the contributions of TDA, VGA, and teacher--student consistency. Qualitative decoding results also demonstrate that the proposed CTC-based temporal modeling learns meaningful viseme-aware structure from event streams.

Sparsely-activated Mixture-of-Experts (MoE) language models achieve remarkable inference efficiency by activating only a small fraction of parameters per token, yet their full expert banks reside in memory at all times, creating a prohibitive deployment bottleneck. Existing structured pruning methods, largely designed for dense transformers, assess expert importance using locally derived heuristics that are blind to the interdependent nature of MoE routing. We introduce MAESTRO (Markov-chain Approximated Expert Sparsification via Transition-based ROuting), a structured pruning framework designed for MoE architectures that models autoregressive expert activation trajectories as Ergodic Markov chains whose stationary distributions encode cross-layer dependencies, yielding a globally aware importance heuristic. Evaluated across five diverse domains including Safety, Bias, and Ethics, MAESTRO outperforms state-of-the-art baselines by up to 10.61% in average performance retention under a strict 50% compression regime, while exhibiting substantially lower cross-task variance, indicating that global, routing-congruent pruning produces models that generalize more consistently across heterogeneous tasks.

Sparse autoencoders (SAEs) have emerged as a promising technique for mechanistic interpretability by learning a set of sparse latent features in large models, each of which encodes a distinct concept. However, in vision-language models (VLMs), vanilla SAEs struggle to learn modality-consistent concepts, with concepts often exhibiting fragmented coverage (i.e., disjoint regions) in the visual modality. To address this challenge, we propose a Structured Sparse AutoEncoder ($S^2AE$) that enforces concept consistency from both semantic and spatial perspectives in the visual modality. Specifically, we group image patches based on Transformer attention similarity and spatial proximity, and introduce a structured sparsity regularization when training the vanilla SAE. The regularization consists of exclusive sparsity for inter-group concept disentanglement and group sparsity for intra-group concept consistency, which drives the latent neurons by SAEs to specialize in distinct, semantically grounded concepts. Evaluated on the Qwen2.5-VL-7B-Instruct model, the method achieves 6.06% average improvement in semantic alignment (mIoU) and 60.81 in representational efficiency (lower l0 norm) while maintaining near-perfect reconstruction fidelity with an Explained Variance above 99%. Cross-modal analysis further demonstrates that $S^2AE$ enhances neuronal monosemanticity by this visual structural prior, achieving a 3.08% average gain in semantic consistency and a 2.37% average gain in monosemanticity scores for both modalities of multimodal features, thereby fostering more coherent and disentangled representations.

Speculative decoding accelerates sampling from an autoregressive LLM by using a faster auxiliary model to draft tokens which are then verified in parallel by the LLM. Standard speculative decoding is lossless: its rejection and resampling steps exactly preserve the LLM's sampling distribution. Recent work argues that relaxing this strict guarantee can yield further speed-ups, controlled capability-speed trade-offs, or even capability gains. We practically investigate training-free relaxed speculative decoding techniques, unify existing approaches within a shared framework, benchmark them on contemporary settings, and distil takeaways and empirical findings for practitioners. Important takeaways include: relaxation can require considerable capability evaluation unlike lossless speculative decoding, and many relaxed approaches rely on a drafter that is a good language model, making them unsuited for lightweight dedicated multi-token-prediction drafters.

Modern Video Object Segmentation (VOS) involves tracking and segmenting user-specified targets. While recent approaches have achieved remarkable performance in single-target scenarios, extending them to multi-target settings typically involves replicating the single-target processing for each individual object, resulting in reduced frame rates (FPS) with unbounded latency as target count increases. Built upon Segment Anything 2 (SAM2), we propose SAM-MT, which addresses this by transforming the model into an interactive framework for real-time Multi-Target video segmentation. SAM-MT uses explicit queries to represent different individual targets, in parallel with a shared representation for global context. It employs decoupled masked attention to keep individual identities distinct from cross-target interference, and sparse memory for stable temporal evolution, along with specialized strategies for occlusion handling and overlap prevention. SAM-MT successfully decouples latency from the number of targets, achieving real-time speed on par with single-target baselines (>36 FPS for 10 targets) while maintaining SAM2's robust video segmentation performance.

Speculative decoding accelerates LLM inference by drafting several tokens and verifying them in parallel. Block-diffusion drafters such as DFlash produce a draft block in one pass but model only per-position marginals, and best-first tree methods such as DDTree expand candidate trees from those marginals. The released Domino drafter adds a GRU-based causal correction that makes each draft token distribution path-dependent, a structure DDTree's factorized formulation cannot represent. We introduce DominoTree, a training-free best-first draft tree scored by Domino's conditional, non-factorized correction along each root-to-node path, made practical by restricting the per-node correction to a candidate top-M set. On Qwen3-4B across eight benchmarks, DominoTree reaches up to 6.6x speedup over autoregressive decoding and the highest mean accepted length of any evaluated method, up to 10.7 tokens per round, at every tested temperature. DominoTree constructs its tree with a GPU-native CUDA-graph builder that is bit-identical to a reference Python implementation, so acceptance is unchanged, while keeping per-round tree construction cheap. With this builder as default, DominoTree improves throughput over the released Domino decoder, the drafter it builds on, at every tested temperature: 9% to 10% overall on Qwen3-4B and up to 22% on Alpaca. It also outperforms DDTree and CaDDTree at every tested temperature, not only under greedy decoding. On Qwen3-8B, DominoTree keeps the highest accepted length at every temperature and gives a 24% throughput gain over DDTree at T=0; at higher temperature its edge over DDTree and CaDDTree narrows to a tie and a small loss, while its aggregate gains over DFlash and Domino persist.

We propose OPSD-V, an on-policy self-distillation paradigm for post-training few-step autoregressive (AR) video diffusion models. Existing few-step AR video generators can produce long videos with low latency, but still suffer from error accumulation and weakened motion dynamics during long autoregressive rollout. OPSD-V reduces long-horizon degradation while preserving the original few-step inference path. The key idea is to introduce real long-video data as temporal context during training and use it to provide dense trajectory-level supervision. Specifically, the student follows the exact inference-time rollout, generating each chunk conditioned on its own previously generated KV cache. In parallel, the teacher is evaluated at the same student-visited denoising states, but uses a cleaner AR-consistent temporal cache in which older history can be replaced by real-video context. This provides dense denoising-level corrective targets under on-policy AR cache dynamics, without changing the sampler, number of denoising steps, or inference-time cache mechanism. We apply OPSD-V to representative few-step AR video models, including Self-Forcing and LongLive. Experiments show consistent improvements in visual quality, motion dynamics, and VBenchLong scores. A user study with 10 participants comparing 20 video pairs shows that OPSD-V is preferred over the base models in 66.0% of overall-preference judgments (82.5% excluding ties).

Deploying classifier-free guidance (CFG) diffusion models under real-world compute budgets requires quantization, yet existing post-training quantization (PTQ) methods treat CFG models as single-branch networks, ignoring the paired conditional/unconditional structure that CFG inference fundamentally relies on. This structural blind spot has two consequences. At the system level, the two-pass CFG execution pattern imposes a latency overhead that parameter-count and bit-operation metrics conceal entirely, and commodity INT8 inference stacks fail to realize the theoretical efficiency gains that BOPs calculations promise. At the algorithmic level, calibrating against the guidance gap alone admits an exact null space: a quantized model can achieve perfect gap-fidelity diagnostics while the unconditional branch drifts arbitrarily, corrupting every guided prediction at inference time. This paper terms this the branch-drift trap, proves its existence analytically, and confirms it empirically through a false-positive result in which the best-calibrated model by standard diagnostics simultaneously produces the worst sample quality. To close the trap, Guidance-Aware Mixed Precision (GAMP) is proposed, which calibrates directly on the guided prediction, derives per-layer activation-bit sensitivity from guided-output degradation, and allocates bits via a greedy knapsack — provably preventing unconditional branch drift by construction.

Non-GPU AI accelerators are increasingly adopted as alternatives to general-purpose GPUs for large-model inference, but the real engineering cost of migrating demanding workloads beyond CUDA remains poorly documented. We present a field study of deploying two large inference workloads on a 16-device Huawei Ascend 910 system using CANN and vLLM-Ascend: an LLM-as-a-judge safety and alignment evaluation pipeline based on a W8A8 MoE judge model, DeepSeek-V4-Flash, and a multimodal medical vision--language benchmark based on DeepSeek-V4-Flash-Vision for MMMU and MMMU-Pro. Making these workloads reliable required twelve source-level patches to the vendor inference plugin, disabling several high-throughput features to preserve numerical correctness, and adding operational safeguards for recurring device-level failures. We summarize the main platform limitations in eight categories: incomplete operator and feature support, fragile parallelism, numerical faults in low-level kernels, immature graph compilation, unstable advanced features, limited scalability, weak observability, and ecosystem fragmentation. For each category, we report the symptoms, evidence, and likely causes. We also quantify the integration effort, concurrency behavior, and benchmark quality to show that both workloads were served correctly. Our study provides a reproducible reference for teams evaluating or operating non-GPU accelerators for large-model inference.

Large language models (LLMs) are increasingly constrained by memory capacity, weight bandwidth, and checkpoint storage during deployment. Existing low-bit compression methods mainly follow two directions. Scalar or group-wise quantization is simple and compatible with efficient low-precision kernels, but its representation capacity becomes limited when the target budget approaches 2 bits per weight. Vector-quantized weight compression provides a richer block-level representation, but usually introduces explicit codebooks, index lookup, and additional storage accounting. This paper presents BiSCo-LLM, a codebook-free binary spherical coding framework for extreme low-bit LLM weight compression. The core pipeline is built on three components. First, local weight chunks are mapped onto a unit hypersphere and binarized into compact spherical codes, so that the main payload is a bit-packed sign stream rather than explicit VQ centroids. Second, a residual BSQ stage encodes the reconstruction error left by the base spherical codec, providing an explicit rate-distortion path without stored codebooks. Third, category-wise recovery distillation is performed after replacing each Transformer module category, reducing the mismatch between local weight reconstruction and assembled model behavior. A small 8-bit protected-channel path is used as an auxiliary stabilization mechanism for sensitive channels and is counted separately from the BSQ payload. The reported storage budget includes binary codes, neural decoders, protected-channel payloads, LoRA adapters, and metadata.

Post-training quantization is widely used to deploy large language models in resource-constrained settings, yet its evaluation relies almost exclusively on accuracy and perplexity. We show that these metrics fail to capture behavioral changes induced by quantization. We introduce correctness agreement, a decision-level metric that measures overlap in correct predictions between a base model and its quantized variants, independent of absolute accuracy. Across multiple models and quantization schemes from 8-bit to 2-bit, we find that behavioral divergence emerges under moderate quantization even when task performance appears preserved. To explain this effect, we analyze quantization as a structural operator on attention weights and quantify layer-wise distortions using statistical and distributional measures. Our results reveal non-linear breakpoints at low bit-widths and show that query and key projections are consistently more sensitive than value and output projections. These findings expose an illusion of equivalence between base and quantized models and motivate behavioral evaluation beyond conventional performance metrics.

The emergence of vision language models with fewer than 3 billion parameters has accelerated the implementation of on-device multimodal intelligence. However, a detailed understanding of component-wise quantization remains a bottleneck for optimal deployment. This paper presents a systematic evaluation framework for empirically validating five hypotheses across six quantization configurations on the Jetson Orin NX and AGX. By separating the vision encoder, projector, and large language model backbone yields the following results: (1) Quantization sensitivity is governed by the structural paradigm (MoE vs. dense) rather than scale alone, with MoE backbones mitigating INT4 noise where dense backbones degrade; (2) SigLIP encoders incur disproportionate INT8 latency on Jetson Ampere--a deployment-specific encoder-kernel-hardware interaction, not a SigLIP flaw; (3) Although INT4 quantization of LLMs greatly reduces VRAM consumption, it also causes slower token generation due to dequantization overhead; (4) Composite quantization errors are largely additive, except along the modality-alignment path, which is architecture-dependent; (5) The intelligence-per-joule profile varies significantly across platforms owing to memory bandwidth constraints.

Crossbar-based In-Memory Processing (IMP) accelerators achieve high-speed, low-power computing for deep neural networks (DNNs), but face three obstacles. First, floating-point (FP) arithmetic is incompatible with crossbars, and existing quantization schemes still require FP processors for scaling factors, incurring hardware overhead. Second, redundant DNN parameters occupy too many crossbars, and current IMP-aware pruning methods require data aligning across crossbars, introducing significant memory and computing overhead. Third, non-ideal crossbar behaviors such as write variations degrade the accuracy of deployed models, and existing compensation methods add substantial overhead. In this paper, we address all three problems within a single training process. We reuse bit-shift units in crossbars to approximately multiply scaling factors, avoiding FP processors. We apply kernel-group pruning and crossbar pruning to remove the hardware units needed for data aligning. We adopt runtime-aware non-ideality adaptation to relieve the impact of device non-ideality from the training stage by exploiting crossbar features. Integrating these three optimizations into one comprehensive learning framework reduces training overhead and improves accuracy. Experiments show that our quantization incurs a negligible accuracy drop, and our pruning achieves higher sparsity and accuracy than state-of-the-art methods. Our framework produces integer-only, pruned, and reliable VGG-16 and ResNet-56 models for CIFAR-10 on IMP accelerators, with accuracy drops of only 2.19% and 1.26%, respectively, without hardware overhead.