生成模型与LLM推理优化

Bimanual manipulation policies trained with imitation learning are typically evaluated on workstation or datacenter-class GPUs, leaving the cost of deploying them on embedded hardware largely uncharacterized. We present a bimanual SO-101 system running entirely on an NVIDIA Jetson Orin Nano Super (8 GB), the entry-level tier of NVIDIA's embedded line, using a desktop GPU (RTX 3070) only for offline training, evaluated on pick-and-place of a deformable beanbag. First, we build a GStreamer capture pipeline backed by NVMM buffers that removes redundant host-device copies from three-camera sensing. Contrary to expectation, the conventional path fit the memory budget and dropped no frames; what zero-copy sensing recovers is CPU headroom (peak single-core utilization 98.0% to 77.0%) and worst-case latency (117.31 ms to 101.52 ms). Second, we train ACT and Diffusion Policy on identical demonstrations, each at its own reference budget (100k gradient steps for ACT, 200k for Diffusion Policy). ACT converges to a task-competent policy (19/20 trials) while Diffusion Policy does not converge to a usable one (0/10) even at twice the step count, which we attribute to differing convergence costs rather than an accuracy ceiling. Third, we convert ACT to TensorRT. FP16 reduces mean inference latency from 114.02 ms to 17.93 ms (6.4x) and INT8 to 12.65 ms (9.0x), with task success preserved at all three precisions (19/20, 18/20, 19/20). We report two findings not previously documented for ACT: TensorRT's general-purpose INT8 calibration quantizes the ResNet18 backbone but accepts zero of 145 transformer layers, explaining INT8's negligible size reduction over FP16 (0.9%) despite a further 28% latency gain; and the need for quantization is conditional on ACT's action-chunking configuration, feasible in full precision at n_action_steps = 100 but not at the per-step re-prediction temporal ensembling requires.

Modern large language models (LLMs) exhibit activation sparsity, wherein only a subset of their neurons is activated for given input tokens. Researchers have leveraged this property to optimize LLM serving systems by omitting weight accesses and computations pertaining to inactive neurons. Unfortunately, however, such optimizations create input-dependent weight accesses, which can be leaked over side channels. We present SparSEEty, a new token extraction attack that exploits input-dependent neuron weight accesses introduced by sparsity-exploiting LLM serving systems. SparSEEty first constructs a neuron-activation oracle using neuron weight access side channels during LLM inference, and then inverts the activation traces to reconstruct the input tokens, forming an end-to-end token extraction attack. We instantiate SparSEEty against an LLM serving system protected inside an Intel TDX confidential virtual machine (CVM), addressing three key challenges: (i) constructing a neuron-activation oracle using a combination of side channels exposed by CVMs, (ii) reducing inference-time overheads of neuron activation monitoring for covertness, and (iii) accurately inverting partial binary activation traces back to tokens. Our evaluation shows that SparSEEty can reconstruct both prompt and response tokens with consistently high BLEU scores (>0.95) across various models and datasets, while incurring monitoring overheads of 3.7% to 7.2%.

Codebook-driven generative compression uses a pretrained image or video generator as a zero-shot visual prior and transmits compact codebook indices to guide reconstruction at ultra-low bitrate. Current codecs tie each finite-rate correction to a fresh prior evaluation, so shortening the sampler also removes correction slots that carry target-dependent information. We propose GVCCTurbo, a BPP-driven scheduler that separates expensive prior refreshes from codebook corrections: after calibrating an atom-count operating point and skip-gap ratio once per protocol, it maps a target codebook-payload bitrate to a trajectory length and refresh period, making BPP a schedule input instead of a fixed consequence of sampler length. The same endpoint-prediction and finite-rate steering interface covers GVCC-style rectified-flow video and DDCM-style diffusion image compression, preserving zero-training deployment and compatibility with future distilled priors. Native 1080p curves position the complete zero-shot codec in the ultra-low-bitrate regime. In a controlled 720p Wan-GVCC study, the scheduler cuts prior evaluations from 20 to 9 for a $\sim\!44%$ measured decoding-time reduction shared across the whole schedule family, at a small shared LPIPS cost on high-motion content; within that family, uniform refresh thinning (pure-skip) is a boundary point, and the BPP-aware interior point trades $2.9%$ fewer codebook-payload bits for consistently higher PSNR at comparable LPIPS. These results support BPP-to-compute scheduling as a controllable extension of sampler-length tuning, without requiring the allocated point to dominate every boundary point.

Modern deep learning workloads increasingly comprise heterogeneous computation graphs that combine compute-intensive operators with memory-intensive subgraphs. Existing deep learning compilers typically optimize these operator classes separately, creating rigid fusion boundaries that limit cross-operator optimization and on-chip data reuse. We observe that downstream memory-intensive operations can execute concurrently with compute-intensive operators, allowing their execution to be hidden behind computation; however, automatically exploiting this opportunity poses new compilation challenges. In this paper, we present ComFuse, an automated GPU compilation system that employs a novel operator fusion strategy to generate high-performance kernels for complex graph structures comprising compute-intensive operators and dependency-rich, memory-intensive elementwise-reduction subgraphs. ComFuse further supports the fusion of back-to-back GEMM (B2BGEMM) patterns, extending its applicability to more complex compute-memory interaction patterns. Additionally, it automatically lowers high-level tensor subprograms into optimized fused kernels, reducing the need for manual kernel engineering. Experimental results show that the fused kernels generated by ComFuse outperform those produced by TorchInductor across post-norm workloads and various complex computation scenarios, while supporting more flexible fusion patterns.

The increasing scale and computational demands of large artificial intelligence models (LAIMs) present significant challenges for efficient inference in resource-constrained distributed environments. In this paper, we propose a multi-cluster LAIM co-inference framework, where an edge server equipped with multiple graphics processing units (GPUs) coordinates multiple user clusters to execute inference tasks collaboratively. Within each cluster, devices capture data from diverse perspectives and employ lightweight on-device LAIMs to extract local features. These features are then transmitted to the edge server, where they are aggregated and fused to generate a more accurate inference outcome. To reveal the fundamental trade-off between model pruning and collaborative inference performance, we develop a theoretical framework that characterizes the impact of pruning ratios and device contributions using rate-distortion theory and partial information decomposition. Based on this analysis, we formulate a joint optimization problem that determines the model pruning ratio, the task scheduling strategy, the bandwidth allocation, and the transmission power, with the goal of minimizing the inference distortion while satisfying the constraints of latency, energy consumption, and server capacity. Extensive simulation results demonstrate that the proposed framework significantly outperforms existing benchmark schemes, achieving superior inference accuracy and resource efficiency in multi-cluster edge intelligence networks.

Static quantization assigns one weight precision to every denoising step. To preserve quality, that precision must accommodate the most quantization-sensitive step, even though many other steps can tolerate fewer bits. The resulting model may satisfy its memory budget, but it repeatedly pays worst-case arithmetic throughout the denoising trajectory. We introduce Temporal-Adaptive Bit Sparsification Quantization (TASQ) to separate these two costs. TASQ stores one shared maximum-precision weight buffer and learns a Temporal-Spatial LSB Mask that selects a lower effective precision for each layer and denoising stage by truncating least-significant bits. Storage therefore remains fixed by the worst case, while BitOPs decrease at less sensitive stages without per-stage weight copies or runtime search. A Temporal-Precision Engine maps the learned schedule to bit-serial execution, where cycles scale with effective precision and switching precision has no measured cycle overhead. On PixArt-Sigma, SANA-1.6B, and SDXL-Turbo, TASQ achieves quality comparable to static quantization with less computation. Together with the Temporal-Precision Engine, it reduces execution cycles by 25 to 50 percent over static quantization and by 6.1 to 7.5x over a naive static 8-bit bit-serial execution. Code is available at https://github.com/seokho-han/tasq.

The growing demand for mobile video streaming requires edge delivery systems that adapt efficiently to rapid network fluctuations and diverse user preferences. Existing approaches such as FlyCache rely on reactive ABR heuristics and loosely coupled cache policies, limiting their responsiveness and coordination under real-world wireless dynamics. We propose ProCAVE (Proactive Caching with Adaptive Video Experience), a self-adaptive DRL-based framework that unifies predictive bandwidth modeling, proactive bitrate selection, and preference-aware cache control. ProCAVE employs: (i) a lightweight Transformer for short-term throughput forecasting; (ii) a PPO-driven ABR agent; and (iii) a DDPG-based continuous cache controller operating on a high-dimensional global state. Experiments using MovieLens preference traces and Ghent 4G bandwidth measurements show that ProCAVE improves byte hit rate, reduces backhaul load, and enhances QoE compared with FlyCache and other baselines. These results highlight the benefits of predictive, DRL-coordinated control for efficient and user-centric edge video delivery.

Fewer visual tokens do not guarantee lower end-to-end latency. We evaluate break-even with a reproducible protocol that accounts for decision overhead, shared work, and the operators each policy can avoid. A stage-level decomposition reconciles these components with measured end-to-end latency. In a 30-example pilot, the two tested autoregressive probes remain slower than Full despite state reuse. A lightweight post-vision predictor yields paired confidence intervals below zero on RTX 3090 and A100 and remains significant after a conservative all-pairs Holm correction. A pre-vision image-size rule also yields intervals below zero on both GPUs, although neither comparison remains significant after the same correction. Pre-vision routing has a structural opportunity unavailable to post-vision pruning: it can avoid preprocessing and vision encoding. On A100, this opportunity outweighs a nearly eightfold larger downstream token reduction by the post-vision policy. Reported quality is conditional on examples answered correctly by Full and is not benchmark accuracy.

Ternary LLMs make their weight-dominated projections compact and efficient, but attention remains a mismatch: its K/V cache is created online and is typically processed by a separate higher-precision engine. Compressing this cache alone does not resolve the mismatch. To execute attention with the same lookup-table machinery as ternary projections, values accumulated in one reduction must retain a compatible representation and scale. This requirement also differs for keys and values during causal decoding, because newly generated values may belong to an unfinished cache block. This work develops a unified lookup-table inference approach for ternary LLMs. It stores runtime K/V states as scaled multi-plane signed digits organized around the reduction structure of attention. The resulting digit planes are consumed directly by activation-derived tables, avoiding dense K/V materialization between cache storage and attention computation. The design combines online K/V formation, bounded handling of incomplete value blocks, and a shared multi-stream datapath for Linear projections and attention. A constraint-guided search selects the representation and execution policy for a target quality--efficiency trade-off. Experiments on native and post-training ternary models validate the approach across cache capacity, model quality, and hardware efficiency.

Low-bit quantization suffers severe accuracy degradation on compact networks, rooted in the dominant full-parameter coupled training paradigm that ignores parameter subspace heterogeneity. Their limited feature redundancy leaves little room to absorb quantization errors. Conventional pipelines adopt monolithic optimization: PTQ reconstructs fixed pretrained models without improving inherent quantization friendliness; QAT updates all parameters jointly, suffering from gradient coupling between backbone weights and calibration parameters. In this paper, we identify normalization affine parameters as a low-dimensional high-leverage subspace dominating quantization robustness, and propose Normalization Affine Preconditioning (NAP) for targeted subspace optimization. For PTQ, NAP freezes backbone weights and fine-tunes only affine parameters under the target fake-quantization graph on full-precision models, proactively boosting quantization friendliness before downstream reconstruction. For QAT, we introduce an alternating QAT-NAP schema that decouples feature learning and numerical calibration, breaking the performance ceiling of saturated joint training. Theoretical analysis confirms BN affine parameters fully cancel the channel-wise affine component of quantization distortion, while nonlinear rounding and clipping residuals form the irreducible error boundary; distillation-guided NAP acts as directional flatness optimization, projecting teacher-student logit mismatch onto the restricted subspace. Experiments on ImageNet and CIFAR-100 show NAP recovers severely collapsed low-bit quantization, consistently boosts reconstruction-based PTQ, and outperforms saturated full-parameter QAT with negligible tuning cost. This work reveals the principle of targeted low-dimensional subspace optimization, offering a new perspective beyond full-parameter coupled training for efficient deep learning.

Existing low rank KV cache methods preserve either model weights or key variance, neither of which directly reflects the attention scores used during inference. We derive the expected attention score distortion caused by rank r key compression and show that it yields a covariance weighted low rank objective. Under a margin condition, controlling this distortion also improves top k recall. The optimal rank r solution has a closed form asymmetric factorization obtained from the SVD of the covariance weighted query key operator. This motivates SAKI, a training free KV cache index that directly preserves attention scores rather than key reconstruction quality. Across LLaMA 3.1 8B, Qwen 2.5 7B, Mistral 7B v0.1, and Llama 3.2 3B, SAKI outperforms key PCA at every tested rank. At rank 32, it removes 13 to 30 percent of PCA's remaining top 64 recall error, including improvements from 0.748 to 0.799 on LLaMA 3.1 8B and from 0.786 to 0.850 on Qwen 2.5 7B. It improves 68 to 89 percent of attention heads per model, with the largest gains in deeper layers. Predicted score MSE reductions closely match empirical measurements, with a Pearson correlation of 0.997, while ablation studies confirm that the gains arise from optimizing the attention score objective rather than covariance weighting alone. Analysis of the scoring operator further explains why weight only, invariant subspace, and key reconstruction methods can be suboptimal. SAKI uses random-matrix theory to separate genuine covariance signal from autocorrelated sampling noise, matching PCA with only 512 calibration tokens and adding value exactly where PCA sees no reliable signal.

Physical AI policies require inference throughout their lifecycle, including model evaluation, cloud reinforcement learning rollout, edge GPU serving, and onboard deployment. Although these settings share the same checkpoint and action semantics, they often rely on separate inference programs. To unify them, we build PhyAI, a Physical AI inference engine with a single runtime that keeps architecture-specific conditioning, solver, cache, and output logic in model adapters while sharing graph execution, kernels, memory management, and parallel services. The same codebase runs vision-language-action (VLA) models and world-action models (WAMs) on single or multiple GPUs across onboard, edge, and cloud deployments. We used the adapter interface to add MiniCPM-Robot on the day of its release. PhyAI achieves 1.40x-4.65x speedups over the official implementations of pi0, pi0.5, GR00T N1.7, and MiniCPM-Robot. On Cosmos3-Nano-Policy-DROID it reduces latency from 2.46 to 1.18 s on eight H20 GPUs (CFG=2, TP=4), a 2.08x speedup. Specialized runtimes remain faster in several configurations, so our goal is one runtime with competitive latency rather than the fastest result in every case. Detailed profiles reveal why different models need different execution policies: on a Hopper-series GPU at batch size one, the pi0.5 action expert accounts for 8.8% of FLOPs but 57.2% of latency; at batch size 32 its share drops to 13.5% and throughput reaches about 100 samples/s. Cosmos3 remains generation-dominated and gains only 14.3% throughput as batch size increases from 1 to 16. We further introduce the control-time Roofline, which distinguishes inference-bound from environment-bound control; the measured pi0.5 points on four LIBERO suites are environment-bound while Cosmos3 stays inference-bound. Code and benchmarks: https://github.com/mingti-org/phyai.

Small language models are often the only option for deployment under tight latency, cost, and on-premises constraints, but they are rarely trained from scratch: a compressed model is usually recovered through knowledge distillation (KD). This recovery step largely decides the final quality, yet it is expensive. We present a practitioner's study of how to make distillation training efficient, organised around two systems contributions. First, we show that offline KD (caching the teacher's top-$K$ logits once and training the student against the cache) matches online distillation at near-identical training loss while removing the teacher from memory, running about 29% faster per iteration, and reaching up to 41% higher throughput on a single H200 GPU. Second, we introduce a fused, chunked KL loss that never materialises the full vocabulary-sized logit tensor, making peak memory linear in the sequence length. This removes the memory spike that otherwise caps context length and lets us train at four times the context (32{,}768 tokens) on a single GPU. A separate output-head-only toy benchmark isolates the loss kernel and confirms its memory and iteration-rate scaling from 4K to 256K tokens. Together these make large-scale healing and hundreds of ablations affordable. We also report supporting ablations on loss design and sequence packing. We release our chunked-loss implementation: https://github.com/CompactifAI/Full-Chunked-KL-Loss.

Video diffusion transformers (vDiTs) generate high quality but pay quadratic self-attention cost, making inference prohibitive at video-token scales. The challenge is input-adaptive sparsity: selecting critical Q/K/V tokens with negligible overhead and executing them for end-to-end gains. We present SPADE, a training-free sparse-attention engine of three parts: (i) vDiT-SSR, a specification defining 3D blocking candidates and formalizing dynamic masks via Summarizer/Estimator expressions; (ii) runtime scheme generation using SICS and a head-wise policy; and (iii) an executor with low-overhead index search, flash block-sparse attention, and kernel grouping. Across Hunyuan-Video and Wan 2.1/2.2 for text-to-video and image-to-video generation, SPADE raises sparsity and speed while preserving quality, accelerating attention by 2.26x-3.40x and end-to-end inference by 1.49x-1.80x. Our code is open-sourced at https://github.com/6somehow/DAC-SPADE.

Production deployments often swap between different-sized models in a family for cost-quality cascading, mid-conversation switching, and routing, and each swap forces the receiver to repay the prefill from scratch. We propose cross-model KV cache transfer, where the receiver reuses the source's KV cache, skipping prefill. We find that cross-model KV has substantial linear structure across matched-KV pairs, where source and target share KV head count and per-head dimension. On Qwen3 14B->32B, one source layer explains 56% of variance in the target's keys and 32% in values, rising to 79% and 65% with multiple source layers. Building on this, we design a closed-form ridge mapper that operates per head and proceeds in three steps. First, for each target layer we select the top-k most predictive source layers and concatenate their KV as input. Second, we strip RoPE from the keys before mapping, so the fit is position-free and reusable across context lengths. Third, we fit ridge regression on a small calibration set of 500 FineWeb-Edu sequences of 1,024 tokens each. Surprisingly, across six pairs in three families, this linear mapper retains 73-98% of the receiver's standalone-prefill accuracy on four pairs, while two degrade sharply. A nonlinear MLP recovers up to +37 pp HellaSwag retention on the failures. The mapper runs 2.7-25x faster than re-prefill and remains stable across multi-turn handoff, making cross-model KV cache transfer practical.

Speculative decoding verifies a tree of draft tokens in one target-model forward pass. For a mixture-of-experts (MoE) target, however, parallel verification can activate the union of the experts selected by all tree nodes, even though only a small subset of those nodes reaches the accepted output. Token count, activated-expert union size, and expert-weight traffic are therefore distinct cost measures: reducing the token workload need not shrink the expert union proportionally, and under offloading, transfer traffic also depends on cache residency. We introduce AcceptMoE, a verifier-side expert selector that combines target-router scores with offline-estimated commitment probabilities and automatically adjusts the number of eligible experts for each verification block, eliminating the need for a user-specified expert budget. Under offloading, AcceptMoE conditions expert eligibility on cache residency instead of predicting natural routes and prefetching the corresponding expert weights. Although constraining target-expert eligibility changes the model distribution, across 12 model-task pairs spanning three MoE targets and four benchmarks, AcceptMoE's mean accuracy is 0.27 percentage points lower than that of EAGLE-3 speculative decoding with natural routing. Served with SGLang at batch size one, it reaches 1.290 times the throughput of this baseline with all expert weights in GPU memory, and 2.06 times under physical expert offloading, while reducing host-to-device traffic by 73.6 percent to 77.1 percent.

Large Language Models (LLMs) are integrated into software systems and AI services, making efficient LLM serving a concern for software engineering. Serving LLMs is challenging because inference requires computation, memory, GPU resources, and execution while maintaining latency and throughput. Although prior research has proposed LLM inference, optimization, and serving techniques and frameworks, little is known about how they are adopted in practice. In this study, we investigate the use of LLM serving frameworks and serving methods in open-source software systems. We identify and analyze five LLM-specific frameworks: vLLM, SGLang, TensorRT-LLM, LMDeploy, and FlashInfer. We examine how these frameworks and techniques are adopted individually and in combination, how adoption varies across categories of LLMs, and how repositories differ in intent, focus, use case, and architectural design. Our results show that vLLM is the most visible framework in popularity and adoption, while parallel computation, memory management, and network pruning are the most frequently used serving-method categories. Multi-framework usage is limited, suggesting that developers rely on a single serving framework; however, combined frameworks connect complementary capabilities across the serving stack. Framework adoption varies across model families, modalities, model sizes, domain specializations, and deployment settings. Repository-level analysis shows that LLM serving frameworks support applications and architectures, including Reinforcement Learning (RL)-based reasoning, multimodal generation and understanding, microservices, and cloud infrastructure. Overall, this study provides a large-scale empirical characterization of LLM serving framework adoption in practice and offers insights for researchers, framework maintainers, and practitioners working on LLM systems.

Agentic inference now dominates the LLM inference landscape, requiring LLMs to actively engage in multi-turn interactions with tool-calling capabilities. This introduces a more complex workload for the underlying inference system: serving stages such as prefill and decode exhibit substantially different behaviors and demand distinct compute and memory-bandwidth capabilities. As a result, a single homogeneous GPU system now struggles to support agentic inference, motivating an industry shift toward heterogeneous systems with disaggregated serving capabilities, such as the emerging Vera-Rubin platform with GPUs and Groq LPUs. However, the question of what the optimal hardware should look like for each component in a heterogeneous system remains underexplored. To this end, we propose a novel simulation framework for disaggregated serving, termed HeteroPanacea, that enables system-level simulation across three dimensions: 1) disaggregated quantization, 2) automated intra- and inter-device parallelization scheduling, and 3) PDAF (prefill-decode-attention-FFN) NPU architectural heterogeneity. By combining these three axes, we provide a cross-stack simulation framework for future heterogeneous agentic serving systems. We confirm the benefit of Prefill Decode disaggregation, simulating increased serving throughput by up to 75% compared to traditional serving with current GPUs and demonstrate 4 way Prefill Decode Attention FFN disaggregation is the most consistent for increasing throughput across different models, assuming custom NPUs. We also investigate the relationship between model architecture and gain from disaggregation by running a set of ablation studies.

Long-context inference with large language models is constrained by the linear growth of the key-value cache to sequence length. While pruning offers mitigation, prevailing methods determine query-specific token importance that cannot be reused across unseen queries. In contrast, we introduce TaskPress, a framework for task-guided, query-agnostic KV cache eviction. Instead of optimizing the cache for a single query, TaskPress constructs a reusable memory representation conditioned on a high-level task guide. The guide functions as a meta-query during prefill to filter irrelevant tokens before downstream queries are issued. In addition, TaskPress leverages quantization scale factors as a zero-cost signal for detecting influential representation outliers, providing an efficient proxy for token importance. Experiments on conducted on various tasks with long context input demonstrate that TaskPress efficiently creates a compact, reusable cache across diverse queries.

This paper presents a heterogeneous decode-phase serving system that relocates the KV cache out of GPU memory, motivated by the retrieval-based sparse attention that recent frontier LLMs adopt to serve million-token contexts. It partitions a decode step by operation type: GPU nodes hold the model weights and execute the projections and MoE layers, while processing-near-memory (PNM) nodes hold the KV cache and index keys and execute every operation that reads them. We first show that the assumptions behind prior PIM and PNM designs no longer hold for these operations, and derive four design requirements for such a node. From these requirements, we propose KARAT (KV-cache-resident Accelerator for Retrieval-based ATtention), a general-purpose PNM design that is the design point meeting all four. A KARAT device combines large LPDDR capacity with general-purpose compute sized for the retrieval indexer, serving an operational intensity beyond what PIM/PNM designs built for low-intensity GEMV target while accommodating diverse sparse attention algorithms that fixed-function units cannot support as they evolve. To reduce pipeline bubbles as the two device types alternate between micro-batches, we further propose opportunistic, fine-grained micro-batch scheduling (OFMS), which hides expert all-to-all behind the other micro-batch's GEMMs, and context-length-aware micro-batch rebalancing (CMR), which equalizes their token counts despite the variance in context length. Across three state-of-the-art models and real agentic traces, our proposed system improves throughput per TDP under a service-level objective by 2.09-6.13x over a GPU-only baseline and runs training-free sparse attention methods with 1.36-3.21x improvements.

When decoder language models are used as classifiers, predicted class probabilities depend on implementation choices, including the prompt template, verbalizer (label-to-token mapping), and scoring rule, that are rarely treated as experimental variables. We present a controlled evaluation of three Mistral-7B variants (Base, BioMistral, and Instruct) on PubMed RCT sentence classification (n=2000) under FP16, INT8, and INT4 precision using four answer-text prompt templates. Our primary finding is that the probability extraction protocol dominates apparent calibration. Switching from summed to mean token log-likelihood scoring reverses the calibration ranking between models: BioMistral average expected calibration error increases from 0.097 to 0.289, whereas Instruct decreases from 0.237 to 0.096, while accuracy changes by less than 1 percentage point for the specialized models but 4-6 percentage points for the base model. Prompt template choice produces accuracy differences of 7-24 percentage points, comparable to or larger than model-level effects. On one template, BioMistral outperforms Instruct although the overall mean favors Instruct by only 1.3 percentage points. For BioMistral and Instruct, INT8 quantization changes accuracy and F1 by only 1-2 percentage points relative to FP16, whereas the base model shows larger INT8 effects on some templates (up to +4.2 percentage points). INT4 produces heterogeneous but non-catastrophic effects. Temperature scaling reduces expected calibration error under summed scoring for both models but only for that scoring rule. A fine-tuned PubMedBERT reference achieves 82.7% accuracy but uses about 176000 labeled training examples, precluding direct comparison. These results demonstrate that prompt template design and scoring normalization are first-order experimental decisions when evaluating decoder language model calibration.

Microscaling (MX) is now the standard for low-bit large language model (LLM) inference. Its 4-bit form MXFP4 still loses substantial accuracy, because existing MX formats fix either the element format or the precision-recovery scheme across blocks, and thus capture only limited quantization heterogeneity. Quantization heterogeneity appears at two levels: 1) across blocks, the preferred element format and precision-recovery scheme vary; 2) across operands, weights and activations require different encoding. We introduce AdaMX (Adaptive Microscaling), a heterogeneity-aware format and accelerator. It selects the precision-recovery scheme per block and the representation per operand, at no increase in equivalent bit width (EBW). One design covers two block sizes, giving a higher-accuracy operating point and a lower-EBW operating point that saves storage. We implement a 22nm FD-SOI AI accelerator prototype with the proposed decoder, computing unit, and quantization logic. Against an otherwise identical MXFP4 accelerator with FP4-only multipliers, AdaMX adds about 1% system energy. At the lower-EBW point, AdaMX stays more accurate than the baseline while lowering both memory footprint and energy. Across LLMs from 3B to 70B, AdaMX removes 83% of the MXFP4 accuracy loss on commonsense and 82% on MMLU, and 43% and 27% of the NVFP4 loss. AdaMX also generalizes to multimodal models. On Gemma-4 12B, it leads MXFP4 on all four vision-language benchmarks and keeps up to 96% of FP16 accuracy.

Kernel generation for hardware accelerators such as GPUs and NPUs has become a proving ground for large language models (LLMs), and state-of-the-art systems raise correctness through pipelines that couple LLMs with agentic reinforcement learning and evolutionary search. Such pipelines generate, compile, and execute large numbers of candidate kernels, discarding most of them and forgoing the opportunity to distill failures into reusable knowledge. Many discarded candidates are near-miss operators that compile and run but fail numerical validation; each embodies genuine domain knowledge and a nontrivial investment in LLM inference, cross-compilation, and hardware execution. We argue for a paradigm shift: rather than regenerate, debug. Debugging is far more constrained than generating from scratch: the search space is small and feedback is dense. We present a domain-specific debug agent that addresses three core challenges in autonomous repair: mitigating knowledge scarcity through retrieved patterns and diagnostic instrumentation, ensuring integrity through anti-cheat detection and full-coverage evaluation, and controlling cost via convergence guards and bounded iteration. Debugging serves two complementary roles: it extends the capability frontier by recovering operators that repeated regeneration fails to produce, and it lowers cost per deliverable operator. Debug Pass@1 achieves 66.7% versus Regenerate Avg Pass@1's 25.9% and Regenerate Pass@3's 40.7%, while consuming 92.8% fewer tokens per success than three-trial regeneration. Component ablations show that the knowledge base drives recovery, while integrity gates reject 12.5-33.3% of the successes the workflow itself accepted.

MLLM-based segmentation faces a core segmentation trilemma: high segmentation performance, preserved dialogue ability, and fast inference. Embedding-prediction methods may disrupt language modeling through pixel-level objectives, whereas next-token generation is inefficient for dense masks. We propose All-Mask Prediction, decoupling autoregressive dialogue from non-autoregressive mask prediction. Its binary instantiation, STAMP (Simultaneous Textual All-Mask Prediction), emits an in-vocabulary <SEG> trigger, fuses image-aligned mask tokens with corresponding patch features, and uses hybrid attention to classify all tokens as foreground or background in one pass. It thereby combines strong referring and reasoning segmentation with preserved multimodal ability and efficient inference. However, binary masks cannot retain multiple semantic or instance identities without repeated target-specific predictions. We therefore propose Structured All-Mask Prediction and develop STAMPlus. It generates a target list with explicit IDs and optional boxes, binds these IDs to a shared multi-class mask space, and jointly predicts all targets in one non-autoregressive pass. A single unified checkpoint retains STAMP's referring and reasoning capabilities while extending to open-vocabulary semantic, instance-aware, and remote-sensing small-target segmentation, where high-resolution mask-token scaling preserves finer spatial evidence. Across these settings, STAMPlus achieves state-of-the-art segmentation performance, preserves general multimodal instruction following, and reduces 12-category latency from 13.50s for repeated STAMP inference to 5.16s. Further analyses show that accurate target cues improve segmentation and learned spatial grounding benefits look-twice reasoning. Overall, STAMPlus resolves the trilemma beyond single-target prediction.

3D point cloud registration in laparoscopic surgery estimates the transformation between an intraoperative organ reconstructed from video and its preoperative mesh. Because ground-truth transformations are unavailable for real data, supervised networks are trained on synthetic organ pairs. At test time, real reconstructions differ from synthetic data and are noisy, sparse, and occluded, which degrades correspondence estimation. Test-time adaptation (TTA) can reduce this domain shift, but existing methods mainly rely on logits, entropy, class prototypes, or cache memories unavailable in registration. Registration also involves paired inputs with an asymmetric shift that primarily affects the intraoperative cloud. We analyse and modify state-of-the-art TTA methods from three families to 3D registration: model, normalization, and input adaptation. We analyze four representative approaches based on auxiliary-task model updates, backpropagation-free token purging, feature alignment, and layer-normalization calibration. We modify them to handle asymmetric shifts between preoperative and intraoperative point clouds and replace classification-based entropy objectives. Using a correspondence-based model trained on clean synthetic source data, we evaluate adaptation to corrupted synthetic and real target data on P2P and P2ILReg. For synthetic targets, we apply eight corruptions, including uniform noise and global density reduction, at five severity levels. All methods improve registration on P2P, whereas normalization adaptation degrades performance on P2ILReg. Considering the computational overhead of backpropagation-based adaptation, input adaptation is the most promising option for laparoscopic surgery, providing low inference latency and consistent error reductions across datasets. Code: https://github.com/ninaa-git/survey_pc_registration_tta

State Space Models(SSMs) such as Mamba and Mamba-2 achieve linear-time autoregressive inference, making them attractive for latency-sensitive and resource-constrained deployment. Yet their large input and output projection layers impose quadratic weight memory and off-chip bandwidth costs that bottleneck practical FPGA deployment, accounting for over 60% per-token runtime at sequence lengths of 1,024 and beyond. Existing accelerators reduce this overhead through quantization or activation sparsity, but none treat projection rank as an explicit hardware design variable, leaving a systematic accuracy-throughput trade-off unexplored. We present LowRank-SSM, a hardware-software co-design framework that closes this gap. On the software side, we decompose the input and output projection weights via post-training truncated SVD and introduce a greedy bandwise rank-allocation algorithm that searches for the per-band rank vector that minimizes weight storage while respecting a user-specified accuracy constraint. On the hardware side, we map the resulting factored projections onto a fully-pipelined accelerator on an FPGA, featuring a dual-path projection(low-rank path and full-rank path), a fused selective-scan unit, and five independent AXI master bundles that saturate DDR4 bandwidth without bus contention. A per-band runtime rank mask enables mixed-rank execution across all 64 layers with zero architectural overhead. On Xilinx Versal VC1902 at 400 MHz, the deployed mixed-rank INT8 design achieves 7.89~tokens/s, representing a ${2.19\times}$ throughput improvement and ${2.03\times}$ energy-efficiency improvement over SOTA at comparable power and accuracy.

Analog compute-in-memory (CIM) enables energy-efficient neural network inference, but device variation and read noise can severely degrade low-bit quantized models. Existing CIM-oriented quantization methods mainly minimize ideal quantization error, ignoring the hardware noise floor and thus causing inefficient precision allocation. We propose NANQ, a noise-aware mixed-precision non-uniform quantization framework for analog CIM. NANQ models magnitude-dependent weight noise from measured responses of an eFlash CIM array and converts the noise profile into an adaptive quantization density, assigning finer resolution to low-noise regions while avoiding ineffective precision in noise-dominated regions. It further assigns layer-wise bit-widths by identifying each layer's precision saturation point under hardware noise using a unified threshold. On-chip experiments on an eFlash CIM SoC show that, under 2-bit weight-magnitude quantization, NANQ improves vision-model accuracy by 8.05 percentage points and reduces language-model PPL by 54.7% on average over PowerQuant. Mixed-precision NANQ captures most of the gains obtainable from additional quantization resources with only 3.2-3.8 equivalent bits.

The attention score with rotary position embeddings (RoPE) decomposes exactly into a sum over its 2D-rotation frequency pairs, and each pair's wavelength limits how far it can discriminate position. Aligned with this structure, we propose the per-RoPE-wavelength distance window: it prunes the query--key inner-product terms beyond a wavelength-proportional distance. Unlike a sliding window, every key remains reachable, at least through the low-frequency pairs. The reduction rate is input-independent, with a closed form logarithmic in the sequence length $N$, in contrast to dynamic-sparse methods like MInference. Such token-level selection is orthogonal to our frequency-level pruning. The window can therefore be applied on top of those methods. On Qwen2.5-0.5B and Llama-3.2-3B, the window prunes 37--48% of the query--key inner-product terms within each model's native context length. Relative to full attention, the top-1 match rate stays at 96--98% and the mean output-distribution KL at the $10^{-3}$-nat level on LongBench-v2 contexts. We examine absolute scores on long-context benchmarks such as RULER, OpenAI-MRCR, LongCodeQA, and $\infty$Bench: they are broadly preserved. We implement the window as a slice of the query--key contraction axis, leaving the online-softmax recurrences untouched, and port it with minimal diffs into the released FlashAttention-4 prefill and FlashInfer decode. On RTX PRO 6000 with Llama, both ports outpace stock with gains growing with context length, up to $1.29\times$ at 128K. End to end on Qwen2.5-7B-1M, with 57% of the inner-product terms pruned, the speedup reaches $1.31\times$ at a 1M-token context.

The key-value (KV) cache is the primary memory bottleneck in long-context LLM inference. Existing approaches attack it from opposite ends: eviction methods permanently discard tokens, degrading performance whenever a discarded token later proves essential, while quantization methods retain all tokens at low precision but offer limited compression. We propose AnchorKV, a compression scheme that shrinks the cache by $20\times$ without discarding a single token. AnchorKV represents the cache using a small set of anchors stored exactly, expresses every other token through its most similar anchor, and refines only those whose approximation most affects the model's output. AnchorKV consistently preserves accuracy across models and datasets, retaining 99% of the full-cache score at the 70B scale, while keeping the entire context at a fraction of its cost.

The key-value (KV) cache has become a major memory and bandwidth bottleneck in long-context large language model inference, making ultra-low-bit quantization increasingly important. However, existing rotation-based INT2 methods optimize cache statistics or proxy errors before the complete attention readout, even though the model is ultimately affected by the error propagated through attention and the output projection $W_O$. To address this mismatch, we propose OptR, an output-aware rotation method that minimizes post-$W_O$ attention-output error. OptR decomposes the post-$W_O$ attention-output error into key- and value-induced terms and learns per-head orthogonal corrections through the full INT2 quantization and attention path. OptR further applies an attention-equivalent key reparameterization to reduce large channel-wise offsets without changing the softmax distribution. Across three models and five reasoning and coding benchmarks, OptR consistently improves both QuaRot and OSCAR and strengthens long-context retrieval, while preserving the paged KV-cache format with negligible inference overhead.