生成模型与LLM推理优化

As LLM inference shifts to multi-tenant GPU clusters, co-batching improves throughput but obscures per-tenant usage and limits control. Enabling fractional sharing of the inference engine requires a real-time, per-request attribution primitive that is accurate and light enough to run inside the scheduling loop. We present LLMVisor, a roofline-guided latency attribution model that captures the memory-bound and compute-bound phases via a concise piecewise-linear form over features proportional to FLOPs and memory I/O traffic. LLMVisor decomposes batch latency into additive, per-request shares and runs efficiently at microsecond scale. We evaluate LLMVisor across Llama 3.1-8B and Qwen 2.5-14B/32B on A100/H100 GPUs under varying tensor parallelism and workload mixes. Compared to a token-count baseline, LLMVisor attains near-perfect R-squared and reduces relative error by up to 2.5x and 3.3x at p90 and p99, respectively, for prefill, and by up to 3.5x and 4.4x for decode, despite batching variability and sequence divergence.

Long-context LLM inference is bottlenecked by KV cache memory, yet distributing a limited cache budget across layers remains challenging. Existing methods rely on proxies such as layer depth, attention statistics, or representation change. These proxies do not measure how perturbations at each layer propagate to the output and may therefore cause sensitive layers to be underallocated while tolerant layers are overallocated. To address this issue, we propose RippleKV, which allocates cache across layers by estimating how perturbations to each layer's value cache affect the final predictive distribution. RippleKV independently injects norm-adaptive perturbations into each layer's value cache and measures the induced KL divergence at the model output over a small calibration set. Averaging these responses yields a sensitivity profile specific to the model that need not vary monotonically with depth. RippleKV then converts the sensitivity profile into layer budget multipliers by normalizing the sensitivity scores and applying an exponential mapping. A ratio parameter controls the allocation disparity between sensitive and tolerant layers, while a final normalization preserves the KV cache budget. Experiments on LongBench demonstrate that RippleKV achieves the highest average performance among the evaluated KV cache compression methods under matched cache budgets.

Omni-modal LLMs jointly process audio, video, and text, but long multimodal sequences incur substantial prefill and KV-cache costs. Existing omni-modal compression methods primarily focus on pre-LLM token reduction, leaving modality-specific compression across the LLM boundary underexplored. We propose A-PACK, a two-stage framework that defers audio pruning until query-conditioned multimodal interactions emerge. Our analysis shows that audio exhibits higher task-relevant information density and representational diversity per token than video. We further find that local audio-visual dynamics provide a more effective cue for visual selection than token-wise matching. We therefore preserve audio and compress video with local dynamics before the LLM, then progressively prune low-relevance audio and visual tokens and their KV-cache entries inside the LLM. Across four benchmarks on Qwen2.5-Omni-7B/3B, A-PACK achieves the strongest average performance among the evaluated prior methods while reducing prefill FLOPs by up to 78% and improving decoding throughput by up to 2.21x.

Recent advancements in Speech Large Language Models have demonstrated remarkable capabilities in understanding complex audio tasks. Despite this progress, their long-context inference remains severely bottlenecked by prohibitive KV cache memory demands. Existing text-centric compression methods struggle here, often disrupting speech continuity or discarding crucial semantic cues. To address this, we propose VoxZip, a train-free, two-stage semantic-anchored KV cache compression framework. The first stage uses automatic speech recognition (ASR) transcriptions as explicit semantic anchors to temporally align, compress, and fuse audio tokens, significantly reducing the initial KV cache while elevating token information density. To further improve the compression ratio, the second stage employs a dynamic filtering strategy based on temporally decayed accumulated attention to evict non-essential tokens while mitigating early-token bias. Comprehensive evaluations on Qwen3-Omni across six diverse audio benchmarks demonstrate the superiority of our approach. VoxZip excels in long-audio reasoning and consistently maintains high-fidelity perception on short-form tasks. Notably, it sustains over 90% of the uncompressed baseline performance even under an aggressive 20x KV cache compression in long-context scenarios. Furthermore, at a 4x compression ratio, VoxZip yields a 1.9x increase in inference throughput alongside a 3.3x reduction in peak memory overhead. Code and models will be available at https://github.com/MM-Speech/VoxZip.

Transformer-based large language models (LLMs) achieve strong performance across many tasks, but their Key-Value (KV) cache grows linearly with sequence length, creating a severe memory bottleneck for long-context inference. Existing heuristic eviction methods (e.g., H$_2$O and SnapKV) rely on static attention or positional signals that often fail to capture a token's future predictive influence. We propose DistillCache, a reinforcement learning framework that formulates KV-cache eviction as a sequential decision problem. DistillCache learns a lightweight policy network using rich internal model signals (attention statistics, value norms, entropy, and position) and trains it with REINFORCE via a per-step KL-divergence reward to preserve the full-cache output distribution. On a 7B-parameter instruction-tuned Transformer (Mistral-7B-Instruct-v0.3), DistillCache retains 94.2% of full-cache accuracy on LongBench at a 25% cache budget, outperforming both strong heuristic baselines (H$_2$O, SnapKV) by up to 2.7 absolute points and, under our re-implementations, concurrent RL-based methods (ForesightKV, RLKV) by up to 1.4 points on long-context tasks. On reasoning benchmarks, DistillCache is competitive with the best concurrent method and surpasses it under aggressive compression. It also delivers up to 2.1x full-cache throughput while maintaining competitive practical efficiency. These results highlight the effectiveness of learned, distribution-aware policies for memory-efficient long-context LLM inference.

Large Language Models are deployed to multiple types of environments, from internet browsers to edge devices, and WebGPU serves as a modern cross-platform standard. The engines for browser-based LLM inference have proliferated, yet the overhead of WebGPU per-operation dispatch remains poorly characterized. In this work, we introduce a sequential-dispatch measurement method and show that naive single-operation measurements overestimate per-dispatch cost by conflating dispatch with synchronization. Using our method, we measure the per-dispatch cost and show that it is independent of data type used. We show that the dispatch overhead, not kernel quality, is the bottleneck at batch size 1, and isolate the dispatch count as the cause. Therefore, we conclude that at batch size 1, the effective approach to LLM inference optimization in WebGPU is reducing dispatch count. Our findings point to dispatch amortization, in the inference engines and in the WebGPU specification, as a path to practical browser-based inference.

Current LLM safety guardrails face a fundamental tension: fine-tuning distorts pre-trained representations while generative judges incur prohibitive inference costs. We challenge the prevailing paradigm by asking: can safety be achieved through pure geometric reasoning over frozen semantic representations? We present HoloAegis, a minimally parametric topological inference framework that decouples representation from reasoning. We term our approach minimally parametric because the only free parameters are the anchor count K and the temperature tau, both fixed after construction and requiring no gradient-based training. An un-fine-tuned encoder maps text to a unit sphere, after which all decisions are purely geometric. We formalize safety evaluation as a Gibbs-Boltzmann Free Energy computation over a pre-computed System Topology Anchor Bank, and we introduce Dual Time-Scale Exponential Moving Averages to detect progressive multi-turn semantic drift. Our key theoretical insight is a Topological Boundary Stability Conjecture: we provide theoretical motivation and strong empirical evidence that sparse anchor centroids stabilize the decision boundary against high-frequency lexical perturbations far better than full vector space methods. Evaluated across 8 benchmarks, HoloAegis achieves state-of-the-art accuracy (1.0000 AUC on AuthenHallu, 0.9802 on HarmBench) with sub-millisecond latency, zero cold-start data, and cross-lingual transfer (0.9758 AUC on Chinese CHIFRAUD).

Establishing the provenance of a language model---including its base checkpoint and possible overlap in training distributions---is a governance challenge that metadata alone cannot resolve. We introduce a training-free fingerprint based on the top-$k$ vocabulary projections of late hidden states elicited by 250 fixed knowledge probes, compared using Jaccard overlap over decoded token strings. We evaluate the method on 32 open-weight models from nine families (0.6B--32B) with documented relationships. (1)~A similarity ladder broadly follows model relatedness: independently trained models on identical data score 0.48 raw (0.35 vocabulary-corrected), followed by shared-base fine-tunes (0.39/0.33), same-developer relatives (0.38/0.28), and models with no documented relationship (0.22/0.17). This identical-data signal persists across three organizations, two tokenizer families, and two architecture classes, and emerges within the first 1% of training before measurable task competence, suggesting a contribution from shared training data beyond capability convergence. (2)~As a nearest-neighbor lineage-retrieval method, the fingerprint ranks the exact documented base among the top two candidates for all five R1 distillations (mean rank 1.8, MRR 0.60), including a math-specialized base not identifiable from coarse metadata. (3)~A depth ablation shows that lineage group discrimination strengthens toward the output distribution, with AUC increasing from 0.72 at quarter depth to 0.90 at the output; using only the top 5 output tokens retains AUC 0.87. (4)~The fingerprint remains stable under quantization, with Jaccard similarity of 0.92 under int8 and 0.82--0.85 under int4, compared with a maximum cross-model similarity of 0.81 in the calibration pool. We release the probes, code, and fingerprints.

Post-training quantization reduces the deployment cost of large language models, yet how severely a quantized model degrades is not determined by bit-width alone. We systematically study weight-only post-training quantization across bit-widths, quantization methods, model scales and downstream tasks on multiple model families. We observe that such degradation varies substantially across these factors: 4-bit quantization usually preserves performance, 2-bit often causes broad degradation, and at 3-bit, degradation becomes apparent but varies markedly with task type, quantization method and model scale. To explain this variability, we use the signal-to-noise ratio (SNR) to measure how strongly quantization perturbs full-precision representations. We trace degradation back to two linked processes: how quantization errors arise within individual modules, and how they accumulate across layers. First, a source SNR decomposition shows that newly introduced errors depend on three factors: the magnitude of the weight error, the strength of the task-specific signal, and how strongly the quantization error aligns with task-specific activations. Different factors affect these components in distinct ways. Second, a cross-layer propagation analysis shows that these errors can be attenuated, preserved, or amplified as they pass across layers, and that larger models benefit from weaker error amplification. Together, these results establish that quantization degradation is governed by how errors are introduced at the source and how they accumulate across the network.

Mixture-of-Experts (MoE) models have outgrown accelerator memory, and offloading expert weights to host memory is now standard. This makes expert cache management an attractive lever: a policy that raised the hit rate would cut expert traffic per token. Evaluating that is a measurement problem, and we find the measurement fragile. With a trace-driven, event-atomic simulator over three MoE models (40, 64, 128 experts), we isolate three evaluation axes that change conclusions, not just numbers. Replay semantics: under a fused-event traffic contract, an inconsistent per-access replay inflates recency-based policies by 27-29% while leaving frequency-based and static ones within 4%, inverting the policy ranking. Workload contamination: probe sets using one instruction template per category produce verbatim-identical generation prefixes; a matched-pair rendering intervention moves the measured early-window effect by 19.4-31.9 points and reverses which workloads look most cache-friendly. Operating regimes: normalized miss fractions do not transfer across models, so the per-step expert union relative to per-layer capacity must be reported — yet permuting only the temporal order of an identical event stream moves the offline-optimal gap from 44.9% to 30.8%, so it is not sufficient. Corrected, a stable gap to the offline optimum remains (44.2-45.9% over 13 frozen workload compositions). A forced-admission oracle attributes 84.3-96.6% of it to knowing which resident expert is used furthest in the future. A causal next-use predictor, used as an eviction rule, recovers -11.4% of the gap; it picks an optimal victim 3.4% of the time, against 2.4% for a random resident block and 20.6-22.1% for LRU and LFRU. Our position is narrow: in our evaluated settings a large offline-optimal gap substantially overstates the gains recovered by representative lightweight causal mechanisms.

Agentic retrieval-augmented generation (RAG) systems combine preprocessing, embedding, retrieval, memory access, context construction, generation, and vector-index updates. Although LLM decoding is GPU-bound, the surrounding orchestration layer can still limit end-to-end performance through serialization overhead, fragmented scheduling, inefficient batching, and CPU--GPU pipeline stalls. Existing frameworks provide flexible control flow, while distributed runtimes provide scalable task parallelism, but neither exposes RAG stages as resource-aware operators with deterministic execution semantics. We present OpRAG, a resource-deterministic distributed runtime for GPU-backed multi-stage RAG workflows. OpRAG models embedding, retrieval, reasoning, memory, and upsert as first-class operators and lowers them into communication-aware execution graphs. It combines an Arrow zero-copy data plane, persistent workers, bounded queues, CPU tokenizer prefetching, batched GPU embedding, and overlapped retrieval/generation execution to reduce non-model overhead around LLM inference. We evaluate OpRAG using Llama3-8B and Mistral-7B with FlashAttention~2, BF16 execution, and 32K RAG chunks. In end-to-end GPU pipeline experiments, OpRAG improves over the nearest competitor by 16.16% for Llama3-8B and 15.66% for Mistral-7B, and over RayScalableRAG by 20.57% and 20.71%, respectively. Against LangChain, LangGraph, CrewAI, and AutoGen, OpRAG is 17.77% and 17.48% faster than the best framework baseline. In Higress-style query serving, OpRAG reduces hybrid retrieval latency by 59.20--59.62% and generation-scenario latency by 52.48--53.55%, while preserving 100% Recall@5. These results show that optimizing the distributed orchestration layer can substantially improve GPU-backed multi-stage RAG without modifying the LLM decoding kernel.

Large mixture-of-experts (MoE) language models with 26--120 billion parameters exceed the memory capacity of consumer devices through three simultaneous pressures: resident weight matrices, key-value (KV) cache state that grows linearly with context, and dozens of expert sublayers that must be paged on demand. We present RotaryQuant, a three-axis compression system that addresses all three. Mixed-precision weight quantization assigns bit-widths by architectural role: 4-bit for dense layers, 2-bit for routed experts, and 8-bit for the shared expert whose high activation kurtosis resists aggressive compression. LRU expert offloading pages non-resident experts to disk under genuine memory pressure. The novel axis is IsoQuant, a KV cache compression method that applies a Walsh--Hadamard transform followed by block-diagonal SO(4) rotations to isotropize activation distributions before 3-bit scalar quantization, requiring $O(d \log d)$ operations and 256 stored parameters per head versus $O(d^2)$ and 16{,}384 for dense rotation methods. A fused four-kernel Metal GPU pipeline performs attention directly on packed 3-bit tensors without materializing full-precision KV state---a different execution model, not just a quantization scheme. The combined system fits Gemma 4-26B-A4B and Qwen3-30B-A3B within a 16\,GB budget and Nemotron-H 120B within 32\,GB, running interactively at 9--19 tok/s with near-zero perplexity degradation ($Δ$PPL $\leq +0.0012$) and 100% retrieval accuracy at 32K context.

Large language models (LLMs) increasingly read long inputs in the agentic era, from whole documents and codebases to conversations across many turns. Their inference memory is then dominated by the key-value (KV) cache, the stored attention keys and values of every token the model has read and generated. Because the cache grows with context length and is re-read in full at every generated token, a longer context means more GPU memory. To reduce this cost, most existing methods compress the KV cache by lowering every stored value to the same low precision, a technique known as quantization. They can push this to nearly two bits per value, but rarely further, because quality drops sharply at this 2-bit cliff: four levels are too few for the cache's outlier-heavy values, where a few large entries consume the levels and collapse the rest into noise. A natural remedy is to spend more bits on the channels (feature dimensions) that matter and fewer on the rest, but the raw cache offers no handle: its channels are strongly correlated, so none stands out as more important. Our analysis shows that this handle appears once the cache is rotated into a coordinate system computed from its own statistics, removing these correlations. There, a small fraction of channels carries almost all the information, and spending the budget on those few is far more accurate than spreading it evenly. Guided by this analysis, we develop SPECTRA, a training-free, drop-in codec that re-encodes the cache into this coordinate system and concentrates the bit budget on the channels that carry the signal. On Llama-3.1-8B and Qwen2.5-7B over long-context benchmarks, SPECTRA is near-lossless at 4x compression, competitive at 8x where uniform quantization has collapsed, and reaches up to 12x, pushing usable compression past the 2-bit cliff so the same GPU holds longer contexts and larger batches.

Large language model (LLM) inference serving is increasingly constrained by memory rather than compute. As long-context and long-form reasoning workloads become more prevalent, the key-value (KV) cache dominates both memory footprint and memory traffic during LLM token generation, i.e., decode. In particular, HBM capacity has become a scarce and costly resource that heavily limits inference batch size and system throughput. This paper presents OasisKV, a memory-centric LLM inference system design that alleviates HBM capacity pressure by decoupling full KV-cache storage from HBM during LLM decoding. Because decode-time attention is naturally sparse, OasisKV keeps only the KV entries of the most relevant tokens in HBMs for attention computation. We observe that future important tokens can be predicted accurately in advance using lookahead tokens drafted by speculative decoding (SD). OasisKV employs an efficient attention background pipeline to identify important KV blocks. They are then prefetched from higher-capacity memory tiers (e.g., host or remote memory) and staged in HBMs before being used in the next decode step. We implement OasisKV based on vLLM. The lookahead prediction is accurate enough to keep accuracy within 0.7 points of full attention under a 2,048-token KV budget. This lets OasisKV turn sparsity into throughput gain: $1.69\times$ over dense vLLM on the reasoning workload at 0.1 points of accuracy loss, and up to $2.1\times$ on multi-GPU long-context serving. Under prefill--decode disaggregation, OasisKV reaches about $2\times$ dense throughput while admitting each request with $6.5$--$9.7\times$ less KV and holding $2.2$-$2.6$ less decode-node host memory than full KV transfer.

Diffusion language models (DLMs) iteratively refine a sequence, allowing earlier predictions to be revised as context evolves. This rollback capability distinguishes them from irreversible autoregressive generation, but makes inference costly. Every denoising update alters the global context, forcing both prompt and response states to be recomputed even though only response tokens are revisable. Key-value (KV) caching could reduce this cost, yet conventional caching assumes immutable historical states and is therefore difficult to reconcile with rollback. In this paper, we introduce Adaptive Reuse of Cached Hidden States for Efficient Rollback (Archer), a training-free KV caching method for rollback-capable DLMs. Archer asymmetrically keeps the mutable response synchronized with the current hypothesis while reusing prompt K/V within a bounded state neighborhood. Although prompt representations also change under bidirectional attention, their token identities remain fixed; bounded reuse therefore amortizes repeated prompt computation without caching mutable response states. It also delays feedback from tentative tokens, reducing premature reinforcement of transient high-confidence errors and giving rollback more opportunity to correct them. Our analysis characterizes prompt reuse as a reversibility-aligned cache boundary, bounds its state-dependent approximation error, and gives a decoder-margin condition for preserving full-refresh decisions. Existing DLM acceleration often trades quality for speed. Archer shifts this frontier, attaining the best mean performance of 33.63% together with a 2.57x mean speedup on the main suite. Across evaluated settings, it improves Pass@1 by up to 3.05 points and reaches up to 2.95x speedup. Controlled analyses connect the quality gain to delayed prompt feedback and validate state-aware refresh. Our code is available at https://github.com/Hxnng/Archer.

Graph Neural Networks (GNNs) are widely used across domains such as natural sciences, social network analysis, chip design, and recommendation systems. However, as graph sizes grow, storing and processing them entirely on a single-node CPU-GPU system becomes increasingly impractical. A promising approach is to distribute the graph across multiple remote memory nodes, though this introduces a major bottleneck: inter-node network congestion during training. To address this, we propose LGNNIC, a novel inter-node system architecture that leverages SmartNICs co-located with remote memory nodes-a configuration already available in modern systems-to reduce communication overhead in distributed GNN training. LGNNIC offloads key preprocessing tasks to SmartNICs, reducing the volume of data transferred to computational (training) nodes and alleviating network congestion. We introduce two complementary techniques executed on the SmartNICs during the preprocessing phase: Neighbor Sampling, which performs mini-batch sampling, and Quantization of the sampled batches. To evaluate LGNNIC under different communication infrastructures, we designed both an optimized low-overhead DMA-based synchronization mechanism and a high-overhead socket-based alternative used as a benchmark. We evaluate the core SmartNIC offloading mechanisms across standard GNN workloads and sampling hyperparameters using a proof-of-concept (PoC) system comprising one remote-memory node with an NVIDIA BlueField-2 SmartNIC and one compute node with an A100 GPU. Both Neighbor Sampling and Quantization on the remote node demonstrated substantial training speedups in most configurations. Neighbor Sampling achieved up to 62.4x and 17.5x speedups with Sockets and DOCA-DMA, respectively, primarily due to reduced data transaction time. Quantization provided additional speedups of up to 3.6x and 1.3x, respectively, by reducing data transfer.

Long-term memory enables language agents to reuse past facts, preferences, and task experience. Persistence also creates a central falsifiability problem: when the world changes, stale memories can remain retrievable and pollute the prompt. We characterize this failure mode as memory pollution: degradation caused by active memories that newer conflicting evidence has superseded. We introduce TEPA, a revocable evidence-memory mechanism that makes validity an explicit state of memory. TEPA represents observations as keyed precedents and revokes active precedents when fresh evidence contradicts them under the same key, allowing retrieval to draw from current evidence while preserving revoked history for audit. Across controlled hidden-regime drift, real file-backed executable drift, and preference-update streams, revocation prevents stale active memory from remaining in the retrieval set after reversal. In controlled drift over 50 seeds, append-only and last-write-wins memory fell below no memory during full reversal (append-only and last-write-wins both 0.210, no memory 0.309, TEPA 0.950), and the same pattern reproduced under real file execution (append-only 0.203, no memory 0.298, TEPA 0.950). On clean MemoryAgentBench SH-6k, TEPA matches a strong last-write-wins cache, confirming that current-key replacement is the decisive operation for single-hop fact consolidation. Boundary tests on multi-hop and very long-context MemoryAgentBench settings expose retrieval-chain and context-selection bottlenecks beyond fact-level validity tracking. Together, these results establish lifecycle revocation as a core memory operation for agents that must falsify, audit, and later re-promote evolving knowledge.

We study visual persistence in interactive video world models. These models rely on a Key-Value (KV) cache as a growing visual memory to carry forward previously generated frames. However, we find that models can no longer reliably address stored content once rollouts extend beyond the training horizon, because temporal Rotary Positional Embeddings (RoPE) offsets then fall outside the range seen during training and the model struggles to retrieve the relevant visual information through attention. Moreover, naively compressing the cache in the RoPE-rotated space corrupts memory by averaging together incompatible positional phases. To address this, we propose WorldTrace, a training-free memory framework for long-horizon visual persistence. WorldTrace keeps compressed memory addressable by assigning each summary slot a distinct, in-distribution virtual position. Within this addressable cache, we study two memory compression approaches: WorldTrace-Field compresses history for temporal coherence, while WorldTrace-Landmark stores verbatim scene traces at detected transitions for episodic recall. We further introduce LoopBench, a benchmark evaluating whether a compressed cache can reconstruct a previously visited scene after a long detour. WorldTrace-Field improves temporal consistency by +15.5%, and WorldTrace-Landmark improves episodic recall by +19.5% on LoopBench, extending visually persistent generation without retraining.

Recent optimization studies on Retrieval-Augmented Generation (RAG) have exploited chunk-level KV cache reuse to avoid processing long retrieved contexts for higher efficiency, while significant information redundancy and noise still remain in the coarse-grained chunks. This paper optimizes the Pareto frontier under low prefill latency constraints while maximizing accuracy by proposing CoinRAG (Contextualized Information Nugget KV Cache Reuse for Long-Context RAG). The name metaphorically reflects our core mechanism: much like assembling small tokens (or "coins") to accumulate a larger value, CoinRAG compositionally reuses offline-computed, fine-grained nugget caches to form a learned contextual representation efficiently in a more semantically relevant but compact manner. Specifically, instead of full-chunk encoding, CoinRAG identifies query-relevant semantic units within retrieved chunks through two-stage retrieval and seamlessly assembles their sliced KV representations with a chunk-level context. Extensive evaluations on LongBench multi-hop question answering tasks demonstrate that CoinRAG significantly reduces operational costs and outperforms the other baselines with a new Pareto frontier and an average 5.3% relative improvement in answer quality (F1) under a standard fast prefill latency budget.

This paper proposes a method and its implementation for emulating FP64 matrix multiplication (DGEMM) by constructing, on FP4 (E2M1; 2 exponent bits and 1 mantissa bit) Tensor Cores, Ozaki schemes I and II, which realize high-precision matrix multiplication on low-precision arithmetic units. Prior implementations were based on INT8 and FP8, and the use of the faster FP4 had not been realized. The key property is that every FP4 value becomes an integer when doubled, and that shifting this integer set by multiples of 13 covers all integers. Converting an arbitrary integer into base-13 FP4 limbs by this property keeps intermediate sums error-free in FP32 accumulators, which makes FP4 Tensor Cores usable for Ozaki schemes I and II. By the same principle, the integer GEMM of INT8 Tensor Cores can also be emulated bit-exactly on FP4 Tensor Cores. When FP4 Tensor Cores have twice the throughput of FP8, Ozaki scheme II on FP4 theoretically achieves slightly higher performance than its FP8 counterpart. This paper further proposes kernel implementation optimizations raising the attained fraction of peak performance, obtaining a measured speedup on top of the theoretical advantage. We verify this on an RTX PRO 6000 Blackwell, achieving performance competitive with that of an existing FP8-based implementation of Ozaki scheme II, and actually exceeding it at a large problem size ($16384^3$).

Layout-guided multi-instance generation is essential for controllable image synthesis in Multi-Modal Diffusion Transformers (MM-DiTs). However, integrating this capability into unified architectures remains challenging. Prior frameworks rely on redundant full-resolution canvas padding and Shifted-RoPE to manage multiple reference images. This mechanism drastically inflates computational overhead for sparse layouts and disrupts critical low-frequency RoPE features, creating a severe spatial-frequency compromise that blurs absolute spatial correspondence. To overcome these limitations, we propose ControlRef, a highly efficient and precise multi-instance synthesis framework. ControlRef utilizes a Unified Instance-Layout Control (UILC) attention mask to strictly decouple inter-instance semantic interactions and enforce precise regional binding. To further promote region-level spatial alignment, we introduce Anchored 4D-RoPE, a novel positional encoding mechanism that directly anchors tokens to their absolute geometric centers. By pre-aligning reference images to their corresponding bounding box resolutions, physically anchoring both layout and reference tokens to their absolute geometric centers, and stacking the references along the z-axis, Anchored 4D-RoPE natively preserves spatial priors and mitigates the spatial-frequency compromise without lossy shifting. Extensive experiments demonstrate that ControlRef achieves state-of-the-art visual fidelity and localization accuracy, while concurrently slashing inference latency by over 80% in sparse layouts and reducing memory overhead by 50% in dense scenarios.

Vision-language models (VLMs) have achieved remarkable generalization across diverse multimodal tasks through large-scale pre-training, yet their rapidly increasing computational and memory requirements pose significant challenges for deployment in constrained environments. Existing pruning strategies often depend on task-specific criteria or LLM-oriented importance measures, making them unsuitable for task-agnostic pruning, where no task-specific samples are available at pruning time and the pruned model remains broadly applicable. We introduce a retraining-free VLM pruning framework called PORTA that derives a task- and modality-agnostic importance formulation based on activation variation, estimated from generic calibration data, which reliably captures feature-level representation utility across modalities. PORTA further incorporates an adaptive sparsity allocation mechanism that assigns layer-wise pruning ratios based on output feature variability, avoiding the limitations of uniform sparsity and reducing performance degradation at high compression levels. Extensive experiments across VLM architectures, such as CLIP, BLIP, and Qwen2-VL, demonstrate that PORTA achieves competitive downstream performance under high sparsity without requiring any retraining, supporting efficient VLM compression. Code is available at https://github.com/cau-hai-lab/PORTA.git.

Real-time human action recognition on Internet-of-Things (IoT) edge devices requires models that capture rich spatio-temporal cues within strict latency, memory, and power envelopes. Current 3D CNNs, video transformers, and shift-based ViT deliver high accuracy but come at computational costs that preclude edge IoT deployment. This paper proposes CoDAT, a Collaborative Dual-Attention Transformer that replaces conventional multi-head attention with a lightweight dual-branch module: Spatial Convolutional Attention (SCA) for local aggregation and Strided Single-Head Attention (SSHA) for global context. SSHA jointly compresses the spatial resolution and channel dimensions of the query, key, and value tensors via stride-based sparse projection, then fuses the resulting global and local features at a markedly reduced cost. To enable temporal communication across frames, a parameter-free TShift module is embedded in each block. Extensive experiments on Jetson AGX Orin and Raspberry Pi 5 demonstrate that CoDAT achieves an energy-accuracy balance in both image and action recognition. On ImageNet-1K, CoDAT-M runs 2x faster than EfficientViT384 and FastViT-S12 at comparable accuracy, and CoDAT-L matches ViT-S with 3x fewer parameters at 2x higher throughput. On Kinetics-400 and MA-52, CoDAT achieves competitive Top-1 accuracy against state-of-the-art CNN, transformer, and hybrid baselines while running up to 2.9x faster than VSwin-T, 2x faster than ViT-Temporal-Shift variants, and 5x faster than UniFormer-B. On UCF-101, CoDAT-S384 matches TokShift and LAPS while being 6x faster and requiring up to 13x fewer FLOPs, establishing an efficiency-accuracy balance for real-time action recognition in edge IoT perception systems. Code is available at https://github.com/novendrastywn/CoDAT .

The top-$k$ operation is a fundamental building block of modern sparse computation, enabling token routing, expert activation, memory selection, and attention pruning. Yet standard hard top-$k$ blocks gradients, while existing continuous (soft) relaxations remain too costly for large-scale models. We introduce Fast LapSum, an exact-budget soft top-$k$ primitive whose GPU solver runs in linear time after sorting. Unlike prior linear-time methods such as DFTopK, which relax the normalization constraint, Fast LapSum is, to our knowledge, the first method to preserve an exact selection mass of $k$ while remaining fully differentiable end-to-end. Our solver combines a linear-time threshold computation with an analytical vector--Jacobian product, and for extreme scales employs probabilistic bracketing to sort only the uncertain middle band of kernel-noised scores. The resulting overhead is almost negligible: the solver processes $10^6$, $10^7$, and $10^8$ scores in $0.41$, $1.15$, and $5.23$\,ms, respectively. This makes exact soft top-$k$ practical for sparse routing, retrieval, and large-scale optimization. We demonstrate Fast LapSum on two demanding applications operating over millions of coordinates inside the training loop: generating megapixel sparse adversarial examples with an exact soft budget of ${\sim}0.02%$ of an image's pixels, achieving an order-of-magnitude speedup over state-of-the-art methods, and training a fully differentiable sparse image coder from scratch.

Post-training quantization (PTQ) is widely used to reduce the memory and computational cost of large language models. Existing PTQ methods typically obtain an initial quantized model through heuristic rules or greedy optimization, and once quantization is completed the resulting integer assignments are usually treated as final. This observation motivates a complementary optimization stage within PTQ that keeps quantized weights improvable after an executable quantized model has been produced, while preserving the quantized format. We introduce ReQuant, a backpropagation-free fixed-grid refinement procedure for this stage. Agnostic to the PTQ initializer, ReQuant takes an existing quantized model as a feasible starting point and iteratively revisits its discrete weight assignments on the fixed quantization grid. Accepted updates strictly reduce the mean squared reconstruction error and remain on the original grid. In this way, ReQuant turns the initially fixed PTQ output into an iteratively optimizable discrete solution and serves as a plug-and-play post-processing stage for existing PTQ pipelines. Experiments across diverse model families, bit-widths, and downstream tasks show that ReQuant consistently improves quantized models from heterogeneous PTQ initializers, with especially large gains on simple initializers and lower bit-widths. Notably, ReQuant can refine a simple round-to-nearest initialization across multiple sweeps until it approaches or surpasses GPTAQ under the same quantization format. These results establish ReQuant as a practical complementary stage for further improving existing PTQ pipelines.

Heterogeneous architectures that combine neural processing unit (NPU) and processing-in-memory (PIM) are increasingly adopted to accelerate LLM inference. Prior work focuses on building a unified memory that allows NPUs and PIM to share data without duplication. However, these designs implicitly assume that each tensor is bound to a fixed execution device, and therefore rely on static, device-biased data mappings. We observe that this assumption does not hold in modern LLM workloads. Due to phase changes (e.g., prefill vs. decode) and dynamic behaviors such as MoE routing, the optimal execution device for the same tensor can change at runtime. Under such dynamic execution, device-biased mappings become mismatched to access patterns, leading to substantial bandwidth underutilization and performance loss. This paper presents PFM (PIM-as-Flexible-Memory), a dual-view memory system that decouples physical data layout from accessor-visible logical views. PFM stores data in a jointly optimized physical layout and exposes different logical interpretations to NPUs and PIM, enabling efficient access across devices without data duplication or relayout. We further design accessor-aware address translation and runtime scheduling mechanisms to support dynamic execution when LLM workloads fluctuate and the optimal execution device dynamically changes. Our evaluation across LLMs shows that PFM improves end-to-end throughput by up to 2.32$\times$, demonstrating its effectiveness and broad applicability as a unified memory management solution for NPU-PIM systems.

The rapid scaling of Large Language Models (LLMs) has significantly increased computational cost, energy consumption, and inference latency, making accurate estimation essential for sustainable artificial intelligence deployment and hardware-aware design. In this work, we introduce Hybrid Modeling for Energy and Latency of LLMs (HYMELL), a hybrid three-level framework for estimating LLM inference latency and energy by combining analytical modeling with machine learning (ML). HYMELL models LLM execution through a three-level hierarchy: analytical estimation of primitive operations, ML prediction of higher-level components, and an end-to-end model that captures system-level overheads across both prefill and decode phases. The framework supports diverse architectures, including dense and mixture-of-experts (MoE) feed-forward networks (FFNs), as well as multi-head attention (MHA) and grouped-query attention (GQA) mechanisms. Evaluated on an NVIDIA H100 graphics processing unit (GPU), HYMELL achieves high predictive accuracy; notably, for LLaMA 3 8B, it attains less than 5% error for both prefill and decode phases. By predicting execution costs directly from architectural parameters, it enables fast, hardware-free design space exploration and energy-efficient optimization.

Quantized Neural Networks~(QNN) with low-bitwidth data have proven promising in efficient storage and computation on edge devices. To mitigate accuracy degradation while maximizing speedup, layer-wise mixed-precision quantization~(MPQ) becomes a popular solution. However, existing algorithms for exploring MPQ schemes are limited in flexibility and efficiency. Comprehending the complex impacts of different MPQ schemes on post-training quantization and quantization-aware training results is a challenge for conventional methods. Furthermore, an end-to-end framework for the optimization and deployment of MPQ models is missing in existing work. To address these challenges, we propose the MiCo framework, a holistic MPQ exploration and deployment framework for edge AI applications. The framework adopts a novel optimization algorithm to search for accuracy-optimal quantization configurations under strict latency constraints. We further extended the framework to MiCoPro, which introduces a robust Hardware-Aware Proxy (HAP) model to enhance prediction accuracy and hardware versatility. By leveraging target-specific latency modeling, MiCoPro enables rapid exploration and direct deployment from PyTorch models to bare-metal C code. We demonstrate the versatility of our framework on both the BitFusion accelerator and SIMD-extended RISC-V processors, achieving up to 40% of latency reduction with less than 3% of accuracy drop.

Multimodal large language models (MLLMs) encode images as long visual token sequences, making prefilling and KV-cache storage expensive. Existing training-free pruning methods select tokens by importance, diversity, or spatial coverage, but treat retained tokens as interchangeable and do not explicitly track which object-related regions are already covered. We present RoRA, a training-free framework that casts visual token pruning as role-oriented regional evidence allocation. Given a fixed budget, RoRA partitions tokens into a protected semantic core, complementary context, and fine-grained detail. It first calibrates text-conditioned attention with a positional prior and a prompt-calibrated object prior, then builds Attention-Anchored Regions (AARs) from high-confidence anchors as lightweight proxies for covered object support. Context is explored mainly outside AARs, while a small AAR-guided budget restores local detail; pairwise similarity is used only for context-stage redundancy filtering. Under matched budgets, RoRA consistently outperforms strong training-free baselines across LLaVA and Qwen-VL families, retaining most of the unpruned accuracy even at aggressive pruning ratios, e.g., 96.5% of full performance at 88.9% pruning on LLaVA-1.5, and improving over D2Pruner by about 5% on Qwen3-VL at 75-90% pruning. At a 66.7% pruning ratio, RoRA requires only 0.7 ms for token selection and reduces end-to-end inference time by 24.6%, corresponding to a 1.33x speedup over unpruned inference on an NVIDIA H800.

Weight quantization for large-language-model inference must balance adaptive reconstruction levels with representations regular enough for efficient GPU execution. Uniform integers constrain each group to a linear grid. Low-bit floating-point formats use a fixed exponent-mantissa structure, while learned codebooks gain flexibility at the cost of irregular decoding and additional metadata. We introduce CubicQuant, a parametric non-uniform scalar format that preserves a dense integer code stream while adapting reconstruction levels within each weight group. A monotonic cubic curve, specified by two shape parameters and one scale, maps uniformly spaced magnitude codes to non-uniform levels. The family spans 1-8-bit weight payloads, contains symmetric uniform integer quantization as an exact special case, and has effective width B + 64/G bits per weight for payload width B and group size G. We derive population distortion under Uniform, Gaussian, and Laplace distributions, formulate continuous and Dynamic-A8-carrier-aware fitting objectives, and describe direct packed-weight GPU execution. For finite groups of G=128 with 15,360 samples per distribution, W4 CubicQuant reduced reconstruction RMSE relative to optimally clipped four-bit uniform integer quantization by 3.90% on Uniform, 13.49% on Gaussian, and 28.14% on Laplace samples. Relative to the best enumerated four-bit finite floating-point format, the reductions were 3.90%, 9.44%, and 6.27%. Preliminary H200 kernel measurements show a workload-dependent crossover: model-dtype execution is faster for narrow GEMV, while Dynamic A8 becomes favorable as row count grows. The results establish the format's representational promise and direct executability; downstream model quality and cross-device end-to-end performance remain open evaluation questions.