生成模型与LLM推理优化

Recently, residual reconstruction-based model quantization methods have achieved promising performance in low-bit post-training quantization (PTQ) by introducing cross-layer residuals to reduce error accumulated from previous layers.However, these residuals may also introduce additional bias arising from the Hessian-approximation (HA) assumption underlying reconstruction-based PTQ, leading to suboptimal quantization performance.In this work, we analyze that multiplying the residual term by a scaling coefficient provides a direct way to mitigate the HA bias associated with residual strength, while preserving accumulated-error correction. More importantly, we observe that this trade-off is module-dependent, making a single global residual strength insufficient to balance effective correction and residual-related bias across modules.Based on these observations, we propose Module-Adaptive Residual Reconstruction (MARR), which assigns a module-specific scaling coefficient to adaptively balance accumulated-error correction and residual-related HA bias for each module.To avoid expensive per-module coefficient search and obtain a stable coefficient estimate, we design a Proportional-Integral-Derivative (PID)-based adaptive update strategy that uses reconstruction error as feedback to progressively refine this coefficient. Experiments on several typical large language models (LLMs) and vision transformers (ViTs) demonstrate the effectiveness of MARR under low-bit quantization (less than or equal to 4-bit), achieving up to 20.2% performance gains on LLMs and up to 4.6% relative gains on ViTs over the residual reconstruction state-of-the-art methods.Code will be made publicly available upon acceptance.

Supporting long-context LLMs is challenging due to the substantial memory demands of the key-value (KV) cache. Existing offloading systems store the full cache in host memory and selectively fetch critical entries during decoding, but this strategy quickly hits a ceiling: sparsity cannot be pushed further without degrading accuracy. As a result, when context length and batch size grow, the volume of KV transfers rises sharply and becomes the dominant source of decoding latency. We present KVDrive, a holistic multi-tier KV cache management system spanning GPU memory, host DRAM, and SSD. Unlike prior work that pursues greater sparsity through algorithmic refinements, KVDrive tackles the problem from a systems perspective - jointly orchestrating cache placement, pipeline scheduling, and cross-tier coordination to sustain high-throughput inference under tight GPU budgets. KVDrive advances three fundamental capabilities: it adapts cache management to attention behavior to maximize reuse and minimize redundant data movement; it restructures the decoding pipeline to overlap I/O- and CPU/GPU compute-bound stages, eliminating stalls across heterogeneous resources; and it harmonizes data movement across memory tiers to unlock scalable long-context inference far beyond GPU and DRAM limits. We have implemented a fully functional prototype of KVDrive and evaluated it on long-context benchmarks with popular LLMs. The system achieves up to 1.74x higher throughput compared to state-of-the-art works while preserving accuracy.

INT2 KV-cache quantization is attractive for long-context LLM serving, but it remains difficult to make both accurate and deployable. Simple rotations such as Hadamard transforms reduce outliers, but still degrade at INT2 because they are not aligned with downstream attention. We propose OSCAR, an Ultra-low-bit KV Cache quantization method that estimates attention-aware covariance structures offline and uses them to derive fixed rotations and clipping thresholds for quantization. In this way, it aligns KV quantization with the covariance structures that attention actually consumes. More importantly, we not only provide theoretical justification but also develop a fully deployable OSCAR system with a custom INT2 attention kernel that remains compatible with paged KV-cache serving and fused kernel pipelines, enabling seamless integration into modern LLM serving frameworks such as SGLang and vLLM. We evaluate our methods on recent reasoning models with reasoning traces of up to 32k tokens across 5 tasks. On Qwen3-4B-Thinking-2507 and Qwen3-8B, OSCAR reduces the BF16 accuracy gap to 3.78 and 1.42 points, respectively, while naive rotation INT2 collapses to nearly zero. We further scale OSCAR to Qwen3-32B and GLM-4.7 (358B params), where it remains effectively on par with BF16. On long context - RULER-NIAH up to 128K, OSCAR remains robust on both Qwen3 models, while naive rotation INT2 collapses. System-wise, OSCAR reduces KV-cache memory by approximately 8x, improves throughput by up to 7x at large batch sizes under the same memory budget, and accelerates batch-size-1 decoding by up to 3x over BF16 due to reduced memory bandwidth overhead.

Video diffusion models have recently enabled high-quality video generation with ViT-based architectures, but remain computationally intensive because generation requires attention computation over long spatiotemporal sequences. Token pruning has proven effective for ViTs and VLMs. However, most prior pruning methods are attention-based and operate per frame, failing to ensure the vital temporal coherence across frames in video generation tasks. In practice, naively adopting attention-only pruning causes noticeable degradation due to worsened background consistency, flickering, and reduced image quality. To address this, we propose TAPE, a training-free Temporal Aware Pruning for Efficient diffusion-based video generation. TAPE (i) applies temporal smoothing to align token-importance across adjacent frames and suppress selection jitter; and (ii) performs token reselection in selected layers to align token pruning with layers' diverse semantic focus and avoid error accumulation in specific areas; it also (iii) adopt a timestep-level budget scheduling that prunes aggressively at early noisy steps and relaxes pruning during fidelity-critical refinement. The experimental results show that TAPE delivers significant speedups while preserving high visual fidelity, outperforming prior token reduction approaches.

Recent advances in autoregressive video diffusion have enabled sequential and streaming video generation. However, long-horizon generation requires increasingly large KV caches, making efficient compression without sacrificing quality challenging. Existing methods mostly select historical frames based on attention scores, but their context decisions remain coarse. When multiple frames are generated in the same chunk, these methods often apply a shared history selection to the whole chunk, score historical frames solely by attention, and assign head-wise budgets either uniformly or by attention-pattern heuristics rather than explicit head-importance estimation. We show that frames within the same generated chunk can depend on distinct historical frames, that the same historical frame can receive different attention scores as its relative temporal distance to the current frames changes, and that masking different heads induces unequal generation degradation. Motivated by these findings, we propose Focused Forcing, a training-free KV selection method that focuses cached history along both generated-frame and head dimensions. For each generated frame, Focused Forcing preserves the most relevant and distinctive historical frames by combining attention scores with diversity scores of historical frames, while assigning larger budgets to heads with higher estimated importance. Across multiple autoregressive generation paradigms, Focused Forcing achieves up to $1.48\times$ end-to-end acceleration without training, while improving visual quality and text alignment. Our code will be released on GitHub.

In video generation models, particularly world models, training large-scale video diffusion Transformers (such as DiT and MMDiT) poses significant computational challenges due to the extreme variance in sequence lengths within mixed-mode datasets. Existing bucket-based data loading strategies typically rely on "equal token length" constraints. This approach fails to account for the quadratic complexity of self-attention mechanisms, leading to severe load imbalance and underutilization of GPU resources. This paper proposes AdaptiveLoad, an integrated optimization framework consisting of two core components: (1) A dual-constraint adaptive load balancing system, which eliminates long-sequence bottlenecks by simultaneously limiting memory consumption and computational load ($B \times S^p \le M_{\text{comp}}$); (2) A fused LayerNorm-Modulate CUDA kernel, which utilizes a D-tile coalesced reduction strategy to increase throughput and alleviate memory pressure. Experimental results on the Wan 2.1 world model demonstrate that our method reduces the computational imbalance rate from 39% to 18.9%, improves peak VRAM utilization efficiency by 22.7%, and achieves an overall training throughput increase of 27.2%.

Modern interactive video world models have achieved impressive visual fidelity, yet lack fine-grained multi-entity control and cross-entity, cross-world generalization. We trace this gap to the action interface: standard control protocols (e.g. animation IDs, device inputs, scene-level captions) bind action semantics to specific entities or engines at design time. We propose natural language as the interface to unlock expressiveness that no prior interface can achieve, and we present Incantation, the first interactive video world model with per-latent-frame (0.25 s) natural-language conditioning that supports simultaneous multi-entity control and concept-level cross-entity transfer beyond any fixed rendering pipeline. We pair a pretrained bidirectional video backbone with frame-local text cross-attention, and enable real-time long-horizon streaming through ODE-initialized Self-Forcing distillation with a RoPE-decoupled sliding KV-cache. We surpass the Action-Index baseline on cross-entity transfer (89% vs. 43%) and out-of-vocabulary prompts (90% vs. 0%), and our 2-step student sustains 19.7 FPS at 480p with stable FVD over 2-hour rollouts. We further apply the same architecture and training recipe to The King of Fighters, changing only the per-entity action vocabulary slots. We have released a preview subset of the Incantation dataset at https://huggingface.co/datasets/zhush/incantation-elden-ring-scenes, containing manually collected Elden Ring player-boss combat clips with structured action-oriented metadata. Larger-scale Elden Ring and KOF data will be released with the full project.

Unstructured sparsity is now natively accelerated by recent GPU kernels and dataflow hardware, shifting the bottleneck from inference execution to the pruning algorithm. State-of-the-art methods for unstructured LLM pruning are layer-wise surrogates derived from the Optimal Brain Surgeon principle, and they sacrifice end-to-end accuracy, especially under aggressive sparsity. End-to-end alternatives such as MaskLLM and PATCH show that learnable masks can close this gap, but their categorical-over-patterns parameterization scales with the number of valid masks per row and does not port to the unstructured setting. We introduce LEAP, which replaces this intractable parameterization with a per-weight Bernoulli-via-Gumbel- sigmoid relaxation that makes end-to-end unstructured mask learning tractable. Across five LLM families from 0.5B to 8B parameters at 50% and 60% sparsity, LEAP improves six-task average zero-shot accuracy by +2.59 points on average over ADMM, the best layer-wise baseline in our sweep.

Binary quantization (BQ) compresses high-dimensional embeddings into one or two bits per coordinate, enabling nearest neighbor search at extreme speed. Yet a striking puzzle persists: BQ achieves competitive recall on contrastive embeddings but fails on others — and two leading systems adopt diametrically opposite strategies (random rotation vs. preserving coordinate axes) without a common theory explaining when each is appropriate. We resolve this puzzle by connecting the Gaussian structure recently established for InfoNCE-trained representations to a complete analytical framework for BQ quality. The key insight is that coordinate heterogeneity — the non-uniformity of per-coordinate variances — governs the key aspects of BQ performance. We derive closed-form expressions for ranking fidelity, prove that the magnitude bit carries information proportional to heterogeneity, and show that random rotation destroys precisely the signal that one paradigm exploits while creating the isotropy that the other requires. A two-parameter scaling law predicts fidelity across models and dimensions. Experiments on 13 datasets and 6 embedding families validate all predictions and provide the first principled design guide for binary quantization systems.

Vision-Language Models (VLMs) have shown strong promise on Optical Character Recognition (OCR), yet the sheer number of visual tokens required to encode dense documents incurs prohibitive inference cost. Existing pruning methods rely on physical eviction, e.g., permanently discarding visual tokens during the prefill stage. While effective for natural images, this strategy fundamentally breaks down on OCR, where virtually every visual token may correspond to a character or structural element, and any irreversible loss leads to catastrophic accuracy degradation. We observe that, although document images appear globally dense and seemingly unprunable, the model's attention to them is in fact temporally sparse: at each decoding step it concentrates on a small region that shifts gradually across steps, much as a human reader fixates on successive words rather than perceiving an entire page at once. Motivated by this Dynamic Visual Fixation phenomenon, we recast the intractable global pruning problem as a tractable local, dynamic one and propose FastOCR, a training-free framework with two complementary modules. Specifically, Focal-Guided Pruning identifies a small set of focal layers and selects the most task-relevant visual tokens from them at each step, while Cross-Step Fixation Reuse exploits the gradual shift of fixation to warm-start each step from the previous one. By dynamically adjusting which tokens are attended rather than evicting any from the cache, FastOCR avoids permanent information loss. Extensive experiments show that FastOCR serves as a plug-and-play acceleration module, generalizing consistently across five VLMs of varying sizes and architectures. On Qwen2.5-VL, FastOCR retains 98% of the unpruned model's accuracy while attending to only 5% of the visual tokens per decoding step, reducing attention latency by 3.0$\times$.

The large size of the KV cache has become a major bottleneck for serving LLMs with increasing context lengths. In response, many KV cache compression methods, such as token dropping and quantization, have been proposed. However, almost all of these methods are inherently lossy-despite minimal accuracy degradation for short outputs, their outputs increasingly diverge from full-KV-cache outputs as more tokens are decoded, which leads to catastrophic failures in code generation and tool calling. We present VeriCache, the first inference framework that ensures the same output as full-KV-cache decoding but largely preserves the high decoding throughput of a range of KV cache compression algorithms. VeriCache uses the compressed KV cache to draft tokens, then verifies them against the full KV cache. While it may seem like just speculative decoding, VeriCache requires addressing a key system challenge to work-keeping the full KV cache out of GPU memory and minimizing the overhead of swapping it in for verification. The insight is two-fold: (1) compressed-KV decoding can be parallelized with full-KV swap, because one is HBM-bandwidth-bound and the other is PCIe/network-bound, and (2) the compressed KV cache often produces output similar to the full KV cache, allowing a long drafting horizon to amortize each full-KV swap. VeriCache applies to both long-context decoding and remote prefix caching, supports a broad family of token-dropping and quantization methods through a uniform compressor interface, and composes with traditional speculative decoding. Experimental results show that VeriCache achieves up to 4X higher throughput than full-KV inference while producing identical outputs.

Quantization-aware training (QAT) is widely adopted to quantize language models by training full-precision weights using gradients from the quantized model. The main bottleneck is its slow convergence and early performance plateau, particularly below 4-bit-widths. While this problem has been observed in prior work, its precise cause remains unclear. In this paper, we analyze the convergence of QAT by estimating the spectrum of the loss-surface Hessians. We find that the weights converge to flat regions around saddle points, where a large fraction of the Hessian eigenvalues are both positive and negative. During training, an increasing fraction of Hessian eigenvalues concentrates around zero, whose magnitude decreases. At lower bit-widths, the magnitude of eigenvalues in the Hessian spectrum is significantly smaller. To mitigate these issues, we propose an algorithm called WinQ to accelerate QAT, which involves: (1) periodically resetting weights to the linear interpolation of full-precision and quantized weights, reducing the distance to the quantization grid and increasing eigenvalue magnitude, and (2) computing gradients of noise-injected weights to regularize the Hessian. Extensive experiments show that WinQ accelerates QAT by up to 4 times across various quantization methods and models. Under the same training cost, WinQ improves state-of-the-art sub-4-bit quantization by up to 8.8%. These results are consistent across 16 settings with different language models, quantization methods, and bit widths.

Diffusion Transformers (DiTs) achieve state-of-the-art image generation quality but incur substantial memory and computational costs at inference. While aggressive Post-Training Quantization (PTQ) to 4-bit precision offers significant efficiency gains, it typically results in severe quality degradation. Existing approaches, including smoothing-based methods, mixed-precision schemes, rotation techniques, and low-rank residual methods, partially mitigate this issue but still leave a noticeable gap to FP16/BF16 performance. In this work, we introduce DiRotQ, a W4A4 PTQ framework that mitigates this degradation through rotation-aware activation quantization. DiRotQ identifies a low-rank subspace capturing dominant activation variance via Principal Component Analysis (PCA), preserving coefficients in this subspace at higher precision while quantizing the remaining components to 4-bit. Activations are rotated into the PCA basis at inference time using calibration-derived orthogonal transformations, while the inverse rotation is fused into the layer weights offline. Combined with GPTQ-based weight quantization, DiRotQ achieves an FID (lower is better) of 15.9 and PSNR (higher is better) of 19.1 dB on PixArt-Σ over the MJHQ-30K dataset, outperforming the prior state-of-the-art SVDQuant (FID 18.9, PSNR 17.6) under the same INT W4A4 setting. Beyond standard metrics, we introduce a VLM-as-a-Judge evaluation protocol for diffusion model quantization, the first such evaluation in this setting, providing a more holistic assessment of perceptual quality and prompt alignment under aggressive compression. On the systems side, we implement a Triton-based custom kernel to enable efficient end-to-end inference, reducing memory usage of the 12B FLUX.1-dev model by 2.1x and delivering 2.3x speedup over the BF16 baseline, on a 24 GB RTX 4090 GPU.

The classical simulation of quantum algorithms is a crucial tool for circuit development, testing, and validation. Although acceleration using GPUs significantly reduces simulation time, most high-performance simulators rely on vendor-specific frameworks that target data-center hardware. To broaden access to quantum simulation, this work proposes a vendor-agnostic approach targeting the integrated GPUs commonly found in consumer-grade laptops. A primary challenge in state-vector simulation is its inherently poor spatial locality, which creates a memory bandwidth bottleneck. Consequently, baseline implementations experience a severe degradation in relative GPU speedup as the number of simulated qubits increases. To address this limitation, we introduce a state partitioning optimization that reorganizes the quantum state vector to maximize the last-level cache locality and minimize costly main memory fetches. We evaluate this strategy using a Quantum Phase Estimation algorithm across diverse architectures from Intel, AMD, and Apple. The experimental results demonstrate that the proposed optimization successfully mitigates performance degradation at larger qubit scales. In particular, for a 28-qubit simulation, the optimization reversed a performance deficit on an Intel Core i5, improving the GPU speedup over the CPU from 0.95x to 1.89x, and increased the Apple M1 Pro speedup from 3.71x to 5.88x. Overall, this approach yields consistent execution time improvements, demonstrating the viability of integrated GPUs for efficient quantum circuit simulation.

Large language models can now be personalised efficiently at scale using parameter efficient finetuning methods (PEFTs), but serving user-specific PEFTs harms throughput, even with specialised kernels and memory management techniques. This is because, theoretically and empirically, a mismatch exists between prefill (processing a large number of tokens at once) and decode (generating a single token autoregressively): the latter has far lower throughput when serving multiple adapters. Rather than optimising performance relative to parameter count, for efficient multi-adapter serving, we instead ought to optimise performance relative to serving throughput. We therefore propose PreFT (Prefill-only Finetuning), wherein we only apply the adapter to prefill tokens and discard it afterwards. PreFT significantly increases throughput with minimal effect on performance. We develop and release an efficient implementation of two prefill-only PEFTs, LoRA and ReFT, on the vLLM inference engine. We first show that serving multi-user PreFTs is more efficient than traditional PEFTs ($1.9\times$ the throughput when serving $512$ adapters on Llama 3.1 70B). Then, we compare the performance of prefill-only vs. all-token adapters on a variety of supervised finetuning and reinforcement learning tasks with LMs at varying scales. On SFT, we observe that the evaluation loss of PreFTs is higher than PEFTs, but can be compensated by increasing rank with nearly no reduction in throughput. On RL, we consistently find that PreFTs approach parity with standard PEFTs. Together, this work validates prefill-only adaptation of LLMs as a more favourable accuracy-throughput tradeoff than existing PEFTs for personalised serving.

Discrete diffusion language models improve generation efficiency through parallel token prediction, but standard $X_0$ prediction methods introduce factorization errors by approximating the clean token posterior with independent token-wise distributions. This paper proposes Factorization-Error-Free Discrete Diffusion Language Modeling (FeF-DLLM), which replaces independent clean-token prediction with an exact prefix-conditioned factorization of the clean posterior to better preserve token dependencies. To reduce the sequential cost introduced by prefix conditioning, FeF-DLLM further incorporates speculative decoding within diffusion denoising, accelerating inference while maintaining the parallel prediction and re-masking properties of DLLMs. Theoretically, we prove that FeF-DLLM generates from the true joint distribution and derive its expected acceleration ratio. Experiments on GSM8K, MATH, HumanEval, and MBPP demonstrate that our method improves accuracy by an average of 5.04 percentage points while achieving an average inference speedup of $3.86\times$.

We present EnergyLens, an end-to-end framework for energy-aware large language model (LLM) inference optimization. As LLMs scale, predicting and reducing their energy footprint has become critical for sustainability and datacenter operations, yet existing approaches either require production-level code and expensive profiling or fail to accurately capture multi-GPU energy behavior. As a result, practitioners lack tools for deciding which optimizations to prioritize and for selecting among existing deployment configurations when exhaustive profiling is impractical. EnergyLens addresses this gap with an intuitive einsum-based interface that captures LLM specifications including fusion, parallelism, and compute-communication overlap, combined with load-imbalance-aware MoE modeling and an empirically driven communication energy model for multi-GPU settings. We validate EnergyLens on Llama3 and Qwen3-MoE across tensor-parallel and expert-parallel configurations, achieving mean absolute percentage errors (MAPEs) between 9.25% and 13.19% for multi-GPU prefill and decode energy, and 12.97% across SM allocations for Megatron-style overlap. Our energy-driven exploration reveals up to 1.47x and 52.9x energy variation across configurations in prefill and decode efficiency and motivates distributed serving. We further show that compute-communication overlap is difficult to optimize with intuition alone, but EnergyLens correctly identifies Pareto-optimal overlap configurations.

Streaming video understanding with large vision-language models (VLMs) requires a compact memory that can support future reasoning over an ever-growing visual history. A common solution is to compress the key-value (KV) cache, but existing streaming methods typically rely on local token-wise heuristics, such as recency, temporal redundancy, or saliency, which do not explicitly optimize whether the retained cache is representative of the accumulated history. We propose to view KV-cache compression as a coreset selection problem: rather than scoring tokens independently for retention, we select a small subset that covers the geometry of the accumulated visual cache. Our method operates in a joint KV representation and introduces a bicriteria objective that balances coverage in key and value spaces, preserving both retrieval structure and output-relevant information. To encourage a more diverse retained subset, we further introduce an orthogonality-driven diversity criterion that favors candidates contributing new directions beyond the current selection, and connect this criterion to log-determinant subset selection. Across four open-source VLMs and five long-video and streaming-video benchmarks, our method improves over heuristic streaming compression baselines under a fixed cache budget. These results highlight that representative coreset selection offers a more effective principle, than token-wise pruning, for memory-constrained streaming video understanding.

Standard unlearning evaluations measure behavioral suppression in full precision, immediately after training, despite every deployed language model being quantized first. Recent work has shown that 4-bit post-training quantization can reverse machine unlearning; we show this is not a tuning artefact but a systematic dual failure: gradient-based methods that achieve meaningful forgetting lose it under compression, while methods that survive quantization barely change the model. Both failures trace to the same root cause: across all baselines, per-parameter updates lie 47-828x below the NF4 quantization bin width; updates diffused across billions of parameters cannot clear quantization bin boundaries, a consequence we formalize as a sparsity-permanence tradeoff. We present MANSU (Mechanistic-Aligned Null-Space Unlearning), which resolves both modes by combining causal circuit attribution to isolate the minimal forget-set subgraph, circuit-restricted null-space projection with a diagonal-Fisher retain bound, and a per-parameter magnitude floor guaranteeing quantization survival by construction. We additionally introduce Circuit Attribution Divergence (CAD), a mechanistic verification metric distinguishing structural erasure from behavioral suppression, a distinction existing metrics cannot make. Across multiple model families and hazard benchmarks, MANSU is the first method to jointly satisfy all four properties with margin on each (meaningful forgetting, retain preservation, non-positive PTQ gap, and structural erasure), while gradient-based baselines recover up to +0.05 accuracy under compression.

We introduce XFP, a dynamic weight quantizer for LLM inference that inverts the conventional workflow: the operator specifies reconstruction quality floors on per-channel cosine similarity (one strict floor for attention and shared experts, one lazy floor for routed-expert MoE); XFP determines codebook size, outlier budget, and packing per layer automatically — no Hessian, no calibration data, no manual bit-width selection. Each weight matrix is decomposed into a sparse fp16 outlier residual and a dense sub-byte index tensor into a per-group learned codebook. Two storage modes share one auto-select frontend and one fused decode kernel: V2 (per-channel Lloyd) and V2a (shared library of L=32 codebooks per layer). On Qwen3.5-122B-A10B under V2, XFP reaches 138 tok/s single-stream decode on workstation hardware (RTX PRO 6000 Blackwell, TP=2) at 94.49% GSM8K strict-match (3 seeds, n=3957), and is 49% faster than Marlin INT4 at TP=1. For models that do not fit in the target memory envelope, we present the H-Process: a quality-driven iteration over the two cosine thresholds that finds the operating point at which the model just fits while still producing sensible output. Three constraints define its search space: the operator-set thresholds, an OOM boundary at quantize-on-load, and a garbage boundary in generation (cosine similarity steers; benches verify). On Qwen3.5-397B-A17B (512 routed experts/layer), the H-Process fits the full expert population into 2x96 GB at ~3.4 effective bits and delivers 100.9 tok/s long-output decode at 66.72% GSM8K strict-match on the full 1319-problem set (single seed at submission; multi-seed evaluation in progress), exceeding INT4 with routed-expert pruning on memory, throughput, and accuracy simultaneously.

Visual Autoregressive (VAR) models have recently demonstrated impressive image generation quality while maintaining low latency. However, they suffer from severe KV-cache memory constraints, often requiring gigabytes of memory per generated image. We introduce HeatKV, a novel compression method that adapts cache allocation in each head based on its attention to previously generated scales. Using a small offline calibration set, the attention heads are ranked according to their attention scores over prior scales. Based on this ranking, we construct a static pruning schedule tailored to a given memory budget. Applied to the Infinity-2B model, HeatKV achieves $2 \times$ higher compression ratio in memory allocation for KV cache compared to existing methods, while maintaining similar or better image fidelity, prompt alignment and human perception score. Our method achieves a new state-of-the-art (SOTA) for VAR model KV-cache compression, showcasing the effectiveness of fine-grained, head-specific cache allocation.

Scaled Outer Product (SOP) is a post-training quantization methodology for large language model weights, designed to deliver near-lossless fidelity at 4.5--6 bits per weight on hardware with per-layer LUT decode. The methodology combines per-layer search of fixed and dynamic codebook pairs selected by a per-block selection bit, signed per-block scales, activation-weighted cosine selection, and multiple-choice knapsack promotion of sensitive layers with outlier and sparse-residual correction. Fixed codebooks include NF4, BOF4, Split87, and SH4; per-layer optimized codebooks (DD4) are hosted in LUT SRAM. A new hardware-efficient LUT output format (HIF) is proposed to improve performance, energy, and cost. Across six open model families, the recommended FP6 operating point (E2M3sUE4M4, 6.5 bpw) achieves lower weight reconstruction error than the conventional per-layer-POT FP8 baseline (E4M3, 8.0 bpw) at 1.5 bpw lower storage cost, demonstrating that block-scaled small atoms with carefully chosen scale precision can replace conventionally-deployed FP8. Full evaluation across the 4.5--6 bpw range, including layer promotion and sparse residual correction, is reported in a companion paper.

Speculative decoding (SD) accelerates large language model (LLM) inference by using a smaller draft model to propose multiple tokens that are verified by a larger target model in parallel. While prior work demonstrates substantial speedups in isolated or fixed-batch settings, the behavior of SD in production serving systems remains poorly understood: request load varies over time, and effective batch size emerges from the serving system rather than being directly controlled or observed. In this work, we develop a simple and interpretable latency model for SD in LLM serving. We infer effective batch size from request rate using Little's Law and decompose per-request demand into load-independent and load-dependent components for prefill, drafting, and verification. We validate our model using extensive measurements from vLLM across verifier and drafter model sizes, prefill and decode lengths, request rates, draft lengths, and acceptance probabilities. The model accurately describes observed latency, explains why speedups often diminish as server load increases, and characterizes how draft length, acceptance rate, and verifier-drafter size shape latency across serving conditions, with implications for configuring SD in deployed systems. We further show how the framework extends to mixture of experts models, where sparse expert activation changes the effective service costs across load regimes. Together, our results provide a structured framework for understanding SD in real LLM serving systems.

Diffusion-based video generation has advanced substantially in visual fidelity and temporal coherence, but practical deployment remains limited by the quadratic complexity of full attention. Training-free sparse attention is attractive because it accelerates pretrained models without retraining, yet existing online top-$p$ sparse attention still spends non-negligible cost on mask prediction and applies shared thresholds despite strong head-level heterogeneity. We show that these two overlooked factors limit the practical speed-quality trade-off of training-free sparse attention in Video DiTs. To address them, we introduce a head-wise adaptive framework with two plug-in components: Temporal Mask Reuse, which skips unnecessary mask prediction based on query-key drift, and Error-guided Budgeted Calibration, which assigns per-head top-$p$ thresholds by minimizing measured model-output error under a global sparsity budget. On Wan2.1-1.3B and Wan2.1-14B, our method consistently improves XAttention and SVG2, achieving up to 1.93 times speedup at 720P while maintaining competitive video quality and similarity metrics.

Speech-driven gestures and facial animations are fundamental to expressive digital avatars in games, virtual production, and interactive media. However, existing methods are either limited to a single modality for audio motion alignment, failing to fully utilize the potential of massive human motion data, or are constrained by the representation ability and throughput of multimodal models, which makes it difficult to achieve high-quality motion generation or real-time performance. We present UMo, a unified sparse motion modeling architecture for real-time co-speech avatars, which processes text, audio, and motion tokens within a unified formulation. Leveraging a spatially sparse Mixture-of-Experts framework and a temporally sparse, keyframe-centric design, UMo efficiently performs real-time dense reconstruction, enabling temporally coherent and high-fidelity animation generation for both facial expressions and gestures. Furthermore, we implement a multi-stage training strategy with targeted audio augmentation to enhance acoustic diversity and semantic consistency. Consequently, UMo preserves fine-grained speech-motion alignment even under strict latency constraints. Extensive quantitative and qualitative evaluations show that UMo achieves better output quality under low latency and real-time performance constraints, offering a practical solution for high-fidelity real-time co-speech avatars.

We introduce SANA-WM, an efficient 2.6B-parameter open-source world model natively trained for one-minute generation, synthesizing high-fidelity, 720p, minute-scale videos with precise camera control. SANA-WM achieves visual quality comparable to large-scale industrial baselines such as LingBot-World and HY-WorldPlay, while significantly improving efficiency. Four core designs drive our architecture: (1) Hybrid Linear Attention combines frame-wise Gated DeltaNet (GDN) with softmax attention for memory-efficient long-context modeling. (2) Dual-Branch Camera Control ensures precise 6-DoF trajectory adherence. (3) Two-Stage Generation Pipeline applies a long-video refiner to stage-1 outputs, improving quality and consistency across sequences. (4) Robust Annotation Pipeline extracts accurate metric-scale 6-DoF camera poses from public videos to yield high-quality, spatiotemporally consistent action labels. Driven by these designs, SANA-WMdemonstrates remarkable efficiency across data, training compute, and inference hardware: it uses only $\sim$213K public video clips with metric-scale pose supervision, completes training in 15 days on 64 H100s, and generates each 60s clip on a single GPU; its distilled variant can be deployed on a single RTX 5090 with NVFP4 quantization to denoise a 60s 720p clip in 34s. On our one-minute world-model benchmark, SANA-WM demonstrates stronger action-following accuracy than prior open-source baselines and achieves comparable visual quality at $36\times$ higher throughput for scalable world modeling.

Aligning streaming autoregressive (AR) video generators with human preferences is challenging. Existing reinforcement learning methods predominantly rely on noise-based exploration and SDE-based surrogate policies that are mismatched to the deterministic ODE dynamics of distilled AR models, and tend to perturb low-level appearance rather than the high-level semantic storyline progression critical for long-horizon coherence. To address these limitations, we present KVPO, an ODE-native online Group Relative Policy Optimization (GRPO) framework for aligning streaming video generators. For diversity exploration, KVPO introduces a causal-semantic exploration paradigm that relocates the source of variation from stochastic noise to the historical KV cache. By stochastically routing historical KV entries, it constructs semantically diverse generation branches that remain strictly on the data manifold. For policy modeling, KVPO introduces a velocity-field surrogate policy based on Trajectory Velocity Energy (TVE), which quantifies branch likelihood in flow-matching velocity space and yields a reward-weighted contrastive objective fully consistent with the native ODE formulation. Experiments on multiple distilled AR video generators demonstrate consistent gains in visual quality, motion quality, and text-video alignment across both single-prompt short-video and multi-prompt long-video settings.

Autoregressive video diffusion models support real-time synthesis but suffer from error accumulation and context loss over long horizons. We discover that attention heads in AR video diffusion transformers serve functionally distinct roles as local heads for detail refinement, anchor heads for structural stabilization, and memory heads for long-range context aggregation, yet existing methods treat them uniformly, leading to suboptimal KV cache allocation. We propose Head Forcing, a training-free framework that assigns each head type a tailored KV cache strategy: local and anchor heads retain only essential tokens, while memory heads employ a hierarchical memory system with dynamic episodic updates for long-range consistency. A head-wise RoPE re-encoding scheme further ensures positional encodings remain within the pretrained range. Without additional training, Head Forcing extends generation from 5 seconds to minute-level duration, supports multi-prompt interactive synthesis, and consistently outperforms existing baselines. Project Page: https://jiahaotian-sjtu.github.io/headforcing.github.io/.

Real-time interactive video generation requires low-latency, streaming, and controllable rollout. Existing autoregressive (AR) diffusion distillation methods have achieved strong results in the chunk-wise 4-step regime by distilling bidirectional base models into few-step AR students, but they remain limited by coarse response granularity and non-negligible sampling latency. In this paper, we study a more aggressive setting: frame-wise autoregression with only 1--2 sampling steps. In this regime, we identify the initialization of a few-step AR student as the key bottleneck: existing strategies are either target-misaligned, incapable of few-step generation, or too costly to scale. We propose Causal Forcing++, a principled and scalable pipeline that uses causal consistency distillation (causal CD) for few-step AR initialization. The core idea is that causal CD learns the same AR-conditional flow map as causal ODE distillation, but obtains supervision from a single online teacher ODE step between adjacent timesteps, avoiding the need to precompute and store full PF-ODE trajectories. This makes the initialization both more efficient and easier to optimize. The resulting pipeline, \ours, surpasses the SOTA 4-step chunk-wise Causal Forcing under the frame-wise 2-step setting by 0.1 in VBench Total, 0.3 in VBench Quality, and 0.335 in VisionReward, while reducing first-frame latency by 50% and Stage 2 training cost by $\sim$$4\times$. We further extend the pipeline to action-conditioned world model generation in the spirit of Genie3. Project Page: https://github.com/thu-ml/Causal-Forcing and https://github.com/shengshu-ai/minWM .

Diffusion Transformers (DiTs) deliver remarkable image and video generation quality but incur high computational cost, limiting scalability and on-device deployment. We introduce CoReDiT, a structured token pruning framework for DiTs across vision tasks. CoReDiT uses a linear-time spatial coherence score to estimate local redundancy in the latent token lattice and skips high coherence (redundant) tokens in self-attention. To maintain a dense representation and avoid visual discontinuities, we reconstruct skipped attention outputs via coherence-guided aggregation of spatially neighboring retained tokens. We further introduce a progressive, block-adaptive pruning schedule that increases pruning gradually and allocates larger budgets to blocks and denoising steps with higher redundancy. Across state-of-the-art diffusion backbones including PixArt-α and MagicDrive-V2, CoReDiT achieves up to 55% self-attention FLOPs reduction and inference speedups of 1.33x on cloud GPUs and 1.72x on mobile NPUs, while maintaining high visual quality. Notably, CoReDiT also increases on-device memory head-room, enabling higher-resolution generation.