arXiv subjects

Xinyu Duan

Publications and source records attributed to Xinyu Duan.

At least 19 recordsLinked to original sources

arXiv2026

Magnitude-Direction Decoupling for Fast Video Generation with Flow Matching Models

Flow matching models for video generation achieve impressive performance but suffer from high computational overhead due to iterative denoising. In fact, the original model is not necessary for all denoising steps, allowing some steps to use lightweight alternatives for faster sampling. However, directly using caching or lightweight models can deviate from the original denoising trajectory, resulting in suboptimal performance. Through empirical analysis, we find that lightweight models can robustly capture the magnitude components of the original model's output, while caching provides reliable directional guidance. Building on this insight, we propose the Magnitude-Direction Decoupling (MDD) method, which adaptively employs a direction-calibrated lightweight model as a substitute for the original model to accelerate inference and effectively correct deviations in the denoising trajectory. Moreover, MDD further reduces inference costs by reusing magnitude information under classifier-free guidance (CFG). As a result, MDD offers a more reliable and lightweight solution to accelerate sampling. Experiments show that MDD outperforms existing acceleration methods, delivering promising speedups (e.g., up to 2.95x on Wan2.1) while preserving high visual fidelity and content richness.

cs.CV↗

arXiv2026

MLLM-Guided Semantic Correction for Text-to-Video Generation

Recent advances in diffusion models and Transformer architectures have led to significant progress in text-to-video generation. However, these models often suffer from semantic errors such as missing objects, incorrect attributes, or mismatched actions. Although some semantic correction methods perform optimization before sampling or refinement after sampling, how to detect and correct semantic deviations during the video generation process remains underexplored. In this paper, we introduce a training-free, interpretable mid-generation correction framework that integrates multimodal large language model (MLLM) feedback directly into the diffusion sampling loop. Our framework achieves diffusion trajectory correction by injecting semantic evaluation signals during video synthesis, enabling the model to optimize the generated content through continuous self-reflection. We propose two key modules: a Semantic Assessment Supervisor that generates intermediate preview frames for semantic evaluations and deviation diagnostics, and a Semantic Modification Assistant that corrects semantic drift during inference via a controllable latent trajectory intervention. Our method improves semantic alignment, visual fidelity, and temporal consistency without modifying model parameters. We validate the effectiveness of our approach through extensive experiments across multiple benchmarks.

cs.CV↗

arXiv2026

Cross-Resolution Distribution Matching for Diffusion Distillation

Diffusion distillation is central to accelerating image and video generation, yet existing methods are fundamentally limited by the denoising process, where step reduction has largely saturated. Partial timestep low-resolution generation can further accelerate inference, but it suffers noticeable quality degradation due to cross-resolution distribution gaps. We propose Cross-Resolution Distribution Matching Distillation (RMD), a novel distillation framework that bridges cross-resolution distribution gaps for high-fidelity, few-step multi-resolution cascaded inference. Specifically, RMD divides the timestep intervals for each resolution using logarithmic signal-to-noise ratio (logSNR) curves, and introduces logSNR-based mapping to compensate for resolution-induced shifts. Distribution matching is conducted along resolution trajectories to reduce the gap between low-resolution generator distributions and the teacher's high-resolution distribution. In addition, a predicted-noise re-injection mechanism is incorporated during upsampling to stabilize training and improve synthesis quality. Quantitative and qualitative results show that RMD preserves high-fidelity generation while accelerating inference across various backbones. Notably, RMD achieves up to 33.4X speedup on SDXL and 25.6X on Wan2.1-14B, while preserving high visual fidelity.

cs.CV↗

arXiv2025

$A^3$: Attention-Aware Accurate KV Cache Fusion for Fast Large Language Model Serving

Large language models (LLMs) have demonstrated strong capabilities in processing long contexts, enabling them to tackle tasks involving long textual inputs such as multi-turn conversations, legal documents, or retrieved documents in Retrieval-Augmented Generation (RAG) systems. However, despite their ability to handle long sequences, the resulting decoding latency and memory overhead remain substantial, posing challenges for real-world deployment. Recent advances in KV Cache reuse have shown potential to mitigate these costs, but still suffer from notable performance degradation. To address this issue, we conduct an in-depth investigation of recomputation-based reuse methods and observe that the recomputed tokens often fail to align with the context segments most relevant to the question. This misalignment hinders proper updates to the critical contextual representations. Therefore, we propose the $\textbf{A}$ttention-$\textbf{A}$ware $\textbf{A}$ccurate KV Cache Fusion algorithm ($A^3$), which precomputes and selectively fuses the KV Cache of text chunks based on their relevance to the question, achieving accurate integration with minimal computational overhead. Extensive experiments on various benchmarks and LLMs demonstrate that $A^3$ achieves the best task performance compared to four baselines while reducing the time-to-first-token (TTFT) by 2$\times$.

cs.CL↗

arXiv2025

Huawei Cloud Model-as-a-Service on the CloudMatrix384 SuperPod

Scaled-out MoE LLMs and scaled-up SuperPods create new systems challenges for production Model-as-a-Service (MaaS), requiring disaggregation, low-latency communication, and decentralized serving. This report presents xDeepServe, the production serving system behind Huawei Cloud's MaaS offering on CloudMatrix384, a 48-server SuperPod with 384 Ascend 910C chips connected by a high-bandwidth UB fabric and global shared memory. It serves models including DeepSeek, Kimi, GLM, Qwen, and MiniMax, among others. xDeepServe is built around Transformerless, a disaggregated execution architecture that decomposes transformer inference into modular units -- attention, feedforward, and MoE -- and supports disaggregated Prefill-Decode and MoE-Attention deployments. To enable disaggregation, we develop XCCL, a memory-semantic communication layer providing microsecond-level point-to-point and scalable all-to-all primitives, and we extend FlowServe with decentralized DP groups and techniques to mitigate stragglers and synchronization variance. In a peak decoding configuration, xDeepServe reaches 2400 tokens/s per Ascend 910C chip at ~50ms time-per-output-token (TPOT).

cs.DC↗

arXiv2025

CaliDrop: KV Cache Compression with Calibration

Large Language Models (LLMs) require substantial computational resources during generation. While the Key-Value (KV) cache significantly accelerates this process by storing attention intermediates, its memory footprint grows linearly with sequence length, batch size, and model size, creating a bottleneck in long-context scenarios. Various KV cache compression techniques, including token eviction, quantization, and low-rank projection, have been proposed to mitigate this bottleneck, often complementing each other. This paper focuses on enhancing token eviction strategies. Token eviction leverages the observation that the attention patterns are often sparse, allowing for the removal of less critical KV entries to save memory. However, this reduction usually comes at the cost of notable accuracy degradation, particularly under high compression ratios. To address this issue, we propose \textbf{CaliDrop}, a novel strategy that enhances token eviction through calibration. Our preliminary experiments show that queries at nearby positions exhibit high similarity. Building on this observation, CaliDrop performs speculative calibration on the discarded tokens to mitigate the accuracy loss caused by token eviction. Extensive experiments demonstrate that CaliDrop significantly improves the accuracy of existing token eviction methods.

cs.CL↗

arXiv2025

Rhythm Controllable and Efficient Zero-Shot Voice Conversion via Shortcut Flow Matching

Zero-Shot Voice Conversion (VC) aims to transform the source speaker's timbre into an arbitrary unseen one while retaining speech content. Most prior work focuses on preserving the source's prosody, while fine-grained timbre information may leak through prosody, and transferring target prosody to synthesized speech is rarely studied. In light of this, we propose R-VC, a rhythm-controllable and efficient zero-shot voice conversion model. R-VC employs data perturbation techniques and discretize source speech into Hubert content tokens, eliminating much content-irrelevant information. By leveraging a Mask Generative Transformer for in-context duration modeling, our model adapts the linguistic content duration to the desired target speaking style, facilitating the transfer of the target speaker's rhythm. Furthermore, R-VC introduces a powerful Diffusion Transformer (DiT) with shortcut flow matching during training, conditioning the network not only on the current noise level but also on the desired step size, enabling high timbre similarity and quality speech generation in fewer sampling steps, even in just two, thus minimizing latency. Experimental results show that R-VC achieves comparable speaker similarity to state-of-the-art VC methods with a smaller dataset, and surpasses them in terms of speech naturalness, intelligibility and style transfer performance.

eess.AS↗

arXiv2025

Alignment-Augmented Speculative Decoding with Alignment Sampling and Conditional Verification

Recent works have revealed the great potential of speculative decoding in accelerating the autoregressive generation process of large language models. The success of these methods relies on the alignment between draft candidates and the sampled outputs of the target model. Existing methods mainly achieve draft-target alignment with training-based methods, e.g., EAGLE, Medusa, involving considerable training costs. In this paper, we present a training-free alignment-augmented speculative decoding algorithm. We propose alignment sampling, which leverages output distribution obtained in the prefilling phase to provide more aligned draft candidates. To further benefit from high-quality but non-aligned draft candidates, we also introduce a simple yet effective flexible verification strategy. Through an adaptive probability threshold, our approach can improve generation accuracy while further improving inference efficiency. Experiments on 8 datasets (including question answering, summarization and code completion tasks) show that our approach increases the average generation score by 3.3 points for the LLaMA3 model. Our method achieves a mean acceptance length up to 2.39 and speed up generation by 2.23.

cs.CL↗

arXiv2025

Accurate KV Cache Quantization with Outlier Tokens Tracing

The impressive capabilities of Large Language Models (LLMs) come at the cost of substantial computational resources during deployment. While KV Cache can significantly reduce recomputation during inference, it also introduces additional memory overhead. KV Cache quantization presents a promising solution, striking a good balance between memory usage and accuracy. Previous research has shown that the Keys are distributed by channel, while the Values are distributed by token. Consequently, the common practice is to apply channel-wise quantization to the Keys and token-wise quantization to the Values. However, our further investigation reveals that a small subset of unusual tokens exhibit unique characteristics that deviate from this pattern, which can substantially impact quantization accuracy. To address this, we develop a simple yet effective method to identify these tokens accurately during the decoding process and exclude them from quantization as outlier tokens, significantly improving overall accuracy. Extensive experiments show that our method achieves significant accuracy improvements under 2-bit quantization and can deliver a 6.4 times reduction in memory usage and a 2.3 times increase in throughput.

cs.CL↗

arXiv2025

Taming the Titans: A Survey of Efficient LLM Inference Serving

Large Language Models (LLMs) for Generative AI have achieved remarkable progress, evolving into sophisticated and versatile tools widely adopted across various domains and applications. However, the substantial memory overhead caused by their vast number of parameters, combined with the high computational demands of the attention mechanism, poses significant challenges in achieving low latency and high throughput for LLM inference services. Recent advancements, driven by groundbreaking research, have significantly accelerated progress in this field. This paper provides a comprehensive survey of these methods, covering fundamental instance-level approaches, in-depth cluster-level strategies, emerging scenario directions, and other miscellaneous but important areas. At the instance level, we review model placement, request scheduling, decoding length prediction, storage management, and the disaggregation paradigm. At the cluster level, we explore GPU cluster deployment, multi-instance load balancing, and cloud service solutions. For emerging scenarios, we organize the discussion around specific tasks, modules, and auxiliary methods. To ensure a holistic overview, we also highlight several niche yet critical areas. Finally, we outline potential research directions to further advance the field of LLM inference serving.

cs.CL↗

arXiv2025

Adaptive Feature-based Low-Rank Compression of Large Language Models via Bayesian Optimization

In recent years, large language models (LLMs) have driven advances in natural language processing. Still, their growing scale has increased the computational burden, necessitating a balance between efficiency and performance. Low-rank compression, a promising technique, reduces non-essential parameters by decomposing weight matrices into products of two low-rank matrices. Yet, its application in LLMs has not been extensively studied. The key to low-rank compression lies in low-rank factorization and low-rank dimensions allocation. To address the challenges of low-rank compression in LLMs, we conduct empirical research on the low-rank characteristics of large models. We propose a low-rank compression method suitable for LLMs. This approach involves precise estimation of feature distributions through pooled covariance matrices and a Bayesian optimization strategy for allocating low-rank dimensions. Experiments on the LLaMA-2 models demonstrate that our method outperforms existing strong structured pruning and low-rank compression techniques in maintaining model performance at the same compression ratio.

cs.CL↗

arXiv2024

Beware of Calibration Data for Pruning Large Language Models

As large language models (LLMs) are widely applied across various fields, model compression has become increasingly crucial for reducing costs and improving inference efficiency. Post-training pruning is a promising method that does not require resource-intensive iterative training and only needs a small amount of calibration data to assess the importance of parameters. Recent research has enhanced post-training pruning from different aspects but few of them systematically explore the effects of calibration data, and it is unclear if there exist better calibration data construction strategies. We fill this blank and surprisingly observe that calibration data is also crucial to post-training pruning, especially for high sparsity. Through controlled experiments on important influence factors of calibration data, including the pruning settings, the amount of data, and its similarity with pre-training data, we observe that a small size of data is adequate, and more similar data to its pre-training stage can yield better performance. As pre-training data is usually inaccessible for advanced LLMs, we further provide a self-generating calibration data synthesis strategy to construct feasible calibration data. Experimental results on recent strong open-source LLMs (e.g., DCLM, and LLaMA-3) show that the proposed strategy can enhance the performance of strong pruning methods (e.g., Wanda, DSnoT, OWL) by a large margin (up to $2.68\%$). Code is available at https://github.com/Dereck0602/calibration_data.

cs.CL↗

arXiv2024

OPT-Tree: Speculative Decoding with Adaptive Draft Tree Structure

Autoregressive language models demonstrate excellent performance in various scenarios. However, the inference efficiency is limited by its one-step-one-word generation mode, which has become a pressing problem recently as the models become increasingly larger. Speculative decoding employs a "draft and then verify" mechanism to allow multiple tokens to be generated in one step, realizing lossless acceleration. Existing methods mainly adopt fixed heuristic draft structures, which fail to adapt to different situations to maximize the acceptance length during verification. To alleviate this dilemma, we proposed OPT-Tree, an algorithm to construct adaptive and scalable draft trees. It searches the optimal tree structure that maximizes the mathematical expectation of the acceptance length in each decoding step. Experimental results reveal that OPT-Tree outperforms the existing draft structures and achieves a speed-up ratio of up to 3.2 compared with autoregressive decoding. If the draft model is powerful enough and the node budget is sufficient, it can generate more than ten tokens in a single step. Our code is available at https://github.com/Jikai0Wang/OPT-Tree.

cs.CL↗

arXiv2024

High-order Joint Constituency and Dependency Parsing

This work revisits the topic of jointly parsing constituency and dependency trees, i.e., to produce compatible constituency and dependency trees simultaneously for input sentences, which is attractive considering that the two types of trees are complementary in representing syntax. The original work of Zhou and Zhao (2019) performs joint parsing only at the inference phase. They train two separate parsers under the multi-task learning framework (i.e., one shared encoder and two independent decoders). They design an ad-hoc dynamic programming-based decoding algorithm of $O(n^5)$ time complexity for finding optimal compatible tree pairs. Compared to their work, we make progress in three aspects: (1) adopting a much more efficient decoding algorithm of $O(n^4)$ time complexity, (2) exploring joint modeling at the training phase, instead of only at the inference phase, (3) proposing high-order scoring components to promote constituent-dependency interaction. We conduct experiments and analysis on seven languages, covering both rich-resource and low-resource scenarios. Results and analysis show that joint modeling leads to a modest overall performance boost over separate modeling, but substantially improves the complete matching ratio of whole trees, thanks to the explicit modeling of tree compatibility.

cs.CL↗

arXiv2024

AuG-KD: Anchor-Based Mixup Generation for Out-of-Domain Knowledge Distillation

Due to privacy or patent concerns, a growing number of large models are released without granting access to their training data, making transferring their knowledge inefficient and problematic. In response, Data-Free Knowledge Distillation (DFKD) methods have emerged as direct solutions. However, simply adopting models derived from DFKD for real-world applications suffers significant performance degradation, due to the discrepancy between teachers' training data and real-world scenarios (student domain). The degradation stems from the portions of teachers' knowledge that are not applicable to the student domain. They are specific to the teacher domain and would undermine students' performance. Hence, selectively transferring teachers' appropriate knowledge becomes the primary challenge in DFKD. In this work, we propose a simple but effective method AuG-KD. It utilizes an uncertainty-guided and sample-specific anchor to align student-domain data with the teacher domain and leverages a generative method to progressively trade off the learning process between OOD knowledge distillation and domain-specific information learning via mixup learning. Extensive experiments in 3 datasets and 8 settings demonstrate the stability and superiority of our approach. Code available at https://github.com/IshiKura-a/AuG-KD .

cs.LG↗

arXiv2023

TransFace: Unit-Based Audio-Visual Speech Synthesizer for Talking Head Translation

Direct speech-to-speech translation achieves high-quality results through the introduction of discrete units obtained from self-supervised learning. This approach circumvents delays and cascading errors associated with model cascading. However, talking head translation, converting audio-visual speech (i.e., talking head video) from one language into another, still confronts several challenges compared to audio speech: (1) Existing methods invariably rely on cascading, synthesizing via both audio and text, resulting in delays and cascading errors. (2) Talking head translation has a limited set of reference frames. If the generated translation exceeds the length of the original speech, the video sequence needs to be supplemented by repeating frames, leading to jarring video transitions. In this work, we propose a model for talking head translation, \textbf{TransFace}, which can directly translate audio-visual speech into audio-visual speech in other languages. It consists of a speech-to-unit translation model to convert audio speech into discrete units and a unit-based audio-visual speech synthesizer, Unit2Lip, to re-synthesize synchronized audio-visual speech from discrete units in parallel. Furthermore, we introduce a Bounded Duration Predictor, ensuring isometric talking head translation and preventing duplicate reference frames. Experiments demonstrate that our proposed Unit2Lip model significantly improves synchronization (1.601 and 0.982 on LSE-C for the original and generated audio speech, respectively) and boosts inference speed by a factor of 4.35 on LRS2. Additionally, TransFace achieves impressive BLEU scores of 61.93 and 47.55 for Es-En and Fr-En on LRS3-T and 100% isochronous translations.

cs.SD↗

arXiv2023

StyleSinger: Style Transfer for Out-of-Domain Singing Voice Synthesis

Style transfer for out-of-domain (OOD) singing voice synthesis (SVS) focuses on generating high-quality singing voices with unseen styles (such as timbre, emotion, pronunciation, and articulation skills) derived from reference singing voice samples. However, the endeavor to model the intricate nuances of singing voice styles is an arduous task, as singing voices possess a remarkable degree of expressiveness. Moreover, existing SVS methods encounter a decline in the quality of synthesized singing voices in OOD scenarios, as they rest upon the assumption that the target vocal attributes are discernible during the training phase. To overcome these challenges, we propose StyleSinger, the first singing voice synthesis model for zero-shot style transfer of out-of-domain reference singing voice samples. StyleSinger incorporates two critical approaches for enhanced effectiveness: 1) the Residual Style Adaptor (RSA) which employs a residual quantization module to capture diverse style characteristics in singing voices, and 2) the Uncertainty Modeling Layer Normalization (UMLN) to perturb the style attributes within the content representation during the training phase and thus improve the model generalization. Our extensive evaluations in zero-shot style transfer undeniably establish that StyleSinger outperforms baseline models in both audio quality and similarity to the reference singing voice samples. Access to singing voice samples can be found at https://aaronz345.github.io/StyleSingerDemo/.

eess.AS↗

arXiv2023

How Well Do Large Language Models Understand Syntax? An Evaluation by Asking Natural Language Questions

While recent advancements in large language models (LLMs) bring us closer to achieving artificial general intelligence, the question persists: Do LLMs truly understand language, or do they merely mimic comprehension through pattern recognition? This study seeks to explore this question through the lens of syntax, a crucial component of sentence comprehension. Adopting a natural language question-answering (Q&A) scheme, we craft questions targeting nine syntactic knowledge points that are most closely related to sentence comprehension. Experiments conducted on 24 LLMs suggest that most have a limited grasp of syntactic knowledge, exhibiting notable discrepancies across different syntactic knowledge points. In particular, questions involving prepositional phrase attachment pose the greatest challenge, whereas those concerning adjectival modifier and indirect object are relatively easier for LLMs to handle. Furthermore, a case study on the training dynamics of the LLMs reveals that the majority of syntactic knowledge is learned during the initial stages of training, hinting that simply increasing the number of training tokens may not be the `silver bullet' for improving the comprehension ability of LLMs.

cs.CL↗