Daily Papers

We introduce Fish Audio S2, an open-sourced text-to-speech system featuring multi-speaker, multi-turn generation, and, most importantly, instruction-following control via natural-language descriptions. To scale training, we develop a multi-stage training recipe together with a staged data pipeline covering video captioning and speech captioning, voice-quality assessment, and reward modeling. To push the frontier of open-source TTS, we release our model weights, fine-tuning code, and an SGLang-based inference engine. The inference engine is production-ready for streaming, achieving an RTF of 0.195 and a time-to-first-audio below 100 ms.Our code and weights are available on GitHub (https://github.com/fishaudio/fish-speech) and Hugging Face (https://huggingface.co/fishaudio/s2-pro). We highly encourage readers to visit https://fish.audio to try custom voices.

This technical report presents MOSS-TTS, a speech generation foundation model built on a scalable recipe: discrete audio tokens, autoregressive modeling, and large-scale pretraining. Built on MOSS-Audio-Tokenizer, a causal Transformer tokenizer that compresses 24 kHz audio to 12.5 fps with variable-bitrate RVQ and unified semantic-acoustic representations, we release two complementary generators: MOSS-TTS, which emphasizes structural simplicity, scalability, and long-context/control-oriented deployment, and MOSS-TTS-Local-Transformer, which introduces a frame-local autoregressive module for higher modeling efficiency, stronger speaker preservation, and a shorter time to first audio. Across multilingual and open-domain settings, MOSS-TTS supports zero-shot voice cloning, token-level duration control, phoneme-/pinyin-level pronunciation control, smooth code-switching, and stable long-form generation. This report summarizes the design, training recipe, and empirical characteristics of the released models.

We present a technical tutorial for building enterprise-grade realtime voice agents from first principles. While end-to-end speech-to-speech models may ultimately provide the best latency for voice agents, fully self-hosted end-to-end solutions are not yet available. We evaluate the closest candidate, Qwen3-Omni, across three configurations: its cloud-only DashScope Realtime API achieves sim702ms audio-to-audio latency with streaming, but is not self-hostable; its local vLLM deployment supports only the Thinker (text generation from audio, 516ms), not the Talker (audio synthesis); and its local Transformers deployment runs the full pipeline but at sim146s -- far too slow for realtime. The cascaded streaming pipeline (STT rightarrow LLM rightarrow TTS) therefore remains the practical architecture for self-hosted realtime voice agents, and the focus of this tutorial. We build a complete voice agent using Deepgram (streaming STT), vLLM-served LLMs with function calling (streaming text generation), and ElevenLabs (streaming TTS), achieving a measured time-to-first-audio of 755ms (best case 729ms) with full function calling support. We release the full codebase as a 9-chapter progressive tutorial with working, tested code for every component.

Current approaches for large audio language models (LALMs) often rely on closed data sources or proprietary models, limiting their generalization and accessibility. This paper introduces MiDashengLM, a novel open audio-language model designed for efficient and comprehensive audio understanding through the use of general audio captions using our novel ACAVCaps training dataset. MiDashengLM exclusively relies on publicly available pretraining and supervised fine-tuning (SFT) datasets, ensuring full transparency and reproducibility. At its core, MiDashengLM integrates Dasheng, an open-source audio encoder, specifically engineered to process diverse auditory information effectively. Unlike previous works primarily focused on Automatic Speech Recognition (ASR) based audio-text alignment, our strategy centers on general audio captions, fusing speech, sound and music information into one textual representation, enabling a holistic textual representation of complex audio scenes. Lastly, MiDashengLM provides an up to 4x speedup in terms of time-to-first-token (TTFT) and up to 20x higher throughput than comparable models. Checkpoints are available online at https://huggingface.co/mispeech/midashenglm-7b and https://github.com/xiaomi-research/dasheng-lm.

Large Audio Language Models (LALMs) still struggle in complex acoustic scenes because they often fail to preserve task-relevant acoustic evidence before reasoning begins. We call this failure the evidence bottleneck: state-of-the-art systems show larger deficits in evidence extraction than in downstream reasoning, suggesting that the main limitation lies in upstream perception rather than reasoning policy. To address this problem, we propose EvA (Evidence-First Audio), a dual-path architecture that combines Whisper and CED-Base through non-compressive, time-aligned fusion. EvA first aggregates intermediate CED layers to preserve multi-scale acoustic cues, then aligns the aggregated CED features to the Whisper timeline and adds the two streams without changing sequence length. We also build EvA-Perception, a large-scale open-source training set with about 54K event-ordered captions (150 h) and about 500K QA pairs. Under a unified zero-shot protocol, EvA achieves the best open-source Perception scores on MMAU, MMAR, and MMSU, and improves over Kimi-Audio-7B on all reported metrics, with the largest gains on perception-heavy splits. These results support the evidence-first hypothesis: stronger audio understanding depends on preserving acoustic evidence before reasoning.

Omnimodal large language models (OmniLLMs) have attracted increasing research attention of late towards unified audio-video understanding, wherein processing audio-video token sequences creates a significant computational bottleneck, however. Existing token compression methods have yet to accommodate this emerging need of jointly compressing multimodal tokens. To bridge this gap, we present OmniZip, a training-free, audio-guided audio-visual token-compression framework that optimizes multimodal token representation and accelerates inference. Specifically, OmniZip first identifies salient audio tokens, then computes an audio retention score for each time group to capture information density, thereby dynamically guiding video token pruning and preserving cues from audio anchors enhanced by cross-modal similarity. For each time window, OmniZip compresses the video tokens using an interleaved spatio-temporal scheme. Extensive empirical results demonstrate the merits of OmniZip - it achieves 3.42X inference speedup and 1.4X memory reduction over other top-performing counterparts, while maintaining performance with no training.

We introduce Moshi, a speech-text foundation model and full-duplex spoken dialogue framework. Current systems for spoken dialogue rely on pipelines of independent components, namely voice activity detection, speech recognition, textual dialogue and text-to-speech. Such frameworks cannot emulate the experience of real conversations. First, their complexity induces a latency of several seconds between interactions. Second, text being the intermediate modality for dialogue, non-linguistic information that modifies meaning -- such as emotion or non-speech sounds -- is lost in the interaction. Finally, they rely on a segmentation into speaker turns, which does not take into account overlapping speech, interruptions and interjections. Moshi solves these independent issues altogether by casting spoken dialogue as speech-to-speech generation. Starting from a text language model backbone, Moshi generates speech as tokens from the residual quantizer of a neural audio codec, while modeling separately its own speech and that of the user into parallel streams. This allows for the removal of explicit speaker turns, and the modeling of arbitrary conversational dynamics. We moreover extend the hierarchical semantic-to-acoustic token generation of previous work to first predict time-aligned text tokens as a prefix to audio tokens. Not only this "Inner Monologue" method significantly improves the linguistic quality of generated speech, but we also illustrate how it can provide streaming speech recognition and text-to-speech. Our resulting model is the first real-time full-duplex spoken large language model, with a theoretical latency of 160ms, 200ms in practice, and is available at https://github.com/kyutai-labs/moshi.

As an increasing majority of global video content is consumed on social platforms for interactive social purposes, video generation models built for social worlds are important but largely overlooked by previous studies. In this work, we define the position of social world models and build a prototype model as the first step towards this goal. While previous world models successfully simulate physical environments or gaming world exploration, they remain fundamentally detached from human-centric social dynamics. To bridge this gap as the first step to social world models, we present MaineCoon, the first real-time audio-visual autoregressive model that has 22B parameters and is capable of real-time streaming generation and sub-second interaction, with a record-breaking frame rate of up to 47.5 FPS, on a single GPU. To the best of our knowledge, MaineCoon is also the first real-time audio-visual generation model specifically optimized for social-interactive applications. To enable efficient and stable training, we introduce several novel techniques into MaineCoon, including self-resampling, cross-modal representation alignment, domain-aware preference optimization, and reinforced online-policy distillation (ROPD). We also design the first agentic streaming inference framework that supports thousand-second-scale or even longer generation while mitigating drift with agentic cache management and prompt planing. These innovations significantly accelerate training while optimizing real-time inference performance. We believe this work not only sets a new state-of-the-art (SOTA) performance benchmark for high-quality, low-latency, and long-horizon audio-visual autoregressive models, but also points out the paradigm shift desired for next-generation AI-native social platforms.

In this work, we introduce the first autoregressive framework for real-time, audio-driven portrait animation, a.k.a, talking head. Beyond the challenge of lengthy animation times, a critical challenge in realistic talking head generation lies in preserving the natural movement of diverse body parts. To this end, we propose Teller, the first streaming audio-driven protrait animation framework with autoregressive motion generation. Specifically, Teller first decomposes facial and body detail animation into two components: Facial Motion Latent Generation (FMLG) based on an autoregressive transfromer, and movement authenticity refinement using a Efficient Temporal Module (ETM).Concretely, FMLG employs a Residual VQ model to map the facial motion latent from the implicit keypoint-based model into discrete motion tokens, which are then temporally sliced with audio embeddings. This enables the AR tranformer to learn real-time, stream-based mappings from audio to motion. Furthermore, Teller incorporate ETM to capture finer motion details. This module ensures the physical consistency of body parts and accessories, such as neck muscles and earrings, improving the realism of these movements. Teller is designed to be efficient, surpassing the inference speed of diffusion-based models (Hallo 20.93s vs. Teller 0.92s for one second video generation), and achieves a real-time streaming performance of up to 25 FPS. Extensive experiments demonstrate that our method outperforms recent audio-driven portrait animation models, especially in small movements, as validated by human evaluations with a significant margin in quality and realism.

Existing diffusion-based video generation methods are fundamentally constrained by sequential computation and long-horizon inconsistency, limiting their practical adoption in real-time, streaming audio-driven avatar synthesis. We present Live Avatar, an algorithm-system co-designed framework that enables efficient, high-fidelity, and infinite-length avatar generation using a 14-billion-parameter diffusion model. Our approach introduces Timestep-forcing Pipeline Parallelism (TPP), a distributed inference paradigm that pipelines denoising steps across multiple GPUs, effectively breaking the autoregressive bottleneck and ensuring stable, low-latency real-time streaming. To further enhance temporal consistency and mitigate identity drift and color artifacts, we propose the Rolling Sink Frame Mechanism (RSFM), which maintains sequence fidelity by dynamically recalibrating appearance using a cached reference image. Additionally, we leverage Self-Forcing Distribution Matching Distillation to facilitate causal, streamable adaptation of large-scale models without sacrificing visual quality. Live Avatar demonstrates state-of-the-art performance, reaching 20 FPS end-to-end generation on 5 H800 GPUs, and, to the best of our knowledge, is the first to achieve practical, real-time, high-fidelity avatar generation at this scale. Our work establishes a new paradigm for deploying advanced diffusion models in industrial long-form video synthesis applications.

Real-time text-driven joint audio-video avatar generation requires jointly synthesizing portrait video and speech with high fidelity and precise synchronization, yet existing audio-visual diffusion models remain too slow for interactive use and often degrade noticeably after aggressive acceleration. We present Hallo-Live, a streaming framework for joint audio-visual avatar generation that combines asynchronous dual-stream diffusion with human-centric preference-guided distillation. To reduce articulation lag in causal generation, we introduce Future-Expanding Attention, which allows each video block to access synchronous audio together with a short horizon of future phonetic cues. To mitigate the quality loss of few-step distillation, we further propose Human-Centric Preference-Guided DMD (HP-DMD), which reweights training samples using rewards from visual fidelity, speech naturalness, and audio-visual synchronization. On two NVIDIA H200 GPUs, Hallo-Live runs at 20.38 FPS with 0.94 seconds latency, yielding 16.0x higher throughput and 99.3x lower latency than the teacher model Ovi. Despite this speedup, it retains strong generation quality, reaching comparable VideoAlign overall score and Sync Confidence score while outperforming other accelerated baselines in the overall quality-efficiency trade-off. Qualitative results further show robust generalization across photorealistic, multi-speaker, and stylized scenarios. To the best of our knowledge, Hallo-Live is the first framework to combine streaming dual-stream diffusion with preference-guided distillation for real-time, text-driven audio-visual generation.

Recent joint audio-visual diffusion models achieve remarkable generation quality but suffer from high latency due to their bidirectional attention dependencies, hindering real-time applications. We propose OmniForcing, the first framework to distill an offline, dual-stream bidirectional diffusion model into a high-fidelity streaming autoregressive generator. However, naively applying causal distillation to such dual-stream architectures triggers severe training instability, due to the extreme temporal asymmetry between modalities and the resulting token sparsity. We address the inherent information density gap by introducing an Asymmetric Block-Causal Alignment with a zero-truncation Global Prefix that prevents multi-modal synchronization drift. The gradient explosion caused by extreme audio token sparsity during the causal shift is further resolved through an Audio Sink Token mechanism equipped with an Identity RoPE constraint. Finally, a Joint Self-Forcing Distillation paradigm enables the model to dynamically self-correct cumulative cross-modal errors from exposure bias during long rollouts. Empowered by a modality-independent rolling KV-cache inference scheme, OmniForcing achieves state-of-the-art streaming generation at sim25 FPS on a single GPU, maintaining multi-modal synchronization and visual quality on par with the bidirectional teacher.Project Page: https://omniforcing.com{https://omniforcing.com}

Audio effects are extensively used at every stage of audio and music content creation. The majority of differentiable audio effects modeling approaches fall into the black-box or gray-box paradigms; and most models have been proposed and applied to nonlinear effects like guitar amplifiers, overdrive, distortion, fuzz and compressor. Although a plethora of architectures have been introduced for the task at hand there is still lack of understanding on the state of the art, since most publications experiment with one type of nonlinear audio effect and a very small number of devices. In this work we aim to shed light on the audio effects modeling landscape by comparing black-box and gray-box architectures on a large number of nonlinear audio effects, identifying the most suitable for a wide range of devices. In the process, we also: introduce time-varying gray-box models and propose models for compressor, distortion and fuzz, publish a large dataset for audio effects research - ToneTwist AFx https://github.com/mcomunita/tonetwist-afx-dataset - that is also the first open to community contributions, evaluate models on a variety of metrics and conduct extensive subjective evaluation. Code https://github.com/mcomunita/nablafx and supplementary material https://github.com/mcomunita/nnlinafx-supp-material are also available.

Research on multi-modal contrastive learning strategies for audio and text has rapidly gained interest. Contrastively trained Audio-Language Models (ALMs), such as CLAP, which establish a unified representation across audio and language modalities, have enhanced the efficacy in various subsequent tasks by providing good text aligned audio encoders and vice versa. These improvements are evident in areas like zero-shot audio classification and audio retrieval, among others. However, the ability of these models to understand natural language and temporal relations is still a largely unexplored and open field for research. In this paper, we propose to equip the multi-modal ALMs with temporal understanding without loosing their inherent prior capabilities of audio-language tasks with a temporal instillation method TeminAL. We implement a two-stage training scheme TeminAL A & B, where the model first learns to differentiate between multiple sounds in TeminAL A, followed by a phase that instills a sense of time, thereby enhancing its temporal understanding in TeminAL B. This approach results in an average performance gain of 5.28% in temporal understanding on the ESC-50 dataset, while the model remains competitive in zero-shot retrieval and classification tasks on the AudioCap/Clotho datasets. We also note the lack of proper evaluation techniques for contrastive ALMs and propose a strategy for evaluating ALMs in zero-shot settings. The general-purpose zero-shot model evaluation strategy ZSTE, is used to evaluate various prior models. ZSTE demonstrates a general strategy to evaluate all ZS contrastive models. The model trained with TeminAL successfully outperforms current models on most downstream tasks.

Deep generative models applied to audio have improved by a large margin the state-of-the-art in many speech and music related tasks. However, as raw waveform modelling remains an inherently difficult task, audio generative models are either computationally intensive, rely on low sampling rates, are complicated to control or restrict the nature of possible signals. Among those models, Variational AutoEncoders (VAE) give control over the generation by exposing latent variables, although they usually suffer from low synthesis quality. In this paper, we introduce a Realtime Audio Variational autoEncoder (RAVE) allowing both fast and high-quality audio waveform synthesis. We introduce a novel two-stage training procedure, namely representation learning and adversarial fine-tuning. We show that using a post-training analysis of the latent space allows a direct control between the reconstruction fidelity and the representation compactness. By leveraging a multi-band decomposition of the raw waveform, we show that our model is the first able to generate 48kHz audio signals, while simultaneously running 20 times faster than real-time on a standard laptop CPU. We evaluate synthesis quality using both quantitative and qualitative subjective experiments and show the superiority of our approach compared to existing models. Finally, we present applications of our model for timbre transfer and signal compression. All of our source code and audio examples are publicly available.

End-to-end Spoken Language Understanding (SLU) is proposed to infer the semantic meaning directly from audio features without intermediate text representation. Although the acoustic model component of an end-to-end SLU system can be pre-trained with Automatic Speech Recognition (ASR) targets, the SLU component can only learn semantic features from limited task-specific training data. In this paper, for the first time we propose to do large-scale unsupervised pre-training for the SLU component of an end-to-end SLU system, so that the SLU component may preserve semantic features from massive unlabeled audio data. As the output of the acoustic model component, i.e. phoneme posterior sequences, has much different characteristic from text sequences, we propose a novel pre-training model called BERT-PLM, which stands for Bidirectional Encoder Representations from Transformers through Permutation Language Modeling. BERT-PLM trains the SLU component on unlabeled data through a regression objective equivalent to the partial permutation language modeling objective, while leverages full bi-directional context information with BERT networks. The experiment results show that our approach out-perform the state-of-the-art end-to-end systems with over 12.5% error reduction.

Text-to-audio (TTA) generation can significantly benefit the media industry by reducing production costs and enhancing work efficiency. However, most current TTA models (primarily diffusion-based) suffer from slow inference speeds and high computational costs. In this paper, we introduce AudioGAN, the first successful Generative Adversarial Networks (GANs)-based TTA framework that generates audio in a single pass, thereby reducing model complexity and inference time. To overcome the inherent difficulties in training GANs, we integrate multiple ,contrastive losses and propose innovative components Single-Double-Triple (SDT) Attention and Time-Frequency Cross-Attention (TF-CA). Extensive experiments on the AudioCaps dataset demonstrate that AudioGAN achieves state-of-the-art performance while using 90% fewer parameters and running 20 times faster, synthesizing audio in under one second. These results establish AudioGAN as a practical and powerful solution for real-time TTA.

Recent work has studied text-to-audio synthesis using large amounts of paired text-audio data. However, audio recordings with high-quality text annotations can be difficult to acquire. In this work, we approach text-to-audio synthesis using unlabeled videos and pretrained language-vision models. We propose to learn the desired text-audio correspondence by leveraging the visual modality as a bridge. We train a conditional diffusion model to generate the audio track of a video, given a video frame encoded by a pretrained contrastive language-image pretraining (CLIP) model. At test time, we first explore performing a zero-shot modality transfer and condition the diffusion model with a CLIP-encoded text query. However, we observe a noticeable performance drop with respect to image queries. To close this gap, we further adopt a pretrained diffusion prior model to generate a CLIP image embedding given a CLIP text embedding. Our results show the effectiveness of the proposed method, and that the pretrained diffusion prior can reduce the modality transfer gap. While we focus on text-to-audio synthesis, the proposed model can also generate audio from image queries, and it shows competitive performance against a state-of-the-art image-to-audio synthesis model in a subjective listening test. This study offers a new direction of approaching text-to-audio synthesis that leverages the naturally-occurring audio-visual correspondence in videos and the power of pretrained language-vision models.

Joint audio-video generation aims to synthesize realistic audio-video pairs that are both semantically aligned with text prompts and precisely synchronized. While existing joint audio-video generation models often require substantial training resources to improve fidelity, Inference-Time Scaling (ITS) has recently emerged as a promising training-free alternative in single-modality domains. However, extending ITS from a single modality to multimodal domains is non-trivial, as it requires balancing multiple heterogeneous objectives. In this paper, we present the first comprehensive study of ITS for joint audio-video generation. We first demonstrate that a multi-verifier framework is essential to address the limitations of single-objective guidance, including asymmetric performance trade-offs and verifier hacking. Through systematic analysis, we then identify an optimal multi-verifier combination that yields balanced improvements across all quality dimensions. Finally, to effectively aggregate diverse reward signals, we propose Adaptive Reward Weighting (ARW), a novel test-time optimization algorithm. ARW treats reward aggregation as an online optimization problem, utilizing learnable parameters to calibrate reward variances without requiring prior knowledge of reward distributions, thereby ensuring robust multi-objective selection. Experimental results on VGGSound and JavisBench-mini benchmarks demonstrate that our framework significantly enhances semantic alignment, perceptual quality, and audio-visual synchronization of generated outputs. Synthesized samples and code are available on the project page: https://jung-jaemin.github.io/ITS-AVGen-Proj.

This paper presents EGSTalker, a real-time audio-driven talking head generation framework based on 3D Gaussian Splatting (3DGS). Designed to enhance both speed and visual fidelity, EGSTalker requires only 3-5 minutes of training video to synthesize high-quality facial animations. The framework comprises two key stages: static Gaussian initialization and audio-driven deformation. In the first stage, a multi-resolution hash triplane and a Kolmogorov-Arnold Network (KAN) are used to extract spatial features and construct a compact 3D Gaussian representation. In the second stage, we propose an Efficient Spatial-Audio Attention (ESAA) module to fuse audio and spatial cues, while KAN predicts the corresponding Gaussian deformations. Extensive experiments demonstrate that EGSTalker achieves rendering quality and lip-sync accuracy comparable to state-of-the-art methods, while significantly outperforming them in inference speed. These results highlight EGSTalker's potential for real-time multimedia applications.

Person-generic audio-driven face generation is a challenging task in computer vision. Previous methods have achieved remarkable progress in audio-visual synchronization, but there is still a significant gap between current results and practical applications. The challenges are two-fold: 1) Preserving unique individual traits for achieving high-precision lip synchronization. 2) Generating high-quality facial renderings in real-time performance. In this paper, we propose a novel generalized audio-driven framework RealTalk, which consists of an audio-to-expression transformer and a high-fidelity expression-to-face renderer. In the first component, we consider both identity and intra-personal variation features related to speaking lip movements. By incorporating cross-modal attention on the enriched facial priors, we can effectively align lip movements with audio, thus attaining greater precision in expression prediction. In the second component, we design a lightweight facial identity alignment (FIA) module which includes a lip-shape control structure and a face texture reference structure. This novel design allows us to generate fine details in real-time, without depending on sophisticated and inefficient feature alignment modules. Our experimental results, both quantitative and qualitative, on public datasets demonstrate the clear advantages of our method in terms of lip-speech synchronization and generation quality. Furthermore, our method is efficient and requires fewer computational resources, making it well-suited to meet the needs of practical applications.

While Large Audio Language Models (LALMs) achieve strong performance on short audio, they degrade on long-form inputs. This degradation is more severe in temporal awareness tasks, where temporal alignment becomes increasingly inaccurate as audio duration grows. We attribute these limitations to the lack of data, benchmarks, and modeling approaches tailored for long-form temporal awareness. To bridge this gap, we first construct LAT-Chronicle, a 1.2k hour long-form audio dataset with temporal annotations across real-world scenarios. We further develop LAT-Bench, the first human-verified benchmark supporting audio up to 30 minutes while covering three core tasks: Dense Audio Caption, Temporal Audio Grounding, and Targeted Audio Caption. Leveraging these resources, we propose LAT-Audio, formulating temporal awareness as a progressive global-to-local reasoning paradigm. A global timeline is first constructed as an aligned temporal-semantic context,and the Think-With-Audio Chain-of-Thought (TWA-CoT) is then introduced to perform iterative reasoning by incorporating local audio information via tool use. Experiments show that LAT-Audio surpasses existing models on long-form audio temporal awareness tasks and improves robustness to input duration. We release the dataset, benchmark, and model to facilitate future research at https://github.com/alanshaoTT/LAT-Audio-Repo.

Deploying massive diffusion models for real-time, infinite-duration, audio-driven avatar generation presents a significant engineering challenge, primarily due to the conflict between computational load and strict latency constraints. Existing approaches often compromise visual fidelity by enforcing strictly unidirectional attention mechanisms or reducing model capacity. To address this problem, we introduce SoulX-FlashTalk, a 14B-parameter framework optimized for high-fidelity real-time streaming. Diverging from conventional unidirectional paradigms, we use a Self-correcting Bidirectional Distillation strategy that retains bidirectional attention within video chunks. This design preserves critical spatiotemporal correlations, significantly enhancing motion coherence and visual detail. To ensure stability during infinite generation, we incorporate a Multi-step Retrospective Self-Correction Mechanism, enabling the model to autonomously recover from accumulated errors and preventing collapse. Furthermore, we engineered a full-stack inference acceleration suite incorporating hybrid sequence parallelism, Parallel VAE, and kernel-level optimizations. Extensive evaluations confirm that SoulX-FlashTalk is the first 14B-scale system to achieve a sub-second start-up latency (0.87s) while reaching a real-time throughput of 32 FPS, setting a new standard for high-fidelity interactive digital human synthesis.

Recent years have witnessed remarkable advances in talking head generation, owing to its potential to revolutionize the human-AI interaction from text interfaces into realistic video chats. However, research on text-driven talking heads remains underexplored, with existing methods predominantly adopting a cascaded pipeline that combines TTS systems with audio-driven talking head models. This conventional pipeline not only introduces system complexity and latency overhead but also fundamentally suffers from asynchronous audiovisual output and stylistic discrepancies between generated speech and visual expressions. To address these limitations, we introduce OmniTalker, an end-to-end unified framework that simultaneously generates synchronized speech and talking head videos from text and reference video in real-time zero-shot scenarios, while preserving both speech style and facial styles. The framework employs a dual-branch diffusion transformer architecture: the audio branch synthesizes mel-spectrograms from text, while the visual branch predicts fine-grained head poses and facial dynamics. To bridge modalities, we introduce a novel audio-visual fusion module that integrates cross-modal information to ensure temporal synchronization and stylistic coherence between audio and visual outputs. Furthermore, our in-context reference learning module effectively captures both speech and facial style characteristics from a single reference video without introducing an extra style extracting module. To the best of our knowledge, OmniTalker presents the first unified framework that jointly models speech style and facial style in a zero-shot setting, achieving real-time inference speed of 25 FPS. Extensive experiments demonstrate that our method surpasses existing approaches in generation quality, particularly excelling in style preservation and audio-video synchronization.

We consider the task of generating diverse and realistic videos guided by natural audio samples from a wide variety of semantic classes. For this task, the videos are required to be aligned both globally and temporally with the input audio: globally, the input audio is semantically associated with the entire output video, and temporally, each segment of the input audio is associated with a corresponding segment of that video. We utilize an existing text-conditioned video generation model and a pre-trained audio encoder model. The proposed method is based on a lightweight adaptor network, which learns to map the audio-based representation to the input representation expected by the text-to-video generation model. As such, it also enables video generation conditioned on text, audio, and, for the first time as far as we can ascertain, on both text and audio. We validate our method extensively on three datasets demonstrating significant semantic diversity of audio-video samples and further propose a novel evaluation metric (AV-Align) to assess the alignment of generated videos with input audio samples. AV-Align is based on the detection and comparison of energy peaks in both modalities. In comparison to recent state-of-the-art approaches, our method generates videos that are better aligned with the input sound, both with respect to content and temporal axis. We also show that videos produced by our method present higher visual quality and are more diverse.

Despite impressive advancements in video understanding, most efforts remain limited to coarse-grained or visual-only video tasks. However, real-world videos encompass omni-modal information (vision, audio, and speech) with a series of events forming a cohesive storyline. The lack of multi-modal video data with fine-grained event annotations and the high cost of manual labeling are major obstacles to comprehensive omni-modality video perception. To address this gap, we propose an automatic pipeline consisting of high-quality multi-modal video filtering, semantically coherent omni-modal event boundary detection, and cross-modal correlation-aware event captioning. In this way, we present LongVALE, the first-ever Vision-Audio-Language Event understanding benchmark comprising 105K omni-modal events with precise temporal boundaries and detailed relation-aware captions within 8.4K high-quality long videos. Further, we build a baseline that leverages LongVALE to enable video large language models (LLMs) for omni-modality fine-grained temporal video understanding for the first time. Extensive experiments demonstrate the effectiveness and great potential of LongVALE in advancing comprehensive multi-modal video understanding.

Large-scale multimodal generative modeling has created milestones in text-to-image and text-to-video generation. Its application to audio still lags behind for two main reasons: the lack of large-scale datasets with high-quality text-audio pairs, and the complexity of modeling long continuous audio data. In this work, we propose Make-An-Audio with a prompt-enhanced diffusion model that addresses these gaps by 1) introducing pseudo prompt enhancement with a distill-then-reprogram approach, it alleviates data scarcity with orders of magnitude concept compositions by using language-free audios; 2) leveraging spectrogram autoencoder to predict the self-supervised audio representation instead of waveforms. Together with robust contrastive language-audio pretraining (CLAP) representations, Make-An-Audio achieves state-of-the-art results in both objective and subjective benchmark evaluation. Moreover, we present its controllability and generalization for X-to-Audio with "No Modality Left Behind", for the first time unlocking the ability to generate high-definition, high-fidelity audios given a user-defined modality input. Audio samples are available at https://Text-to-Audio.github.io

This paper introduces FlowMAC, a novel neural audio codec for high-quality general audio compression at low bit rates based on conditional flow matching (CFM). FlowMAC jointly learns a mel spectrogram encoder, quantizer and decoder. At inference time the decoder integrates a continuous normalizing flow via an ODE solver to generate a high-quality mel spectrogram. This is the first time that a CFM-based approach is applied to general audio coding, enabling a scalable, simple and memory efficient training. Our subjective evaluations show that FlowMAC at 3 kbps achieves similar quality as state-of-the-art GAN-based and DDPM-based neural audio codecs at double the bit rate. Moreover, FlowMAC offers a tunable inference pipeline, which permits to trade off complexity and quality. This enables real-time coding on CPU, while maintaining high perceptual quality.

To the best of our knowledge, we first present a live system that generates personalized photorealistic talking-head animation only driven by audio signals at over 30 fps. Our system contains three stages. The first stage is a deep neural network that extracts deep audio features along with a manifold projection to project the features to the target person's speech space. In the second stage, we learn facial dynamics and motions from the projected audio features. The predicted motions include head poses and upper body motions, where the former is generated by an autoregressive probabilistic model which models the head pose distribution of the target person. Upper body motions are deduced from head poses. In the final stage, we generate conditional feature maps from previous predictions and send them with a candidate image set to an image-to-image translation network to synthesize photorealistic renderings. Our method generalizes well to wild audio and successfully synthesizes high-fidelity personalized facial details, e.g., wrinkles, teeth. Our method also allows explicit control of head poses. Extensive qualitative and quantitative evaluations, along with user studies, demonstrate the superiority of our method over state-of-the-art techniques.

We present a novel multi-modal unspoken punctuation prediction system for the English language which combines acoustic and text features. We demonstrate for the first time, that by relying exclusively on synthetic data generated using a prosody-aware text-to-speech system, we can outperform a model trained with expensive human audio recordings on the unspoken punctuation prediction problem. Our model architecture is well suited for on-device use. This is achieved by leveraging hash-based embeddings of automatic speech recognition text output in conjunction with acoustic features as input to a quasi-recurrent neural network, keeping the model size small and latency low.

Understanding and reasoning over non-speech sounds and music are crucial for both humans and AI agents to interact effectively with their environments. In this paper, we introduce Audio Flamingo 2 (AF2), an Audio-Language Model (ALM) with advanced audio understanding and reasoning capabilities. AF2 leverages (i) a custom CLAP model, (ii) synthetic Audio QA data for fine-grained audio reasoning, and (iii) a multi-stage curriculum learning strategy. AF2 achieves state-of-the-art performance with only a 3B parameter small language model, surpassing large open-source and proprietary models across over 20 benchmarks. Next, for the first time, we extend audio understanding to long audio segments (30 secs to 5 mins) and propose LongAudio, a large and novel dataset for training ALMs on long audio captioning and question-answering tasks. Fine-tuning AF2 on LongAudio leads to exceptional performance on our proposed LongAudioBench, an expert annotated benchmark for evaluating ALMs on long audio understanding capabilities. We conduct extensive ablation studies to confirm the efficacy of our approach. Project Website: https://research.nvidia.com/labs/adlr/AF2/.

With the growing requirement for natural human-computer interaction, speech-based systems receive increasing attention as speech is one of the most common forms of daily communication. However, the existing speech models still experience high latency when generating the first audio token during streaming, which poses a significant bottleneck for deployment. To address this issue, we propose VITA-Audio, an end-to-end large speech model with fast audio-text token generation. Specifically, we introduce a lightweight Multiple Cross-modal Token Prediction (MCTP) module that efficiently generates multiple audio tokens within a single model forward pass, which not only accelerates the inference but also significantly reduces the latency for generating the first audio in streaming scenarios. In addition, a four-stage progressive training strategy is explored to achieve model acceleration with minimal loss of speech quality. To our knowledge, VITA-Audio is the first multi-modal large language model capable of generating audio output during the first forward pass, enabling real-time conversational capabilities with minimal latency. VITA-Audio is fully reproducible and is trained on open-source data only. Experimental results demonstrate that our model achieves an inference speedup of 3~5x at the 7B parameter scale, but also significantly outperforms open-source models of similar model size on multiple benchmarks for automatic speech recognition (ASR), text-to-speech (TTS), and spoken question answering (SQA) tasks.

Large Audio-Language Models (LALMs) are becoming essential as a powerful multimodal backbone for real-world applications. However, recent studies show that audio inputs can more easily elicit harmful responses than text, exposing new risks toward deployment. While safety alignment has made initial advances in LLMs and Large Vision-Language Models (LVLMs), we find that vanilla adaptation of these approaches to LALMs faces two key limitations: 1) LLM-based steering fails under audio input due to the large distributional gap between activations, and 2) prompt-based defenses induce over-refusals on benign-speech queries. To address these challenges, we propose Safe-Ablated Refusal Steering (SARSteer), the first inference-time defense framework for LALMs. Specifically, SARSteer leverages text-derived refusal steering to enforce rejection without manipulating audio inputs and introduces decomposed safe-space ablation to mitigate over-refusal. Extensive experiments demonstrate that SARSteer significantly improves harmful-query refusal while preserving benign responses, establishing a principled step toward safety alignment in LALMs.

In traditional audio captioning methods, a model is usually trained in a fully supervised manner using a human-annotated dataset containing audio-text pairs and then evaluated on the test sets from the same dataset. Such methods have two limitations. First, these methods are often data-hungry and require time-consuming and expensive human annotations to obtain audio-text pairs. Second, these models often suffer from performance degradation in cross-domain scenarios, i.e., when the input audio comes from a different domain than the training set, which, however, has received little attention. We propose an effective audio captioning method based on the contrastive language-audio pre-training (CLAP) model to address these issues. Our proposed method requires only textual data for training, enabling the model to generate text from the textual feature in the cross-modal semantic space.In the inference stage, the model generates the descriptive text for the given audio from the audio feature by leveraging the audio-text alignment from CLAP.We devise two strategies to mitigate the discrepancy between text and audio embeddings: a mixed-augmentation-based soft prompt and a retrieval-based acoustic-aware hard prompt. These approaches are designed to enhance the generalization performance of our proposed model, facilitating the model to generate captions more robustly and accurately. Extensive experiments on AudioCaps and Clotho benchmarks show the effectiveness of our proposed method, which outperforms other zero-shot audio captioning approaches for in-domain scenarios and outperforms the compared methods for cross-domain scenarios, underscoring the generalization ability of our method.

The sound effects that designers add to videos are designed to convey a particular artistic effect and, thus, may be quite different from a scene's true sound. Inspired by the challenges of creating a soundtrack for a video that differs from its true sound, but that nonetheless matches the actions occurring on screen, we propose the problem of conditional Foley. We present the following contributions to address this problem. First, we propose a pretext task for training our model to predict sound for an input video clip using a conditional audio-visual clip sampled from another time within the same source video. Second, we propose a model for generating a soundtrack for a silent input video, given a user-supplied example that specifies what the video should "sound like". We show through human studies and automated evaluation metrics that our model successfully generates sound from video, while varying its output according to the content of a supplied example. Project site: https://xypb.github.io/CondFoleyGen/

Open-vocabulary audio language models (ALMs), like Contrastive Language Audio Pretraining (CLAP), represent a promising new paradigm for audio-text retrieval using natural language queries. In this paper, for the first time, we perform controlled experiments on various benchmarks to show that existing ALMs struggle to generalize to linguistic variations in textual queries. To address this issue, we propose RobustCLAP, a novel and compute-efficient technique to learn audio-language representations agnostic to linguistic variations. Specifically, we reformulate the contrastive loss used in CLAP architectures by introducing a multi-view contrastive learning objective, where paraphrases are treated as different views of the same audio scene and use this for training. Our proposed approach improves the text-to-audio retrieval performance of CLAP by 0.8%-13% across benchmarks and enhances robustness to linguistic variation.

In recent years, DeepFake technology has achieved unprecedented success in high-quality video synthesis, but these methods also pose potential and severe security threats to humanity. DeepFake can be bifurcated into entertainment applications like face swapping and illicit uses such as lip-syncing fraud. However, lip-forgery videos, which neither change identity nor have discernible visual artifacts, present a formidable challenge to existing DeepFake detection methods. Our preliminary experiments have shown that the effectiveness of the existing methods often drastically decrease or even fail when tackling lip-syncing videos. In this paper, for the first time, we propose a novel approach dedicated to lip-forgery identification that exploits the inconsistency between lip movements and audio signals. We also mimic human natural cognition by capturing subtle biological links between lips and head regions to boost accuracy. To better illustrate the effectiveness and advances of our proposed method, we create a high-quality LipSync dataset, AVLips, by employing the state-of-the-art lip generators. We hope this high-quality and diverse dataset could be well served the further research on this challenging and interesting field. Experimental results show that our approach gives an average accuracy of more than 95.3% in spotting lip-syncing videos, significantly outperforming the baselines. Extensive experiments demonstrate the capability to tackle deepfakes and the robustness in surviving diverse input transformations. Our method achieves an accuracy of up to 90.2% in real-world scenarios (e.g., WeChat video call) and shows its powerful capabilities in real scenario deployment. To facilitate the progress of this research community, we release all resources at https://github.com/AaronComo/LipFD.

Diffusion-based lip synchronization models achieve strong visual quality and audio-visual alignment, but full-sequence bidirectional attention and many denoising steps make them impractical for real-time inference. We present Lip Forcing, to our knowledge the first autoregressive diffusion method for video-to-video (V2V) lip synchronization, which distills a 14B audio-conditioned bidirectional video diffusion teacher into causal students. At inference, the students generate each chunk in only two denoising steps without inference-time CFG, enabling real-time lip synchronization. A lip-sync-specific teacher-trajectory analysis reveals a CFG fidelity-sync tradeoff: no-CFG predictions favor reference fidelity, whereas CFG-guided predictions favor synchronization within a mid-trajectory band. Lip Forcing translates this finding into three analysis-derived components: Sync-Window DMD, a two-step inference schedule, and a SyncNet-based reward. We validate Lip Forcing at two student scales, both distilled from the 14B teacher. The 1.3B student crosses into real-time streaming at 31 FPS, 17.6times faster than its same-scale bidirectional model. The 14B student, the largest diffusion model reported for V2V lip synchronization, runs 39.8times faster than its teacher at comparable reference fidelity. Time-to-first-frame is sub-millisecond at both scales, far below every diffusion baseline.

Spoken dialogue is a primary source of information in videos; therefore, accurately identifying who spoke what and when is essential for deep video understanding. We introduce D-ORCA, a dialogue-centric omni-modal large language model optimized for robust audio-visual captioning. We further curate DVD, a large-scale, high-quality bilingual dataset comprising nearly 40,000 multi-party dialogue videos for training and 2000 videos for evaluation in English and Mandarin, addressing a critical gap in the open-source ecosystem. To ensure fine-grained captioning accuracy, we adopt group relative policy optimization with three novel reward functions that assess speaker attribution accuracy, global speech content accuracy, and sentence-level temporal boundary alignment. These rewards are derived from evaluation metrics widely used in speech processing and, to our knowledge, are applied for the first time as reinforcement learning objectives for audio-visual captioning. Extensive experiments demonstrate that D-ORCA substantially outperforms existing open-source models in speaker identification, speech recognition, and temporal grounding. Notably, despite having only 8 billion parameters, D-ORCA achieves performance competitive with Qwen3-Omni across several general-purpose audio-visual understanding benchmarks. Demos are available at https://d-orca-llm.github.io/{https://d-orca-llm.github.io/}. Our code, data, and checkpoints will be available at https://github.com/WeChatCV/D-ORCA/{https://github.com/WeChatCV/D-ORCA/}.

Context-sensitive two-point layer 5 pyramidal cells (L5PCs) were discovered as long ago as 1999. However, the potential of this discovery to provide useful neural computation has yet to be demonstrated. Here we show for the first time how a transformative L5PCs-driven deep neural network (DNN), termed the multisensory cooperative computing (MCC) architecture, can effectively process large amounts of heterogeneous real-world audio-visual (AV) data, using far less energy compared to best available 'point' neuron-driven DNNs. A novel highly-distributed parallel implementation on a Xilinx UltraScale+ MPSoC device estimates energy savings up to 245759 times 50000 muJ (i.e., 62% less than the baseline model in a semi-supervised learning setup) where a single synapse consumes 8e^{-5}muJ. In a supervised learning setup, the energy-saving can potentially reach up to 1250x less (per feedforward transmission) than the baseline model. The significantly reduced neural activity in MCC leads to inherently fast learning and resilience against sudden neural damage. This remarkable performance in pilot experiments demonstrates the embodied neuromorphic intelligence of our proposed cooperative L5PC that receives input from diverse neighbouring neurons as context to amplify the transmission of most salient and relevant information for onward transmission, from overwhelmingly large multimodal information utilised at the early stages of on-chip training. Our proposed approach opens new cross-disciplinary avenues for future on-chip DNN training implementations and posits a radical shift in current neuromorphic computing paradigms.

Motion retrieval is crucial for motion acquisition, offering superior precision, realism, controllability, and editability compared to motion generation. Existing approaches leverage contrastive learning to construct a unified embedding space for motion retrieval from text or visual modality. However, these methods lack a more intuitive and user-friendly interaction mode and often overlook the sequential representation of most modalities for improved retrieval performance. To address these limitations, we propose a framework that aligns four modalities -- text, audio, video, and motion -- within a fine-grained joint embedding space, incorporating audio for the first time in motion retrieval to enhance user immersion and convenience. This fine-grained space is achieved through a sequence-level contrastive learning approach, which captures critical details across modalities for better alignment. To evaluate our framework, we augment existing text-motion datasets with synthetic but diverse audio recordings, creating two multi-modal motion retrieval datasets. Experimental results demonstrate superior performance over state-of-the-art methods across multiple sub-tasks, including an 10.16% improvement in R@10 for text-to-motion retrieval and a 25.43% improvement in R@1 for video-to-motion retrieval on the HumanML3D dataset. Furthermore, our results show that our 4-modal framework significantly outperforms its 3-modal counterpart, underscoring the potential of multi-modal motion retrieval for advancing motion acquisition.

Historically, neuroscience has progressed by fragmenting into specialized domains, each focusing on isolated modalities, tasks, or brain regions. While fruitful, this approach hinders the development of a unified model of cognition. Here, we introduce TRIBE, the first deep neural network trained to predict brain responses to stimuli across multiple modalities, cortical areas and individuals. By combining the pretrained representations of text, audio and video foundational models and handling their time-evolving nature with a transformer, our model can precisely model the spatial and temporal fMRI responses to videos, achieving the first place in the Algonauts 2025 brain encoding competition with a significant margin over competitors. Ablations show that while unimodal models can reliably predict their corresponding cortical networks (e.g. visual or auditory networks), they are systematically outperformed by our multimodal model in high-level associative cortices. Currently applied to perception and comprehension, our approach paves the way towards building an integrative model of representations in the human brain. Our code is available at https://github.com/facebookresearch/algonauts-2025.

We consider the generative modeling of speech over multiple minutes, a requirement for long-form multimedia generation and audio-native voice assistants. However, current spoken language models struggle to generate plausible speech past tens of seconds, from high temporal resolution of speech tokens causing loss of coherence, to architectural issues with long-sequence training or extrapolation, to memory costs at inference time. With these considerations we propose SpeechSSM, the first speech language model to learn from and sample long-form spoken audio (e.g., 16 minutes of read or extemporaneous speech) in a single decoding session without text intermediates, based on recent advances in linear-time sequence modeling. Furthermore, to address growing challenges in spoken language evaluation, especially in this new long-form setting, we propose: new embedding-based and LLM-judged metrics; quality measurements over length and time; and a new benchmark for long-form speech processing and generation, LibriSpeech-Long. Speech samples and the dataset are released at https://google.github.io/tacotron/publications/speechssm/

We present Qwen3-Omni, a single multimodal model that, for the first time, maintains state-of-the-art performance across text, image, audio, and video without any degradation relative to single-modal counterparts. Qwen3-Omni matches the performance of same-sized single-modal models within the Qwen series and excels particularly on audio tasks. Across 36 audio and audio-visual benchmarks, Qwen3-Omni achieves open-source SOTA on 32 benchmarks and overall SOTA on 22, outperforming strong closed-source models such as Gemini-2.5-Pro, Seed-ASR, and GPT-4o-Transcribe. Qwen3-Omni adopts a Thinker-Talker MoE architecture that unifies perception and generation across text, images, audio, and video, yielding fluent text and natural real-time speech. It supports text interaction in 119 languages, speech understanding in 19 languages, and speech generation in 10 languages. To reduce first-packet latency in streaming synthesis, Talker autoregressively predicts discrete speech codecs using a multi-codebook scheme. Leveraging the representational capacity of these codebooks, we replace computationally intensive block-wise diffusion with a lightweight causal ConvNet, enabling streaming from the first codec frame. In cold-start settings, Qwen3-Omni achieves a theoretical end-to-end first-packet latency of 234 ms. To further strengthen multimodal reasoning, we introduce a Thinking model that explicitly reasons over inputs from any modality. Since the research community currently lacks a general-purpose audio captioning model, we fine-tuned Qwen3-Omni-30B-A3B to obtain Qwen3-Omni-30B-A3B-Captioner, which produces detailed, low-hallucination captions for arbitrary audio inputs. Qwen3-Omni-30B-A3B, Qwen3-Omni-30B-A3B-Thinking, and Qwen3-Omni-30B-A3B-Captioner are publicly released under the Apache 2.0 license.

The Model Context Protocol (MCP) defines a schema bound execution model for agent-tool interaction, enabling modular computer vision workflows without retraining. To our knowledge, this is the first protocol level, deployment scale audit of MCP in vision systems, identifying systemic weaknesses in schema semantics, interoperability, and runtime coordination. We analyze 91 publicly registered vision centric MCP servers, annotated along nine dimensions of compositional fidelity, and develop an executable benchmark with validators to detect and categorize protocol violations. The audit reveals high prevalence of schema format divergence, missing runtime schema validation, undeclared coordinate conventions, and reliance on untracked bridging scripts. Validator based testing quantifies these failures, with schema format checks flagging misalignments in 78.0 percent of systems, coordinate convention checks detecting spatial reference errors in 24.6 percent, and memory scope checks issuing an average of 33.8 warnings per 100 executions. Security probes show that dynamic and multi agent workflows exhibit elevated risks of privilege escalation and untyped tool connections. The proposed benchmark and validator suite, implemented in a controlled testbed and to be released on GitHub, establishes a reproducible framework for measuring and improving the reliability and security of compositional vision workflows.

Spiking Neural Networks (SNNs), with their event-driven and biologically inspired mechanisms, are well-suited for energy-efficient neuromorphic hardware. Neural coding, which is critical to SNNs, determines how information is represented via spikes. While Time-to-First-Spike (TTFS) coding uses a single spike per neuron to offer extreme sparsity and energy efficiency, it often suffers from unstable training and low accuracy due to its sparse firing. To address these challenges, we propose a training framework that incorporates parameter initialization, training normalization, a temporal output decoder, and a re-evaluation of the pooling layer. The proposed parameter initialization and training normalization mitigate signal diminishing and gradient vanishing, which helps stabilize training. Our output decoder aggregates temporal spikes to encourage earlier firing, thereby reducing latency. The re-evaluation of the pooling layer demonstrates that max-pooling violates single-spike constraints, which should be avoided, whereas average-pooling preserves them. Experiments show that our framework stabilizes and accelerates training, reduces latency, and achieves state-of-the-art accuracy for step-by-step TTFS SNNs on MNIST (99.48%), Fashion-MNIST (92.90%), CIFAR10 (90.56%), CIFAR100 (70.27%) and DVS Gesture (95.83%).

Inference with transformer-based language models begins with a prompt processing step. In this step, the model generates the first output token and stores the KV cache needed for future generation steps. This prompt processing step can be computationally expensive, taking 10s of seconds or more for billion-parameter models on edge devices when prompt lengths or batch sizes rise. This degrades user experience by introducing significant latency into the model's outputs. To reduce the time spent producing the first output (known as the ``time to first token'', or TTFT) of a pretrained model, we introduce a novel method called KV Prediction. In our method, a small auxiliary model is used to process the prompt and produce an approximation of the KV cache used by a base model. This approximated KV cache is then used with the base model for autoregressive generation without the need to query the auxiliary model again. We demonstrate that our method produces a pareto-optimal efficiency-accuracy trade-off when compared to baselines. On TriviaQA, we demonstrate relative accuracy improvements in the range of 15%-50% across a range of TTFT FLOPs budgets. We also demonstrate accuracy improvements of up to 30% on HumanEval python code completion at fixed TTFT FLOPs budgets. Additionally, we benchmark models on an Apple M2 Pro CPU and demonstrate that our improvement in FLOPs translates to a TTFT speedup on hardware. We release our code at https://github.com/apple/corenet/tree/main/projects/kv-prediction .

Context retrieval systems for LLM inference face a critical challenge: high retrieval latency creates a fundamental tension between waiting for complete context (poor time-to-first-token) and proceeding without it (reduced quality). Streaming context incrementally--overlapping retrieval with inference--can mitigate this latency, but doing so with concurrent requests introduces new challenges: requests contend for GPU compute and memory, and scheduling must adapt to dynamic context arrivals. We present Stream2LLM, a streaming-aware LLM serving system for concurrent prefill-decode disaggregated deployments. Stream2LLM introduces adaptive scheduling and preemption for two distinct retrieval patterns: append-mode (progressive context accumulation) and update-mode (iterative refinement with cache invalidation). It decouples scheduling decisions from resource acquisition, enabling flexible preemption strategies guided by hardware-specific cost models, and uses longest common prefix matching to minimize redundant computation when input changes dynamically. To evaluate Stream2LLM, we collect two large-scale, real-world streaming workloads based on web crawling and approximate nearest neighbor search. Our evaluation demonstrates that streaming architecture delivers up to 11x TTFT improvements, with cost-aware scheduling providing critical benefits under memory pressure, all while maintaining throughput parity with non-streaming baselines. Code: https://github.com/rajveerb/stream2llm/tree/mlsys_artifact

Retrieval-Augmented Generation enhances Large Language Models by integrating external knowledge, which reduces hallucinations but increases prompt length. This increase leads to higher computational costs and longer Time to First Token (TTFT). To mitigate this issue, existing solutions aim to reuse the preprocessed KV cache of each retrieved chunk to accelerate RAG. However, the lack of cross-chunk contextual information leads to a significant drop in generation quality, leaving the potential benefits of KV cache reuse largely unfulfilled. The challenge lies in how to reuse the precomputed KV cache of chunks while preserving generation quality. We propose FusionRAG, a novel inference framework that optimizes both the preprocessing and reprocessing stages of RAG. In the offline preprocessing stage, we embed information from other related text chunks into each chunk, while in the online reprocessing stage, we recompute the KV cache for tokens that the model focuses on. As a result, we achieve a better trade-off between generation quality and efficiency. According to our experiments, FusionRAG significantly improves generation quality at the same recomputation ratio compared to previous state-of-the-art solutions. By recomputing fewer than 15% of the tokens, FusionRAG achieves up to 70% higher normalized F1 scores than baselines and reduces TTFT by 2.66x-9.39x compared to Full Attention.

Real-time execution is crucial for deploying Vision-Language-Action (VLA) models in the physical world. Existing asynchronous inference methods primarily optimize trajectory smoothness, but neglect the critical latency in reacting to environmental changes. By rethinking the notion of reaction in action chunking policies, this paper presents a systematic analysis of the factors governing reaction time. We show that reaction time follows a uniform distribution determined jointly by the Time to First Action (TTFA) and the execution horizon. Moreover, we reveal that the standard practice of applying a constant schedule in flow-based VLAs can be inefficient and forces the system to complete all sampling steps before any movement can start, forming the bottleneck in reaction latency. To overcome this issue, we propose Fast Action Sampling for ImmediaTE Reaction (FASTER). By introducing a Horizon-Aware Schedule, FASTER adaptively prioritizes near-term actions during flow sampling, compressing the denoising of the immediate reaction by tenfold (e.g., in π_{0.5} and X-VLA) into a single step, while preserving the quality of long-horizon trajectory. Coupled with a streaming client-server pipeline, FASTER substantially reduces the effective reaction latency on real robots, especially when deployed on consumer-grade GPUs. Real-world experiments, including a highly dynamic table tennis task, prove that FASTER unlocks unprecedented real-time responsiveness for generalist policies, enabling rapid generation of accurate and smooth trajectories.

Large Language Model (LLM) inference in production must meet stringent service-level objectives for both time-to-first-token (TTFT) and time-between-token (TBT) while maximizing throughput under fixed compute, memory, and interconnect budgets. Modern serving systems adopt stall-free scheduling techniques such as chunked prefill, which splits long prompt processing along the token dimension and interleaves prefill with ongoing decode iterations. While effective at stabilizing TBT, chunked prefill incurs substantial overhead in Mixture-of-Experts (MoE) models: redundant expert weight loads increase memory traffic by up to 39% and inflate energy consumption. We propose layered prefill, a new scheduling paradigm that treats transformer layer groups as the primary scheduling unit. By vertically partitioning the model into contiguous layer groups and interleaving prefill and decode across the groups, layered prefill sustains stall-free decoding while eliminating chunk-induced MoE weight reloads. It reduces off-chip bandwidth demand, lowering TTFT by up to 70%, End-to-End latency by 41% and per-token energy by up to 22%. Evaluations show that layered prefill consistently improves the TTFT--TBT Pareto frontier over chunked prefill, reducing expert-load traffic and energy cost while maintaining stall-free decoding. Overall, shifting the scheduling axis from tokens to layers unlocks a new operating regime for high-efficiency, energy-aware LLM serving in co-located environments.

Voice agents face a fundamental tension: the reasoning, retrieval, and tool use that make foundation models capable are iterative and slow, while conversational interaction demands responses on a millisecond timescale. Smaller, real-time models meet the latency bar but cannot match foundation models on complex tasks, leaving current voice agents to trade away either responsiveness or capability. We introduce conversational infill, where a small talker model both immediately generates contextually grounded responses to hide the latency of an external reasoner model and fluently integrates streamed reasoner knowledge into its responses during inference. We curate a 290,571-example synthetic dataset spanning six domains and demonstrate that this task is learnable across seven widely used small language models ranging from 135M to 1.7B parameters. Our system implementation, ConvFill, sustains millisecond-level time-to-first-response while closing the accuracy gap to within 6.3% of the corresponding frontier reasoner performance. In a live user study (n=18) with talker deployments running on an Apple M2 SoC, participants rank ConvFill on par with frontier models overall, prefer it for retrieval-heavy tasks, and rate it significantly more responsive. These results show that conversational infill unlocks a new point on the latency-capability Pareto frontier, offering a practical path toward voice agents that are both responsive and highly capable. Code, models, and datasets are available at https://github.com/vysri/conversational-infill.

Real-time speech synthesis requires balancing inference latency and acoustic fidelity for interactive applications. Conventional continuous text-to-speech pipelines require computationally intensive neural vocoders to reconstruct phase information, creating a significant streaming bottleneck. Furthermore, regression-based acoustic modeling frequently induces spectral over-smoothing artifacts. To address these limitations, this paper proposes a novel end-to-end non-autoregressive architecture optimized for ultra-low latency block-wise generation, directly modeling the highly compressed discrete latent space of the Mimi neural audio codec. Integrating a modified FastSpeech 2 backbone with a progressive depth-wise sequential decoding strategy, the architecture dynamically conditions 32 layers of residual vector quantization codes. This mechanism resolves phonetic alignment degradation and manages the complexity of high-fidelity discrete representations without temporal autoregressive overhead. Experimental evaluations on English and Malay datasets validate its language-independent deployment capability. Compared to conventional continuous regression models, the proposed architecture demonstrates quantitative improvements in fundamental voicing accuracy and mitigates high-frequency spectral degradation. It achieves ultra-low latency inference, translating to a 10.6-fold absolute acceleration over conventional cascaded pipelines. Crucially, the system achieves an average time-to-first-byte latency of 48.99 milliseconds, falling significantly below the human perception threshold for real-time interactive streaming. These results firmly establish the proposed architecture as a highly optimized solution for deploying real-time streaming speech interfaces.

With the emergence of wearable devices and other embedded systems, deploying large language models (LLMs) on edge platforms has become an urgent need. However, this is challenging because of their high computational and memory demands. Although recent low-bit quantization methods (e.g., BitNet, DeepSeek) compress weights to as low as 1.58~bits with minimal accuracy loss, edge deployment is still constrained by limited on-chip resources, power budgets, and the often-neglected long latency of the prefill stage. We present TeLLMe, the first table-lookup-based ternary LLM accelerator for low-power edge FPGAs that fully supports both prefill and autoregressive decoding using 1.58-bit weights and 8-bit activations. TeLLMe incorporates several novel techniques, including (1) a table-lookup-based ternary matrix multiplication (TLMM) engine utilizing grouped activations and online precomputation for low resource utilization and high throughput; (2) a fine-grained analytic URAM-based weight buffer management scheme for efficient loading and compute engine access; (3) a streaming dataflow architecture that fuses floating-point element-wise operations with linear computations to hide latency; (4) a reversed-reordered prefill stage attention with fused attention operations for high memory efficiency; and (5) a resource-efficient specialized decoding stage attention. Under a 5~W power budget, TeLLMe delivers up to 25~tokens/s decoding throughput and 0.45--0.96~s time-to-first-token (TTFT) for 64--128 token prompts, marking a significant energy-efficiency advancement in LLM inference on edge FPGAs.

Video understanding is shifting from the offline paradigm -- taking a fully recorded video as input and producing a single answer after it ends -- toward real-time interaction, in which the model perceives new frames while still replying, revises its answer as new evidence appears, and remains silent when there is nothing to say. We present MOSS-Video-Preview to validate this paradigm. Our central claim is that perception must not be blocked by generation; its natural realization is a two-channel architecture. We argue that a cross-attention backbone is better suited to real-time vision-language fusion than the prevailing decoder-only design: visual features enter through a side channel rather than joining the autoregressive sequence, so perception and generation run on separate, non-blocking pathways -- reducing the frequency of visual processing and exposing a clean channel-wise interface for independent compression. We complement this with a data synthesis pipeline that converts dense captions into real-time understanding QA whose answers are revised to match what the model has perceived so far, and we specialize an offline model on these data to elicit real-time behavior. Our model trails the strong Qwen2.5-VL-7B baseline overall -- a gap we attribute primarily to data and scale rather than the architecture -- yet attains competitive offline video and multimodal understanding, remains robust on the spatial and fine-grained temporal reasoning central to real-time use, and acquires behaviors that offline models lack: continuous perception, answer revision, and timely silence. On a single H200 with 256 frames per video, it achieves about a 5x speedup in time to first token and 2.7x higher decoding throughput, with negligible degradation in offline ability. Our study of paradigm, architecture, and data outlines a viable path toward real-time video understanding.

The growing demand for large language models (LLMs) requires serving systems to handle many concurrent requests with diverse service level objectives (SLOs). This exacerbates head-of-line (HoL) blocking during the compute-intensive prefill phase, where long-running requests monopolize resources and delay higher-priority ones, leading to widespread time-to-first-token (TTFT) SLO violations. While chunked prefill enables interruptibility, it introduces an inherent trade-off between responsiveness and throughput: reducing chunk size improves response latency but degrades computational efficiency, whereas increasing chunk size maximizes throughput but exacerbates blocking. This necessitates an adaptive preemption mechanism. However, dynamically balancing execution granularity against scheduling overheads remains a key challenge. In this paper, we propose FlowPrefill, a TTFT-goodput-optimized serving system that resolves this conflict by decoupling preemption granularity from scheduling frequency. To achieve adaptive prefill scheduling, FlowPrefill introduces two key innovations: 1) Operator-Level Preemption, which leverages operator boundaries to enable fine-grained execution interruption without the efficiency loss associated with fixed small chunking; and 2) Event-Driven Scheduling, which triggers scheduling decisions only upon request arrival or completion events, thereby supporting efficient preemption responsiveness while minimizing control-plane overhead. Evaluation on real-world production traces shows that FlowPrefill improves maximum goodput by up to 5.6times compared to state-of-the-art systems while satisfying heterogeneous SLOs.

Improving time-to-first-token (TTFT) is an essentially important objective in modern large language model (LLM) inference engines. Optimizing TTFT directly results in higher maximal QPS and meets the requirements of many critical applications. However, boosting TTFT is notoriously challenging since it is compute-bounded and the performance bottleneck shifts from the self-attention that many prior works focus on to the MLP part. In this work, we present SpecPrefill, a training free framework that accelerates the inference TTFT for both long and medium context queries based on the following insight: LLMs are generalized enough to preserve the quality given only a carefully chosen subset of prompt tokens. At its core, SpecPrefill leverages a lightweight model to speculate locally important tokens based on the context. These tokens, along with the necessary positional information, are then sent to the main model for processing. We evaluate SpecPrefill with a diverse set of tasks, followed by a comprehensive benchmarking of performance improvement both in a real end-to-end setting and ablation studies. SpecPrefill manages to serve Llama-3.1-405B-Instruct-FP8 with up to 7times maximal end-to-end QPS on real downstream tasks and 7.66times TTFT improvement.

Video Language Models (VideoLMs) empower AI systems to understand temporal dynamics in videos. To fit to the maximum context window constraint, current methods use keyframe sampling which can miss both macro-level events and micro-level details due to the sparse temporal coverage. Furthermore, processing full images and their tokens for each frame incurs substantial computational overhead. To address these limitations, we propose to leverage video codec primitives (specifically motion vectors and residuals) which natively encode video redundancy and sparsity without requiring expensive full-image encoding for most frames. To this end, we introduce lightweight transformer-based encoders that aggregate codec primitives and align their representations with image encoder embeddings through a pre-training strategy that accelerates convergence during end-to-end fine-tuning. Our approach reduces the time-to-first-token by up to 86% and token usage by up to 93% compared to standard VideoLMs. Moreover, by varying the keyframe and codec primitive densities we are able to maintain or exceed performance on 14 diverse video understanding benchmarks spanning general question answering, temporal reasoning, long-form understanding, and spatial scene understanding.

Role specialization in multi-LLM agent systems is often realized via multi-LoRA, where agents share a pretrained backbone and differ only through lightweight adapters. Despite sharing base model weights, each agent independently builds and stores its own KV cache for the same long, tool-augmented trajectories, incurring substantial memory and compute overhead. Existing KV cache sharing methods largely overlook this multi-LoRA setting. We observe that, across agents, cache differences are dominated by adapter outputs, while activations from the shared pretrained backbone remain highly similar. Based on this observation, we propose LRAgent, a KV cache sharing framework for multi-LoRA agents that decomposes the cache into a shared base component from the pretrained weights and an adapter-dependent component from LoRA weights. LRAgent reduces memory overhead by sharing the base component and storing the adapter component in its inherent low-rank form, and further reduces compute overhead, enabled by shared-A multi-LoRA architectures, by also sharing the low-rank cache and avoiding redundant computations for contexts already processed by other agents. To efficiently reconstruct adapter contributions at runtime, we introduce Flash-LoRA-Attention, a kernel that reorders attention computation to avoid materializing the low-rank cache to full dimension. LRAgent achieves throughput and time-to-first-token latency close to fully shared caching, while preserving accuracy near the non-shared caching baseline across agentic question-answering benchmarks.

We present a single-GPU (H100, bf16) evaluation of GPT-OSS-20B (Mixture-of-Experts; 20.9B total, approx. 3.61B active) against dense baselines Qwen3-32B and Yi-34B across multiple dimensions. We measure true time-to-first-token (TTFT), full-decode throughput (TPOT), end-to-end latency percentiles, peak VRAM with past key values (PKV) held, and energy via a consistent nvidia-smi-based sampler. At a 2048-token context with 64-token decode, GPT-OSS-20B delivers higher decode throughput and tokens per Joule than dense baselines Qwen3-32B and Yi-34B, while substantially reducing peak VRAM and energy per 1000 generated tokens; its TTFT is higher due to MoE routing overhead. With only 17.3% of parameters active (3.61B of 20.9B), GPT-OSS-20B provides about 31.8% higher decode throughput and 25.8% lower energy per 1000 generated tokens than Qwen3-32B at 2048/64, while using 31.7% less peak VRAM. Normalized by active parameters, GPT-OSS-20B shows markedly stronger per-active-parameter efficiency (APE), underscoring MoE's deployment advantages. We do not evaluate accuracy; this is a deployment-focused study. We release code and consolidated results to enable replication and extension.

Current Retrieval-Augmented Generation (RAG) systems concatenate and process numerous retrieved document chunks for prefill which requires a large volume of computation, therefore leading to significant latency in time-to-first-token (TTFT). To reduce the computation overhead as well as TTFT, we introduce TurboRAG, a novel RAG system that redesigns the inference paradigm of the current RAG system by first pre-computing and storing the key-value (KV) caches of documents offline, and then directly retrieving the saved KV cache for prefill. Hence, online computation of KV caches is eliminated during inference. In addition, we provide a number of insights into the mask matrix and positional embedding mechanisms, plus fine-tune a pretrained language model to maintain model accuracy of TurboRAG. Our approach is applicable to most existing large language models and their applications without any requirement in modification of models and inference systems. Experimental results across a suite of RAG benchmarks demonstrate that TurboRAG reduces TTFT by up to 9.4x compared to the conventional RAG systems (on an average of 8.6x), but reserving comparable performance to the standard RAG systems.

As large language models (LLMs) continue to advance rapidly, they are becoming increasingly capable while simultaneously demanding ever-longer context lengths. To improve the inference efficiency of long-context processing, several novel low-complexity hybrid architectures have recently been proposed, effectively alleviating the computational burden of long-context inference. However, existing research on long-context prefill acceleration remains predominantly focused on sparse attention mechanisms, which achieve their maximum speedup only on full-attention models. When transferred to emerging architectures--such as linear/full attention hybrids or sliding window/full attention hybrids--these prefill acceleration approaches suffer significant performance degradation. Furthermore, such methods are generally incompatible with continuous batching, making them difficult to integrate into modern inference engines such as vLLM. To this end, we propose UniPrefill, a prefill acceleration framework applicable to virtually any model architecture, which directly accelerates the model's computation at the token level. We further implement UniPrefill as a continuous batching operator and extend vLLM's scheduling strategy to natively support prefill-decode co-processing and tensor parallel for UniPrefill, enabling its seamless integration into vLLM. UniPrefill achieves up to 2.1x speedup in Time-To-First-Token (TTFT), with the acceleration becoming increasingly pronounced as the number of concurrent requests grows.

The Key-Value (KV) cache of Large Language Models (LLMs) is prefix-based, making it highly inefficient for processing contexts retrieved in arbitrary order. Position-Independent Caching (PIC) has been proposed to enable KV reuse without positional constraints; however, existing approaches often incur substantial accuracy degradation, limiting their practical adoption. To address this issue, we propose native PIC by reintroducing the encoder to prevalent decoder-only LLMs and explicitly training it to support PIC. We further develop COMB, a PIC-aware caching system that integrates seamlessly with existing inference frameworks. Experimental results show that COMB reduces Time-to-First-Token (TTFT) by 51-94% and increases throughput by 3times with comparable accuracy. Furthermore, the quality improvement when using DeepSeek-V2-Lite-Chat demonstrates the applicability of COMB to other types of decoder-only LLMs. Our code is available at https://github.com/shijuzhao/Comb.

Transformer-based large language model (LLM) inference serving is now the backbone of many cloud services. LLM inference consists of a prefill phase and a decode phase. However, existing LLM deployment practices often overlook the distinct characteristics of these phases, leading to significant interference. To mitigate interference, our insight is to carefully schedule and group inference requests based on their characteristics. We realize this idea in TetriInfer through three pillars. First, it partitions prompts into fixed-size chunks so that the accelerator always runs close to its computationsaturated limit. Second, it disaggregates prefill and decode instances so each can run independently. Finally, it uses a smart two-level scheduling algorithm augmented with predicted resource usage to avoid decode scheduling hotspots. Results show that TetriInfer improves time-to-first-token (TTFT), job completion time (JCT), and inference efficiency in turns of performance per dollar by a large margin, e.g., it uses 38% less resources all the while lowering average TTFT and average JCT by 97% and 47%, respectively.

We present Prompt Cache, an approach for accelerating inference for large language models (LLM) by reusing attention states across different LLM prompts. Many input prompts have overlapping text segments, such as system messages, prompt templates, and documents provided for context. Our key insight is that by precomputing and storing the attention states of these frequently occurring text segments on the inference server, we can efficiently reuse them when these segments appear in user prompts. Prompt Cache employs a schema to explicitly define such reusable text segments, called prompt modules. The schema ensures positional accuracy during attention state reuse and provides users with an interface to access cached states in their prompt. Using a prototype implementation, we evaluate Prompt Cache across several LLMs. We show that Prompt Cache significantly reduce latency in time-to-first-token, especially for longer prompts such as document-based question answering and recommendations. The improvements range from 8x for GPU-based inference to 60x for CPU-based inference, all while maintaining output accuracy and without the need for model parameter modifications.

Large language model (LLM) serving has transformed from stateless to stateful systems, utilizing techniques like context caching and disaggregated inference. These optimizations extend the lifespan and domain of the KV cache, necessitating a new architectural approach. We present MemServe, a unified system that integrates both inter-request and intra-request optimizations. MemServe introduces MemPool, an elastic memory pool managing distributed memory and KV caches across serving instances. Using MemPool APIs, MemServe combines context caching with disaggregated inference for the first time, supported by a global scheduler that enhances cache reuse through a global prompt tree-based locality-aware policy. Tests show that MemServe significantly improves job completion time and time-to-first-time.

DistServe improves the performance of large language models (LLMs) serving by disaggregating the prefill and decoding computation. Existing LLM serving systems colocate the two phases and batch the computation of prefill and decoding across all users and requests. We find that this strategy not only leads to strong prefill-decoding interferences but also couples the resource allocation and parallelism plans for both phases. LLM applications often emphasize individual latency for each phase: time to first token (TTFT) for the prefill phase and time per output token (TPOT) of each request for the decoding phase. In the presence of stringent latency requirements, existing systems have to prioritize one latency over the other, or over-provision compute resources to meet both. DistServe assigns prefill and decoding computation to different GPUs, hence eliminating prefill-decoding interferences. Given the application's TTFT and TPOT requirements, DistServe co-optimizes the resource allocation and parallelism strategy tailored for each phase. DistServe also places the two phases according to the serving cluster's bandwidth to minimize the communication caused by disaggregation. As a result, DistServe significantly improves LLM serving performance in terms of the maximum rate that can be served within both TTFT and TPOT constraints on each GPU. Our evaluations show that on various popular LLMs, applications, and latency requirements, DistServe can serve 4.48x more requests or 10.2x tighter SLO, compared to state-of-the-art systems, while staying within latency constraints for > 90% of requests.

Large Multimodal Models (LMMs) extend Large Language Models (LLMs) by handling diverse inputs such as images, audio, and video, but at the cost of adding a multimodal encoding stage that increases both computational and memory overhead. This step negatively affects key Service Level Objectives (SLOs), such as time to first token (TTFT) and time per output token (TPOT). We introduce Encode-Prefill-Decode (EPD) Disaggregation, a novel framework that separates the encoding, prefill, and decode stages onto dedicated resources. Unlike current systems, which bundle encoding and prefill together, our approach decouples these steps, unlocking new opportunities and optimizations. These include a mechanism to cache multimedia tokens for efficient transfer, a novel way to parallelize the encoding load within a request, a module for optimal resource allocation for disaggregated serving, and a novel role-switching method to handle changing workload characteristics. Experimental evaluations with popular LMMs show substantial gains in memory efficiency (up to 15x lower peak memory utilization), batch sizes (up to 22x larger), 10x more images per request, and 2.2x larger KV caches. Furthermore, it leads to significant improvements in SLO attainment (up to 90-100% improvement) and TTFT (up to 71% reduction), compared to systems that do not disaggregate. The code is available at https://github.com/vbdi/epdserve.

We present OmniVLM, a sub-billion-parameter vision-language model for efficient on-device inference. OmniVLM introduces a token compression mechanism that reduces visual token sequence length from 729 to 81 tokens, significantly reducing computational overhead while preserving visual-semantic fidelity. Through a multi-stage training pipeline of pretraining, supervised fine-tuning, and minimal-edit Direct Preference Optimization (DPO), OmniVLM matches the performance of larger models. On multiple benchmarks including ScienceQA, POPE, and MMMU, OmniVLM outperforms existing baselines like nanoLLAVA within a 968M-parameter footprint. Empirical results on the same laptop demonstrate 9.1x faster time-to-first-token (0.75s vs 6.82s) and 1.5x higher decoding speed (29.41 vs 19.20 tokens/s) compared to nanoLLAVA, enabling efficient deployment on edge devices. The model weights can be accessed on huggingface: https://huggingface.co/NexaAIDev/OmniVLM-968M, and the inference examples can be find in Appendix B.

Large Language Models (LLMs) have achieved strong performance across natural language and multimodal tasks, yet their practical deployment remains constrained by inference latency and kernel launch overhead, particularly in interactive, short-sequence settings. This paper presents a hybrid runtime framework that combines Just-In-Time (JIT) compilation with CUDA Graph execution to reduce launch overhead while preserving runtime flexibility during autoregressive decoding. The framework partitions transformer inference into static components executed via CUDA Graph replay and dynamic components handled through JIT-compiled kernels, enabling asynchronous graph capture and reuse across decoding steps. We evaluate the proposed approach on LLaMA-2 7B using single-GPU, batch-size-one inference across prompt lengths from 10 to 500 tokens. Experimental results show that the hybrid runtime reduces Time-to-First-Token (TTFT) by up to 66.0% and achieves lower P99 latency compared with TensorRT-LLM in this regime. These results indicate that hybrid JIT-CUDA Graph execution can effectively reduce inference latency and variance for short-sequence LLM workloads, making it a practical optimization strategy for latency-sensitive AI applications.

Despite the computational efficiency of MoE models, the excessive memory footprint and I/O overhead inherent in multi-expert architectures pose formidable challenges for real-time inference on resource-constrained edge platforms. While existing static methods struggle with a rigid latency-accuracy trade-off, we observe that expert importance is highly skewed and depth-dependent. Motivated by these insights, we propose DyMoE, a dynamic mixed-precision quantization framework designed for high-performance edge inference. Leveraging insights into expert importance skewness and depth-dependent sensitivity, DyMoE introduces: (1) importance-aware prioritization to dynamically quantize experts at runtime; (2) depth-adaptive scheduling to preserve semantic integrity in critical layers; and (3) look-ahead prefetching to overlap I/O stalls. Experimental results on commercial edge hardware show that DyMoE reduces Time-to-First-Token (TTFT) by 3.44x-22.7x and up to a 14.58x speedup in Time-Per-Output-Token (TPOT) compared to state-of-the-art offloading baselines, enabling real-time, accuracy-preserving MoE inference on resource-constrained edge devices.

Retrieval-Augmented Generation (RAG) has shown significant improvements in various natural language processing tasks by integrating the strengths of large language models (LLMs) and external knowledge databases. However, RAG introduces long sequence generation and leads to high computation and memory costs. We propose RAGCache, a novel multilevel dynamic caching system tailored for RAG. Our analysis benchmarks current RAG systems, pinpointing the performance bottleneck (i.e., long sequence due to knowledge injection) and optimization opportunities (i.e., caching knowledge's intermediate states). Based on these insights, we design RAGCache, which organizes the intermediate states of retrieved knowledge in a knowledge tree and caches them in the GPU and host memory hierarchy. RAGCache proposes a replacement policy that is aware of LLM inference characteristics and RAG retrieval patterns. It also dynamically overlaps the retrieval and inference steps to minimize the end-to-end latency. We implement RAGCache and evaluate it on vLLM, a state-of-the-art LLM inference system and Faiss, a state-of-the-art vector database. The experimental results show that RAGCache reduces the time to first token (TTFT) by up to 4x and improves the throughput by up to 2.1x compared to vLLM integrated with Faiss.

Large Language Model (LLM) serving systems remain fundamentally fragile, where frequent hardware faults in hyperscale clusters trigger disproportionate service outages in the software stack. Current recovery mechanisms are prohibitively slow, often requiring up to 10 minutes to reinitialize resources and reload massive model weights. We introduce KevlarFlow, a fault tolerant serving architecture designed to bridge the gap between hardware unreliability and service availability. KevlarFlow leverages 1) decoupled model parallelism initialization, 2) dynamic traffic rerouting, and 3) background KV cache replication to maintain high throughput during partial failures. Our evaluation demonstrates that KevlarFlow reduces mean-time-to-recovery (MTTR) by 20x and, under failure conditions, improves average latency by 3.1x, 99th percentile (p99) latency by 2.8x, average time-to-first-token (TTFT) by 378.9x, and p99 TTFT by 574.6x with negligible runtime overhead in comparison to state-of-the-art LLM serving systems.

Multimodal large language models (MLLMs) extend LLMs with visual understanding through a three-stage pipeline: multimodal preprocessing, vision encoding, and LLM inference. While these stages enhance capability, they introduce significant system bottlenecks. First, multimodal preprocessing-especially video decoding-often dominates Time-to-First-Token (TTFT). Most systems rely on CPU-based decoding, which severely limits throughput, while existing GPU-based approaches prioritize throughput-oriented parallelism and fail to meet the latency-sensitive requirements of MLLM inference. Second, the vision encoder is a standalone, compute-intensive stage that produces visual embeddings and cannot be co-batched with LLM prefill or decoding. This heterogeneity forces inter-stage blocking and increases token-generation latency. Even when deployed on separate GPUs, these stages underutilize available compute and memory resources, reducing overall utilization and constraining system throughput. To address these challenges, we present FlashCodec and UnifiedServe, two complementary designs that jointly optimize the end-to-end MLLM pipeline. FlashCodec accelerates the multimodal preprocessing stage through collaborative multi-GPU video decoding, reducing decoding latency while preserving high throughput. UnifiedServe optimizes the vision-to-text and inference stages using a logically decoupled their execution to eliminate inter-stage blocking, yet physically sharing GPU resources to maximize GPU system utilization. By carefully orchestrating execution across stages and minimizing interference, UnifiedServe Together, our proposed framework forms an end-to-end optimized stack that can serve up to 3.0times more requests or enforce 1.5times tighter SLOs, while achieving up to 4.4times higher throughput compared to state-of-the-art systems.

Serverless computing has grown rapidly for serving Large Language Model (LLM) inference due to its pay-as-you-go pricing, fine-grained GPU usage, and rapid scaling. However, our analysis reveals that current serverless can effectively serve general LLM but fail with Low-Rank Adaptation (LoRA) inference due to three key limitations: 1) massive parameter redundancy among functions where 99% of weights are unnecessarily duplicated, 2) costly artifact loading latency beyond LLM loading, and 3) magnified resource contention when serving multiple LoRA LLMs. These inefficiencies lead to massive GPU wastage, increased Time-To-First-Token (TTFT), and high monetary costs. We propose ServerlessLoRA, a novel serverless inference system designed for faster and cheaper LoRA LLM serving. ServerlessLoRA enables secure backbone LLM sharing across isolated LoRA functions to reduce redundancy. We design a pre-loading method that pre-loads comprehensive LoRA artifacts to minimize cold-start latency. Furthermore, ServerlessLoRA employs contention aware batching and offloading to mitigate GPU resource conflicts during bursty workloads. Experiment on industrial workloads demonstrates that ServerlessLoRA reduces TTFT by up to 86% and cuts monetary costs by up to 89% compared to state-of-the-art LLM inference solutions.

Vision-language models (VLMs) have recently expanded from static image understanding to video reasoning, but their scalability is fundamentally limited by the quadratic cost of processing dense frame sequences. Long videos often exceed the token budget of modern language models, leading to severe context limitations and latency issues. We introduce Efficient Video Sampling (EVS), a simple, plug-and-play method for reducing token redundancy in videos by identifying and pruning temporally static patches -- spatial regions that remain unchanged across consecutive frames. EVS preserves positional identity, requires no architectural changes or retraining. We show that EVS substantially reduces token count while maintaining semantic fidelity, enabling faster inference and longer input sequences. Applied at inference time, EVS reduces large language model (LLM) time-to-first-token (TTFT) by up to 4x with minimal accuracy loss. When combined with an uptraining phase using stochastic pruning rates, EVS yields models that are robust to varying compression levels and retain full performance under aggressive pruning. Extensive experiments demonstrate that EVS consistently improves efficiency-accuracy trade-offs, unlocking scalable video-language understanding without sacrificing quality.

Long-context inference for Large Language Models (LLMs) is heavily limited by high computational demands. While several existing methods optimize attention computation, they still process the full set of hidden states at each layer, limiting overall efficiency. In this work, we propose SlimInfer, an innovative framework that aims to accelerate inference by directly pruning less critical prompt tokens during the forward pass. Our key insight is an information diffusion phenomenon: As information from critical tokens propagates through layers, it becomes distributed across the entire sequence. This diffusion process suggests that LLMs can maintain their semantic integrity when excessive tokens, even including these critical ones, are pruned in hidden states. Motivated by this, SlimInfer introduces a dynamic fine-grained pruning mechanism that accurately removes redundant tokens of hidden state at intermediate layers. This layer-wise pruning naturally enables an asynchronous KV cache manager that prefetches required token blocks without complex predictors, reducing both memory usage and I/O costs. Extensive experiments show that SlimInfer can achieve up to 2.53times time-to-first-token (TTFT) speedup and 1.88times end-to-end latency reduction for LLaMA3.1-8B-Instruct on a single RTX 4090, without sacrificing performance on LongBench. Our code will be released upon acceptance.

We introduce Block-Attention, an attention mechanism designed to address the increased inference latency and cost in Retrieval-Augmented Generation (RAG) scenarios. Traditional approaches often encode the entire context. Instead, Block-Attention divides retrieved documents into discrete blocks, with each block independently calculating key-value (KV) states except for the final block. In RAG scenarios, by defining each passage as a block, Block-Attention enables us to reuse the KV states of passages that have been seen before, thereby significantly reducing the latency and the computation overhead during inference. The implementation of Block-Attention involves block segmentation, position re-encoding, and fine-tuning the LLM to adapt to the Block-Attention mechanism. Experiments on four RAG benchmarks demonstrate that after block fine-tuning, the Block-Attention model achieves performance comparable to self-attention models (68.4\% vs 67.9\% on Llama3) or even superior performance (62.8\% vs 59.6\% on Mistral). Notably, Block-Attention significantly reduces the time to first token (TTFT) and floating point operations (FLOPs) to a very low level. It only takes 45 ms to output the first token for an input sequence with a total length of 32K. Compared to the self-attention models, the time consumption and corresponding FLOPs are reduced by 98.7\% and 99.8\%, respectively.

In recent years, Large Language Models (LLM) such as ChatGPT, CoPilot, and Gemini have been widely adopted in different areas. As the use of LLMs continues to grow, many efforts have focused on reducing the massive training overheads of these models. But it is the environmental impact of handling user requests to LLMs that is increasingly becoming a concern. Recent studies estimate that the costs of operating LLMs in their inference phase can exceed training costs by 25x per year. As LLMs are queried incessantly, the cumulative carbon footprint for the operational phase has been shown to far exceed the footprint during the training phase. Further, estimates indicate that 500 ml of fresh water is expended for every 20-50 requests to LLMs during inference. To address these important sustainability issues with LLMs, we propose a novel framework called SLIT to co-optimize LLM quality of service (time-to-first token), carbon emissions, water usage, and energy costs. The framework utilizes a machine learning (ML) based metaheuristic to enhance the sustainability of LLM hosting across geo-distributed cloud datacenters. Such a framework will become increasingly vital as LLMs proliferate.

Video large language models (Video-LLMs) have demonstrated strong capabilities in video understanding tasks. However, their practical deployment is still hindered by the inefficiency introduced by processing massive amounts of visual tokens. Although recent approaches achieve extremely low token retention ratios while maintaining accuracy comparable to full-token baselines, most of them perform compression only at the late stage of prefilling, leaving the efficiency of the vision encoder unoptimized. In this paper, we first show that vision encoding contributes a large portion to the time-to-first-token (TTFT). Therefore, instead of compressing visual tokens only after the vision encoder, performing compression inside the encoder still leaves substantial room for exploration. Based on this insight, we propose EarlyTom, a training-free token compression framework that performs early-stage visual token compression inside the vision encoder, enabling significantly better TTFT reduction and higher throughput. In addition, we introduce a decoupled spatial token selection strategy that improves the overall compression effectiveness. EarlyTom reduces TTFT by up to 2.65x and FLOPs by up to 61% on a single NVIDIA A100 GPU for the LLaVA-OneVision-7B model, while maintaining accuracy comparable to the full-token baseline. These improvements substantially enhance the practicality of deploying Video-LLMs in real-world production scenarios.

Video large language models (video LLMs) excel at video comprehension but face significant computational inefficiency due to redundant video tokens. Existing token pruning methods offer solutions. However, approaches operating within the LLM (inner-LLM pruning), such as FastV, incur intrinsic computational overhead in shallow layers. In contrast, methods performing token pruning before the LLM (outer-LLM pruning) primarily address spatial redundancy within individual frames or limited temporal windows, neglecting the crucial global temporal dynamics and correlations across longer video sequences. This leads to sub-optimal spatio-temporal reduction and does not leverage video compressibility fully. Crucially, the synergistic potential and mutual influence of combining these strategies remain unexplored. To further reduce redundancy, we introduce HoliTom, a novel training-free holistic token merging framework. HoliTom employs outer-LLM pruning through global redundancy-aware temporal segmentation, followed by spatial-temporal merging to reduce visual tokens by over 90%, significantly alleviating the LLM's computational burden. Complementing this, we introduce a robust inner-LLM token similarity-based merging approach, designed for superior performance and compatibility with outer-LLM pruning. Evaluations demonstrate our method's promising efficiency-performance trade-off on LLaVA-OneVision-7B, reducing computational costs to 6.9% of FLOPs while maintaining 99.1% of the original performance. Furthermore, we achieve a 2.28x reduction in Time-To-First-Token (TTFT) and a 1.32x acceleration in decoding throughput, highlighting the practical benefits of our integrated pruning approach for efficient video LLMs inference.

Blocking communication presents a major hurdle in running MoEs efficiently in distributed settings. To address this, we present FarSkip-Collective which modifies the architecture of modern models to enable overlapping of their computation with communication. Our approach modifies the architecture to skip connections in the model and it is unclear a priori whether the modified model architecture can remain as capable, especially for large state-of-the-art models and while modifying all of the model layers. We answer this question in the affirmative and fully convert a series of state-of-the-art models varying from 16B to 109B parameters to enable overlapping of their communication while achieving accuracy that is comparable with their original open-source releases. For example, we convert Llama 4 Scout (109B) via self-distillation and achieve average accuracy within 1% of its instruction tuned release averaged over a wide range of downstream evaluations. In addition to demonstrating retained accuracy of the large modified models, we realize the benefits of FarSkip-Collective through optimized implementations that explicitly overlap communication with computation, accelerating both training and inference in existing frameworks. For inference, we demonstrate 32.6% speedup in Time To First Token when serving a converted DeepSeek-V3 architecture with expert parallelism in SGLang and achieve 97.3% communication-computation overlap during the prefill stage. During training, our approach enables 88.9% communication overlap of the all-to-all communication collectives when pre-training DeepSeek-V3 MoE layers with expert parallelism.

Long chain-of-thought (CoT) significantly enhances large language models' (LLM) reasoning capabilities. However, the extensive reasoning traces lead to inefficiencies and an increased time-to-first-token (TTFT). We propose a novel training paradigm that uses reinforcement learning (RL) to guide reasoning LLMs to interleave thinking and answering for multi-hop questions. We observe that models inherently possess the ability to perform interleaved reasoning, which can be further enhanced through RL. We introduce a simple yet effective rule-based reward to incentivize correct intermediate steps, which guides the policy model toward correct reasoning paths by leveraging intermediate signals generated during interleaved reasoning. Extensive experiments conducted across five diverse datasets and three RL algorithms (PPO, GRPO, and REINFORCE++) demonstrate consistent improvements over traditional think-answer reasoning, without requiring external tools. Specifically, our approach reduces TTFT by over 80% on average and improves up to 19.3% in Pass@1 accuracy. Furthermore, our method, trained solely on question answering and logical reasoning datasets, exhibits strong generalization ability to complex reasoning datasets such as MATH, GPQA, and MMLU. Additionally, we conduct in-depth analysis to reveal several valuable insights into conditional reward modeling.

Multimodal large language models (MLLMs) extend LLMs to handle images, videos, and audio by incorporating feature extractors and projection modules. However, these additional components -- combined with complex inference pipelines and heterogeneous workloads -- introduce significant inference overhead. Therefore, efficiently serving MLLMs remains a major challenge. Current tightly coupled serving architectures struggle to distinguish between mixed request types or adapt parallelism strategies to different inference stages, leading to increased time-to-first-token (TTFT) latency and poor resource utilization. To address this, we introduce Elastic Multimodal Parallelism (EMP), a new serving paradigm that elastically adapts to resource heterogeneity across request types and inference stages. Building upon EMP, we develop ElasticMM, an MLLM serving system that (1) separates requests into independent modality groups with dynamic resource allocation via a modality-aware load balancer; (2) decouples inference stages and enables parallelism adjustment and adaptive scaling via elastic partition scheduling; and (3) improves inference efficiency through unified multimodal prefix caching and non-blocking encoding. Experiments on diverse real-world datasets show that ElasticMM outperforms state-of-the-art (SOTA) serving systems, reducing TTFT by up to 4.2x and achieving 3.2-4.5x higher throughput while meeting service-level objectives (SLOs).

Recent advances in long-text understanding have pushed the context length of large language models (LLMs) up to one million tokens. It boosts LLMs's accuracy and reasoning capacity but causes exorbitant computational costs and unsatisfactory Time to First Token (TTFT). KV cache reuse, which reuses the exact same KV cache of prefixes and templates or shares similar ones but with extra selective recomputation, offers a promising way to tackle this issue. However, prior studies overlook the cross-request KV reuse and the attention deviations introduced by new tokens during the decoding stage. In this paper, we present a KV cache management module that shares the KV cache across requests under multi-tenant scenarios without sacrificing model accuracy. Our system, KVShare, enables accurate and efficient LLM serving by 1) a Dual-Stage High Deviation algorithm (DHD) that conditionally selects a small portion of KV cache to be recomputed during both prefill and decode phases, and 2) a cache-aware scheduler that prioritizes requests based on their KV cache hit rates and orchestrates continuous batching to achieve enhanced system efficiency and faster TTFT. Multi-task experiments conducted on models such as Qwen2.5-7B,Llama3.1-8B and Yi1.5-9B demonstrate that KVShare reduces TTFT by up to 9.39x and increases 1.2x of the throughput compared to the full KV recompute. Moreover, KVShare achieves 20.38% boost in terms of accuracy compared to SOTA methods.

Multi-modal Large Language Models (MLLMs) serving systems commonly employ KV-cache compression to reduce memory footprint. However, existing compression methods introduce significant processing overhead and queuing delays, particularly in concurrent serving scenarios. We present FastCache, a novel serving framework that effectively addresses these challenges through two key innovations: (1) a dynamic batching strategy that optimizes request scheduling across prefill, compression, and decode stages, and (2) an efficient KV-cache memory pool mechanism that eliminates memory fragmentation while maintaining high GPU utilization. Our comprehensive experiments on the GQA and MileBench datasets demonstrate that FastCache achieves up to 19.3times reduction in Time-To-First-Token (TTFT) and 12.1times improvement in throughput compared to state-of-the-art baselines. The system maintains stable performance under high-concurrency scenarios (up to 40 req/s) while reducing average memory consumption by 20\%. These results establish FastCache as an efficient solution for real-world LLM serving systems with KV-cache compression.

Scaling the input image resolution is essential for enhancing the performance of Vision Language Models (VLMs), particularly in text-rich image understanding tasks. However, popular visual encoders such as ViTs become inefficient at high resolutions due to the large number of tokens and high encoding latency caused by stacked self-attention layers. At different operational resolutions, the vision encoder of a VLM can be optimized along two axes: reducing encoding latency and minimizing the number of visual tokens passed to the LLM, thereby lowering overall latency. Based on a comprehensive efficiency analysis of the interplay between image resolution, vision latency, token count, and LLM size, we introduce FastVLM, a model that achieves an optimized trade-off between latency, model size and accuracy. FastVLM incorporates FastViTHD, a novel hybrid vision encoder designed to output fewer tokens and significantly reduce encoding time for high-resolution images. Unlike previous methods, FastVLM achieves the optimal balance between visual token count and image resolution solely by scaling the input image, eliminating the need for additional token pruning and simplifying the model design. In the LLaVA-1.5 setup, FastVLM achieves 3.2times improvement in time-to-first-token (TTFT) while maintaining similar performance on VLM benchmarks compared to prior works. Compared to LLaVa-OneVision at the highest resolution (1152times1152), FastVLM achieves comparable performance on key benchmarks like SeedBench and MMMU, using the same 0.5B LLM, but with 85times faster TTFT and a vision encoder that is 3.4times smaller.

Large language models (LLMs) exhibit exceptional performance across a wide range of tasks; however, their token-by-token autoregressive generation process significantly hinders inference speed. Speculative decoding presents a promising draft-then-verify framework that reduces generation latency while maintaining output distribution fidelity. Nevertheless, the draft model introduces additional computational overhead, becoming a performance bottleneck and increasing the time to first token (TTFT). Previous approaches to mitigate draft model overhead have primarily relied on heuristics and generally failed to match the quality of the draft language models. To address these challenges, we propose DuoDecoding, a novel approach that strategically deploys the draft and target models on the CPU and GPU respectively, enabling parallel decoding while preserving draft quality. Our method incorporates a hardware-aware optimal draft budget to minimize idle times and employs dynamic multi-sequence drafting to enhance draft quality. Extensive experiments across seven tasks show that DuoDecoding achieves up to 2.61x speedup in generation latency, while reducing TTFT to 83% of that in conventional speculative decoding. The Code is available at https://github.com/KaiLv69/DuoDecoding.

Large Language Models (LLMs) have demonstrated remarkable capabilities in leveraging extensive external knowledge to enhance responses in multi-turn and agentic applications, such as retrieval-augmented generation (RAG). However, processing long-context inputs introduces significant system latency and demands substantial memory for the key-value cache, resulting in reduced throughput and a fundamental trade-off between knowledge enrichment and system efficiency. While minimizing latency for long-context inputs is a primary objective for LLMs, we contend that RAG require specialized consideration. In RAG, much of the LLM context consists of concatenated passages from retrieval, with only a small subset directly relevant to the query. These passages often exhibit low semantic similarity due to diversity or deduplication during re-ranking, leading to block-diagonal attention patterns that differ from those in standard LLM generation tasks. Based on this observation, we argue that most computations over the RAG context during decoding are unnecessary and can be eliminated with minimal impact on performance. To this end, we propose REFRAG, an efficient decoding framework that compresses, senses, and expands to improve latency in RAG applications. By exploiting the sparsity structure, we demonstrate a 30.85 the time-to-first-token acceleration (3.75 improvement to previous work) without loss in perplexity. In addition, our optimization framework for large context enables REFRAG to extend the context size of LLMs by 16. We provide rigorous validation of REFRAG across diverse long-context tasks, including RAG, multi-turn conversations, and long document summarization, spanning a wide range of datasets. Experimental results confirm that REFRAG delivers substantial speedup with no loss in accuracy compared to LLaMA models and other state-of-the-art baselines across various context sizes.

Mainstream Transformer-based large language models face major efficiency bottlenecks: training computation scales quadratically with sequence length, and inference memory grows linearly, limiting long-context processing. Building large models on non-NVIDIA platforms also poses challenges for stable and efficient training. To address this, we introduce SpikingBrain, a family of brain-inspired models designed for efficient long-context training and inference. SpikingBrain leverages the MetaX GPU cluster and focuses on three aspects: (1) Model Architecture: linear and hybrid-linear attention architectures with adaptive spiking neurons; (2) Algorithmic Optimizations: an efficient, conversion-based training pipeline and a dedicated spike coding framework; (3) System Engineering: customized training frameworks, operator libraries, and parallelism strategies tailored to MetaX hardware. Using these techniques, we develop two models: SpikingBrain-7B, a linear LLM, and SpikingBrain-76B, a hybrid-linear MoE LLM. These models demonstrate the feasibility of large-scale LLM development on non-NVIDIA platforms. SpikingBrain achieves performance comparable to open-source Transformer baselines while using only about 150B tokens for continual pre-training. Our models significantly improve long-sequence training efficiency and deliver inference with (partially) constant memory and event-driven spiking behavior. For example, SpikingBrain-7B attains over 100x speedup in Time to First Token for 4M-token sequences. Training remains stable for weeks on hundreds of MetaX C550 GPUs, with the 7B model reaching a Model FLOPs Utilization of 23.4 percent. The proposed spiking scheme achieves 69.15 percent sparsity, enabling low-power operation. Overall, this work demonstrates the potential of brain-inspired mechanisms to drive the next generation of efficient and scalable large model design.

This technical report introduces AngelSlim, a comprehensive and versatile toolkit for large model compression developed by the Tencent Hunyuan team. By consolidating cutting-edge algorithms, including quantization, speculative decoding, token pruning, and distillation. AngelSlim provides a unified pipeline that streamlines the transition from model compression to industrial-scale deployment. To facilitate efficient acceleration, we integrate state-of-the-art FP8 and INT8 Post-Training Quantization (PTQ) algorithms alongside pioneering research in ultra-low-bit regimes, featuring HY-1.8B-int2 as the first industrially viable 2-bit large model. Beyond quantization, we propose a training-aligned speculative decoding framework compatible with multimodal architectures and modern inference engines, achieving 1.8x to 2.0x throughput gains without compromising output correctness. Furthermore, we develop a training-free sparse attention framework that reduces Time-to-First-Token (TTFT) in long-context scenarios by decoupling sparse kernels from model architectures through a hybrid of static patterns and dynamic token selection. For multimodal models, AngelSlim incorporates specialized pruning strategies, namely IDPruner for optimizing vision tokens via Maximal Marginal Relevance and Samp for adaptive audio token merging and pruning. By integrating these compression strategies from low-level implementations, AngelSlim enables algorithm-focused research and tool-assisted deployment.

The rapid advancement of Large Language Models (LLMs) has driven the need for more efficient serving strategies. In this context, efficiency refers to the proportion of requests that meet their Service Level Objectives (SLOs), particularly for Time To First Token (TTFT) and Time Between Tokens (TBT). However, existing systems often prioritize one metric at the cost of the other. We present Ascendra, an LLM serving system designed to meet both TTFT and TBT SLOs simultaneously. The core insight behind Ascendra is that a request's urgency evolves as it approaches its deadline. To leverage this, Ascendra partitions GPU resources into two types of instances: low-priority and high-priority. Low-priority instances maximize throughput by processing requests out of arrival order, but at the risk of request starvation. To address this, Ascendra employs a performance model to predict requests at risk of missing their SLOs and proactively offloads them to high-priority instances. High-priority instances are optimized for low-latency execution and handle urgent requests nearing their deadlines. This partitioned architecture enables Ascendra to effectively balance high throughput and low latency. Extensive evaluation shows that Ascendra improves system throughput by up to 1.7x compared to vLLM and Sarathi-Serve while meeting both TTFT and TBT SLOs.

Generative models are reshaping the live-streaming industry by redefining how content is created, styled, and delivered. Previous image-based streaming diffusion models have powered efficient and creative live streaming products but have hit limits on temporal consistency due to the foundation of image-based designs. Recent advances in video diffusion have markedly improved temporal consistency and sampling efficiency for offline generation. However, offline generation systems primarily optimize throughput by batching large workloads. In contrast, live online streaming operates under strict service-level objectives (SLOs): time-to-first-frame must be minimal, and every frame must meet a per-frame deadline with low jitter. Besides, scalable multi-GPU serving for real-time streams remains largely unresolved so far. To address this, we present StreamDiffusionV2, a training-free pipeline for interactive live streaming with video diffusion models. StreamDiffusionV2 integrates an SLO-aware batching scheduler and a block scheduler, together with a sink-token--guided rolling KV cache, a motion-aware noise controller, and other system-level optimizations. Moreover, we introduce a scalable pipeline orchestration that parallelizes the diffusion process across denoising steps and network layers, achieving near-linear FPS scaling without violating latency guarantees. The system scales seamlessly across heterogeneous GPU environments and supports flexible denoising steps (e.g., 1--4), enabling both ultra-low-latency and higher-quality modes. Without TensorRT or quantization, StreamDiffusionV2 renders the first frame within 0.5s and attains 58.28 FPS with a 14B-parameter model and 64.52 FPS with a 1.3B-parameter model on four H100 GPUs, making state-of-the-art generative live streaming practical and accessible--from individual creators to enterprise-scale platforms.

Large Language Models have revolutionized natural language processing, yet serving them efficiently in data centers remains challenging due to mixed workloads comprising latency-sensitive (LS) and best-effort (BE) jobs. Existing inference systems employ iteration-level first-come-first-served scheduling, causing head-of-line blocking when BE jobs delay LS jobs. We introduce QLLM, a novel inference system designed for Mixture of Experts (MoE) models, featuring a fine-grained, priority-aware preemptive scheduler. QLLM enables expert-level preemption, deferring BE job execution while minimizing LS time-to-first-token (TTFT). Our approach removes iteration-level scheduling constraints, enabling the scheduler to preempt jobs at any layer based on priority. Evaluations on an Nvidia A100 GPU show that QLLM significantly improves performance. It reduces LS TTFT by an average of 65.5times and meets the SLO at up to 7 requests/sec, whereas the baseline fails to do so under the tested workload. Additionally, it cuts LS turnaround time by up to 12.8times without impacting throughput. QLLM is modular, extensible, and seamlessly integrates with Hugging Face MoE models.

We investigate the feasibility of using Multimodal Large Language Models (MLLMs) for real-time online episodic memory question answering. While cloud offloading is common, it raises privacy and latency concerns for wearable assistants, hence we investigate implementation on the edge. We integrated streaming constraints into our question answering pipeline, which is structured into two asynchronous threads: a Descriptor Thread that continuously converts video into a lightweight textual memory, and a Question Answering (QA) Thread that reasons over the textual memory to answer queries. Experiments on the QAEgo4D-Closed benchmark analyze the performance of Multimodal Large Language Models (MLLMs) within strict resource boundaries, showing promising results also when compared to clound-based solutions. Specifically, an end-to-end configuration running on a consumer-grade 8GB GPU achieves 51.76% accuracy with a Time-To-First-Token (TTFT) of 0.41s. Scaling to a local enterprise-grade server yields 54.40% accuracy with a TTFT of 0.88s. In comparison, a cloud-based solution obtains an accuracy of 56.00%. These competitive results highlight the potential of edge-based solutions for privacy-preserving episodic memory retrieval.

Large Language Models (LLMs) rely on attention mechanisms whose time complexity grows quadratically with input sequence length, creating significant computational bottlenecks during the prefilling stage. Existing static sparse attention methods typically degrade accuracy, while dynamic sparsity methods introduce additional computational overhead due to runtime sparse index estimation. To address these limitations, we propose TriangleMix, a novel training-free static attention pattern. TriangleMix employs dense attention in shallow layers and switches to a triangle-shaped sparse pattern in deeper layers. Extensive experiments demonstrate that TriangleMix reduces attention overhead by 3.7x to 15.3x in deep layers, and decreases overall Time-to-First-Token (TTFT) by 12% to 32% for sequence lengths ranging from 32K to 128K, without sacrificing model accuracy. Moreover, TriangleMix can be seamlessly integrated with dynamic sparsity methods to achieve further speedup, e.g. accelerating MInference by 19% at 128K, highlighting its potential to enhance LLM inference efficiency.

Efficiently serving Large Language Models (LLMs) requires selecting an optimal parallel execution plan, balancing computation, memory, and communication overhead. However, determining the best strategy is challenging due to varying parallelism techniques (data, pipeline, tensor) and workload characteristics (e.g., compute-intensive tasks with long prompts vs. memory-intensive tasks with long generation). We propose APEX, an LLM serving system simulator that efficiently identifies optimal parallel execution plans by considering key factors of LLM serving systems, such as memory usage, batching behavior, etc. APEX performs dynamism-aware simulation to model iteration-level batching, and leverages LLMs' repetitive structure to reduce design space, scaling efficiently to trillion-scale models. APEX abstracts the key components of LLM serving systems, including the model, batching module, quantization formats, and device clusters, enabling the simulator to be general and extensible. Simulating on a CPU, APEX evaluates execution plans for various device clusters, covering diverse LLMs and workloads. APEX finds plans up to 3.37x faster than heuristics, and also plans that reduce energy consumption by up to 45% compared to latency-optimal plans. APEX performs comprehensive evaluations, reporting key system metrics like time per output token and time to first token, which can help service providers meet SLOs. APEX identifies an optimal plan within 15 minutes on a CPU, making it 71x faster and 1234x more cost-effective than cloud-based GPU deployment. APEX can be accessed at https://github.com/microsoft/apex_plus

Large Transformer networks are increasingly used in settings where low inference latency can improve the end-user experience and enable new applications. However, autoregressive inference is resource intensive and requires parallelism for efficiency. Parallelism introduces collective communication that is both expensive and represents a phase when hardware resources are underutilized. Towards mitigating this, Kraken is an evolution of the standard Transformer architecture that is designed to complement existing tensor parallelism schemes for efficient inference on multi-device systems. By introducing a fixed degree of intra-layer model parallelism, the architecture allows collective operations to be overlapped with compute, decreasing latency and increasing hardware utilization. When trained on OpenWebText, Kraken models reach a similar perplexity as standard Transformers while also preserving their language modeling capabilities when evaluated on the SuperGLUE benchmark. Importantly, when tested on multi-GPU systems using TensorRT-LLM engines, Kraken speeds up Time To First Token by a mean of 35.6% across a range of model sizes, context lengths, and degrees of tensor parallelism.

Large model inference is shifting from cloud to edge due to concerns about the privacy of user interaction data. However, edge devices often struggle with limited computing power, memory, and bandwidth, requiring collaboration across multiple devices to run and speed up LLM inference. Pipeline parallelism, the mainstream solution, is inefficient for single-user scenarios, while tensor parallelism struggles with frequent communications. In this paper, we argue that tensor parallelism can be more effective than pipeline on low-resource devices, and present a compute- and memory-efficient tensor parallel inference system, named TPI-LLM, to serve 70B-scale models. TPI-LLM keeps sensitive raw data local in the users' devices and introduces a sliding window memory scheduler to dynamically manage layer weights during inference, with disk I/O latency overlapped with the computation and communication. This allows larger models to run smoothly on memory-limited devices. We analyze the communication bottleneck and find that link latency, not bandwidth, emerges as the main issue, so a star-based allreduce algorithm is implemented. Through extensive experiments on both emulated and real testbeds, TPI-LLM demonstrated over 80% less time-to-first-token and token latency compared to Accelerate, and over 90% compared to Transformers and Galaxy, while cutting the peak memory footprint of Llama 2-70B by 90%, requiring only 3.1 GB of memory for 70B-scale models.

We present a systematic, empirical evaluation of five local large language model (LLM) runtimes on Apple Silicon: MLX, MLC-LLM, llama.cpp, Ollama, and PyTorch MPS. Experiments were conducted on a Mac Studio equipped with an M2 Ultra processor and 192 GB of unified memory. Using the Qwen-2.5 model family across prompts ranging from a few hundred to 100,000 tokens, we measure time-to-first-token (TTFT), steady-state throughput, latency percentiles, long-context behavior (key-value and prompt caching), quantization support, streaming performance, batching and concurrency behavior, and deployment complexity. Under our settings, MLX achieves the highest sustained generation throughput, while MLC-LLM delivers consistently lower TTFT for moderate prompt sizes and offers stronger out-of-the-box inference features. llama.cpp is highly efficient for lightweight single-stream use, Ollama emphasizes developer ergonomics but lags in throughput and TTFT, and PyTorch MPS remains limited by memory constraints on large models and long contexts. All frameworks execute fully on-device with no telemetry, ensuring strong privacy guarantees. We release scripts, logs, and plots to reproduce all results. Our analysis clarifies the design trade-offs in Apple-centric LLM deployments and provides evidence-based recommendations for interactive and long-context processing. Although Apple Silicon inference frameworks still trail NVIDIA GPU-based systems such as vLLM in absolute performance, they are rapidly maturing into viable, production-grade solutions for private, on-device LLM inference.

Large Reasoning Models (LRMs) are criticized for the excessively lengthy Chain-of-Thought (CoT) to derive the final answer, suffering from high first-token and overall latency. Typically, the CoT of LRMs mixes multiple thinking units; each unit attempts to produce a candidate answer to the original query. Hence, a natural idea to improve efficiency is to reduce the unit number. Yet, the fact that the thinking units in vanilla CoT cannot be explicitly managed renders doing so challenging. This paper introduces Multi-Turn Decomposition (MinD) to decode conventional CoT into a sequence of explicit, structured, and turn-wise interactions to bridge the gap. In MinD, the model provides a multi-turn response to the query, where each turn embraces a thinking unit and yields a corresponding answer. The subsequent turns can reflect, verify, revise, or explore alternative approaches to both the thinking and answer parts of earlier ones. This not only makes the answer delivered more swiftly, but also enables explicit controls over the iterative reasoning process (i.e., users may halt or continue at any turn). We follow a supervised fine-tuning (SFT) then reinforcement learning (RL) paradigm to realize MinD. We first rephrase the outputs of an LRM into multi-turn formats by prompting another LLM, and then tune the LRM with such data. Observing that the tuned model tends to consume even more tokens than the original one (probably due to that the multi-turn formats introduce additional answer tokens), we advocate leveraging RL algorithms like GRPO to prioritize correct outputs with fewer turns. Trained on the MATH dataset using R1-Distill models, MinD can achieve up to ~70% reduction in both output token usage and time to first token (TTFT), while maintaining competitive performance on reasoning benchmarks such as MATH-500, AIME24, AMC23, and GPQA-Diamond.

Large Language Models (LLMs) face significant computational challenges when processing long contexts due to the quadratic complexity of self-attention. While soft context compression methods, which map input text to smaller latent representations, have shown promise, their real-world adoption is limited. Existing techniques typically compress the context as a single unit, which leads to quadratic compression complexity and an inability to reuse computations across queries with overlapping contexts. In this work, we introduce CompLLM, a soft compression technique designed for practical deployment. Instead of processing the context holistically, CompLLM divides it into segments and compresses each one independently. This simple design choice yields three critical properties: efficiency, as the compression step scales linearly with the context length; scalability, enabling models trained on short sequences (e.g., 1k tokens) to generalize to contexts of 100k tokens; and reusability, allowing compressed segments to be cached and reused across different queries. Our experiments show that with a 2x compression rate, at high context lengths CompLLM speeds up Time To First Token (TTFT) by up to 4x and reduces the KV cache size by 50%. Furthermore, CompLLM achieves performance comparable to that obtained with the uncompressed context, and even surpasses it on very long sequences, demonstrating its effectiveness and practical utility.

Large language models face significant computational and memory challenges when processing long contexts. During inference, efficient management of the key-value (KV) cache, which stores intermediate activations for autoregressive generation, is critical to reducing memory overhead and improving computational efficiency. Traditional token-level efficient KV caching methods overlook semantic information, treating tokens independently without considering their semantic relationships. Meanwhile, existing semantic-preserving KV cache management approaches often suffer from substantial memory usage and high time-to-first-token. To address these limitations, we propose SentenceKV, a novel sentence-level semantic KV caching approach designed to enhance inference efficiency while preserving semantic coherence. During prefilling, SentenceKV groups tokens based on sentence-level semantic similarity, compressing sentence representations into concise semantic vectors stored directly on the GPU, while individual KV pairs are offloaded to CPU. During decoding, SentenceKV generates tokens by selectively retrieving semantically relevant sentence-level KV entries, leveraging the semantic similarity between the prefilling-stage semantic vectors and decoding-stage queries. This ensures efficient and contextually accurate predictions, minimizing the loading of redundant or irrelevant data into GPU memory and significantly reducing memory overhead while maintaining stable inference latency, even for extremely long contexts. Extensive evaluations on benchmarks including PG-19, LongBench, and Needle-In-A-Haystack demonstrate that SentenceKV significantly outperforms state-of-the-art methods in both efficiency and memory usage, without compromising model accuracy.

Mixture of experts (MoE) models achieve state-of-the-art results in language modeling but suffer from inefficient hardware utilization due to imbalanced token routing and communication overhead. While prior work has focused on optimizing MoE training and decoder architectures, inference for encoder-based MoE models in a multi-GPU with expert parallelism setting remains underexplored. We introduce MoEShard, an inference system that achieves perfect load balancing through tensor sharding of MoE experts. Unlike existing approaches that rely on heuristic capacity factors or drop tokens, MoEShard evenly distributes computation across GPUs and ensures full token retention, maximizing utilization regardless of routing skewness. We achieve this through a strategic row- and column-wise decomposition of expert matrices. This reduces idle time and avoids bottlenecks caused by imbalanced expert assignments. Furthermore, MoEShard minimizes kernel launches by fusing decomposed expert computations, significantly improving throughput. We evaluate MoEShard against DeepSpeed on encoder-based architectures, demonstrating speedups of up to 6.4times in time to first token (TTFT). Our results show that tensor sharding, when properly applied to experts, is a viable and effective strategy for efficient MoE inference.

Converting a pretrained Transformer into a more efficient hybrid model through distillation offers a promising approach to reducing inference costs. However, achieving high-quality generation in distilled models requires careful joint design of both the student architecture and the distillation process. Many prior distillation works evaluate downstream multiple-choice benchmarks by ranking candidate answers with log-likelihood rather than requiring autoregressive generation, which can obscure important differences in model quality. For example, we show that a 7B parameter distilled model that nearly matches its teacher to within 0.2\,pp under log-likelihood scoring actually falls behind by 20.8\,pp when the model must generate answers autoregressively. We propose a Hybrid Kimi Delta Attention (Hybrid-KDA) architecture paired with GenDistill, a multi-stage distillation pipeline, and use generation-based evaluation throughout to guide design decisions. Applying this approach to Qwen3-0.6B, we systematically ablate six design axes: training objective, loss masking, training duration, dataset selection, parameter freezing, and architecture choice. We find that log-likelihood-based evaluation consistently underestimates the gap between teacher and student, and can in some cases reverse the ranking of design choices, meaning that conclusions drawn from perplexity-only evaluation may be misleading. Among the factors we study, dataset selection, completion-only masking, and freezing attention layers during post-training have the largest impact on generation quality. Our best Hybrid-KDA model retains 86--90\% of teacher accuracy on knowledge benchmarks while reducing KV cache memory by up to 75\% and improving time-to-first-token by 2--4times at 128K-token contexts.

The prefill stage in long-context LLM inference remains a computational bottleneck. Recent token-ranking heuristics accelerate inference by selectively processing a subset of semantically relevant tokens. However, existing methods suffer from unstable token importance estimation, often varying between layers. Evaluating token-ranking quality independently from heuristic-specific architectures is challenging. To address this, we introduce an Answer-Informed Oracle, which defines ground-truth token importance by measuring attention from generated answers back to the prompt. This oracle reveals that existing heuristics exhibit high variance across layers: rankings can degrade sharply at specific layers, a failure mode invisible to end-to-end benchmarks. The diagnosis suggests a simple fix: aggregate scores across layers rather than relying on any single one. We implement this as Cross-Layer Attention Aggregation (CLAA), which closes the gap to the oracle upper bound and reduces Time-to-First-Token (TTFT) by up to 39\% compared to the Full KV Cache baseline.

We present Chatterbox-Flash, a zero-shot text-to-speech model obtained by fine-tuning a pretrained autoregressive TTS decoder into a block-diffusion decoder, enabling parallel token generation within each block while retaining block-by-block streaming. We find that naively transferring mainstream block-diffusion decoding to discrete speech tokens degrades quality, as a long-tail token distribution biases parallel position selection toward a few high-frequency tokens. To mitigate this without architectural modification, we introduce two inference-time techniques: prior-calibrated scoring, which subtracts the block-level marginal token distribution, and an early-decoding schedule, which adaptively terminates iteration based on calibrated confidence. On standard zero-shot TTS benchmarks, Chatterbox-Flash attains high-fidelity synthesis comparable to strong autoregressive and non-autoregressive baselines, while supporting streaming inference with time-to-first-packet on par with streaming AR systems and substantially lower real-time factor. Code and audio samples are available at https://github.com/resemble-ai/chatterbox-flash.

Deploying multiple models within shared GPU clusters is a key strategy to improve resource efficiency in large language model (LLM) serving. Existing multi-LLM serving systems improve GPU utilization at the cost of degraded inference performance, particularly time-to-first-token (TTFT). We attribute this degradation to the lack of awareness regarding future workload characteristics. In contrast, recent analyses have shown the strong periodicity and long-term predictability of real-world LLM serving workloads. In this paper, we propose one-for-many GPU prewarming, which proactively loads parameters from multiple models onto GPUs based on workload forecasts. These prewarmed weights enable the system to promptly instantiate serving instances upon encountering request bursts. We design and implement WarmServe, a multi-LLM serving system incorporating three key techniques: (1) a model placement algorithm that optimizes prewarming decisions to minimize cross-model prewarming interference, (2) a KV cache reservation strategy that repurposes idle KV cache space on running GPUs for prewarming new models, and (3) an efficient GPU memory switching mechanism for tensor management. Evaluation on real-world datasets shows that WarmServe reduces tail TTFT by up to 50.8times compared to the state-of-the-art autoscaling-based system, while supporting up to 2.5times higher request throughput than the GPU-sharing system.

Moltbook is a social network where every participant is an AI agent. We analyze 1,312,238 posts, 6.7~million comments, and over 120,000 agent profiles across 5,400 communities, collected over 40 days (January 27 to March 9, 2026). We evaluate the platform through three layers. At the interaction layer, 91.4% of post authors never return to their own threads, 85.6% of conversations are flat (no reply ever receives a reply), the median time-to-first-comment is 55 seconds, and 97.3% of comments receive zero upvotes. Interaction reciprocity is 3.3%, compared to 22-60% on human platforms. An argumentation analysis finds that 64.6% of comment-to-post relations carry no argumentative connection. At the content layer, 97.9% of agents never post in a community matching their bio, 92.5% of communities contain every topic in roughly equal proportions, and over 80% of shared URLs point to the platform's own infrastructure. At the instruction layer, we use 41 Wayback Machine snapshots to identify six instruction changes during the observation window. Hard constraints (rate limit, content filters) produce immediate behavioral shifts. Soft guidance (``upvote good posts'', ``stay on topic'') is ignored until it becomes an explicit step in the executable checklist. The platform also poses technological risks. We document credential leaks (API keys, JWT tokens), 12,470 unique Ethereum addresses with 3,529 confirmed transaction histories, and attack discourse ranging from template-based SSH brute-forcing to multi-agent offensive security architectures. These persist unmoderated because the quality-filtering mechanisms are themselves non-functional. Moltbook is a socio-technical system where the technical layer responds to changes, but the social layer largely fails to emerge. The form of social media is reproduced in full. The function is absent.

We describe KVLink, an approach for efficient key-value (KV) cache reuse in large language models (LLMs). In many LLM applications, different inputs can share overlapping context, such as the same retrieved document appearing in multiple queries. However, the LLMs still need to encode the entire context for each query, leading to redundant computation. In this paper, we investigate a new strategy to eliminate such inefficiency, where the KV cache of each document is precomputed independently. During inference, the KV caches of retrieved documents are concatenated, allowing the model to reuse cached representations instead of recomputing them. To mitigate the performance degradation when using KV caches computed independently for each document, KVLink introduces two key techniques: adjusting positional embeddings of the KV cache at inference to match the global position after concatenation, and using trainable special tokens to restore self-attention across independently encoded documents. Experiments across 7 datasets demonstrate that KVLink improves question answering accuracy by an average of 4% over state-of-the-art methods. Furthermore, by leveraging precomputed KV caches, our approach reduces time-to-first-token by up to 96% compared to standard LLM inference, making it a scalable and efficient solution for context reuse. Additionally, KVLink can be combined with KV cache compression to further save cache loading and storage overhead while outperforming the baselines.

As large language models (LLMs) evolve to handle increasingly longer contexts, serving inference requests for context lengths in the range of millions of tokens presents unique challenges. While existing techniques are effective for training, they fail to address the unique challenges of inference, such as varying prefill and decode phases and their associated latency constraints - like Time to First Token (TTFT) and Time Between Tokens (TBT). Furthermore, there are no long context inference solutions that allow batching requests to increase the hardware utilization today. In this paper, we propose three key innovations for efficient interactive long context LLM inference, without resorting to any approximation: adaptive chunking to reduce prefill overheads in mixed batching, Sequence Pipeline Parallelism (SPP) to lower TTFT, and KV Cache Parallelism (KVP) to minimize TBT. These contributions are combined into a 3D parallelism strategy, enabling Mnemosyne to scale interactive inference to context lengths at least up to 10 million tokens with high throughput enabled with batching. To our knowledge, Mnemosyne is the first to be able to achieve support for 10 million long context inference efficiently, while satisfying production-grade SLOs on TBT (30ms) on contexts up to and including 10 million.

Large Language Model or LLM inference has two phases, the prompt (or prefill) phase to output the first token and the extension (or decoding) phase to the generate subsequent tokens. In this work, we propose an efficient parallelization scheme, KV-Runahead to accelerate the prompt phase. The key observation is that the extension phase generates tokens faster than the prompt phase because of key-value cache (KV-cache). Hence, KV-Runahead parallelizes the prompt phase by orchestrating multiple processes to populate the KV-cache and minimizes the time-to-first-token (TTFT). Dual-purposing the KV-cache scheme has two main benefits. Fist, since KV-cache is designed to leverage the causal attention map, we minimize computation and computation automatically. Second, since it already exists for the exten- sion phase, KV-Runahead is easy to implement. We further propose context-level load-balancing to handle uneven KV-cache generation (due to the causal attention) and to optimize TTFT. Compared with an existing parallelization scheme such as tensor or sequential parallelization where keys and values are locally generated and exchanged via all-gather collectives, our experimental results demonstrate that KV-Runahead can offer over 1.4x and 1.6x speedups for Llama 7B and Falcon 7B respectively.

Interactive humanoid video generation aims to synthesize lifelike visual agents that can engage with humans through continuous and responsive video. Despite recent advances in video synthesis, existing methods often grapple with the trade-off between high-fidelity synthesis and real-time interaction requirements. In this paper, we propose FlowAct-R1, a framework specifically designed for real-time interactive humanoid video generation. Built upon a MMDiT architecture, FlowAct-R1 enables the streaming synthesis of video with arbitrary durations while maintaining low-latency responsiveness. We introduce a chunkwise diffusion forcing strategy, complemented by a novel self-forcing variant, to alleviate error accumulation and ensure long-term temporal consistency during continuous interaction. By leveraging efficient distillation and system-level optimizations, our framework achieves a stable 25fps at 480p resolution with a time-to-first-frame (TTFF) of only around 1.5 seconds. The proposed method provides holistic and fine-grained full-body control, enabling the agent to transition naturally between diverse behavioral states in interactive scenarios. Experimental results demonstrate that FlowAct-R1 achieves exceptional behavioral vividness and perceptual realism, while maintaining robust generalization across diverse character styles.

Hybrid models that combine the language modeling capabilities of Attention layers with the efficiency of Recurrent layers (e.g., State Space Models) have gained traction in practically supporting long contexts in Large Language Model serving. Yet, the unique properties of these models complicate the usage of complementary efficiency optimizations such as prefix caching that skip redundant computations across requests. Most notably, their use of in-place state updates for recurrent layers precludes rolling back cache entries for partial sequence overlaps, and instead mandates only exact-match cache hits; the effect is a deluge of (large) cache entries per sequence, most of which yield minimal reuse opportunities. We present Marconi, the first system that supports efficient prefix caching with Hybrid LLMs. Key to Marconi are its novel admission and eviction policies that more judiciously assess potential cache entries based not only on recency, but also on (1) forecasts of their reuse likelihood across a taxonomy of different hit scenarios, and (2) the compute savings that hits deliver relative to memory footprints. Across diverse workloads and Hybrid models, Marconi achieves up to 34.4times higher token hit rates (71.1% or 617 ms lower TTFT) compared to state-of-the-art prefix caching systems.

Large language models (LLMs) often incorporate multiple text chunks in their inputs to provide the necessary contexts. To speed up the prefill of the long LLM inputs, one can pre-compute the KV cache of a text and re-use the KV cache when the context is reused as the prefix of another LLM input. However, the reused text chunks are not always the input prefix, and when they are not, their precomputed KV caches cannot be directly used since they ignore the text's cross-attention with the preceding text in the LLM input. Thus, the benefits of reusing KV caches remain largely unrealized. This paper tackles just one question: when an LLM input contains multiple text chunks, how to quickly combine their precomputed KV caches in order to achieve the same generation quality as the expensive full prefill (i.e., without reusing KV cache)? We present CacheBlend, a scheme that reuses the pre-computed KV caches, regardless prefix or not, and selectively recomputes the KV values of a small subset of tokens to partially update each reused KV cache. In the meantime,the small extra delay for recomputing some tokens can be pipelined with the retrieval of KV caches within the same job,allowing CacheBlend to store KV caches in slower devices with more storage capacity while retrieving them without increasing the inference delay. By comparing CacheBlend with the state-of-the-art KV cache reusing schemes on three open-source LLMs of various sizes and four popular benchmark datasets of different tasks, we show that CacheBlend reduces time-to-first-token (TTFT) by 2.2-3.3X and increases the inference throughput by 2.8-5X, compared with full KV recompute, without compromising generation quality or incurring more storage cost.

Visual encoding followed by token condensing has become the standard architectural paradigm in multi-modal large language models (MLLMs). Many recent MLLMs increasingly favor global native- resolution visual encoding over slice-based methods. To investigate this trend, we systematically compare their behavior on vision-language understanding and attention patterns, revealing that global encoding enhances overall capability but at the expense of greater computational overhead. To address this issue, we present LLaVA-UHD v3, an MLLM centered upon our proposed Progressive Visual Compression (PVC) method, which can be seamlessly integrated into standard Vision Transformer (ViT) to enable efficient native-resolution encoding. The PVC approach consists of two key modules: (i) refined patch embedding, which supports flexible patch-size scaling for fine-grained visual model- ing, (ii) windowed token compression, hierarchically deployed across ViT layers to progressively aggregate local token representations. Jointly modulated by these two modules, a widely pretrained ViT can be reconfigured into an efficient architecture while largely preserving generality. Evaluated across extensive benchmarks, the transformed ViT, termed ViT-UHD, demonstrates competitive performance with MoonViT while reducing TTFT (time-to-first-token) by 2.4x, when developed within an identical MLLM architecture. Building upon ViT-UHD, LLaVA-UHD v3 also achieves competitive performance to Qwen2-VL, while further reducing TTFT by 1.9x. We will release all code and checkpoints to support future research on efficient MLLMs.

As large language models become increasingly prevalent, memory bandwidth constraints significantly limit inference throughput, motivating post-training quantization (PTQ). In this paper, we propose FireQ, a co-designed PTQ framework and an INT4-FP8 matrix multiplication kernel that accelerates LLM inference across all linear layers. Specifically, FireQ quantizes linear layer weights and key-values to INT4, and activations and queries to FP8, significantly enhancing throughput. Additionally, we introduce a three-stage pipelining for the prefill phase, which modifies the FlashAttention-3 kernel, effectively reducing time-to-first-token in the prefill phase. To minimize accuracy loss from quantization, we develop novel outlier smoothing techniques tailored separately for linear and attention layers. In linear layers, we explicitly use per-tensor scaling to prevent underflow caused by the FP8 quantization scaling factor of INT4 quantization, and channel-wise scaling to compensate for coarse granularity of INT4. In attention layers, we address quantization challenges posed by rotary positional embeddings (RoPE) by combining pre-RoPE and post-RoPE scaling strategies. FireQ significantly outperforms state-of-the-art methods, achieving 1.68x faster inference in feed-forward network layers on Llama2-7B and 1.26x faster prefill phase performance on Llama3-8B compared to QServe, with negligible accuracy loss.

Serving numerous users and requests concurrently requires good fairness in Large Language Models (LLMs) serving system. This ensures that, at the same cost, the system can meet the Service Level Objectives (SLOs) of more users , such as time to first token (TTFT) and time between tokens (TBT), rather than allowing a few users to experience performance far exceeding the SLOs. To achieve better fairness, the preemption-based scheduling policy dynamically adjusts the priority of each request to maintain balance during runtime. However, existing systems tend to overly prioritize throughput, overlooking the overhead caused by preemption-induced context switching, which is crucial for maintaining fairness through priority adjustments. In this work, we identify three main challenges that result in this overhead. 1) Inadequate I/O utilization. 2) GPU idleness. 3) Unnecessary I/O transmission during multi-turn conversations. Our key insight is that the block-based KV cache memory policy in existing systems, while achieving near-zero memory waste, leads to discontinuity and insufficient granularity in the KV cache memory. To respond, we introduce FastSwitch, a fairness-aware serving system that not only aligns with existing KV cache memory allocation policy but also mitigates context switching overhead. Our evaluation shows that FastSwitch outperforms the state-of-the-art LLM serving system vLLM with speedups of 1.4-11.2x across different tail TTFT and TBT.

On-device large language models (LLMs) are catalyzing novel mobile applications such as UI task automation and personalized email auto-reply, without giving away users' private data. However, on-device LLMs still suffer from unacceptably long inference latency, especially the time to first token (prefill stage) due to the need of long context for accurate, personalized content generation, as well as the lack of parallel computing capacity of mobile CPU/GPU. To enable practical on-device LLM, we present mllm-NPU, the first-of-its-kind LLM inference system that efficiently leverages on-device Neural Processing Unit (NPU) offloading. Essentially, mllm-NPU is an algorithm-system co-design that tackles a few semantic gaps between the LLM architecture and contemporary NPU design. Specifically, it re-constructs the prompt and model in three levels: (1) At prompt level, it divides variable-length prompts into multiple fixed-sized chunks while maintaining data dependencies; (2) At tensor level, it identifies and extracts significant outliers to run on the CPU/GPU in parallel with minimal overhead; (3) At block level, it schedules Transformer blocks in an out-of-order manner to the CPU/GPU and NPU based on their hardware affinity and sensitivity to accuracy. Compared to competitive baselines, mllm-NPU achieves 22.4x faster prefill speed and 30.7x energy savings on average, and up to 32.8x speedup in an end-to-end real-world application. For the first time, mllm-NPU achieves more than 1,000 tokens/sec prefilling for a billion-sized model (Qwen1.5-1.8B), paving the way towards practical on-device LLM.

Large language models (LLMs) now support extremely long context windows, but the quadratic complexity of vanilla attention results in significantly long Time-to-First-Token (TTFT) latency. Existing approaches to address this complexity require additional pretraining or finetuning, and often sacrifice model accuracy. In this paper, we first provide both theoretical and empirical foundations for near-lossless sparse attention. We find dynamically capturing head-specific sparse patterns at runtime with low overhead is crucial. To address this, we propose SampleAttention, an adaptive structured and near-lossless sparse attention. Leveraging observed significant sparse patterns, SampleAttention attends to a fixed percentage of adjacent tokens to capture local window patterns, and employs a two-stage query-guided key-value filtering approach, which adaptively select a minimum set of key-values with low overhead, to capture column stripe patterns. Comprehensive evaluations show that SampleAttention can seamlessly replace vanilla attention in off-the-shelf LLMs with nearly no accuracy loss, and reduces TTFT by up to 2.42times compared with FlashAttention.