Daily Papers

Recent advancements in speech-to-speech dialogue systems leverage LLMs for multimodal interactions, yet they remain hindered by fine-tuning requirements, high computational overhead, and text-speech misalignment. Existing speech-enabled LLMs often degrade conversational quality by modifying the LLM, thereby compromising its linguistic capabilities. In contrast, we propose LLMVoX, a lightweight 30M-parameter, LLM-agnostic, autoregressive streaming TTS system that generates high-quality speech with low latency, while fully preserving the capabilities of the base LLM. Our approach achieves a significantly lower Word Error Rate compared to speech-enabled LLMs, while operating at comparable latency and UTMOS score. By decoupling speech synthesis from LLM processing via a multi-queue token streaming system, LLMVoX supports seamless, infinite-length dialogues. Its plug-and-play design also facilitates extension to various tasks with different backbones. Furthermore, LLMVoX generalizes to new languages with only dataset adaptation, attaining a low Character Error Rate on an Arabic speech task. Additionally, we have integrated LLMVoX with a Vision-Language Model to create an omni-model with speech, text, and vision capabilities, without requiring additional multimodal training. Our code base and project page is available at https://mbzuai-oryx.github.io/LLMVoX .

Distributed inference of large language models (LLMs) can introduce overheads of up to 20% even over GPUs connected via high-speed interconnects such as NVLINK. Multiple techniques have been proposed to mitigate these overheads by decomposing computations into finer-grained tasks and overlapping communication with sub-tasks as they complete. However, fine-grained decomposition of a large computation into many smaller computations on GPUs results in overheads. Further, the communication itself uses many streaming multiprocessors (SMs), adding to the overhead. We present TokenWeave to address these challenges. TokenWeave proposes a Token-Splitting technique that divides the tokens in the inference batch into two approximately equal subsets in a wave-aware manner. The computation of one subset is then overlapped with the communication of the other. In addition, TokenWeave optimizes the order of the layer normalization computation with respect to communication operations and implements a novel fused AllReduce-RMSNorm kernel carefully leveraging Multimem instruction support available on NVIDIA Hopper GPUs. These optimizations allow TokenWeave to perform communication and RMSNorm using only 2-8 SMs. Moreover, our kernel enables the memory bound RMSNorm to be overlapped with the other batch's computation, providing additional gains. Our evaluations demonstrate up to 29% latency gains and up to 26% throughput gains across multiple models and workloads. In several settings, TokenWeave results in better performance compared to an equivalent model with all communication removed.

Deploying Large Language Models (LLMs) in streaming applications such as multi-round dialogue, where long interactions are expected, is urgently needed but poses two major challenges. Firstly, during the decoding stage, caching previous tokens' Key and Value states (KV) consumes extensive memory. Secondly, popular LLMs cannot generalize to longer texts than the training sequence length. Window attention, where only the most recent KVs are cached, is a natural approach -- but we show that it fails when the text length surpasses the cache size. We observe an interesting phenomenon, namely attention sink, that keeping the KV of initial tokens will largely recover the performance of window attention. In this paper, we first demonstrate that the emergence of attention sink is due to the strong attention scores towards initial tokens as a ``sink'' even if they are not semantically important. Based on the above analysis, we introduce StreamingLLM, an efficient framework that enables LLMs trained with a finite length attention window to generalize to infinite sequence lengths without any fine-tuning. We show that StreamingLLM can enable Llama-2, MPT, Falcon, and Pythia to perform stable and efficient language modeling with up to 4 million tokens and more. In addition, we discover that adding a placeholder token as a dedicated attention sink during pre-training can further improve streaming deployment. In streaming settings, StreamingLLM outperforms the sliding window recomputation baseline by up to 22.2x speedup. Code and datasets are provided at https://github.com/mit-han-lab/streaming-llm.

Many LLM tasks are performed in large batches or even offline, and the performance indictor for which is throughput. These tasks usually show the characteristic of prefix sharing, where different prompt input can partially show the common prefix. However, the existing LLM inference engines tend to optimize the streaming requests and show limitations of supporting the large batched tasks with the prefix sharing characteristic. The existing solutions use the LRU-based cache to reuse the KV context of common prefix. The KV context that is about to be reused may prematurely be evicted with the implicit cache management. Even if not evicted, the lifetime of the shared KV context is extended since requests sharing the same context are not scheduled together, resulting in larger memory usage. These streaming oriented systems schedule the requests in the first-come-first-serve or similar order. As a result, the requests with larger ratio of decoding steps may be scheduled too late to be able to mix with the prefill chunks to increase the hardware utilization. Besides, the token and request number based batching can limit the size of token-batch, which keeps the GPU from saturating for the iterations dominated by decoding tokens. We propose BatchLLM to address the above problems. BatchLLM explicitly identifies the common prefixes globally. The requests sharing the same prefix will be scheduled together to reuse the KV context the best, which also shrinks the lifetime of common KV memory. BatchLLM reorders the requests and schedules the requests with larger ratio of decoding first to better mix the decoding tokens with the latter prefill chunks and applies memory-centric token batching to enlarge the token-batch sizes, which helps to increase the GPU utilization. Extensive evaluation shows that BatchLLM outperforms vLLM by 1.1x to 2x on a set of microbenchmarks and two typical industry workloads.

The advent of large language models (LLMs) has transformed text-based services, enabling capabilities ranging from real-time translation to AI-driven chatbots. However, existing serving systems primarily focus on optimizing server-side aggregate metrics like token generation throughput, ignoring individual user experience with streamed text. As a result, under high and/or bursty load, a significant number of users can receive unfavorable service quality or poor Quality-of-Experience (QoE). In this paper, we first formally define QoE of text streaming services, where text is delivered incrementally and interactively to users, by considering the end-to-end token delivery process throughout the entire interaction with the user. Thereafter, we propose Andes, a QoE-aware serving system that enhances user experience for LLM-enabled text streaming services. At its core, Andes strategically allocates contended GPU resources among multiple requests over time to optimize their QoE. Our evaluations demonstrate that, compared to the state-of-the-art LLM serving systems like vLLM, Andes improves the average QoE by up to 3.2times under high request rate, or alternatively, it attains up to 1.6times higher request rate while preserving high QoE.

Efficient parallelization of Large Language Models (LLMs) with long sequences is essential but challenging due to their significant computational and memory demands, particularly stemming from communication bottlenecks in attention mechanisms. While sequence parallelism (SP) has been introduced as a potential solution, existing methods often suffer from limited scalability or inefficiency, rendering their effectiveness. Ring-Attention demonstrates the potential for scaling sequence processing but faces significant limitations due to its reliance on peer-to-peer (P2P) communication and inefficient utilization of network resources. As the degree of SP increases, the quadratic decrease in computation time per step contrasts sharply with the linear reduction in communication volume, exacerbating communication bottlenecks. To address these challenges, we propose TokenRing, a fine-grained parallel framework that leverages bidirectional P2P communication to effectively overlap computation and data transmission. By partitioning the attention block and concurrently transmitting Query and block outputs (i.e., block_out and block_lse) within a fully connected mesh topology, TokenRing achieves significant reductions in communication overhead and better load balancing. These innovations improve the scalability and efficiency of distributed Transformer models, particularly for long-context sequences. Experimental results demonstrate that TokenRing enhances throughput and reduces communication latency. Moreover, its design adapts seamlessly to various multi-GPU interconnect solutions, such as Huawei Ascend, ensuring broad compatibility and cost-effectiveness for distributed LLM inference and training. The code is available at: https://github.com/ACA-Lab-SJTU/token-ring.

Large language models (LLMs) have achieved impressive success, but their high memory requirements present challenges for long-context token generation. In this paper we study the streaming complexity of attention approximation, a key computational primitive underlying token generation. Our main contribution is BalanceKV, a streaming algorithm for epsilon-approximating attention computations based on geometric process for selecting a balanced collection of Key and Value tokens as per Banaszczyk's vector balancing theory. We complement our algorithm with space lower bounds for streaming attention computation. Besides strong theoretical guarantees, BalanceKV exhibits empirically validated performance improvements over existing methods, both for attention approximation and end-to-end performance on various long context benchmarks.

Large Language Models have shown remarkable efficacy in generating streaming data such as text and audio, thanks to their temporally uni-directional attention mechanism, which models correlations between the current token and previous tokens. However, video streaming remains much less explored, despite a growing need for live video processing. State-of-the-art video diffusion models leverage bi-directional temporal attention to model the correlations between the current frame and all the surrounding (i.e. including future) frames, which hinders them from processing streaming videos. To address this problem, we present Live2Diff, the first attempt at designing a video diffusion model with uni-directional temporal attention, specifically targeting live streaming video translation. Compared to previous works, our approach ensures temporal consistency and smoothness by correlating the current frame with its predecessors and a few initial warmup frames, without any future frames. Additionally, we use a highly efficient denoising scheme featuring a KV-cache mechanism and pipelining, to facilitate streaming video translation at interactive framerates. Extensive experiments demonstrate the effectiveness of the proposed attention mechanism and pipeline, outperforming previous methods in terms of temporal smoothness and/or efficiency.

Speculative decoding accelerates LLM inference by using a draft model to look ahead, but gains are capped by the cost of autoregressive draft generation: increasing draft size elevates acceptance rates but introduces additional latency overhead exacerbating the speed-accuracy tradeoff. Prior methods (Medusa, Hydra, EAGLE) partially reduce draft cost but either degrade acceptance or introduce overheads that limit scaling. We present Mirror Speculative Decoding (Mirror-SD), an inference algorithm that breaks the latency-acceptance tradeoff. Mirror-SD launches branch-complete rollouts from early-exit signals in parallel with the target model's suffix and explicitly maps computation across heterogeneous accelerators (GPU and NPU) to exploit cross-device parallelism. The draft speculates forward continuations for the target to verify, while the target simultaneously speculates correction paths for the draft, converting speculation into two complementary execution pipelines. To further cut draft latency without weakening acceptance semantics, we add speculative streaming so the draft emits multiple tokens per step. This dual strategy of parallel heterogeneous execution plus multi-token speculative streaming pushes speculative decoding toward its ideal regime of high acceptance with low overhead. On SpecBench with server-scale models from 14B to 66B parameters, Mirror-SD delivers consistent end-to-end gains, achieving 2.8x-5.8x wall-time speedups across diverse tasks and a 30% average relative improvement over the strongest baseline, EAGLE3.

Speculative decoding is a prominent technique to speed up the inference of a large target language model based on predictions of an auxiliary draft model. While effective, in application-specific settings, it often involves fine-tuning both draft and target models to achieve high acceptance rates. As the number of downstream tasks grows, these draft models add significant complexity to inference systems. We propose Speculative Streaming, a single-model speculative decoding method that fuses drafting into the target model by changing the fine-tuning objective from next token prediction to future n-gram prediction. Speculative Streaming speeds up decoding by 1.8 - 3.1X in a diverse set of tasks, such as Summarization, Structured Queries, and Meaning Representation, without sacrificing generation quality. Additionally, Speculative Streaming is parameter-efficient. It achieves on-par/higher speed-ups than Medusa-style architectures while using ~10000X fewer extra parameters, making it well-suited for resource-constrained devices.

The growing need for instant spoken language transcription and translation is driven by increased global communication and cross-lingual interactions. This has made offering translations in multiple languages essential for user applications. Traditional approaches to automatic speech recognition (ASR) and speech translation (ST) have often relied on separate systems, leading to inefficiencies in computational resources, and increased synchronization complexity in real time. In this paper, we propose a streaming Transformer-Transducer (T-T) model able to jointly produce many-to-one and one-to-many transcription and translation using a single decoder. We introduce a novel method for joint token-level serialized output training based on timestamp information to effectively produce ASR and ST outputs in the streaming setting. Experiments on {it,es,de}->en prove the effectiveness of our approach, enabling the generation of one-to-many joint outputs with a single decoder for the first time.

Large language models (LLMs) propel the prosperity of interactive AI applications showcased by ChatGPT that demand timely response of inference services. However, LLM inference is computation intensive and memory intensive, and improper parameter configuration at LLM platforms may exacerbate the inference time. In this paper, we analyze the impact of LLM output token distribution on the inference queueing delay, where the max-token clipping and the batched inference are considered. By formulating an M/G/1 model, we observe that enforcing a maximum output token limit on a very small fraction of inference requests can significantly reduce the queueing delay, and our model facilitates the selection of the optimal limit. For the batch inference, we model the service process as a bulk queue in which the batch processing time is affected by the batch size and the maximum token size inside this batch jointly. The queueing delays of the batching of all buffered requests (dynamic batching), the batching of constant number of requests (fixed batching), and the batching without intra-batch waiting (elastic batching) are derived. Experimental results show that our mathematical models coincide with the event-driven simulations well.

Recent advances in Large Language Models (LLMs) have enabled the development of Video-LLMs, advancing multimodal learning by bridging video data with language tasks. However, current video understanding models struggle with processing long video sequences, supporting multi-turn dialogues, and adapting to real-world dynamic scenarios. To address these issues, we propose StreamChat, a training-free framework for streaming video reasoning and conversational interaction. StreamChat leverages a novel hierarchical memory system to efficiently process and compress video features over extended sequences, enabling real-time, multi-turn dialogue. Our framework incorporates a parallel system scheduling strategy that enhances processing speed and reduces latency, ensuring robust performance in real-world applications. Furthermore, we introduce StreamBench, a versatile benchmark that evaluates streaming video understanding across diverse media types and interactive scenarios, including multi-turn interactions and complex reasoning tasks. Extensive evaluations on StreamBench and other public benchmarks demonstrate that StreamChat significantly outperforms existing state-of-the-art models in terms of accuracy and response times, confirming its effectiveness for streaming video understanding. Code is available at StreamChat: https://github.com/hmxiong/StreamChat.

We propose TETRIS, a novel method that optimizes the total throughput of batch speculative decoding in multi-request settings. Unlike existing methods that optimize for a single request or a group of requests as a whole, TETRIS actively selects the most promising draft tokens (for every request in a batch) to be accepted when verified in parallel, resulting in fewer rejected tokens and hence less wasted computing resources. Such an effective resource utilization to achieve fast inference in large language models (LLMs) is especially important to service providers with limited inference capacity. Compared to baseline speculative decoding, TETRIS yields a consistently higher acceptance rate and more effective utilization of the limited inference capacity. We show theoretically and empirically that TETRIS outperforms baseline speculative decoding and existing methods that dynamically select draft tokens, leading to a more efficient batch inference in LLMs.

Each LLM serving request goes through two phases. The first is prefill which processes the entire input prompt to produce one output token and the second is decode which generates the rest of output tokens, one-at-a-time. Prefill iterations have high latency but saturate GPU compute due to parallel processing of the input prompt. In contrast, decode iterations have low latency but also low compute utilization because a decode iteration processes only a single token per request. This makes batching highly effective for decodes and consequently for overall throughput. However, batching multiple requests leads to an interleaving of prefill and decode iterations which makes it challenging to achieve both high throughput and low latency. We introduce an efficient LLM inference scheduler Sarathi-Serve inspired by the techniques we originally proposed for optimizing throughput in Sarathi. Sarathi-Serve leverages chunked-prefills from Sarathi to create stall-free schedules that can add new requests in a batch without pausing ongoing decodes. Stall-free scheduling unlocks the opportunity to improve throughput with large batch sizes while minimizing the effect of batching on latency. Our evaluation shows that Sarathi-Serve improves serving throughput within desired latency SLOs of Mistral-7B by up to 2.6x on a single A100 GPU and up to 6.9x for Falcon-180B on 8 A100 GPUs over Orca and vLLM.

Existing zero-shot text-to-speech (TTS) systems are typically designed to process complete sentences and are constrained by the maximum duration for which they have been trained. However, in many streaming applications, texts arrive continuously in short chunks, necessitating instant responses from the system. We identify the essential capabilities required for chunk-level streaming and introduce LiveSpeech 2, a stream-aware model that supports infinitely long speech generation, text-audio stream synchronization, and seamless transitions between short speech chunks. To achieve these, we propose (1) adopting Mamba, a class of sequence modeling distinguished by linear-time decoding, which is augmented by cross-attention mechanisms for conditioning, (2) utilizing rotary positional embeddings in the computation of cross-attention, enabling the model to process an infinite text stream by sliding a window, and (3) decoding with semantic guidance, a technique that aligns speech with the transcript during inference with minimal overhead. Experimental results demonstrate that our models are competitive with state-of-the-art language model-based zero-shot TTS models, while also providing flexibility to support a wide range of streaming scenarios.

The auto-regressive decoding of Large Language Models (LLMs) results in significant overheads in their hardware performance. While recent research has investigated various speculative decoding techniques for multi-token generation, these efforts have primarily focused on improving processing speed such as throughput. Crucially, they often neglect other metrics essential for real-life deployments, such as memory consumption and training cost. To overcome these limitations, we propose a novel parallel prompt decoding that requires only 0.0002% trainable parameters, enabling efficient training on a single A100-40GB GPU in just 16 hours. Inspired by the human natural language generation process, PPD approximates outputs generated at future timesteps in parallel by using multiple prompt tokens. This approach partially recovers the missing conditional dependency information necessary for multi-token generation, resulting in up to a 28% higher acceptance rate for long-range predictions. Furthermore, we present a hardware-aware dynamic sparse tree technique that adaptively optimizes this decoding scheme to fully leverage the computational capacities on different GPUs. Through extensive experiments across LLMs ranging from MobileLlama to Vicuna-13B on a wide range of benchmarks, our approach demonstrates up to 2.49times speedup and maintains a minimal runtime memory overhead of just 0.0004%. More importantly, our parallel prompt decoding can serve as an orthogonal optimization for synergistic integration with existing speculative decoding, showing up to 1.22times further speed improvement. Our code is available at https://github.com/hmarkc/parallel-prompt-decoding.

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.

Sparse Mixture of Experts (MoE) models, while outperforming dense Large Language Models (LLMs) in terms of performance, face significant deployment challenges during inference due to their high memory demands. Existing offloading techniques, which involve swapping activated and idle experts between the GPU and CPU, often suffer from rigid expert caching mechanisms. These mechanisms fail to adapt to dynamic routing, leading to inefficient cache utilization, or incur prohibitive costs for prediction training. To tackle these inference-specific challenges, we introduce ExpertFlow, a comprehensive system specifically designed to enhance inference efficiency by accommodating flexible routing and enabling efficient expert scheduling between CPU and GPU. This reduces overhead and boosts system performance. Central to our approach is a predictive routing path-based offloading mechanism that utilizes a lightweight predictor to accurately forecast routing paths before computation begins. This proactive strategy allows for real-time error correction in expert caching, significantly increasing cache hit ratios and reducing the frequency of expert transfers, thereby minimizing I/O overhead. Additionally, we implement a dynamic token scheduling strategy that optimizes MoE inference by rearranging input tokens across different batches. This method not only reduces the number of activated experts per batch but also improves computational efficiency. Our extensive experiments demonstrate that ExpertFlow achieves up to 93.72\% GPU memory savings and enhances inference speed by 2 to 10 times compared to baseline methods, highlighting its effectiveness and utility as a robust solution for resource-constrained inference scenarios.

Recently, simultaneous translation has gathered a lot of attention since it enables compelling applications such as subtitle translation for a live event or real-time video-call translation. Some of these translation applications allow editing of partial translation giving rise to re-translation approaches. The current re-translation approaches are based on autoregressive sequence generation models (ReTA), which generate tar-get tokens in the (partial) translation sequentially. The multiple re-translations with sequential generation inReTAmodelslead to an increased inference time gap between the incoming source input and the corresponding target output as the source input grows. Besides, due to the large number of inference operations involved, the ReTA models are not favorable for resource-constrained devices. In this work, we propose a faster re-translation system based on a non-autoregressive sequence generation model (FReTNA) to overcome the aforementioned limitations. We evaluate the proposed model on multiple translation tasks and our model reduces the inference times by several orders and achieves a competitive BLEUscore compared to the ReTA and streaming (Wait-k) models.The proposed model reduces the average computation time by a factor of 20 when compared to the ReTA model by incurring a small drop in the translation quality. It also outperforms the streaming-based Wait-k model both in terms of computation time (1.5 times lower) and translation quality.

Scaling the size of language models to tens of billions of parameters has led to impressive performance on a wide range of tasks. At generation, these models are used auto-regressively, requiring a forward pass for each generated token, and thus reading the full set of parameters from memory. This memory access forms the primary bottleneck for generation and it worsens as the model size increases. Moreover, executing a forward pass for multiple tokens in parallel often takes nearly the same time as it does for just one token. These two observations lead to the development of speculative sampling, where a second smaller model is used to draft a few tokens, that are then validated or rejected using a single forward pass of the large model. Unfortunately, this method requires two models that share the same tokenizer and thus limits its adoption. As an alternative, we propose to use parallel decoding as a way to draft multiple tokens from a single model with no computational cost, nor the need for a second model. Our approach only requires an additional input token that marks the words that will be generated simultaneously. We show promising performance (up to 30% speed-up) while requiring only as few as O(d_{emb}) additional parameters.

The increasing number of different, incompatible congestion control algorithms has led to an increased deployment of fair queuing. Fair queuing isolates each network flow and can thus guarantee fairness for each flow even if the flows' congestion controls are not inherently fair. So far, each queue in the fair queuing system either has a fixed, static maximum size or is managed by an Active Queue Management (AQM) algorithm like CoDel. In this paper we design an AQM mechanism (Learning Fair Qdisc (LFQ)) that dynamically learns the optimal buffer size for each flow according to a specified reward function online. We show that our Deep Learning based algorithm can dynamically assign the optimal queue size to each flow depending on its congestion control, delay and bandwidth. Comparing to competing fair AQM schedulers, it provides significantly smaller queues while achieving the same or higher throughput.

Quantization is a widely-used compression technology to reduce the overhead of serving large language models (LLMs) on terminal devices and in cloud data centers. However, prevalent quantization methods, such as 8-bit weight-activation or 4-bit weight-only quantization, achieve limited performance improvements due to poor support for low-precision (e.g., 4-bit) activation. This work, for the first time, realizes practical W4A4KV4 serving for LLMs, fully utilizing the INT4 tensor cores on modern GPUs and reducing the memory bottleneck caused by the KV cache. Specifically, we propose a novel fine-grained mixed-precision quantization algorithm (FMPQ) that compresses most activations into 4-bit with negligible accuracy loss. To support mixed-precision matrix multiplication for W4A4 and W4A8, we develop a highly optimized W4Ax kernel. Our approach introduces a novel mixed-precision data layout to facilitate access and fast dequantization for activation and weight tensors, utilizing the GPU's software pipeline to hide the overhead of data loading and conversion. Additionally, we propose fine-grained streaming multiprocessor (SM) scheduling to achieve load balance across different SMs. We integrate the optimized W4Ax kernel into our inference framework, COMET, and provide efficient management to support popular LLMs such as LLaMA-3-70B. Extensive evaluations demonstrate that, when running LLaMA family models on a single A100-80G-SMX4, COMET achieves a kernel-level speedup of 2.88times over cuBLAS and a 2.02 times throughput improvement compared to TensorRT-LLM from an end-to-end framework perspective.

Large Language Models (LLMs) typically generate outputs token by token using a fixed compute budget, leading to inefficient resource utilization. To address this shortcoming, recent advancements in mixture of expert (MoE) models, speculative decoding, and early exit strategies leverage the insight that computational demands can vary significantly based on the complexity and nature of the input. However, identifying optimal routing patterns for dynamic execution remains an open challenge, limiting the full potential of these adaptive methods. To address this need, we study adaptive computation in LLMs more systematically. We propose a novel framework that integrates smaller auxiliary modules within each Feed-Forward Network layer of the LLM. This design enables dynamic routing of tokens based on task complexity: tokens can be processed by either the small or big modules at each layer, or even bypass certain layers entirely. This allows us to introduce a novel notion of a token's difficulty, defined by its potential to benefit from additional computational resources. Importantly, by employing oracles to identify optimal patterns of adaptive computations, we gain valuable insights into the internal workings of LLMs and the routing processes in a simplified heterogeneous MoE setup. We show that trained routers operate differently from oracles and often yield suboptimal solutions. Notably, activating a large module in just one layer outperforms models that use large modules across all layers, underscoring the gap between practical implementations of routing in MoE models and theoretical optima for adaptive computation.

Reinforcement learning (RL) has become a pivotal technology in the post-training phase of large language models (LLMs). Traditional task-colocated RL frameworks suffer from significant scalability bottlenecks, while task-separated RL frameworks face challenges in complex dataflows and the corresponding resource idling and workload imbalance. Moreover, most existing frameworks are tightly coupled with LLM training or inference engines, making it difficult to support custom-designed engines. To address these challenges, we propose AsyncFlow, an asynchronous streaming RL framework for efficient post-training. Specifically, we introduce a distributed data storage and transfer module that provides a unified data management and fine-grained scheduling capability in a fully streamed manner. This architecture inherently facilitates automated pipeline overlapping among RL tasks and dynamic load balancing. Moreover, we propose a producer-consumer-based asynchronous workflow engineered to minimize computational idleness by strategically deferring parameter update process within staleness thresholds. Finally, the core capability of AsynFlow is architecturally decoupled from underlying training and inference engines and encapsulated by service-oriented user interfaces, offering a modular and customizable user experience. Extensive experiments demonstrate an average of 1.59 throughput improvement compared with state-of-the-art baseline. The presented architecture in this work provides actionable insights for next-generation RL training system designs.

Recently, great progress has been achieved in text-to-video (T2V) generation by scaling transformer-based diffusion models to billions of parameters, which can generate high-quality videos. However, existing models typically produce only short clips offline, restricting their use cases in interactive and real-time applications. This paper addresses these challenges by proposing StreamDiT, a streaming video generation model. StreamDiT training is based on flow matching by adding a moving buffer. We design mixed training with different partitioning schemes of buffered frames to boost both content consistency and visual quality. StreamDiT modeling is based on adaLN DiT with varying time embedding and window attention. To practice the proposed method, we train a StreamDiT model with 4B parameters. In addition, we propose a multistep distillation method tailored for StreamDiT. Sampling distillation is performed in each segment of a chosen partitioning scheme. After distillation, the total number of function evaluations (NFEs) is reduced to the number of chunks in a buffer. Finally, our distilled model reaches real-time performance at 16 FPS on one GPU, which can generate video streams at 512p resolution. We evaluate our method through both quantitative metrics and human evaluation. Our model enables real-time applications, e.g. streaming generation, interactive generation, and video-to-video. We provide video results and more examples in our project website: <a href="https://cumulo-autumn.github.io/StreamDiT/">this https URL.</a>

Autoregressive models, despite their commendable performance in a myriad of generative tasks, face challenges stemming from their inherently sequential structure. Inference on these models, by design, harnesses a temporal dependency, where the current token's probability distribution is conditioned on preceding tokens. This inherent characteristic severely impedes computational efficiency during inference as a typical inference request can require more than thousands of tokens, where generating each token requires a load of entire model weights, making the inference more memory-bound. The large overhead becomes profound in real deployment where requests arrive randomly, necessitating various generation lengths. Existing solutions, such as dynamic batching and concurrent instances, introduce significant response delays and bandwidth contention, falling short of achieving optimal latency and throughput. To address these shortcomings, we propose Flover -- a temporal fusion framework for efficiently inferring multiple requests in parallel. We deconstruct the general generation pipeline into pre-processing and token generation, and equip the framework with a dedicated work scheduler for fusing the generation process temporally across all requests. By orchestrating the token-level parallelism, Flover exhibits optimal hardware efficiency and significantly spares the system resources. By further employing a fast buffer reordering algorithm that allows memory eviction of finished tasks, it brings over 11x inference speedup on GPT and 16x on LLAMA compared to the cutting-edge solutions provided by NVIDIA FasterTransformer. Crucially, by leveraging the advanced tensor parallel technique, Flover proves efficacious across diverse computational landscapes, from single-GPU setups to distributed scenarios, thereby offering robust performance optimization that adapts to variable use cases.

Sparsely-gated mixture-of-experts (MoE) has been widely adopted to scale deep learning models to trillion-plus parameters with fixed computational cost. The algorithmic performance of MoE relies on its token routing mechanism that forwards each input token to the right sub-models or experts. While token routing dynamically determines the amount of expert workload at runtime, existing systems suffer inefficient computation due to their static execution, namely static parallelism and pipelining, which does not adapt to the dynamic workload. We present Flex, a highly scalable stack design and implementation for MoE with dynamically adaptive parallelism and pipelining. Flex designs an identical layout for distributing MoE model parameters and input data, which can be leveraged by all possible parallelism or pipelining methods without any mathematical inequivalence or tensor migration overhead. This enables adaptive parallelism/pipelining optimization at zero cost during runtime. Based on this key design, Flex also implements various MoE acceleration techniques. Aggregating all techniques, Flex finally delivers huge speedup at any scale -- 4.96x and 5.75x speedup of a single MoE layer over 16 and 2,048 A100 GPUs, respectively, over the previous state-of-the-art. Our evaluation shows that Flex efficiently and effectively runs a real-world MoE-based model named SwinV2-MoE, built upon Swin Transformer V2, a state-of-the-art computer vision architecture. On efficiency, Flex accelerates SwinV2-MoE, achieving up to 1.55x and 2.11x speedup in training and inference over Fairseq, respectively. On effectiveness, the SwinV2-MoE model achieves superior accuracy in both pre-training and down-stream computer vision tasks such as COCO object detection than the counterpart dense model, indicating the readiness of Flex for end-to-end real-world model training and inference.

Prior works have demonstrated zero-shot text-to-speech by using a generative language model on audio tokens obtained via a neural audio codec. It is still challenging, however, to adapt them to low-latency scenarios. In this paper, we present LiveSpeech - a fully autoregressive language model-based approach for zero-shot text-to-speech, enabling low-latency streaming of the output audio. To allow multiple token prediction within a single decoding step, we propose (1) using adaptive codebook loss weights that consider codebook contribution in each frame and focus on hard instances, and (2) grouping codebooks and processing groups in parallel. Experiments show our proposed models achieve competitive results to state-of-the-art baselines in terms of content accuracy, speaker similarity, audio quality, and inference speed while being suitable for low-latency streaming applications.

The Mixture of Experts (MoE) is an effective architecture for scaling large language models by leveraging sparse expert activation, optimizing the trade-off between performance and efficiency. However, under expert parallelism, MoE suffers from inference inefficiencies due to imbalanced token-to-expert assignment, where some experts are overloaded while others remain underutilized. This imbalance leads to poor resource utilization and increased latency, as the most burdened expert dictates the overall delay, a phenomenon we define as the \textit{Straggler Effect}. To mitigate this, we propose Capacity-Aware Inference, including two key techniques: (1) \textit{Capacity-Aware Token Drop}, which discards overloaded tokens to regulate the maximum latency of MoE, and (2) \textit{Capacity-Aware Token Reroute}, which reallocates overflowed tokens to underutilized experts, balancing the token distribution. These techniques collectively optimize both high-load and low-load expert utilization, leading to a more efficient MoE inference pipeline. Extensive experiments demonstrate the effectiveness of our methods, showing significant improvements in inference efficiency, e.g., 0.2\% average performance increase and a 1.94times inference speedup on Mixtral-8times7B-Instruct.

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.

The rapid growth of online video platforms, particularly live streaming services, has created an urgent need for real-time video understanding systems. These systems must process continuous video streams and respond to user queries instantaneously, presenting unique challenges for current Video Large Language Models (VideoLLMs). While existing VideoLLMs excel at processing complete videos, they face significant limitations in streaming scenarios due to their inability to handle dense, redundant frames efficiently. We introduce TimeChat-Online, a novel online VideoLLM that revolutionizes real-time video interaction. At its core lies our innovative Differential Token Drop (DTD) module, which addresses the fundamental challenge of visual redundancy in streaming videos. Drawing inspiration from human visual perception's Change Blindness phenomenon, DTD preserves meaningful temporal changes while filtering out static, redundant content between frames. Remarkably, our experiments demonstrate that DTD achieves an 82.8% reduction in video tokens while maintaining 98% performance on StreamingBench, revealing that over 80% of visual content in streaming videos is naturally redundant without requiring language guidance. To enable seamless real-time interaction, we present TimeChat-Online-139K, a comprehensive streaming video dataset featuring diverse interaction patterns including backward-tracing, current-perception, and future-responding scenarios. TimeChat-Online's unique Proactive Response capability, naturally achieved through continuous monitoring of video scene transitions via DTD, sets it apart from conventional approaches. Our extensive evaluation demonstrates TimeChat-Online's superior performance on streaming benchmarks (StreamingBench and OvOBench) and maintaining competitive results on long-form video tasks such as Video-MME and MLVU.

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.

A naive application of state-of-the-art bidirectional encoders for streaming sequence tagging would require encoding each token from scratch for each new token in an incremental streaming input (like transcribed speech). The lack of re-usability of previous computation leads to a higher number of Floating Point Operations (or FLOPs) and higher number of unnecessary label flips. Increased FLOPs consequently lead to higher wall-clock time and increased label flipping leads to poorer streaming performance. In this work, we present a Hybrid Encoder with Adaptive Restart (HEAR) that addresses these issues while maintaining the performance of bidirectional encoders over the offline (or complete) inputs while improving performance on streaming (or incomplete) inputs. HEAR has a Hybrid unidirectional-bidirectional encoder architecture to perform sequence tagging, along with an Adaptive Restart Module (ARM) to selectively guide the restart of bidirectional portion of the encoder. Across four sequence tagging tasks, HEAR offers FLOP savings in streaming settings upto 71.1% and also outperforms bidirectional encoders for streaming predictions by upto +10% streaming exact match.

Large Language Models (LLMs) have exhibited exceptional performance across a spectrum of natural language processing tasks. However, their substantial sizes pose considerable challenges, particularly in computational demands and inference speed, due to their quadratic complexity. In this work, we have identified a key pattern: certain seemingly meaningless special tokens (i.e., separators) contribute disproportionately to attention scores compared to semantically meaningful tokens. This observation suggests that information of the segments between these separator tokens can be effectively condensed into the separator tokens themselves without significant information loss. Guided by this insight, we introduce SepLLM, a plug-and-play framework that accelerates inference by compressing these segments and eliminating redundant tokens. Additionally, we implement efficient kernels for training acceleration. Experimental results across training-free, training-from-scratch, and post-training settings demonstrate SepLLM's effectiveness. Notably, using the Llama-3-8B backbone, SepLLM achieves over 50% reduction in KV cache on the GSM8K-CoT benchmark while maintaining comparable performance. Furthermore, in streaming settings, SepLLM effectively processes sequences of up to 4 million tokens or more while maintaining consistent language modeling capabilities.

The deployment and scaling of large language models (LLMs) have become critical as they permeate various applications, demanding high-throughput and low-latency serving systems. Existing frameworks struggle to balance these requirements, especially for workloads with long prompts. This paper introduces DeepSpeed-FastGen, a system that employs Dynamic SplitFuse, a novel prompt and generation composition strategy, to deliver up to 2.3x higher effective throughput, 2x lower latency on average, and up to 3.7x lower (token-level) tail latency, compared to state-of-the-art systems like vLLM. We leverage a synergistic combination of DeepSpeed-MII and DeepSpeed-Inference to provide an efficient and easy-to-use serving system for LLMs. DeepSpeed-FastGen's advanced implementation supports a range of models and offers both non-persistent and persistent deployment options, catering to diverse user scenarios from interactive sessions to long-running applications. We present a detailed benchmarking methodology, analyze the performance through latency-throughput curves, and investigate scalability via load balancing. Our evaluations demonstrate substantial improvements in throughput and latency across various models and hardware configurations. We discuss our roadmap for future enhancements, including broader model support and new hardware backends. The DeepSpeed-FastGen code is readily available for community engagement and contribution.

Deployment of autoregressive large language models (LLMs) is costly, and as these models increase in size, the associated costs will become even more considerable. Consequently, different methods have been proposed to accelerate the token generation process and reduce costs. Speculative decoding (SD) is among the most promising approaches to speed up the LLM decoding process by verifying multiple tokens in parallel and using an auxiliary smaller draft model to generate the possible tokens. In SD, usually, one draft model is used to serve a specific target model; however, in practice, LLMs are diverse, and we might need to deal with many target models or more than one target model simultaneously. In this scenario, it is not clear which draft model should be used for which target model, and searching among different draft models or training customized draft models can further increase deployment costs. In this paper, we first introduce a novel multi-target scenario for the deployment of draft models for faster inference. Then, we present a novel, more efficient sorted speculative decoding mechanism that outperforms regular baselines in multi-target settings. We evaluated our method on Spec-Bench in different settings, including base models such as Vicuna 7B, 13B, and LLama Chat 70B. Our results suggest that our draft models perform better than baselines for multiple target models at the same time.

Mixture of Experts (MoE) models enhance neural network scalability by dynamically selecting relevant experts per input token, enabling larger model sizes while maintaining manageable computation costs. However, efficient training of large-scale MoE models across thousands of GPUs presents significant challenges due to limitations in existing parallelism strategies. We introduce an end-to-end training framework for large-scale MoE models that utilizes five-dimensional hybrid parallelism: Tensor Parallelism, Expert Parallelism, Context Parallelism, Data Parallelism, and Pipeline Parallelism. Central to our approach is MoE Parallel Folding, a novel strategy that decouples the parallelization of attention and MoE layers in Transformer models, allowing each layer type to adopt optimal parallel configurations. Additionally, we develop a flexible token-level dispatcher that supports both token-dropping and token-dropless MoE training across all five dimensions of parallelism. This dispatcher accommodates dynamic tensor shapes and coordinates different parallelism schemes for Attention and MoE layers, facilitating complex parallelism implementations. Our experiments demonstrate significant improvements in training efficiency and scalability. We achieve up to 49.3% Model Flops Utilization (MFU) for the Mixtral 8x22B model and 39.0% MFU for the Qwen2-57B-A14B model on H100 GPUs, outperforming existing methods. The framework scales efficiently up to 1,024 GPUs and maintains high performance with sequence lengths up to 128K tokens, validating its effectiveness for large-scale MoE model training. The code is available in Megatron-Core.

The rapid evolution of large language models (LLMs), driven by growing parameter scales, adoption of mixture-of-experts (MoE) architectures, and expanding context lengths, imposes unprecedented demands on AI infrastructure. Traditional AI clusters face limitations in compute intensity, memory bandwidth, inter-chip communication, and latency, compounded by variable workloads and strict service-level objectives. Addressing these issues requires fundamentally redesigned hardware-software integration. This paper introduces Huawei CloudMatrix, a next-generation AI datacenter architecture, realized in the production-grade CloudMatrix384 supernode. It integrates 384 Ascend 910C NPUs and 192 Kunpeng CPUs interconnected via an ultra-high-bandwidth Unified Bus (UB) network, enabling direct all-to-all communication and dynamic pooling of resources. These features optimize performance for communication-intensive operations, such as large-scale MoE expert parallelism and distributed key-value cache access. To fully leverage CloudMatrix384, we propose CloudMatrix-Infer, an advanced LLM serving solution incorporating three core innovations: a peer-to-peer serving architecture that independently scales prefill, decode, and caching; a large-scale expert parallelism strategy supporting EP320 via efficient UB-based token dispatch; and hardware-aware optimizations including specialized operators, microbatch-based pipelining, and INT8 quantization. Evaluation with the DeepSeek-R1 model shows CloudMatrix-Infer achieves state-of-the-art efficiency: prefill throughput of 6,688 tokens/s per NPU and decode throughput of 1,943 tokens/s per NPU (<50 ms TPOT). It effectively balances throughput and latency, sustaining 538 tokens/s even under stringent 15 ms latency constraints, while INT8 quantization maintains model accuracy across benchmarks.

Large language models such as GPT and Llama are trained with a next-token prediction loss. In this work, we suggest that training language models to predict multiple future tokens at once results in higher sample efficiency. More specifically, at each position in the training corpus, we ask the model to predict the following n tokens using n independent output heads, operating on top of a shared model trunk. Considering multi-token prediction as an auxiliary training task, we measure improved downstream capabilities with no overhead in training time for both code and natural language models. The method is increasingly useful for larger model sizes, and keeps its appeal when training for multiple epochs. Gains are especially pronounced on generative benchmarks like coding, where our models consistently outperform strong baselines by several percentage points. Our 13B parameter models solves 12 % more problems on HumanEval and 17 % more on MBPP than comparable next-token models. Experiments on small algorithmic tasks demonstrate that multi-token prediction is favorable for the development of induction heads and algorithmic reasoning capabilities. As an additional benefit, models trained with 4-token prediction are up to 3 times faster at inference, even with large batch sizes.

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.

Streaming visual transformers like StreamVGGT achieve strong 3D perception but suffer from unbounded growth of key value (KV) memory, which limits scalability. We propose a training-free, inference-time token eviction policy that bounds memory by discarding redundant tokens while keeping the most informative ones. Our method uses significantly less memory with little to no drop in accuracy: on 7-Scenes with long sequences it reduces peak memory from 18.63 GB to 9.39 GB while accuracy and completeness drop by only 0.003. Under strict memory budgets, eviction enables denser frame sampling, which improves reconstruction accuracy compared to the baseline. Experiments across video depth estimation (Sintel, KITTI), 3D reconstruction (7-Scenes, NRGBD), and camera pose estimation (Sintel, TUM-dynamics) show that our approach closely matches StreamVGGT at a fraction of the memory and makes long-horizon streaming inference more practical.

Generating ultra-long sequences with large language models (LLMs) has become increasingly crucial but remains a highly time-intensive task, particularly for sequences up to 100K tokens. While traditional speculative decoding methods exist, simply extending their generation limits fails to accelerate the process and can be detrimental. Through an in-depth analysis, we identify three major challenges hindering efficient generation: frequent model reloading, dynamic key-value (KV) management and repetitive generation. To address these issues, we introduce TOKENSWIFT, a novel framework designed to substantially accelerate the generation process of ultra-long sequences while maintaining the target model's inherent quality. Experimental results demonstrate that TOKENSWIFT achieves over 3 times speedup across models of varying scales (1.5B, 7B, 8B, 14B) and architectures (MHA, GQA). This acceleration translates to hours of time savings for ultra-long sequence generation, establishing TOKENSWIFT as a scalable and effective solution at unprecedented lengths. Code can be found at https://github.com/bigai-nlco/TokenSwift.

It is well known that LLM-based systems are data-hungry. Recent LLM-based TTS works typically employ complex data processing pipelines to obtain high-quality training data. These sophisticated pipelines require excellent models at each stage (e.g., speech denoising, speech enhancement, speaker diarization, and punctuation models), which themselves demand high-quality training data and are rarely open-sourced. Even with state-of-the-art models, issues persist, such as incomplete background noise removal and misalignment between punctuation and actual speech pauses. Moreover, the stringent filtering strategies often retain only 10-30\% of the original data, significantly impeding data scaling efforts. In this work, we leverage a noise-robust audio tokenizer (S3Tokenizer) to design a simplified yet effective TTS data processing pipeline that maintains data quality while substantially reducing data acquisition costs, achieving a data retention rate of over 50\%. Beyond data scaling challenges, LLM-based TTS systems also incur higher deployment costs compared to conventional approaches. Current systems typically use LLMs solely for text-to-token generation, while requiring separate models (e.g., flow matching models) for token-to-waveform generation, which cannot be directly executed by LLM inference engines, further complicating deployment. To address these challenges, we eliminate redundant modules in both LLM and flow components, replacing the flow model backbone with an LLM architecture. Building upon this simplified flow backbone, we propose a unified architecture for both streaming and non-streaming inference, significantly reducing deployment costs. Finally, we explore the feasibility of unifying TTS and ASR tasks using the same data for training, thanks to the simplified pipeline and the S3Tokenizer that reduces the quality requirements for TTS training data.

Transformer models serve as the backbone of many state-ofthe-art language models, and most use the scaled dot-product attention (SDPA) mechanism to capture relationships between tokens. However, the straightforward implementation of SDPA has quadratic compute and memory complexity with respect to the sequence length. On processor architectures such as GPUs and TPUs, there is a robust body of prior work. However, little work has been performed on non-processor architectures.In this work, we show how the architecture and execution model of Streaming Dataflow Accelerators can help tackle this challenge. We first define abstract hardware that adopts a streaming execution model, and we implement a cycle-accurate simulator of the abstract hardware using the Dataflow Abstract Machine simulation framework. Second, we implement the naive SDPA algorithm on this abstract hardware and show it requires linear (O(N)) intermediate memory. Third, we then modify the naive algorithm, taking inspiration from prior processor-oriented works, by reordering the multiplication and division operations. Finally, we map the modified algorithm to abstract hardware, and confirm that the implementation computes SDPA at full throughput while only using a constant amount (O(1)) of intermediate memory.

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.

Discrete audio tokens are compact representations that aim to preserve perceptual quality, phonetic content, and speaker characteristics while enabling efficient storage and inference, as well as competitive performance across diverse downstream tasks.They provide a practical alternative to continuous features, enabling the integration of speech and audio into modern large language models (LLMs). As interest in token-based audio processing grows, various tokenization methods have emerged, and several surveys have reviewed the latest progress in the field. However, existing studies often focus on specific domains or tasks and lack a unified comparison across various benchmarks. This paper presents a systematic review and benchmark of discrete audio tokenizers, covering three domains: speech, music, and general audio. We propose a taxonomy of tokenization approaches based on encoder-decoder, quantization techniques, training paradigm, streamability, and application domains. We evaluate tokenizers on multiple benchmarks for reconstruction, downstream performance, and acoustic language modeling, and analyze trade-offs through controlled ablation studies. Our findings highlight key limitations, practical considerations, and open challenges, providing insight and guidance for future research in this rapidly evolving area. For more information, including our main results and tokenizer database, please refer to our website: https://poonehmousavi.github.io/dates-website/.

Transformer-based models have achieved great success in various NLP, vision, and speech tasks. However, the core of Transformer, the self-attention mechanism, has a quadratic time and memory complexity with respect to the sequence length, which hinders applications of Transformer-based models to long sequences. Many approaches have been proposed to mitigate this problem, such as sparse attention mechanisms, low-rank matrix approximations and scalable kernels, and token mixing alternatives to self-attention. We propose a novel Pooling Network (PoNet) for token mixing in long sequences with linear complexity. We design multi-granularity pooling and pooling fusion to capture different levels of contextual information and combine their interactions with tokens. On the Long Range Arena benchmark, PoNet significantly outperforms Transformer and achieves competitive accuracy, while being only slightly slower than the fastest model, FNet, across all sequence lengths measured on GPUs. We also conduct systematic studies on the transfer learning capability of PoNet and observe that PoNet achieves 95.7% of the accuracy of BERT on the GLUE benchmark, outperforming FNet by 4.5% relative. Comprehensive ablation analysis demonstrates effectiveness of the designed multi-granularity pooling and pooling fusion for token mixing in long sequences and efficacy of the designed pre-training tasks for PoNet to learn transferable contextualized language representations.

While large language models (LLMs) have revolutionized text-to-speech (TTS) synthesis through discrete tokenization paradigms, current architectures exhibit fundamental tensions between three critical dimensions: 1) irreversible loss of acoustic characteristics caused by quantization of speech prompts; 2) stringent dependence on precisely aligned prompt speech-text pairs that limit real-world deployment; and 3) catastrophic forgetting of the LLM's native text comprehension during optimization for speech token generation. To address these challenges, we propose an LLM-based text-to-speech Generation approach Optimized via a novel dual-branch ArchiTecture (GOAT-TTS). Our framework introduces two key innovations: (1) The modality-alignment branch combines a speech encoder and projector to capture continuous acoustic embeddings, enabling bidirectional correlation between paralinguistic features (language, timbre, emotion) and semantic text representations without transcript dependency; (2) The speech-generation branch employs modular fine-tuning on top-k layers of an LLM for speech token prediction while freezing the bottom-k layers to preserve foundational linguistic knowledge. Moreover, multi-token prediction is introduced to support real-time streaming TTS synthesis. Experimental results demonstrate that our GOAT-TTS achieves performance comparable to state-of-the-art TTS models while validating the efficacy of synthesized dialect speech data.

Live animation has gained immense popularity for enhancing online engagement, yet achieving high-quality, real-time, and stable animation with diffusion models remains challenging, especially on consumer-grade GPUs. Existing methods struggle with generating long, consistent video streams efficiently, often being limited by latency issues and degraded visual quality over extended periods. In this paper, we introduce RAIN, a pipeline solution capable of animating infinite video streams in real-time with low latency using a single RTX 4090 GPU. The core idea of RAIN is to efficiently compute frame-token attention across different noise levels and long time-intervals while simultaneously denoising a significantly larger number of frame-tokens than previous stream-based methods. This design allows RAIN to generate video frames with much shorter latency and faster speed, while maintaining long-range attention over extended video streams, resulting in enhanced continuity and consistency. Consequently, a Stable Diffusion model fine-tuned with RAIN in just a few epochs can produce video streams in real-time and low latency without much compromise in quality or consistency, up to infinite long. Despite its advanced capabilities, the RAIN only introduces a few additional 1D attention blocks, imposing minimal additional burden. Experiments in benchmark datasets and generating super-long videos demonstrating that RAIN can animate characters in real-time with much better quality, accuracy, and consistency than competitors while costing less latency. All code and models will be made publicly available.

Recent innovations in generative large language models (LLMs) have made their applications and use-cases ubiquitous. This has led to large-scale deployments of these models, using complex, expensive, and power-hungry AI accelerators, most commonly GPUs. These developments make LLM inference efficiency an important challenge. Based on our extensive characterization, we find that there are two main phases during an LLM inference request: a compute-intensive prompt computation, and a memory-intensive token generation, each with distinct latency, throughput, memory, and power characteristics. Despite state-of-the-art batching and scheduling, the token generation phase underutilizes compute resources. Specifically, unlike compute-intensive prompt computation phases, token generation phases do not require the compute capability of the latest GPUs, and can be run with lower power and cost. With Splitwise, we propose splitting the two phases of a LLM inference request on to separate machines. This allows us to use hardware that is well-suited for each phase, and provision resources independently per phase. However, splitting an inference request across machines requires state transfer from the machine running prompt computation over to the machine generating tokens. We implement and optimize this state transfer using the fast back-plane interconnects available in today's GPU clusters. We use the Splitwise technique to design LLM inference clusters using the same or different types of machines for the prompt computation and token generation phases. Our clusters are optimized for three key objectives: throughput, cost, and power. In particular, we show that we can achieve 1.4x higher throughput at 20% lower cost than current designs. Alternatively, we can achieve 2.35x more throughput with the same cost and power budgets.

We present StreamBridge, a simple yet effective framework that seamlessly transforms offline Video-LLMs into streaming-capable models. It addresses two fundamental challenges in adapting existing models into online scenarios: (1) limited capability for multi-turn real-time understanding, and (2) lack of proactive response mechanisms. Specifically, StreamBridge incorporates (1) a memory buffer combined with a round-decayed compression strategy, supporting long-context multi-turn interactions, and (2) a decoupled, lightweight activation model that can be effortlessly integrated into existing Video-LLMs, enabling continuous proactive responses. To further support StreamBridge, we construct Stream-IT, a large-scale dataset tailored for streaming video understanding, featuring interleaved video-text sequences and diverse instruction formats. Extensive experiments show that StreamBridge significantly improves the streaming understanding capabilities of offline Video-LLMs across various tasks, outperforming even proprietary models such as GPT-4o and Gemini 1.5 Pro. Simultaneously, it achieves competitive or superior performance on standard video understanding benchmarks.

We consider multi-draft speculative sampling, where the proposal sequences are sampled independently from different draft models. At each step, a token-level draft selection scheme takes a list of valid tokens as input and produces an output token whose distribution matches that of the target model. Previous works have demonstrated that the optimal scheme (which maximizes the probability of accepting one of the input tokens) can be cast as a solution to a linear program. In this work we show that the optimal scheme can be decomposed into a two-step solution: in the first step an importance sampling (IS) type scheme is used to select one intermediate token; in the second step (single-draft) speculative sampling is applied to generate the output token. For the case of two identical draft models we further 1) establish a necessary and sufficient condition on the distributions of the target and draft models for the acceptance probability to equal one and 2) provide an explicit expression for the optimal acceptance probability. Our theoretical analysis also motives a new class of token-level selection scheme based on weighted importance sampling. Our experimental results demonstrate consistent improvements in the achievable block efficiency and token rates over baseline schemes in a number of scenarios.

Zero-shot streaming text-to-speech is an important research topic in human-computer interaction. Existing methods primarily use a lookahead mechanism, relying on future text to achieve natural streaming speech synthesis, which introduces high processing latency. To address this issue, we propose SMLLE, a streaming framework for generating high-quality speech frame-by-frame. SMLLE employs a Transducer to convert text into semantic tokens in real time while simultaneously obtaining duration alignment information. The combined outputs are then fed into a fully autoregressive (AR) streaming model to reconstruct mel-spectrograms. To further stabilize the generation process, we design a Delete < Bos > Mechanism that allows the AR model to access future text introducing as minimal delay as possible. Experimental results suggest that the SMLLE outperforms current streaming TTS methods and achieves comparable performance over sentence-level TTS systems. Samples are available on https://anonymous.4open.science/w/demo_page-48B7/.

Large Language Models (LLMs) are primarily designed for batch processing. Existing methods for adapting LLMs to streaming rely either on expensive re-encoding or specialized architectures with limited scalability. This work identifies three key mismatches in adapting batch-oriented LLMs to streaming: (1) input-attention, (2) output-attention, and (3) position-ID mismatches. While it is commonly assumed that the latter two mismatches require frequent re-encoding, our analysis reveals that only the input-attention mismatch significantly impacts performance, indicating re-encoding outputs is largely unnecessary. To better understand this discrepancy with the common assumption, we provide the first comprehensive analysis of the impact of position encoding on LLMs in streaming, showing that preserving relative positions within source and target contexts is more critical than maintaining absolute order. Motivated by the above analysis, we introduce a group position encoding paradigm built on batch architectures to enhance consistency between streaming and batch modes. Extensive experiments on cross-lingual and cross-modal tasks demonstrate that our method outperforms existing approaches. Our method requires no architectural modifications, exhibits strong generalization in both streaming and batch modes. The code is available at repository https://github.com/EIT-NLP/StreamingLLM.

Large language models (LLMs) power a new generation of interactive AI applications exemplified by ChatGPT. The interactive nature of these applications demand low job completion time (JCT) for model inference. Existing LLM serving systems use run-to-completion processing for inference jobs, which suffers from head-of-line blocking and long JCT. We present FastServe, a distributed inference serving system for LLMs. FastServe exploits the autoregressive pattern of LLM inference to enable preemption at the granularity of each output token. FastServe uses preemptive scheduling to minimize JCT with a novel skip-join Multi-Level Feedback Queue scheduler. Based on the new semi information-agnostic setting of LLM inference, the scheduler leverages the input length information to assign an appropriate initial queue for each arrival job to join. The higher priority queues than the joined queue are skipped to reduce demotions. We design an efficient GPU memory management mechanism that proactively offloads and uploads intermediate states between GPU memory and host memory for LLM inference. We build a system prototype of FastServe based on NVIDIA FasterTransformer. Experimental results show that compared to the state-of-the-art solution Orca, FastServe improves the average and tail JCT by up to 5.1times and 6.4times, respectively.

The autoregressive nature of large language models (LLMs) limits inference speed. Each forward pass generates only a single token and is often bottlenecked by memory bandwidth. Speculative decoding alleviates this issue using a draft-then-verify approach to accelerate token generation. However, the overhead introduced during the draft phase and the training cost of the draft model limit the efficiency and adaptability of speculative decoding. In this work, we introduce PARallel Draft (PARD), a novel speculative decoding method that enables low-cost adaptation of autoregressive draft models into parallel draft models. PARD enhances inference efficiency by predicting multiple future tokens in a single forward pass of the draft phase, and incorporates a conditional drop token method to accelerate training. Its target-independence property allows a single draft model to be applied to an entire family of different models, minimizing the adaptation cost. Our proposed conditional drop token method can improves draft model training efficiency by 3x. On our optimized inference framework, PARD accelerates LLaMA3.1-8B inference by 4.08x, achieving 311.5 tokens per second.

We introduce Guard Vector, a safety task vector computed as the parameter difference between a guardrail model (Guard Model) and a same-architecture pretrained language model. Composing this vector with a target language model yields a Target Guard Model (TGM). We then adapt TGM with a streaming-aware approach that combines prefix-based training and evaluation with a classifier that produces a single-token output. With this composition alone, TGM improves classification quality over established Guard Models across standard safety suites and enables language extensibility to Chinese, Japanese, and Korean, requiring neither additional training nor target language labels. It also demonstrates model portability across two widely used public guardrail backbones, Llama and Gemma. With prefix SFT (supervised fine-tuning), TGM preserves classification quality under streaming by aligning the behavior between prefix inputs and full-text inputs. The single-token output design increases throughput and reduces latency. Together, these components reduce data and compute requirements while promoting streaming-aware evaluation practices, thereby contributing to a more responsible AI ecosystem.

Deploying long-context large language models (LLMs) is essential but poses significant computational and memory challenges. Caching all Key and Value (KV) states across all attention heads consumes substantial memory. Existing KV cache pruning methods either damage the long-context capabilities of LLMs or offer only limited efficiency improvements. In this paper, we identify that only a fraction of attention heads, a.k.a, Retrieval Heads, are critical for processing long contexts and require full attention across all tokens. In contrast, all other heads, which primarily focus on recent tokens and attention sinks--referred to as Streaming Heads--do not require full attention. Based on this insight, we introduce DuoAttention, a framework that only applies a full KV cache to retrieval heads while using a light-weight, constant-length KV cache for streaming heads, which reduces both LLM's decoding and pre-filling memory and latency without compromising its long-context abilities. DuoAttention uses a lightweight, optimization-based algorithm with synthetic data to identify retrieval heads accurately. Our method significantly reduces long-context inference memory by up to 2.55x for MHA and 1.67x for GQA models while speeding up decoding by up to 2.18x and 1.50x and accelerating pre-filling by up to 1.73x and 1.63x for MHA and GQA models, respectively, with minimal accuracy loss compared to full attention. Notably, combined with quantization, DuoAttention enables Llama-3-8B decoding with 3.3 million context length on a single A100 GPU. Code is provided in https://github.com/mit-han-lab/duo-attention.

This paper presents a streaming extension for the Sortformer speaker diarization framework, whose key property is the arrival-time ordering of output speakers. The proposed approach employs an Arrival-Order Speaker Cache (AOSC) to store frame-level acoustic embeddings of previously observed speakers. Unlike conventional speaker-tracing buffers, AOSC orders embeddings by speaker index corresponding to their arrival time order, and is dynamically updated by selecting frames with the highest scores based on the model's past predictions. Notably, the number of stored embeddings per speaker is determined dynamically by the update mechanism, ensuring efficient cache utilization and precise speaker tracking. Experiments on benchmark datasets confirm the effectiveness and flexibility of our approach, even in low-latency setups. These results establish Streaming Sortformer as a robust solution for real-time multi-speaker tracking and a foundation for streaming multi-talker speech processing.

With the emergence of neural audio codecs, which encode multiple streams of discrete tokens from audio, large language models have recently gained attention as a promising approach for zero-shot Text-to-Speech (TTS) synthesis. Despite the ongoing rush towards scaling paradigms, audio tokenization ironically amplifies the scalability challenge, stemming from its long sequence length and the complexity of modelling the multiple sequences. To mitigate these issues, we present CLaM-TTS that employs a probabilistic residual vector quantization to (1) achieve superior compression in the token length, and (2) allow a language model to generate multiple tokens at once, thereby eliminating the need for cascaded modeling to handle the number of token streams. Our experimental results demonstrate that CLaM-TTS is better than or comparable to state-of-the-art neural codec-based TTS models regarding naturalness, intelligibility, speaker similarity, and inference speed. In addition, we examine the impact of the pretraining extent of the language models and their text tokenization strategies on performances.

Deep learning time-series processing often relies on convolutional neural networks with overlapping windows. This overlap allows the network to produce an output faster than the window length. However, it introduces additional computations. This work explores the potential to optimize computational efficiency during inference by exploiting convolution's shift-invariance properties to skip the calculation of layer activations between successive overlapping windows. Although convolutions are shift-invariant, zero-padding and pooling operations, widely used in such networks, are not efficient and complicate efficient streaming inference. We introduce StreamiNNC, a strategy to deploy Convolutional Neural Networks for online streaming inference. We explore the adverse effects of zero padding and pooling on the accuracy of streaming inference, deriving theoretical error upper bounds for pooling during streaming. We address these limitations by proposing signal padding and pooling alignment and provide guidelines for designing and deploying models for StreamiNNC. We validate our method in simulated data and on three real-world biomedical signal processing applications. StreamiNNC achieves a low deviation between streaming output and normal inference for all three networks (2.03 - 3.55% NRMSE). This work demonstrates that it is possible to linearly speed up the inference of streaming CNNs processing overlapping windows, negating the additional computation typically incurred by overlapping windows.

We present VoXtream, a fully autoregressive, zero-shot streaming text-to-speech (TTS) system for real-time use that begins speaking from the first word. VoXtream directly maps incoming phonemes to audio tokens using a monotonic alignment scheme and a dynamic look-ahead that does not delay onset. Built around an incremental phoneme transformer, a temporal transformer predicting semantic and duration tokens, and a depth transformer producing acoustic tokens, VoXtream achieves, to our knowledge, the lowest initial delay among publicly available streaming TTS: 102 ms on GPU. Despite being trained on a mid-scale 9k-hour corpus, it matches or surpasses larger baselines on several metrics, while delivering competitive quality in both output- and full-streaming settings. Demo and code are available at https://herimor.github.io/voxtream.

In this work, we aim to simultaneously enhance the effectiveness and efficiency of Mixture-of-Experts (MoE) methods. To achieve this, we propose MoE++, a general and heterogeneous MoE framework that integrates both Feed-Forward Network~(FFN) and zero-computation experts. Specifically, we introduce three types of zero-computation experts: the zero expert, copy expert, and constant expert, which correspond to discard, skip, and replace operations, respectively. This design offers three key advantages: (i) Low Computing Overhead: Unlike the uniform mixing mechanism for all tokens within vanilla MoE, MoE++ allows each token to engage with a dynamic number of FFNs, be adjusted by constant vectors, or even skip the MoE layer entirely. (ii) High Performance: By enabling simple tokens to utilize fewer FFN experts, MoE++ allows more experts to focus on challenging tokens, thereby unlocking greater performance potential than vanilla MoE. (iii) Deployment Friendly: Given that zero-computation experts have negligible parameters, we can deploy all zero-computation experts on each GPU, eliminating the significant communication overhead and expert load imbalance associated with FFN experts distributed across different GPUs. Moreover, we leverage gating residuals, enabling each token to consider the pathway taken in the previous layer when selecting the appropriate experts. Extensive experimental results demonstrate that MoE++ achieves better performance while delivering 1.1-2.1x expert forward throughput compared to a vanilla MoE model of the same size, which lays a solid foundation for developing advanced and efficient MoE-related models.

As large language models (LLMs) become increasingly powerful, the sequential nature of autoregressive generation creates a fundamental throughput bottleneck that limits the practical deployment. While Multi-Token Prediction (MTP) has demonstrated remarkable benefits for model training efficiency and performance, its inherent potential for inference acceleration remains largely unexplored. This paper introduces FastMTP, a simple yet effective method that improves multi-step draft quality by aligning MTP training with its inference pattern, significantly enhancing speculative decoding performance. Our approach fine-tunes a single MTP head with position-shared weights on self-distilled data, enabling it to capture dependencies among consecutive future tokens and maintain high acceptance rates across multiple recursive draft steps. By integrating language-aware dynamic vocabulary compression into the MTP head, we further reduce computational overhead in the drafting process. Experimental results across seven diverse benchmarks demonstrate that FastMTP achieves an average of 2.03x speedup compared to standard next token prediction with lossless output quality, outperforming vanilla MTP by 82%. FastMTP requires only lightweight training and seamlessly integrates with existing inference frameworks, offering a practical and rapidly deployable solution for accelerating LLM inference.

We present TokenFlow, a novel unified image tokenizer that bridges the long-standing gap between multimodal understanding and generation. Prior research attempt to employ a single reconstruction-targeted Vector Quantization (VQ) encoder for unifying these two tasks. We observe that understanding and generation require fundamentally different granularities of visual information. This leads to a critical trade-off, particularly compromising performance in multimodal understanding tasks. TokenFlow addresses this challenge through an innovative dual-codebook architecture that decouples semantic and pixel-level feature learning while maintaining their alignment via a shared mapping mechanism. This design enables direct access to both high-level semantic representations crucial for understanding tasks and fine-grained visual features essential for generation through shared indices. Our extensive experiments demonstrate TokenFlow's superiority across multiple dimensions. Leveraging TokenFlow, we demonstrate for the first time that discrete visual input can surpass LLaVA-1.5 13B in understanding performance, achieving a 7.2\% average improvement. For image reconstruction, we achieve a strong FID score of 0.63 at 384*384 resolution. Moreover, TokenFlow establishes state-of-the-art performance in autoregressive image generation with a GenEval score of 0.55 at 256*256 resolution, achieving comparable results to SDXL.

The Transformer self-attention network has shown promising performance as an alternative to recurrent neural networks in end-to-end (E2E) automatic speech recognition (ASR) systems. However, Transformer has a drawback in that the entire input sequence is required to compute both self-attention and source--target attention. In this paper, we propose a novel blockwise synchronous beam search algorithm based on blockwise processing of encoder to perform streaming E2E Transformer ASR. In the beam search, encoded feature blocks are synchronously aligned using a block boundary detection technique, where a reliability score of each predicted hypothesis is evaluated based on the end-of-sequence and repeated tokens in the hypothesis. Evaluations of the HKUST and AISHELL-1 Mandarin, LibriSpeech English, and CSJ Japanese tasks show that the proposed streaming Transformer algorithm outperforms conventional online approaches, including monotonic chunkwise attention (MoChA), especially when using the knowledge distillation technique. An ablation study indicates that our streaming approach contributes to reducing the response time, and the repetition criterion contributes significantly in certain tasks. Our streaming ASR models achieve comparable or superior performance to batch models and other streaming-based Transformer methods in all tasks considered.

In real-world applications, users often require both translations and transcriptions of speech to enhance their comprehension, particularly in streaming scenarios where incremental generation is necessary. This paper introduces a streaming Transformer-Transducer that jointly generates automatic speech recognition (ASR) and speech translation (ST) outputs using a single decoder. To produce ASR and ST content effectively with minimal latency, we propose a joint token-level serialized output training method that interleaves source and target words by leveraging an off-the-shelf textual aligner. Experiments in monolingual (it-en) and multilingual (\{de,es,it\}-en) settings demonstrate that our approach achieves the best quality-latency balance. With an average ASR latency of 1s and ST latency of 1.3s, our model shows no degradation or even improves output quality compared to separate ASR and ST models, yielding an average improvement of 1.1 WER and 0.4 BLEU in the multilingual case.

Prevalent semantic speech tokenizers, designed to capture linguistic content, are surprisingly fragile. We find they are not robust to meaning-irrelevant acoustic perturbations; even at high Signal-to-Noise Ratios (SNRs) where speech is perfectly intelligible, their output token sequences can change drastically, increasing the learning burden for downstream LLMs. This instability stems from two flaws: a brittle single-path quantization architecture and a distant training signal indifferent to intermediate token stability. To address this, we introduce StableToken, a tokenizer that achieves stability through a consensus-driven mechanism. Its multi-branch architecture processes audio in parallel, and these representations are merged via a powerful bit-wise voting mechanism to form a single, stable token sequence. StableToken sets a new state-of-the-art in token stability, drastically reducing Unit Edit Distance (UED) under diverse noise conditions. This foundational stability translates directly to downstream benefits, significantly improving the robustness of SpeechLLMs on a variety of tasks.

Models for low-latency, streaming applications could benefit from the knowledge capacity of larger models, but edge devices cannot run these models due to resource constraints. A possible solution is to transfer hints during inference from a large model running remotely to a small model running on-device. However, this incurs a communication delay that breaks real-time requirements and does not guarantee that both models will operate on the same data at the same time. We propose knowledge boosting, a novel technique that allows a large model to operate on time-delayed input during inference, while still boosting small model performance. Using a streaming neural network that processes 8 ms chunks, we evaluate different speech separation and enhancement tasks with communication delays of up to six chunks or 48 ms. Our results show larger gains where the performance gap between the small and large models is wide, demonstrating a promising method for large-small model collaboration for low-latency applications. Code, dataset, and audio samples available at https://knowledgeboosting.cs.washington.edu/.

Current dialogue generation approaches typically require the complete dialogue text before synthesis and produce a single, inseparable speech containing all voices, making them unsuitable for interactive chat; moreover, they suffer from unstable synthesis, inaccurate speaker transitions, and incoherent prosody. In this work, we present FireRedTTS-2, a long-form streaming TTS system for multi-speaker dialogue generation, delivering stable, natural speech with reliable speaker switching and context-aware prosody. A new 12.5Hz streaming speech tokenizer accelerates training and inference, extends maximum dialogue length, encodes richer semantics to stabilize text-to-token modeling and supports high-fidelity streaming generation for real-time applications. We adopt a text-speech interleaved format, concatenating speaker-labeled text with aligned speech tokens in chronological order, and model it with a dual-transformer: a large decoder-only transformer predicts tokens at the first layer, and a smaller one completes subsequent layers. Experimental results show that FireRedTTS-2 integrates seamlessly with chat frameworks and, with minimal fine-tuning, produces emotionally expressive speech guided by implicit contextual cues. In podcast generation, it surpasses existing systems including MoonCast, Zipvoice-Dialogue, and MOSS-TTSD in objective intelligibility, speaker-turn reliability, and perceived naturalness with context-consistent prosody. Our demos are available at https://fireredteam.github.io/demos/firered_tts_2.

The rapid advancement of multi-modal language models (MLLMs) like GPT-4o has propelled the development of Omni language models, designed to process and proactively respond to continuous streams of multi-modal data. Despite their potential, evaluating their real-world interactive capabilities in streaming video contexts remains a formidable challenge. In this work, we introduce OmniMMI, a comprehensive multi-modal interaction benchmark tailored for OmniLLMs in streaming video contexts. OmniMMI encompasses over 1,121 videos and 2,290 questions, addressing two critical yet underexplored challenges in existing video benchmarks: streaming video understanding and proactive reasoning, across six distinct subtasks. Moreover, we propose a novel framework, Multi-modal Multiplexing Modeling (M4), designed to enable an inference-efficient streaming model that can see, listen while generating.

Recent advances in deep learning have mainly relied on Transformers due to their data dependency and ability to learn at scale. The attention module in these architectures, however, exhibits quadratic time and space in input size, limiting their scalability for long-sequence modeling. Despite recent attempts to design efficient and effective architecture backbone for multi-dimensional data, such as images and multivariate time series, existing models are either data independent, or fail to allow inter- and intra-dimension communication. Recently, State Space Models (SSMs), and more specifically Selective State Space Models, with efficient hardware-aware implementation, have shown promising potential for long sequence modeling. Motivated by the success of SSMs, we present MambaMixer, a new architecture with data-dependent weights that uses a dual selection mechanism across tokens and channels, called Selective Token and Channel Mixer. MambaMixer connects selective mixers using a weighted averaging mechanism, allowing layers to have direct access to early features. As a proof of concept, we design Vision MambaMixer (ViM2) and Time Series MambaMixer (TSM2) architectures based on the MambaMixer block and explore their performance in various vision and time series forecasting tasks. Our results underline the importance of selective mixing across both tokens and channels. In ImageNet classification, object detection, and semantic segmentation tasks, ViM2 achieves competitive performance with well-established vision models and outperforms SSM-based vision models. In time series forecasting, TSM2 achieves outstanding performance compared to state-of-the-art methods while demonstrating significantly improved computational cost. These results show that while Transformers, cross-channel attention, and MLPs are sufficient for good performance in time series forecasting, neither is necessary.

Modern video generation frameworks based on Latent Diffusion Models suffer from inefficiencies in tokenization due to the Frame-Proportional Information Assumption. Existing tokenizers provide fixed temporal compression rates, causing the computational cost of the diffusion model to scale linearly with the frame rate. The paper proposes the Duration-Proportional Information Assumption: the upper bound on the information capacity of a video is proportional to the duration rather than the number of frames. Based on this insight, the paper introduces VFRTok, a Transformer-based video tokenizer, that enables variable frame rate encoding and decoding through asymmetric frame rate training between the encoder and decoder. Furthermore, the paper proposes Partial Rotary Position Embeddings (RoPE) to decouple position and content modeling, which groups correlated patches into unified tokens. The Partial RoPE effectively improves content-awareness, enhancing the video generation capability. Benefiting from the compact and continuous spatio-temporal representation, VFRTok achieves competitive reconstruction quality and state-of-the-art generation fidelity while using only 1/8 tokens compared to existing tokenizers.

We present Seed Diffusion Preview, a large-scale language model based on discrete-state diffusion, offering remarkably fast inference speed. Thanks to non-sequential, parallel generation, discrete diffusion models provide a notable speedup to mitigate the inherent latency of token-by-token decoding, as demonstrated recently (e.g., Mercury Coder, Gemini Diffusion). Seed Diffusion Preview achieves an inference speed of 2,146 token/s over H20 GPUs while maintaining competitive performance across a sweep of standard code evaluation benchmarks, significantly faster than contemporary Mercury and Gemini Diffusion, establishing new state of the art on the speed-quality Pareto frontier for code models.

Federated Learning (FL) enables edge devices or clients to collaboratively train machine learning (ML) models without sharing their private data. Much of the existing work in FL focuses on efficiently learning a model for a single task. In this paper, we study simultaneous training of multiple FL models using a common set of clients. The few existing simultaneous training methods employ synchronous aggregation of client updates, which can cause significant delays because large models and/or slow clients can bottleneck the aggregation. On the other hand, a naive asynchronous aggregation is adversely affected by stale client updates. We propose FedAST, a buffered asynchronous federated simultaneous training algorithm that overcomes bottlenecks from slow models and adaptively allocates client resources across heterogeneous tasks. We provide theoretical convergence guarantees for FedAST for smooth non-convex objective functions. Extensive experiments over multiple real-world datasets demonstrate that our proposed method outperforms existing simultaneous FL approaches, achieving up to 46.0% reduction in time to train multiple tasks to completion.

MoE (Mixture of Experts) prevails as a neural architecture that can scale modern transformer-based LLMs (Large Language Models) to unprecedented scales. Nevertheless, large MoEs' great demands of computing power, memory capacity and memory bandwidth make scalable serving a fundamental challenge and efficient parallel inference has become a requisite to attain adequate throughput under latency constraints. DeepSpeed-MoE, one state-of-the-art MoE inference framework, adopts a 3D-parallel paradigm including EP (Expert Parallelism), TP (Tensor Parallel) and DP (Data Parallelism). However, our analysis shows DeepSpeed-MoE's inference efficiency is largely bottlenecked by EP, which is implemented with costly all-to-all collectives to route token activation. Our work aims to boost DeepSpeed-MoE by strategically reducing EP's communication overhead with a technique named Speculative MoE. Speculative MoE has two speculative parallelization schemes, speculative token shuffling and speculative expert grouping, which predict outstanding tokens' expert routing paths and pre-schedule tokens and experts across devices to losslessly trim EP's communication volume. Besides DeepSpeed-MoE, we also build Speculative MoE into a prevailing MoE inference engine SGLang. Experiments show Speculative MoE can significantly boost state-of-the-art MoE inference frameworks on fast homogeneous and slow heterogeneous interconnects.

We introduce Delayed Streams Modeling (DSM), a flexible formulation for streaming, multimodal sequence-to-sequence learning. Sequence-to-sequence generation is often cast in an offline manner, where the model consumes the complete input sequence before generating the first output timestep. Alternatively, streaming sequence-to-sequence rely on learning a policy for choosing when to advance on the input stream, or write to the output stream. DSM instead models already time-aligned streams with a decoder-only language model. By moving the alignment to a pre-processing step,and introducing appropriate delays between streams, DSM provides streaming inference of arbitrary output sequences, from any input combination, making it applicable to many sequence-to-sequence problems. In particular, given text and audio streams, automatic speech recognition (ASR) corresponds to the text stream being delayed, while the opposite gives a text-to-speech (TTS) model. We perform extensive experiments for these two major sequence-to-sequence tasks, showing that DSM provides state-of-the-art performance and latency while supporting arbitrary long sequences, being even competitive with offline baselines. Code, samples and demos are available at https://github.com/kyutai-labs/delayed-streams-modeling

Transformer-based Mixture-of-Experts (MoE) models have been driving several recent technological advancements in Natural Language Processing (NLP). These MoE models adopt a router mechanism to determine which experts to activate for routing input tokens. However, existing router mechanisms allocate a fixed number of experts to each token, which neglects the varying importance of different input tokens. In this study, we propose a novel dynamic router mechanism that Dynamically Allocates a variable number of experts for Mixture-of-Experts (DA-MoE) models based on an effective token importance measure. First, we show that the Transformer attention mechanism provides a natural and effective way of calculating token importance. Second, we propose a dynamic router mechanism that effectively decides the optimal number of experts (K) and allocates the top-K experts for each input token. Third, comprehensive experiments on several benchmark datasets demonstrate that our DA-MoE approach consistently outperforms the state-of-the-art Transformer based MoE model on the popular GLUE benchmark.

Speculative decoding (SD) has emerged as a promising technique to accelerate LLM inference by employing a small draft model to propose draft tokens in advance, and validating them in parallel with the large target model. However, the existing SD methods still remain constrained by their serialized execution, which causes the mutual waiting bubbles between the draft and target models. To address this challenge, we draw inspiration from branch prediction in modern processors and propose a novel framework SpecBranch to unlock branch parallelism in SD. Specifically, we first take an in-depth analysis of the potential of branch parallelism in SD, and recognize that the key challenge lies in the trade-offs between parallelization and token rollback. Based on the analysis, we introduce parallel speculative branches to preemptively hedge against likely rejections. Meanwhile, to enhance parallelism, we jointly orchestrate adaptive draft lengths with a hybrid combination of the implicit draft model confidence and explicit reusing of target model features. Extensive experiments across various models and benchmarks show that SpecBranch achieves over 1.8times sim 4.5times speedups against the auto-regressive decoding and reduces rollback tokens by 50\% for poorly aligned models, while maintaining an identical sampling distribution.

Non-autoregressive (NAR) modeling has gained significant interest in speech processing since these models achieve dramatically lower inference time than autoregressive (AR) models while also achieving good transcription accuracy. Since NAR automatic speech recognition (ASR) models must wait for the completion of the entire utterance before processing, some works explore streaming NAR models based on blockwise attention for low-latency applications. However, streaming NAR models significantly lag in accuracy compared to streaming AR and non-streaming NAR models. To address this, we propose a streaming "semi-autoregressive" ASR model that incorporates the labels emitted in previous blocks as additional context using a Language Model (LM) subnetwork. We also introduce a novel greedy decoding algorithm that addresses insertion and deletion errors near block boundaries while not significantly increasing the inference time. Experiments show that our method outperforms the existing streaming NAR model by 19% relative on Tedlium2, 16%/8% on Librispeech-100 clean/other test sets, and 19%/8% on the Switchboard(SWB) / Callhome(CH) test sets. It also reduced the accuracy gap with streaming AR and non-streaming NAR models while achieving 2.5x lower latency. We also demonstrate that our approach can effectively utilize external text data to pre-train the LM subnetwork to further improve streaming ASR accuracy.

We introduce ProVideLLM, an end-to-end framework for real-time procedural video understanding. ProVideLLM integrates a multimodal cache configured to store two types of tokens - verbalized text tokens, which provide compressed textual summaries of long-term observations, and visual tokens, encoded with DETR-QFormer to capture fine-grained details from short-term observations. This design reduces token count by 22x over existing methods in representing one hour of long-term observations while effectively encoding fine-granularity of the present. By interleaving these tokens in our multimodal cache, ProVideLLM ensures sub-linear scaling of memory and compute with video length, enabling per-frame streaming inference at 10 FPS and streaming dialogue at 25 FPS, with a minimal 2GB GPU memory footprint. ProVideLLM also sets new state-of-the-art results on six procedural tasks across four datasets.

Speculative decoding (SPD) aims to accelerate the auto-regressive token generation process of a target Large Language Model (LLM). Some approaches employ a draft model with multiple heads to predict a sequence of future tokens, where each head handles a token in the sequence. The target LLM verifies the predicted sequence and accepts aligned tokens, enabling efficient multi-token generation. However, existing methods assume that all tokens within a sequence are equally important, employing identical head structures and relying on a single-generation paradigm, either serial or parallel. To this end, we theoretically demonstrate that initial tokens in the draft sequence are more important than later ones. Building on this insight, we propose Gumiho, a hybrid model combining serial and parallel heads. Specifically, given the critical importance of early tokens, we employ a sophisticated Transformer architecture for the early draft heads in a serial configuration to improve accuracy. For later tokens, we utilize multiple lightweight MLP heads operating in parallel to enhance efficiency. By allocating more advanced model structures and longer running times to the early heads, Gumiho achieves improved overall performance. The experimental results demonstrate that our method outperforms existing approaches, fully validating its effectiveness.

Multimodal large language models (MLLMs) have made significant progress in visual-language reasoning, but their ability to efficiently handle long videos remains limited. Despite recent advances in long-context MLLMs, storing and attending to the key-value (KV) cache for long visual contexts incurs substantial memory and computational overhead. Existing visual compression methods require either encoding the entire visual context before compression or having access to the questions in advance, which is impractical for long video understanding and multi-turn conversational settings. In this work, we propose StreamMem, a query-agnostic KV cache memory mechanism for streaming video understanding. Specifically, StreamMem encodes new video frames in a streaming manner, compressing the KV cache using attention scores between visual tokens and generic query tokens, while maintaining a fixed-size KV memory to enable efficient question answering (QA) in memory-constrained, long-video scenarios. Evaluation on three long video understanding and two streaming video question answering benchmarks shows that StreamMem achieves state-of-the-art performance in query-agnostic KV cache compression and is competitive with query-aware compression approaches.

Accelerating inference in Large Language Models (LLMs) is critical for real-time interactions, as they have been widely incorporated into real-world services. Speculative decoding, a fully algorithmic solution, has gained attention for improving inference speed by drafting and verifying tokens, thereby generating multiple tokens in a single forward pass. However, current drafting strategies usually require significant fine-tuning or have inconsistent performance across tasks. To address these challenges, we propose Hierarchy Drafting (HD), a novel lossless drafting approach that organizes various token sources into multiple databases in a hierarchical framework based on temporal locality. In the drafting step, HD sequentially accesses multiple databases to obtain draft tokens from the highest to the lowest locality, ensuring consistent acceleration across diverse tasks and minimizing drafting latency. Our experiments on Spec-Bench using LLMs with 7B and 13B parameters demonstrate that HD outperforms existing database drafting methods, achieving robust inference speedups across model sizes, tasks, and temperatures.

Autoregressive large language models (LLMs) have made remarkable progress in various natural language generation tasks. However, they incur high computation cost and latency resulting from the autoregressive token-by-token generation. To address this issue, several approaches have been proposed to reduce computational cost using early-exit strategies. These strategies enable faster text generation using reduced computation without applying the full computation graph to each token. While existing token-level early exit methods show promising results for online inference, they cannot be readily applied for batch inferencing and Key-Value caching. This is because they have to wait until the last token in a batch exits before they can stop computing. This severely limits the practical application of such techniques. In this paper, we propose a simple and effective token-level early exit method, SkipDecode, designed to work seamlessly with batch inferencing and KV caching. It overcomes prior constraints by setting up a singular exit point for every token in a batch at each sequence position. It also guarantees a monotonic decrease in exit points, thereby eliminating the need to recompute KV Caches for preceding tokens. Rather than terminating computation prematurely as in prior works, our approach bypasses lower to middle layers, devoting most of the computational resources to upper layers, allowing later tokens to benefit from the compute expenditure by earlier tokens. Our experimental results show that SkipDecode can obtain 2x to 5x inference speedups with negligible regression across a variety of tasks. This is achieved using OPT models of 1.3 billion and 6.7 billion parameters, all the while being directly compatible with batching and KV caching optimization techniques.

Multi-token prediction has emerged as a promising objective for improving language model pretraining, but its benefits have not consistently generalized to other settings such as fine-tuning. In this paper, we propose MuToR, a simple and effective approach to multi-token prediction that interleaves learnable register tokens into the input sequence, each tasked with predicting future targets. Compared to existing methods, MuToR offers several key advantages: it introduces only a negligible number of additional parameters, requires no architectural changes--ensuring compatibility with off-the-shelf pretrained language models--and remains aligned with the next-token pretraining objective, making it especially well-suited for supervised fine-tuning. Moreover, it naturally supports scalable prediction horizons. We demonstrate the effectiveness and versatility of MuToR across a range of use cases, including supervised fine-tuning, parameter-efficient fine-tuning (PEFT), and pretraining, on challenging generative tasks in both language and vision domains. Our code will be available at: https://github.com/nasosger/MuToR.

The generation speed of LLMs are bottlenecked by autoregressive decoding, where tokens are predicted sequentially one by one. Alternatively, diffusion large language models (dLLMs) theoretically allow for parallel token generation, but in practice struggle to achieve the speed of autoregressive models without significantly sacrificing quality. We therefore introduce adaptive parallel decoding (APD), a novel method that dynamically adjusts the number of tokens sampled in parallel. We achieve this by defining a multiplicative mixture between the dLLM marginal probabilities and the joint probability of sequences under a small auxiliary autoregressive model. This inverts the standard setup of speculative decoding, where the goal is to sample from a large autoregressive verifier by drafting from a smaller model. We further optimize APD by enabling KV caching and limiting the size of the masked input. Altogether, our method puts forward three tunable parameters to flexibly tradeoff throughput and quality. We show that APD provides markedly higher throughput with minimal quality degradations on downstream benchmarks.

Training large language models (LLMs) with increasingly long and varying sequence lengths introduces severe load imbalance challenges in large-scale data-parallel training. Recent frameworks attempt to mitigate these issues through data reorganization or hybrid parallel strategies. However, they often overlook how computational and communication costs scale with sequence length, resulting in suboptimal performance. We identify three critical challenges: (1) varying computation-to-communication ratios across sequences of different lengths in distributed attention, (2) mismatch between static NIC-GPU affinity and dynamic parallel workloads, and (3) distinct optimal partitioning strategies required for quadratic attention versus linear components. To address these challenges, we present Zeppelin, a novel training system that integrates three key techniques: (1) a hierarchical sequence partitioning method for the attention module that reduces communication overhead and balances computation, supported by an efficient attention engine that applies divergent parallel strategies; (2) a routing layer that orchestrates inter-node transfers to fully utilize NIC bandwidth; and (3) a remapping layer that transforms sequence layouts between attention and linear modules, ensuring high computational efficiency across both. Comprehensive evaluations across diverse configurations show that Zeppelin delivers an average 2.80x speedup over state-of-the-art methods.

Recent advances in zero-shot text-to-speech (TTS) synthesis have achieved high-quality speech generation for unseen speakers, but most systems remain unsuitable for real-time applications because of their offline design. Current streaming TTS paradigms often rely on multi-stage pipelines and discrete representations, leading to increased computational cost and suboptimal system performance. In this work, we propose StreamMel, a pioneering single-stage streaming TTS framework that models continuous mel-spectrograms. By interleaving text tokens with acoustic frames, StreamMel enables low-latency, autoregressive synthesis while preserving high speaker similarity and naturalness. Experiments on LibriSpeech demonstrate that StreamMel outperforms existing streaming TTS baselines in both quality and latency. It even achieves performance comparable to offline systems while supporting efficient real-time generation, showcasing broad prospects for integration with real-time speech large language models. Audio samples are available at: https://aka.ms/StreamMel.

Training language models on long sequence data is a demanding requirement for enhancing the model's capability on complex tasks, e.g., long-chain reasoning. However, as the sequence length scales up, the memory cost for storing activation values becomes huge during the Backpropagation (BP) process, even with the application of gradient checkpointing technique. To tackle this challenge, we propose a memory-efficient and exact BP method called StreamBP, which performs a linear decomposition of the chain rule along the sequence dimension in a layer-wise manner, significantly reducing the memory cost of activation values and logits. The proposed method is applicable to common objectives such as SFT, GRPO, and DPO. From an implementation perspective, StreamBP achieves less computational FLOPs and faster BP speed by leveraging the causal structure of the language model. Compared to gradient checkpointing, StreamBP scales up the maximum sequence length of BP by 2.8-5.5 times larger, while using comparable or even less BP time. Note that StreamBP's sequence length scaling ability can be directly transferred to batch size scaling for accelerating training. We further develop a communication-efficient distributed StreamBP to effectively support multi-GPU training and broaden its applicability. Our code can be easily integrated into the training pipeline of any transformer models and is available at https://github.com/Ledzy/StreamBP.

Automatic Speech Recognition (ASR) has seen remarkable progress, with models like OpenAI Whisper and NVIDIA Canary achieving state-of-the-art (SOTA) performance in offline transcription. However, these models are not designed for streaming (online or real-time) transcription, due to limitations in their architecture and training methodology. We propose a method to turn the transformer encoder-decoder model into a low-latency streaming model that is careless about future context. We present an analysis explaining why it is not straightforward to convert an encoder-decoder transformer to a low-latency streaming model. Our proposed method modifies the existing (non-causal) encoder to a causal encoder by fine-tuning both the encoder and decoder using Low-Rank Adaptation (LoRA) and a weakly aligned dataset. We then propose an updated inference mechanism that utilizes the fine-tune causal encoder and decoder to yield greedy and beam-search decoding, and is shown to be locally optimal. Experiments on low-latency chunk sizes (less than 300 msec) show that our fine-tuned model outperforms existing non-fine-tuned streaming approaches in most cases, while using a lower complexity. Additionally, we observe that our training process yields better alignment, enabling a simple method for extracting word-level timestamps. We release our training and inference code, along with the fine-tuned models, to support further research and development in streaming ASR.

With the development of large language models (LLMs), the ability to handle longer contexts has become a key capability for Web applications such as cross-document understanding and LLM-powered search systems. However, this progress faces two major challenges: performance degradation due to sequence lengths out-of-distribution, and excessively long inference times caused by the quadratic computational complexity of attention. These issues hinder the application of LLMs in long-context scenarios. In this paper, we propose Dynamic Token-Level KV Cache Selection (TokenSelect), a model-agnostic, training-free method for efficient and accurate long-context inference. TokenSelect builds upon the observation of non-contiguous attention sparsity, using Query-Key dot products to measure per-head KV Cache criticality at token-level. By per-head soft voting mechanism, TokenSelect selectively involves a small number of critical KV cache tokens in the attention calculation without sacrificing accuracy. To further accelerate TokenSelect, we designed the Selection Cache based on observations of consecutive Query similarity and implemented efficient dot product kernel, significantly reducing the overhead of token selection. A comprehensive evaluation of TokenSelect demonstrates up to 23.84x speedup in attention computation and up to 2.28x acceleration in end-to-end latency, while providing superior performance compared to state-of-the-art long-context inference methods.

Efficiently modeling sequences with infinite context length has been a long-standing problem. Past works suffer from either the quadratic computation complexity or the limited extrapolation ability on length generalization. In this work, we present Samba, a simple hybrid architecture that layer-wise combines Mamba, a selective State Space Model (SSM), with Sliding Window Attention (SWA). Samba selectively compresses a given sequence into recurrent hidden states while still maintaining the ability to precisely recall memories with the attention mechanism. We scale Samba up to 3.8B parameters with 3.2T training tokens and show that Samba substantially outperforms the state-of-the-art models based on pure attention or SSMs on a wide range of benchmarks. When trained on 4K length sequences, Samba can be efficiently extrapolated to 256K context length with perfect memory recall and show improved token predictions up to 1M context length. As a linear-time sequence model, Samba enjoys a 3.73x higher throughput compared to Transformers with grouped-query attention when processing user prompts of 128K length, and 3.64x speedup when generating 64K tokens with unlimited streaming. A sample implementation of Samba is publicly available in https://github.com/microsoft/Samba.

Transformer-based models excel in speech recognition. Existing efforts to optimize Transformer inference, typically for long-context applications, center on simplifying attention score calculations. However, streaming speech recognition models usually process a limited number of tokens each time, making attention score calculation less of a bottleneck. Instead, the bottleneck lies in the linear projection layers of multi-head attention and feedforward networks, constituting a substantial portion of the model size and contributing significantly to computation, memory, and power usage. To address this bottleneck, we propose folding attention, a technique targeting these linear layers, significantly reducing model size and improving memory and power efficiency. Experiments on on-device Transformer-based streaming speech recognition models show that folding attention reduces model size (and corresponding memory consumption) by up to 24% and power consumption by up to 23%, all without compromising model accuracy or computation overhead.

We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance. We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully harness its capabilities. Comprehensive evaluations reveal that DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models. Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training. In addition, its training process is remarkably stable. Throughout the entire training process, we did not experience any irrecoverable loss spikes or perform any rollbacks. The model checkpoints are available at https://github.com/deepseek-ai/DeepSeek-V3.

Autoregressive next token prediction language models offer powerful capabilities but face significant challenges in practical deployment due to the high computational and memory costs of inference, particularly during the decoding stage. We introduce Set Block Decoding (SBD), a simple and flexible paradigm that accelerates generation by integrating standard next token prediction (NTP) and masked token prediction (MATP) within a single architecture. SBD allows the model to sample multiple, not necessarily consecutive, future tokens in parallel, a key distinction from previous acceleration methods. This flexibility allows the use of advanced solvers from the discrete diffusion literature, offering significant speedups without sacrificing accuracy. SBD requires no architectural changes or extra training hyperparameters, maintains compatibility with exact KV-caching, and can be implemented by fine-tuning existing next token prediction models. By fine-tuning Llama-3.1 8B and Qwen-3 8B, we demonstrate that SBD enables a 3-5x reduction in the number of forward passes required for generation while achieving same performance as equivalent NTP training.

In this paper, we propose an efficient and accurate streaming speech recognition model based on the FastConformer architecture. We adapted the FastConformer architecture for streaming applications through: (1) constraining both the look-ahead and past contexts in the encoder, and (2) introducing an activation caching mechanism to enable the non-autoregressive encoder to operate autoregressively during inference. The proposed model is thoughtfully designed in a way to eliminate the accuracy disparity between the train and inference time which is common for many streaming models. Furthermore, our proposed encoder works with various decoder configurations including Connectionist Temporal Classification (CTC) and RNN-Transducer (RNNT) decoders. Additionally, we introduced a hybrid CTC/RNNT architecture which utilizes a shared encoder with both a CTC and RNNT decoder to boost the accuracy and save computation. We evaluate the proposed model on LibriSpeech dataset and a multi-domain large scale dataset and demonstrate that it can achieve better accuracy with lower latency and inference time compared to a conventional buffered streaming model baseline. We also showed that training a model with multiple latencies can achieve better accuracy than single latency models while it enables us to support multiple latencies with a single model. Our experiments also showed the hybrid architecture would not only speedup the convergence of the CTC decoder but also improves the accuracy of streaming models compared to single decoder models.

In this short report, we introduce joint multi-token prediction (JTP), a lightweight modification of standard next-token prediction designed to enrich hidden state representations by jointly predicting multiple future tokens. Unlike previous multi-token prediction approaches, JTP strategically employs teacher forcing of future-tokens through a carefully designed representation bottleneck, allowing the model to encode rich predictive information with minimal computational overhead during training. We show that the JTP approach achieves a short-horizon belief state representation, while popular alternatives for multi-token prediction fail to do so. We demonstrate the effectiveness of our method on the synthetic star graph navigation task from from Bachmann and Nagarajan [2024], highlighting a significant performance improvement over existing methods. This manuscript presents promising preliminary results intended to stimulate further research.

Recurrent Neural Network (RNN) inference exhibits low hardware utilization due to the strict data dependencies across time-steps. Batching multiple requests can increase throughput. However, RNN batching requires a large amount of padding since the batched input sequences may largely differ in length. Schemes that dynamically update the batch every few time-steps avoid padding. However, they require executing different RNN layers in a short timespan, decreasing energy efficiency. Hence, we propose E-BATCH, a low-latency and energy-efficient batching scheme tailored to RNN accelerators. It consists of a runtime system and effective hardware support. The runtime concatenates multiple sequences to create large batches, resulting in substantial energy savings. Furthermore, the accelerator notifies it when the evaluation of a sequence is done, so that a new sequence can be immediately added to a batch, thus largely reducing the amount of padding. E-BATCH dynamically controls the number of time-steps evaluated per batch to achieve the best trade-off between latency and energy efficiency for the given hardware platform. We evaluate E-BATCH on top of E-PUR and TPU. In E-PUR, E-BATCH improves throughput by 1.8x and energy-efficiency by 3.6x, whereas in TPU, it improves throughput by 2.1x and energy-efficiency by 1.6x, over the state-of-the-art.

Large Language Models (LLMs) have demonstrated success across many benchmarks. However, they still exhibit limitations in long-context scenarios, primarily due to their short effective context length, quadratic computational complexity, and high memory overhead when processing lengthy inputs. To mitigate these issues, we introduce a novel context compression pipeline, called Intermediate Layer Retrieval (ILRe), which determines one intermediate decoder layer offline, encodes context by streaming chunked prefill only up to that layer, and recalls tokens by the attention scores between the input query and full key cache in that specified layer. In particular, we propose a multi-pooling kernels allocating strategy in the token recalling process to maintain the completeness of semantics. Our approach not only reduces the prefilling complexity from O(L^2) to O(L), but also achieves performance comparable to or better than the full context in the long context scenarios. Without additional post training or operator development, ILRe can process a single 1M tokens request in less than half a minute (speedup approx 180times) and scores RULER-1M benchmark of approx 79.8 with model Llama-3.1-UltraLong-8B-1M-Instruct on a Huawei Ascend 910B NPU.

In this work we explore the latency and accuracy of keyword spotting (KWS) models in streaming and non-streaming modes on mobile phones. NN model conversion from non-streaming mode (model receives the whole input sequence and then returns the classification result) to streaming mode (model receives portion of the input sequence and classifies it incrementally) may require manual model rewriting. We address this by designing a Tensorflow/Keras based library which allows automatic conversion of non-streaming models to streaming ones with minimum effort. With this library we benchmark multiple KWS models in both streaming and non-streaming modes on mobile phones and demonstrate different tradeoffs between latency and accuracy. We also explore novel KWS models with multi-head attention which reduce the classification error over the state-of-art by 10% on Google speech commands data sets V2. The streaming library with all experiments is open-sourced.

Deploying ASR models at an industrial scale poses significant challenges in hardware resource management, especially for long-form transcription tasks where audio may last for hours. Large Conformer models, despite their capabilities, are limited to processing only 15 minutes of audio on an 80GB GPU. Furthermore, variable input lengths worsen inefficiencies, as standard batching leads to excessive padding, increasing resource consumption and execution time. To address this, we introduce ChunkFormer, an efficient ASR model that uses chunk-wise processing with relative right context, enabling long audio transcriptions on low-memory GPUs. ChunkFormer handles up to 16 hours of audio on an 80GB GPU, 1.5x longer than the current state-of-the-art FastConformer, while also boosting long-form transcription performance with up to 7.7% absolute reduction on word error rate and maintaining accuracy on shorter tasks compared to Conformer. By eliminating the need for padding in standard batching, ChunkFormer's masked batching technique reduces execution time and memory usage by more than 3x in batch processing, substantially reducing costs for a wide range of ASR systems, particularly regarding GPU resources for models serving in real-world applications.

Transformer models have been successful in various sequence processing tasks, but the self-attention mechanism's computational cost limits its practicality for long sequences. Although there are existing attention variants that improve computational efficiency, they have a limited ability to abstract global information effectively based on their hand-crafted mixing strategies. On the other hand, state-space models (SSMs) are tailored for long sequences but cannot capture complicated local information. Therefore, the combination of them as a unified token mixer is a trend in recent long-sequence models. However, the linearized attention degrades performance significantly even when equipped with SSMs. To address the issue, we propose a new method called LongVQ. LongVQ uses the vector quantization (VQ) technique to compress the global abstraction as a length-fixed codebook, enabling the linear-time computation of the attention matrix. This technique effectively maintains dynamic global and local patterns, which helps to complement the lack of long-range dependency issues. Our experiments on the Long Range Arena benchmark, autoregressive language modeling, and image and speech classification demonstrate the effectiveness of LongVQ. Our model achieves significant improvements over other sequence models, including variants of Transformers, Convolutions, and recent State Space Models.

Large language model (LLM) inference workload dominates a wide variety of modern AI applications, ranging from multi-turn conversation to document analysis. Balancing fairness and efficiency is critical for managing diverse client workloads with varying prefix patterns. Unfortunately, existing fair scheduling algorithms for LLM serving, such as Virtual Token Counter (VTC), fail to take prefix locality into consideration and thus suffer from poor performance. On the other hand, locality-aware scheduling algorithms in existing LLM serving frameworks tend to maximize the prefix cache hit rate without considering fair sharing among clients. This paper introduces the first locality-aware fair scheduling algorithm, Deficit Longest Prefix Match (DLPM), which can maintain a high degree of prefix locality with a fairness guarantee. We also introduce a novel algorithm, Double Deficit LPM (D^2LPM), extending DLPM for the distributed setup that can find a balance point among fairness, locality, and load-balancing. Our extensive evaluation demonstrates the superior performance of DLPM and D^2LPM in ensuring fairness while maintaining high throughput (up to 2.87times higher than VTC) and low per-client (up to 7.18times lower than state-of-the-art distributed LLM serving system) latency.

As large language models gain widespread adoption, running them efficiently becomes crucial. Recent works on LLM inference use speculative decoding to achieve extreme speedups. However, most of these works implicitly design their algorithms for high-end datacenter hardware. In this work, we ask the opposite question: how fast can we run LLMs on consumer machines? Consumer GPUs can no longer fit the largest available models (50B+ parameters) and must offload them to RAM or SSD. When running with offloaded parameters, the inference engine can process batches of hundreds or thousands of tokens at the same time as just one token, making it a natural fit for speculative decoding. We propose SpecExec (Speculative Execution), a simple parallel decoding method that can generate up to 20 tokens per target model iteration for popular LLM families. It utilizes the high spikiness of the token probabilities distribution in modern LLMs and a high degree of alignment between model output probabilities. SpecExec takes the most probable tokens continuation from the draft model to build a "cache" tree for the target model, which then gets validated in a single pass. Using SpecExec, we demonstrate inference of 50B+ parameter LLMs on consumer GPUs with RAM offloading at 4-6 tokens per second with 4-bit quantization or 2-3 tokens per second with 16-bit weights.

This technical report describes the design and training of novel speculative decoding draft models, for accelerating the inference speeds of large language models in a production environment. By conditioning draft predictions on both context vectors and sampled tokens, we can train our speculators to efficiently predict high-quality n-grams, which the base model then accepts or rejects. This allows us to effectively predict multiple tokens per inference forward pass, accelerating wall-clock inference speeds of highly optimized base model implementations by a factor of 2-3x. We explore these initial results and describe next steps for further improvements.

Transformer models have demonstrated high accuracy in numerous applications but have high complexity and lack sequential processing capability making them ill-suited for many streaming applications at the edge where devices are heavily resource-constrained. Thus motivated, many researchers have proposed reformulating the transformer models as RNN modules which modify the self-attention computation with explicit states. However, these approaches often incur significant performance degradation. The ultimate goal is to develop a model that has the following properties: parallel training, streaming and low-cost inference, and SOTA performance. In this paper, we propose a new direction to achieve this goal. We show how architectural modifications to a recurrent model can help push its performance toward Transformer models while retaining its sequential processing capability. Specifically, inspired by the recent success of Legendre Memory Units (LMU) in sequence learning tasks, we propose LMUFormer, which augments the LMU with convolutional patch embedding and convolutional channel mixer. Moreover, we present a spiking version of this architecture, which introduces the benefit of states within the patch embedding and channel mixer modules while simultaneously reducing the computing complexity. We evaluated our architectures on multiple sequence datasets. In comparison to SOTA transformer-based models within the ANN domain on the SCv2 dataset, our LMUFormer demonstrates comparable performance while necessitating a remarkable 53 times reduction in parameters and a substantial 65 times decrement in FLOPs. Additionally, owing to our model's proficiency in real-time data processing, we can achieve a 32.03% reduction in sequence length, all while incurring an inconsequential decline in performance. Our code is publicly available at https://github.com/zeyuliu1037/LMUFormer.git.

Recent advances in GPT-4o like multi-modality models have demonstrated remarkable progress for direct speech-to-speech conversation, with real-time speech interaction experience and strong speech understanding ability. However, current research focuses on discrete speech tokens to align with discrete text tokens for language modelling, which depends on an audio codec with residual connections or independent group tokens, such a codec usually leverages large scale and diverse datasets training to ensure that the discrete speech codes have good representation for varied domain, noise, style data reconstruction as well as a well-designed codec quantizer and encoder-decoder architecture for discrete token language modelling. This paper introduces Flow-Omni, a continuous speech token based GPT-4o like model, capable of real-time speech interaction and low streaming latency. Specifically, first, instead of cross-entropy loss only, we combine flow matching loss with a pretrained autoregressive LLM and a small MLP network to predict the probability distribution of the continuous-valued speech tokens from speech prompt. second, we incorporated the continuous speech tokens to Flow-Omni multi-modality training, thereby achieving robust speech-to-speech performance with discrete text tokens and continuous speech tokens together. Experiments demonstrate that, compared to discrete text and speech multi-modality training and its variants, the continuous speech tokens mitigate robustness issues by avoiding the inherent flaws of discrete speech code's representation loss for LLM.

Over the last few years, multi-vector retrieval methods, spearheaded by ColBERT, have become an increasingly popular approach to Neural IR. By storing representations at the token level rather than at the document level, these methods have demonstrated very strong retrieval performance, especially in out-of-domain settings. However, the storage and memory requirements necessary to store the large number of associated vectors remain an important drawback, hindering practical adoption. In this paper, we introduce a simple clustering-based token pooling approach to aggressively reduce the number of vectors that need to be stored. This method can reduce the space & memory footprint of ColBERT indexes by 50% with virtually no retrieval performance degradation. This method also allows for further reductions, reducing the vector count by 66%-to-75% , with degradation remaining below 5% on a vast majority of datasets. Importantly, this approach requires no architectural change nor query-time processing, and can be used as a simple drop-in during indexation with any ColBERT-like model.

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.

Popular video training methods mainly operate on a fixed number of tokens sampled from a predetermined spatiotemporal grid, resulting in sub-optimal accuracy-computation trade-offs due to inherent video redundancy. They also lack adaptability to varying computational budgets for downstream tasks, hindering applications of the most competitive model in real-world scenes. We thus propose a new test setting, Token Optimization, for maximized input information across budgets, which optimizes the size-limited set of input tokens through token selection from more suitably sampled videos. To this end, we propose a novel augmentation tool termed Flux. By making the sampling grid flexible and leveraging token selection, it is easily adopted in most popular video training frameworks, boosting model robustness with nearly no additional cost. We integrate Flux in large-scale video pre-training, and the resulting FluxViT establishes new state-of-the-art results across extensive tasks at standard costs. Notably, with 1/4 tokens only, it can still match the performance of previous state-of-the-art models with Token Optimization, yielding nearly 90\% savings. All models and data are available at https://github.com/OpenGVLab/FluxViT.

Simultaneous translation of unbounded streaming speech remains a challenging problem due to the need for effectively processing the history speech context and past translations so that quality and latency, including computation overhead, can be balanced. Most prior works assume pre-segmented speech, limiting their real-world applicability. In this paper, we propose InfiniSST, a novel approach that formulates SST as a multi-turn dialogue task, enabling seamless translation of unbounded speech. We construct translation trajectories and robust segments from MuST-C with multi-latency augmentation during training and develop a key-value (KV) cache management strategy to facilitate efficient inference. Experiments on MuST-C En-Es, En-De, and En-Zh demonstrate that InfiniSST reduces computation-aware latency by 0.5 to 1 second while maintaining the same translation quality compared to baselines. Ablation studies further validate the contributions of our data construction and cache management strategy. We release the code and demo at https://github.com/LeiLiLab/InfiniSST

In distributed computing, slower nodes (stragglers) usually become a bottleneck. Gradient Coding (GC), introduced by Tandon et al., is an efficient technique that uses principles of error-correcting codes to distribute gradient computation in the presence of stragglers. In this paper, we consider the distributed computation of a sequence of gradients {g(1),g(2),ldots,g(J)}, where processing of each gradient g(t) starts in round-t and finishes by round-(t+T). Here Tgeq 0 denotes a delay parameter. For the GC scheme, coding is only across computing nodes and this results in a solution where T=0. On the other hand, having T>0 allows for designing schemes which exploit the temporal dimension as well. In this work, we propose two schemes that demonstrate improved performance compared to GC. Our first scheme combines GC with selective repetition of previously unfinished tasks and achieves improved straggler mitigation. In our second scheme, which constitutes our main contribution, we apply GC to a subset of the tasks and repetition for the remainder of the tasks. We then multiplex these two classes of tasks across workers and rounds in an adaptive manner, based on past straggler patterns. Using theoretical analysis, we demonstrate that our second scheme achieves significant reduction in the computational load. In our experiments, we study a practical setting of concurrently training multiple neural networks over an AWS Lambda cluster involving 256 worker nodes, where our framework naturally applies. We demonstrate that the latter scheme can yield a 16\% improvement in runtime over the baseline GC scheme, in the presence of naturally occurring, non-simulated stragglers.

The growth in the number of parameters of Large Language Models (LLMs) has led to a significant surge in computational requirements, making them challenging and costly to deploy. Speculative decoding (SD) leverages smaller models to efficiently propose future tokens, which are then verified by the LLM in parallel. Small models that utilise activations from the LLM currently achieve the fastest decoding speeds. However, we identify several limitations of SD models including the lack of on-policyness during training and partial observability. To address these shortcomings, we propose a more grounded architecture for small models by introducing a Mixture of Attentions for SD. Our novel architecture can be applied in two scenarios: a conventional single device deployment and a novel client-server deployment where the small model is hosted on a consumer device and the LLM on a server. In a single-device scenario, we demonstrate state-of-the-art speedups improving EAGLE-2 by 9.5% and its acceptance length by 25%. In a client-server setting, our experiments demonstrate: 1) state-of-the-art latencies with minimal calls to the server for different network conditions, and 2) in the event of a complete disconnection, our approach can maintain higher accuracy compared to other SD methods and demonstrates advantages over API calls to LLMs, which would otherwise be unable to continue the generation process.

In this report, we present Qwen2.5-Omni, an end-to-end multimodal model designed to perceive diverse modalities, including text, images, audio, and video, while simultaneously generating text and natural speech responses in a streaming manner. To enable the streaming of multimodal information inputs, both audio and visual encoders utilize a block-wise processing approach. To synchronize the timestamps of video inputs with audio, we organize the audio and video sequentially in an interleaved manner and propose a novel position embedding approach, named TMRoPE(Time-aligned Multimodal RoPE). To concurrently generate text and speech while avoiding interference between the two modalities, we propose Thinker-Talker architecture. In this framework, Thinker functions as a large language model tasked with text generation, while Talker is a dual-track autoregressive model that directly utilizes the hidden representations from the Thinker to produce audio tokens as output. Both the Thinker and Talker models are designed to be trained and inferred in an end-to-end manner. For decoding audio tokens in a streaming manner, we introduce a sliding-window DiT that restricts the receptive field, aiming to reduce the initial package delay. Qwen2.5-Omni is comparable with the similarly sized Qwen2.5-VL and outperforms Qwen2-Audio. Furthermore, Qwen2.5-Omni achieves state-of-the-art performance on multimodal benchmarks like Omni-Bench. Notably, Qwen2.5-Omni's performance in end-to-end speech instruction following is comparable to its capabilities with text inputs, as evidenced by benchmarks such as MMLU and GSM8K. As for speech generation, Qwen2.5-Omni's streaming Talker outperforms most existing streaming and non-streaming alternatives in robustness and naturalness.

Practical large language model (LLM) services may involve a long system prompt, which specifies the instructions, examples, and knowledge documents of the task and is reused across numerous requests. However, the long system prompt causes throughput/latency bottlenecks as the cost of generating the next token grows w.r.t. the sequence length. This paper aims to improve the efficiency of LLM services that involve long system prompts. Our key observation is that handling these system prompts requires heavily redundant memory accesses in existing causal attention computation algorithms. Specifically, for batched requests, the cached hidden states (i.e., key-value pairs) of system prompts are transferred from off-chip DRAM to on-chip SRAM multiple times, each corresponding to an individual request. To eliminate such a redundancy, we propose RelayAttention, an attention algorithm that allows reading these hidden states from DRAM exactly once for a batch of input tokens. RelayAttention is a free lunch: it maintains the generation quality while requiring no model retraining, as it is based on a mathematical reformulation of causal attention.

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.

Streaming speech recognition architectures are employed for low-latency, real-time applications. Such architectures are often characterized by their causality. Causal architectures emit tokens at each frame, relying only on current and past signal, while non-causal models are exposed to a window of future frames at each step to increase predictive accuracy. This dichotomy amounts to a trade-off for real-time Automatic Speech Recognition (ASR) system design: profit from the low-latency benefit of strictly-causal architectures while accepting predictive performance limitations, or realize the modeling benefits of future-context models accompanied by their higher latency penalty. In this work, we relax the constraints of this choice and present the Adaptive Non-Causal Attention Transducer (ANCAT). Our architecture is non-causal in the traditional sense, but executes in a low-latency, streaming manner by dynamically choosing when to rely on future context and to what degree within the audio stream. The resulting mechanism, when coupled with our novel regularization algorithms, delivers comparable accuracy to non-causal configurations while improving significantly upon latency, closing the gap with their causal counterparts. We showcase our design experimentally by reporting comparative ASR task results with measures of accuracy and latency on both publicly accessible and production-scale, voice-assistant datasets.

Transformer-based large language models (LLMs) exhibit impressive performance in generative tasks but introduce significant challenges in real-world serving due to inefficient use of the expensive, computation-optimized accelerators. This mismatch arises from the autoregressive nature of LLMs, where the generation phase comprises operators with varying resource demands. Specifically, the attention operator is memory-intensive, exhibiting a memory access pattern that clashes with the strengths of modern accelerators, especially as context length increases. To enhance the efficiency and cost-effectiveness of LLM serving, we introduce the concept of attention offloading. This approach leverages a collection of cheap, memory-optimized devices for the attention operator while still utilizing high-end accelerators for other parts of the model. This heterogeneous setup ensures that each component is tailored to its specific workload, maximizing overall performance and cost efficiency. Our comprehensive analysis and experiments confirm the viability of splitting the attention computation over multiple devices. Also, the communication bandwidth required between heterogeneous devices proves to be manageable with prevalent networking technologies. To further validate our theory, we develop Lamina, an LLM inference system that incorporates attention offloading. Experimental results indicate that Lamina can provide 1.48x-12.1x higher estimated throughput per dollar than homogeneous solutions.

Mixture-of-Experts (MoE) model architecture has emerged as a promising solution for scaling transformer models efficiently, offering sparse activation that reduces computational costs while increasing model capacity. However, as MoE models scale, they need to be distributed across GPU devices, thus face critical performance bottlenecks due to their large memory footprint. Expert parallelism distributes experts across GPUs, however, faces key challenges including an unbalanced token routing and expert activation, resulting in communication tail latency and processing inefficiencies. While existing solutions address some of these issues, they fail to resolve the dual challenges of load imbalance and communication skew. The imbalance in token processing load across experts causes uneven processing times on different GPUs, while communication skew between GPUs leads to unbalanced inter-GPU data transfers. These factors degrade the performance of MoE models by increasing tail latency and reducing overall throughput. To address these limitations, we propose an Integer Linear Programming (ILP) formulation to optimize expert placement by jointly considering token load, communication, and computation costs. We exploit the property that there is a token routing dependency across layers, where tokens routed to a specific expert in one layer are likely to be routed to a limited set of experts in the subsequent layer. Our solution, MoETuner, offers an optimal expert-to-GPU assignment that minimizes inter-GPU token routing costs and balances token processing across devices, thereby reducing tail latency and end-to-end execution time. Experimental results demonstrate 9.3% and 17.5% of end-to-end speedups for single-node and multi-node inference respectively, showcasing the potential of our ILP-based optimization for offering expert parallel solutions for next-generation MoEs.

While speculative decoding has recently appeared as a promising direction for accelerating the inference of large language models (LLMs), the speedup and scalability are strongly bounded by the token acceptance rate. Prevalent methods usually organize predicted tokens as independent chains or fixed token trees, which fails to generalize to diverse query distributions. In this paper, we propose DySpec, a faster speculative decoding algorithm with a novel dynamic token tree structure. We begin by bridging the draft distribution and acceptance rate from intuitive and empirical clues, and successfully show that the two variables are strongly correlated. Based on this, we employ a greedy strategy to dynamically expand the token tree at run time. Theoretically, we show that our method can achieve optimal results under mild assumptions. Empirically, DySpec yields a higher acceptance rate and speedup than fixed trees. DySpec can drastically improve the throughput and reduce the latency of token generation across various data distribution and model sizes, which significantly outperforms strong competitors, including Specinfer and Sequoia. Under low temperature setting, DySpec can improve the throughput up to 9.1times and reduce the latency up to 9.4times on Llama2-70B. Under high temperature setting, DySpec can also improve the throughput up to 6.21times, despite the increasing difficulty of speculating more than one token per step for draft model.

Neural Temporal Point Processes (TPPs) are the prevalent paradigm for modeling continuous-time event sequences, such as user activities on the web and financial transactions. In real-world applications, event data is typically received in a streaming manner, where the distribution of patterns may shift over time. Additionally, privacy and memory constraints are commonly observed in practical scenarios, further compounding the challenges. Therefore, the continuous monitoring of a TPP to learn the streaming event sequence is an important yet under-explored problem. Our work paper addresses this challenge by adopting Continual Learning (CL), which makes the model capable of continuously learning a sequence of tasks without catastrophic forgetting under realistic constraints. Correspondingly, we propose a simple yet effective framework, PromptTPPOur code is available at {\small \url{ https://github.com/yanyanSann/PromptTPP}}, by integrating the base TPP with a continuous-time retrieval prompt pool. The prompts, small learnable parameters, are stored in a memory space and jointly optimized with the base TPP, ensuring that the model learns event streams sequentially without buffering past examples or task-specific attributes. We present a novel and realistic experimental setup for modeling event streams, where PromptTPP consistently achieves state-of-the-art performance across three real user behavior datasets.

Large Language Models (LLMs) achieve impressive reasoning capabilities at the cost of substantial inference overhead, posing substantial deployment challenges. Although distilled Small Language Models (SLMs) significantly enhance efficiency, their performance suffers as they fail to follow LLMs' reasoning paths. Luckily, we reveal that only a small fraction of tokens genuinely diverge reasoning paths between LLMs and SLMs. Most generated tokens are either identical or exhibit neutral differences, such as minor variations in abbreviations or expressions. Leveraging this insight, we introduce **Roads to Rome (R2R)**, a neural token routing method that selectively utilizes LLMs only for these critical, path-divergent tokens, while leaving the majority of token generation to the SLM. We also develop an automatic data generation pipeline that identifies divergent tokens and generates token-level routing labels to train the lightweight router. We apply R2R to combine R1-1.5B and R1-32B models from the DeepSeek family, and evaluate on challenging math, coding, and QA benchmarks. With an average activated parameter size of 5.6B, R2R surpasses the average accuracy of R1-7B by 1.6x, outperforming even the R1-14B model. Compared to R1-32B, it delivers a 2.8x wall-clock speedup with comparable performance, advancing the Pareto frontier of test-time scaling efficiency. Our code is available at https://github.com/thu-nics/R2R.

The rapid growth in the parameters of large language models (LLMs) has made inference latency a fundamental bottleneck, limiting broader application of LLMs. Speculative decoding represents a lossless approach to accelerate inference through a guess-and-verify paradigm, leveraging the parallel capabilities of modern hardware. Some speculative decoding methods rely on additional structures to guess draft tokens, such as small models or parameter-efficient architectures, which need extra training before use. Alternatively, retrieval-based train-free techniques build libraries from pre-existing corpora or by n-gram generation. However, they face challenges like large storage requirements, time-consuming retrieval, and limited adaptability. Observing that candidate tokens generated during the decoding process are likely to reoccur in future sequences, we propose Token Recycling. This approach stores candidate tokens in an adjacency matrix and employs a breadth-first search (BFS)-like algorithm on the matrix to construct a draft tree. The tree is then validated through tree attention. New candidate tokens from the decoding process are then used to update the matrix. Token Recycling requires \textless2MB of additional storage and achieves approximately 2x speedup across all sizes of LLMs. It significantly outperforms existing train-free methods by 30\% and even a training method by 25\%. It can be directly applied to any existing LLMs and tasks without the need for adaptation.

Large language models have shown impressive capabilities across a variety of NLP tasks, yet their generating text autoregressively is time-consuming. One way to speed them up is speculative decoding, which generates candidate segments (a sequence of tokens) from a fast draft model that is then verified in parallel by the target model. However, the acceptance rate of candidate tokens receives limitations from several factors, such as the model, the dataset, and the decoding setup. This paper proposes sampling multiple candidates from a draft model and then organising them in batches for verification. We design algorithms for efficient multi-candidate verification while maintaining the distribution of the target model. Our approach shows significant improvements in acceptance rates on multiple datasets and models, consistently outperforming standard speculative decoding.

Attention for transformers is a critical workload that has recently received significant "attention" as a target for custom acceleration. Yet, while prior work succeeds in reducing attention's memory-bandwidth requirements, it creates load imbalance between attention operators (resulting in severe compute under-utilization) and requires on-chip memory that scales with sequence length (which is expected to grow over time). This paper ameliorates these issues, enabling attention with nearly 100% compute utilization, no off-chip memory traffic bottlenecks, and on-chip buffer size requirements that are independent of sequence length. The main conceptual contribution is to use a recently proposed abstraction -- the cascade of Einsums -- to describe, formalize and taxonomize the space of attention algorithms that appear in the literature. In particular, we show how Einsum cascades can be used to infer non-trivial lower bounds on the number of passes a kernel must take through its input data, which has implications for either required on-chip buffer capacity or memory traffic. We show how this notion can be used to meaningfully divide the space of attention algorithms into several categories and use these categories to inform our design process. Based on the above characterization, we propose FuseMax -- a novel mapping of attention onto a spatial array-style architecture. On attention, in an iso-area comparison, FuseMax achieves an average 6.7times speedup over the prior state-of-the-art FLAT while using 79% of the energy. Similarly, on the full end-to-end transformer inference, FuseMax achieves an average 5.3times speedup over FLAT using 83% of the energy.

Speculative Decoding has gained popularity as an effective technique for accelerating the auto-regressive inference process of Large Language Models (LLMs). However, Speculative Decoding entirely relies on the availability of efficient draft models, which are often lacking for many existing language models due to a stringent constraint of vocabulary incompatibility. In this work we introduce FastDraft, a novel and efficient approach for pre-training and aligning a draft model to any large language model by incorporating efficient pre-training, followed by fine-tuning over synthetic datasets generated by the target model. We demonstrate FastDraft by training two highly parameter efficient drafts for the popular Phi-3-mini and Llama-3.1-8B models. Using FastDraft, we were able to produce a draft with approximately 10 billion tokens on a single server with 8 Intel^circledR Gaudi^circledR 2 accelerators in under 24 hours. Our results show that the draft model achieves impressive results in key metrics of acceptance rate, block efficiency and up to 3x memory bound speed up when evaluated on code completion and up to 2x in summarization, text completion and instruction tasks. We validate our theoretical findings through benchmarking on the latest Intel^circledR Core^{tiny TM} Ultra, achieving a wall-clock time speedup of up to 2x, indicating a significant reduction in runtime. Due to its high quality, FastDraft unlocks large language models inference on AI-PC and other edge-devices.

Full-duplex dialog models are designed to listen and speak simultaneously with rapid responses to fast-changing user input. Among existing approaches, native full-duplex models merges different channels (e.g. listen and speak) in a single time step, overcoming the high response latency inherent to time-division multiplexing time-division multiplexing (TDM) alternatives. Yet, a key challenge remains: aligning textual monologues with audio streams that operate at different bitrates. The prevailing solution relies on word-level alignment, but this can degrade the language ability of large pre-trained models. Moreover, it requires highly accurate timestamps for every token, which introduces cascading errors and increases pre-processing costs. In this paper, we propose textual monologues in continuous tokens sequence, namely "natural" monologues, which mimics humanoid cognitive behavior in dialogs. For temporal alignment, we alternate the position of the natural monologue - leading or trailing the audio - across different training stages. This "dual" training paradigm proves highly effective in building FLM-Audio, our 7B spoken dialog model that demonstrates superior responsiveness, duplexity, and chatting experiences, as confirmed by experimental results.

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.

Inference of Large Language Models (LLMs) across computer clusters has become a focal point of research in recent times, with many acceleration techniques taking inspiration from CPU speculative execution. These techniques reduce bottlenecks associated with memory bandwidth, but also increase end-to-end latency per inference run, requiring high speculation acceptance rates to improve performance. Combined with a variable rate of acceptance across tasks, speculative inference techniques can result in reduced performance. Additionally, pipeline-parallel designs require many user requests to maintain maximum utilization. As a remedy, we propose PipeInfer, a pipelined speculative acceleration technique to reduce inter-token latency and improve system utilization for single-request scenarios while also improving tolerance to low speculation acceptance rates and low-bandwidth interconnects. PipeInfer exhibits up to a 2.15times improvement in generation speed over standard speculative inference. PipeInfer achieves its improvement through Continuous Asynchronous Speculation and Early Inference Cancellation, the former improving latency and generation speed by running single-token inference simultaneously with several speculative runs, while the latter improves speed and latency by skipping the computation of invalidated runs, even in the middle of inference.

Effective attention modules have played a crucial role in the success of Transformer-based large language models (LLMs), but the quadratic time and memory complexities of these attention modules also pose a challenge when processing long sequences. One potential solution for the long sequence problem is to utilize distributed clusters to parallelize the computation of attention modules across multiple devices (e.g., GPUs). However, adopting a distributed approach inevitably introduces extra memory overheads to store local attention results and incurs additional communication costs to aggregate local results into global ones. In this paper, we propose a distributed attention framework named ``BurstAttention'' to optimize memory access and communication operations at both the global cluster and local device levels. In our experiments, we compare BurstAttention with other competitive distributed attention solutions for long sequence processing. The experimental results under different length settings demonstrate that BurstAttention offers significant advantages for processing long sequences compared with these competitive baselines, reducing 40% communication overheads and achieving 2 X speedup during training 32K sequence length on 8 X A100.

Recent Large Language Models have been enhanced with vision capabilities, enabling them to comprehend images, videos, and interleaved vision-language content. However, the learning methods of these large multimodal models typically treat videos as predetermined clips, making them less effective and efficient at handling streaming video inputs. In this paper, we propose a novel Learning-In-Video-Stream (LIVE) framework, which enables temporally aligned, long-context, and real-time conversation within a continuous video stream. Our LIVE framework comprises comprehensive approaches to achieve video streaming dialogue, encompassing: (1) a training objective designed to perform language modeling for continuous streaming inputs, (2) a data generation scheme that converts offline temporal annotations into a streaming dialogue format, and (3) an optimized inference pipeline to speed up the model responses in real-world video streams. With our LIVE framework, we built VideoLLM-online model upon Llama-2/Llama-3 and demonstrate its significant advantages in processing streaming videos. For instance, on average, our model can support streaming dialogue in a 5-minute video clip at over 10 FPS on an A100 GPU. Moreover, it also showcases state-of-the-art performance on public offline video benchmarks, such as recognition, captioning, and forecasting. The code, model, data, and demo have been made available at https://showlab.github.io/videollm-online.

Diffusion language models offer parallel token generation and inherent bidirectionality, promising more efficient and powerful sequence modeling compared to autoregressive approaches. However, state-of-the-art diffusion models (e.g., Dream 7B, LLaDA 8B) suffer from slow inference. While they match the quality of similarly sized Autoregressive (AR) Models (e.g., Qwen2.5 7B, Llama3 8B), their iterative denoising requires multiple full-sequence forward passes, resulting in high computational costs and latency, particularly for long input prompts and long-context scenarios. Furthermore, parallel token generation introduces token incoherence problems, and current sampling heuristics suffer from significant quality drops with decreasing denoising steps. We address these limitations with two training-free techniques. First, we propose FreeCache, a Key-Value (KV) approximation caching technique that reuses stable KV projections across denoising steps, effectively reducing the computational cost of DLM inference. Second, we introduce Guided Diffusion, a training-free method that uses a lightweight pretrained autoregressive model to supervise token unmasking, dramatically reducing the total number of denoising iterations without sacrificing quality. We conduct extensive evaluations on open-source reasoning benchmarks, and our combined methods deliver up to a 34x end-to-end speedup without compromising accuracy. For the first time, diffusion language models achieve a comparable and even faster latency as the widely adopted autoregressive models. Our work successfully paved the way for scaling up the diffusion language model to a broader scope of applications across different domains.

Speculative decoding has emerged as a powerful method to improve latency and throughput in hosting large language models. However, most existing implementations focus on generating a single sequence. Real-world generative AI applications often require multiple responses and how to perform speculative decoding in a batched setting while preserving its latency benefits poses non-trivial challenges. This paper describes a system of batched speculative decoding that sets a new state of the art in multi-sequence generation latency and that demonstrates superior GPU utilization as well as quality of generations within a time budget. For example, for a 7.8B-size model on a single A100 GPU and with a batch size of 8, each sequence is generated at an average speed of 5.8ms per token, the overall throughput being 1.1K tokens per second. These results represent state-of-the-art latency and a 2.15X speed-up over optimized regular decoding. Within a time budget that regular decoding does not finish, our system is able to generate sequences with HumanEval Pass@First of 43% and Pass@All of 61%, far exceeding what's feasible with single-sequence speculative decoding. Our peak GPU utilization during decoding reaches as high as 15.8%, more than 3X the highest of that of regular decoding and around 10X of single-sequence speculative decoding.

Linear sequence modeling approaches, such as linear attention, provide advantages like linear-time training and constant-memory inference over sequence lengths. However, existing sequence parallelism (SP) methods are either not optimized for the right-product-first feature of linear attention or use a ring-style communication strategy, which results in lower computation parallelism, limits their scalability for longer sequences in distributed systems. In this paper, we introduce LASP-2, a new SP method to enhance both communication and computation parallelism when training linear attention transformer models with very-long input sequences. Compared to previous work LASP, LASP-2 rethinks the minimal communication requirement for SP on linear attention layers, reorganizes the whole communication-computation workflow of LASP. In this way, only one single AllGather collective communication is needed on intermediate memory states, whose sizes are independent of the sequence length, leading to significant improvements of both communication and computation parallelism, as well as their overlap. Additionally, we extend LASP-2 to LASP-2H by applying similar communication redesign to standard attention modules, offering an efficient SP solution for hybrid models that blend linear and standard attention layers. Our evaluation on a Linear-Llama3 model, a variant of Llama3 with linear attention replacing standard attention, demonstrates the effectiveness of LASP-2 and LASP-2H. Specifically, LASP-2 achieves training speed improvements of 15.2% over LASP and 36.6% over Ring Attention, with a sequence length of 2048K across 64 GPUs. The Code is released as a part of: https://github.com/OpenSparseLLMs/Linear-MoE.

Due to the high resource demands of Large Language Models (LLMs), achieving widespread deployment on consumer-grade devices presents significant challenges. Typically, personal or consumer-grade devices, including servers configured prior to the era of large-scale models, generally have relatively weak GPUs and relatively strong CPUs. However, most current methods primarily depend on GPUs for computation. Therefore, we propose Dovetail, an approach that deploys the draft model on the GPU to generate draft tokens while allowing the target model to perform parallel verification on the CPU, thereby improving the utilization of all available hardware resources and occupying less inter-device communication bandwidth. Accordingly, we have redesigned the draft model to better align with heterogeneous hardware characteristics. To this end, we implemented several optimizations: reducing the number of draft tokens to mitigate latency in parallel verification, increasing the depth of the draft model to enhance its predictive capacity, and introducing DGF (Dynamic Gating Fusion) to improve the integration of features and token embeddings. In the HumanEval benchmark, Dovetail achieved an inference speed of 5.86 tokens per second for LLaMA2-Chat-7B using 3GB of VRAM, representing an approximately 2.77x improvement over CPU-only inference. Furthermore, the inference speed was increased to 8 tokens per second when utilizing 7GB of VRAM.

The attention mechanism is becoming increasingly popular in Natural Language Processing (NLP) applications, showing superior performance than convolutional and recurrent architectures. However, attention becomes the compution bottleneck because of its quadratic computational complexity to input length, complicated data movement and low arithmetic intensity. Moreover, existing NN accelerators mainly focus on optimizing convolutional or recurrent models, and cannot efficiently support attention. In this paper, we present SpAtten, an efficient algorithm-architecture co-design that leverages token sparsity, head sparsity, and quantization opportunities to reduce the attention computation and memory access. Inspired by the high redundancy of human languages, we propose the novel cascade token pruning to prune away unimportant tokens in the sentence. We also propose cascade head pruning to remove unessential heads. Cascade pruning is fundamentally different from weight pruning since there is no trainable weight in the attention mechanism, and the pruned tokens and heads are selected on the fly. To efficiently support them on hardware, we design a novel top-k engine to rank token and head importance scores with high throughput. Furthermore, we propose progressive quantization that first fetches MSBs only and performs the computation; if the confidence is low, it fetches LSBs and recomputes the attention outputs, trading computation for memory reduction. Extensive experiments on 30 benchmarks show that, on average, SpAtten reduces DRAM access by 10.0x with no accuracy loss, and achieves 1.6x, 3.0x, 162x, 347x speedup, and 1,4x, 3.2x, 1193x, 4059x energy savings over A3 accelerator, MNNFast accelerator, TITAN Xp GPU, Xeon CPU, respectively.

Transformers have demonstrated strong performance across a wide range of sequence modeling tasks, but their quadratic attention complexity limits scalability to long sequences. Linear models such as Mamba and sliding-window attention (SWA) address this by mixing tokens through recurrent or localized operations with fixed-size memory, achieving efficient inference. However, these methods risk degrading performance on long sequences due to their inability to retain detailed information from distant tokens. We propose SCOUT (Segment Compression for Optimized Utility in Transformers), a hybrid architecture that compresses tokens locally within fixed-size segments and applies attention only over these compressed representations. Each token embedding is first enriched via a linear local mixer, Mamba or SWA, that integrates recent context. Then, instead of attending to all previous tokens, each token sparsely attends to a small number of compressed checkpoint tokens that summarize the input history. This design retains much of the expressivity of full attention while substantially reducing the computational and memory cost. By attending to compressed history rather than all previous tokens, SCOUT incurs slightly higher memory than purely linear models, but its growth rate remains sub-quadratic and far more scalable than that of full Transformers. We analyze SCOUT's computational and memory efficiency and evaluate it empirically on long-context language modeling and reasoning tasks. SCOUT with both Mamba and SWA mixers outperforms strong long-sequence baselines under the same computational budget, matches full-attention Transformers on language modeling and common-sense reasoning tasks at 400M and 1.3B scales. Moreover, our SCOUT achieves higher end-to-end throughput than SOTA models, while delivering comparable results on long sequence benchmarks.

Vision-language models (VLMs) could power real-time assistants and autonomous agents, but they face a critical challenge: understanding near-infinite video streams without escalating latency and memory usage. Processing entire videos with full attention leads to quadratic computational costs and poor performance on long videos. Meanwhile, simple sliding window methods are also flawed, as they either break coherence or suffer from high latency due to redundant recomputation. In this paper, we introduce StreamingVLM, a model designed for real-time, stable understanding of infinite visual input. Our approach is a unified framework that aligns training with streaming inference. During inference, we maintain a compact KV cache by reusing states of attention sinks, a short window of recent vision tokens, and a long window of recent text tokens. This streaming ability is instilled via a simple supervised fine-tuning (SFT) strategy that applies full attention on short, overlapped video chunks, which effectively mimics the inference-time attention pattern without training on prohibitively long contexts. For evaluation, we build Inf-Streams-Eval, a new benchmark with videos averaging over two hours that requires dense, per-second alignment between frames and text. On Inf-Streams-Eval, StreamingVLM achieves a 66.18% win rate against GPT-4O mini and maintains stable, real-time performance at up to 8 FPS on a single NVIDIA H100. Notably, our SFT strategy also enhances general VQA abilities without any VQA-specific fine-tuning, improving performance on LongVideoBench by +4.30 and OVOBench Realtime by +5.96. Code is available at https://github.com/mit-han-lab/streaming-vlm.

Large-scale distributed training in production datacenters constitutes a challenging workload bottlenecked by network communication. In response, both major industry players (e.g., Ultra Ethernet Consortium) and parts of academia have surprisingly, and almost unanimously, agreed that packet spraying is necessary to improve the performance of large-scale distributed training workloads. In this paper, we challenge this prevailing belief and pose the question: How close can a singlepath transport approach an optimal multipath transport? We demonstrate that singlepath transport (from a NIC's perspective) is sufficient and can perform nearly as well as an ideal multipath transport with packet spraying, particularly in the context of distributed training in leaf-spine topologies. Our assertion is based on four key observations about workloads driven by collective communication patterns: (i) flows within a collective start almost simultaneously, (ii) flow sizes are nearly equal, (iii) the completion time of a collective is more crucial than individual flow completion times, and (iv) flows can be split upon arrival. We analytically prove that singlepath transport, using minimal flow splitting (at the application layer), is equivalent to an ideal multipath transport with packet spraying in terms of maximum congestion. Our preliminary evaluations support our claims. This paper suggests an alternative agenda for developing next-generation transport protocols tailored for large-scale distributed training.

The increasing usage of Large Language Models (LLMs) has resulted in a surging demand for planet-scale serving systems, where tens of thousands of GPUs continuously serve hundreds of millions of users. Consequently, throughput (under reasonable latency constraints) has emerged as a key metric that determines serving systems' performance. To boost throughput, various methods of inter-device parallelism (e.g., data, tensor, pipeline) have been explored. However, existing methods do not consider overlapping the utilization of different resources within a single device, leading to underutilization and sub-optimal performance. We propose NanoFlow, a novel serving framework that exploits intra-device parallelism, which overlaps the usage of resources including compute, memory, and network within a single device through operation co-scheduling. To exploit intra-device parallelism, NanoFlow introduces two key innovations: First, NanoFlow splits requests into nano-batches at the granularity of operations, which breaks the dependency of sequential operations in LLM inference and enables overlapping; then, to get benefit from overlapping, NanoFlow uses an operation-level pipeline with execution unit scheduling, which partitions the device's functional units and simultaneously executes different operations in each unit. NanoFlow automates the pipeline setup using a parameter search algorithm, which enables easily porting NanoFlow to different models. We implement NanoFlow on NVIDIA GPUs and evaluate end-to-end serving throughput on several popular models such as LLaMA-2-70B, Mixtral 8x7B, LLaMA-3-8B, etc.. With practical workloads, NanoFlow provides 1.91x throughput boost compared to state-of-the-art serving systems achieving 59% to 72% of optimal throughput across ported models.

We propose ReKV, a novel training-free approach that enables efficient streaming video question-answering (StreamingVQA), by seamlessly integrating with existing Video Large Language Models (Video-LLMs). Traditional VideoQA systems struggle with long videos, as they must process entire videos before responding to queries, and repeat this process for each new question. In contrast, our approach analyzes long videos in a streaming manner, allowing for prompt responses as soon as user queries are received. Building on a common Video-LLM, we first incorporate a sliding-window attention mechanism, ensuring that input frames attend to a limited number of preceding frames, thereby reducing computational overhead. To prevent information loss, we store processed video key-value caches (KV-Caches) in RAM and disk, reloading them into GPU memory as needed. Additionally, we introduce a retrieval method that leverages an external retriever or the parameters within Video-LLMs to retrieve only query-relevant KV-Caches, ensuring both efficiency and accuracy in question answering. ReKV enables the separation of video encoding and question-answering across different processes and GPUs, significantly enhancing the efficiency of StreamingVQA. Through comprehensive experimentation, we validate the efficacy and practicality of our approach, which significantly boosts efficiency and enhances applicability over existing VideoQA models.

Generating consistent long videos is a complex challenge: while diffusion-based generative models generate visually impressive short clips, extending them to longer durations often leads to memory bottlenecks and long-term inconsistency. In this paper, we propose TokensGen, a novel two-stage framework that leverages condensed tokens to address these issues. Our method decomposes long video generation into three core tasks: (1) inner-clip semantic control, (2) long-term consistency control, and (3) inter-clip smooth transition. First, we train To2V (Token-to-Video), a short video diffusion model guided by text and video tokens, with a Video Tokenizer that condenses short clips into semantically rich tokens. Second, we introduce T2To (Text-to-Token), a video token diffusion transformer that generates all tokens at once, ensuring global consistency across clips. Finally, during inference, an adaptive FIFO-Diffusion strategy seamlessly connects adjacent clips, reducing boundary artifacts and enhancing smooth transitions. Experimental results demonstrate that our approach significantly enhances long-term temporal and content coherence without incurring prohibitive computational overhead. By leveraging condensed tokens and pre-trained short video models, our method provides a scalable, modular solution for long video generation, opening new possibilities for storytelling, cinematic production, and immersive simulations. Please see our project page at https://vicky0522.github.io/tokensgen-webpage/ .

Self-attention is an essential component of large language models(LLMs) but a significant source of inference latency for long sequences. In multi-tenant LLMs serving scenarios, the compute and memory operation cost of self-attention can be optimized by using the probability that multiple LLM requests have shared system prompts in prefixes. In this paper, we introduce ChunkAttention, a prefix-aware self-attention module that can detect matching prompt prefixes across multiple requests and share their key/value tensors in memory at runtime to improve the memory utilization of KV cache. This is achieved by breaking monolithic key/value tensors into smaller chunks and structuring them into the auxiliary prefix tree. Consequently, on top of the prefix-tree based KV cache, we design an efficient self-attention kernel, where a two-phase partition algorithm is implemented to improve the data locality during self-attention computation in the presence of shared system prompts. Experiments show that ChunkAttention can speed up the self-attention kernel by 3.2-4.8times compared to the start-of-the-art implementation, with the length of the system prompt ranging from 1024 to 4096.

Google's congestion control (GCC) has become a cornerstone for real-time video and audio communication, yet its performance remains fragile in emerging Low Earth Orbit (LEO) networks. Sparse direct-to-device constellations offer longer duration links and reduced handover frequency compared to dense deployments, presenting a unique opportunity for high-quality real-time communication (RTC) in environments with limited terrestrial network infrastructure. In this paper, we study the behavior of videoconferencing systems in sparse LEO constellations. We observe that video quality degrades due to inherent delays and network instability introduced by the high altitude and rapid movement of LEO satellites, with these effects exacerbated by WebRTC's conventional ``one-size-fits-all'' sender-side pacing queue management. To boost RTC performance, we introduce a data-driven queue management mechanism that adapts the maximum pacing queue capacity based on predicted handover activity. Specifically, our approach employs shorter queue limits during stable, no-handover phases to prioritize low latency communication, and preemptively increases pacing queue capacity when entering periods of increased handover activity to absorb disruptions. Our method yields up to 3x improvements in video bitrate and reduces freeze rate by 62% compared to default WebRTC.

Image tokenization has enabled major advances in autoregressive image generation by providing compressed, discrete representations that are more efficient to process than raw pixels. While traditional approaches use 2D grid tokenization, recent methods like TiTok have shown that 1D tokenization can achieve high generation quality by eliminating grid redundancies. However, these methods typically use a fixed number of tokens and thus cannot adapt to an image's inherent complexity. We introduce FlexTok, a tokenizer that projects 2D images into variable-length, ordered 1D token sequences. For example, a 256x256 image can be resampled into anywhere from 1 to 256 discrete tokens, hierarchically and semantically compressing its information. By training a rectified flow model as the decoder and using nested dropout, FlexTok produces plausible reconstructions regardless of the chosen token sequence length. We evaluate our approach in an autoregressive generation setting using a simple GPT-style Transformer. On ImageNet, this approach achieves an FID<2 across 8 to 128 tokens, outperforming TiTok and matching state-of-the-art methods with far fewer tokens. We further extend the model to support to text-conditioned image generation and examine how FlexTok relates to traditional 2D tokenization. A key finding is that FlexTok enables next-token prediction to describe images in a coarse-to-fine "visual vocabulary", and that the number of tokens to generate depends on the complexity of the generation task.

We present xGen-MM-Vid (BLIP-3-Video): a multimodal language model for videos, particularly designed to efficiently capture temporal information over multiple frames. BLIP-3-Video takes advantage of the 'temporal encoder' in addition to the conventional visual tokenizer, which maps a sequence of tokens over multiple frames into a compact set of visual tokens. This enables BLIP3-Video to use much fewer visual tokens than its competing models (e.g., 32 vs. 4608 tokens). We explore different types of temporal encoders, including learnable spatio-temporal pooling as well as sequential models like Token Turing Machines. We experimentally confirm that BLIP-3-Video obtains video question-answering accuracies comparable to much larger state-of-the-art models (e.g., 34B), while being much smaller (i.e., 4B) and more efficient by using fewer visual tokens. The project website is at https://www.salesforceairesearch.com/opensource/xGen-MM-Vid/index.html

Sparse mixture of experts (SMoE) offers an appealing solution to scale up the model complexity beyond the mean of increasing the network's depth or width. However, we argue that effective SMoE training remains challenging because of the suboptimal routing process where experts that perform computation do not directly contribute to the routing process. In this work, we propose competition, a novel mechanism to route tokens to experts with the highest neural response. Theoretically, we show that the competition mechanism enjoys a better sample efficiency than the traditional softmax routing. Furthermore, we develop CompeteSMoE, a simple yet effective algorithm to train large language models by deploying a router to learn the competition policy, thus enjoying strong performances at a low training overhead. Our extensive empirical evaluations on both the visual instruction tuning and language pre-training tasks demonstrate the efficacy, robustness, and scalability of CompeteSMoE compared to state-of-the-art SMoE strategies. We have made the implementation available at: https://github.com/Fsoft-AIC/CompeteSMoE. This work is an improved version of the previous study at arXiv:2402.02526

Inference accounts for the majority of latency and energy consumption in large language model (LLM) deployments, often exceeding 90% of total cost. While training-time efficiency has seen extensive progress, runtime optimization remains a key bottleneck, particularly under autoregressive decoding. Existing approaches -- such as pruning, quantization, early exits, and speculative decoding -- often require retraining, architectural changes, or disrupt decoding compatibility. We introduce QuickSilver, a modular, token-level framework that enables semantic adaptivity at inference time without altering model weights or structure. QuickSilver integrates four synergistic mechanisms: (i) Dynamic Token Halting, which halts computation for tokens with converged representations; (ii) KV Cache Skipping, which selectively suppresses memory writes to reduce attention overhead; and (iii) Contextual Token Fusion, which collapses redundant tokens into shared paths to shrink sequence length. Unlike speculative decoding or MoE routing, QuickSilver operates entirely on frozen, dense models and requires no auxiliary networks. Applied to GPT-2 and Llama-2 across WikiText-103 and C4, QuickSilver achieves up to 39.6% FLOP reduction with negligible perplexity degradation (<=0.2).

We present Eagle (RWKV-5) and Finch (RWKV-6), sequence models improving upon the RWKV (RWKV-4) architecture. Our architectural design advancements include multi-headed matrix-valued states and a dynamic recurrence mechanism that improve expressivity while maintaining the inference efficiency characteristics of RNNs. We introduce a new multilingual corpus with 1.12 trillion tokens and a fast tokenizer based on greedy matching for enhanced multilinguality. We trained four Eagle models, ranging from 0.46 to 7.5 billion parameters, and two Finch models with 1.6 and 3.1 billion parameters and find that they achieve competitive performance across a wide variety of benchmarks. We release all our models on HuggingFace under the Apache 2.0 license. Models at: https://huggingface.co/RWKV Training code at: https://github.com/RWKV/RWKV-LM Inference code at: https://github.com/RWKV/ChatRWKV Time-parallel training code at: https://github.com/RWKV/RWKV-infctx-trainer

Speculative decoding enhances the efficiency of large language models (LLMs) by leveraging a draft model to draft for a larger target model to review. However, drafting in speculative decoding involves slow autoregressive generation and generating tokens of different importance with the same time allocation. These two inefficiencies lead to its suboptimal performance. To address this issue, we introduce Cascade Speculative Drafting (CS. Drafting), a novel approach that employs two types of cascades. The Vertical Cascade eliminates autoregressive generation from neural models. The Horizontal Cascade constitutes efficient time allocation in drafting with its optimality supported by our theoretical analysis. Combining both cascades, our CS. Drafting algorithm has achieved up to 72 percent additional speedup over speculative decoding in our experiments while keeping the same output distribution.

Large Language Models (LLMs) rely on the Key-Value (KV) Cache to store token history, enabling efficient decoding of tokens. As the KV-Cache grows, it becomes a major memory and computation bottleneck, however, there is an opportunity to alleviate this bottleneck, especially because prior research has shown that only a small subset of tokens contribute meaningfully to each decoding step. A key challenge in finding these critical tokens is that they are dynamic, and heavily input query-dependent. Existing methods either risk quality by evicting tokens permanently, or retain the full KV-Cache but rely on retrieving chunks (pages) of tokens at generation, failing at dense, context-rich tasks. Additionally, many existing KV-Cache sparsity methods rely on inaccurate proxies for token importance. To address these limitations, we introduce TokenButler, a high-granularity, query-aware predictor that learns to identify these critical tokens. By training a light-weight predictor with less than 1.2% parameter overhead, TokenButler prioritizes tokens based on their contextual, predicted importance. This improves perplexity & downstream accuracy by over 8% relative to SoTA methods for estimating token importance. We evaluate TokenButler on a novel synthetic small-context co-referential retrieval task, demonstrating near-oracle accuracy. Code, models and benchmarks: https://github.com/abdelfattah-lab/TokenButler

Existing large language model (LLM) serving systems fall into two categories: 1) a unified system where prefill phase and decode phase are co-located on the same GPU, sharing the unified computational resource and storage, and 2) a disaggregated system where the two phases are disaggregated to different GPUs. The design of the disaggregated system addresses the latency interference and sophisticated scheduling issues in the unified system but leads to storage challenges including 1) replicated weights for both phases that prevent flexible deployment, 2) KV cache transfer overhead between the two phases, 3) storage imbalance that causes substantial wasted space of the GPU capacity, and 4) suboptimal resource adjustment arising from the difficulties in migrating KV cache. Such storage inefficiency delivers poor serving performance under high request rates. In this paper, we identify that the advantage of the disaggregated system lies in the disaggregated computation, i.e., partitioning the computational resource to enable the asynchronous computation of two phases. Thus, we propose a novel LLM serving system, semi-PD, characterized by disaggregated computation and unified storage. In semi-PD, we introduce a computation resource controller to achieve disaggregated computation at the streaming multi-processor (SM) level, and a unified memory manager to manage the asynchronous memory access from both phases. semi-PD has a low-overhead resource adjustment mechanism between the two phases, and a service-level objective (SLO) aware dynamic partitioning algorithm to optimize the SLO attainment. Compared to state-of-the-art systems, semi-PD maintains lower latency at higher request rates, reducing the average end-to-end latency per request by 1.27-2.58x on DeepSeek series models, and serves 1.55-1.72x more requests adhering to latency constraints on Llama series models.

Transformers have shown great potential in computer vision tasks. A common belief is their attention-based token mixer module contributes most to their competence. However, recent works show the attention-based module in Transformers can be replaced by spatial MLPs and the resulted models still perform quite well. Based on this observation, we hypothesize that the general architecture of the Transformers, instead of the specific token mixer module, is more essential to the model's performance. To verify this, we deliberately replace the attention module in Transformers with an embarrassingly simple spatial pooling operator to conduct only basic token mixing. Surprisingly, we observe that the derived model, termed as PoolFormer, achieves competitive performance on multiple computer vision tasks. For example, on ImageNet-1K, PoolFormer achieves 82.1% top-1 accuracy, surpassing well-tuned Vision Transformer/MLP-like baselines DeiT-B/ResMLP-B24 by 0.3%/1.1% accuracy with 35%/52% fewer parameters and 50%/62% fewer MACs. The effectiveness of PoolFormer verifies our hypothesis and urges us to initiate the concept of "MetaFormer", a general architecture abstracted from Transformers without specifying the token mixer. Based on the extensive experiments, we argue that MetaFormer is the key player in achieving superior results for recent Transformer and MLP-like models on vision tasks. This work calls for more future research dedicated to improving MetaFormer instead of focusing on the token mixer modules. Additionally, our proposed PoolFormer could serve as a starting baseline for future MetaFormer architecture design. Code is available at https://github.com/sail-sg/poolformer.

The Key-Value (KV) cache in generative large language models (LLMs) introduces substantial memory overhead. Existing works mitigate this burden by offloading or compressing the KV cache. However, loading the entire cache incurs significant latency due to PCIe bandwidth bottlenecks in CPU-GPU communication, while aggressive compression causes notable performance degradation. We identify that certain layers in the LLM need to maintain global information and are unsuitable for selective loading. In contrast, other layers primarily focus on a few tokens with dominant activations that potentially incur substantial quantization error. This observation leads to a key insight that loading dominant tokens and quantizing all tokens can complement each other. Building on this insight, we propose a hybrid compression method, TailorKV, which seamlessly integrates quantization and offloading. TailorKV develops an inference framework along with a hardware-friendly implementation that leverages these complementary characteristics. Extensive long-context evaluations exhibit that TailorKV achieves nearly lossless performance under aggressive compression settings, outperforming the state-of-the-art. Particularly, the Llama-3.1-8B with 128k context can be served within a single RTX 3090 GPU, reaching 82 ms per token during decoding.

The mixture of Expert (MoE) parallelism is a recent advancement that scales up the model size with constant computational cost. MoE selects different sets of parameters (i.e., experts) for each incoming token, resulting in a sparsely-activated model. Despite several successful applications of MoE, its training efficiency degrades significantly as the number of experts increases. The routing stage in MoE relies on the efficiency of the All2All communication collective, which suffers from network congestion and has poor scalability. To mitigate these issues, we introduce SMILE, which exploits heterogeneous network bandwidth and splits a single-step routing into bi-level routing. Our experimental results show that the proposed method obtains a 2.5x speedup over Switch Transformer in terms of pretraining throughput on the Colossal Clean Crawled Corpus without losing any convergence speed.

We introduce VoiceFilter-Lite, a single-channel source separation model that runs on the device to preserve only the speech signals from a target user, as part of a streaming speech recognition system. Delivering such a model presents numerous challenges: It should improve the performance when the input signal consists of overlapped speech, and must not hurt the speech recognition performance under all other acoustic conditions. Besides, this model must be tiny, fast, and perform inference in a streaming fashion, in order to have minimal impact on CPU, memory, battery and latency. We propose novel techniques to meet these multi-faceted requirements, including using a new asymmetric loss, and adopting adaptive runtime suppression strength. We also show that such a model can be quantized as a 8-bit integer model and run in realtime.

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.

Inference serving for large language models (LLMs) is the key to unleashing their potential in people's daily lives. However, efficient LLM serving remains challenging today because the requests are inherently heterogeneous and unpredictable in terms of resource and latency requirements, as a result of the diverse applications and the dynamic execution nature of LLMs. Existing systems are fundamentally limited in handling these characteristics and cause problems such as severe queuing delays, poor tail latencies, and SLO violations. We introduce Llumnix, an LLM serving system that reacts to such heterogeneous and unpredictable requests by runtime rescheduling across multiple model instances. Similar to context switching across CPU cores in modern operating systems, Llumnix reschedules requests to improve load balancing and isolation, mitigate resource fragmentation, and differentiate request priorities and SLOs. Llumnix implements the rescheduling with an efficient and scalable live migration mechanism for requests and their in-memory states, and exploits it in a dynamic scheduling policy that unifies the multiple rescheduling scenarios elegantly. Our evaluations show that Llumnix improves tail latencies by an order of magnitude, accelerates high-priority requests by up to 1.5x, and delivers up to 36% cost savings while achieving similar tail latencies, compared against state-of-the-art LLM serving systems. Llumnix is publicly available at https://github.com/AlibabaPAI/llumnix.

This paper introduces Interleaved Speech-Text Language Model (IST-LM) for streaming zero-shot Text-to-Speech (TTS). Unlike many previous approaches, IST-LM is directly trained on interleaved sequences of text and speech tokens with a fixed ratio, eliminating the need for additional efforts in duration prediction and grapheme-to-phoneme alignment. The ratio of text chunk size to speech chunk size is crucial for the performance of IST-LM. To explore this, we conducted a comprehensive series of statistical analyses on the training data and performed correlation analysis with the final performance, uncovering several key factors: 1) the distance between speech tokens and their corresponding text tokens, 2) the number of future text tokens accessible to each speech token, and 3) the frequency of speech tokens precedes their corresponding text tokens. Experimental results demonstrate how to achieve an optimal streaming TTS system without complicated engineering optimization, which has a limited gap with the non-streaming system. IST-LM is conceptually simple and empirically powerful, paving the way for streaming TTS with minimal overhead while largely maintaining performance, showcasing broad prospects coupled with real-time text stream from LLMs.

We present significant extensions to diffusion-based sequence generation models, blurring the line with autoregressive language models. We introduce hyperschedules, which assign distinct noise schedules to individual token positions, generalizing both autoregressive models (e.g., GPT) and conventional diffusion models (e.g., SEDD, MDLM) as special cases. Second, we propose two hybrid token-wise noising processes that interpolate between absorbing and uniform processes, enabling the model to fix past mistakes, and we introduce a novel inference algorithm that leverages this new feature in a simplified context inspired from MDLM. To support efficient training and inference, we design attention masks compatible with KV-caching. Our methods achieve state-of-the-art perplexity and generate diverse, high-quality sequences across standard benchmarks, suggesting a promising path for autoregressive diffusion-based sequence generation.

With the increasing scale of models, the need for efficient distributed training has become increasingly urgent. Recently, many synchronous pipeline parallelism approaches have been proposed to improve training throughput. However, these approaches still suffer from two major issues, i.e., pipeline bubbles caused by periodic flushing and extra communication due to the increasing number of pipeline stages. To this end, we propose BitPipe, a bidirectional interleaved pipeline parallelism for accelerating large models training. Specifically, a hybrid scheme of fusing interleaved pipelines with bidirectional pipelines is proposed to reduce the computational time of each single micro-batch and multiply the number of devices executing simultaneously. A V-shaped schedule with eager gradient synchronization is introduced to reduce and overlap the communication between devices. Experiments conducted on up to 32 GPUs show that BitPipe improves the training throughput of GPT-style and BERT-style models by 1.05x-1.28x compared to the state-of-the-art synchronous approaches. The code of our implementation is available at https://github.com/wuhouming/BitPipe.

As a robust and large-scale multilingual speech recognition model, Whisper has demonstrated impressive results in many low-resource and out-of-distribution scenarios. However, its encoder-decoder structure hinders its application to streaming speech recognition. In this paper, we introduce Simul-Whisper, which uses the time alignment embedded in Whisper's cross-attention to guide auto-regressive decoding and achieve chunk-based streaming ASR without any fine-tuning of the pre-trained model. Furthermore, we observe the negative effect of the truncated words at the chunk boundaries on the decoding results and propose an integrate-and-fire-based truncation detection model to address this issue. Experiments on multiple languages and Whisper architectures show that Simul-Whisper achieves an average absolute word error rate degradation of only 1.46% at a chunk size of 1 second, which significantly outperforms the current state-of-the-art baseline.

Parameter-efficient finetuning (PEFT) is a widely used technique to adapt large language models for different tasks. Service providers typically create separate systems for users to perform PEFT model finetuning and inference tasks. This is because existing systems cannot handle workloads that include a mix of inference and PEFT finetuning requests. As a result, shared GPU resources are underutilized, leading to inefficiencies. To address this problem, we present FlexLLM, the first system that can serve inference and parameter-efficient finetuning requests in the same iteration. Our system leverages the complementary nature of these two tasks and utilizes shared GPU resources to run them jointly, using a method called co-serving. To achieve this, FlexLLM introduces a novel token-level finetuning mechanism, which breaks down the finetuning computation of a sequence into smaller token-level computations and uses dependent parallelization and graph pruning, two static compilation optimizations, to minimize the memory overhead and latency for co-serving. Compared to existing systems, FlexLLM's co-serving approach reduces the activation GPU memory overhead by up to 8x, and the end-to-end GPU memory requirement of finetuning by up to 36% while maintaining a low inference latency and improving finetuning throughput. For example, under a heavy inference workload, FlexLLM can still preserve more than 80% of the peak finetuning throughput, whereas existing systems cannot make any progress with finetuning. The source code of FlexLLM is publicly available at https://github.com/flexflow/FlexFlow.

Autoregressive sampling from large language models has led to state-of-the-art results in several natural language tasks. However, autoregressive sampling generates tokens one at a time making it slow, and even prohibitive in certain tasks. One way to speed up sampling is speculative decoding: use a small model to sample a draft (block or sequence of tokens), and then score all tokens in the draft by the large language model in parallel. A subset of the tokens in the draft are accepted (and the rest rejected) based on a statistical method to guarantee that the final output follows the distribution of the large model. In this work, we provide a principled understanding of speculative decoding through the lens of optimal transport (OT) with membership cost. This framework can be viewed as an extension of the well-known maximal-coupling problem. This new formulation enables us to generalize the speculative decoding method to allow for a set of k candidates at the token-level, which leads to an improved optimal membership cost. We show that the optimal draft selection algorithm (transport plan) can be computed via linear programming, whose best-known runtime is exponential in k. We then propose a valid draft selection algorithm whose acceptance probability is (1-1/e)-optimal multiplicatively. Moreover, it can be computed in time almost linear with size of domain of a single token. Using this new draft selection algorithm, we develop a new autoregressive sampling algorithm called SpecTr, which provides speedup in decoding while ensuring that there is no quality degradation in the decoded output. We experimentally demonstrate that for state-of-the-art large language models, the proposed approach achieves a wall clock speedup of 2.13X, a further 1.37X speedup over speculative decoding on standard benchmarks.

Soft attention is a critical mechanism powering LLMs to locate relevant parts within a given context. However, individual attention weights are determined by the similarity of only a single query and key token vector. This "single token attention" bottlenecks the amount of information used in distinguishing a relevant part from the rest of the context. To address this issue, we propose a new attention method, Multi-Token Attention (MTA), which allows LLMs to condition their attention weights on multiple query and key vectors simultaneously. This is achieved by applying convolution operations over queries, keys and heads, allowing nearby queries and keys to affect each other's attention weights for more precise attention. As a result, our method can locate relevant context using richer, more nuanced information that can exceed a single vector's capacity. Through extensive evaluations, we demonstrate that MTA achieves enhanced performance on a range of popular benchmarks. Notably, it outperforms Transformer baseline models on standard language modeling tasks, and on tasks that require searching for information within long contexts, where our method's ability to leverage richer information proves particularly beneficial.

As text generation has become a core capability of modern Large Language Models (LLMs), it underpins a wide range of downstream applications. However, most existing LLMs rely on autoregressive (AR) generation, producing one token at a time based on previously generated context-resulting in limited generation speed due to the inherently sequential nature of the process. To address this challenge, an increasing number of researchers have begun exploring parallel text generation-a broad class of techniques aimed at breaking the token-by-token generation bottleneck and improving inference efficiency. Despite growing interest, there remains a lack of comprehensive analysis on what specific techniques constitute parallel text generation and how they improve inference performance. To bridge this gap, we present a systematic survey of parallel text generation methods. We categorize existing approaches into AR-based and Non-AR-based paradigms, and provide a detailed examination of the core techniques within each category. Following this taxonomy, we assess their theoretical trade-offs in terms of speed, quality, and efficiency, and examine their potential for combination and comparison with alternative acceleration strategies. Finally, based on our findings, we highlight recent advancements, identify open challenges, and outline promising directions for future research in parallel text generation. We have also created a GitHub repository for indexing relevant papers and open resources available at https://github.com/zhanglingzhe0820/Awesome-Parallel-Text-Generation.

This report presents the Prime Collective Communications Library (PCCL), a novel fault-tolerant collective communication library designed for distributed ML workloads over the public internet. PCCL introduces a new programming model that enables dynamic peer joining and failure recovery. The library implements efficient collective operations like all-reduce while providing robust fault tolerance mechanisms that allow the system to continue operating even when peers fail or join during ongoing operations. We demonstrate that PCCL's design enables practical solutions to dynamic membership challenges in workloads with repeated operations and deterministic state advancement. Our implementation passes extensive stress tests across all major operating systems, showing reliable operation even under rapid peer churn and concurrent collective operations. By dispatching to multiple connections, we can efficiently utilize cross-continental long-fat-pipe TCP WAN links, in our experiments achieving up to 45 Gbit/s of bandwidth utilization across Europe and 25 Gbit/s across North America and Europe. PCCL's architecture enables easy implementation of distributed low-communication optimization strategies like DiLoCo, which significantly reduce communication frequency. Combined with quantization, this leads to a significant reduction in the bandwidth required for distributed training workloads. PCCL also allows for concurrent collective operations, which enables optimization strategies like async DiLoCo, which can completely hide communication overhead by implementing one-step delayed parameter updates. PCCL can facilitate exact bit-parity of the shared state across peers in all cases induced by graceful or abrupt peer churn. While PCCL exposes a C99 API, Python bindings are available which are compatible with PyTorch alongside FSDP. PCCL is available under the open source MIT license.

The generation quality of large language models (LLMs) is often improved by utilizing inference-time sequence-level scaling methods (e.g., Chain-of-Thought). We introduce hyper-parallel scaling, a complementary framework that improves prediction quality at the token level. Hyper-parallel scaling computes and aggregates multiple output proposals for a single token from the model. We implement this concept in Mixture-of-Experts (MoE) models, which we refer to as Roster of Experts (RoE). RoE is a training-free inference algorithm that turns a single MoE into a dynamic ensemble of MoEs. RoE injects controlled stochasticity into the expert routing mechanism, enabling it to sample multiple diverse experts for each token and aggregate their outputs for a more accurate final prediction.To overcome the computational cost, we introduce an efficient batching strategy and a specialized KV-caching mechanism that minimizes compute and memory overhead. For example, RoE enables a 7B MoE model to match the performance of a 10.5B MoE model while using 30% less compute for inference. These gains are achieved without any fine-tuning of model parameters.

The Diffusion Transformer plays a pivotal role in advancing text-to-image and text-to-video generation, owing primarily to its inherent scalability. However, existing controlled diffusion transformer methods incur significant parameter and computational overheads and suffer from inefficient resource allocation due to their failure to account for the varying relevance of control information across different transformer layers. To address this, we propose the Relevance-Guided Efficient Controllable Generation framework, RelaCtrl, enabling efficient and resource-optimized integration of control signals into the Diffusion Transformer. First, we evaluate the relevance of each layer in the Diffusion Transformer to the control information by assessing the "ControlNet Relevance Score"-i.e., the impact of skipping each control layer on both the quality of generation and the control effectiveness during inference. Based on the strength of the relevance, we then tailor the positioning, parameter scale, and modeling capacity of the control layers to reduce unnecessary parameters and redundant computations. Additionally, to further improve efficiency, we replace the self-attention and FFN in the commonly used copy block with the carefully designed Two-Dimensional Shuffle Mixer (TDSM), enabling efficient implementation of both the token mixer and channel mixer. Both qualitative and quantitative experimental results demonstrate that our approach achieves superior performance with only 15% of the parameters and computational complexity compared to PixArt-delta. More examples are available at https://relactrl.github.io/RelaCtrl/.

Despite great progress, text-driven long video editing is still notoriously challenging mainly due to excessive memory overhead. Although recent efforts have simplified this task into a two-step process of keyframe translation and interpolation generation, the token-wise keyframe translation still plagues the upper limit of video length. In this paper, we propose a novel and training-free approach towards efficient and effective long video editing, termed AdaFlow. We first reveal that not all tokens of video frames hold equal importance for keyframe translation, based on which we propose an Adaptive Attention Slimming scheme for AdaFlow to squeeze the KV sequence, thus increasing the number of keyframes for translations by an order of magnitude. In addition, an Adaptive Keyframe Selection scheme is also equipped to select the representative frames for joint editing, further improving generation quality. With these innovative designs, AdaFlow achieves high-quality long video editing of minutes in one inference, i.e., more than 1k frames on one A800 GPU, which is about ten times longer than the compared methods, e.g., TokenFlow. To validate AdaFlow, we also build a new benchmark for long video editing with high-quality annotations, termed LongV-EVAL. Our code is released at: https://github.com/jidantang55/AdaFlow.

When the complete source sentence is provided, Large Language Models (LLMs) perform excellently in offline machine translation even with a simple prompt "Translate the following sentence from [src lang] into [tgt lang]:". However, in many real scenarios, the source tokens arrive in a streaming manner and simultaneous machine translation (SiMT) is required, then the efficiency and performance of decoder-only LLMs are significantly limited by their auto-regressive nature. To enable LLMs to achieve high-quality SiMT as efficiently as offline translation, we propose a novel paradigm that includes constructing supervised fine-tuning (SFT) data for SiMT, along with new training and inference strategies. To replicate the token input/output stream in SiMT, the source and target tokens are rearranged into an interleaved sequence, separated by special tokens according to varying latency requirements. This enables powerful LLMs to learn read and write operations adaptively, based on varying latency prompts, while still maintaining efficient auto-regressive decoding. Experimental results show that, even with limited SFT data, our approach achieves state-of-the-art performance across various SiMT benchmarks, and preserves the original abilities of offline translation. Moreover, our approach generalizes well to document-level SiMT setting without requiring specific fine-tuning, even beyond the offline translation model.

Speculative decoding has demonstrated its effectiveness in accelerating the inference of large language models while maintaining a consistent sampling distribution. However, the conventional approach of training a separate draft model to achieve a satisfactory token acceptance rate can be costly. Drawing inspiration from early exiting, we propose a novel self-speculative decoding framework Kangaroo, which uses a fixed shallow sub-network as a self-draft model, with the remaining layers serving as the larger target model. We train a lightweight and efficient adapter module on top of the sub-network to bridge the gap between the sub-network and the full model's representation ability. It is noteworthy that the inference latency of the self-draft model may no longer be negligible compared to the large model, necessitating strategies to increase the token acceptance rate while minimizing the drafting steps of the small model. To address this challenge, we introduce an additional early exiting mechanism for generating draft tokens. Specifically, we halt the small model's subsequent prediction during the drafting phase once the confidence level for the current token falls below a certain threshold. Extensive experiments on the Spec-Bench demonstrate the effectiveness of Kangaroo. Under single-sequence verification, Kangaroo achieves speedups up to 1.68times on Spec-Bench, outperforming Medusa-1 with 88.7\% fewer additional parameters (67M compared to 591M). The code for Kangaroo is available at https://github.com/Equationliu/Kangaroo.

KV cache stores key and value states from previous tokens to avoid re-computation, yet it demands substantial storage space, especially for long sequences. Adaptive KV cache compression seeks to discern the saliency of tokens, preserving vital information while aggressively compressing those of less importance. However, previous methods of this approach exhibit significant performance degradation at high compression ratios due to inaccuracies in identifying salient tokens. In this paper, we present ZipCache, an accurate and efficient KV cache quantization method for LLMs. First, we construct a strong baseline for quantizing KV cache. Through the proposed channel-separable tokenwise quantization scheme, the memory overhead of quantization parameters are substantially reduced compared to fine-grained groupwise quantization. To enhance the compression ratio, we propose normalized attention score as an effective metric for identifying salient tokens by considering the lower triangle characteristics of the attention matrix. Moreover, we develop an efficient approximation method that decouples the saliency metric from full attention scores, enabling compatibility with fast attention implementations like FlashAttention. Extensive experiments demonstrate that ZipCache achieves superior compression ratios, fast generation speed and minimal performance losses compared with previous KV cache compression methods. For instance, when evaluating Mistral-7B model on GSM8k dataset, ZipCache is capable of compressing the KV cache by 4.98times, with only a 0.38% drop in accuracy. In terms of efficiency, ZipCache also showcases a 37.3% reduction in prefill-phase latency, a 56.9% reduction in decoding-phase latency, and a 19.8% reduction in GPU memory usage when evaluating LLaMA3-8B model with a input length of 4096.

Modern RL-based post-training for large language models (LLMs) co-locate trajectory sampling and policy optimisation on the same GPU cluster, forcing the system to switch between inference and training workloads. This serial context switching violates the single-program-multiple-data (SPMD) assumption underlying today's distributed training systems. We present Echo, the RL system that cleanly decouples these two phases across heterogeneous "inference" and "training" swarms while preserving statistical efficiency. Echo introduces two lightweight synchronization protocols: a sequential pull mode that refreshes policy weights according to API call for minimal bias, and an asynchronous push-pull mode that streams version-tagged rollouts through a replay buffer to maximise hardware utilisation. Training four representative RL workloads with Qwen3-4B, Qwen2.5-7B, Qwen3-30B-A3B-Thinking-2507 and Qwen3-32B on a geographically distributed cluster, Echo matches a fully co-located Verl baseline in convergence speed and final reward while off-loading trajectory generation to commodity edge hardware. These promising results demonstrate that large-scale RL for LLMs could achieve datacentre-grade performance using decentralised, heterogeneous resources.

To alleviate the computational burden of large language models (LLMs), architectures with activation sparsity, represented by mixture-of-experts (MoE), have attracted increasing attention. However, the non-differentiable and inflexible routing of vanilla MoE hurts model performance. Moreover, while each token activates only a few parameters, these sparsely-activated architectures exhibit low chunk-level sparsity, indicating that the union of multiple consecutive tokens activates a large ratio of parameters. Such a sparsity pattern is unfriendly for acceleration under low-resource conditions (e.g., end-side devices) and incompatible with mainstream acceleration techniques (e.g., speculative decoding). To address these challenges, we introduce a novel MoE architecture, BlockFFN, as well as its efficient training and deployment techniques. Specifically, we use a router integrating ReLU activation and RMSNorm for differentiable and flexible routing. Next, to promote both token-level sparsity (TLS) and chunk-level sparsity (CLS), CLS-aware training objectives are designed, making BlockFFN more acceleration-friendly. Finally, we implement efficient acceleration kernels, combining activation sparsity and speculative decoding for the first time. The experimental results demonstrate the superior performance of BlockFFN over other MoE baselines, achieving over 80% TLS and 70% 8-token CLS. Our kernels achieve up to 3.67times speedup on real end-side devices than dense models. All codes and checkpoints are available publicly (https://github.com/thunlp/BlockFFN).

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.

Transformers have emerged as the underpinning architecture for Large Language Models (LLMs). In generative language models, the inference process involves two primary phases: prompt processing and token generation. Token generation, which constitutes the majority of the computational workload, primarily entails vector-matrix multiplications and interactions with the Key-Value (KV) Cache. This phase is constrained by memory bandwidth due to the overhead of transferring weights and KV cache values from the memory system to the computing units. This memory bottleneck becomes particularly pronounced in applications that require long-context and extensive text generation, both of which are increasingly crucial for LLMs. This paper introduces "Keyformer", an innovative inference-time approach, to mitigate the challenges associated with KV cache size and memory bandwidth utilization. Keyformer leverages the observation that approximately 90% of the attention weight in generative inference focuses on a specific subset of tokens, referred to as "key" tokens. Keyformer retains only the key tokens in the KV cache by identifying these crucial tokens using a novel score function. This approach effectively reduces both the KV cache size and memory bandwidth usage without compromising model accuracy. We evaluate Keyformer's performance across three foundational models: GPT-J, Cerebras-GPT, and MPT, which employ various positional embedding algorithms. Our assessment encompasses a variety of tasks, with a particular emphasis on summarization and conversation tasks involving extended contexts. Keyformer's reduction of KV cache reduces inference latency by 2.1x and improves token generation throughput by 2.4x, while preserving the model's accuracy.

Recent large language models (LLMs) have tended to leverage sparsity to reduce computations, employing the sparsely activated mixture-of-experts (MoE) technique. MoE introduces four modules, including token routing, token communication, expert computation, and expert parallelism, that impact model quality and training efficiency. To enable versatile usage of MoE models, we introduce FSMoE, a flexible training system optimizing task scheduling with three novel techniques: 1) Unified abstraction and online profiling of MoE modules for task scheduling across various MoE implementations. 2) Co-scheduling intra-node and inter-node communications with computations to minimize communication overheads. 3) To support near-optimal task scheduling, we design an adaptive gradient partitioning method for gradient aggregation and a schedule to adaptively pipeline communications and computations. We conduct extensive experiments with configured MoE layers and real-world MoE models on two GPU clusters. Experimental results show that 1) our FSMoE supports four popular types of MoE routing functions and is more efficient than existing implementations (with up to a 1.42times speedup), and 2) FSMoE outperforms the state-of-the-art MoE training systems (DeepSpeed-MoE and Tutel) by 1.18times-1.22times on 1458 MoE layers and 1.19times-3.01times on real-world MoE models based on GPT-2 and Mixtral using a popular routing function.

Emerging Large Language Model (LLM) system patterns, such as disaggregated inference, Mixture-of-Experts (MoE) routing, and asynchronous reinforcement fine-tuning, require flexible point-to-point communication beyond simple collectives. Existing implementations are locked to specific Network Interface Controllers (NICs), hindering integration into inference engines and portability across hardware providers. We present TransferEngine, which bridges the functionality of common NICs to expose a uniform interface. TransferEngine exposes one-sided WriteImm operations with a ImmCounter primitive for completion notification, without ordering assumptions of network transport, transparently managing multiple NICs per GPU. We demonstrate peak throughput of 400 Gbps on both NVIDIA ConnectX-7 and AWS Elastic Fabric Adapter (EFA). We showcase TransferEngine through three production systems: (1) KvCache transfer for disaggregated inference with dynamic scaling, (2) RL weight updates achieving 1.3 seconds for trillion-parameter models, and (3) MoE dispatch/combine implementation exceeding DeepEP decode latency on ConnectX-7, with the first viable latencies on EFA. We demonstrate that our portable point-to-point communication complements collectives while avoiding lock-in.

Video Diffusion Transformers (DiTs) have demonstrated significant potential for generating high-fidelity videos but are computationally intensive. Existing acceleration methods include distillation, which requires costly retraining, and feature caching, which is highly sensitive to network architecture. Recent token reduction methods are training-free and architecture-agnostic, offering greater flexibility and wider applicability. However, they enforce the same sequence length across different components, constraining their acceleration potential. We observe that intra-sequence redundancy in video DiTs varies across features, blocks, and denoising timesteps. Building on this observation, we propose Asymmetric Reduction and Restoration (AsymRnR), a training-free approach to accelerate video DiTs. It offers a flexible and adaptive strategy that reduces the number of tokens based on their redundancy to enhance both acceleration and generation quality. We further propose matching cache to facilitate faster processing. Integrated into state-of-the-art video DiTs, AsymRnR achieves a superior speedup without compromising the quality.