We introduce SeerAttention-R, a sparse attention framework specifically tailored for the long decoding of reasoning models. Extended from SeerAttention, SeerAttention-R retains the design of learning attention sparsity through a self-distilled gating mechanism, while removing query pooling to accommodate auto-regressive decoding. With a lightweight plug-in gating, SeerAttention-R is flexible and can be easily integrated into existing pretrained model without modifying the original parameters. We demonstrate that SeerAttention-R, trained on just 0.4B tokens, maintains near-lossless reasoning accuracy with 4K token budget in AIME benchmark under large sparse attention block sizes (64/128). Using TileLang, we develop a highly optimized sparse decoding kernel that achieves near-theoretical speedups of up to 9x over FlashAttention-3 on H100 GPU at 90% sparsity. Code is available at: https://github.com/microsoft/SeerAttention.
As integrated circuit scale grows and design complexity rises, effective circuit representation helps support logic synthesis, formal verification, and other automated processes in electronic design automation. And-Inverter Graphs (AIGs), as a compact and canonical structure, are widely adopted for representing Boolean logic in these workflows. However, the increasing complexity and integration density of modern circuits introduce structural heterogeneity and global logic information loss in AIGs, posing significant challenges to accurate circuit modeling. To address these issues, we propose FuncGNN, which integrates hybrid feature aggregation to extract multi-granularity topological patterns, thereby mitigating structural heterogeneity and enhancing logic circuit representations. FuncGNN further introduces gate-aware normalization that adapts to circuit-specific gate distributions, improving robustness to structural heterogeneity. Finally, FuncGNN employs multi-layer integration to merge intermediate features across layers, effectively synthesizing local and global semantic information for comprehensive logic representations. Experimental results on two logic-level analysis tasks (i.e., signal probability prediction and truth-table distance prediction) demonstrate that FuncGNN outperforms existing state-of-the-art methods, achieving improvements of 2.06% and 18.71%, respectively, while reducing training time by approximately 50.6% and GPU memory usage by about 32.8%.
Transformer models struggle with long-context inference due to their quadratic time and linear memory complexity. Recurrent Memory Transformers (RMTs) offer a solution by reducing the asymptotic cost to linear time and constant memory usage. However, their memory update mechanism leads to sequential execution, causing a performance bottleneck. We introduce Diagonal Batching, a scheduling scheme that unlocks parallelism across segments in RMTs while preserving exact recurrence. This approach eliminates the sequential constraint, enabling efficient GPU inference even for single long-context inputs without complex batching and pipelining techniques. Because the technique is purely a run-time computation reordering, existing RMT models adopt it with no retraining. Applied to a LLaMA-1B ARMT model, Diagonal Batching yields a 3.3x speedup over standard full-attention LLaMA-1B and a 1.8x speedup over the sequential RMT implementation on 131,072-token sequences. By removing sequential bottleneck, Diagonal Batching reduces inference cost and latency, thereby strengthening RMTs as a practical solution for real-world, long-context applications.
We introduce MonkeyOCR, a vision-language model for document parsing that advances the state of the art by leveraging a Structure-Recognition-Relation (SRR) triplet paradigm. This design simplifies what would otherwise be a complex multi-tool pipeline (as in MinerU's modular approach) and avoids the inefficiencies of processing full pages with giant end-to-end models (e.g., large multimodal LLMs like Qwen-VL). In SRR, document parsing is abstracted into three fundamental questions - "Where is it?" (structure), "What is it?" (recognition), and "How is it organized?" (relation) - corresponding to layout analysis, content identification, and logical ordering. This focused decomposition balances accuracy and speed: it enables efficient, scalable processing without sacrificing precision. To train and evaluate this approach, we introduce the MonkeyDoc (the most comprehensive document parsing dataset to date), with 3.9 million instances spanning over ten document types in both Chinese and English. Experiments show that MonkeyOCR outperforms MinerU by an average of 5.1%, with particularly notable improvements on challenging content such as formulas (+15.0%) and tables (+8.6%). Remarkably, our 3B-parameter model surpasses much larger and top-performing models, including Qwen2.5-VL (72B) and Gemini 2.5 Pro, achieving state-of-the-art average performance on English document parsing tasks. In addition, MonkeyOCR processes multi-page documents significantly faster (0.84 pages per second compared to 0.65 for MinerU and 0.12 for Qwen2.5-VL-7B). The 3B model can be efficiently deployed for inference on a single NVIDIA 3090 GPU. Code and models will be released at https://github.com/Yuliang-Liu/MonkeyOCR.
Vision-language models (VLMs) pretrained on large-scale multimodal datasets encode rich visual and linguistic knowledge, making them a strong foundation for robotics. Rather than training robotic policies from scratch, recent approaches adapt VLMs into vision-language-action (VLA) models that enable natural language-driven perception and control. However, existing VLAs are typically massive--often with billions of parameters--leading to high training costs and limited real-world deployability. Moreover, they rely on academic and industrial datasets, overlooking the growing availability of community-collected data from affordable robotic platforms. In this work, we present SmolVLA, a small, efficient, and community-driven VLA that drastically reduces both training and inference costs, while retaining competitive performance. SmolVLA is designed to be trained on a single GPU and deployed on consumer-grade GPUs or even CPUs. To further improve responsiveness, we introduce an asynchronous inference stack decoupling perception and action prediction from action execution, allowing higher control rates with chunked action generation. Despite its compact size, SmolVLA achieves performance comparable to VLAs that are 10x larger. We evaluate SmolVLA on a range of both simulated as well as real-world robotic benchmarks and release all code, pretrained models, and training data.
Efficient inference of Multi-Head Latent Attention (MLA) is challenged by deploying the DeepSeek-R1 671B model on a single Multi-GPU server. This paper introduces FlashMLA-ETAP, a novel framework that enhances MLA inference for the single-instance deployment scenario on NVIDIA H20 GPUs. We propose the Efficient Transpose Attention Pipeline (ETAP), which reconfigures attention computation through transposition to align the KV context length with the \(M\)-dimension in WGMMA operations, significantly reducing redundant computations. FlashMLA-ETAP achieves a 2.78x speedup over FlashMLA at 64K sequence length (batch size 16), with 5.24x and 4.94x improvements over FlashAttention-3 and FlashInfer, respectively, while maintaining numerical stability with a 15.2x lower RMSE (\(1.25 \times 10^{-5}\)) than FlashAttention-3. Furthermore, ETAP's design enables seamless integration into frameworks like FlashAttention-3 and FlashInfer, supported by a detailed theoretical analysis. Our work addresses a critical gap in resource-constrained inference, offering a scalable solution for mid-tier GPUs and paving the way for broader adoption in hardware-aware optimization. Code is available at https://github.com/pengcuo/FlashMLA-ETAP.
The computational sparsity of Mixture-of-Experts (MoE) models enables sub-linear growth in compute cost as model size increases, thus offering a scalable path to training massive neural networks. However, existing implementations suffer from \emph{low GPU utilization}, \emph{significant latency overhead}, and a fundamental \emph{inability to leverage task locality}, primarily due to CPU-managed scheduling, host-initiated communication, and frequent kernel launches. To overcome these limitations, we develop FlashDMoE, a fully GPU-resident MoE operator that fuses expert computation and inter-GPU communication into a \emph{single persistent GPU kernel}. FlashDMoE enables fine-grained pipelining of dispatch, compute, and combine phases, eliminating launch overheads and reducing idle gaps. Unlike existing work, FlashDMoE obviates bulk-synchronous collectives for one-sided, device-initiated, inter-GPU (R)DMA transfers, thus unlocking \emph{payload efficiency}, where we eliminate bloated or redundant network payloads in sparsely activated layers. When evaluated on a single 8-H100 GPU node with MoE models having up to 128 experts and 16K token sequences, FlashDMoE achieves up to \textbf{9}$\times$ higher GPU utilization, \textbf{6}$\times$ lower latency, \textbf{5.7}$\times$ higher throughput, and \textbf{4}$\times$ better overlap efficiency compared to state-of-the-art baselines, despite using FP32 while baselines use FP16. FlashDMoE demonstrates that principled GPU kernel-hardware co-design is key to unlocking the performance ceiling of large-scale distributed ML workloads.
LiDAR semantic segmentation plays a vital role in autonomous driving. Existing voxel-based methods for LiDAR semantic segmentation apply uniform partition to the 3D LiDAR point cloud to form a structured representation based on cartesian/cylindrical coordinates. Although these methods show impressive performance, the drawback of existing voxel-based methods remains in two aspects: (1) it requires a large enough input voxel resolution, which brings a large amount of computation cost and memory consumption. (2) it does not well handle the unbalanced point distribution of LiDAR point cloud. In this paper, we propose a non-uniform cylindrical partition network named NUC-Net to tackle the above challenges. Specifically, we propose the Arithmetic Progression of Interval (API) method to non-uniformly partition the radial axis and generate the voxel representation which is representative and efficient. Moreover, we propose a non-uniform multi-scale aggregation method to improve contextual information. Our method achieves state-of-the-art performance on SemanticKITTI and nuScenes datasets with much faster speed and much less training time. And our method can be a general component for LiDAR semantic segmentation, which significantly improves both the accuracy and efficiency of the uniform counterpart by $4 \times$ training faster and $2 \times$ GPU memory reduction and $3 \times$ inference speedup. We further provide theoretical analysis towards understanding why NUC is effective and how point distribution affects performance. Code is available at \href{https://github.com/alanWXZ/NUC-Net}{https://github.com/alanWXZ/NUC-Net}.
Reinforcement learning (RL) has become a dominant paradigm for training large language models (LLMs), particularly for reasoning tasks. Effective RL for LLMs requires massive parallelization and poses an urgent need for efficient training systems. Most existing large-scale RL systems for LLMs are synchronous, alternating generation and training in a batch setting where rollouts in each training batch are generated by the same model. This approach stabilizes RL training but suffers from severe system-level inefficiency: generation must wait until the longest output in the batch is completed before model updates, resulting in GPU underutilization. We present AReaL, a fully asynchronous RL system that completely decouples generation from training. Rollout workers in AReaL continuously generate new outputs without waiting, while training workers update the model whenever a batch of data is collected. AReaL also incorporates a collection of system-level optimizations, leading to substantially higher GPU utilization. To stabilize RL training, AReaL balances the workload of rollout and training workers to control data staleness, and adopts a staleness-enhanced PPO variant to better handle outdated training samples. Extensive experiments on math and code reasoning benchmarks show that AReaL achieves up to 2.77$\times$ training speedup compared to synchronous systems with the same number of GPUs and matched or improved final performance. The code of AReaL is available at https://github.com/inclusionAI/AReaL/.
Speculative decoding and quantization effectively accelerate memory-bound inference of large language models. Speculative decoding mitigates the memory bandwidth bottleneck by verifying multiple tokens within a single forward pass, which increases computational effort. Quantization achieves this optimization by compressing weights and activations into lower bit-widths and also reduces computations via low-bit matrix multiplications. To further leverage their strengths, we investigate the integration of these two techniques. Surprisingly, experiments applying the advanced speculative decoding method EAGLE-2 to various quantized models reveal that the memory benefits from 4-bit weight quantization are diminished by the computational load from speculative decoding. Specifically, verifying a tree-style draft incurs significantly more time overhead than a single-token forward pass on 4-bit weight quantized models. This finding led to our new speculative decoding design: a hierarchical framework that employs a small model as an intermediate stage to turn tree-style drafts into sequence drafts, leveraging the memory access benefits of the target quantized model. Experimental results show that our hierarchical approach achieves a 2.78$\times$ speedup across various tasks for the 4-bit weight Llama-3-70B model on an A100 GPU, outperforming EAGLE-2 by 1.31$\times$. Code available at https://github.com/AI9Stars/SpecMQuant.
Diffusion Transformer (DiT)-based video diffusion models generate high-quality videos at scale but incur prohibitive processing latency and memory costs for long videos. To address this, we propose a novel distributed inference strategy, termed DualParal. The core idea is that, instead of generating an entire video on a single GPU, we parallelize both temporal frames and model layers across GPUs. However, a naive implementation of this division faces a key limitation: since diffusion models require synchronized noise levels across frames, this implementation leads to the serialization of original parallelisms. We leverage a block-wise denoising scheme to handle this. Namely, we process a sequence of frame blocks through the pipeline with progressively decreasing noise levels. Each GPU handles a specific block and layer subset while passing previous results to the next GPU, enabling asynchronous computation and communication. To further optimize performance, we incorporate two key enhancements. Firstly, a feature cache is implemented on each GPU to store and reuse features from the prior block as context, minimizing inter-GPU communication and redundant computation. Secondly, we employ a coordinated noise initialization strategy, ensuring globally consistent temporal dynamics by sharing initial noise patterns across GPUs without extra resource costs. Together, these enable fast, artifact-free, and infinitely long video generation. Applied to the latest diffusion transformer video generator, our method efficiently produces 1,025-frame videos with up to 6.54$\times$ lower latency and 1.48$\times$ lower memory cost on 8$\times$RTX 4090 GPUs.
Recent advances in 3D Gaussian Splatting have shown remarkable potential for novel view synthesis. However, most existing large-scale scene reconstruction methods rely on the divide-and-conquer paradigm, which often leads to the loss of global scene information and requires complex parameter tuning due to scene partitioning and local optimization. To address these limitations, we propose MixGS, a novel holistic optimization framework for large-scale 3D scene reconstruction. MixGS models the entire scene holistically by integrating camera pose and Gaussian attributes into a view-aware representation, which is decoded into fine-detailed Gaussians. Furthermore, a novel mixing operation combines decoded and original Gaussians to jointly preserve global coherence and local fidelity. Extensive experiments on large-scale scenes demonstrate that MixGS achieves state-of-the-art rendering quality and competitive speed, while significantly reducing computational requirements, enabling large-scale scene reconstruction training on a single 24GB VRAM GPU. The code will be released at https://github.com/azhuantou/MixGS.
Diffusion Transformers (DiT) have become the de-facto model for generating high-quality visual content like videos and images. A huge bottleneck is the attention mechanism where complexity scales quadratically with resolution and video length. One logical way to lessen this burden is sparse attention, where only a subset of tokens or patches are included in the calculation. However, existing techniques fail to preserve visual quality at extremely high sparsity levels and might even incur non-negligible compute overheads. % To address this concern, we propose Re-ttention, which implements very high sparse attention for visual generation models by leveraging the temporal redundancy of Diffusion Models to overcome the probabilistic normalization shift within the attention mechanism. Specifically, Re-ttention reshapes attention scores based on the prior softmax distribution history in order to preserve the visual quality of the full quadratic attention at very high sparsity levels. % Experimental results on T2V/T2I models such as CogVideoX and the PixArt DiTs demonstrate that Re-ttention requires as few as 3.1\% of the tokens during inference, outperforming contemporary methods like FastDiTAttn, Sparse VideoGen and MInference. Further, we measure latency to show that our method can attain over 45\% end-to-end % and over 92\% self-attention latency reduction on an H100 GPU at negligible overhead cost. Code available online here: \href{https://github.com/cccrrrccc/Re-ttention}{https://github.com/cccrrrccc/Re-ttention}
Large language models (LLMs) have revolutionized natural language processing and are increasingly applied to other sequential data types, including genetic sequences. However, adapting LLMs to genomics presents significant challenges. Capturing complex genomic interactions requires modeling long-range dependencies within DNA sequences, where interactions often span over 10,000 base pairs, even within a single gene, posing substantial computational burdens under conventional model architectures and training paradigms. Moreover, standard LLM training approaches are suboptimal for DNA: autoregressive training, while efficient, supports only unidirectional understanding. However, DNA is inherently bidirectional, e.g., bidirectional promoters regulate transcription in both directions and account for nearly 11% of human gene expression. Masked language models (MLMs) allow bidirectional understanding but are inefficient, as only masked tokens contribute to the loss per step. To address these limitations, we introduce JanusDNA, the first bidirectional DNA foundation model built upon a novel pretraining paradigm that combines the optimization efficiency of autoregressive modeling with the bidirectional comprehension of masked modeling. JanusDNA adopts a hybrid Mamba, Attention and Mixture of Experts (MoE) architecture, combining long-range modeling of Attention with efficient sequential learning of Mamba. MoE layers further scale model capacity via sparse activation while keeping computational cost low. Notably, JanusDNA processes up to 1 million base pairs at single nucleotide resolution on a single 80GB GPU. Extensive experiments and ablations show JanusDNA achieves new SOTA results on three genomic representation benchmarks, outperforming models with 250x more activated parameters. Code: https://github.com/Qihao-Duan/JanusDNA
Distributed machine learning workloads use data and tensor parallelism for training and inference, both of which rely on the AllReduce collective to synchronize gradients or activations. However, bulk-synchronous AllReduce algorithms can be delayed by a persistent straggler that is slower to reach the synchronization barrier required to begin the collective. To address this challenge, we propose StragglAR: an AllReduce algorithm that accelerates distributed training and inference in the presence of persistent stragglers. StragglAR implements a ReduceScatter among the remaining GPUs during the straggler-induced delay, and then executes a novel collective algorithm to complete the AllReduce once the straggler reaches the synchronization barrier. StragglAR achieves a 2x theoretical speedup over popular bandwidth-efficient AllReduce algorithms (e.g., Ring) for large GPU clusters with persistent stragglers. On an 8-GPU server, our implementation of StragglAR yields a 22% speedup over state-of-the-art AllReduce algorithms.
Feed-forward 3D Gaussian Splatting (3DGS) models have recently emerged as a promising solution for novel view synthesis, enabling one-pass inference without the need for per-scene 3DGS optimization. However, their scalability is fundamentally constrained by the limited capacity of their encoders, leading to degraded performance or excessive memory consumption as the number of input views increases. In this work, we analyze feed-forward 3DGS frameworks through the lens of the Information Bottleneck principle and introduce ZPressor, a lightweight architecture-agnostic module that enables efficient compression of multi-view inputs into a compact latent state $Z$ that retains essential scene information while discarding redundancy. Concretely, ZPressor enables existing feed-forward 3DGS models to scale to over 100 input views at 480P resolution on an 80GB GPU, by partitioning the views into anchor and support sets and using cross attention to compress the information from the support views into anchor views, forming the compressed latent state $Z$. We show that integrating ZPressor into several state-of-the-art feed-forward 3DGS models consistently improves performance under moderate input views and enhances robustness under dense view settings on two large-scale benchmarks DL3DV-10K and RealEstate10K. The video results, code and trained models are available on our project page: https://lhmd.top/zpressor.
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.
In the past decade, image foundation models (IFMs) have achieved unprecedented progress. However, the potential of directly using IFMs for video self-supervised representation learning has largely been overlooked. In this study, we propose an advancing video self-supervised learning (AdViSe) approach, aimed at significantly reducing the training overhead of video representation models using pre-trained IFMs. Specifically, we first introduce temporal modeling modules (ResNet3D) to IFMs, constructing a video representation model. We then employ a video self-supervised learning approach, playback rate perception, to train temporal modules while freezing the IFM components. Experiments on UCF101 demonstrate that AdViSe achieves performance comparable to state-of-the-art methods while reducing training time by $3.4\times$ and GPU memory usage by $8.2\times$. This study offers fresh insights into low-cost video self-supervised learning based on pre-trained IFMs. Code is available at https://github.com/JingwWu/advise-video-ssl.
Recent high-capacity vision-language-action (VLA) models have demonstrated impressive performance on a range of robotic manipulation tasks by imitating human demonstrations. However, exploiting offline data with limited visited states will cause execution failure in out-of-distribution scenarios. Intuitively, an exploration-based method that improves on online collected data at test time could address this limitation. We present VLA-RL, an algorithmic and systematic framework that leverages online reinforcement learning (RL) to improve pretrained auto-regressive VLAs in downstream tasks. Within a unified perspective, we first introduce a trajectory-level RL formulation for auto-regressive VLA training, which models general robotic manipulation trajectory as multi-modal multi-turn conversation. To address the challenge of sparse rewards, we fine-tune a pretrained vision-language model as a robotic process reward model, which is trained on pseudo reward labels annotated on automatically extracted task segments. To scale up, we identify several implementation findings that improve the stability and efficiency including curriculum selection strategy, GPU-balanced vectorized environments, batch decoding, and critic warmup. VLA-RL enables OpenVLA-7B to surpass the strongest finetuned baseline by 4.5% on 40 challenging robotic manipulation tasks in LIBERO, and even matches the performance of advanced commercial models such as $\pi_0$-FAST. Notably, we observe that VLA-RL benefits from increased test-time optimization, indicating an early spark of inference scaling laws in robotics.
We present the first deep-learning solver for backward stochastic Volterra integral equations (BSVIEs) and their fully-coupled forward-backward variants. The method trains a neural network to approximate the two solution fields in a single stage, avoiding the use of nested time-stepping cycles that limit classical algorithms. For the decoupled case we prove a non-asymptotic error bound composed of an a posteriori residual plus the familiar square root dependence on the time step. Numerical experiments confirm this rate and reveal two key properties: \emph{scalability}, in the sense that accuracy remains stable from low dimension up to 500 spatial variables while GPU batching keeps wall-clock time nearly constant; and \emph{generality}, since the same method handles coupled systems whose forward dynamics depend on the backward solution. These results open practical access to a family of high-dimensional, path-dependent problems in stochastic control and quantitative finance.