Song Han

Amala Sanjay Deshmukh, Kateryna Chumachenko, Tuomas Rintamaki et al. · nvidia

We introduce Nemotron Nano V2 VL, the latest model of the Nemotron vision-language series designed for strong real-world document understanding, long video comprehension, and reasoning tasks. Nemotron Nano V2 VL delivers significant improvements over our previous model, Llama-3.1-Nemotron-Nano-VL-8B, across all vision and text domains through major enhancements in model architecture, datasets, and training recipes. Nemotron Nano V2 VL builds on Nemotron Nano V2, a hybrid Mamba-Transformer LLM, and innovative token reduction techniques to achieve higher inference throughput in long document and video scenarios. We are releasing model checkpoints in BF16, FP8, and FP4 formats and sharing large parts of our datasets, recipes and training code.

Aaron Blakeman, Aarti Basant, Abhinav Khattar et al. · nvidia

As inference-time scaling becomes critical for enhanced reasoning capabilities, it is increasingly becoming important to build models that are efficient to infer. We introduce Nemotron-H, a family of 8B and 56B/47B hybrid Mamba-Transformer models designed to reduce inference cost for a given accuracy level. To achieve this goal, we replace the majority of self-attention layers in the common Transformer model architecture with Mamba layers that perform constant computation and require constant memory per generated token. We show that Nemotron-H models offer either better or on-par accuracy compared to other similarly-sized state-of-the-art open-sourced Transformer models (e.g., Qwen-2.5-7B/72B and Llama-3.1-8B/70B), while being up to 3$\times$ faster at inference. To further increase inference speed and reduce the memory required at inference time, we created Nemotron-H-47B-Base from the 56B model using a new compression via pruning and distillation technique called MiniPuzzle. Nemotron-H-47B-Base achieves similar accuracy to the 56B model, but is 20% faster to infer. In addition, we introduce an FP8-based training recipe and show that it can achieve on par results with BF16-based training. This recipe is used to train the 56B model. We are releasing Nemotron-H base model checkpoints with support in Hugging Face and NeMo.

Guangxuan Xiao, Ji Lin, Mickael Seznec et al.

Large language models (LLMs) show excellent performance but are compute- and memory-intensive. Quantization can reduce memory and accelerate inference. However, existing methods cannot maintain accuracy and hardware efficiency at the same time. We propose SmoothQuant, a training-free, accuracy-preserving, and general-purpose post-training quantization (PTQ) solution to enable 8-bit weight, 8-bit activation (W8A8) quantization for LLMs. Based on the fact that weights are easy to quantize while activations are not, SmoothQuant smooths the activation outliers by offline migrating the quantization difficulty from activations to weights with a mathematically equivalent transformation. SmoothQuant enables an INT8 quantization of both weights and activations for all the matrix multiplications in LLMs, including OPT, BLOOM, GLM, MT-NLG, Llama-1/2, Falcon, Mistral, and Mixtral models. We demonstrate up to 1.56x speedup and 2x memory reduction for LLMs with negligible loss in accuracy. SmoothQuant enables serving 530B LLM within a single node. Our work offers a turn-key solution that reduces hardware costs and democratizes LLMs. Code is available at https://github.com/mit-han-lab/smoothquant.

Ji Lin, Ligeng Zhu, Wei-Ming Chen et al. · mit

On-device training enables the model to adapt to new data collected from the sensors by fine-tuning a pre-trained model. Users can benefit from customized AI models without having to transfer the data to the cloud, protecting the privacy. However, the training memory consumption is prohibitive for IoT devices that have tiny memory resources. We propose an algorithm-system co-design framework to make on-device training possible with only 256KB of memory. On-device training faces two unique challenges: (1) the quantized graphs of neural networks are hard to optimize due to low bit-precision and the lack of normalization; (2) the limited hardware resource does not allow full back-propagation. To cope with the optimization difficulty, we propose Quantization-Aware Scaling to calibrate the gradient scales and stabilize 8-bit quantized training. To reduce the memory footprint, we propose Sparse Update to skip the gradient computation of less important layers and sub-tensors. The algorithm innovation is implemented by a lightweight training system, Tiny Training Engine, which prunes the backward computation graph to support sparse updates and offload the runtime auto-differentiation to compile time. Our framework is the first solution to enable tiny on-device training of convolutional neural networks under 256KB SRAM and 1MB Flash without auxiliary memory, using less than 1/1000 of the memory of PyTorch and TensorFlow while matching the accuracy on tinyML application VWW. Our study enables IoT devices not only to perform inference but also to continuously adapt to new data for on-device lifelong learning. A video demo can be found here: https://youtu.be/0pUFZYdoMY8.

Ji Lin, Jiaming Tang, Haotian Tang et al. · mit

Large language models (LLMs) have transformed numerous AI applications. On-device LLM is becoming increasingly important: running LLMs locally on edge devices can reduce the cloud computing cost and protect users' privacy. However, the astronomical model size and the limited hardware resource pose significant deployment challenges. We propose Activation-aware Weight Quantization (AWQ), a hardware-friendly approach for LLM low-bit weight-only quantization. AWQ finds that not all weights in an LLM are equally important. Protecting only 1% salient weights can greatly reduce quantization error. To identify salient weight channels, we should refer to the activation distribution, not weights. To avoid the hardware-inefficient mix-precision quantization, we mathematically derive that scaling up the salient channels can reduce the quantization error. AWQ employs an equivalent transformation to scale the salient weight channels to protect them. The scale is determined by collecting the activation statistics offline. AWQ does not rely on any backpropagation or reconstruction, so it generalizes to different domains and modalities without overfitting the calibration set. AWQ outperforms existing work on various language modeling and domain-specific benchmarks (coding and math). Thanks to better generalization, it achieves excellent quantization performance for instruction-tuned LMs and, for the first time, multi-modal LMs. Alongside AWQ, we implement TinyChat, an efficient and flexible inference framework tailored for 4-bit on-device LLM/VLMs. With kernel fusion and platform-aware weight packing, TinyChat offers more than 3x speedup over the Huggingface FP16 implementation on both desktop and mobile GPUs. It also democratizes the deployment of the 70B Llama-2 model on mobile GPUs.

Zhijian Liu, Haotian Tang, Alexander Amini et al.

Multi-sensor fusion is essential for an accurate and reliable autonomous driving system. Recent approaches are based on point-level fusion: augmenting the LiDAR point cloud with camera features. However, the camera-to-LiDAR projection throws away the semantic density of camera features, hindering the effectiveness of such methods, especially for semantic-oriented tasks (such as 3D scene segmentation). In this paper, we break this deeply-rooted convention with BEVFusion, an efficient and generic multi-task multi-sensor fusion framework. It unifies multi-modal features in the shared bird's-eye view (BEV) representation space, which nicely preserves both geometric and semantic information. To achieve this, we diagnose and lift key efficiency bottlenecks in the view transformation with optimized BEV pooling, reducing latency by more than 40x. BEVFusion is fundamentally task-agnostic and seamlessly supports different 3D perception tasks with almost no architectural changes. It establishes the new state of the art on nuScenes, achieving 1.3% higher mAP and NDS on 3D object detection and 13.6% higher mIoU on BEV map segmentation, with 1.9x lower computation cost. Code to reproduce our results is available at https://github.com/mit-han-lab/bevfusion.

Yaosheng Fu, Guangxuan Xiao, Xin Dong et al.

Sparse attention reduces compute and memory bandwidth for long-context LLM inference. However, two key challenges remain: (1) KV cache capacity still grows with sequence length, and offloading to CPU memory introduces a PCIe transfer bottleneck; (2) the sparse selection step itself retains $O(T^2)$ complexity and can dominate attention cost at long contexts. We propose SparDA, a decoupled sparse attention architecture that introduces a fourth per-layer projection, the Forecast, alongside Query, Key, and Value. The Forecast predicts the KV blocks needed by the next layer, enabling lookahead selection that overlaps CPU-to-GPU prefetch with current-layer execution. Because Forecast is decoupled from the attention query, our GQA implementation uses one Forecast head per GQA group, reducing selection overhead versus the original multi-head selector. SparDA adds $<$0.5% parameters and trains only the Forecast projections by matching the original selector's attention distribution. On two sparse-pretrained 8B models, SparDA matches or slightly improves accuracy and delivers up to 1.25$\times$ prefill speedup and 1.7$\times$ decode speedup over the sparse-attention offload baseline. By enabling larger feasible batch sizes on a single GPU, SparDA further reaches up to 5.3$\times$ higher decode throughput than the non-offload sparse baseline. Our source code is available at https://github.com/NVlabs/SparDA.

Hanrui Wang, Pengyu Liu, Jinglei Cheng et al. · mit

Among different quantum algorithms, PQC for QML show promises on near-term devices. To facilitate the QML and PQC research, a recent python library called TorchQuantum has been released. It can construct, simulate, and train PQC for machine learning tasks with high speed and convenient debugging supports. Besides quantum for ML, we want to raise the community's attention on the reversed direction: ML for quantum. Specifically, the TorchQuantum library also supports using data-driven ML models to solve problems in quantum system research, such as predicting the impact of quantum noise on circuit fidelity and improving the quantum circuit compilation efficiency. This paper presents a case study of the ML for quantum part. Since estimating the noise impact on circuit reliability is an essential step toward understanding and mitigating noise, we propose to leverage classical ML to predict noise impact on circuit fidelity. Inspired by the natural graph representation of quantum circuits, we propose to leverage a graph transformer model to predict the noisy circuit fidelity. We firstly collect a large dataset with a variety of quantum circuits and obtain their fidelity on noisy simulators and real machines. Then we embed each circuit into a graph with gate and noise properties as node features, and adopt a graph transformer to predict the fidelity. Evaluated on 5 thousand random and algorithm circuits, the graph transformer predictor can provide accurate fidelity estimation with RMSE error 0.04 and outperform a simple neural network-based model by 0.02 on average. It can achieve 0.99 and 0.95 R$^2$ scores for random and algorithm circuits, respectively. Compared with circuit simulators, the predictor has over 200X speedup for estimating the fidelity.

Zhijian Liu, Haotian Tang, Shengyu Zhao et al. · mit

3D neural networks are widely used in real-world applications (e.g., AR/VR headsets, self-driving cars). They are required to be fast and accurate; however, limited hardware resources on edge devices make these requirements rather challenging. Previous work processes 3D data using either voxel-based or point-based neural networks, but both types of 3D models are not hardware-efficient due to the large memory footprint and random memory access. In this paper, we study 3D deep learning from the efficiency perspective. We first systematically analyze the bottlenecks of previous 3D methods. We then combine the best from point-based and voxel-based models together and propose a novel hardware-efficient 3D primitive, Point-Voxel Convolution (PVConv). We further enhance this primitive with the sparse convolution to make it more effective in processing large (outdoor) scenes. Based on our designed 3D primitive, we introduce 3D Neural Architecture Search (3D-NAS) to explore the best 3D network architecture given a resource constraint. We evaluate our proposed method on six representative benchmark datasets, achieving state-of-the-art performance with 1.8-23.7x measured speedup. Furthermore, our method has been deployed to the autonomous racing vehicle of MIT Driverless, achieving larger detection range, higher accuracy and lower latency.

Guangxuan Xiao, Yuandong Tian, Beidi Chen et al.

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.

Han Cai, Ji Lin, Yujun Lin et al. · mit

Deep neural networks (DNNs) have achieved unprecedented success in the field of artificial intelligence (AI), including computer vision, natural language processing and speech recognition. However, their superior performance comes at the considerable cost of computational complexity, which greatly hinders their applications in many resource-constrained devices, such as mobile phones and Internet of Things (IoT) devices. Therefore, methods and techniques that are able to lift the efficiency bottleneck while preserving the high accuracy of DNNs are in great demand in order to enable numerous edge AI applications. This paper provides an overview of efficient deep learning methods, systems and applications. We start from introducing popular model compression methods, including pruning, factorization, quantization as well as compact model design. To reduce the large design cost of these manual solutions, we discuss the AutoML framework for each of them, such as neural architecture search (NAS) and automated pruning and quantization. We then cover efficient on-device training to enable user customization based on the local data on mobile devices. Apart from general acceleration techniques, we also showcase several task-specific accelerations for point cloud, video and natural language processing by exploiting their spatial sparsity and temporal/token redundancy. Finally, to support all these algorithmic advancements, we introduce the efficient deep learning system design from both software and hardware perspectives.

Wei Shi, Hanrui Wang, Jiaqi Gu et al. · mit

Analog/mixed-signal circuit design is one of the most complex and time-consuming stages in the whole chip design process. Due to various process, voltage, and temperature (PVT) variations from chip manufacturing, analog circuits inevitably suffer from performance degradation. Although there has been plenty of work on automating analog circuit design under the typical condition, limited research has been done on exploring robust designs under real and unpredictable silicon variations. Automatic analog design against variations requires prohibitive computation and time costs. To address the challenge, we present RobustAnalog, a robust circuit design framework that involves the variation information in the optimization process. Specifically, circuit optimizations under different variations are considered as a set of tasks. Similarities among tasks are leveraged and competitions are alleviated to realize a sample-efficient multi-task training. Moreover, RobustAnalog prunes the task space according to the current performance in each iteration, leading to a further simulation cost reduction. In this way, RobustAnalog can rapidly produce a set of circuit parameters that satisfies diverse constraints (e.g. gain, bandwidth, noise...) across variations. We compare RobustAnalog with Bayesian optimization, Evolutionary algorithm, and Deep Deterministic Policy Gradient (DDPG) and demonstrate that RobustAnalog can significantly reduce required optimization time by 14-30 times. Therefore, our study provides a feasible method to handle various real silicon conditions.

Yukang Chen, Shengju Qian, Haotian Tang et al. · mit

We present LongLoRA, an efficient fine-tuning approach that extends the context sizes of pre-trained large language models (LLMs), with limited computation cost. Typically, training LLMs with long context sizes is computationally expensive, requiring extensive training hours and GPU resources. For example, training on the context length of 8192 needs 16x computational costs in self-attention layers as that of 2048. In this paper, we speed up the context extension of LLMs in two aspects. On the one hand, although dense global attention is needed during inference, fine-tuning the model can be effectively and efficiently done by sparse local attention. The proposed shifted sparse attention effectively enables context extension, leading to non-trivial computation saving with similar performance to fine-tuning with vanilla attention. Particularly, it can be implemented with only two lines of code in training, while being optional in inference. On the other hand, we revisit the parameter-efficient fine-tuning regime for context expansion. Notably, we find that LoRA for context extension works well under the premise of trainable embedding and normalization. LongLoRA combines this improved LoRA with S^2-Attn. LongLoRA demonstrates strong empirical results on various tasks on Llama2 models from 7B/13B to 70B. LongLoRA extends Llama2 7B from 4k context to 100k, or Llama2 70B to 32k on a single 8x A100 machine. LongLoRA extends models' context while retaining their original architectures, and is compatible with most existing techniques, like Flash-Attention2. In addition, we further conduct supervised fine-tuning with LongLoRA and our long instruction-following LongAlpaca dataset.

Guangxuan Xiao, Ji Lin, Song Han

Transfer learning is important for foundation models to adapt to downstream tasks. However, many foundation models are proprietary, so users must share their data with model owners to fine-tune the models, which is costly and raise privacy concerns. Moreover, fine-tuning large foundation models is computation-intensive and impractical for most downstream users. In this paper, we propose Offsite-Tuning, a privacy-preserving and efficient transfer learning framework that can adapt billion-parameter foundation models to downstream data without access to the full model. In offsite-tuning, the model owner sends a light-weight adapter and a lossy compressed emulator to the data owner, who then fine-tunes the adapter on the downstream data with the emulator's assistance. The fine-tuned adapter is then returned to the model owner, who plugs it into the full model to create an adapted foundation model. Offsite-tuning preserves both parties' privacy and is computationally more efficient than the existing fine-tuning methods that require access to the full model weights. We demonstrate the effectiveness of offsite-tuning on various large language and vision foundation models. Offsite-tuning can achieve comparable accuracy as full model fine-tuning while being privacy-preserving and efficient, achieving 6.5x speedup and 5.6x memory reduction. Code is available at https://github.com/mit-han-lab/offsite-tuning.

Aditi, Niket Agarwal, Arslan Ali et al.

We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture. By supporting highly flexible input-output configurations, Cosmos 3 seamlessly unifies critical modalities for Physical AI -- effectively subsuming vision-language models, video generators, world simulators, and world-action models into a single framework. Our evaluation demonstrates that Cosmos 3 establishes a new state-of-the-art across a diverse suite of understanding and generation tasks, demonstrating omnimodal world models as scalable, general-purpose backbones for embodied agents. Our post-trained Cosmos 3 models were ranked as the best open-source Text-to-Image and Image-to-Video models by Artificial Analysis, and the best policy model by RoboArena at the time the technical report was written. To accelerate open research and deployment in Physical AI, we make our code, model checkpoints, curated synthetic datasets, and evaluation benchmark available under the Linux Foundation's OpenMDW-1.1 https://openmdw.ai/license/1-1/ License at https://github.com/nvidia/cosmos}{github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3 . The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos3 .

Qixin Hu, Shuai Yang, Wei Huang et al.

Autoregressive (AR) video diffusion enables variable-length synthesis, but long-horizon generation often suffers from accumulated errors and identity drift. For efficiency, existing methods commonly adopt sliding-window attention during generation. This creates an irreversible generation trajectory: once the active window accumulates appearance errors, subsequent generations can only condition on this degraded trajectory and drift further away. We address this limitation by formulating long video generation as a retrieval-augmented generation (RAG) problem. Rather than relying solely on the recent window, we treat previously generated latents as a dynamic, searchable history. We propose LongLive-RAG, a general retrieval framework for AR video generation. At each new block, LongLive-RAG uses a query embedding to retrieve relevant historical latents. This lightweight retrieval step adds only a small overhead relative to generation and lets the generator condition on non-local context instead of only the recent window. To make retrieval more discriminative, we introduce the Window Temporal Delta Loss that suppresses redundant local similarity and encourages embeddings to capture meaningful temporal changes. Together, these components help reduce error accumulation caused by sliding-window attention. Experiments across multiple AR backbones and generation lengths show improved long-video quality and the best average VBench-Long rank. To our knowledge, among open-ended AR long video generation methods, LongLive-RAG is the first to formulate self-generated latent history as content-addressable retrieval memory. Code is available at https://github.com/qixinhu11/LongLive-RAG.

Yuyang Zhao, Yicheng Pan, Qiyuan He et al.

Real-time streaming video-to-video editing (V2V) is critical for interactive applications such as live broadcasting and gaming, yet it remains a formidable challenge due to the stringent requirements for temporal consistency and inference throughput. In this paper, we present SANA-Streaming, a system-algorithm co-designed framework for high-resolution, real-time streaming video editing on consumer GPUs, with the following three core designs: (1) Hybrid Diffusion Transformer architecture introduces softmax attention in part of the blocks to improve local modeling capabilities while preserving the efficiency of linear layers. (2) Cycle-Reverse Regularization is a novel training strategy that enforces semantic consistency by predicting source frames from generated content via flow matching, improving temporal consistency without requiring paired long edited videos. (3) Efficient System Co-design combines fused GDN kernels and Mixed-Precision Quantization (MPQ) optimized for the NVIDIA Blackwell (RTX 5090) architecture. By profiling real-world throughput, our MPQ maximizes Tensor Core utilization while maintaining generation quality. The resulting system achieves real-time 1280 x 704 resolution editing at 24 end-to-end FPS on a single RTX 5090 GPU, with the DiT core running at 58 FPS. Experimental results demonstrate that our co-design approach significantly outperforms existing SOTA methods in both temporal coherence and system throughput.

Haotian Tang, Zhijian Liu, Xiuyu Li et al. · mit

Deep learning on point clouds has received increased attention thanks to its wide applications in AR/VR and autonomous driving. These applications require low latency and high accuracy to provide real-time user experience and ensure user safety. Unlike conventional dense workloads, the sparse and irregular nature of point clouds poses severe challenges to running sparse CNNs efficiently on the general-purpose hardware. Furthermore, existing sparse acceleration techniques for 2D images do not translate to 3D point clouds. In this paper, we introduce TorchSparse, a high-performance point cloud inference engine that accelerates the sparse convolution computation on GPUs. TorchSparse directly optimizes the two bottlenecks of sparse convolution: irregular computation and data movement. It applies adaptive matrix multiplication grouping to trade computation for better regularity, achieving 1.4-1.5x speedup for matrix multiplication. It also optimizes the data movement by adopting vectorized, quantized and fused locality-aware memory access, reducing the memory movement cost by 2.7x. Evaluated on seven representative models across three benchmark datasets, TorchSparse achieves 1.6x and 1.5x measured end-to-end speedup over the state-of-the-art MinkowskiEngine and SpConv, respectively.

Zhijian Liu, Xinyu Yang, Haotian Tang et al. · mit

Transformer, as an alternative to CNN, has been proven effective in many modalities (e.g., texts and images). For 3D point cloud transformers, existing efforts focus primarily on pushing their accuracy to the state-of-the-art level. However, their latency lags behind sparse convolution-based models (3x slower), hindering their usage in resource-constrained, latency-sensitive applications (such as autonomous driving). This inefficiency comes from point clouds' sparse and irregular nature, whereas transformers are designed for dense, regular workloads. This paper presents FlatFormer to close this latency gap by trading spatial proximity for better computational regularity. We first flatten the point cloud with window-based sorting and partition points into groups of equal sizes rather than windows of equal shapes. This effectively avoids expensive structuring and padding overheads. We then apply self-attention within groups to extract local features, alternate sorting axis to gather features from different directions, and shift windows to exchange features across groups. FlatFormer delivers state-of-the-art accuracy on Waymo Open Dataset with 4.6x speedup over (transformer-based) SST and 1.4x speedup over (sparse convolutional) CenterPoint. This is the first point cloud transformer that achieves real-time performance on edge GPUs and is faster than sparse convolutional methods while achieving on-par or even superior accuracy on large-scale benchmarks.

Zhiding Liang, Jinglei Cheng, Hang Ren et al. · mit

Variational quantum algorithms (VQAs) have demonstrated great potentials in the Noisy Intermediate Scale Quantum (NISQ) era. In the workflow of VQA, the parameters of ansatz are iteratively updated to approximate the desired quantum states. We have seen various efforts to draft better ansatz with less gates. Some works consider the physical meaning of the underlying circuits, while others adopt the ideas of neural architecture search (NAS) for ansatz generator. However, these designs do not exploit the full advantages of VQAs. Because most techniques target gate ansatz, and the parameters are usually rotation angles of the gates. In quantum computers, the gate ansatz will eventually be transformed into control signals such as microwave pulses on superconducting qubits. These control pulses need elaborate calibrations to minimize the errors such as over-rotation and under-rotation. In the case of VQAs, this procedure will introduce redundancy, but the variational properties of VQAs can naturally handle problems of over-rotation and under-rotation by updating the amplitude and frequency parameters. Therefore, we propose NAPA, a native-pulse ansatz generator framework for VQAs. We generate native-pulse ansatz with trainable parameters for amplitudes and frequencies. In our proposed NAPA, we are tuning parametric pulses, which are natively supported on NISQ computers. Given the limited availability of gradient-based optimizers for pulse-level quantum programs, we choose to deploy non-gradient optimizers in our framework. To constrain the number of parameters sent to the optimizer, we adopt a progressive way to generate our native-pulse ansatz. Experiments are conducted on both simulators and quantum devices for Variational Quantum Eigensolver (VQE) tasks to evaluate our methods.

Xuanyao Chen, Zhijian Liu, Haotian Tang et al. · mit

High-resolution images enable neural networks to learn richer visual representations. However, this improved performance comes at the cost of growing computational complexity, hindering their usage in latency-sensitive applications. As not all pixels are equal, skipping computations for less-important regions offers a simple and effective measure to reduce the computation. This, however, is hard to be translated into actual speedup for CNNs since it breaks the regularity of the dense convolution workload. In this paper, we introduce SparseViT that revisits activation sparsity for recent window-based vision transformers (ViTs). As window attentions are naturally batched over blocks, actual speedup with window activation pruning becomes possible: i.e., ~50% latency reduction with 60% sparsity. Different layers should be assigned with different pruning ratios due to their diverse sensitivities and computational costs. We introduce sparsity-aware adaptation and apply the evolutionary search to efficiently find the optimal layerwise sparsity configuration within the vast search space. SparseViT achieves speedups of 1.5x, 1.4x, and 1.3x compared to its dense counterpart in monocular 3D object detection, 2D instance segmentation, and 2D semantic segmentation, respectively, with negligible to no loss of accuracy.

Yihan Wang, Muyang Li, Han Cai et al.

Pose estimation plays a critical role in human-centered vision applications. However, it is difficult to deploy state-of-the-art HRNet-based pose estimation models on resource-constrained edge devices due to the high computational cost (more than 150 GMACs per frame). In this paper, we study efficient architecture design for real-time multi-person pose estimation on edge. We reveal that HRNet's high-resolution branches are redundant for models at the low-computation region via our gradual shrinking experiments. Removing them improves both efficiency and performance. Inspired by this finding, we design LitePose, an efficient single-branch architecture for pose estimation, and introduce two simple approaches to enhance the capacity of LitePose, including Fusion Deconv Head and Large Kernel Convs. Fusion Deconv Head removes the redundancy in high-resolution branches, allowing scale-aware feature fusion with low overhead. Large Kernel Convs significantly improve the model's capacity and receptive field while maintaining a low computational cost. With only 25% computation increment, 7x7 kernels achieve +14.0 mAP better than 3x3 kernels on the CrowdPose dataset. On mobile platforms, LitePose reduces the latency by up to 5.0x without sacrificing performance, compared with prior state-of-the-art efficient pose estimation models, pushing the frontier of real-time multi-person pose estimation on edge. Our code and pre-trained models are released at https://github.com/mit-han-lab/litepose.

Haotian Tang, Shang Yang, Zhijian Liu et al. · mit

Sparse convolution plays a pivotal role in emerging workloads, including point cloud processing in AR/VR, autonomous driving, and graph understanding in recommendation systems. Since the computation pattern is sparse and irregular, specialized high-performance kernels are required. Existing GPU libraries offer two dataflow types for sparse convolution. The gather-GEMM-scatter dataflow is easy to implement but not optimal in performance, while the dataflows with overlapped computation and memory access (e.g.implicit GEMM) are highly performant but have very high engineering costs. In this paper, we introduce TorchSparse++, a new GPU library that achieves the best of both worlds. We create a highly efficient Sparse Kernel Generator that generates performant sparse convolution kernels at less than one-tenth of the engineering cost of the current state-of-the-art system. On top of this, we design the Sparse Autotuner, which extends the design space of existing sparse convolution libraries and searches for the best dataflow configurations for training and inference workloads. Consequently, TorchSparse++ achieves 2.9x, 3.3x, 2.2x and 1.7x measured end-to-end speedup on an NVIDIA A100 GPU over state-of-the-art MinkowskiEngine, SpConv 1.2, TorchSparse and SpConv v2 in inference; and is 1.2-1.3x faster than SpConv v2 in mixed precision training across seven representative autonomous driving benchmarks. It also seamlessly supports graph convolutions, achieving 2.6-7.6x faster inference speed compared with state-of-the-art graph deep learning libraries.

Lvmin Zhang, Shengqu Cai, Muyang Li et al. · stanford

We present PFP, a neural network structure to compress long videos into short contexts, with an explicit pretraining objective to preserve the high-frequency details of single frames at arbitrary temporal positions. The baseline model can compress a 20-second video into a context at about 5k length, where random frames can be retrieved with perceptually preserved appearances. Such pretrained models can be directly fine-tuned as memory encoders for autoregressive video models, enabling long history memory with low context cost and relatively low fidelity loss. We evaluate the framework with ablative settings and discuss the trade-offs of possible neural architecture designs.

Muyang Li, Ji Lin, Chenlin Meng et al.

During image editing, existing deep generative models tend to re-synthesize the entire output from scratch, including the unedited regions. This leads to a significant waste of computation, especially for minor editing operations. In this work, we present Spatially Sparse Inference (SSI), a general-purpose technique that selectively performs computation for edited regions and accelerates various generative models, including both conditional GANs and diffusion models. Our key observation is that users prone to gradually edit the input image. This motivates us to cache and reuse the feature maps of the original image. Given an edited image, we sparsely apply the convolutional filters to the edited regions while reusing the cached features for the unedited areas. Based on our algorithm, we further propose Sparse Incremental Generative Engine (SIGE) to convert the computation reduction to latency reduction on off-the-shelf hardware. With about $1\%$-area edits, SIGE accelerates DDPM by $3.0\times$ on NVIDIA RTX 3090 and $4.6\times$ on Apple M1 Pro GPU, Stable Diffusion by $7.2\times$ on 3090, and GauGAN by $5.6\times$ on 3090 and $5.2\times$ on M1 Pro GPU. Compared to our conference version, we extend SIGE to accommodate attention layers and apply it to Stable Diffusion. Additionally, we offer support for Apple M1 Pro GPU and include more results with large and sequential edits.

Amala Sanjay Deshmukh, Kateryna Chumachenko, Tuomas Rintamaki et al. · amazon-science, nvidia

We introduce Nemotron 3 Nano Omni, the latest model in the Nemotron multimodal series and the first to natively support audio inputs alongside text, images, and video. Nemotron 3 Nano Omni delivers consistent accuracy improvements over its predecessor, Nemotron Nano V2 VL, across all modalities, enabled by advances in architecture, training data and recipes. In particular, Nemotron 3 delivers leading results in real-world document understanding, long audio-video comprehension, and agentic computer use. Built on the highly efficient Nemotron 3 Nano 30B-A3B backbone, Nemotron 3 Nano Omni further incorporates innovative multimodal token-reduction techniques to deliver substantially lower inference latency and higher throughput than other models of similar size. We are releasing model checkpoints in BF16, FP8, and FP4 formats, along with portions of the training data and codebase to facilitate further research and development.

Yukang Chen, Fuzhao Xue, Dacheng Li et al.

Long-context capability is critical for multi-modal foundation models, especially for long video understanding. We introduce LongVILA, a full-stack solution for long-context visual-language models by co-designing the algorithm and system. For model training, we upgrade existing VLMs to support long video understanding by incorporating two additional stages, i.e., long context extension and long video supervised fine-tuning. However, training on long video is computationally and memory intensive. We introduce the long-context Multi-Modal Sequence Parallelism (MM-SP) system that efficiently parallelizes long video training and inference, enabling 2M context length training on 256 GPUs without any gradient checkpointing. LongVILA efficiently extends the number of video frames of VILA from 8 to 2048, achieving 99.8% accuracy in 6,000-frame (more than 1 million tokens) video needle-in-a-haystack. LongVILA-7B demonstrates strong accuracy on 9 popular video benchmarks, e.g. 65.1% VideoMME with subtitle. Besides, MM-SP is 2.1x - 5.7x faster than ring style sequence parallelism and 1.1x - 1.4x faster than Megatron with a hybrid context and tensor parallelism. Moreover, it seamlessly integrates with Hugging Face Transformers.

Han Cai, Junyan Li, Muyan Hu et al.

High-resolution dense prediction enables many appealing real-world applications, such as computational photography, autonomous driving, etc. However, the vast computational cost makes deploying state-of-the-art high-resolution dense prediction models on hardware devices difficult. This work presents EfficientViT, a new family of high-resolution vision models with novel multi-scale linear attention. Unlike prior high-resolution dense prediction models that rely on heavy softmax attention, hardware-inefficient large-kernel convolution, or complicated topology structure to obtain good performances, our multi-scale linear attention achieves the global receptive field and multi-scale learning (two desirable features for high-resolution dense prediction) with only lightweight and hardware-efficient operations. As such, EfficientViT delivers remarkable performance gains over previous state-of-the-art models with significant speedup on diverse hardware platforms, including mobile CPU, edge GPU, and cloud GPU. Without performance loss on Cityscapes, our EfficientViT provides up to 13.9$\times$ and 6.2$\times$ GPU latency reduction over SegFormer and SegNeXt, respectively. For super-resolution, EfficientViT delivers up to 6.4x speedup over Restormer while providing 0.11dB gain in PSNR. For Segment Anything, EfficientViT delivers 48.9x higher throughput on A100 GPU while achieving slightly better zero-shot instance segmentation performance on COCO.

Yecheng Wu, Zhuoyang Zhang, Junyu Chen et al.

VILA-U is a Unified foundation model that integrates Video, Image, Language understanding and generation. Traditional visual language models (VLMs) use separate modules for understanding and generating visual content, which can lead to misalignment and increased complexity. In contrast, VILA-U employs a single autoregressive next-token prediction framework for both tasks, eliminating the need for additional components like diffusion models. This approach not only simplifies the model but also achieves near state-of-the-art performance in visual language understanding and generation. The success of VILA-U is attributed to two main factors: the unified vision tower that aligns discrete visual tokens with textual inputs during pretraining, which enhances visual perception, and autoregressive image generation can achieve similar quality as diffusion models with high-quality dataset. This allows VILA-U to perform comparably to more complex models using a fully token-based autoregressive framework.

Yulu Gan, Ligeng Zhu, Dandan Shan et al.

Motion understanding is fundamental to physical reasoning, enabling models to infer dynamics and predict future states. However, state-of-the-art models still struggle on recent motion benchmarks, primarily due to the scarcity of large-scale, fine-grained motion datasets. Existing motion datasets are often constructed from costly manual annotation, severely limiting scalability. To address this challenge, we introduce FoundationMotion, a fully automated data curation pipeline that constructs large-scale motion datasets. Our approach first detects and tracks objects in videos to extract their trajectories, then leverages these trajectories and video frames with Large Language Models (LLMs) to generate fine-grained captions and diverse question-answer pairs about motion and spatial reasoning. Using datasets produced by this pipeline, we fine-tune open-source models including NVILA-Video-15B and Qwen2.5-7B, achieving substantial improvements in motion understanding without compromising performance on other tasks. Notably, our models outperform strong closed-source baselines like Gemini-2.5 Flash and large open-source models such as Qwen2.5-VL-72B across diverse motion understanding datasets and benchmarks. FoundationMotion thus provides a scalable solution for curating fine-grained motion datasets that enable effective fine-tuning of diverse models to enhance motion understanding and spatial reasoning capabilities.

Ruitao Liu, Qinghao Hu, Alex Hu et al.

Reinforcement learning with verifiable rewards (RLVR) has become a powerful paradigm for improving language models on reasoning-intensive tasks, but its effectiveness is often limited by exploration. For example, models often fail on hard problems, leaving little useful reward signal. External expert traces offer a natural source of guidance, yet they may also expose reward-relevant content along the critical path to the verifier target, such as final answers, intermediate values, executable implementations, or answer-related entities. This content can create an unintended reward hacking channel, allowing the policy to obtain reward by copying the trace rather than learning the underlying reasoning or agentic behavior. Existing guided-RL methods reduce this risk by using partial trajectories, but they mainly control how much expert information is shown heuristically rather than which parts should be hidden. To this end, we propose Semantic Masked Expert Policy Optimization (SMEPO), a fine-grained semantic masking strategy for expert-guided RLVR. Instead of truncating traces coarsely or revealing them unchanged, SMEPO masks reward-relevant semantic spans along the critical path while preserving the expert's decomposition, plan, and procedural structure. This turns hard problems from reasoning from scratch into a fill-in-the-blank process: the policy can follow the expert's problem-solving route, but must still reconstruct the missing values, code, or entities by itself. SMEPO is simple to apply and requires no changes to the reward function or RL objective. Across diverse domains, including math, code, and agentic search, SMEPO improves accuracy by up to 3.2 points over GRPO and reduces training time by up to 4.2x. The code is available at https://github.com/mit-han-lab/SMEPO.

Jiaming Tang, Yufei Sun, Yilong Zhao et al. · mit

Vision-Language-Action models (VLAs) are becoming increasingly capable across diverse robotic tasks. However, their real-world deployment remains slow and inefficient: demonstration videos are often sped up by 5-10x to appear smooth, with noticeable action stalls and delayed reactions to environmental changes. Asynchronous inference offers a promising solution to achieve continuous and low-latency control by enabling robots to execute actions and perform inference simultaneously. However, because the robot and environment continue to evolve during inference, a temporal misalignment arises between the prediction and execution intervals. This leads to significant action instability, while existing methods either degrade accuracy or introduce runtime overhead to mitigate it. We propose VLASH, a general asynchronous inference framework for VLAs that delivers smooth, accurate, and fast reaction control without additional overhead or architectural changes. VLASH estimates the future execution-time state by rolling the robot state forward with the previously generated action chunk, thereby bridging the gap between prediction and execution. Experiments show that VLASH achieves up to 2.03x speedup and reduces reaction latency by up to 17.4x compared to synchronous inference while fully preserving the original accuracy. Moreover, it empowers VLAs to handle fast-reaction, high-precision tasks such as playing ping-pong and playing whack-a-mole, where traditional synchronous inference fails. Code is available at https://github.com/mit-han-lab/vlash

Yunhao Fang, Ligeng Zhu, Yao Lu et al.

While visual language model architectures and training infrastructures advance rapidly, data curation remains under-explored where quantity and quality become a bottleneck. Existing work either crawls extra Internet data with a loose guarantee of quality or distills from black-box proprietary models, e.g., GPT-4V / Gemini that are API frequency and performance bounded. This work enables a VLM to improve itself via data enhancement, exploiting its generative nature. We introduce a simple yet effective VLM augmentation scheme that includes a self-augment step and a specialist-augment step to iteratively improve data quality and hence, model performance. In the self-augment step, the instruction-finetuned VLM recaptions its pretraining caption datasets and then retrains from scratch leveraging refined data. Without any expensive human-in-the-loop annotation, we observe improvements in data quality and downstream accuracy boosts with three self-augmentation rounds -- a viable free lunch to the current VLM training recipe. When self-augmentation saturates, we augment the caption diversity by leveraging specialty skills picked up from instruction finetuning. We finetune VLM specialists from the self-augmented VLM with domain-specific experts, including spatial, grounding, and OCR, to fuse task-aware synthetic data into the pretraining stage. Data quality improvements and hallucination reductions are cross-checked by VLM (GPT-4V, Gemini) and human judges. Combining self-augmentation and specialist-augmented training, VILA$^2$ consistently improves the accuracy on a wide range of benchmarks over the prior art, producing a reusable pretraining dataset that is 300x more cost-efficient than human labeling.

Shuo Yang, Haocheng Xi, Yilong Zhao et al. · tsinghua

$k$-means has historically been positioned primarily as an offline processing primitive, typically used for dataset organization or embedding preprocessing rather than as a first-class component in online systems. In this work, we revisit this classical algorithm under the lens of modern AI system design and enable $k$-means as an online primitive. We point out that existing GPU implementations of $k$-means remain fundamentally bottlenecked by low-level system constraints rather than theoretical algorithmic complexity. Specifically, the assignment stage suffers from a severe IO bottleneck due to the massive explicit materialization of the $N \times K$ distance matrix in High Bandwidth Memory (HBM). Simultaneously, the centroid update stage is heavily penalized by hardware-level atomic write contention caused by irregular, scatter-style token aggregations. To bridge this performance gap, we propose flash-kmeans, an IO-aware and contention-free $k$-means implementation for modern GPU workloads. Flash-kmeans introduces two core kernel-level innovations: (1) FlashAssign, which fuses distance computation with an online argmin to completely bypass intermediate memory materialization; (2) sort-inverse update, which explicitly constructs an inverse mapping to transform high-contention atomic scatters into high-bandwidth, segment-level localized reductions. Furthermore, we integrate algorithm-system co-designs, including chunked-stream overlap and cache-aware compile heuristics, to ensure practical deployability. Extensive evaluations on NVIDIA H200 GPUs demonstrate that flash-kmeans achieves up to 17.9$\times$ end-to-end speedup over best baselines, while outperforming industry-standard libraries like cuML and FAISS by 33$\times$ and over 200$\times$, respectively.

Guangxuan Xiao, Junxian Guo, Kasra Mazaheri et al.

Mixture of Block Attention (MoBA) (Lu et al., 2025) is a promising building block for efficiently processing long contexts in LLMs by enabling queries to sparsely attend to a small subset of key-value blocks, drastically reducing computational cost. However, the design principles governing MoBA's performance are poorly understood, and it lacks an efficient GPU implementation, hindering its practical adoption. In this paper, we first develop a statistical model to analyze MoBA's underlying mechanics. Our model reveals that performance critically depends on the router's ability to accurately distinguish relevant from irrelevant blocks based on query-key affinities. We derive a signal-to-noise ratio that formally connects architectural parameters to this retrieval accuracy. Guided by our analysis, we identify two key pathways for improvement: using smaller block sizes and applying a short convolution on keys to cluster relevant signals, which enhances routing accuracy. While theoretically better, small block sizes are inefficient on GPUs. To bridge this gap, we introduce FlashMoBA, a hardware-aware CUDA kernel that enables efficient MoBA execution even with the small block sizes our theory recommends. We validate our insights by training LLMs from scratch, showing that our improved MoBA models match the performance of dense attention baselines. FlashMoBA achieves up to 14.7x speedup over FlashAttention-2 for small blocks, making our theoretically-grounded improvements practical. Code is available at: https://github.com/mit-han-lab/flash-moba.

Hanrui Wang, Pengyu Liu, Kevin Shao et al.

Quantum computing has the potential to solve problems that are intractable for classical systems, yet the high error rates in contemporary quantum devices often exceed tolerable limits for useful algorithm execution. Quantum Error Correction (QEC) mitigates this by employing redundancy, distributing quantum information across multiple data qubits and utilizing syndrome qubits to monitor their states for errors. The syndromes are subsequently interpreted by a decoding algorithm to identify and correct errors in the data qubits. This task is complex due to the multiplicity of error sources affecting both data and syndrome qubits as well as syndrome extraction operations. Additionally, identical syndromes can emanate from different error sources, necessitating a decoding algorithm that evaluates syndromes collectively. Although machine learning (ML) decoders such as multi-layer perceptrons (MLPs) and convolutional neural networks (CNNs) have been proposed, they often focus on local syndrome regions and require retraining when adjusting for different code distances. We introduce a transformer-based QEC decoder which employs self-attention to achieve a global receptive field across all input syndromes. It incorporates a mixed loss training approach, combining both local physical error and global parity label losses. Moreover, the transformer architecture's inherent adaptability to variable-length inputs allows for efficient transfer learning, enabling the decoder to adapt to varying code distances without retraining. Evaluation on six code distances and ten different error configurations demonstrates that our model consistently outperforms non-ML decoders, such as Union Find (UF) and Minimum Weight Perfect Matching (MWPM), and other ML decoders, thereby achieving best logical error rates. Moreover, the transfer learning can save over 10x of training cost.

Ligeng Zhu, Lanxiang Hu, Ji Lin et al.

On-device learning and efficient fine-tuning enable continuous and privacy-preserving customization (e.g., locally fine-tuning large language models on personalized data). However, existing training frameworks are designed for cloud servers with powerful accelerators (e.g., GPUs, TPUs) and lack the optimizations for learning on the edge, which faces challenges of resource limitations and edge hardware diversity. We introduce PockEngine: a tiny, sparse and efficient engine to enable fine-tuning on various edge devices. PockEngine supports sparse backpropagation: it prunes the backward graph and sparsely updates the model with measured memory saving and latency reduction while maintaining the model quality. Secondly, PockEngine is compilation first: the entire training graph (including forward, backward and optimization steps) is derived at compile-time, which reduces the runtime overhead and brings opportunities for graph transformations. PockEngine also integrates a rich set of training graph optimizations, thus can further accelerate the training cost, including operator reordering and backend switching. PockEngine supports diverse applications, frontends and hardware backends: it flexibly compiles and tunes models defined in PyTorch/TensorFlow/Jax and deploys binaries to mobile CPU/GPU/DSPs. We evaluated PockEngine on both vision models and large language models. PockEngine achieves up to 15 $\times$ speedup over off-the-shelf TensorFlow (Raspberry Pi), 5.6 $\times$ memory saving back-propagation (Jetson AGX Orin). Remarkably, PockEngine enables fine-tuning LLaMav2-7B on NVIDIA Jetson AGX Orin at 550 tokens/s, 7.9$\times$ faster than the PyTorch.

Baifeng Shi, Stephanie Fu, Long Lian et al.

Multi-modal large language models (MLLMs) have advanced general-purpose video understanding but struggle with long, high-resolution videos -- they process every pixel equally in their vision transformers (ViTs) or LLMs despite significant spatiotemporal redundancy. We introduce AutoGaze, a lightweight module that removes redundant patches before processed by a ViT or an MLLM. Trained with next-token prediction and reinforcement learning, AutoGaze autoregressively selects a minimal set of multi-scale patches that can reconstruct the video within a user-specified error threshold, eliminating redundancy while preserving information. Empirically, AutoGaze reduces visual tokens by 4x-100x and accelerates ViTs and MLLMs by up to 19x, enabling scaling MLLMs to 1K-frame 4K-resolution videos and achieving superior results on video benchmarks (e.g., 67.0% on VideoMME). Furthermore, we introduce HLVid: the first high-resolution, long-form video QA benchmark with 5-minute 4K-resolution videos, where an MLLM scaled with AutoGaze improves over the baseline by 10.1% and outperforms the previous best MLLM by 4.5%. Project page: https://autogaze.github.io/.

Jack Cook, Hyemin S. Lee, Kathryn Le et al.

NVFP4 has grown increasingly popular as a 4-bit format for quantizing large language models due to its hardware support and its ability to retain useful information with relatively few bits per parameter. However, the format is not without limitations: recent work has shown that NVFP4 suffers from its error distribution, resulting in large amounts of quantization error on near-maximal values in each group of 16 values. In this work, we leverage this insight to design new Adaptive Block-Scaled Data Types that can adapt to the distribution of their input values. For four-bit quantization, our proposed IF4 (Int/Float 4) data type selects between FP4 and INT4 representations for each group of 16 values, which are then scaled by an E4M3 scale factor as is done with NVFP4. The selected data type is denoted using the scale factor's sign bit, which is currently unused in NVFP4, and we apply the same insight to design formats for other bit-widths, including IF3 and IF6. When used to quantize language models, we find that IF4 outperforms existing 4-bit block-scaled formats, achieving lower loss during quantized training and achieving higher accuracy on many tasks in post-training quantization. We additionally design and evaluate an IF4 Multiply-Accumulate (MAC) unit to demonstrate that IF4 can be implemented efficiently in next-generation hardware accelerators. Our code is available at https://github.com/mit-han-lab/fouroversix.

Tianrui Feng, Zhi Li, Shuo Yang et al.

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

An-Chieh Cheng, Yang Fu, Yatai Ji et al.

We present GR3D, a spatial vision language model equipped with three complementary grounding capabilities--explicit 2D grounding, implicit 2D grounding, and monocular 3D grounding--within a single framework. GR3D introduces an implicit grounding mechanism that identifies entity mentions during generation and inserts the corresponding region tokens into the text stream, allowing the model to reference visual evidence on the fly when producing spatial chain-of-thought responses. In parallel, a region-prompted monocular 3D grounding design predicts 3D bounding boxes in the camera view from grounded region queries, supported by intrinsic-aware normalization and dense geometric supervision. Together, these grounding capabilities enable GR3D to decompose complex spatial understanding problems into grounded 2D perception followed by 3D inference. GR3D achieves consistent improvements across grounded and non-grounded spatial benchmarks, demonstrating grounding as an effective inductive bias for strengthening spatial understanding in VLMs. These grounding capabilities collectively enhance general spatial understanding beyond the grounding task itself.

Hanrui Wang, Yilian Liu, Pengyu Liu et al.

Quantum state preparation, a crucial subroutine in quantum computing, involves generating a target quantum state from initialized qubits. Arbitrary state preparation algorithms can be broadly categorized into arithmetic decomposition (AD) and variational quantum state preparation (VQSP). AD employs a predefined procedure to decompose the target state into a series of gates, whereas VQSP iteratively tunes ansatz parameters to approximate target state. VQSP is particularly apt for Noisy-Intermediate Scale Quantum (NISQ) machines due to its shorter circuits. However, achieving noise-robust parameter optimization still remains challenging. We present RobustState, a novel VQSP training methodology that combines high robustness with high training efficiency. The core idea involves utilizing measurement outcomes from real machines to perform back-propagation through classical simulators, thus incorporating real quantum noise into gradient calculations. RobustState serves as a versatile, plug-and-play technique applicable for training parameters from scratch or fine-tuning existing parameters to enhance fidelity on target machines. It is adaptable to various ansatzes at both gate and pulse levels and can even benefit other variational algorithms, such as variational unitary synthesis. Comprehensive evaluation of RobustState on state preparation tasks for 4 distinct quantum algorithms using 10 real quantum machines demonstrates a coherent error reduction of up to 7.1 $\times$ and state fidelity improvement of up to 96\% and 81\% for 4-Q and 5-Q states, respectively. On average, RobustState improves fidelity by 50\% and 72\% for 4-Q and 5-Q states compared to baseline approaches.

Haoyi Zhu, Haozhe Liu, Yuyang Zhao et al.

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

Boyi Li, Ligeng Zhu, Ran Tian et al.

We propose Wolf, a WOrLd summarization Framework for accurate video captioning. Wolf is an automated captioning framework that adopts a mixture-of-experts approach, leveraging complementary strengths of Vision Language Models (VLMs). By utilizing both image and video models, our framework captures different levels of information and summarizes them efficiently. Our approach can be applied to enhance video understanding, auto-labeling, and captioning. To evaluate caption quality, we introduce CapScore, an LLM-based metric to assess the similarity and quality of generated captions compared to the ground truth captions. We further build four human-annotated datasets in three domains: autonomous driving, general scenes, and robotics, to facilitate comprehensive comparisons. We show that Wolf achieves superior captioning performance compared to state-of-the-art approaches from the research community (VILA1.5, CogAgent) and commercial solutions (Gemini-Pro-1.5, GPT-4V). For instance, in comparison with GPT-4V, Wolf improves CapScore both quality-wise by 55.6% and similarity-wise by 77.4% on challenging driving videos. Finally, we establish a benchmark for video captioning and introduce a leaderboard, aiming to accelerate advancements in video understanding, captioning, and data alignment. Webpage: https://wolfv0.github.io/.

Hanrui Wang, Pengyu Liu, Yilian Liu et al.

Quantum hardware suffers from high error rates and noise, which makes directly running applications on them ineffective. Quantum Error Correction (QEC) is a critical technique towards fault tolerance which encodes the quantum information distributively in multiple data qubits and uses syndrome qubits to check parity. Minimum-Weight-Perfect-Matching (MWPM) is a popular QEC decoder that takes the syndromes as input and finds the matchings between syndromes that infer the errors. However, there are two paramount challenges for MWPM decoders. First, as noise in real quantum systems can drift over time, there is a potential misalignment with the decoding graph's initial weights, leading to a severe performance degradation in the logical error rates. Second, while the MWPM decoder addresses independent errors, it falls short when encountering correlated errors typical on real hardware, such as those in the 2Q depolarizing channel. We propose DGR, an efficient decoding graph edge re-weighting strategy with no quantum overhead. It leverages the insight that the statistics of matchings across decoding iterations offer rich information about errors on real quantum hardware. By counting the occurrences of edges and edge pairs in decoded matchings, we can statistically estimate the up-to-date probabilities of each edge and the correlations between them. The reweighting process includes two vital steps: alignment re-weighting and correlation re-weighting. The former updates the MWPM weights based on statistics to align with actual noise, and the latter adjusts the weight considering edge correlations. Extensive evaluations on surface code and honeycomb code under various settings show that DGR reduces the logical error rate by 3.6x on average-case noise mismatch with exceeding 5000x improvement under worst-case mismatch.

Dongyun Zou, Zhuoyang Zhang, Junyu Chen et al.

We introduce JetViT, a novel family of hybrid-architecture Vision Transformer (ViT) models that match the accuracy of state-of-the-art full-attention vision foundation models while achieving substantially higher inference efficiency on high-resolution images. At the core of our approach is Post-Training Attention Search, a post-training acceleration framework that converts pre-trained full-attention ViTs into efficient hybrid-attention variants by identifying and replacing redundant full-attention blocks with linear or window-attention blocks. By inheriting the MLP and attention weights from the base model, Post-Training Attention Search efficiently explores the architectural design space through three key steps: (1) optimizing the linear-attention block design; (2) finding the best combination of linear-attention and window-attention blocks; and (3) identifying and preserving critical full-attention blocks. We evaluate JetViT on two representative high-resolution vision foundation models, DINOv3 and DepthAnythingV2. On the NVIDIA H100 GPU, JetViT achieves up to 1.79x higher throughput and up to 44.81% lower latency without sacrificing accuracy. We will release our code and accelerated ViT models soon.

Keshawn Smith, Zhili Zhang, H M Sabbir Ahmad et al.

Deep multi-agent reinforcement learning (MARL) has been demonstrated effectively in simulations for multi-robot problems. For autonomous vehicles, the development of vehicle-to-vehicle (V2V) communication technologies provide opportunities to further enhance system safety. However, zero-shot transfer of simulator-trained MARL policies to dynamic hardware systems remains challenging, and how to leverage communication and shared information for MARL has limited demonstrations on hardware. This problem is challenged by discrepancies between simulated and physical states, system state and model uncertainties, practical shared information design, and the need for safety guarantees in both simulation and hardware. This paper designs RSR-RSMARL, a novel Robust and Safe MARL framework that supports Real-Sim-Real (RSR) policy adaptation for multi-agent systems with communication among agents, with both simulation and hardware demonstrations. RSR-RSMARL leverages state (includes shared state information among agents) and action representations considering real system complexities for MARL formulation. The MARL policy is trained with robust MARL algorithm to enable zero-shot transfer to hardware considering the sim-to-real gap. A safety shield module using Control Barrier Functions (CBFs) provides safety guarantee for each individual agent. Experimental results on 1/10th-scale autonomous vehicles with V2V communication demonstrate the ability of RSR-RSMARL framework to enhance driving safety and coordination across multiple configurations. These findings emphasize the importance of jointly designing robust policy representations and modular safety architectures to enable scalable, generalizable RSR transfer in multi-agent autonomy.

Yonggan Fu, Lexington Whalen, Zhifan Ye et al.

Diffusion language models (dLMs) have emerged as a promising paradigm that enables parallel, non-autoregressive generation, but their learning efficiency lags behind that of autoregressive (AR) language models when trained from scratch. To this end, we study AR-to-dLM conversion to transform pretrained AR models into efficient dLMs that excel in speed while preserving AR models' task accuracy. We achieve this by identifying limitations in the attention patterns and objectives of existing AR-to-dLM methods and then proposing principles and methodologies for more effective AR-to-dLM conversion. Specifically, we first systematically compare different attention patterns and find that maintaining pretrained AR weight distributions is critical for effective AR-to-dLM conversion. As such, we introduce a continuous pretraining scheme with a block-wise attention pattern, which remains causal across blocks while enabling bidirectional modeling within each block. We find that this approach can better preserve pretrained AR models' weight distributions than fully bidirectional modeling, in addition to its known benefit of enabling KV caching, and leads to a win-win in accuracy and efficiency. Second, to mitigate the training-test gap in mask token distributions (uniform vs. highly left-to-right), we propose a position-dependent token masking strategy that assigns higher masking probabilities to later tokens during training to better mimic test-time behavior. Leveraging this framework, we conduct extensive studies of dLMs' attention patterns, training dynamics, and other design choices, providing actionable insights into scalable AR-to-dLM conversion. These studies lead to the Efficient-DLM family, which outperforms state-of-the-art AR models and dLMs, e.g., our Efficient-DLM 8B achieves +5.4%/+2.7% higher accuracy with 4.5x/2.7x higher throughput compared to Dream 7B and Qwen3 4B, respectively.

Zhijian Liu, Zhuoyang Zhang, Samir Khaki et al.

Semantic segmentation empowers numerous real-world applications, such as autonomous driving and augmented/mixed reality. These applications often operate on high-resolution images (e.g., 8 megapixels) to capture the fine details. However, this comes at the cost of considerable computational complexity, hindering the deployment in latency-sensitive scenarios. In this paper, we introduce SparseRefine, a novel approach that enhances dense low-resolution predictions with sparse high-resolution refinements. Based on coarse low-resolution outputs, SparseRefine first uses an entropy selector to identify a sparse set of pixels with high entropy. It then employs a sparse feature extractor to efficiently generate the refinements for those pixels of interest. Finally, it leverages a gated ensembler to apply these sparse refinements to the initial coarse predictions. SparseRefine can be seamlessly integrated into any existing semantic segmentation model, regardless of CNN- or ViT-based. SparseRefine achieves significant speedup: 1.5 to 3.7 times when applied to HRNet-W48, SegFormer-B5, Mask2Former-T/L and SegNeXt-L on Cityscapes, with negligible to no loss of accuracy. Our "dense+sparse" paradigm paves the way for efficient high-resolution visual computing.

Yujun Lin, Haotian Tang, Shang Yang et al.

Quantization can accelerate large language model (LLM) inference. Going beyond INT8 quantization, the research community is actively exploring even lower precision, such as INT4. Nonetheless, state-of-the-art INT4 quantization techniques only accelerate low-batch, edge LLM inference, failing to deliver performance gains in large-batch, cloud-based LLM serving. We uncover a critical issue: existing INT4 quantization methods suffer from significant runtime overhead (20-90%) when dequantizing either weights or partial sums on GPUs. To address this challenge, we introduce QoQ, a W4A8KV4 quantization algorithm with 4-bit weight, 8-bit activation, and 4-bit KV cache. QoQ stands for quattuor-octo-quattuor, which represents 4-8-4 in Latin. QoQ is implemented by the QServe inference library that achieves measured speedup. The key insight driving QServe is that the efficiency of LLM serving on GPUs is critically influenced by operations on low-throughput CUDA cores. Building upon this insight, in QoQ algorithm, we introduce progressive quantization that can allow low dequantization overhead in W4A8 GEMM. Additionally, we develop SmoothAttention to effectively mitigate the accuracy degradation incurred by 4-bit KV quantization. In the QServe system, we perform compute-aware weight reordering and take advantage of register-level parallelism to reduce dequantization latency. We also make fused attention memory-bound, harnessing the performance gain brought by KV4 quantization. As a result, QServe improves the maximum achievable serving throughput of Llama-3-8B by 1.2x on A100, 1.4x on L40S; and Qwen1.5-72B by 2.4x on A100, 3.5x on L40S, compared to TensorRT-LLM. Remarkably, QServe on L40S GPU can achieve even higher throughput than TensorRT-LLM on A100. Thus, QServe effectively reduces the dollar cost of LLM serving by 3x. Code is available at https://github.com/mit-han-lab/omniserve.