Qinghao Hu

Qiang Chen, Qiman Wu, Jian Wang et al.

While local-window self-attention performs notably in vision tasks, it suffers from limited receptive field and weak modeling capability issues. This is mainly because it performs self-attention within non-overlapped windows and shares weights on the channel dimension. We propose MixFormer to find a solution. First, we combine local-window self-attention with depth-wise convolution in a parallel design, modeling cross-window connections to enlarge the receptive fields. Second, we propose bi-directional interactions across branches to provide complementary clues in the channel and spatial dimensions. These two designs are integrated to achieve efficient feature mixing among windows and dimensions. Our MixFormer provides competitive results on image classification with EfficientNet and shows better results than RegNet and Swin Transformer. Performance in downstream tasks outperforms its alternatives by significant margins with less computational costs in 5 dense prediction tasks on MS COCO, ADE20k, and LVIS. Code is available at \url{https://github.com/PaddlePaddle/PaddleClas}.

Qinghao Hu, Gang Li, Qiman Wu et al.

Recently low-precision deep learning accelerators (DLAs) have become popular due to their advantages in chip area and energy consumption, yet the low-precision quantized models on these DLAs bring in severe accuracy degradation. One way to achieve both high accuracy and efficient inference is to deploy high-precision neural networks on low-precision DLAs, which is rarely studied. In this paper, we propose the PArallel Low-precision Quantization (PalQuant) method that approximates high-precision computations via learning parallel low-precision representations from scratch. In addition, we present a novel cyclic shuffle module to boost the cross-group information communication between parallel low-precision groups. Extensive experiments demonstrate that PalQuant has superior performance to state-of-the-art quantization methods in both accuracy and inference speed, e.g., for ResNet-18 network quantization, PalQuant can obtain 0.52\% higher accuracy and 1.78$\times$ speedup simultaneously over their 4-bit counter-part on a state-of-the-art 2-bit accelerator. Code is available at \url{https://github.com/huqinghao/PalQuant}.

Wei Gao, Qinghao Hu, Zhisheng Ye et al.

Deep learning (DL) shows its prosperity in a wide variety of fields. The development of a DL model is a time-consuming and resource-intensive procedure. Hence, dedicated GPU accelerators have been collectively constructed into a GPU datacenter. An efficient scheduler design for such GPU datacenter is crucially important to reduce the operational cost and improve resource utilization. However, traditional approaches designed for big data or high performance computing workloads can not support DL workloads to fully utilize the GPU resources. Recently, substantial schedulers are proposed to tailor for DL workloads in GPU datacenters. This paper surveys existing research efforts for both training and inference workloads. We primarily present how existing schedulers facilitate the respective workloads from the scheduling objectives and resource consumption features. Finally, we prospect several promising future research directions. More detailed summary with the surveyed paper and code links can be found at our project website: https://github.com/S-Lab-System-Group/Awesome-DL-Scheduling-Papers

Xiangyu Chen, Qinghao Hu, Kaidong Li et al.

Vision Transformers has demonstrated competitive performance on computer vision tasks benefiting from their ability to capture long-range dependencies with multi-head self-attention modules and multi-layer perceptron. However, calculating global attention brings another disadvantage compared with convolutional neural networks, i.e. requiring much more data and computations to converge, which makes it difficult to generalize well on small datasets, which is common in practical applications. Previous works are either focusing on transferring knowledge from large datasets or adjusting the structure for small datasets. After carefully examining the self-attention modules, we discover that the number of trivial attention weights is far greater than the important ones and the accumulated trivial weights are dominating the attention in Vision Transformers due to their large quantity, which is not handled by the attention itself. This will cover useful non-trivial attention and harm the performance when trivial attention includes more noise, e.g. in shallow layers for some backbones. To solve this issue, we proposed to divide attention weights into trivial and non-trivial ones by thresholds, then Suppressing Accumulated Trivial Attention (SATA) weights by proposed Trivial WeIghts Suppression Transformation (TWIST) to reduce attention noise. Extensive experiments on CIFAR-100 and Tiny-ImageNet datasets show that our suppressing method boosts the accuracy of Vision Transformers by up to 2.3%. Code is available at https://github.com/xiangyu8/SATA.

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.

Yiqun Chen, Qiang Chen, Qinghao Hu et al.

Fully convolutional detectors discard the one-to-many assignment and adopt a one-to-one assigning strategy to achieve end-to-end detection but suffer from the slow convergence issue. In this paper, we revisit these two assignment methods and find that bringing one-to-many assignment back to end-to-end fully convolutional detectors helps with model convergence. Based on this observation, we propose {\em \textbf{D}ual \textbf{A}ssignment} for end-to-end fully convolutional de\textbf{TE}ction (DATE). Our method constructs two branches with one-to-many and one-to-one assignment during training and speeds up the convergence of the one-to-one assignment branch by providing more supervision signals. DATE only uses the branch with the one-to-one matching strategy for model inference, which doesn't bring inference overhead. Experimental results show that Dual Assignment gives nontrivial improvements and speeds up model convergence upon OneNet and DeFCN. Code: https://github.com/YiqunChen1999/date.

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.

Weixiang Xu, Xiangyu He, Tianli Zhao et al.

Large neural networks are difficult to deploy on mobile devices because of intensive computation and storage. To alleviate it, we study ternarization, a balance between efficiency and accuracy that quantizes both weights and activations into ternary values. In previous ternarized neural networks, a hard threshold Δ is introduced to determine quantization intervals. Although the selection of Δ greatly affects the training results, previous works estimate Δ via an approximation or treat it as a hyper-parameter, which is suboptimal. In this paper, we present the Soft Threshold Ternary Networks (STTN), which enables the model to automatically determine quantization intervals instead of depending on a hard threshold. Concretely, we replace the original ternary kernel with the addition of two binary kernels at training time, where ternary values are determined by the combination of two corresponding binary values. At inference time, we add up the two binary kernels to obtain a single ternary kernel. Our method dramatically outperforms current state-of-the-arts, lowering the performance gap between full-precision networks and extreme low bit networks. Experiments on ImageNet with ResNet-18 (Top-1 66.2%) achieves new state-of-the-art. Update: In this version, we further fine-tune the experimental hyperparameters and training procedure. The latest STTN shows that ResNet-18 with ternary weights and ternary activations achieves up to 68.2% Top-1 accuracy on ImageNet. Code is available at: github.com/WeixiangXu/STTN.

Changwang Mei, Peisong Wang, Zekun Li et al.

Visual Autoregressive (VAR) models deliver high-quality image generation but suffer from significant inference latency at high resolutions. Recent acceleration approaches most rely on heuristic measures with layer features to prune tokens. Such heuristics are sensitive to complex contextual semantics, leading to inaccurate identification of redundant computation and poor adaptability across prompts. We rethink redundancy in VAR from the perspective of its impact on pixel-space generation and introduce Latent Discrepancy. This unified metric quantifies a token's contribution by measuring the change in model states during generation. Our analysis shows that redundancy is more accurately identified when guided by image latent or pixel-space signals. We further observed that in classifier-free guidance (CFG), the convergence trend of the discrepancy between conditional and unconditional branches exhibits high dynamics with different prompts. Based on these findings, we propose LD-Pruning (Latent Discrepancy Pruning), a training-free framework that removes redundancy via latent discrepancy by integrating decoding-free region selection and adaptive unconditional-branch skipping. Extensive experiments show that LD-Pruning substantially reduces inference latency while maintaining high generation quality, achieving up to 2.35x speedup on Infinity-8B.

Zeyu Zhu, Peisong Wang, Qinghao Hu et al.

Graph Neural Networks (GNNs) have shown great superiority on non-Euclidean graph data, achieving ground-breaking performance on various graph-related tasks. As a practical solution to train GNN on large graphs with billions of nodes and edges, the sampling-based training is widely adopted by existing training frameworks. However, through an in-depth analysis, we observe that the efficiency of existing sampling-based training frameworks is still limited due to the key bottlenecks lying in all three phases of sampling-based training, i.e., subgraph sample, memory IO, and computation. To this end, we propose FastGL, a GPU-efficient Framework for accelerating sampling-based training of GNN at Large scale by simultaneously optimizing all above three phases, taking into account both GPU characteristics and graph structure. Specifically, by exploiting the inherent overlap within graph structures, FastGL develops the Match-Reorder strategy to reduce the data traffic, which accelerates the memory IO without incurring any GPU memory overhead. Additionally, FastGL leverages a Memory-Aware computation method, harnessing the GPU memory's hierarchical nature to mitigate irregular data access during computation. FastGL further incorporates the Fused-Map approach aimed at diminishing the synchronization overhead during sampling. Extensive experiments demonstrate that FastGL can achieve an average speedup of 11.8x, 2.2x and 1.5x over the state-of-the-art frameworks PyG, DGL, and GNNLab, respectively.Our code is available at https://github.com/a1bc2def6g/fastgl-ae.

Xingting Yao, Qinghao Hu, Fei Zhou et al.

In this paper, we propose SpikingNeRF, which aligns the temporal dimension of spiking neural networks (SNNs) with the radiance rays, to seamlessly accommodate SNNs to the reconstruction of neural radiance fields (NeRF). Thus, the computation turns into a spike-based, multiplication-free manner, reducing energy consumption and making high-quality 3D rendering, for the first time, accessible to neuromorphic hardware. In SpikingNeRF, each sampled point on the ray is matched to a particular time step and represented in a hybrid manner where the voxel grids are maintained as well. Based on the voxel grids, sampled points are determined whether to be masked out for faster training and inference. However, this masking operation also incurs irregular temporal length, making it intractable for hardware processors, e.g., GPUs, to conduct parallel training. To address this problem, we develop the temporal padding strategy to tackle the masked samples to maintain regular temporal length, i.e., regular tensors, and further propose the temporal condensing strategy to form a denser data structure for hardware-friendly computation. Experiments on various datasets demonstrate that our method can reduce energy consumption by an average of 70.79\% and obtain comparable synthesis quality with the ANN baseline. Verification on the neuromorphic hardware accelerator also shows that SpikingNeRF can further benefit from neuromorphic computing over the ANN baselines on energy efficiency. Codes and the appendix are in \url{https://github.com/Ikarosy/SpikingNeRF-of-CASIA}.

Zeyu Zhu, Fanrong Li, Zitao Mo et al.

As graph data size increases, the vast latency and memory consumption during inference pose a significant challenge to the real-world deployment of Graph Neural Networks (GNNs). While quantization is a powerful approach to reducing GNNs complexity, most previous works on GNNs quantization fail to exploit the unique characteristics of GNNs, suffering from severe accuracy degradation. Through an in-depth analysis of the topology of GNNs, we observe that the topology of the graph leads to significant differences between nodes, and most of the nodes in a graph appear to have a small aggregation value. Motivated by this, in this paper, we propose the Aggregation-Aware mixed-precision Quantization ($\rm A^2Q$) for GNNs, where an appropriate bitwidth is automatically learned and assigned to each node in the graph. To mitigate the vanishing gradient problem caused by sparse connections between nodes, we propose a Local Gradient method to serve the quantization error of the node features as the supervision during training. We also develop a Nearest Neighbor Strategy to deal with the generalization on unseen graphs. Extensive experiments on eight public node-level and graph-level datasets demonstrate the generality and robustness of our proposed method. Compared to the FP32 models, our method can achieve up to a 18.6x (i.e., 1.70bit) compression ratio with negligible accuracy degradation. Morever, compared to the state-of-the-art quantization method, our method can achieve up to 11.4\% and 9.5\% accuracy improvements on the node-level and graph-level tasks, respectively, and up to 2x speedup on a dedicated hardware accelerator.

Meng Zhang, Qinghao Hu, Peng Sun et al.

Training Graph Neural Networks (GNNs) on large graphs is challenging due to the conflict between the high memory demand and limited GPU memory. Recently, distributed full-graph GNN training has been widely adopted to tackle this problem. However, the substantial inter-GPU communication overhead can cause severe throughput degradation. Existing communication compression techniques mainly focus on traditional DNN training, whose bottleneck lies in synchronizing gradients and parameters. We find they do not work well in distributed GNN training as the barrier is the layer-wise communication of features during the forward pass & feature gradients during the backward pass. To this end, we propose an efficient distributed GNN training framework Sylvie, which employs one-bit quantization technique in GNNs and further pipelines the curtailed communication with computation to enormously shrink the overhead while maintaining the model quality. In detail, Sylvie provides a lightweight Low-bit Module to quantize the sent data and dequantize the received data back to full precision values in each layer. Additionally, we propose a Bounded Staleness Adaptor to control the introduced staleness to achieve further performance enhancement. We conduct theoretical convergence analysis and extensive experiments on various models & datasets to demonstrate Sylvie can considerably boost the training throughput by up to 28.1x.

Meng Zhang, Jie Sun, Qinghao Hu et al.

Graph Transformer is a new architecture that surpasses GNNs in graph learning. While there emerge inspiring algorithm advancements, their practical adoption is still limited, particularly on real-world graphs involving up to millions of nodes. We observe existing graph transformers fail on large-scale graphs mainly due to heavy computation, limited scalability and inferior model quality. Motivated by these observations, we propose TorchGT, the first efficient, scalable, and accurate graph transformer training system. TorchGT optimizes training at different levels. At algorithm level, by harnessing the graph sparsity, TorchGT introduces a Dual-interleaved Attention which is computation-efficient and accuracy-maintained. At runtime level, TorchGT scales training across workers with a communication-light Cluster-aware Graph Parallelism. At kernel level, an Elastic Computation Reformation further optimizes the computation by reducing memory access latency in a dynamic way. Extensive experiments demonstrate that TorchGT boosts training by up to 62.7x and supports graph sequence lengths of up to 1M.

Shengxiao Zhou, Chenghua Li, Jianhao Huang et al.

Reference-guided instance editing is fundamentally limited by semantic entanglement, where a reference's intrinsic appearance is intertwined with its extrinsic attributes. The key challenge lies in disentangling what information should be borrowed from the reference, and determining how to apply it appropriately to the target. To tackle this challenge, we propose GENIE, a Generalizable Instance Editing framework capable of achieving explicit disentanglement. GENIE first corrects spatial misalignments with a Spatial Alignment Module (SAM). Then, an Adaptive Residual Scaling Module (ARSM) learns what to borrow by amplifying salient intrinsic cues while suppressing extrinsic attributes, while a Progressive Attention Fusion (PAF) mechanism learns how to render this appearance onto the target, preserving its structure. Extensive experiments on the challenging AnyInsertion dataset demonstrate that GENIE achieves state-of-the-art fidelity and robustness, setting a new standard for disentanglement-based instance editing.

Shang Yang, Junxian Guo, Haotian Tang et al. · mit

Large language models (LLMs) have shown remarkable potential in processing long sequences and complex reasoning tasks, yet efficiently serving these models remains challenging due to the quadratic computational complexity of attention in the prefilling stage and the large memory footprint of the KV cache in the decoding stage. To address these issues, we introduce LServe, an efficient system that accelerates long-sequence LLM serving via hybrid sparse attention. This method unifies different hardware-friendly, structured sparsity patterns for both prefilling and decoding attention into a single framework, where computations on less important tokens are skipped block-wise. LServe demonstrates the compatibility of static and dynamic sparsity in long-context LLM attention. This design enables multiplicative speedups by combining these optimizations. Specifically, we convert half of the attention heads to nearly free streaming heads in both the prefilling and decoding stages. Additionally, we find that only a constant number of KV pages is required to preserve long-context and reasoning capabilities, irrespective of context length. We then design a hierarchical KV page selection policy that dynamically prunes KV pages based on query-centric similarity. On average, LServe accelerates LLM prefilling by up to 2.9x and decoding by 1.3-2.1x over vLLM, maintaining long-context accuracy. Code is released at https://github.com/mit-han-lab/omniserve.

Weifan Guan, Huasen Xi, Chenxiao Zhang et al.

Current embodied AI systems face severe engineering impediments, primarily characterized by poor cross-scenario adaptability, rigid inter-module coupling, and fragmented inference acceleration. To overcome these limitations, we propose RoboNeuron, a universal deployment framework for embodied intelligence. RoboNeuron is the first framework to deeply integrate the cognitive capabilities of Large Language Models (LLMs) and Vision-Language-Action (VLA) models with the real-time execution backbone of the Robot Operating System (ROS). We utilize the Model Context Protocol (MCP) as a semantic bridge, enabling the LLM to dynamically orchestrate underlying robotic tools. The framework establishes a highly modular architecture that strictly decouples sensing, reasoning, and control by leveraging ROS's unified communication interfaces. Crucially, we introduce an automated tool to translate ROS messages into callable MCP functions, significantly streamlining development. RoboNeuron significantly enhances cross-scenario adaptability and component flexibility, while establishing a systematic platform for horizontal performance benchmarking, laying a robust foundation for scalable real-world embodied applications.

Zhuoyang Zhang, Shang Yang, Qinghao Hu et al.

Vision-Language-Action (VLA) models convert high-level language instructions into concrete, executable actions, a task that is especially challenging in open-world environments. We present Visual Foresight Planning (ForeAct), a general and efficient planner that guides a VLA step-by-step using imagined future observations and subtask descriptions. With an imagined future observation, the VLA can focus on visuo-motor inference rather than high-level semantic reasoning, leading to improved accuracy and generalization. Our planner comprises a highly efficient foresight image generation module that predicts a high-quality 640$\times$480 future observation from the current visual input and language instruction within only 0.33s on an H100 GPU, together with a vision-language model that reasons over the task and produces subtask descriptions for both the generator and the VLA. Importantly, state-of-the-art VLAs can integrate our planner seamlessly by simply augmenting their visual inputs, without any architectural modification. The foresight generator is pretrained on over 1 million multi-task, cross-embodiment episodes, enabling it to learn robust embodied dynamics. We evaluate our framework on a benchmark that consists of 11 diverse, multi-step real-world tasks. It achieves an average success rate of 87.4%, demonstrating a +40.9% absolute improvement over the $π_0$ baseline (46.5%) and a +30.3% absolute improvement over $π_0$ augmented with textual subtask guidance (57.1%).

Meng Li, Peisong Wang, Yuantian Shao et al.

Large Language Models (LLMs) achieve strong performance across diverse tasks but face deployment challenges due to their massive size. Structured pruning offers acceleration benefits but leads to significant performance degradation. Recent PCA-based pruning methods have alleviated this issue by retaining key activation components, but are only applied between modules in order to fuse the transformation matrix, which introduces extra parameters and severely disrupts activation distributions due to residual connections. To address these issues, we propose IntraSlice, a framework that applies block-wise module-intra PCA compression pruning. By leveraging the structural characteristics of Transformer modules, we design an approximate PCA method whose transformation matrices can be fully fused into the model without additional parameters. We also introduce a PCA-based global pruning ratio estimator that further considers the distribution of compressed activations, building on conventional module importance. We validate our method on Llama2, Llama3, and Phi series across various language benchmarks. Experimental results demonstrate that our approach achieves superior compression performance compared to recent baselines at the same compression ratio or inference speed.

Luke Huang, Zhuoyang Zhang, Qinghao Hu et al.

Reinforcement learning (RL) is widely used to improve large language models on reasoning tasks, and asynchronous RL training is attractive because it increases end-to-end throughput. However, for widely adopted critic-free policy-gradient methods such as REINFORCE and GRPO, high asynchrony makes the policy-gradient estimator markedly $\textbf{higher variance}$: training on stale rollouts creates heavy-tailed importance ratios, causing a small fraction of samples to dominate updates. This amplification makes gradients noisy and learning unstable relative to matched on-policy training. Across math and general reasoning benchmarks, we find collapse is reliably predicted by effective sample size (ESS) and unstable gradient norms. Motivated by this diagnosis, we propose $\textbf{V}$ariance $\textbf{C}$ontrolled $\textbf{P}$olicy $\textbf{O}$ptimization ($\textbf{VCPO}$), a general stabilization method for REINFORCE/GRPO-style algorithms that (i) scales learning rate based on effective sample size to dampen unreliable updates, and (ii) applies a closed-form minimum-variance baseline for the off-policy setting, avoiding an auxiliary value model and adding minimal overhead. Empirically, VCPO substantially improves robustness for asynchronous training across math, general reasoning, and tool-use tasks, outperforming a broad suite of baselines spanning masking/clipping stabilizers and algorithmic variants. This reduces long-context, multi-turn training time by 2.5$\times$ while matching synchronous performance, demonstrating that explicit control of policy-gradient variance is key for reliable asynchronous RL at scale.

Qinghao Hu, Shang Yang, Junxian Guo et al.

The emergence of Large Language Models (LLMs) with strong reasoning capabilities marks a significant milestone, unlocking new frontiers in complex problem-solving. However, training these reasoning models, typically using Reinforcement Learning (RL), encounters critical efficiency bottlenecks: response generation during RL training exhibits a persistent long-tail distribution, where a few very long responses dominate execution time, wasting resources and inflating costs. To address this, we propose TLT, a system that accelerates reasoning RL training losslessly by integrating adaptive speculative decoding. Applying speculative decoding in RL is challenging due to the dynamic workloads, evolving target model, and draft model training overhead. TLT overcomes these obstacles with two synergistic components: (1) Adaptive Drafter, a lightweight draft model trained continuously on idle GPUs during long-tail generation to maintain alignment with the target model at no extra cost; and (2) Adaptive Rollout Engine, which maintains a memory-efficient pool of pre-captured CUDAGraphs and adaptively select suitable SD strategies for each input batch. Evaluations demonstrate that TLT achieves over 1.7x end-to-end RL training speedup over state-of-the-art systems, preserves the model accuracy, and yields a high-quality draft model as a free byproduct suitable for efficient deployment. Code is released at https://github.com/mit-han-lab/fastrl.

Xingting Yao, Qinghao Hu, Fei Zhou et al.

Multi-bit spiking neural networks (SNNs) have recently become a heated research spot, pursuing energy-efficient and high-accurate AI. However, with more bits involved, the associated memory and computation demands escalate to the point where the performance improvements become disproportionate. Based on the insight that different layers demonstrate different importance and extra bits could be wasted and interfering, this paper presents an adaptive bit allocation strategy for direct-trained SNNs, achieving fine-grained layer-wise allocation of memory and computation resources. Thus, SNN's efficiency and accuracy can be improved. Specifically, we parametrize the temporal lengths and the bit widths of weights and spikes, and make them learnable and controllable through gradients. To address the challenges caused by changeable bit widths and temporal lengths, we propose the refined spiking neuron, which can handle different temporal lengths, enable the derivation of gradients for temporal lengths, and suit spike quantization better. In addition, we theoretically formulate the step-size mismatch problem of learnable bit widths, which may incur severe quantization errors to SNN, and accordingly propose the step-size renewal mechanism to alleviate this issue. Experiments on various datasets, including the static CIFAR and ImageNet datasets and the dynamic CIFAR-DVS, DVS-GESTURE, and SHD datasets, demonstrate that our methods can reduce the overall memory and computation cost while achieving higher accuracy. Particularly, our SEWResNet-34 can achieve a 2.69% accuracy gain and 4.16x lower bit budgets over the advanced baseline work on ImageNet. This work will be open-sourced.

Xiangyu He, Ke Cheng, Qiang Chen et al.

Long-range dependencies modeling, widely used in capturing spatiotemporal correlation, has shown to be effective in CNN dominated computer vision tasks. Yet neither stacks of convolutional operations to enlarge receptive fields nor recent nonlocal modules is computationally efficient. In this paper, we present a generic family of lightweight global descriptors for modeling the interactions between positions across different dimensions (e.g., channels, frames). This descriptor enables subsequent convolutions to access the informative global features with negligible computational complexity and parameters. Benchmark experiments show that the proposed method can complete state-of-the-art long-range mechanisms with a significant reduction in extra computing cost. Code available at https://github.com/HolmesShuan/Compact-Global-Descriptor.

Qinghao Hu, Zhisheng Ye, Zerui Wang et al.

Large Language Models (LLMs) have presented impressive performance across several transformative tasks. However, it is non-trivial to efficiently utilize large-scale cluster resources to develop LLMs, often riddled with numerous challenges such as frequent hardware failures, intricate parallelization strategies, and imbalanced resource utilization. In this paper, we present an in-depth characterization study of a six-month LLM development workload trace collected from our GPU datacenter Acme. Specifically, we investigate discrepancies between LLMs and prior task-specific Deep Learning (DL) workloads, explore resource utilization patterns, and identify the impact of various job failures. Our analysis summarizes hurdles we encountered and uncovers potential opportunities to optimize systems tailored for LLMs. Furthermore, we introduce our system efforts: (1) fault-tolerant pretraining, which enhances fault tolerance through LLM-involved failure diagnosis and automatic recovery. (2) decoupled scheduling for evaluation, which achieves timely performance feedback via trial decomposition and scheduling optimization.

Yukang Chen, Wei Huang, Baifeng Shi et al.

We introduce a full-stack framework that scales up reasoning in vision-language models (VLMs) to long videos, leveraging reinforcement learning. We address the unique challenges of long video reasoning by integrating three critical components: (1) a large-scale dataset, LongVideo-Reason, comprising 104K long video QA pairs with high-quality reasoning annotations across diverse domains such as sports, games, and vlogs; (2) a two-stage training pipeline that extends VLMs with chain-of-thought supervised fine-tuning (CoT-SFT) and reinforcement learning (RL); and (3) a training infrastructure for long video RL, named Multi-modal Reinforcement Sequence Parallelism (MR-SP), which incorporates sequence parallelism and a vLLM-based engine tailored for long video, using cached video embeddings for efficient rollout and prefilling. In our experiments, LongVILA-R1-7B achieves strong performance on video benchmarks, reaching 65.1% and 71.1% accuracy on VideoMME without and with subtitles, respectively, and consistently outperforming LongVILA-7B across multiple benchmarks. Moreover, LongVILA-R1-7B supports processing up to 8,192 video frames per video, and configurable FPS settings. Notably, our MR-SP system achieves up to 2.1x speedup on long video RL training. In addition, we release our training system for public availability that supports RL training on various modalities (video, text, and audio), various models (VILA and Qwen series), and even image and video generation models. On a single A100 node (8 GPUs), it supports RL training on hour-long videos (e.g., 3,600 frames).

Xiaozhe Yao, Qinghao Hu, Ana Klimovic

Fine-tuning large language models (LLMs) greatly improves model quality for downstream tasks. However, serving many fine-tuned LLMs concurrently is challenging due to the sporadic, bursty, and varying request patterns of different LLMs. To bridge this gap, we present DeltaZip, an LLM serving system that efficiently serves multiple full-parameter fine-tuned models concurrently by aggressively compressing model deltas by up to 10x while maintaining high model quality. The key insight behind this design is that fine-tuning results in small-magnitude changes to the pre-trained model. By co-designing the serving system with the compression algorithm, DeltaZip achieves 2x to 12x improvement in throughput compared to the state-of-the-art systems.

Yuxian Gu, Qinghao Hu, Shang Yang et al.

We present Jet-Nemotron, a new family of hybrid-architecture language models, which matches or exceeds the accuracy of leading full-attention models while significantly improving generation throughput. Jet-Nemotron is developed using Post Neural Architecture Search (PostNAS), a novel neural architecture exploration pipeline that enables efficient model design. Unlike prior approaches, PostNAS begins with a pre-trained full-attention model and freezes its MLP weights, allowing efficient exploration of attention block designs. The pipeline includes four key components: (1) learning optimal full-attention layer placement and elimination, (2) linear attention block selection, (3) designing new attention blocks, and (4) performing hardware-aware hyperparameter search. Our Jet-Nemotron-2B model achieves comparable or superior accuracy to Qwen3, Qwen2.5, Gemma3, and Llama3.2 across a comprehensive suite of benchmarks while delivering up to 53.6x generation throughput speedup and 6.1x prefilling speedup. It also achieves higher accuracy on MMLU and MMLU-Pro than recent advanced MoE full-attention models, such as DeepSeek-V3-Small and Moonlight, despite their larger scale with 15B total and 2.2B activated parameters.

Weifan Guan, Qinghao Hu, Aosheng Li et al.

Vision-Language-Action (VLA) models extend vision-language models to embodied control by mapping natural-language instructions and visual observations to robot actions. Despite their capabilities, VLA systems face significant challenges due to their massive computational and memory demands, which conflict with the constraints of edge platforms such as on-board mobile manipulators that require real-time performance. Addressing this tension has become a central focus of recent research. In light of the growing efforts toward more efficient and scalable VLA systems, this survey provides a systematic review of approaches for improving VLA efficiency, with an emphasis on reducing latency, memory footprint, and training and inference costs. We categorize existing solutions into four dimensions: model architecture, perception feature, action generation, and training/inference strategies, summarizing representative techniques within each category. Finally, we discuss future trends and open challenges, highlighting directions for advancing efficient embodied intelligence.

Zerui Wang, Qinghao Hu, Ana Klimovic et al.

Deep learning (DL) schedulers are pivotal in optimizing resource allocation in GPU clusters, but operate with a critical limitation: they are largely blind to the semantic context of the jobs they manage. This forces them to rely on limited metadata, leading to high profiling overhead, unreliable duration estimation, inadequate failure handling, and poor observability. To this end, we propose SchedMate, a framework that bridges this semantic gap by systematically extracting deep insights from overlooked, unstructured data sources: source code, runtime logs, and historical jobs. SchedMate enhances existing schedulers non-intrusively through three LLM-based components. Our implementation integrates seamlessly with existing deep learning schedulers. Evaluations on a 128-GPU physical cluster and extensive simulations on production traces show SchedMate reduces average job completion times by up to 1.91x, substantially enhancing the scheduling performance, demonstrating the critical role of semantic-awareness in modern DL scheduling.

Jiecheng Zhou, Qinghao Hu, Yuyang Jin et al.

Large Language Models (LLMs) are now widely used across many domains. With their rapid development, Reinforcement Learning with Verifiable Rewards (RLVR) has surged in recent months to enhance their reasoning and understanding abilities. However, its complex data flows and diverse tasks pose substantial challenges to RL training systems, and there is limited understanding of RLVR from a system perspective. To thoroughly understand the system challenges introduced by RLVR, we present a characterization study of RLVR tasks in our LLM deployment. Specifically, we investigate the distribution and variation trends of workloads across different RL tasks across training steps. We identify issues such as GPU idling caused by skewed sequence length distribution, inefficient parallel strategies in dynamically varying workloads, inefficient data management mechanisms, and load imbalance. We describe our observations and call for further investigation into the remaining open challenges. Furthermore, we propose PolyTrace benchmark suite to conduct evaluation with realistic workloads, and a practical use case validates that PolyTrace benchmark suite exhibits 94.7% accuracy.

Qinghao Hu, Peng Sun, Shengen Yan et al.

Modern GPU datacenters are critical for delivering Deep Learning (DL) models and services in both the research community and industry. When operating a datacenter, optimization of resource scheduling and management can bring significant financial benefits. Achieving this goal requires a deep understanding of the job features and user behaviors. We present a comprehensive study about the characteristics of DL jobs and resource management. First, we perform a large-scale analysis of real-world job traces from SenseTime. We uncover some interesting conclusions from the perspectives of clusters, jobs and users, which can facilitate the cluster system designs. Second, we introduce a general-purpose framework, which manages resources based on historical data. As case studies, we design: a Quasi-Shortest-Service-First scheduling service, which can minimize the cluster-wide average job completion time by up to 6.5x; and a Cluster Energy Saving service, which improves overall cluster utilization by up to 13%.

Tianli Zhao, Qinghao Hu, Xiangyu He et al.

Acceleration of deep neural networks to meet a specific latency constraint is essential for their deployment on mobile devices. In this paper, we design an architecture aware latency constrained sparse (ALCS) framework to prune and accelerate CNN models. Taking modern mobile computation architectures into consideration, we propose Single Instruction Multiple Data (SIMD)-structured pruning, along with a novel sparse convolution algorithm for efficient computation. Besides, we propose to estimate the run time of sparse models with piece-wise linear interpolation. The whole latency constrained pruning task is formulated as a constrained optimization problem that can be efficiently solved with Alternating Direction Method of Multipliers (ADMM). Extensive experiments show that our system-algorithm co-design framework can achieve much better Pareto frontier among network accuracy and latency on resource-constrained mobile devices.

Xing Lan, Qinghao Hu, Qiang Chen et al.

Heatmap-based regression overcomes the lack of spatial and contextual information of direct coordinate regression, and has revolutionized the task of face alignment. Yet it suffers from quantization errors caused by neglecting subpixel coordinates in image resizing and network downsampling. In this paper, we first quantitatively analyze the quantization error on benchmarks, which accounts for more than 1/3 of the whole prediction errors for state-of-the-art methods. To tackle this problem, we propose a novel Heatmap In Heatmap(HIH) representation and a coordinate soft-classification (CSC) method, which are seamlessly integrated into the classic hourglass network. The HIH representation utilizes nested heatmaps to jointly represent the coordinate label: one heatmap called integer heatmap stands for the integer coordinate, and the other heatmap named decimal heatmap represents the subpixel coordinate. The range of a decimal heatmap makes up one pixel in the corresponding integer heatmap. Besides, we transfer the offset regression problem to an interval classification task, and CSC regards the confidence of the pixel as the probability of the interval. Meanwhile, CSC applying the distribution loss leverage the soft labels generated from the Gaussian distribution function to guide the offset heatmap training, which makes it easier to learn the distribution of coordinate offsets. Extensive experiments on challenging benchmark datasets demonstrate that our HIH can achieve state-of-the-art results. In particular, our HIH reaches 4.08 NME (Normalized Mean Error) on WFLW, and 3.21 on COFW, which exceeds previous methods by a significant margin.

Xiangyu He, Qinghao Hu, Peisong Wang et al.

Convolutional neural networks are able to learn realistic image priors from numerous training samples in low-level image generation and restoration. We show that, for high-level image recognition tasks, we can further reconstruct "realistic" images of each category by leveraging intrinsic Batch Normalization (BN) statistics without any training data. Inspired by the popular VAE/GAN methods, we regard the zero-shot optimization process of synthetic images as generative modeling to match the distribution of BN statistics. The generated images serve as a calibration set for the following zero-shot network quantizations. Our method meets the needs for quantizing models based on sensitive information, \textit{e.g.,} due to privacy concerns, no data is available. Extensive experiments on benchmark datasets show that, with the help of generated data, our approach consistently outperforms existing data-free quantization methods.

Fanrong Li, Zitao Mo, Peisong Wang et al.

Object detection has made impressive progress in recent years with the help of deep learning. However, state-of-the-art algorithms are both computation and memory intensive. Though many lightweight networks are developed for a trade-off between accuracy and efficiency, it is still a challenge to make it practical on an embedded device. In this paper, we present a system-level solution for efficient object detection on a heterogeneous embedded device. The detection network is quantized to low bits and allows efficient implementation with shift operators. In order to make the most of the benefits of low-bit quantization, we design a dedicated accelerator with programmable logic. Inside the accelerator, a hybrid dataflow is exploited according to the heterogeneous property of different convolutional layers. We adopt a straightforward but resource-friendly column-prior tiling strategy to map the computation-intensive convolutional layers to the accelerator that can support arbitrary feature size. Other operations can be performed on the low-power CPU cores, and the entire system is executed in a pipelined manner. As a case study, we evaluate our object detection system on a real-world surveillance video with input size of 512x512, and it turns out that the system can achieve an inference speed of 18 fps at the cost of 6.9W (with display) with an mAP of 66.4 verified on the PASCAL VOC 2012 dataset.

Xianyang Li, Feng Wang, Qinghao Hu et al.

With the development of convolutional neural network, significant progress has been made in computer vision tasks. However, the commonly used loss function softmax loss and highly efficient network architecture for common visual tasks are not as effective for face recognition. In this paper, we propose a novel loss function named Li-ArcFace based on ArcFace. Li-ArcFace takes the value of the angle through linear function as the target logit rather than through cosine function, which has better convergence and performance on low dimensional embedding feature learning for face recognition. In terms of network architecture, we improved the the perfomance of MobileFaceNet by increasing the network depth, width and adding attention module. Besides, we found some useful training tricks for face recognition. With all the above results, we won the second place in the deepglint-light challenge of LFR2019.

Qinghao Hu, Peisong Wang, Jian Cheng

Deep convolutional neural networks (CNNs) have shown appealing performance on various computer vision tasks in recent years. This motivates people to deploy CNNs to realworld applications. However, most of state-of-art CNNs require large memory and computational resources, which hinders the deployment on mobile devices. Recent studies show that low-bit weight representation can reduce much storage and memory demand, and also can achieve efficient network inference. To achieve this goal, we propose a novel approach named BWNH to train Binary Weight Networks via Hashing. In this paper, we first reveal the strong connection between inner-product preserving hashing and binary weight networks, and show that training binary weight networks can be intrinsically regarded as a hashing problem. Based on this perspective, we propose an alternating optimization method to learn the hash codes instead of directly learning binary weights. Extensive experiments on CIFAR10, CIFAR100 and ImageNet demonstrate that our proposed BWNH outperforms current state-of-art by a large margin.

Jian Cheng, Peisong Wang, Gang Li et al.

Deep neural networks have evolved remarkably over the past few years and they are currently the fundamental tools of many intelligent systems. At the same time, the computational complexity and resource consumption of these networks also continue to increase. This will pose a significant challenge to the deployment of such networks, especially in real-time applications or on resource-limited devices. Thus, network acceleration has become a hot topic within the deep learning community. As for hardware implementation of deep neural networks, a batch of accelerators based on FPGA/ASIC have been proposed in recent years. In this paper, we provide a comprehensive survey of recent advances in network acceleration, compression and accelerator design from both algorithm and hardware points of view. Specifically, we provide a thorough analysis of each of the following topics: network pruning, low-rank approximation, network quantization, teacher-student networks, compact network design and hardware accelerators. Finally, we will introduce and discuss a few possible future directions.

Jiaxiang Wu, Cong Leng, Yuhang Wang et al.

Recently, convolutional neural networks (CNN) have demonstrated impressive performance in various computer vision tasks. However, high performance hardware is typically indispensable for the application of CNN models due to the high computation complexity, which prohibits their further extensions. In this paper, we propose an efficient framework, namely Quantized CNN, to simultaneously speed-up the computation and reduce the storage and memory overhead of CNN models. Both filter kernels in convolutional layers and weighting matrices in fully-connected layers are quantized, aiming at minimizing the estimation error of each layer's response. Extensive experiments on the ILSVRC-12 benchmark demonstrate 4~6x speed-up and 15~20x compression with merely one percentage loss of classification accuracy. With our quantized CNN model, even mobile devices can accurately classify images within one second.