Fine-tuning large language models (LLMs) often faces GPU memory bottlenecks: the backward pass of first-order optimizers like Adam increases memory usage to more than 10 times the inference level (e.g., 633 GB for OPT-30B). Zeroth-order (ZO) optimizers avoid this cost by estimating gradients only from forward passes, yet existing methods like MeZO usually require many more steps to converge. Can this trade-off between speed and memory in ZO be fundamentally improved? Normalized-SGD demonstrates strong empirical performance with greater memory efficiency than Adam. In light of this, we introduce FZOO, a Fast Zeroth-Order Optimizer toward Adam-Scale Speed. FZOO reduces the total forward passes needed for convergence by employing batched one-sided estimates that adapt step sizes based on the standard deviation of batch losses. It also accelerates per-batch computation through the use of Rademacher random vector perturbations coupled with CUDA's parallel processing. Extensive experiments on diverse models, including RoBERTa-large, OPT (350M-66B), Phi-2, and Llama3, across 11 tasks validate FZOO's effectiveness. On average, FZOO outperforms MeZO by 3 percent in accuracy while requiring 3 times fewer forward passes. For RoBERTa-large, FZOO achieves average improvements of 5.6 percent in accuracy and an 18 times reduction in forward passes compared to MeZO, achieving convergence speeds comparable to Adam. We also provide theoretical analysis proving FZOO's formal equivalence to a normalized-SGD update rule and its convergence guarantees. FZOO integrates smoothly into PEFT techniques, enabling even larger memory savings. Overall, our results make single-GPU, high-speed, full-parameter fine-tuning practical and point toward future work on memory-efficient pre-training.
This paper introduces a domain-specific Large Language Model for nuclear applications, built from the publicly accessible Essential CANDU textbook. Drawing on a compact Transformer-based architecture, the model is trained on a single GPU to protect the sensitive data inherent in nuclear operations. Despite relying on a relatively small dataset, it shows encouraging signs of capturing specialized nuclear vocabulary, though the generated text sometimes lacks syntactic coherence. By focusing exclusively on nuclear content, this approach demonstrates the feasibility of in-house LLM solutions that align with rigorous cybersecurity and data confidentiality standards. Early successes in text generation underscore the model's utility for specialized tasks, while also revealing the need for richer corpora, more sophisticated preprocessing, and instruction fine-tuning to enhance domain accuracy. Future directions include extending the dataset to cover diverse nuclear subtopics, refining tokenization to reduce noise, and systematically evaluating the model's readiness for real-world applications in nuclear domain.
Most existing structured pruning methods for Large Language Models (LLMs) require substantial computational and data resources for retraining to reestablish the corrupted correlations, making them prohibitively expensive. To address this, we propose a pruning framework for LLMs called Orthogonal decomposition and Linear Calibration (Olica), which eliminates the need for retraining. A key observation is that the multi-head attention (MHA) layer depends on two types of matrix products. By treating these matrix products as unified entities and applying principal component analysis (PCA), we extract the most important information to compress LLMs without sacrificing accuracy or disrupting their original structure. Consequently, retraining becomes unnecessary. A fast decomposition method is devised, reducing the complexity of PCA by a factor of the square of the number of attention heads. Additionally, to mitigate error accumulation problem caused by pruning the feed-forward network (FFN) layer, we introduce a linear calibration method to reconstruct the residual errors of pruned layers using low-rank matrices. By leveraging singular value decomposition (SVD) on the solution of the least-squares problem, these matrices are obtained without requiring retraining. Extensive experiments show that the proposed Olica is efficient in terms of data usage, GPU memory, and running time, while delivering superior performance across multiple benchmarks.
While large language models (LLMs) have achieved remarkable performance in various tasks including mathematical reasoning, their development typically demands prohibitive computational resources. Recent advancements have reduced costs for training capable models, yet even these approaches rely on high-end hardware clusters. In this paper, we demonstrate that a single average gaming GPU can train a solid mathematical reasoning model, by integrating reinforcement learning and memory optimization techniques. Specifically, we train a 1.5B parameter mathematical reasoning model on RTX 3080 Ti of 16GB memory that achieves comparable or better performance on mathematical reasoning benchmarks than models several times larger, in resource-constrained environments. Our results challenge the paradigm that state-of-the-art mathematical reasoning necessitates massive infrastructure, democratizing access to high-performance AI research. https://github.com/shinandrew/YouronMath.
We introduce SeerAttention-R, a sparse attention framework specifically tailored for the long decoding of reasoning models. Extended from SeerAttention, SeerAttention-R retains the design of learning attention sparsity through a self-distilled gating mechanism, while removing query pooling to accommodate auto-regressive decoding. With a lightweight plug-in gating, SeerAttention-R is flexible and can be easily integrated into existing pretrained model without modifying the original parameters. We demonstrate that SeerAttention-R, trained on just 0.4B tokens, maintains near-lossless reasoning accuracy with 4K token budget in AIME benchmark under large sparse attention block sizes (64/128). Using TileLang, we develop a highly optimized sparse decoding kernel that achieves near-theoretical speedups of up to 9x over FlashAttention-3 on H100 GPU at 90% sparsity. Code is available at: https://github.com/microsoft/SeerAttention.
In the past decade, Convolutional Neural Networks (CNNs) and Transformers have achieved wide applicaiton in semantic segmentation tasks. Although CNNs with Transformer models greatly improve performance, the global context modeling remains inadequate. Recently, Mamba achieved great potential in vision tasks, showing its advantages in modeling long-range dependency. In this paper, we propose a lightweight Efficient CNN-Mamba Network for semantic segmentation, dubbed as ECMNet. ECMNet combines CNN with Mamba skillfully in a capsule-based framework to address their complementary weaknesses. Specifically, We design a Enhanced Dual-Attention Block (EDAB) for lightweight bottleneck. In order to improve the representations ability of feature, We devise a Multi-Scale Attention Unit (MSAU) to integrate multi-scale feature aggregation, spatial aggregation and channel aggregation. Moreover, a Mamba enhanced Feature Fusion Module (FFM) merges diverse level feature, significantly enhancing segmented accuracy. Extensive experiments on two representative datasets demonstrate that the proposed model excels in accuracy and efficiency balance, achieving 70.6% mIoU on Cityscapes and 73.6% mIoU on CamVid test datasets, with 0.87M parameters and 8.27G FLOPs on a single RTX 3090 GPU platform.
Precision scaling - using fewer bits to represent model parameters and related tensors during pre-training - has emerged as a compelling technique for improving GPU efficiency without sacrificing accuracy. Microscaling (MX) formats in NVIDIA's latest Blackwell GPUs represent a major leap in enabling this precision scaling aspect. These formats combine narrow floating-point data types with per-block scaling factors, offering a fine-grained approach to quantizing tensors. Although MX-formats offer the promise of improved numeric stability compared to other reduced-precision representations, in practice they must be used carefully in order to successfully converge an LLM on a multi-trillion token dataset. In this paper, we show that the rounding mode suggested in OCP specification can lead to divergence when pre-training an LLM. We show an improved rounding mode, which uses round-to-infinity to compute scaling factors, enables successful pre-training in MXFP8 for an 8B model on 15T tokens.
In multi-GPU Mixture-of-Experts (MoE) network, experts are distributed across different GPUs, which creates load imbalance as each expert processes different number of tokens. Recent works improve MoE inference load balance by dynamically duplicating popular experts to more GPUs to process excessive tokens, which requires predicting the distribution before routing. In this paper, we discuss the tradeoff of prediction strategies, accuracies, overhead, and end-to-end system performance. We propose MoE-GPS, a framework that guides the selection of the optimal predictor design under various system configurations, by quantifying the performance impact to system-level model runtime. Specifically, we advocate for Distribution-Only Prediction, a prediction strategy that only predicts overall token distribution which significantly reduces overhead compared to the traditional Token-to-Expert Prediction. On Mixtral 8x7B MMLU dataset, MoE-GPS suggests Distribution-Only Prediction which improves end-to-end inference performance by more than 23% compared with Token-to-Expert Prediction.
As integrated circuit scale grows and design complexity rises, effective circuit representation helps support logic synthesis, formal verification, and other automated processes in electronic design automation. And-Inverter Graphs (AIGs), as a compact and canonical structure, are widely adopted for representing Boolean logic in these workflows. However, the increasing complexity and integration density of modern circuits introduce structural heterogeneity and global logic information loss in AIGs, posing significant challenges to accurate circuit modeling. To address these issues, we propose FuncGNN, which integrates hybrid feature aggregation to extract multi-granularity topological patterns, thereby mitigating structural heterogeneity and enhancing logic circuit representations. FuncGNN further introduces gate-aware normalization that adapts to circuit-specific gate distributions, improving robustness to structural heterogeneity. Finally, FuncGNN employs multi-layer integration to merge intermediate features across layers, effectively synthesizing local and global semantic information for comprehensive logic representations. Experimental results on two logic-level analysis tasks (i.e., signal probability prediction and truth-table distance prediction) demonstrate that FuncGNN outperforms existing state-of-the-art methods, achieving improvements of 2.06% and 18.71%, respectively, while reducing training time by approximately 50.6% and GPU memory usage by about 32.8%.
Fractional-order differentiation has many characteristics different from integer-order differentiation. These characteristics can be applied to the optimization algorithms of artificial neural networks to obtain better results. However, due to insufficient theoretical research, at present, there is no fractional-order matrix differentiation method that is perfectly compatible with automatic differentiation (Autograd) technology. Therefore, we propose a fractional-order matrix differentiation calculation method. This method is introduced by the definition of the integer-order Jacobian matrix. We denote it as fractional-order Jacobian matrix differentiation (${{\bf{J}}^\alpha }$). Through ${{\bf{J}}^\alpha }$, we can carry out the matrix-based fractional-order chain rule. Based on the Linear module and the fractional-order differentiation, we design the fractional-order Autograd technology to enable the use of fractional-order differentiation in hidden layers, thereby enhancing the practicality of fractional-order differentiation in deep learning. In the experiment, according to the PyTorch framework, we design fractional-order Linear (FLinear) and replace nn.Linear in the multilayer perceptron with FLinear. Through the qualitative analysis of the training set and validation set $Loss$, the quantitative analysis of the test set indicators, and the analysis of time consumption and GPU memory usage during model training, we verify the superior performance of ${{\bf{J}}^\alpha }$ and prove that it is an excellent fractional-order gradient descent method in the field of deep learning.
Deep neural networks employ specialized architectures for vision, sequential and language tasks, yet this proliferation obscures their underlying commonalities. We introduce a unified matrix-order framework that casts convolutional, recurrent and self-attention operations as sparse matrix multiplications. Convolution is realized via an upper-triangular weight matrix performing first-order transformations; recurrence emerges from a lower-triangular matrix encoding stepwise updates; attention arises naturally as a third-order tensor factorization. We prove algebraic isomorphism with standard CNN, RNN and Transformer layers under mild assumptions. Empirical evaluations on image classification (MNIST, CIFAR-10/100, Tiny ImageNet), time-series forecasting (ETTh1, Electricity Load Diagrams) and language modeling/classification (AG News, WikiText-2, Penn Treebank) confirm that sparse-matrix formulations match or exceed native model performance while converging in comparable or fewer epochs. By reducing architecture design to sparse pattern selection, our matrix perspective aligns with GPU parallelism and leverages mature algebraic optimization tools. This work establishes a mathematically rigorous substrate for diverse neural architectures and opens avenues for principled, hardware-aware network design.
Jointly estimating camera poses and mapping scenes from RGBD images is a fundamental task in simultaneous localization and mapping (SLAM). State-of-the-art methods employ 3D Gaussians to represent a scene, and render these Gaussians through splatting for higher efficiency and better rendering. However, these methods cannot scale up to extremely large scenes, due to the inefficient tracking and mapping strategies that need to optimize all 3D Gaussians in the limited GPU memories throughout the training to maintain the geometry and color consistency to previous RGBD observations. To resolve this issue, we propose novel tracking and mapping strategies to work with a novel 3D representation, dubbed view-tied 3D Gaussians, for RGBD SLAM systems. View-tied 3D Gaussians is a kind of simplified Gaussians, which is tied to depth pixels, without needing to learn locations, rotations, and multi-dimensional variances. Tying Gaussians to views not only significantly saves storage but also allows us to employ many more Gaussians to represent local details in the limited GPU memory. Moreover, our strategies remove the need of maintaining all Gaussians learnable throughout the training, while improving rendering quality, and tracking accuracy. We justify the effectiveness of these designs, and report better performance over the latest methods on the widely used benchmarks in terms of rendering and tracking accuracy and scalability. Please see our project page for code and videos at https://machineperceptionlab.github.io/VTGaussian-SLAM-Project .
Most datasets for sentiment analysis lack context in which an opinion was expressed, often crucial for emotion understanding, and are mainly limited by a few emotion categories. Foundation large language models (LLMs) like GPT-4 suffer from over-predicting emotions and are too resource-intensive. We design an LLM-based data synthesis pipeline and leverage a large model, Mistral-7b, for the generation of training examples for more accessible, lightweight BERT-type encoder models. We focus on enlarging the semantic diversity of examples and propose grounding the generation into a corpus of narratives to produce non-repetitive story-character-centered utterances with unique contexts over 28 emotion classes. By running 700K inferences in 450 GPU hours, we contribute with the dataset of 100K contextual and also 300K context-less examples to cover both scenarios. We use it for fine-tuning pre-trained encoders, which results in several Emo Pillars models. We show that Emo Pillars models are highly adaptive to new domains when tuned to specific tasks such as GoEmotions, ISEAR, IEMOCAP, and EmoContext, reaching the SOTA performance on the first three. We also validate our dataset, conducting statistical analysis and human evaluation, and confirm the success of our measures in utterance diversification (although less for the neutral class) and context personalization, while pointing out the need for improved handling of out-of-taxonomy labels within the pipeline.
Test-Time Scaling (TTS) methods for enhancing Large Language Model (LLM) reasoning often incur substantial computational costs, primarily due to extensive reliance on external Process Reward Models (PRMs) or sampling methods like Best-of-N (BoN). This paper introduces Guided by Gut (GG), an efficient self-guided TTS framework that achieves PRM-level performance without costly external verifier models. Our method employs a lightweight tree search guided solely by intrinsic LLM signals, token-level confidence and step novelty. One critical innovation is improving the reliability of internal confidence estimates via a targeted reinforcement learning fine-tuning phase. Empirical evaluations on challenging mathematical reasoning benchmarks demonstrate that GG enables smaller models (e.g., 1.5B parameters) to achieve accuracy matching or surpassing significantly larger models (e.g., 32B-70B parameters), while reducing GPU memory usage by up to 10x. Compared to PRM-based methods, GG achieves comparable accuracy with 8x faster inference speeds and 4-5x lower memory usage. Additionally, GG reduces KV cache memory usage by approximately 50% compared to the BoN strategy, facilitating more efficient and practical deployment of TTS techniques.
Large Vision-Language Models (VLMs) deliver exceptional performance but require significant computational resources, limiting their deployment on mobile and edge devices. Smaller VLMs typically mirror design choices of larger models, such as extensive image tokenization, leading to inefficient GPU memory usage and constrained practicality for on-device applications. We introduce SmolVLM, a series of compact multimodal models specifically engineered for resource-efficient inference. We systematically explore architectural configurations, tokenization strategies, and data curation optimized for low computational overhead. Through this, we identify key design choices that yield substantial performance gains on image and video tasks with minimal memory footprints. Our smallest model, SmolVLM-256M, uses less than 1GB GPU memory during inference and outperforms the 300-times larger Idefics-80B model, despite an 18-month development gap. Our largest model, at 2.2B parameters, rivals state-of-the-art VLMs consuming twice the GPU memory. SmolVLM models extend beyond static images, demonstrating robust video comprehension capabilities. Our results emphasize that strategic architectural optimizations, aggressive yet efficient tokenization, and carefully curated training data significantly enhance multimodal performance, facilitating practical, energy-efficient deployments at significantly smaller scales.
Real-time decoding of neural activity is central to neuroscience and neurotechnology applications, from closed-loop experiments to brain-computer interfaces, where models are subject to strict latency constraints. Traditional methods, including simple recurrent neural networks, are fast and lightweight but often struggle to generalize to unseen data. In contrast, recent Transformer-based approaches leverage large-scale pretraining for strong generalization performance, but typically have much larger computational requirements and are not always suitable for low-resource or real-time settings. To address these shortcomings, we present POSSM, a novel hybrid architecture that combines individual spike tokenization via a cross-attention module with a recurrent state-space model (SSM) backbone to enable (1) fast and causal online prediction on neural activity and (2) efficient generalization to new sessions, individuals, and tasks through multi-dataset pretraining. We evaluate POSSM's decoding performance and inference speed on intracortical decoding of monkey motor tasks, and show that it extends to clinical applications, namely handwriting and speech decoding in human subjects. Notably, we demonstrate that pretraining on monkey motor-cortical recordings improves decoding performance on the human handwriting task, highlighting the exciting potential for cross-species transfer. In all of these tasks, we find that POSSM achieves decoding accuracy comparable to state-of-the-art Transformers, at a fraction of the inference cost (up to 9x faster on GPU). These results suggest that hybrid SSMs are a promising approach to bridging the gap between accuracy, inference speed, and generalization when training neural decoders for real-time, closed-loop applications.
We have witnessed that strong LLMs like Qwen-Math, MiMo, and Phi-4 possess immense reasoning potential inherited from the pre-training stage. With reinforcement learning (RL), these models can improve dramatically on reasoning tasks. Recent studies have shown that even RL on a single problem can unleash these models' reasoning capabilities. However, RL is not only expensive but also unstable. Even one-shot RL requires hundreds of GPU hours. This raises a critical question: Is there a more efficient way to unleash the reasoning potential of these powerful base LLMs? In this work, we demonstrate that Critique Fine-Tuning (CFT) on only one problem can effectively unleash the reasoning potential of LLMs. Our method constructs critique data by collecting diverse model-generated solutions to a single problem and using teacher LLMs to provide detailed critiques. We fine-tune Qwen and Llama family models, ranging from 1.5B to 14B parameters, on the CFT data and observe significant performance gains across diverse reasoning tasks. For example, with just 5 GPU hours of training, Qwen-Math-7B-CFT show an average improvement of 15% on six math benchmarks and 16% on three logic reasoning benchmarks. These results are comparable to or even surpass the results from RL with 20x less compute. Ablation studies reveal the robustness of one-shot CFT across different prompt problems. These results highlight one-shot CFT as a simple, general, and compute-efficient approach to unleashing the reasoning capabilities of modern LLMs.
Transformer models struggle with long-context inference due to their quadratic time and linear memory complexity. Recurrent Memory Transformers (RMTs) offer a solution by reducing the asymptotic cost to linear time and constant memory usage. However, their memory update mechanism leads to sequential execution, causing a performance bottleneck. We introduce Diagonal Batching, a scheduling scheme that unlocks parallelism across segments in RMTs while preserving exact recurrence. This approach eliminates the sequential constraint, enabling efficient GPU inference even for single long-context inputs without complex batching and pipelining techniques. Because the technique is purely a run-time computation reordering, existing RMT models adopt it with no retraining. Applied to a LLaMA-1B ARMT model, Diagonal Batching yields a 3.3x speedup over standard full-attention LLaMA-1B and a 1.8x speedup over the sequential RMT implementation on 131,072-token sequences. By removing sequential bottleneck, Diagonal Batching reduces inference cost and latency, thereby strengthening RMTs as a practical solution for real-world, long-context applications.
We present Perceive Anything Model (PAM), a conceptually straightforward and efficient framework for comprehensive region-level visual understanding in images and videos. Our approach extends the powerful segmentation model SAM 2 by integrating Large Language Models (LLMs), enabling simultaneous object segmentation with the generation of diverse, region-specific semantic outputs, including categories, label definition, functional explanations, and detailed captions. A key component, Semantic Perceiver, is introduced to efficiently transform SAM 2's rich visual features, which inherently carry general vision, localization, and semantic priors into multi-modal tokens for LLM comprehension. To support robust multi-granularity understanding, we also develop a dedicated data refinement and augmentation pipeline, yielding a high-quality dataset of 1.5M image and 0.6M video region-semantic annotations, including novel region-level streaming video caption data. PAM is designed for lightweightness and efficiency, while also demonstrates strong performance across a diverse range of region understanding tasks. It runs 1.2-2.4x faster and consumes less GPU memory than prior approaches, offering a practical solution for real-world applications. We believe that our effective approach will serve as a strong baseline for future research in region-level visual understanding.
We introduce MonkeyOCR, a vision-language model for document parsing that advances the state of the art by leveraging a Structure-Recognition-Relation (SRR) triplet paradigm. This design simplifies what would otherwise be a complex multi-tool pipeline (as in MinerU's modular approach) and avoids the inefficiencies of processing full pages with giant end-to-end models (e.g., large multimodal LLMs like Qwen-VL). In SRR, document parsing is abstracted into three fundamental questions - "Where is it?" (structure), "What is it?" (recognition), and "How is it organized?" (relation) - corresponding to layout analysis, content identification, and logical ordering. This focused decomposition balances accuracy and speed: it enables efficient, scalable processing without sacrificing precision. To train and evaluate this approach, we introduce the MonkeyDoc (the most comprehensive document parsing dataset to date), with 3.9 million instances spanning over ten document types in both Chinese and English. Experiments show that MonkeyOCR outperforms MinerU by an average of 5.1%, with particularly notable improvements on challenging content such as formulas (+15.0%) and tables (+8.6%). Remarkably, our 3B-parameter model surpasses much larger and top-performing models, including Qwen2.5-VL (72B) and Gemini 2.5 Pro, achieving state-of-the-art average performance on English document parsing tasks. In addition, MonkeyOCR processes multi-page documents significantly faster (0.84 pages per second compared to 0.65 for MinerU and 0.12 for Qwen2.5-VL-7B). The 3B model can be efficiently deployed for inference on a single NVIDIA 3090 GPU. Code and models will be released at https://github.com/Yuliang-Liu/MonkeyOCR.