Recent advances in large language models (LLMs) have driven impressive progress in omni-modal understanding and generation. However, training omni-modal LLMs remains a significant challenge due to the heterogeneous model architectures required to process diverse modalities, necessitating sophisticated system design for efficient large-scale training. Existing frameworks typically entangle model definition with parallel logic, incurring limited scalability and substantial engineering overhead for end-to-end omni-modal training. We present OmniScale, a modular and efficient training framework to accelerate the development of omni-modal LLMs. OmniScale introduces model-centric distributed recipes that decouples communication from computation, enabling efficient 3D parallelism on omni-modal LLMs. OmniScale also features a flexible configuration interface supporting seamless integration of new modalities with minimal code change. Using OmniScale, a omni-modal mixture-of-experts (MoE) model with 30B parameters can be trained with over 2,800 tokens/sec/GPU throughput and scale to 160K context lengths via 3D parallelism on 128 GPUs, showcasing its superior efficiency and scalability for training large omni-modal LLMs.

Lifelong learning investigates how models adapt when exposed to a potentially infinite stream of data.Most conventional approaches focus on updating model parameters (i.e., the neural network weights) as the underlying data distribution evolves over time.However, in natural language processing, model parameters are not the only components that matter.The tokenizer, a foundational part of the system, is usually assumed to remain fixed in lifelong learning scenarios.In this work, we challenge the validity of this assumption: as language evolves, a static tokenizer fragments newly emerging lexical items, reducing compression efficiency and consequently degrading the model performance.We introduce the Temporal Drift Tokenizer (Ted-Tok), which maintains an evolving vocabulary that adapts to emerging linguistic patterns over time.This adaptivity is driven by time-weighted frequency estimators that smooth short-term fluctuations to capture persistent linguistic trends, and a principled addition-deletion strategy targeting sink tokens.Across multiple domains, Ted-Tok consistently improves compression and task performance, with gains increasing under stronger drift, underscoring the role of tokenizer adaptivity in lifelong learning.

Low-Rank Adaptation (LoRA) is widely used for adapting large language models (LLMs) to specific domains due to its efficiency and modularity. However, vanilla LoRA struggles with task conflicts in multi-task scenarios. Recent works adopt Mixture of Experts (MoE) by treating each LoRA module as an expert, thereby mitigating task interference through multiple specialized LoRA modules. While effective, these methods often isolate knowledge within individual tasks, failing to fully exploit the shared knowledge across related tasks. In this paper, we establish a connection between single LoRA and multi-LoRA MoE, integrating them into a unified framework. We demonstrate that the dynamic routing of multiple LoRAs is functionally equivalent to rank partitioning and block-level activation within a single LoRA. To systematically study the role of expert granularity in multi-task learning, we conduct an in-depth investigation within our unified framework. Our empirical results show that a finer-grained expert partitioning not only yields significant performance gains but also captures more diverse parameter patterns. These empirical findings are supported by our theoretical analysis, which proves that finer granularity expands parameter space diversity and tightens the model's error bound. Building on these findings, we propose Single-ranked Mixture of Experts LoRA (SMoRA ), which embeds MoE into LoRA by treating each rank as an independent expert. With a dynamic rank-wise activation mechanism, SMoRA facilitates a flexible composition of knowledge, enabling the model to learn deeper and more diverse features while mitigating task conflicts. Experiments demonstrate that SMoRA activates fewer parameters yet achieves better performance in multi-task scenarios.

In this work, we investigate a typical scenario in code generation where a developer edits existing code in real time and requests a code assistant, e.g., a large language model, to re-predict the next token or next line on the fly. Naively, the LLM needs to re-encode the entire KV cache to provide an accurate prediction. However, this process is computationally expensive, especially when the sequence length is long. Simply encoding the edited subsequence and integrating it to the original KV cache meets the temporal confusion problem, leading to significantly worse performance. We address this efficiency and accuracy trade-off by introducing $\underline{\textbf{P}\text{ositional}\ \textbf{I}\text{ntegrity}\ \textbf{E}\text{ncoding}}$ (PIE). Building upon the rotary positional encoding, PIE first removes the rotary matrices in the Key cache that introduce temporal confusion and then reapplies the correct rotary matrices. This process ensures that positional relationships between tokens are correct and requires only a single round of matrix multiplication. We validate the effectiveness of PIE through extensive experiments on the RepoBench-C-8k dataset, utilizing DeepSeek-Coder models with 1.3B, 6.7B, and 33B parameters. Our evaluation includes three real-world coding tasks: code insertion, code deletion, and multi-place code editing. Results demonstrate that PIE reduces computational overhead by over 85% compared to the standard full recomputation approach across all model sizes and tasks while well approximating the model performance.

Distributed training often suffers from high communication overhead due to large-scale gradient synchronization. Although gradient compression—particularly at 4-bit or even lower precision—significantly reduces transfer volume, it typically results in sacrifice in precision and degradation of the final model accuracy. In this work, we introduce DUO, a distributed training framework designed to mitigate accuracy degradation incurred by gradient compression without involving additional overhead. DUO achieves this by inserting an additional high-precision gradient synchronization step into a previously computation-only phase, so that its communication is fully hidden by computation. We provide a comprehensive theoretical proof of convergence for DUO and validate its effectiveness through extensive pre-training experiments on GPT models. Our results indicate that DUO effectively restores accuracy when using 4-bit gradient compression, achieving performance comparable to uncompressed training. Remarkably, DUO maintains minimal accuracy degradation even under extreme compression scenarios, including 1-bit gradients or complete omission of the low-precision gradient communication step (0-bit transmission).

Low-rank adaptation (LoRA) is widely used to efficiently fine-tune large language models (LLMs), leading to multiple models fine-tuned from the same pre-trained LLM. State-of-the-art LLM serving systems colocate these LoRA models on the same GPU instances for concurrent serving, which decreases memory usage and boosts efficiency. However, the unawareness of the SLO requirements of each LoRA service and the interference between requests from different LoRA services can cause significant SLO violations. This paper presents Cannikin, a multi-LoRA inference serving system that optimizes the minimum of the SLO attainments of all LoRA services in the serving system, denoted as lagger-SLO attainment. We obtain insights from the characterization of a real-world multi-LoRA serving trace, which reveals the stable input/output lengths of the most popular LoRA services. This motivates Cannikin to propose an SLO-aware scheduling algorithm that prioritizes requests based on efficient deadline estimation. Cannikin further detects the influence of interference between different LoRA services on SLO violations and eliminates the bias between these services. The evaluation using real-world traces demonstrates that compared to the state-of-the-art multi-LoRA serving systems, Cannikin can handle up to 3.6× higher rates or 2.8× more burstiness while maintaining the SLO attainment of each LoRA service$> $90% .

In this work, we leverage the intrinsic segmentation of language sequences and design a new positional encoding method called Bilevel Positional Encoding (BiPE). For each position, our BiPE blends an intra-segment encoding and an inter-segment encoding. The intra-segment encoding identifies the locations within a segment and helps the model capture the semantic information therein via absolute positional encoding. The inter-segment encoding specifies the segment index, models the relationships between segments, and aims to improve extrapolation capabilities via relative positional encoding. Theoretical analysis shows this disentanglement of positional information makes learning more effective. The empirical results also show that our BiPE has superior length extrapolation capabilities across a wide range of tasks in diverse text modalities.

In this work, we introduce self-infilling code generation, a general framework that incorporates infilling operations into auto-regressive decoding. Our approach capitalizes on the observation that recent infilling-capable code language models can perform self-infilling: whereas conventional infilling is designed to fill in the middle based on a predefined prefix and suffix, self-infilling sequentially generates both such surrounding context and the infilled content. We utilize self-infilling to introduce novel interruption and looping mechanisms in conventional decoding, evolving it into a non-monotonic process. Interruptions allow for postponing the generation of specific code until a definitive suffix is established, enhancing control during decoding. Meanwhile, the looping mechanism, which leverages the complementary nature of self-infilling and left-to-right decoding, can iteratively update and synchronize each piece of generation cyclically. Extensive experiments across a variety of code generation benchmarks demonstrate that decoding with self-infilling not only improves the output quality but also regularizes the overall generation, which effectively mitigates potential degeneration and scaffolds code to be more consistent with intended functionality.

Recent years have witnessed a clear trend towards language models with an ever-increasing number of parameters, as well as the growing training overhead and memory usage. Distributed training, particularly through Sharded Data Parallelism (ShardedDP) which partitions optimizer states among workers, has emerged as a crucial technique to mitigate training time and memory usage. Yet, a major challenge in the scalability of ShardedDP is the intensive communication of weights and gradients. While compression techniques can alleviate this issue, they often result in worse accuracy. Driven by this limitation, we propose SDP4Bit (Toward 4Bit Communication Quantization in Sharded Data Parallelism for LLM Training), which effectively reduces the communication of weights and gradients to nearly 4 bits via two novel techniques: quantization on weight differences, and two-level gradient smooth quantization. Furthermore, SDP4Bit presents an algorithm-system co-design with runtime optimization to minimize the computation overhead of compression. Additional to the theoretical guarantees of convergence, we empirically evaluate the accuracy of SDP4Bit on the pre-training of GPT models with up to 6.7 billion parameters, and the results demonstrate a negligible impact on training loss. Furthermore, speed experiments show that SDP4Bit achieves up to 4.08× speedup in end-to-end throughput on a scale of 128 GPUs.

Starting from the seminal work of Fully Convolutional Networks (FCN), there has been significant progress on semantic segmentation. However, deep learning models often require large amounts of pixelwise annotations to train accurate and robust models. Given the prohibitively expensive annotation cost of segmentation masks, we introduce a self-training framework in this paper to leverage pseudo labels generated from unlabeled data. In order to handle the data imbalance problem of semantic segmentation, we propose a centroid sampling strategy to uniformly select training samples from every class within each epoch. We also introduce a fast training schedule to alleviate the computational burden. This enables us to explore the usage of large amounts of pseudo labels. Our Centroid Sampling based Self-Training framework (CSST) achieves state-of-the-art results on Cityscapes and CamVid datasets. On PASCAL VOC 2012 test set, our models trained with the original train set even outperform the same models trained on the much bigger augmented train set. This indicates the effectiveness of CSST when there are fewer annotations. We also demonstrate promising few-shot generalization capability from Cityscapes to BDD100K and from Cityscapes to Mapillary datasets.

Traditional normalization techniques (e.g., Batch Normalization and Instance Normalization) generally and simplistically assume that training and test data follow the same distribution. As distribution shifts are inevitable in real-world applications, well-trained models with previous normalization methods can perform badly in new environments. Can we develop new normalization methods to improve generalization robustness under distribution shifts? In this paper, we answer the question by proposing Cross-Norm and SelfNorm. CrossNorm exchanges channel-wise mean and variance between feature maps to enlarge training distribution, while SelfNorm uses attention to recalibrate the statistics to bridge gaps between training and test distributions. CrossNorm and SelfNorm can complement each other, though exploring different directions in statistics usage. Extensive experiments on different fields (vision and language), tasks (classification and segmentation), settings (supervised and semi-supervised), and distribution shift types (synthetic and natural) show the effectiveness. Code is available at https://github.com/amazon-research/crossnorm-selfnorm

Recognizing and localizing objects in the 3D space is a crucial ability for an AI agent to perceive its surrounding environment. While significant progress has been achieved with expensive LiDAR point clouds, it poses a great challenge for 3D object detection given only a monocular image. While there exist different alternatives for tackling this problem, it is found that they are either equipped with heavy networks to fuse RGB and depth information or empirically ineffective to process millions of pseudo-LiDAR points. With in-depth examination, we realize that these limitations are rooted in inaccurate object localization. In this paper, we propose a novel and lightweight approach, dubbed {\em Progressive Coordinate Transforms} (PCT) to facilitate learning coordinate representations. Specifically, a localization boosting mechanism with confidence-aware loss is introduced to progressively refine the localization prediction. In addition, semantic image representation is also exploited to compensate for the usage of patch proposals. Despite being lightweight and simple, our strategy allows us to establish a new state-of-the-art among the monocular 3D detectors on the competitive KITTI benchmark. At the same time, our proposed PCT shows great generalization to most coordinate-based 3D detection frameworks.

We present GluonCV and GluonNLP, the deep learning toolkits for computer vision and natural language processing based on Apache MXNet (incubating). These toolkits provide state-of-the-art pre-trained models, training scripts, and training logs, to facilitate rapid prototyping and promote reproducible research. We also provide modular APIs with flexible building blocks to enable efficient customization. Leveraging the MXNet ecosystem, the deep learning models in GluonCV and GluonNLP can be deployed onto a variety of platforms with different programming languages. The Apache 2.0 license has been adopted by GluonCV and GluonNLP to allow for software distribution, modification, and usage.

Much of the recent progress made in image classification research can be credited to training procedure refinements, such as changes in data augmentations and optimization methods. In the literature, however, most refinements are either briefly mentioned as implementation details or only visible in source code. In this paper, we will examine a collection of such refinements and empirically evaluate their impact on the final model accuracy through ablation study. We will show that, by combining these refinements together, we are able to improve various CNN models significantly. For example, we raise ResNet-50's top-1 validation accuracy from 75.3% to 79.29% on ImageNet. We will also demonstrate that improvement on image classification accuracy leads to better transfer learning performance in other application domains such as object detection and semantic segmentation.

Image representation methods based on deep convolutional neural networks (CNNs) have achieved the state-of-the-art performance in various computer vision tasks, such as image retrieval and person re-identification. We recognize that more discriminative feature embeddings can be learned with supervised deep metric learning and handcrafted features for image retrieval and similar applications. In this paper, we propose a new supervised deep feature embedding with a handcrafted feature model. To fuse handcrafted feature information into CNNs and realize feature embeddings, a general fusion unit is proposed (called Fusion-Net). We also define a network loss function with image label information to realize supervised deep metric learning. Our extensive experimental results on the Stanford online products' data set and the in-shop clothes retrieval data set demonstrate that our proposed methods outperform the existing state-of-the-art methods of image retrieval by a large margin. Moreover, we also explore the applications of the proposed methods in person re-identification and vehicle re-identification; the experimental results demonstrate both the effectiveness and efficiency of the proposed methods.

Human pose estimation using deep neural networks aims to map input images with large variations into multiple body keypoints, which must satisfy a set of geometric constraints and interdependence imposed by the human body model. This is a very challenging nonlinear manifold learning process in a very high dimensional feature space. We believe that the deep neural network, which is inherently an algebraic computation system, is not the most efficient way to capture highly sophisticated human knowledge, for example those highly coupled geometric characteristics and interdependence between keypoints in human poses. In this work, we propose to explore how external knowledge can be effectively represented and injected into the deep neural networks to guide its training process using learned projections that impose proper prior. Specifically, we use the stacked hourglass design and inception-resnet module to construct a fractal network to regress human pose images into heatmaps with no explicit graphical modeling. We encode external knowledge with visual features, which are able to characterize the constraints of human body models and evaluate the fitness of intermediate network output. We then inject these external features into the neural network using a projection matrix learned using an auxiliary cost function. The effectiveness of the proposed inception-resnet module and the benefit in guided learning with knowledge projection is evaluated on two widely used human pose estimation benchmarks. Our approach achieves state-of-the-art performance on both datasets.

In this work, we consider the problem of automatic content retrieval (ACR) using joint audio-video fingerprints. We focus on how to balance the query accuracy and the size of fingerprint, and how to allocate the fingerprint bits to video and audio frames to maximize the query accuracy. By introducing a novel concept called coverage, which is highly correlated to the query accuracy, we are able to construct a rate-coverage model and formulate the joint audio-video fingerprint bit rate allocation into a dynamic programming optimization problem. Our experimental results demonstrate that, compared to existing approaches, our method improves the retrieval accuracy by up to 25% while using 60% of the original fingerprint bit rate.

In this paper, we develop a new approach of spatially supervised recurrent convolutional neural networks for visual object tracking. Our recurrent convolutional network exploits the history of locations as well as the distinctive visual features learned by the deep neural networks. Inspired by recent bounding box regression methods for object detection, we study the regression capability of Long Short-Term Memory (LSTM) in the temporal domain, and propose to concatenate high-level visual features produced by convolutional networks with region information. In contrast to existing deep learning based trackers that use binary classification for region candidates, we use regression for direct prediction of the tracking locations both at the convolutional layer and at the recurrent unit. Our experimental results on challenging benchmark video tracking datasets show that our tracker is competitive with state-of-the-art approaches while maintaining low computational cost.

In this paper, we consider the animal object detection and segmentation from wildlife monitoring videos captured by motion-triggered cameras, called camera-traps. For these types of videos, existing approaches often suffer from low detection rates due to low contrast between the foreground animals and the cluttered background, as well as high false positive rates due to the dynamic background. To address this issue, we first develop a new approach to generate animal object region proposals using multilevel graph cut in the spatiotemporal domain. We then develop a cross-frame temporal patch verification method to determine if these region proposals are true animals or background patches. We construct an efficient feature description for animal detection using joint deep learning and histogram of oriented gradient features encoded with Fisher vectors. Our extensive experimental results and performance comparisons over a diverse set of challenging camera-trap data demonstrate that the proposed spatiotemporal object proposal and patch verification framework outperforms the state-of-the-art methods, including the recent Faster-RCNN method, on animal object detection accuracy by up to 4.5%.

In this paper, we consider animal object segmentation from wildlife monitoring videos captured by motion-triggered cameras, called camera-traps. This is a very challenging task because the wildlife monitoring scenes are often highly cluttered and dynamic. To address this issue, we propose to explore the ideas of coupled ensemble graph cuts and object verification. We consider video object cut as an ensemble of frame-level background-foreground object classifiers which fuse information across frames and refine their segmentation results in a collaborative and iterative manner. To significantly reduce false positives in foreground animal detection and segmentation, we learn an object verification model to further classify if the segmented image patch belongs to the background or the animal. Our extensive experimental results and performance comparisons over a diverse set of challenging camera-trap data, as well as the new Change Detection 2014 benchmark dataset, demonstrate that the proposed framework outperforms various state-of-the-art algorithms and has the capability to handle even the most challenging objects in a wide variety of video sequences.