We introduce LTX-Video, a transformer-based latent diffusion model that adopts a holistic approach to video generation by seamlessly integrating the responsibilities of the Video-VAE and the denoising transformer. Unlike existing methods, which treat these components as independent, LTX-Video aims to optimize their interaction for improved efficiency and quality. At its core is a carefully designed Video-VAE that achieves a high compression ratio of 1:192, with spatiotemporal downscaling of 32 x 32 x 8 pixels per token, enabled by relocating the patchifying operation from the transformer's input to the VAE's input. Operating in this highly compressed latent space enables the transformer to efficiently perform full spatiotemporal self-attention, which is essential for generating high-resolution videos with temporal consistency. However, the high compression inherently limits the representation of fine details. To address this, our VAE decoder is tasked with both latent-to-pixel conversion and the final denoising step, producing the clean result directly in pixel space. This approach preserves the ability to generate fine details without incurring the runtime cost of a separate upsampling module. Our model supports diverse use cases, including text-to-video and image-to-video generation, with both capabilities trained simultaneously. It achieves faster-than-real-time generation, producing 5 seconds of 24 fps video at 768x512 resolution in just 2 seconds on an Nvidia H100 GPU, outperforming all existing models of similar scale. The source code and pre-trained models are publicly available, setting a new benchmark for accessible and scalable video generation.
Missing attribute issues are prevalent in the graph learning, leading to biased outcomes in Graph Neural Networks (GNNs). Existing methods that rely on feature propagation are prone to cold start problem, particularly when dealing with attribute resetting and low-degree nodes, which hinder effective propagation and convergence. To address these challenges, we propose AttriReBoost (ARB), a novel method that incorporates propagation-based method to mitigate cold start problems in attribute-missing graphs. ARB enhances global feature propagation by redefining initial boundary conditions and strategically integrating virtual edges, thereby improving node connectivity and ensuring more stable and efficient convergence. This method facilitates gradient-free attribute reconstruction with lower computational overhead. The proposed method is theoretically grounded, with its convergence rigorously established. Extensive experiments on several real-world benchmark datasets demonstrate the effectiveness of ARB, achieving an average accuracy improvement of 5.11% over state-of-the-art methods. Additionally, ARB exhibits remarkable computational efficiency, processing a large-scale graph with 2.49 million nodes in just 16 seconds on a single GPU. Our code is available at https://github.com/limengran98/ARB.
Despite very fast progress, efficiently training large language models (LLMs) in very long contexts remains challenging. Existing methods fall back to training LLMs with short contexts (a maximum of a few thousands tokens in training) and use inference time techniques when evaluating on long contexts (above 1M tokens context window at inference). As opposed to long-context-inference, training on very long context input prompts is quickly limited by GPU memory availability and by the prohibitively long training times it requires on state-of-the-art hardware. Meanwhile, many real-life applications require not only inference but also training/fine-tuning with long context on specific tasks. Such applications include, for example, augmenting the context with various sources of raw reference information for fact extraction, fact summarization, or fact reconciliation tasks. We propose adjoint sharding, a novel technique that comprises sharding gradient calculation during training to reduce memory requirements by orders of magnitude, making training on very long context computationally tractable. Adjoint sharding is based on the adjoint method and computes equivalent gradients to backpropagation. We also propose truncated adjoint sharding to speed up the algorithm while maintaining performance. We provide a distributed version, and a paralleled version of adjoint sharding to further speed up training. Empirical results show the proposed adjoint sharding algorithm reduces memory usage by up to 3X with a 1.27B parameter large language model on 1M context length training. This allows to increase the maximum context length during training or fine-tuning of a 1.27B parameter model from 35K tokens to above 100K tokens on a training infrastructure composed of five AWS P4 instances.
Solving arithmetic tasks is a simple and fundamental skill, yet modern Large Language Models (LLMs) have great difficulty with them. We introduce the Integrated Gated Calculator (IGC), a module that enables LLMs to perform arithmetic by emulating a calculator on the GPU. We finetune a Llama model with our module and test it on the BigBench Arithmetic benchmark, where it beats the State of the Art, outperforming all models on the benchmark, including models almost two orders of magnitude larger. Our approach takes only a single iteration to run and requires no external tools. It performs arithmetic operations entirely inside the LLM without the need to produce intermediate tokens. It is computationally efficient, interpretable, and avoids side-effects on tasks that do not require arithmetic operations. It reliably achieves 98\% to 99\% accuracy across multiple training runs and for all subtasks, including the substantially harder subtask of multiplication, which was previously unsolved.
Cyberbullying significantly contributes to mental health issues in communities by negatively impacting the psychology of victims. It is a prevalent problem on social media platforms, necessitating effective, real-time detection and monitoring systems to identify harmful messages. However, current cyberbullying detection systems face challenges related to performance, dataset quality, time efficiency, and computational costs. This research aims to conduct a comparative study by adapting and evaluating existing text classification techniques within the cyberbullying detection domain. The study specifically evaluates the effectiveness and performance of these techniques in identifying cyberbullying instances on social media platforms. It focuses on leveraging and assessing large language models, including BERT, RoBERTa, XLNet, DistilBERT, and GPT-2.0, for their suitability in this domain. The results show that BERT strikes a balance between performance, time efficiency, and computational resources: Accuracy of 95%, Precision of 95%, Recall of 95%, F1 Score of 95%, Error Rate of 5%, Inference Time of 0.053 seconds, RAM Usage of 35.28 MB, CPU/GPU Usage of 0.4%, and Energy Consumption of 0.000263 kWh. The findings demonstrate that generative AI models, while powerful, do not consistently outperform fine-tuned models on the tested benchmarks. However, state-of-the-art performance can still be achieved through strategic adaptation and fine-tuning of existing models for specific datasets and tasks.
The integration of Artificial Intelligence (AI) in education requires scalable and efficient frameworks that balance performance, adaptability, and cost. This paper addresses these needs by proposing a shared backbone model architecture enhanced with lightweight LoRA adapters for task-specific fine-tuning, targeting the automated scoring of student responses across 27 mutually exclusive tasks. By achieving competitive performance (average QWK of 0.848 compared to 0.888 for fully fine-tuned models) while reducing GPU memory consumption by 60% and inference latency by 40%, the framework demonstrates significant efficiency gains. This approach aligns with the workshops' focus on improving language models for educational tasks, creating responsible innovations for cost-sensitive deployment, and supporting educators by streamlining assessment workflows. The findings underscore the potential of scalable AI to enhance learning outcomes while maintaining fairness and transparency in automated scoring systems.
Graph neural network universal interatomic potentials (GNN-UIPs) have demonstrated remarkable generalization and transfer capabilities in material discovery and property prediction. These models can accelerate molecular dynamics (MD) simulation by several orders of magnitude while maintaining \textit{ab initio} accuracy, making them a promising new paradigm in material simulations. One notable example is Crystal Hamiltonian Graph Neural Network (CHGNet), pretrained on the energies, forces, stresses, and magnetic moments from the MPtrj dataset, representing a state-of-the-art GNN-UIP model for charge-informed MD simulations. However, training the CHGNet model is time-consuming(8.3 days on one A100 GPU) for three reasons: (i) requiring multi-layer propagation to reach more distant atom information, (ii) requiring second-order derivatives calculation to finish weights updating and (iii) the implementation of reference CHGNet does not fully leverage the computational capabilities. This paper introduces FastCHGNet, an optimized CHGNet, with three contributions: Firstly, we design innovative Force/Stress Readout modules to decompose Force/Stress prediction. Secondly, we adopt massive optimizations such as kernel fusion, redundancy bypass, etc, to exploit GPU computation power sufficiently. Finally, we extend CHGNet to support multiple GPUs and propose a load-balancing technique to enhance GPU utilization. Numerical results show that FastCHGNet reduces memory footprint by a factor of 3.59. The final training time of FastCHGNet can be decreased to \textbf{1.53 hours} on 32 GPUs without sacrificing model accuracy.
Temporal Graph Neural Networks (TGNNs) have emerged as powerful tools for modeling dynamic interactions across various domains. The design space of TGNNs is notably complex, given the unique challenges in runtime efficiency and scalability raised by the evolving nature of temporal graphs. We contend that many of the existing works on TGNN modeling inadequately explore the design space, leading to suboptimal designs. Viewing TGNN models through a performance-focused lens often obstructs a deeper understanding of the advantages and disadvantages of each technique. Specifically, benchmarking efforts inherently evaluate models in their original designs and implementations, resulting in unclear accuracy comparisons and misleading runtime. To address these shortcomings, we propose a practical comparative evaluation framework that performs a design space search across well-known TGNN modules based on a unified, optimized code implementation. Using our framework, we make the first efforts towards addressing three critical questions in TGNN design, spending over 10,000 GPU hours: (1) investigating the efficiency of TGNN module designs, (2) analyzing how the effectiveness of these modules correlates with dataset patterns, and (3) exploring the interplay between multiple modules. Key outcomes of this directed investigative approach include demonstrating that the most recent neighbor sampling and attention aggregator outperform uniform neighbor sampling and MLP-Mixer aggregator; Assessing static node memory as an effective node memory alternative, and showing that the choice between static or dynamic node memory should be based on the repetition patterns in the dataset. Our in-depth analysis of the interplay between TGNN modules and dataset patterns should provide a deeper insight into TGNN performance along with potential research directions for designing more general and effective TGNNs.
Processing long temporal sequences is a key challenge in deep learning. In recent years, Transformers have become state-of-the-art for this task, but suffer from excessive memory requirements due to the need to explicitly store the sequences. To address this issue, structured state-space sequential (S4) models recently emerged, offering a fixed memory state while still enabling the processing of very long sequence contexts. The recurrent linear update of the state in these models makes them highly efficient on modern graphics processing units (GPU) by unrolling the recurrence into a convolution. However, this approach demands significant memory and massively parallel computation, which is only available on the latest GPUs. In this work, we aim to bring the power of S4 models to edge hardware by significantly reducing the size and computational demand of an S4D model through quantization-aware training, even achieving ternary weights for a simple real-world task. To this end, we extend conventional quantization-aware training to tailor it for analog in-memory compute hardware. We then demonstrate the deployment of recurrent S4D kernels on memrisitve crossbar arrays, enabling their computation in an in-memory compute fashion. To our knowledge, this is the first implementation of S4 kernels on in-memory compute hardware.
Quantization of Large Language Models (LLMs) has recently gained popularity, particularly for on-device settings with limited hardware resources. While efficient, quantization inevitably degrades model quality, especially in aggressive low-bit settings such as 3-bit and 4-bit precision. In this paper, we propose QDEC, an inference scheme that improves the quality of low-bit LLMs while preserving the key benefits of quantization: GPU memory savings and inference latency reduction. QDEC stores the residual matrix -- the difference between full-precision and quantized weights -- in CPU, and dynamically fetches the residuals for only a small portion of the weights. This portion corresponds to the salient channels, marked by activation outliers, with the fetched residuals helping to correct quantization errors in these channels. Salient channels are identified dynamically at each decoding step by analyzing the input activations -- this allows for the adaptation to the dynamic nature of activation distribution, and thus maximizes the effectiveness of error compensation. We demonstrate the effectiveness of QDEC by augmenting state-of-the-art quantization methods. For example, QDEC reduces the perplexity of a 3-bit Llama-3-8B-Instruct model from 10.15 to 9.12 -- outperforming its 3.5-bit counterpart -- while adding less than 0.0003\% to GPU memory usage and incurring only a 1.7\% inference slowdown on NVIDIA RTX 4050 Mobile GPU. The code will be publicly available soon.
Deep visual odometry has demonstrated great advancements by learning-to-optimize technology. This approach heavily relies on the visual matching across frames. However, ambiguous matching in challenging scenarios leads to significant errors in geometric modeling and bundle adjustment optimization, which undermines the accuracy and robustness of pose estimation. To address this challenge, this paper proposes MambaVO, which conducts robust initialization, Mamba-based sequential matching refinement, and smoothed training to enhance the matching quality and improve the pose estimation in deep visual odometry. Specifically, when a new frame is received, it is matched with the closest keyframe in the maintained Point-Frame Graph (PFG) via the semi-dense based Geometric Initialization Module (GIM). Then the initialized PFG is processed by a proposed Geometric Mamba Module (GMM), which exploits the matching features to refine the overall inter-frame pixel-to-pixel matching. The refined PFG is finally processed by deep BA to optimize the poses and the map. To deal with the gradient variance, a Trending-Aware Penalty (TAP) is proposed to smooth training by balancing the pose loss and the matching loss to enhance convergence and stability. A loop closure module is finally applied to enable MambaVO++. On public benchmarks, MambaVO and MambaVO++ demonstrate SOTA accuracy performance, while ensuring real-time running performance with low GPU memory requirement. Codes will be publicly available.
In recent years, Convolutional Neural Networks (CNNs) have been widely adopted in computer vision. Complex CNN architecture running on CPU or GPU has either insufficient throughput or prohibitive power consumption. Hence, there is a need to have dedicated hardware to accelerate the computation workload to solve these limitations. In this paper, we accelerate a CNN for image classification with the CIFAR-10 dataset using Vitis-AI on Xilinx Zynq UltraScale+ MPSoC ZCU104 FPGA evaluation board. The work achieves 3.33-5.82x higher throughput and 3.39-6.30x higher energy efficiency than CPU and GPU baselines. It shows the potential to extract 2D features for downstream tasks, such as depth estimation and 3D reconstruction.
The expansion of large language models (LLMs) with hundreds of billions of parameters presents significant challenges to computational resources, particularly data movement and memory bandwidth. Long-context LLMs, which process sequences of tens of thousands of tokens, further increase the demand on the memory system as the complexity in attention layers and key-value cache sizes is proportional to the context length. Processing-in-Memory (PIM) maximizes memory bandwidth by moving compute to the data and can address the memory bandwidth challenges; however, PIM is not necessarily scalable to accelerate long-context LLM because of limited per-module memory capacity and the inflexibility of fixed-functional unit PIM architecture and static memory management. In this work, we propose LoL-PIM which is a multi-node PIM architecture that accelerates long context LLM through hardware-software co-design. In particular, we propose how pipeline parallelism can be exploited across a multi-PIM module while a direct PIM access (DPA) controller (or DMA for PIM) is proposed that enables dynamic PIM memory management and results in efficient PIM utilization across a diverse range of context length. We developed an MLIR-based compiler for LoL-PIM extending a commercial PIM-based compiler where the software modifications were implemented and evaluated, while the hardware changes were modeled in the simulator. Our evaluations demonstrate that LoL-PIM significantly improves throughput and reduces latency for long-context LLM inference, outperforming both multi-GPU and GPU-PIM systems (up to 8.54x and 16.0x speedup, respectively), thereby enabling more efficient deployment of LLMs in real-world applications.
We introduce TangoFlux, an efficient Text-to-Audio (TTA) generative model with 515M parameters, capable of generating up to 30 seconds of 44.1kHz audio in just 3.7 seconds on a single A40 GPU. A key challenge in aligning TTA models lies in the difficulty of creating preference pairs, as TTA lacks structured mechanisms like verifiable rewards or gold-standard answers available for Large Language Models (LLMs). To address this, we propose CLAP-Ranked Preference Optimization (CRPO), a novel framework that iteratively generates and optimizes preference data to enhance TTA alignment. We demonstrate that the audio preference dataset generated using CRPO outperforms existing alternatives. With this framework, TangoFlux achieves state-of-the-art performance across both objective and subjective benchmarks. We open source all code and models to support further research in TTA generation.
The signature kernel is a kernel between time series of arbitrary length and comes with strong theoretical guarantees from stochastic analysis. It has found applications in machine learning such as covariance functions for Gaussian processes. A strength of the underlying signature features is that they provide a structured global description of a time series. However, this property can quickly become a curse when local information is essential and forgetting is required; so far this has only been addressed with ad-hoc methods such as slicing the time series into subsegments. To overcome this, we propose a principled, data-driven approach by introducing a novel forgetting mechanism for signatures. This allows the model to dynamically adapt its context length to focus on more recent information. To achieve this, we revisit the recently introduced Random Fourier Signature Features, and develop Random Fourier Decayed Signature Features (RFDSF) with Gaussian processes (GPs). This results in a Bayesian time series forecasting algorithm with variational inference, that offers a scalable probabilistic algorithm that processes and transforms a time series into a joint predictive distribution over time steps in one pass using recurrence. For example, processing a sequence of length $10^4$ steps in $\approx 10^{-2}$ seconds and in $< 1\text{GB}$ of GPU memory. We demonstrate that it outperforms other GP-based alternatives and competes with state-of-the-art probabilistic time series forecasting algorithms.
Due to the high resource demands of Large Language Models (LLMs), achieving widespread deployment on consumer-grade devices presents significant challenges. Typically, personal or consumer-grade devices, including servers configured prior to the era of large-scale models, generally have relatively weak GPUs and relatively strong CPUs. However, most current methods primarily depend on GPUs for computation. Therefore, we propose Dovetail, an approach that deploys the draft model on the GPU to generate draft tokens while allowing the target model to perform parallel verification on the CPU, thereby improving the utilization of all available hardware resources and occupying less inter-device communication bandwidth. Accordingly, we have redesigned the draft model to better align with heterogeneous hardware characteristics. To this end, we implemented several optimizations: reducing the number of draft tokens to mitigate latency in parallel verification, increasing the depth of the draft model to enhance its predictive capacity, and introducing DGF (Dynamic Gating Fusion) to improve the integration of features and token embeddings. In the HumanEval benchmark, Dovetail achieved an inference speed of 5.86 tokens per second for LLaMA2-Chat-7B using 3GB of VRAM, representing an approximately 2.77x improvement over CPU-only inference. Furthermore, the inference speed was increased to 8 tokens per second when utilizing 7GB of VRAM.
Live animation has gained immense popularity for enhancing online engagement, yet achieving high-quality, real-time, and stable animation with diffusion models remains challenging, especially on consumer-grade GPUs. Existing methods struggle with generating long, consistent video streams efficiently, often being limited by latency issues and degraded visual quality over extended periods. In this paper, we introduce RAIN, a pipeline solution capable of animating infinite video streams in real-time with low latency using a single RTX 4090 GPU. The core idea of RAIN is to efficiently compute frame-token attention across different noise levels and long time-intervals while simultaneously denoising a significantly larger number of frame-tokens than previous stream-based methods. This design allows RAIN to generate video frames with much shorter latency and faster speed, while maintaining long-range attention over extended video streams, resulting in enhanced continuity and consistency. Consequently, a Stable Diffusion model fine-tuned with RAIN in just a few epochs can produce video streams in real-time and low latency without much compromise in quality or consistency, up to infinite long. Despite its advanced capabilities, the RAIN only introduces a few additional 1D attention blocks, imposing minimal additional burden. Experiments in benchmark datasets and generating super-long videos demonstrating that RAIN can animate characters in real-time with much better quality, accuracy, and consistency than competitors while costing less latency. All code and models will be made publicly available.
The evolution of colour vision is captivating, as it reveals the adaptive strategies of extinct species while simultaneously inspiring innovations in modern imaging technology. In this study, we present a simplified model of visual transduction in the retina, introducing a novel opsin layer. We quantify evolutionary pressures by measuring machine vision recognition accuracy on colour images shaped by specific opsins. Building on this, we develop an evolutionary conservation optimisation algorithm to reconstruct the spectral sensitivity of opsins, enabling mutation-driven adaptations to to more effectively spot fruits or predators. This model condenses millions of years of evolution within seconds on GPU, providing an experimental framework to test long-standing hypotheses in evolutionary biology , such as vision of early mammals, primate trichromacy from gene duplication, retention of colour blindness, blue-shift of fish rod and multiple rod opsins with bioluminescence. Moreover, the model enables speculative explorations of hypothetical species, such as organisms with eyes adapted to the conditions on Mars. Our findings suggest a minimalist yet effective approach to task-specific camera filter design, optimising the spectral response function to meet application-driven demands. The code will be made publicly available upon acceptance.
Vision-Language Models (VLMs) have enabled a variety of real-world applications. The large parameter size of VLMs brings large memory and computation overhead which poses significant challenges for deployment. Post-Training Quantization (PTQ) is an effective technique to reduce the memory and computation overhead. Existing PTQ methods mainly focus on large language models (LLMs), without considering the differences across other modalities. In this paper, we discover that there is a significant difference in sensitivity between language and vision tokens in large VLMs. Therefore, treating tokens from different modalities equally, as in existing PTQ methods, may over-emphasize the insensitive modalities, leading to significant accuracy loss. To deal with the above issue, we propose a simple yet effective method, Modality-Balanced Quantization (MBQ), for large VLMs. Specifically, MBQ incorporates the different sensitivities across modalities during the calibration process to minimize the reconstruction loss for better quantization parameters. Extensive experiments show that MBQ can significantly improve task accuracy by up to 4.4% and 11.6% under W3 and W4A8 quantization for 7B to 70B VLMs, compared to SOTA baselines. Additionally, we implement a W3 GPU kernel that fuses the dequantization and GEMV operators, achieving a 1.4x speedup on LLaVA-onevision-7B on the RTX 4090. The code is available at https://github.com/thu-nics/MBQ.
We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance. We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully harness its capabilities. Comprehensive evaluations reveal that DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models. Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training. In addition, its training process is remarkably stable. Throughout the entire training process, we did not experience any irrecoverable loss spikes or perform any rollbacks. The model checkpoints are available at https://github.com/deepseek-ai/DeepSeek-V3.