With the advancements in long-context inference capabilities of large language models (LLMs), the KV cache has become one of the foundational components. However, its substantial GPU memory consumption makes KV cache compression a key technique for enabling efficient LLM inference in industrial scenarios. While recent studies have focused on optimizing the memory occupied by the KV cache, they overlook two critical factors: preserving semantic coherence and considering task-specific characteristic during compression. To address these limitations, we propose a novel task-adaptive KV cache window selection method, WindowKV. WindowKV dynamically selects local semantic windows consisting of consecutive tokens, according to task-specific characteristics, ensuring the retained KV cache captures continuous, essential context. Additionally, we introduce an intra-group layer KV cache indices sharing strategy to reduce computational overhead, achieving a balance between performance and efficiency. We rigorously evaluate WindowKV on the LongBench benchmark, and the results demonstrate that it maintains a performance comparable to full KV cache retention while using only 12% of the original KV cache, significantly reducing memory requirements. Furthermore, our method also achieves state-of-the-art results in the Needle-in-a-Haystack evaluation, highlighting its effectiveness and robustness.
Self-supervised adaptation (SSA) improves foundation model transfer to medical domains but is computationally prohibitive. Although parameter efficient fine-tuning methods such as LoRA have been explored for supervised adaptation, their effectiveness for SSA remains unknown. In this work, we introduce efficient self-supervised adaptation (ESSA), a framework that applies parameter-efficient fine-tuning techniques to SSA with the aim of reducing computational cost and improving adaptation performance. Among the methods tested, Attention Projection Layer Adaptation (APLA) sets a new state-of-the-art, consistently surpassing full-parameter SSA and supervised fine-tuning across diverse medical tasks, while reducing GPU memory by up to 40.1% and increasing training throughput by 25.2%, all while maintaining inference efficiency.
Recognizing arbitrary or previously unseen categories is essential for comprehensive real-world 3D scene understanding. Currently, all existing methods rely on 2D or textual modalities during training, or together at inference. This highlights a clear absence of a model capable of processing 3D data alone for learning semantics end-to-end, along with the necessary data to train such a model. Meanwhile, 3D Gaussian Splatting (3DGS) has emerged as the de facto standard for 3D scene representation across various vision tasks. However, effectively integrating semantic reasoning into 3DGS in a generalizable fashion remains an open challenge. To address these limitations we introduce SceneSplat, to our knowledge the first large-scale 3D indoor scene understanding approach that operates natively on 3DGS. Furthermore, we propose a self-supervised learning scheme that unlocks rich 3D feature learning from unlabeled scenes. In order to power the proposed methods, we introduce SceneSplat-7K, the first large-scale 3DGS dataset for indoor scenes, comprising of 6868 scenes derived from 7 established datasets like ScanNet, Matterport3D, etc. Generating SceneSplat-7K required computational resources equivalent to 119 GPU-days on an L4 GPU, enabling standardized benchmarking for 3DGS-based reasoning for indoor scenes. Our exhaustive experiments on SceneSplat-7K demonstrate the significant benefit of the proposed methods over the established baselines.
Understanding the geometric and semantic properties of the scene is crucial in autonomous navigation and particularly challenging in the case of Unmanned Aerial Vehicle (UAV) navigation. Such information may be by obtained by estimating depth and semantic segmentation maps of the surrounding environment and for their practical use in autonomous navigation, the procedure must be performed as close to real-time as possible. In this paper, we leverage monocular cameras on aerial robots to predict depth and semantic maps in low-altitude unstructured environments. We propose a joint deep-learning architecture that can perform the two tasks accurately and rapidly, and validate its effectiveness on MidAir and Aeroscapes benchmark datasets. Our joint-architecture proves to be competitive or superior to the other single and joint architecture methods while performing its task fast predicting 20.2 FPS on a single NVIDIA quadro p5000 GPU and it has a low memory footprint. All codes for training and prediction can be found on this link: https://github.com/Malga-Vision/Co-SemDepth
LiDARs provide accurate geometric measurements, making them valuable for ego-motion estimation and reconstruction tasks. Although its success, managing an accurate and lightweight representation of the environment still poses challenges. Both classic and NeRF-based solutions have to trade off accuracy over memory and processing times. In this work, we build on recent advancements in Gaussian Splatting methods to develop a novel LiDAR odometry and mapping pipeline that exclusively relies on Gaussian primitives for its scene representation. Leveraging spherical projection, we drive the refinement of the primitives uniquely from LiDAR measurements. Experiments show that our approach matches the current registration performance, while achieving SOTA results for mapping tasks with minimal GPU requirements. This efficiency makes it a strong candidate for further exploration and potential adoption in real-time robotics estimation tasks.
The growing adoption of long-context Large Language Models (LLMs) has introduced significant memory and computational challenges in autoregressive decoding due to the expanding Key-Value (KV) cache. KV cache quantization has emerged as a promising solution, with prior work showing that 4-bit or even 2-bit quantization can maintain model accuracy while reducing memory costs. However, despite these benefits, preliminary implementations for the low-bit KV cache struggle to deliver the expected speedup due to quantization and dequantization overheads and the lack of Tensor Cores utilization. In this work, we propose BitDecoding, a GPU-optimized framework that unlocks Tensor Cores for efficient decoding with low-bit KV cache. Efficiently leveraging Tensor Cores for low-bit KV cache is challenging due to the dynamic nature of KV cache generation at each decoding step. BitDecoding addresses these challenges with a Tensor Cores-Centric BitFusion Scheme that ensures data layout compatibility to enable high utilization of Tensor Cores. Additionally, BitDecoding incorporates a warp-efficient parallel decoding kernel and a fine-grained asynchronous pipeline, minimizing dequantization overhead and improving computational efficiency. Experiments show that BitDecoding achieves up to 7.5x speedup on RTX 4090, 4.8x on A100, and 8.9x on H100, compared to FP16 FlashDecoding-v2. It also outperforms the state-of-the-art low-bit KV cache implementation (QServe) by up to 4.3x. On LLaMA-3.1-8B with a 128K sequence length, BitDecoding reduces single-batch decoding latency by 3x, demonstrating its effectiveness in long-context generation scenarios. The code is available at https://github.com/DD-DuDa/BitDecoding.
The emergent dynamics of complex gene regulatory networks govern various cellular processes. However, understanding these dynamics is challenging due to the difficulty of parameterizing the computational models for these networks, especially as the network size increases. Here, we introduce a simulation library, Gene Regulatory Interaction Network Simulator (GRiNS), to address these challenges. GRiNS integrates popular parameter-agnostic simulation frameworks, RACIPE and Boolean Ising formalism, into a single Python library capable of leveraging GPU acceleration for efficient and scalable simulations. GRiNS extends the ordinary differential equations (ODE) based RACIPE framework with a more modular design, allowing users to choose parameters, initial conditions, and time-series outputs for greater customisability and accuracy in simulations. For large networks, where ODE-based simulation formalisms do not scale well, GRiNS implements Boolean Ising formalism, providing a simplified, coarse-grained alternative, significantly reducing the computational cost while capturing key dynamical behaviours of large regulatory networks. The documentation and installation instructions for GRiNS can be found at https://moltenecdysone09.github.io/GRiNS/.
The advancement of remote sensing technology has improved the spatial resolution of satellite imagery, facilitating more detailed visual representations for diverse interpretations. However, existing methods exhibit limited generalization capabilities across varied applications. While some contemporary foundation models demonstrate potential, they are hindered by insufficient cross-task adaptability and primarily process low-resolution imagery of restricted sizes, thus failing to fully exploit high-resolution data or leverage comprehensive large-scene semantics. Crucially, remote sensing imagery differs fundamentally from natural images, as key foreground targets (eg., maritime objects, artificial structures) often occupy minimal spatial proportions (~1%) and exhibit sparse distributions. Efficiently modeling cross-task generalizable knowledge from lengthy 2D tokens (~100,000) poses a significant challenge yet remains critical for remote sensing image understanding. Motivated by the selective attention mechanisms inherent to the human visual system, we propose DynamicVis, a dynamic visual perception foundation model for remote sensing imagery. The framework integrates a novel dynamic region perception backbone based on the selective state space model, which strategically balances localized detail extraction with global contextual integration, enabling computationally efficient encoding of large-scale data while maintaining architectural scalability. To enhance cross-task knowledge transferring, we introduce a multi-instance learning paradigm utilizing meta-embedding representations, trained on million-scale region-level annotations. Evaluations across nine downstream tasks demonstrate the model's versatility. DynamicVis achieves multi-level feature modeling with exceptional efficiency, processing (2048x2048) pixels with 97 ms latency (6% of ViT's) and 833 MB GPU memory (3% of ViT's).
Linear RNNs with gating recently demonstrated competitive performance compared to Transformers in language modeling. Although their linear compute scaling in sequence length offers theoretical runtime advantages over Transformers, realizing these benefits in practice requires optimized custom kernels, as Transformers rely on the highly efficient Flash Attention kernels (Dao, 2024). Leveraging the chunkwise-parallel formulation of linear RNNs, Flash Linear Attention (FLA) (Yang & Zhang, 2024) shows that linear RNN kernels are faster than Flash Attention, by parallelizing over chunks of the input sequence. However, since the chunk size of FLA is limited, many intermediate states must be materialized in GPU memory. This leads to low arithmetic intensity and causes high memory consumption and IO cost, especially for long-context pre-training. In this work, we present Tiled Flash Linear Attention (TFLA), a novel kernel algorithm for linear RNNs, that enables arbitrary large chunk sizes and high arithmetic intensity by introducing an additional level of sequence parallelization within each chunk. First, we apply TFLA to the xLSTM with matrix memory, the mLSTM (Beck et al., 2024). Second, we propose an mLSTM variant with sigmoid input gate and reduced computation for even faster kernel runtimes at equal language modeling performance. In our speed benchmarks, we show that our new mLSTM kernels based on TFLA outperform highly optimized Flash Attention, Linear Attention and Mamba kernels, setting a new state of the art for efficient long-context sequence modeling primitives.
Leading large language models have demonstrated impressive capabilities in reasoning-intensive tasks, such as standardized educational testing. However, they often require extensive training in low-resource settings with inaccessible infrastructure. Small or compact models, though more efficient, frequently lack sufficient support for underrepresented languages, leaving a performance gap in critical domains. This work explores the potential of parameter-efficient fine-tuning of compact open-weight language models to handle reasoning-intensive tasks in the underrepresented Ukrainian language, building on the findings of the ZNO-Eval benchmark. Parameter-efficient fine-tuning of LLaMA 3.1 (8 billion parameters), LLaMA 3.2 (3 billion parameters), and Gemma 2 (9 billion parameters) models on chain-of-thought solutions resulted in a modest test score improvement of up to 17.4% on complex matching tasks and 1.6% overall compared to tuning on answer letters alone, offering enhanced interpretability and robustness. In addition, the proposed tuning method with joint task topic and step-by-step solution generation outperforms standard chain-of-thought tuning in matching tasks and provides a 5.4% gain over the best LLaMA 3.2 model due to guiding the model to recall and apply domain-relevant information. Contrasting obtained results with zero-shot evaluations of leading open-weight and proprietary models such as Qwen, DeepSeek R1, OpenAI o1 and o3, Gemini, and Claude, highlight that fine-tuning LLaMA and Gemma models with 2,032 step-by-step solutions and 20 to 50 million trainable parameters on a single A100 GPU lets them outperform GPT-4o mini, Mistral Large, and larger open-weight models. This research also evaluates how merging the quantized adapter with the base model influences the generation quality. Source code and tuned models are available at https://github.com/NLPForUA/ZNO.
Top-leading solutions for Video Scene Graph Generation (VSGG) typically adopt an offline pipeline. Though demonstrating promising performance, they remain unable to handle real-time video streams and consume large GPU memory. Moreover, these approaches fall short in temporal reasoning, merely aggregating frame-level predictions over a temporal context. In response, we introduce DIFFVSGG, an online VSGG solution that frames this task as an iterative scene graph update problem. Drawing inspiration from Latent Diffusion Models (LDMs) which generate images via denoising a latent feature embedding, we unify the decoding of object classification, bounding box regression, and graph generation three tasks using one shared feature embedding. Then, given an embedding containing unified features of object pairs, we conduct a step-wise Denoising on it within LDMs, so as to deliver a clean embedding which clearly indicates the relationships between objects. This embedding then serves as the input to task-specific heads for object classification, scene graph generation, etc. DIFFVSGG further facilitates continuous temporal reasoning, where predictions for subsequent frames leverage results of past frames as the conditional inputs of LDMs, to guide the reverse diffusion process for current frames. Extensive experiments on three setups of Action Genome demonstrate the superiority of DIFFVSGG.
Despite the substantial advancements demonstrated by learning-based neural models in the LiDAR Point Cloud Compression (LPCC) task, realizing real-time compression - an indispensable criterion for numerous industrial applications - remains a formidable challenge. This paper proposes RENO, the first real-time neural codec for 3D LiDAR point clouds, achieving superior performance with a lightweight model. RENO skips the octree construction and directly builds upon the multiscale sparse tensor representation. Instead of the multi-stage inferring, RENO devises sparse occupancy codes, which exploit cross-scale correlation and derive voxels' occupancy in a one-shot manner, greatly saving processing time. Experimental results demonstrate that the proposed RENO achieves real-time coding speed, 10 fps at 14-bit depth on a desktop platform (e.g., one RTX 3090 GPU) for both encoding and decoding processes, while providing 12.25% and 48.34% bit-rate savings compared to G-PCCv23 and Draco, respectively, at a similar quality. RENO model size is merely 1MB, making it attractive for practical applications. The source code is available at https://github.com/NJUVISION/RENO.
Fine-tuning large pre-trained LLMs generally demands extensive GPU memory. Traditional first-order optimizers like SGD encounter substantial difficulties due to increased memory requirements from storing activations and gradients during both the forward and backward phases as the model size expands. Alternatively, zeroth-order (ZO) techniques can compute gradients using just forward operations, eliminating the need to store activations. Furthermore, by leveraging CPU capabilities, it's feasible to enhance both the memory and processing power available to a single GPU. We propose a novel framework, ZO2 (Zeroth-Order Offloading), for efficient zeroth-order fine-tuning of LLMs with only limited GPU memory. Our framework dynamically shifts model parameters between the CPU and GPU as required, optimizing computation flow and maximizing GPU usage by minimizing downtime. This integration of parameter adjustments with ZO's double forward operations reduces unnecessary data movement, enhancing the fine-tuning efficacy. Additionally, our framework supports an innovative low-bit precision approach in AMP mode to streamline data exchanges between the CPU and GPU. Employing this approach allows us to fine-tune extraordinarily large models, such as the OPT-175B with more than 175 billion parameters, on a mere 18GB GPU--achievements beyond the reach of traditional methods. Moreover, our framework achieves these results with almost no additional time overhead and absolutely no accuracy loss compared to standard zeroth-order methods. ZO2's code has been open-sourced in https://github.com/liangyuwang/zo2.
With the advancement of RNN models with linear complexity, the quadratic complexity challenge of transformers has the potential to be overcome. Notably, the emerging Mamba-2 has demonstrated competitive performance, bridging the gap between RNN models and transformers. However, due to sequential processing and vanishing gradients, RNN models struggle to capture long-range dependencies, limiting contextual understanding. This results in slow convergence, high resource demands, and poor performance on downstream understanding and complex reasoning tasks. In this work, we present a hybrid model MaTVLM by substituting a portion of the transformer decoder layers in a pre-trained VLM with Mamba-2 layers. Leveraging the inherent relationship between attention and Mamba-2, we initialize Mamba-2 with corresponding attention weights to accelerate convergence. Subsequently, we employ a single-stage distillation process, using the pre-trained VLM as the teacher model to transfer knowledge to the MaTVLM, further enhancing convergence speed and performance. Furthermore, we investigate the impact of differential distillation loss within our training framework. We evaluate the MaTVLM on multiple benchmarks, demonstrating competitive performance against the teacher model and existing VLMs while surpassing both Mamba-based VLMs and models of comparable parameter scales. Remarkably, the MaTVLM achieves up to 3.6x faster inference than the teacher model while reducing GPU memory consumption by 27.5%, all without compromising performance. Code and models are released at http://github.com/hustvl/MaTVLM.
Existing adaptation techniques typically require architectural modifications or added parameters, leading to high computational costs and complexity. We introduce Attention Projection Layer Adaptation (APLA), a simple approach to adapt vision transformers (ViTs) without altering the architecture or adding parameters. Through a systematic analysis, we find that the layer immediately after the attention mechanism is crucial for adaptation. By updating only this projection layer, or even just a random subset of this layer's weights, APLA achieves state-of-the-art performance while reducing GPU memory usage by up to 52.63% and training time by up to 43.0%, with no extra cost at inference. Across 46 datasets covering a variety of tasks including scene classification, medical imaging, satellite imaging, and fine-grained classification, APLA consistently outperforms 17 other leading adaptation methods, including full fine-tuning, on classification, segmentation, and detection tasks. The code is available at https://github.com/MoeinSorkhei/APLA.
The growing demand for computational power is driven by advancements in deep learning, the increasing need for big data processing, and the requirements of scientific simulations for academic and research purposes. Developing countries like Nepal often struggle with the resources needed to invest in new and better hardware for these purposes. However, optimizing and building on existing technology can still meet these computing demands effectively. To address these needs, we built a cluster using four NVIDIA GeForce GTX 1650 GPUs. The cluster consists of four nodes: one master node that controls and manages the entire cluster, and three compute nodes dedicated to processing tasks. The master node is equipped with all necessary software for package management, resource scheduling, and deployment, such as Anaconda and Slurm. In addition, a Network File Storage (NFS) system was integrated to provide the additional storage required by the cluster. Given that the cluster is accessible via ssh by a public domain address, which poses significant cybersecurity risks, we implemented fail2ban to mitigate brute force attacks and enhance security. Despite the continuous challenges encountered during the design and implementation process, this project demonstrates how powerful computational clusters can be built to handle resource-intensive tasks in various demanding fields.
Performance modeling, a pivotal domain in program cost analysis, currently relies on manually crafted models constrained by various program and hardware limitations, especially in the intricate landscape of GPGPU. Meanwhile, Large Language Models (LLMs) have demonstrated their effectiveness in addressing diverse programming challenges. Our work establishes a connection between LLMs and performance modeling, employing the LLM as a performance estimator. Through experimental exploration with carefully designed large-scale OpenCL datasets, we highlight the potential capability as well as the main difficulties of using LLMs in handling performance modeling tasks for OpenCL device source programs. As the first study for this line of work, our LLM-based performance model achieves a mean absolute percentage error of $24.25\%$ for a large-scale generated validation set. On a set of publicly available OpenCL programs, our model achieves a mean absolute percentage error of $46.1\%$.
To overcome computational challenges of Optimal Transport (OT), several variants of Sliced Wasserstein (SW) has been developed in the literature. These approaches exploit the closed-form expression of the univariate OT by projecting measures onto (one-dimensional) lines. However, projecting measures onto low-dimensional spaces can lead to a loss of topological information. Tree-Sliced Wasserstein distance on Systems of Lines (TSW-SL) has emerged as a promising alternative that replaces these lines with a more advanced structure called tree systems. The tree structures enhance the ability to capture topological information of the metric while preserving computational efficiency. However, at the core of TSW-SL, the splitting maps, which serve as the mechanism for pushing forward measures onto tree systems, focus solely on the position of the measure supports while disregarding the projecting domains. Moreover, the specific splitting map used in TSW-SL leads to a metric that is not invariant under Euclidean transformations, a typically expected property for OT on Euclidean space. In this work, we propose a novel class of splitting maps that generalizes the existing one studied in TSW-SL enabling the use of all positional information from input measures, resulting in a novel Distance-based Tree-Sliced Wasserstein (Db-TSW) distance. In addition, we introduce a simple tree sampling process better suited for Db-TSW, leading to an efficient GPU-friendly implementation for tree systems, similar to the original SW. We also provide a comprehensive theoretical analysis of proposed class of splitting maps to verify the injectivity of the corresponding Radon Transform, and demonstrate that Db-TSW is an Euclidean invariant metric. We empirically show that Db-TSW significantly improves accuracy compared to recent SW variants while maintaining low computational cost via a wide range of experiments.
Deep learning methods for point tracking are applicable in 2D echocardiography, but do not yet take advantage of domain specifics that enable extremely fast and efficient configurations. We developed MyoTracker, a low-complexity architecture (0.3M parameters) for point tracking in echocardiography. It builds on the CoTracker2 architecture by simplifying its components and extending the temporal context to provide point predictions for the entire sequence in a single step. We applied MyoTracker to the right ventricular (RV) myocardium in RV-focused recordings and compared the results with those of CoTracker2 and EchoTracker, another specialized point tracking architecture for echocardiography. MyoTracker achieved the lowest average point trajectory error at 2.00 $\pm$ 0.53 mm. Calculating RV Free Wall Strain (RV FWS) using MyoTracker's point predictions resulted in a -0.3$\%$ bias with 95$\%$ limits of agreement from -6.1$\%$ to 5.4$\%$ compared to reference values from commercial software. This range falls within the interobserver variability reported in previous studies. The limits of agreement were wider for both CoTracker2 and EchoTracker, worse than the interobserver variability. At inference, MyoTracker used 67$\%$ less GPU memory than CoTracker2 and 84$\%$ less than EchoTracker on large sequences (100 frames). MyoTracker was 74 times faster during inference than CoTracker2 and 11 times faster than EchoTracker with our setup. Maintaining the entire sequence in the temporal context was the greatest contributor to MyoTracker's accuracy. Slight additional gains can be made by re-enabling iterative refinement, at the cost of longer processing time.
Recent advancements in unified multimodal understanding and visual generation (or multimodal generation) models have been hindered by their quadratic computational complexity and dependence on large-scale training data. We present OmniMamba, the first linear-architecture-based multimodal generation model that generates both text and images through a unified next-token prediction paradigm. The model fully leverages Mamba-2's high computational and memory efficiency, extending its capabilities from text generation to multimodal generation. To address the data inefficiency of existing unified models, we propose two key innovations: (1) decoupled vocabularies to guide modality-specific generation, and (2) task-specific LoRA for parameter-efficient adaptation. Furthermore, we introduce a decoupled two-stage training strategy to mitigate data imbalance between two tasks. Equipped with these techniques, OmniMamba achieves competitive performance with JanusFlow while surpassing Show-o across benchmarks, despite being trained on merely 2M image-text pairs, which is 1,000 times fewer than Show-o. Notably, OmniMamba stands out with outstanding inference efficiency, achieving up to a 119.2 times speedup and 63% GPU memory reduction for long-sequence generation compared to Transformer-based counterparts. Code and models are released at https://github.com/hustvl/OmniMamba