The rapid growth of LLMs has revolutionized natural language processing and AI analysis, but their increasing size and memory demands present significant challenges. A common solution is to spill over to CPU memory; however, traditional GPU-CPU memory swapping often results in higher latency and lower throughput. This paper introduces Pie, an LLM inference framework that addresses these challenges with performance-transparent swapping and adaptive expansion. By leveraging predictable memory access patterns and the high bandwidth of modern hardware like the NVIDIA GH200 Grace Hopper Superchip, Pie enables concurrent data swapping without affecting foreground computation, expanding effective memory without added latency. Adaptive expansion dynamically adjusts CPU memory allocation based on real-time information, optimizing memory usage and performance under varying conditions. Pie maintains low computation latency, high throughput, and high elasticity. Our experimental evaluation demonstrates that Pie achieves optimal swapping policy during cache warmup and effectively balances increased memory capacity with negligible impact on computation. With its extended capacity, Pie outperforms vLLM by up to 1.9X in throughput and 2X in latency. Additionally, Pie can reduce GPU memory usage by up to 1.67X while maintaining the same performance. Compared to FlexGen, an offline profiling-based swapping solution, Pie achieves magnitudes lower latency and 9.4X higher throughput.
Grasping unknown objects with multi-fingered hands at high success rates and in real-time is an unsolved problem. Existing methods are limited in the speed of grasp synthesis or the ability to synthesize a variety of grasps from the same observation. We introduce Five-finger Hand Net (FFHNet), an ML model which can generate a wide variety of high-quality multi-fingered grasps for unseen objects from a single view. Generating and evaluating grasps with FFHNet takes only 30ms on a commodity GPU. To the best of our knowledge, FFHNet is the first ML-based real-time system for multi-fingered grasping with the ability to perform grasp inference at 30 frames per second (FPS). For training, we synthetically generate 180k grasp samples for 129 objects. We are able to achieve 91% grasping success for unknown objects in simulation and we demonstrate the model's capabilities of synthesizing high-quality grasps also for real unseen objects.
Wireless ray-tracing (RT) is emerging as a key tool for three-dimensional (3D) wireless channel modeling, driven by advances in graphical rendering. Current approaches struggle to accurately model beyond 5G (B5G) network signaling, which often operates at higher frequencies and is more susceptible to environmental conditions and changes. Existing online learning solutions require real-time environmental supervision during training, which is both costly and incompatible with GPU-based processing. In response, we propose a novel approach that redefines ray trajectory generation as a sequential decision-making problem, leveraging generative models to jointly learn the optical, physical, and signal properties within each designated environment. Our work introduces the Scene-Aware Neural Decision Wireless Channel Raytracing Hierarchy (SANDWICH), an innovative offline, fully differentiable approach that can be trained entirely on GPUs. SANDWICH offers superior performance compared to existing online learning methods, outperforms the baseline by 4e^-2 radian in RT accuracy, and only fades 0.5 dB away from toplined channel gain estimation.
Audio super-resolution aims to enhance low-resolution signals by creating high-frequency content. In this work, we modify the architecture of AERO (a state-of-the-art system for this task) for music super-resolution. SPecifically, we replace its original Attention and LSTM layers with Mamba, a State Space Model (SSM), across all network layers. Mamba is capable of effectively substituting the mentioned modules, as it offers a mechanism similar to that of Attention while also functioning as a recurrent network. With the proposed AEROMamba, training requires 2-4x less GPU memory, since Mamba exploits the convolutional formulation and leverages GPU memory hierarchy. Additionally, during inference, Mamba operates in constant memory due to recurrence, avoiding memory growth associated with Attention. This results in a 14x speed improvement using 5x less GPU. Subjective listening tests (0 to 100 scale) show that the proposed model surpasses the AERO model. In the MUSDB dataset, degraded signals scored 38.22, while AERO and AEROMamba scored 60.03 and 66.74, respectively. For the PianoEval dataset, scores were 72.92 for degraded signals, 76.89 for AERO, and 84.41 for AEROMamba.
With the increase in the number of parameters in large language models, the process of pre-training and fine-tuning increasingly demands larger volumes of GPU memory. A significant portion of this memory is typically consumed by the optimizer state. To overcome this challenge, recent approaches such as low-rank adaptation (LoRA (Hu et al., 2021)), low-rank gradient projection (GaLore (Zhao et al., 2024)), and blockwise optimization (BAdam (Luo et al., 2024)) have been proposed. However, in all these algorithms, the $\textit{effective rank of the weight updates remains low-rank}$, which can lead to a substantial loss of information from the gradient. This loss can be critically important, especially during the pre-training stage. In this paper, we introduce $\texttt{FRUGAL}$ ($\textbf{F}$ull-$\textbf{R}$ank $\textbf{U}$pdates with $\textbf{G}$r$\textbf{A}$dient sp$\textbf{L}$itting), a new memory-efficient optimization framework. $\texttt{FRUGAL}$ leverages gradient splitting to perform low-dimensional updates using advanced algorithms (such as Adam), while updates along the remaining directions are executed via state-free methods like SGD or signSGD (Bernstein et al., 2018). Our framework can be integrated with various low-rank update selection techniques, including GaLore and BAdam. We provide theoretical convergence guarantees for our framework when using SGDM for low-dimensional updates and SGD for state-free updates. Additionally, our method consistently outperforms concurrent approaches across various fixed memory budgets, achieving state-of-the-art results in pre-training and fine-tuning tasks while balancing memory efficiency and performance metrics.
Inverse Lithography Technology (ILT) has emerged as a promising solution for photo mask design and optimization. Relying on multi-beam mask writers, ILT enables the creation of free-form curvilinear mask shapes that enhance printed wafer image quality and process window. However, a major challenge in implementing curvilinear ILT for large-scale production is mask rule checking, an area currently under development by foundries and EDA vendors. Although recent research has incorporated mask complexity into the optimization process, much of it focuses on reducing e-beam shots, which does not align with the goals of curvilinear ILT. In this paper, we introduce a GPU-accelerated ILT algorithm that improves not only contour quality and process window but also the precision of curvilinear mask shapes. Our experiments on open benchmarks demonstrate a significant advantage of our algorithm over leading academic ILT engines.
LLMs are increasingly used inside database systems and in database applications for better complexity management and decision-making, where LLM inferences require significant GPU costs. LLM inference systems, however, are slow compared to database systems, limiting the expansion of the use of LLMs inside database systems. This paper first analyzes the LLM inference performance and focuses on a data management issue in LLM inference. We reveal that the root of the problem is the lack of an adequate resource cost model and optimization strategy when executing multiple concurrent inference requests. We adapt classic database multi-query optimization techniques by introducing cost models for concurrent inference requests and new scheduling strategies to optimize the use of memory resources by concurrent requests, thereby substantially improving performance.
GraphRAG integrates (knowledge) graphs with large language models (LLMs) to improve reasoning accuracy and contextual relevance. Despite its promising applications and strong relevance to multiple research communities, such as databases and natural language processing, GraphRAG currently lacks modular workflow analysis, systematic solution frameworks, and insightful empirical studies. To bridge these gaps, we propose LEGO-GraphRAG, a modular framework that enables: 1) fine-grained decomposition of the GraphRAG workflow, 2) systematic classification of existing techniques and implemented GraphRAG instances, and 3) creation of new GraphRAG instances. Our framework facilitates comprehensive empirical studies of GraphRAG on large-scale real-world graphs and diverse query sets, revealing insights into balancing reasoning quality, runtime efficiency, and token or GPU cost, that are essential for building advanced GraphRAG systems.
The rapid development of large language models (LLMs) like Llama has significantly advanced information retrieval (IR) systems. However, using LLMs for long documents, as in RankLLaMA, remains challenging due to computational complexity, especially concerning input token length. Furthermore, the internal mechanisms of LLMs during ranking are still not fully understood. In this paper, we first explore the internal workings of LLMs during relevance judgement and identify that specific attention heads play a crucial role in aligning relevant tokens. This observation inspires us to revisit the block pre-ranking strategy used in KeyB, which remains state-of-the-art (SOTA) on the TREC 2019 DL document ranking dataset. Building on these insights, we develop KeyB2, an advanced long document IR approach that integrates block pre-ranking with the performance of LLMs. KeyB2 efficiently identifies and processes the most relevant blocks, reducing computational costs and improving ranking effectiveness. Additionally, we introduce a new bi-encoder block matching strategy for KeyB2. Comprehensive experiments on long-document datasets, including TREC 2019 DL, Robust04, and MLDR-zh, show that KeyB2 outperforms baselines like RankLLaMA and KeyB by reducing reranking time and GPU memory usage while enhancing retrieval performance, achieving new SOTA results on TREC 2019 DL with higher NDCG@10 and MAP scores.
OpenThaiGPT 1.5 is an advanced Thai language chat model based on Qwen v2.5, finetuned on over 2,000,000 Thai instruction pairs. This report provides an engineering perspective on the model's development, capabilities, and performance. We discuss the model's architecture, training process, and key features, including multi-turn conversation support, Retrieval Augmented Generation (RAG) compatibility, and tool-calling functionality. Benchmark results demonstrate OpenThaiGPT 1.5's state-of-the-art performance on various Thai language tasks, outperforming other open-source Thai language models. We also address practical considerations such as GPU memory requirements and deployment strategies.
In large language model (LLM) training, several parallelization strategies, including Tensor Parallelism (TP), Pipeline Parallelism (PP), Data Parallelism (DP), as well as Sequence Parallelism (SP) and Context Parallelism (CP), are employed to distribute model parameters, activations, and optimizer states across devices. Identifying the optimal parallelization configuration for each environment while avoiding GPU memory overflow remains a challenging task. In this study, we provide precise formulas to estimate the memory consumed by parameters, gradients, optimizer states, and activations for 4D parallel training (DP, TP, PP, CP) in the Llama architecture. We conducted 454 experiments on A100 and H100 GPUs, incorporating often neglected factors such as temporary buffers and memory fragmentation into our analysis. Results indicate that when the estimated memory usage is below 80\% of the available GPU memory, the training never encounters out-of-memory errors. This simple yet effective formula allows us to identify parallelization configurations that could lead to memory overflow in advance, significantly reducing the configuration search space. Additionally, through a comprehensive exploration of optimal configurations in 4D parallelism, our analysis of the 454 experimental results provides empirical insights into optimal 4D parallelism configurations.
Deep learning has shown great promise for MRI reconstruction from undersampled data, yet there is a lack of research on validating its performance in 3D parallel imaging acquisitions with non-Cartesian undersampling. In addition, the artifacts and the resulting image quality depend on the under-sampling pattern. To address this uncharted territory, we extend the Non-Cartesian Primal-Dual Network (NC-PDNet), a state-of-the-art unrolled neural network, to a 3D multi-coil setting. We evaluated the impact of channel-specific versus channel-agnostic training configurations and examined the effect of coil compression. Finally, we benchmark four distinct non-Cartesian undersampling patterns, with an acceleration factor of six, using the publicly available Calgary-Campinas dataset. Our results show that NC-PDNet trained on compressed data with varying input channel numbers achieves an average PSNR of 42.98 dB for 1 mm isotropic 32 channel whole-brain 3D reconstruction. With an inference time of 4.95sec and a GPU memory usage of 5.49 GB, our approach demonstrates significant potential for clinical research application.
Diffusion Probabilistic Models (DPMs) have demonstrated exceptional performance in generative tasks, but this comes at the expense of sampling efficiency. To enhance sampling speed without sacrificing quality, various distillation-based accelerated sampling algorithms have been recently proposed. However, they typically require significant additional training costs and model parameter storage, which limit their practical application. In this work, we propose PCA-based Adaptive Search (PAS), which optimizes existing solvers for DPMs with minimal learnable parameters and training costs. Specifically, we first employ PCA to obtain a few orthogonal unit basis vectors to span the high-dimensional sampling space, which enables us to learn just a set of coordinates to correct the sampling direction; furthermore, based on the observation that the cumulative truncation error exhibits an ``S''-shape, we design an adaptive search strategy that further enhances the sampling efficiency and reduces the number of stored parameters to approximately 10. Extensive experiments demonstrate that PAS can significantly enhance existing fast solvers in a plug-and-play manner with negligible costs. For instance, on CIFAR10, PAS requires only 12 parameters and less than 1 minute of training on a single NVIDIA A100 GPU to optimize the DDIM from 15.69 FID (NFE=10) to 4.37.
It is now a common business practice to buy access to large language model (LLM) inference rather than self-host, because of significant upfront hardware infrastructure and energy costs. However, as a buyer, there is no mechanism to verify the authenticity of the advertised service including the serving hardware platform, e.g. that it is actually being served using an NVIDIA H100. Furthermore, there are reports suggesting that model providers may deliver models that differ slightly from the advertised ones, often to make them run on less expensive hardware. That way, a client pays premium for a capable model access on more expensive hardware, yet ends up being served by a (potentially less capable) cheaper model on cheaper hardware. In this paper we introduce \textit{\textbf{hardware and software platform inference (HSPI)}} -- a method for identifying the underlying \GPU{} architecture and software stack of a (black-box) machine learning model solely based on its input-output behavior. Our method leverages the inherent differences of various \GPU{} architectures and compilers to distinguish between different \GPU{} types and software stacks. By analyzing the numerical patterns in the model's outputs, we propose a classification framework capable of accurately identifying the \GPU{} used for model inference as well as the underlying software configuration. Our findings demonstrate the feasibility of inferring \GPU{} type from black-box models. We evaluate HSPI against models served on different real hardware and find that in a white-box setting we can distinguish between different \GPU{}s with between $83.9\%$ and $100\%$ accuracy. Even in a black-box setting we are able to achieve results that are up to three times higher than random guess accuracy.
Diffusion models have been proven highly effective at generating high-quality images. However, as these models grow larger, they require significantly more memory and suffer from higher latency, posing substantial challenges for deployment. In this work, we aim to accelerate diffusion models by quantizing their weights and activations to 4 bits. At such an aggressive level, both weights and activations are highly sensitive, where conventional post-training quantization methods for large language models like smoothing become insufficient. To overcome this limitation, we propose SVDQuant, a new 4-bit quantization paradigm. Different from smoothing which redistributes outliers between weights and activations, our approach absorbs these outliers using a low-rank branch. We first consolidate the outliers by shifting them from activations to weights, then employ a high-precision low-rank branch to take in the weight outliers with Singular Value Decomposition (SVD). This process eases the quantization on both sides. However, na\"{\i}vely running the low-rank branch independently incurs significant overhead due to extra data movement of activations, negating the quantization speedup. To address this, we co-design an inference engine Nunchaku that fuses the kernels of the low-rank branch into those of the low-bit branch to cut off redundant memory access. It can also seamlessly support off-the-shelf low-rank adapters (LoRAs) without the need for re-quantization. Extensive experiments on SDXL, PixArt-$\Sigma$, and FLUX.1 validate the effectiveness of SVDQuant in preserving image quality. We reduce the memory usage for the 12B FLUX.1 models by 3.5$\times$, achieving 3.0$\times$ speedup over the 4-bit weight-only quantized baseline on the 16GB laptop 4090 GPU, paving the way for more interactive applications on PCs. Our quantization library and inference engine are open-sourced.
This paper presents the second-placed solution for task 8 and the participation solution for task 7 of BraTS 2024. The adoption of automated brain analysis algorithms to support clinical practice is increasing. However, many of these algorithms struggle with the presence of brain lesions or the absence of certain MRI modalities. The alterations in the brain's morphology leads to high variability and thus poor performance of predictive models that were trained only on healthy brains. The lack of information that is usually provided by some of the missing MRI modalities also reduces the reliability of the prediction models trained with all modalities. In order to improve the performance of these models, we propose the use of conditional 3D wavelet diffusion models. The wavelet transform enabled full-resolution image training and prediction on a GPU with 48 GB VRAM, without patching or downsampling, preserving all information for prediction. The code for these tasks is available at https://github.com/ShadowTwin41/BraTS_2023_2024_solutions.
Accelerating molecular docking -- the process of predicting how molecules bind to protein targets -- could boost small-molecule drug discovery and revolutionize medicine. Unfortunately, current molecular docking tools are too slow to screen potential drugs against all relevant proteins, which often results in missed drug candidates or unexpected side effects occurring in clinical trials. To address this gap, we introduce RapidDock, an efficient transformer-based model for blind molecular docking. RapidDock achieves at least a $100 \times$ speed advantage over existing methods without compromising accuracy. On the Posebusters and DockGen benchmarks, our method achieves $52.1\%$ and $44.0\%$ success rates ($\text{RMSD}<2$\r{A}), respectively. The average inference time is $0.04$ seconds on a single GPU, highlighting RapidDock's potential for large-scale docking studies. We examine the key features of RapidDock that enable leveraging the transformer architecture for molecular docking, including the use of relative distance embeddings of $3$D structures in attention matrices, pre-training on protein folding, and a custom loss function invariant to molecular symmetries.
Designing quantum circuits for specific tasks is challenging due to the exponential growth of the state space. We introduce gadget reinforcement learning (GRL), which integrates reinforcement learning with program synthesis to automatically generate and incorporate composite gates (gadgets) into the action space. This enhances the exploration of parameterized quantum circuits (PQCs) for complex tasks like approximating ground states of quantum Hamiltonians, an NP-hard problem. We evaluate GRL using the transverse field Ising model under typical computational budgets (e.g., 2- 3 days of GPU runtime). Our results show improved accuracy, hardware compatibility and scalability. GRL exhibits robust performance as the size and complexity of the problem increases, even with constrained computational resources. By integrating gadget extraction, GRL facilitates the discovery of reusable circuit components tailored for specific hardware, bridging the gap between algorithmic design and practical implementation. This makes GRL a versatile framework for optimizing quantum circuits with applications in hardware-specific optimizations and variational quantum algorithms. The code is available at: https://github.com/Aqasch/Gadget_RL
The training or fine-tuning of machine learning, vision, and language models is often implemented as a pipeline: a sequence of stages encompassing data preparation, model training and evaluation. In this paper, we exploit pipeline structures to reduce the cost of hyperparameter tuning for model training/fine-tuning, which is particularly valuable for language models given their high costs in GPU-days. We propose a "memoization-aware" Bayesian Optimization (BO) algorithm, EEIPU, that works in tandem with a pipeline caching system, allowing it to evaluate significantly more hyperparameter candidates per GPU-day than other tuning algorithms. The result is better-quality hyperparameters in the same amount of search time, or equivalently, reduced search time to reach the same hyperparameter quality. In our benchmarks on machine learning (model ensembles), vision (convolutional architecture) and language (T5 architecture) pipelines, we compare EEIPU against recent BO algorithms: EEIPU produces an average of $103\%$ more hyperparameter candidates (within the same budget), and increases the validation metric by an average of $108\%$ more than other algorithms (where the increase is measured starting from the end of warm-up iterations).
High resolution is crucial for precise segmentation in fundus images, yet handling high-resolution inputs incurs considerable GPU memory costs, with diminishing performance gains as overhead increases. To address this issue while tackling the challenge of segmenting tiny objects, recent studies have explored local-global fusion methods. These methods preserve fine details using local regions and capture long-range context information from downscaled global images. However, the necessity of multiple forward passes inevitably incurs significant computational overhead, adversely affecting inference speed. In this paper, we propose HRDecoder, a simple High-Resolution Decoder network for fundus lesion segmentation. It integrates a high-resolution representation learning module to capture fine-grained local features and a high-resolution fusion module to fuse multi-scale predictions. Our method effectively improves the overall segmentation accuracy of fundus lesions while consuming reasonable memory and computational overhead, and maintaining satisfying inference speed. Experimental results on the IDRID and DDR datasets demonstrate the effectiveness of our method. Code is available at https://github.com/CVIU-CSU/HRDecoder.