This paper introduces CARSS (Cooperative Attention-guided Reinforcement Subpath Synthesis), a novel approach to address the Traveling Salesman Problem (TSP) by leveraging cooperative Multi-Agent Reinforcement Learning (MARL). CARSS decomposes the TSP solving process into two distinct yet synergistic steps: "subpath generation" and "subpath merging." In the former, a cooperative MARL framework is employed to iteratively generate subpaths using multiple agents. In the latter, these subpaths are progressively merged to form a complete cycle. The algorithm's primary objective is to enhance efficiency in terms of training memory consumption, testing time, and scalability, through the adoption of a multi-agent divide and conquer paradigm. Notably, attention mechanisms play a pivotal role in feature embedding and parameterization strategies within CARSS. The training of the model is facilitated by the independent REINFORCE algorithm. Empirical experiments reveal CARSS's superiority compared to single-agent alternatives: it demonstrates reduced GPU memory utilization, accommodates training graphs nearly 2.5 times larger, and exhibits the potential for scaling to even more extensive problem sizes. Furthermore, CARSS substantially reduces testing time and optimization gaps by approximately 50% for TSP instances of up to 1000 vertices, when compared to standard decoding methods.
Lowering the memory requirement in full-parameter training on large models has become a hot research area. MeZO fine-tunes the large language models (LLMs) by just forward passes in a zeroth-order SGD optimizer (ZO-SGD), demonstrating excellent performance with the same GPU memory usage as inference. However, the simulated perturbation stochastic approximation for gradient estimate in MeZO leads to severe oscillations and incurs a substantial time overhead. Moreover, without momentum regularization, MeZO shows severe over-fitting problems. Lastly, the perturbation-irrelevant momentum on ZO-SGD does not improve the convergence rate. This study proposes ZO-AdaMU to resolve the above problems by adapting the simulated perturbation with momentum in its stochastic approximation. Unlike existing adaptive momentum methods, we relocate momentum on simulated perturbation in stochastic gradient approximation. Our convergence analysis and experiments prove this is a better way to improve convergence stability and rate in ZO-SGD. Extensive experiments demonstrate that ZO-AdaMU yields better generalization for LLMs fine-tuning across various NLP tasks than MeZO and its momentum variants.
Autoregressive and diffusion models drive the recent breakthroughs on text-to-image generation. Despite their huge success of generating high-realistic images, a common shortcoming of these models is their high inference latency - autoregressive models run more than a thousand times successively to produce image tokens and diffusion models convert Gaussian noise into images with many hundreds of denoising steps. In this work, we explore non-autoregressive text-to-image models that efficiently generate hundreds of image tokens in parallel. We develop many model variations with different learning and inference strategies, initialized text encoders, etc. Compared with autoregressive baselines that needs to run one thousand times, our model only runs 16 times to generate images of competitive quality with an order of magnitude lower inference latency. Our non-autoregressive model with 346M parameters generates an image of 256$\times$256 with about one second on one V100 GPU.
High-definition (HD) cameras for surveillance and road traffic have experienced tremendous growth, demanding intensive computation resources for real-time analytics. Recently, offloading frames from the front-end device to the back-end edge server has shown great promise. In multi-stream competitive environments, efficient bandwidth management and proper scheduling are crucial to ensure both high inference accuracy and high throughput. To achieve this goal, we propose BiSwift, a bi-level framework that scales the concurrent real-time video analytics by a novel adaptive hybrid codec integrated with multi-level pipelines, and a global bandwidth controller for multiple video streams. The lower-level front-back-end collaborative mechanism (called adaptive hybrid codec) locally optimizes the accuracy and accelerates end-to-end video analytics for a single stream. The upper-level scheduler aims to accuracy fairness among multiple streams via the global bandwidth controller. The evaluation of BiSwift shows that BiSwift is able to real-time object detection on 9 streams with an edge device only equipped with an NVIDIA RTX3070 (8G) GPU. BiSwift improves 10%$\sim$21% accuracy and presents 1.2$\sim$9$\times$ throughput compared with the state-of-the-art video analytics pipelines.
Neural Networks can be efficiently compressed through pruning, significantly reducing storage and computational demands while maintaining predictive performance. Simple yet effective methods like Iterative Magnitude Pruning (IMP, Han et al., 2015) remove less important parameters and require a costly retraining procedure to recover performance after pruning. However, with the rise of Large Language Models (LLMs), full retraining has become infeasible due to memory and compute constraints. In this study, we challenge the practice of retraining all parameters by demonstrating that updating only a small subset of highly expressive parameters is often sufficient to recover or even improve performance compared to full retraining. Surprisingly, retraining as little as 0.27%-0.35% of the parameters of GPT-architectures achieves comparable performance to One Shot IMP across various sparsity levels. Our approach, Parameter-Efficient Retraining after Pruning (PERP), drastically reduces compute and memory demands, enabling pruning and retraining of up to 30 billion parameter models on a single NVIDIA A100 GPU within minutes. Despite magnitude pruning being considered as unsuited for pruning LLMs, our findings show that PERP positions it as a strong contender against state-of-the-art retraining-free approaches such as Wanda (Sun et al., 2023) and SparseGPT (Frantar & Alistarh, 2023), opening up a promising alternative to avoiding retraining.
Recent advancements in large language models (LLMs) boasting billions of parameters have generated a significant demand for efficient deployment in inference workloads. The majority of existing approaches rely on temporal architectures that reuse hardware units for different network layers and operators. However, these methods often encounter challenges in achieving low latency due to considerable memory access overhead. This paper investigates the feasibility and potential of model-specific spatial acceleration for LLM inference on FPGAs. Our approach involves the specialization of distinct hardware units for specific operators or layers, facilitating direct communication between them through a dataflow architecture while minimizing off-chip memory accesses. We introduce a comprehensive analytical model for estimating the performance of a spatial LLM accelerator, taking into account the on-chip compute and memory resources available on an FPGA. Through our analysis, we can determine the scenarios in which FPGA-based spatial acceleration can outperform its GPU-based counterpart. To enable more productive implementations of an LLM model on FPGAs, we further provide a library of high-level synthesis (HLS) kernels that are composable and reusable. This library will be made available as open-source. To validate the effectiveness of both our analytical model and HLS library, we have implemented BERT and GPT2 on an AMD Alveo U280 FPGA device. Experimental results demonstrate our approach can achieve up to 13.4x speedup when compared to previous FPGA-based accelerators for the BERT model. For GPT generative inference, we attain a 2.2x speedup compared to DFX, an FPGA overlay, in the prefill stage, while achieving a 1.9x speedup and a 5.7x improvement in energy efficiency compared to the NVIDIA A100 GPU in the decode stage.
Cross-modal medical image-report retrieval task plays a significant role in clinical diagnosis and various medical generative tasks. Eliminating heterogeneity between different modalities to enhance semantic consistency is the key challenge of this task. The current Vision-Language Pretraining (VLP) models, with cross-modal contrastive learning and masked reconstruction as joint training tasks, can effectively enhance the performance of cross-modal retrieval. This framework typically employs dual-stream inputs, using unmasked data for cross-modal contrastive learning and masked data for reconstruction. However, due to task competition and information interference caused by significant differences between the inputs of the two proxy tasks, the effectiveness of representation learning for intra-modal and cross-modal features is limited. In this paper, we propose an efficient VLP framework named Masked Contrastive and Reconstruction (MCR), which takes masked data as the sole input for both tasks. This enhances task connections, reducing information interference and competition between them, while also substantially decreasing the required GPU memory and training time. Moreover, we introduce a new modality alignment strategy named Mapping before Aggregation (MbA). Unlike previous methods, MbA maps different modalities to a common feature space before conducting local feature aggregation, thereby reducing the loss of fine-grained semantic information necessary for improved modality alignment. Qualitative and quantitative experiments conducted on the MIMIC-CXR dataset validate the effectiveness of our approach, demonstrating state-of-the-art performance in medical cross-modal retrieval tasks.
Pre-trained models (PTMs) have achieved great success in various Software Engineering (SE) downstream tasks following the ``pre-train then fine-tune'' paradigm. As fully fine-tuning all parameters of PTMs can be computationally expensive, a widely used solution is parameter-efficient fine-tuning (PEFT), which freezes PTMs while introducing extra parameters. Though work has been done to test PEFT methods in the SE field, a comprehensive evaluation is still lacking. This paper aims to fill in this gap by evaluating the effectiveness of five PEFT methods on eight PTMs and four SE downstream tasks. For different tasks and PEFT methods, we seek answers to the following research questions: 1) Is it more effective to use PTMs trained specifically on source code, or is it sufficient to use PTMs trained on natural language text? 2) What is the impact of varying model sizes? 3) How does the model architecture affect the performance? Besides effectiveness, we also discuss the efficiency of PEFT methods, concerning the costs of required training time and GPU resource consumption. We hope that our findings can provide a deeper understanding of PEFT methods on various PTMs and SE downstream tasks. All the codes and data are available at \url{https://github.com/zwtnju/PEFT.git}.
The key technology to overcome the drawbacks of hyperspectral imaging (expensive, high capture delay, and low spatial resolution) and make it widely applicable is to select only a few representative bands from hundreds of bands. However, current band selection (BS) methods face challenges in fair comparisons due to inconsistent train/validation settings, including the number of bands, dataset splits, and retraining settings. To make BS methods easy and reproducible, this paper presents the first band selection search benchmark (BSS-Bench) containing 52k training and evaluation records of numerous band combinations (BC) with different backbones for various hyperspectral analysis tasks. The creation of BSS-Bench required a significant computational effort of 1.26k GPU days. By querying BSS-Bench, BS experiments can be performed easily and reproducibly, and the gap between the searched result and the best achievable performance can be measured. Based on BSS-Bench, we further discuss the impact of various factors on BS, such as the number of bands, unsupervised statistics, and different backbones. In addition to BSS-Bench, we present an effective one-shot BS method called Single Combination One Shot (SCOS), which learns the priority of any BCs through one-time training, eliminating the need for repetitive retraining on different BCs. Furthermore, the search process of SCOS is flexible and does not require training, making it efficient and effective. Our extensive evaluations demonstrate that SCOS outperforms current BS methods on multiple tasks, even with much fewer bands. Our BSS-Bench and codes are available in the supplementary material and will be publicly available.
We introduce a novel monocular visual odometry (VO) system, NeRF-VO, that integrates learning-based sparse visual odometry for low-latency camera tracking and a neural radiance scene representation for fine-detailed dense reconstruction and novel view synthesis. Our system initializes camera poses using sparse visual odometry and obtains view-dependent dense geometry priors from a monocular prediction network. We harmonize the scale of poses and dense geometry, treating them as supervisory cues to train a neural implicit scene representation. NeRF-VO demonstrates exceptional performance in both photometric and geometric fidelity of the scene representation by jointly optimizing a sliding window of keyframed poses and the underlying dense geometry, which is accomplished through training the radiance field with volume rendering. We surpass SOTA methods in pose estimation accuracy, novel view synthesis fidelity, and dense reconstruction quality across a variety of synthetic and real-world datasets while achieving a higher camera tracking frequency and consuming less GPU memory.
This paper introduces PowerInfer, a high-speed Large Language Model (LLM) inference engine on a personal computer (PC) equipped with a single consumer-grade GPU. The key principle underlying the design of PowerInfer is exploiting the high locality inherent in LLM inference, characterized by a power-law distribution in neuron activation. This distribution indicates that a small subset of neurons, termed hot neurons, are consistently activated across inputs, while the majority, cold neurons, vary based on specific inputs. PowerInfer exploits such an insight to design a GPU-CPU hybrid inference engine: hot-activated neurons are preloaded onto the GPU for fast access, while cold-activated neurons are computed on the CPU, thus significantly reducing GPU memory demands and CPU-GPU data transfers. PowerInfer further integrates adaptive predictors and neuron-aware sparse operators, optimizing the efficiency of neuron activation and computational sparsity. The evaluation shows that PowerInfer significantly outperforms llama.cpp by up to 11.69x while retaining model accuracy across various LLMs (including OPT-175B) on a single NVIDIA RTX 4090 GPU. For the OPT-30B model, PowerInfer achieves performance comparable to that of a high-end server-grade A100 GPU, reaching 82% of its token generation rate on a single consumer-grade RTX 4090 GPU.
Current methodologies in point cloud analysis predominantly explore 3D geometries, often achieved through the introduction of intricate learnable geometric extractors in the encoder or by deepening networks with repeated blocks. However, these approaches inevitably lead to a significant number of learnable parameters, resulting in substantial computational costs and imposing memory burdens on CPU/GPU. Additionally, the existing strategies are primarily tailored for object-level point cloud classification and segmentation tasks, with limited extensions to crucial scene-level applications, such as autonomous driving. In response to these limitations, we introduce PointeNet, an efficient network designed specifically for point cloud analysis. PointeNet distinguishes itself with its lightweight architecture, low training cost, and plug-and-play capability, effectively capturing representative features. The network consists of a Multivariate Geometric Encoding (MGE) module and an optional Distance-aware Semantic Enhancement (DSE) module. The MGE module employs operations of sampling, grouping, and multivariate geometric aggregation to lightweightly capture and adaptively aggregate multivariate geometric features, providing a comprehensive depiction of 3D geometries. The DSE module, designed for real-world autonomous driving scenarios, enhances the semantic perception of point clouds, particularly for distant points. Our method demonstrates flexibility by seamlessly integrating with a classification/segmentation head or embedding into off-the-shelf 3D object detection networks, achieving notable performance improvements at a minimal cost. Extensive experiments on object-level datasets, including ModelNet40, ScanObjectNN, ShapeNetPart, and the scene-level dataset KITTI, demonstrate the superior performance of PointeNet over state-of-the-art methods in point cloud analysis.
Among existing Neural Architecture Search methods, DARTS is known for its efficiency and simplicity. This approach applies continuous relaxation of network representation to construct a weight-sharing supernet and enables the identification of excellent subnets in just a few GPU days. However, performance collapse in DARTS results in deteriorating architectures filled with parameter-free operations and remains a great challenge to the robustness. To resolve this problem, we reveal that the fundamental reason is the biased estimation of the candidate importance in the search space through theoretical and experimental analysis, and more precisely select operations via information-based measurements. Furthermore, we demonstrate that the excessive concern over the supernet and inefficient utilization of data in bi-level optimization also account for suboptimal results. We adopt a more realistic objective focusing on the performance of subnets and simplify it with the help of the information-based measurements. Finally, we explain theoretically why progressively shrinking the width of the supernet is necessary and reduce the approximation error of optimal weights in DARTS. Our proposed method, named IS-DARTS, comprehensively improves DARTS and resolves the aforementioned problems. Extensive experiments on NAS-Bench-201 and DARTS-based search space demonstrate the effectiveness of IS-DARTS.
Diffusion inversion is the problem of taking an image and a text prompt that describes it and finding a noise latent that would generate the exact same image. Most current deterministic inversion techniques operate by approximately solving an implicit equation and may converge slowly or yield poor reconstructed images. We formulate the problem by finding the roots of an implicit equation and devlop a method to solve it efficiently. Our solution is based on Newton-Raphson (NR), a well-known technique in numerical analysis. We show that a vanilla application of NR is computationally infeasible while naively transforming it to a computationally tractable alternative tends to converge to out-of-distribution solutions, resulting in poor reconstruction and editing. We therefore derive an efficient guided formulation that fastly converges and provides high-quality reconstructions and editing. We showcase our method on real image editing with three popular open-sourced diffusion models: Stable Diffusion, SDXL-Turbo, and Flux with different deterministic schedulers. Our solution, Guided Newton-Raphson Inversion, inverts an image within 0.4 sec (on an A100 GPU) for few-step models (SDXL-Turbo and Flux.1), opening the door for interactive image editing. We further show improved results in image interpolation and generation of rare objects.
We introduce the \method, an ultra-efficient approach for monocular 3D object reconstruction. Splatter Image is based on Gaussian Splatting, which allows fast and high-quality reconstruction of 3D scenes from multiple images. We apply Gaussian Splatting to monocular reconstruction by learning a neural network that, at test time, performs reconstruction in a feed-forward manner, at 38 FPS. Our main innovation is the surprisingly straightforward design of this network, which, using 2D operators, maps the input image to one 3D Gaussian per pixel. The resulting set of Gaussians thus has the form an image, the Splatter Image. We further extend the method take several images as input via cross-view attention. Owning to the speed of the renderer (588 FPS), we use a single GPU for training while generating entire images at each iteration to optimize perceptual metrics like LPIPS. On several synthetic, real, multi-category and large-scale benchmark datasets, we achieve better results in terms of PSNR, LPIPS, and other metrics while training and evaluating much faster than prior works. Code, models, demo and more results are available at https://szymanowiczs.github.io/splatter-image.
Large language models (LLMs) have demonstrated remarkable success as foundational models, benefiting various downstream applications through fine-tuning. Recent studies on loss scaling have demonstrated the superior performance of larger LLMs compared to their smaller counterparts. Nevertheless, training LLMs with billions of parameters poses significant challenges and requires considerable computational resources. For example, training a one trillion parameter GPT-style model on 20 trillion tokens requires a staggering 120 million exaflops of computation. This research explores efficient distributed training strategies to extract this computation from Frontier, the world's first exascale supercomputer dedicated to open science. We enable and investigate various model and data parallel training techniques, such as tensor parallelism, pipeline parallelism, and sharded data parallelism, to facilitate training a trillion-parameter model on Frontier. We empirically assess these techniques and their associated parameters to determine their impact on memory footprint, communication latency, and GPU's computational efficiency. We analyze the complex interplay among these techniques and find a strategy to combine them to achieve high throughput through hyperparameter tuning. We have identified efficient strategies for training large LLMs of varying sizes through empirical analysis and hyperparameter tuning. For 22 Billion, 175 Billion, and 1 Trillion parameters, we achieved GPU throughputs of $38.38\%$, $36.14\%$, and $31.96\%$, respectively. For the training of the 175 Billion parameter model and the 1 Trillion parameter model, we achieved $100\%$ weak scaling efficiency on 1024 and 3072 MI250X GPUs, respectively. We also achieved strong scaling efficiencies of $89\%$ and $87\%$ for these two models.
This work presents the PORTALS framework, which leverages surrogate modeling and optimization techniques to enable the prediction of core plasma profiles and performance with nonlinear gyrokinetic simulations at significantly reduced cost, with no loss of accuracy. The efficiency of PORTALS is benchmarked against standard methods, and its full potential is demonstrated on a unique, simultaneous 5-channel (electron temperature, ion temperature, electron density, impurity density and angular rotation) prediction of steady-state profiles in a DIII-D ITER Similar Shape plasma with GPU-accelerated, nonlinear CGYRO. This paper also provides general guidelines for accurate performance predictions in burning plasmas and the impact of transport modeling in fusion pilot plants studies.
As large language models (LLMs) become widespread in various application domains, a critical challenge the AI community is facing is how to train these large AI models in a cost-effective manner. Existing LLM training plans typically employ a heuristic based parallel training strategy which is based on empirical observations rather than grounded upon a thorough examination of the search space of LLM parallelization. Such limitation renders existing systems to leave significant performance left on the table, wasting millions of dollars worth of training cost. This paper presents our profiling-driven simulator called vTrain, providing AI practitioners a fast yet accurate software framework to determine an efficient and cost-effective LLM training system configuration. We demonstrate vTrain's practicality through several case studies, e.g., effectively evaluating optimal training parallelization strategies that balances training time and its associated training cost, efficient multi-tenant GPU cluster schedulers targeting multiple LLM training jobs, and determining a compute-optimal LLM model architecture given a fixed compute budget.
Large language models (LLMs) are central to modern natural language processing, delivering exceptional performance in various tasks. However, their substantial computational and memory requirements present challenges, especially for devices with limited DRAM capacity. This paper tackles the challenge of efficiently running LLMs that exceed the available DRAM capacity by storing the model parameters in flash memory, but bringing them on demand to DRAM. Our method involves constructing an inference cost model that takes into account the characteristics of flash memory, guiding us to optimize in two critical areas: reducing the volume of data transferred from flash and reading data in larger, more contiguous chunks. Within this hardware-informed framework, we introduce two principal techniques. First, "windowing" strategically reduces data transfer by reusing previously activated neurons, and second, "row-column bundling", tailored to the sequential data access strengths of flash memory, increases the size of data chunks read from flash memory. These methods collectively enable running models up to twice the size of the available DRAM, with a 4-5x and 20-25x increase in inference speed compared to naive loading approaches in CPU and GPU, respectively. Our integration of sparsity awareness, context-adaptive loading, and a hardware-oriented design paves the way for effective inference of LLMs on devices with limited memory.
Inspired by the diversity and depth of XLand and the simplicity and minimalism of MiniGrid, we present XLand-MiniGrid, a suite of tools and grid-world environments for meta-reinforcement learning research. Written in JAX, XLand-MiniGrid is designed to be highly scalable and can potentially run on GPU or TPU accelerators, democratizing large-scale experimentation with limited resources. Along with the environments, XLand-MiniGrid provides pre-sampled benchmarks with millions of unique tasks of varying difficulty and easy-to-use baselines that allow users to quickly start training adaptive agents. In addition, we have conducted a preliminary analysis of scaling and generalization, showing that our baselines are capable of reaching millions of steps per second during training and validating that the proposed benchmarks are challenging. XLand-MiniGrid is open-source and available at https://github.com/dunnolab/xland-minigrid.