Domain-specific hardware to solve computationally hard optimization problems has generated tremendous excitement. Here, we evaluate probabilistic bit (p-bit) based Ising Machines (IM) on the 3-regular 3-Exclusive OR Satisfiability (3R3X), as a representative hard optimization problem. We first introduce a multiplexed architecture that emulates all-to-all network functionality while maintaining highly parallelized chromatic Gibbs sampling. We implement this architecture in single Field-Programmable Gate Arrays (FPGA) and show that running the adaptive parallel tempering algorithm demonstrates competitive algorithmic and prefactor advantages over alternative IMs by D-Wave, Toshiba, and Fujitsu. We also implement higher-order interactions that lead to better prefactors without changing algorithmic scaling for the XORSAT problem. Even though FPGA implementations of p-bits are still not quite as fast as the best possible greedy algorithms accelerated on Graphics Processing Units (GPU), scaled magnetic versions of p-bit IMs could lead to orders of magnitude improvements over the state of the art for generic optimization.
Convolutional Neural Networks (CNNs) continue to achieve great success in classification tasks as innovative techniques and complex multi-path architecture topologies are introduced. Neural Architecture Search (NAS) aims to automate the design of these complex architectures, reducing the need for costly manual design work by human experts. Cellular Encoding (CE) is an evolutionary computation technique which excels in constructing novel multi-path topologies of varying complexity and has recently been applied with NAS to evolve CNN architectures for various classification tasks. However, existing CE approaches have severe limitations. They are restricted to only one domain, only partially implement the theme of CE, or only focus on the micro-architecture search space. This paper introduces a new CE representation and algorithm capable of evolving novel multi-path CNN architectures of varying depth, width, and complexity for image and text classification tasks. The algorithm explicitly focuses on the macro-architecture search space. Furthermore, by using a surrogate model approach, we show that the algorithm can evolve a performant CNN architecture in less than one GPU day, thereby allowing a sufficient number of experiment runs to be conducted to achieve scientific robustness. Experiment results show that the approach is highly competitive, defeating several state-of-the-art methods, and is generalisable to both the image and text domains.
Word2Vec remains one of the highly-impactful innovations in the field of Natural Language Processing (NLP) that represents latent grammatical and syntactical information in human text with dense vectors in a low dimension. Word2Vec has high computational cost due to the algorithm's inherent sequentiality, intensive memory accesses, and the large vocabularies it represents. While prior studies have investigated technologies to explore parallelism and improve memory system performance, they struggle to effectively gain throughput on powerful GPUs. We identify memory data access and latency as the primary bottleneck in prior works on GPUs, which prevents highly optimized kernels from attaining the architecture's peak performance. We present a novel algorithm, FULL-W2V, which maximally exploits the opportunities for data reuse in the W2V algorithm and leverages GPU architecture and resources to reduce access to low memory levels and improve temporal locality. FULL-W2V is capable of reducing accesses to GPU global memory significantly, e.g., by more than 89\%, compared to prior state-of-the-art GPU implementations, resulting in significant performance improvement that scales across successive hardware generations. Our prototype implementation achieves 2.97X speedup when ported from Nvidia Pascal P100 to Volta V100 cards, and outperforms the state-of-the-art by 5.72X on V100 cards with the same embedding quality. In-depth analysis indicates that the reduction of memory accesses through register and shared memory caching and high-throughput shared memory reduction leads to a significantly improved arithmetic intensity. FULL-W2V can potentially benefit many applications in NLP and other domains.
Post-training quantization (PTQ) has played a key role in compressing large language models (LLMs) with ultra-low costs. However, existing PTQ methods only focus on handling the outliers within one layer or one block, which ignores the dependency of blocks and leads to severe performance degradation in low-bit settings. In this paper, we propose CBQ, a cross-block reconstruction-based PTQ method for LLMs. CBQ employs a cross-block dependency using a homologous reconstruction scheme, establishing long-range dependencies across multiple blocks to minimize error accumulation. Furthermore, CBQ incorporates a coarse-to-fine preprocessing (CFP) strategy for suppressing weight and activation outliers, coupled with an adaptive LoRA-Rounding technique for precise weight quantization. These innovations enable CBQ to not only handle extreme outliers effectively but also improve overall quantization accuracy. Extensive experiments show that CBQ achieves superior low-bit quantization (W4A4, W4A8, W2A16) and outperforms existing state-of-the-art methods across various LLMs and datasets. Notably, CBQ quantizes the 4-bit LLAMA1-65B model within only 4.3 hours on a single GPU, achieving a commendable tradeoff between performance and quantization efficiency.
Recent advancements in large-scale pre-training of visual-language models on paired image-text data have demonstrated impressive generalization capabilities for zero-shot tasks. Building on this success, efforts have been made to adapt these image-based visual-language models, such as CLIP, for videos extending their zero-shot capabilities to the video domain. While these adaptations have shown promising results, they come at a significant computational cost and struggle with effectively modeling the crucial temporal aspects inherent to the video domain. In this study, we present EZ-CLIP, a simple and efficient adaptation of CLIP that addresses these challenges. EZ-CLIP leverages temporal visual prompting for seamless temporal adaptation, requiring no fundamental alterations to the core CLIP architecture while preserving its remarkable generalization abilities. Moreover, we introduce a novel learning objective that guides the temporal visual prompts to focus on capturing motion, thereby enhancing its learning capabilities from video data. We conducted extensive experiments on five different benchmark datasets, thoroughly evaluating EZ-CLIP for zero-shot learning and base-to-novel video action recognition, and also demonstrating its potential for few-shot generalization.Impressively, with a mere 5.2 million learnable parameters (as opposed to the 71.1 million in the prior best model), EZ-CLIP can be efficiently trained on a single GPU, outperforming existing approaches in several evaluations.
Spiking neural networks (SNNs) are potential competitors to artificial neural networks (ANNs) due to their high energy-efficiency on neuromorphic hardware. However, SNNs are unfolded over simulation time steps during the training process. Thus, SNNs require much more memory than ANNs, which impedes the training of deeper SNN models. In this paper, we propose the reversible spiking neural network to reduce the memory cost of intermediate activations and membrane potentials during training. Firstly, we extend the reversible architecture along temporal dimension and propose the reversible spiking block, which can reconstruct the computational graph and recompute all intermediate variables in forward pass with a reverse process. On this basis, we adopt the state-of-the-art SNN models to the reversible variants, namely reversible spiking ResNet (RevSResNet) and reversible spiking transformer (RevSFormer). Through experiments on static and neuromorphic datasets, we demonstrate that the memory cost per image of our reversible SNNs does not increase with the network depth. On CIFAR10 and CIFAR100 datasets, our RevSResNet37 and RevSFormer-4-384 achieve comparable accuracies and consume 3.79x and 3.00x lower GPU memory per image than their counterparts with roughly identical model complexity and parameters. We believe that this work can unleash the memory constraints in SNN training and pave the way for training extremely large and deep SNNs. The code is available at https://github.com/mi804/RevSNN.git.
Large Language Models (LLMs) are capable of reasoning over diverse input data modalities through pre-trained encoders. However, the growing diversity of input data modalities prevents incorporating all modalities into LLMs, especially when LLMs are deployed on resource-constrained edge devices for embodied AI applications. Instead, a better option is to adaptively involve only the useful modalities at runtime, depending on the current environmental contexts and task requirements. For such modality adaptation, existing work adopts fixed connections between encoders and the LLM's input layer, leading to high training cost at runtime and ineffective cross-modal interaction. In this paper, we address these limitations by presenting mPnP-LLM, a new technique that allows fully elastic, automated and prompt runtime modality adaptation, by connecting unimodal encoders to a flexible set of last LLM blocks and making such latent connections fully trainable at runtime. Experiments over the nuScenes-QA dataset show that mPnP-LLM can achieve up to 3.7x FLOPs reduction and 30% GPU memory usage reduction, while retaining on-par accuracy with the existing schemes. Under the same compute budget, mPnP-LLM improves the task accuracy by up to 4% compared to the best existing scheme.
When pre-trained models become rapidly larger, the cost of fine-tuning on downstream tasks steadily increases, too. To economically fine-tune these models, parameter-efficient transfer learning (PETL) is proposed, which only tunes a tiny subset of trainable parameters to efficiently learn quality representations. However, current PETL methods are facing the dilemma that during training the GPU memory footprint is not effectively reduced as trainable parameters. PETL will likely fail, too, if the full fine-tuning encounters the out-of-GPU-memory issue. This phenomenon happens because trainable parameters from these methods are generally entangled with the backbone, such that a lot of intermediate states have to be stored in GPU memory for gradient propagation. To alleviate this problem, we introduce Disentangled Transfer Learning (DTL), which disentangles the trainable parameters from the backbone using a lightweight Compact Side Network (CSN). By progressively extracting task-specific information with a few low-rank linear mappings and appropriately adding the information back to the backbone, CSN effectively realizes knowledge transfer in various downstream tasks. We conducted extensive experiments to validate the effectiveness of our method. The proposed method not only reduces a large amount of GPU memory usage and trainable parameters, but also outperforms existing PETL methods by a significant margin in accuracy, achieving new state-of-the-art on several standard benchmarks. The code is available at https://github.com/heekhero/DTL.
Flow cytometry is widely used to identify cell populations in patient-derived fluids such as peripheral blood (PB) or cerebrospinal fluid (CSF). While ubiquitous in research and clinical practice, flow cytometry requires gating, i.e. cell type identification which requires labor-intensive and error-prone manual adjustments. To facilitate this process, we designed GateNet, the first neural network architecture enabling full end-to-end automated gating without the need to correct for batch effects. We train GateNet with over 8,000,000 events based on N=127 PB and CSF samples which were manually labeled independently by four experts. We show that for novel, unseen samples, GateNet achieves human-level performance (F1 score ranging from 0.910 to 0.997). In addition we apply GateNet to a publicly available dataset confirming generalization with an F1 score of 0.936. As our implementation utilizes graphics processing units (GPU), gating only needs 15 microseconds per event. Importantly, we also show that GateNet only requires ~10 samples to reach human-level performance, rendering it widely applicable in all domains of flow cytometry.
Due to the substantial scale of Large Language Models (LLMs), the direct application of conventional compression methodologies proves impractical. The computational demands associated with even minimal gradient updates present challenges, particularly on consumer-grade hardware. This paper introduces an innovative approach for the parametric and practical compression of LLMs based on reduced order modelling, which entails low-rank decomposition within the feature space and re-parameterization in the weight space. Notably, this compression technique operates in a layer-wise manner, obviating the need for a GPU device and enabling the compression of billion-scale models within stringent constraints of both memory and time. Our method represents a significant advancement in model compression by leveraging matrix decomposition, demonstrating superior efficacy compared to the prevailing state-of-the-art structured pruning method.
We devise, implement and performance-asses DYAD, a layer which can serve as a faster and more memory-efficient approximate replacement for linear layers, (nn.Linear() in Pytorch). These layers appear in common subcomponents, such as in the ff module of Transformers. DYAD is based on a bespoke near-sparse matrix structure which approximates the dense "weight" matrix W that matrix-multiplies the input in the typical realization of such a layer, a.k.a DENSE. Our alternative near-sparse matrix structure is decomposable to a sum of 2 matrices permutable to a block-sparse counterpart. These can be represented as 3D tensors, which in unison allow a faster execution of matrix multiplication with the mini-batched input matrix X compared to DENSE (O(rows(W ) x cols(W )) --> O( rows(W ) x cols(W ) # of blocks )). As the crux of our experiments, we pretrain both DYAD and DENSE variants of 2 sizes of the OPT arch and 1 size of the Pythia arch, including at different token scales of the babyLM benchmark. We find DYAD to be competitive (>= 90%) of DENSE performance on zero-shot (e.g. BLIMP), few-shot (OPENLM) and finetuning (GLUE) benchmarks, while being >=7-15% faster to train on-GPU even at 125m scale, besides surfacing larger speedups at increasing scale and model width.
Real-time multi-person pose estimation presents significant challenges in balancing speed and precision. While two-stage top-down methods slow down as the number of people in the image increases, existing one-stage methods often fail to simultaneously deliver high accuracy and real-time performance. This paper introduces RTMO, a one-stage pose estimation framework that seamlessly integrates coordinate classification by representing keypoints using dual 1-D heatmaps within the YOLO architecture, achieving accuracy comparable to top-down methods while maintaining high speed. We propose a dynamic coordinate classifier and a tailored loss function for heatmap learning, specifically designed to address the incompatibilities between coordinate classification and dense prediction models. RTMO outperforms state-of-the-art one-stage pose estimators, achieving 1.1% higher AP on COCO while operating about 9 times faster with the same backbone. Our largest model, RTMO-l, attains 74.8% AP on COCO val2017 and 141 FPS on a single V100 GPU, demonstrating its efficiency and accuracy. The code and models are available at https://github.com/open-mmlab/mmpose/tree/main/projects/rtmo.
This work studies a challenging and practical issue known as multi-class unsupervised anomaly detection (MUAD). This problem requires only normal images for training while simultaneously testing both normal and anomaly images across multiple classes. Existing reconstruction-based methods typically adopt pyramidal networks as encoders and decoders to obtain multi-resolution features, often involving complex sub-modules with extensive handcraft engineering. In contrast, a plain Vision Transformer (ViT) showcasing a more straightforward architecture has proven effective in multiple domains, including detection and segmentation tasks. It is simpler, more effective, and elegant. Following this spirit, we explore the use of only plain ViT features for MUAD. We first abstract a Meta-AD concept by synthesizing current reconstruction-based methods. Subsequently, we instantiate a novel ViT-based ViTAD structure, designed incrementally from both global and local perspectives. This model provide a strong baseline to facilitate future research. Additionally, this paper uncovers several intriguing findings for further investigation. Finally, we comprehensively and fairly benchmark various approaches using eight metrics. Utilizing a basic training regimen with only an MSE loss, ViTAD achieves state-of-the-art results and efficiency on MVTec AD, VisA, and Uni-Medical datasets. \Eg, achieving 85.4 mAD that surpasses UniAD by +3.0 for the MVTec AD dataset, and it requires only 1.1 hours and 2.3G GPU memory to complete model training on a single V100 that can serve as a strong baseline to facilitate the development of future research. Full code is available at https://zhangzjn.github.io/projects/ViTAD/.
We introduce "PatchMorph," an new stochastic deep learning algorithm tailored for unsupervised 3D brain image registration. Unlike other methods, our method uses compact patches of a constant small size to derive solutions that can combine global transformations with local deformations. This approach minimizes the memory footprint of the GPU during training, but also enables us to operate on numerous amounts of randomly overlapping small patches during inference to mitigate image and patch boundary problems. PatchMorph adeptly handles world coordinate transformations between two input images, accommodating variances in attributes such as spacing, array sizes, and orientations. The spatial resolution of patches transitions from coarse to fine, addressing both global and local attributes essential for aligning the images. Each patch offers a unique perspective, together converging towards a comprehensive solution. Experiments on human T1 MRI brain images and marmoset brain images from serial 2-photon tomography affirm PatchMorph's superior performance.
Instruction fine-tuning has conventionally been employed to adapt Large Language Models (LLMs) to a variety of tasks. Nonetheless, this technique often necessitates substantial computational resources, making it impractical for deployment by individuals or small-scale entities. Recently, Low-Rank Adaptation (LoRA) has become a promising alternative, offering high capabilities on par with full tuning with reduced resource overhead. However, attaining satisfactory performance through the fine-tuning of LoRA is a non-trivial challenge. In this paper, we propose PILLOW, which aims to improve LoRA's performance by a discrimination-based prompting method, leveraging LLMs' In-Context Learning ability. PILLOW incorporates a matching network that selects prompts from a user-defined prompt pool, concatenates the selected prompts with the user instruction as input, and performs inference using the LoRA-fine-tuned LLMs. Trained with Reinforcement Learning, PILLOW exhibits commensurate performance on various evaluation metrics compared with typical instruction fine-tuning methods, utilizing only consumer-grade GPU resources and exhibiting a large reduction in computational costs.
Large Language Models (LLMs) are wildly popular today and it is important to serve them efficiently. Existing LLM serving systems are stateless across requests. Consequently, when LLMs are used in the common setting of multi-turn conversations, a growing log of the conversation history must be processed alongside any request by the serving system at each turn, resulting in repeated processing. In this paper, we design $Pensieve$, a system optimized for multi-turn conversation LLM serving. $Pensieve$ maintains the conversation state across requests by caching previously processed history to avoid duplicate processing. $Pensieve$'s multi-tier caching strategy can utilize both GPU and CPU memory to efficiently store and retrieve cached data. $Pensieve$ also generalizes the recent PagedAttention kernel to support attention between multiple input tokens with a GPU cache spread over non-contiguous memory. Our evaluation shows that $Pensieve$ can achieve $1.14$-$3.0\times$ the throughput of vLLM and TensorRT-LLM and significantly reduce latency.
Vision-Language Models (VLMs) such as CLIP have demonstrated remarkable generalization capabilities to downstream tasks. However, existing prompt tuning based frameworks need to parallelize learnable textual inputs for all categories, suffering from massive GPU memory consumption when there is a large number of categories in the target dataset. Moreover, previous works require to include category names within prompts, exhibiting subpar performance when dealing with ambiguous category names. To address these shortcomings, we propose Compound Text-Guided Prompt Tuning (TGP-T) that significantly reduces resource demand while achieving superior performance. We introduce text supervision to the optimization of prompts, which enables two benefits: 1) releasing the model reliance on the pre-defined category names during inference, thereby enabling more flexible prompt generation; 2) reducing the number of inputs to the text encoder, which decreases GPU memory consumption significantly. Specifically, we found that compound text supervisions, i.e., category-wise and content-wise, is highly effective, since they provide inter-class separability and capture intra-class variations, respectively. Moreover, we condition the prompt generation on visual features through a module called Bonder, which facilitates the alignment between prompts and visual features. Extensive experiments on few-shot recognition and domain generalization demonstrate that TGP-T achieves superior performance with consistently lower training costs. It reduces GPU memory usage by 93% and attains a 2.5% performance gain on 16-shot ImageNet. The code is available at https://github.com/EricTan7/TGP-T.
Siamese network-based trackers have shown remarkable success in aerial tracking. Most previous works, however, usually perform template matching only between the initial template and the search region and thus fail to deal with rapidly changing targets that often appear in aerial tracking. As a remedy, this work presents Building Appearance Collection Tracking (BACTrack). This simple yet effective tracking framework builds a dynamic collection of target templates online and performs efficient multi-template matching to achieve robust tracking. Specifically, BACTrack mainly comprises a Mixed-Temporal Transformer (MTT) and an appearance discriminator. The former is responsible for efficiently building relationships between the search region and multiple target templates in parallel through a mixed-temporal attention mechanism. At the same time, the appearance discriminator employs an online adaptive template-update strategy to ensure that the collected multiple templates remain reliable and diverse, allowing them to closely follow rapid changes in the target's appearance and suppress background interference during tracking. Extensive experiments show that our BACTrack achieves top performance on four challenging aerial tracking benchmarks while maintaining an impressive speed of over 87 FPS on a single GPU. Speed tests on embedded platforms also validate our potential suitability for deployment on UAV platforms.
In the domain of vision-language integration, generating detailed image captions poses a significant challenge due to the lack of curated and rich datasets. This study introduces PixLore, a novel method that leverages Querying Transformers through the fine-tuning of the BLIP-2 model using the LoRa method on a standard commercial GPU. The followed approach, which involves training on a carefully assembled dataset from state-of-the-art Computer Vision models combined and augmented by ChatGPT, addresses the question of whether intricate image understanding can be achieved with an ensemble of smaller-scale models, referred to as Knowledge Stitching. Comparative evaluations against major models such as GPT-4 and Google Bard demonstrate that PixLore-2.7B, despite having considerably fewer parameters, is rated higher than the existing State-of-the-Art models in over half of the assessments. Precisely, PixLore outperform Bard and BLIP-2, which score approximately 35.18% and 27.98% lower than PixLore in the task of image captioning. This research not only presents a groundbreaking approach but also highlights the importance of well-curated datasets in enhancing the performance of smaller models.
Deep learning (DL) jobs use multi-dimensional parallelism, i.e. combining data, model, and pipeline parallelism, to use large GPU clusters efficiently. Long-running jobs may experience changes to their GPU allocation: (i) resource elasticity during training adds or removes GPUs; (ii) hardware maintenance may require redeployment on different GPUs; and (iii) GPU failures force jobs to run with fewer devices. Current DL frameworks tie jobs to a set of GPUs and thus lack support for these scenarios. In particular, they cannot change the multi-dimensional parallelism of an already-running job in an efficient and model-independent way. We describe Scalai, a state management library for DL systems that enables jobs to change their parallelism dynamically after the GPU allocation is updated at runtime. Scalai achieves this through a new abstraction, a parallelizable tensor collection (PTC), that externalizes the job state during training. After a GPU change, Scalai uses the PTC to transform the job state: the PTC repartitions the dataset state under data parallelism and exposes it to DL workers through a virtual file system; and the PTC obtains the model state as partitioned checkpoints and transforms them to reflect the new parallelization configuration. For efficiency, Scalai executes PTC transformations in parallel with minimum data movement between workers. Our experiments show that Scalai enables DL jobs to support dynamic parallelization with low overhead.