Recent studies have revealed that, during the inference on generative AI models such as transformer, the importance of different weights exhibits substantial context-dependent variations. This naturally manifests a promising potential of adaptively configuring weight quantization to improve the generative AI inference efficiency. Although configurable weight quantization can readily leverage the hardware support of variable-precision arithmetics in modern GPU and AI accelerators, little prior research has studied how one could exploit variable weight quantization to proportionally improve the AI model memory access speed and energy efficiency. Motivated by the rapidly maturing CXL ecosystem, this work develops a CXL-based design solution to fill this gap. The key is to allow CXL memory controllers play an active role in supporting and exploiting runtime configurable weight quantization. Using transformer as a representative generative AI model, we carried out experiments that well demonstrate the effectiveness of the proposed design solution.
Vision Transformer (ViT) acceleration with field programmable gate array (FPGA) is promising but challenging. Existing FPGA-based ViT accelerators mainly rely on temporal architectures, which process different operators by reusing the same hardware blocks and suffer from extensive memory access overhead. Pipelined architectures, either coarse-grained or fine-grained, unroll the ViT computation spatially for memory access efficiency. However, they usually suffer from significant hardware resource constraints and pipeline bubbles induced by the global computation dependency of ViT. In this paper, we introduce HG-PIPE, a pipelined FPGA accelerator for high-throughput and low-latency ViT processing. HG-PIPE features a hybrid-grained pipeline architecture to reduce on-chip buffer cost and couples the computation dataflow and parallelism design to eliminate the pipeline bubbles. HG-PIPE further introduces careful approximations to implement both linear and non-linear operators with abundant Lookup Tables (LUTs), thus alleviating resource constraints. On a ZCU102 FPGA, HG-PIPE achieves 2.78 times better throughput and 2.52 times better resource efficiency than the prior-art accelerators, e.g., AutoViTAcc. With a VCK190 FPGA, HG-PIPE realizes end-to-end ViT acceleration on a single device and achieves 7118 images/s, which is 2.81 times faster than a V100 GPU.
Multi-head-self-attention (MHSA) mechanisms achieve state-of-the-art (SOTA) performance across natural language processing and vision tasks. However, their quadratic dependence on sequence lengths has bottlenecked inference speeds. To circumvent this bottleneck, researchers have proposed various sparse-MHSA models, where a subset of full attention is computed. Despite their promise, current sparse libraries and compilers do not support high-performance implementations for diverse sparse-MHSA patterns due to the underlying sparse formats they operate on. These formats, which are typically designed for high-performance & scientific computing applications, are either curated for extreme amounts of random sparsity (<1% non-zero values), or specific sparsity patterns. However, the sparsity patterns in sparse-MHSA are moderately sparse (10-50% non-zero values) and varied, resulting in existing sparse-formats trading off generality for performance. We bridge this gap, achieving both generality and performance, by proposing a novel sparse format: affine-compressed-sparse-row (ACSR) and supporting code-generation scheme, SPLAT, that generates high-performance implementations for diverse sparse-MHSA patterns on GPUs. Core to our proposed format and code generation algorithm is the observation that common sparse-MHSA patterns have uniquely regular geometric properties. These properties, which can be analyzed just-in-time, expose novel optimizations and tiling strategies that SPLAT exploits to generate high-performance implementations for diverse patterns. To demonstrate SPLAT's efficacy, we use it to generate code for various sparse-MHSA models, achieving geomean speedups of 2.05x and 4.05x over hand-written kernels written in triton and TVM respectively on A100 GPUs. Moreover, its interfaces are intuitive and easy to use with existing implementations of MHSA in JAX.
Enlarging input images is a straightforward and effective approach to promote small object detection. However, simple image enlargement is significantly expensive on both computations and GPU memory. In fact, small objects are usually sparsely distributed and locally clustered. Therefore, massive feature extraction computations are wasted on the non-target background area of images. Recent works have tried to pick out target-containing regions using an extra network and perform conventional object detection, but the newly introduced computation limits their final performance. In this paper, we propose to reuse the detector's backbone to conduct feature-level object-seeking and patch-slicing, which can avoid redundant feature extraction and reduce the computation cost. Incorporating a sparse detection head, we are able to detect small objects on high-resolution inputs (e.g., 1080P or larger) for superior performance. The resulting Efficient Small Object Detection (ESOD) approach is a generic framework, which can be applied to both CNN- and ViT-based detectors to save the computation and GPU memory costs. Extensive experiments demonstrate the efficacy and efficiency of our method. In particular, our method consistently surpasses the SOTA detectors by a large margin (e.g., 8% gains on AP) on the representative VisDrone, UAVDT, and TinyPerson datasets. Code is available at https://github.com/alibaba/esod.
Large language models (LLMs) are crucial in modern natural language processing and artificial intelligence. However, they face challenges in managing their significant memory requirements. Although quantization-aware training (QAT) offers a solution by reducing memory consumption through low-bit representations with minimal accuracy loss, it is impractical due to substantial training resources. To address this, we propose Efficient Quantization-Aware Training (EfficientQAT), a more feasible QAT algorithm. EfficientQAT involves two consecutive phases: Block-wise training of all parameters (Block-AP) and end-to-end training of quantization parameters (E2E-QP). To the best of our knowledge, Block-AP is the first method to enable direct training of all parameters in a block-wise manner, reducing accuracy loss in low-bit scenarios by enhancing the solution space during optimization. E2E-QP then trains only the quantization parameters (step sizes) end-to-end, further improving the performance of quantized models by considering interactions among all sub-modules. Extensive experiments demonstrate that EfficientQAT outperforms previous quantization methods across a range of models, including base LLMs, instruction-tuned LLMs, and multimodal LLMs, with scales from 7B to 70B parameters at various quantization bits. For instance, EfficientQAT obtains a 2-bit Llama-2-70B model on a single A100-80GB GPU in 41 hours, with less than 3 points accuracy degradation compared to the full precision (69.48 vs. 72.41). Code is available at https://github.com/OpenGVLab/EfficientQAT.
Neural architecture search (NAS) enables re-searchers to automatically explore vast search spaces and find efficient neural networks. But NAS suffers from a key bottleneck, i.e., numerous architectures need to be evaluated during the search process, which requires a lot of computing resources and time. In order to improve the efficiency of NAS, a series of methods have been proposed to reduce the evaluation time of neural architectures. However, they are not efficient enough and still only focus on the accuracy of architectures. In addition to the classification accuracy, more efficient and smaller network architectures are required in real-world applications. To address the above problems, we propose the SMEM-NAS, a pairwise com-parison relation-assisted multi-objective evolutionary algorithm based on a multi-population mechanism. In the SMEM-NAS, a surrogate model is constructed based on pairwise compari-son relations to predict the accuracy ranking of architectures, rather than the absolute accuracy. Moreover, two populations cooperate with each other in the search process, i.e., a main population guides the evolution, while a vice population expands the diversity. Our method aims to provide high-performance models that take into account multiple optimization objectives. We conduct a series of experiments on the CIFAR-10, CIFAR-100 and ImageNet datasets to verify its effectiveness. With only a single GPU searching for 0.17 days, competitive architectures can be found by SMEM-NAS which achieves 78.91% accuracy with the MAdds of 570M on the ImageNet. This work makes a significant advance in the important field of NAS.
The application of large-scale models in medical image segmentation demands substantial quantities of meticulously annotated data curated by experts along with high computational resources, both of which are challenges in resource-poor settings. In this study, we present the Medical Segment Anything Model with Galore MedSAGa where we adopt the Segment Anything Model (SAM) to achieve memory-efficient, few-shot medical image segmentation by applying Gradient Low-Rank Projection GaLore to the parameters of the image encoder of SAM. Meanwhile, the weights of the prompt encoder and mask decoder undergo full parameter fine-tuning using standard optimizers. We further assess MedSAGa's few-shot learning capabilities, reporting on its memory efficiency and segmentation performance across multiple standard medical image segmentation datasets. We compare it with several baseline models, including LoRA fine-tuned SAM (SAMed) and DAE-Former. Experiments across multiple datasets and these baseline models with different number of images for fine tuning demonstrated that the GPU memory consumption of MedSAGa is significantly less than that of the baseline models, achieving an average memory efficiency of 66% more than current state-of-the-art (SOTA) models for medical image segmentation. The combination of substantially lower memory requirements and comparable to SOTA results in few-shot learning for medical image segmentation positions MedSAGa as an optimal solution for deployment in resource-constrained settings.
Tuning the parameters and prompts for improving AI-based text-to-image generation has remained a substantial yet unaddressed challenge. Hence we introduce GreenStableYolo, which improves the parameters and prompts for Stable Diffusion to both reduce GPU inference time and increase image generation quality using NSGA-II and Yolo. Our experiments show that despite a relatively slight trade-off (18%) in image quality compared to StableYolo (which only considers image quality), GreenStableYolo achieves a substantial reduction in inference time (266% less) and a 526% higher hypervolume, thereby advancing the state-of-the-art for text-to-image generation.
We introduce NNsight and NDIF, technologies that work in tandem to enable scientific study of the representations and computations learned by very large neural networks. NNsight is an open-source system that extends PyTorch to introduce deferred remote execution. The National Deep Inference Fabric (NDIF) is a scalable inference service that executes NNsight requests, allowing users to share GPU resources and pretrained models. These technologies are enabled by the Intervention Graph, an architecture developed to decouple experimental design from model runtime. Together, this framework provides transparent and efficient access to the internals of deep neural networks such as very large language models (LLMs) without imposing the cost or complexity of hosting customized models individually. We conduct a quantitative survey of the machine learning literature that reveals a growing gap in the study of the internals of large-scale AI. We demonstrate the design and use of our framework to address this gap by enabling a range of research methods on huge models. Finally, we conduct benchmarks to compare performance with previous approaches. Code, documentation, and tutorials are available at https://nnsight.net/.
Public transportation plays a crucial role in our lives, and the road network is a vital component in the implementation of smart cities. Recent advancements in AI have enabled the development of advanced monitoring systems capable of detecting anomalies in road surfaces and road signs, which, if unaddressed, can lead to serious road accidents. This paper presents an innovative approach to enhance road safety through the detection and classification of traffic signs and road surface damage using advanced deep learning techniques. This integrated approach supports proactive maintenance strategies, improving road safety and resource allocation for the Molise region and the city of Campobasso. The resulting system, developed as part of the Casa delle Tecnologie Emergenti (House of Emergent Technologies) Molise (Molise CTE) research project funded by the Italian Minister of Economic Growth (MIMIT), leverages cutting-edge technologies such as Cloud Computing and High Performance Computing with GPU utilization. It serves as a valuable tool for municipalities, enabling quick detection of anomalies and the prompt organization of maintenance operations
Large Language Models (LLMs) are widely used across various domains, processing millions of daily requests. This surge in demand poses significant challenges in optimizing throughput and latency while keeping costs manageable. The Key-Value (KV) cache, a standard method for retaining previous computations, makes LLM inference highly bounded by memory. While batching strategies can enhance performance, they frequently lead to significant memory fragmentation. Even though cutting-edge systems like vLLM mitigate KV cache fragmentation using paged Attention mechanisms, they still suffer from inefficient memory and computational operations due to the tightly coupled page management and computation kernels. This study introduces the vTensor, an innovative tensor structure for LLM inference based on GPU virtual memory management (VMM). vTensor addresses existing limitations by decoupling computation from memory defragmentation and offering dynamic extensibility. Our framework employs a CPU-GPU heterogeneous approach, ensuring efficient, fragmentation-free memory management while accommodating various computation kernels across different LLM architectures. Experimental results indicate that vTensor achieves an average speedup of 1.86x across different models, with up to 2.42x in multi-turn chat scenarios. Additionally, vTensor provides average speedups of 2.12x and 3.15x in kernel evaluation, reaching up to 3.92x and 3.27x compared to SGLang Triton prefix-prefilling kernels and vLLM paged Attention kernel, respectively. Furthermore, it frees approximately 71.25% (57GB) of memory on the NVIDIA A100 GPU compared to vLLM, enabling more memory-intensive workloads.
Graph Neural Networks (GNNs) are widely used today in recommendation systems, fraud detection, and node/link classification tasks. Real world GNNs continue to scale in size and require a large memory footprint for storing graphs and embeddings that often exceed the memory capacities of the target GPUs used for training. To address limited memory capacities, traditional GNN training approaches use graph partitioning and sharding techniques to scale up across multiple GPUs within a node and/or scale out across multiple nodes. However, this approach suffers from the high computational costs of graph partitioning algorithms and inefficient communication across GPUs. To address these overheads, we propose Large-scale Storage-based Multi-GPU GNN framework (LSM-GNN), a storagebased approach to train GNN models that utilizes a novel communication layer enabling GPU software caches to function as a system-wide shared cache with low overheads.LSM-GNN incorporates a hybrid eviction policy that intelligently manages cache space by using both static and dynamic node information to significantly enhance cache performance. Furthermore, we introduce the Preemptive Victim-buffer Prefetcher (PVP), a mechanism for prefetching node feature data from a Victim Buffer located in CPU pinned-memory to further reduce the pressure on the storage devices. Experimental results show that despite the lower compute capabilities and memory capacities, LSM-GNN in a single node with two GPUs offers superior performance over two-node-four-GPU Dist-DGL baseline and provides up to 3.75x speed up on end-to-end epoch time while running large-scale GNN training
The increasing complexity of modern very-large-scale integration (VLSI) design highlights the significance of Electronic Design Automation (EDA) technologies. Chip placement is a critical step in the EDA workflow, which positions chip modules on the canvas with the goal of optimizing performance, power, and area (PPA) metrics of final chip designs. Recent advances have demonstrated the great potential of AI-based algorithms in enhancing chip placement. However, due to the lengthy workflow of chip design, the evaluations of these algorithms often focus on intermediate surrogate metrics, which are easy to compute but frequently reveal a substantial misalignment with the end-to-end performance (i.e., the final design PPA). To address this challenge, we introduce ChiPBench, which can effectively facilitate research in chip placement within the AI community. ChiPBench is a comprehensive benchmark specifically designed to evaluate the effectiveness of existing AI-based chip placement algorithms in improving final design PPA metrics. Specifically, we have gathered 20 circuits from various domains (e.g., CPU, GPU, and microcontrollers). These designs are compiled by executing the workflow from the verilog source code, which preserves necessary physical implementation kits, enabling evaluations for the placement algorithms on their impacts on the final design PPA. We executed six state-of-the-art AI-based chip placement algorithms on these designs and plugged the results of each single-point algorithm into the physical implementation workflow to obtain the final PPA results. Experimental results show that even if intermediate metric of a single-point algorithm is dominant, while the final PPA results are unsatisfactory. We believe that our benchmark will serve as an effective evaluation framework to bridge the gap between academia and industry.
Aligning future system design with the ever-increasing compute needs of large language models (LLMs) is undoubtedly an important problem in today's world. Here, we propose a general performance modeling methodology and workload analysis of distributed LLM training and inference through an analytical framework that accurately considers compute, memory sub-system, network, and various parallelization strategies (model parallel, data parallel, pipeline parallel, and sequence parallel). We validate our performance predictions with published data from literature and relevant industry vendors (e.g., NVIDIA). For distributed training, we investigate the memory footprint of LLMs for different activation re-computation methods, dissect the key factors behind the massive performance gain from A100 to B200 ($\sim$ 35x speed-up closely following NVIDIA's scaling trend), and further run a design space exploration at different technology nodes (12 nm to 1 nm) to study the impact of logic, memory, and network scaling on the performance. For inference, we analyze the compute versus memory boundedness of different operations at a matrix-multiply level for different GPU systems and further explore the impact of DRAM memory technology scaling on inference latency. Utilizing our modeling framework, we reveal the evolution of performance bottlenecks for both LLM training and inference with technology scaling, thus, providing insights to design future systems for LLM training and inference.
Neural networks (NNs) hold great promise for advancing inverse design via topology optimization (TO), yet misconceptions about their application persist. This article focuses on neural topology optimization (neural TO), which leverages NNs to reparameterize the decision space and reshape the optimization landscape. While the method is still in its infancy, our analysis tools reveal critical insights into the NNs' impact on the optimization process. We demonstrate that the choice of NN architecture significantly influences the objective landscape and the optimizer's path to an optimum. Notably, NNs introduce non-convexities even in otherwise convex landscapes, potentially delaying convergence in convex problems but enhancing exploration for non-convex problems. This analysis lays the groundwork for future advancements by highlighting: 1) the potential of neural TO for non-convex problems and dedicated GPU hardware (the "good"), 2) the limitations in smooth landscapes (the "bad"), and 3) the complex challenge of selecting optimal NN architectures and hyperparameters for superior performance (the "ugly").
When implementations of the Transformer's self-attention layer utilize SRAM instead of DRAM, they can achieve significant speedups. The Tenstorrent Grayskull architecture provides a large SRAM, distributed across a grid of cores. This work presents a fused kernel for Grayskull, that exclusively utilizes its large SRAM by combining matrix multiplication, attention score scaling and Softmax operations. Additionally, a dedicated Softmax kernel utilizing the SRAM and a CPU implementation serving as a baseline are presented. The Softmax operation consumes most of the runtime in the computation of attention weights from queries and keys on Grayskull. The speedup of the dedicated Softmax kernel compared to the CPU implementation is up to $10 \times$, and the Softmax implementation inside the fused kernel is approximately $1.8 \times$ faster than the dedicated Softmax kernel. The time and memory complexity of all implementations is quadratic in sequence length. Currently, the Grayskull e150 is approximately $30 \times$ cheaper for the general public than an Nvidia H100 PCIe (a state-of-the-art GPU) and offers approximately $1.5 \times$ more SRAM.
Deep learning kernels exhibit predictable memory accesses and compute patterns, making GPUs' parallel architecture well-suited for their execution. Software and runtime systems for GPUs are optimized to better utilize the stream multiprocessors, on-chip cache, and off-chip high-bandwidth memory. As deep learning models and GPUs evolve, access to newer GPUs is often limited, raising questions about the performance of new model architectures on existing GPUs, existing models on new GPUs, and new model architectures on new GPUs. To address these questions, we introduce NeuSight, a framework to predict the performance of various deep learning models, for both training and inference, on unseen GPUs without requiring actual execution. The framework leverages both GPU hardware behavior and software library optimizations to estimate end-to-end performance. Previous work uses regression models that capture linear trends or multilayer perceptrons to predict the overall latency of deep learning kernels on GPUs. These approaches suffer from higher error percentages when forecasting performance on unseen models and new GPUs. Instead, NeuSight decomposes the prediction problem into smaller problems, bounding the prediction through fundamental performance laws. NeuSight decomposes a single deep learning kernel prediction into smaller working sets called tiles, which are executed independently on the GPU. Tile-granularity predictions are determined using a machine learning approach and aggregated to estimate end-to-end latency. NeuSight outperforms prior work across various deep learning workloads and the latest GPUs. It reduces the percentage error from 121.4% and 30.8% to 2.3% in predicting the latency of GPT3 model for training and inference on H100, compared to state-of-the-art prior work, where both GPT3 and H100 were not used to train the framework.
The increasing demand for deploying large Mixture-of-Experts (MoE) models in resource-constrained environments necessitates efficient approaches to address their high memory and computational requirements challenges. Moreover, given that tasks come in different user-defined constraints and the available resources change over time in multi-tenant environments, it is necessary to design an approach which provides a flexible configuration space. This paper presents an adaptive serving approach for the efficient deployment of MoE models, capitalizing on partial quantization of the experts. By dynamically determining the number of quantized experts and their distribution across CPU and GPU, our approach explores the Pareto frontier and offers a fine-grained range of configurations for tuning throughput and model quality. Our evaluation on an NVIDIA A100 GPU using a Mixtral 8x7B MoE model for three language modelling benchmarks demonstrates that the throughput of token generation can be adjusted from 0.63 to 13.00 token per second. This enhancement comes with a marginal perplexity increase of 3.81 to 4.00, 13.59 to 14.17, and 7.24 to 7.40 for WikiText2, PTB, and C4 datasets respectively under maximum quantization. These results highlight the practical applicability of our approach in dynamic and accuracy-sensitive applications where both memory usage and output quality are important.
As the field of Large Language Models (LLMs) continues to evolve, the context length in inference is steadily growing. Key-Value Cache (KVCache), the intermediate representations of tokens within LLM inference, has now become the primary memory bottleneck due to limited GPU memory. Current methods selectively determine suitable keys and values for self-attention computation in LLMs to address the issue. However, they either fall short in maintaining model quality or result in high serving latency. Drawing inspiration from advanced embedding retrieval techniques prevalent in the data management community, we consider the storage and retrieval of KVCache as a typical embedding retrieval problem. We propose PQCache, which employs Product Quantization (PQ) to manage KVCache, maintaining model quality while ensuring low serving latency. During the prefilling phase, we apply PQ to tokens' keys for each LLM layer and head. During the autoregressive decoding phase, we use PQ codes and centroids to approximately identify important preceding tokens, then fetch the corresponding key-value pairs for self-attention computation. Through meticulous design of overlapping and caching, we minimize any additional computation and communication overhead during both phases. Extensive experiments demonstrate that PQCache achieves both effectiveness and efficiency, with 4.60% score improvement over existing methods on InfiniteBench and low system latency in both prefilling and decoding.
Accurate segmentation of anatomical structures and pathological regions in medical images is crucial for diagnosis, treatment planning, and disease monitoring. While the Segment Anything Model (SAM) and its variants have demonstrated impressive interactive segmentation capabilities on image types not seen during training without the need for domain adaptation or retraining, their practical application in volumetric 3D medical imaging workflows has been hindered by the lack of a user-friendly interface. To address this challenge, we introduce FastSAM-3DSlicer, a 3D Slicer extension that integrates both 2D and 3D SAM models, including SAM-Med2D, MedSAM, SAM-Med3D, and FastSAM-3D. Building on the well-established open-source 3D Slicer platform, our extension enables efficient, real-time segmentation of 3D volumetric medical images, with seamless interaction and visualization. By automating the handling of raw image data, user prompts, and segmented masks, FastSAM-3DSlicer provides a streamlined, user-friendly interface that can be easily incorporated into medical image analysis workflows. Performance evaluations reveal that the FastSAM-3DSlicer extension running FastSAM-3D achieves low inference times of only 1.09 seconds per volume on CPU and 0.73 seconds per volume on GPU, making it well-suited for real-time interactive segmentation. Moreover, we introduce an uncertainty quantification scheme that leverages the rapid inference capabilities of FastSAM-3D for practical implementation, further enhancing its reliability and applicability in medical settings. FastSAM-3DSlicer offers an interactive platform and user interface for 2D and 3D interactive volumetric medical image segmentation, offering a powerful combination of efficiency, precision, and ease of use with SAMs. The source code and a video demonstration are publicly available at https://github.com/arcadelab/FastSAM3D_slicer.