Computing in the Era of Large Generative Models: From Cloud-Native to AI-Native

In this paper, we investigate the intersection of large generative AI models and cloud-native computing architectures. Recent large models such as ChatGPT, while revolutionary in their capabilities, face challenges like escalating costs and demand for high-end GPUs. Drawing analogies between large-model-as-a-service (LMaaS) and cloud database-as-a-service (DBaaS), we describe an AI-native computing paradigm that harnesses the power of both cloud-native technologies (e.g., multi-tenancy and serverless computing) and advanced machine learning runtime (e.g., batched LoRA inference). These joint efforts aim to optimize costs-of-goods-sold (COGS) and improve resource accessibility. The journey of merging these two domains is just at the beginning and we hope to stimulate future research and development in this area.

Atom: Low-bit Quantization for Efficient and Accurate LLM Serving

The growing demand for Large Language Models (LLMs) in applications such as content generation, intelligent chatbots, and sentiment analysis poses considerable challenges for LLM service providers. To efficiently use GPU resources and boost throughput, batching multiple requests has emerged as a popular paradigm; to further speed up batching, LLM quantization techniques reduce memory consumption and increase computing capacity. However, prevalent quantization schemes (e.g., 8-bit weight-activation quantization) cannot fully leverage the capabilities of modern GPUs, such as 4-bit integer operators, resulting in sub-optimal performance. To maximize LLMs' serving throughput, we introduce Atom, a low-bit quantization method that achieves high throughput improvements with negligible accuracy loss. Atom significantly boosts serving throughput by using low-bit operators and considerably reduces memory consumption via low-bit quantization. It attains high accuracy by applying a novel mixed-precision and fine-grained quantization process. We evaluate Atom on 4-bit weight-activation quantization setups in the serving context. Atom improves end-to-end throughput by up to $7.73\times$ compared to the FP16 and by $2.53\times$ compared to INT8 quantization, while maintaining the same latency target.

Punica: Multi-Tenant LoRA Serving

Low-rank adaptation (LoRA) has become an important and popular method to adapt pre-trained models to specific domains. We present Punica, a system to serve multiple LoRA models in a shared GPU cluster. Punica contains a new CUDA kernel design that allows batching of GPU operations for different LoRA models. This allows a GPU to hold only a single copy of the underlying pre-trained model when serving multiple, different LoRA models, significantly enhancing GPU efficiency in terms of both memory and computation. Our scheduler consolidates multi-tenant LoRA serving workloads in a shared GPU cluster. With a fixed-sized GPU cluster, our evaluations show that Punica achieves 12x higher throughput in serving multiple LoRA models compared to state-of-the-art LLM serving systems while only adding 2ms latency per token. Punica is open source at https://github.com/punica-ai/punica .

Symphony: Optimized Model Serving using Centralized Orchestration

The orchestration of deep neural network (DNN) model inference on GPU clusters presents two significant challenges: achieving high accelerator efficiency given the batching properties of model inference while meeting latency service level objectives (SLOs), and adapting to workload changes both in terms of short-term fluctuations and long-term resource allocation. To address these challenges, we propose Symphony, a centralized scheduling system that can scale to millions of requests per second and coordinate tens of thousands of GPUs. Our system utilizes a non-work-conserving scheduling algorithm capable of achieving high batch efficiency while also enabling robust autoscaling. Additionally, we developed an epoch-scale algorithm that allocates models to sub-clusters based on the compute and memory needs of the models. Through extensive experiments, we demonstrate that Symphony outperforms prior systems by up to 4.7x higher goodput.

Multimodal sensor fusion for real-time location-dependent defect detection in laser-directed energy deposition

Real-time defect detection is crucial in laser-directed energy deposition (L-DED) additive manufacturing (AM). Traditional in-situ monitoring approach utilizes a single sensor (i.e., acoustic, visual, or thermal sensor) to capture the complex process dynamic behaviors, which is insufficient for defect detection with high accuracy and robustness. This paper proposes a novel multimodal sensor fusion method for real-time location-dependent defect detection in the robotic L-DED process. The multimodal fusion sources include a microphone sensor capturing the laser-material interaction sound and a visible spectrum CCD camera capturing the coaxial melt pool images. A hybrid convolutional neural network (CNN) is proposed to fuse acoustic and visual data. The key novelty in this study is that the traditional manual feature extraction procedures are no longer required, and the raw melt pool images and acoustic signals are fused directly by the hybrid CNN model, which achieved the highest defect prediction accuracy (98.5 %) without the thermal sensing modality. Moreover, unlike previous region-based quality prediction, the proposed hybrid CNN can detect the onset of defect occurrences. The defect prediction outcomes are synchronized and registered with in-situ acquired robot tool-center-point (TCP) data, which enables localized defect identification. The proposed multimodal sensor fusion method offers a robust solution for in-situ defect detection.

Multisensor fusion-based digital twin in additive manufacturing for in-situ quality monitoring and defect correction

Early detection and correction of defects are critical in additive manufacturing (AM) to avoid build failures. In this paper, we present a multisensor fusion-based digital twin for in-situ quality monitoring and defect correction in a robotic laser direct energy deposition process. Multisensor fusion sources consist of an acoustic sensor, an infrared thermal camera, a coaxial vision camera, and a laser line scanner. The key novelty and contribution of this work are to develop a spatiotemporal data fusion method that synchronizes and registers the multisensor features within the part's 3D volume. The fused dataset can be used to predict location-specific quality using machine learning. On-the-fly identification of regions requiring material addition or removal is feasible. Robot toolpath and auto-tuned process parameters are generated for defecting correction. In contrast to traditional single-sensor-based monitoring, multisensor fusion allows for a more in-depth understanding of underlying process physics, such as pore formation and laser-material interactions. The proposed methods pave the way for self-adaptation AM with higher efficiency, less waste, and cleaner production.

In-situ crack and keyhole pore detection in laser directed energy deposition through acoustic signal and deep learning

Cracks and keyhole pores are detrimental defects in alloys produced by laser directed energy deposition (LDED). Laser-material interaction sound may hold information about underlying complex physical events such as crack propagation and pores formation. However, due to the noisy environment and intricate signal content, acoustic-based monitoring in LDED has received little attention. This paper proposes a novel acoustic-based in-situ defect detection strategy in LDED. The key contribution of this study is to develop an in-situ acoustic signal denoising, feature extraction, and sound classification pipeline that incorporates convolutional neural networks (CNN) for online defect prediction. Microscope images are used to identify locations of the cracks and keyhole pores within a part. The defect locations are spatiotemporally registered with acoustic signal. Various acoustic features corresponding to defect-free regions, cracks, and keyhole pores are extracted and analysed in time-domain, frequency-domain, and time-frequency representations. The CNN model is trained to predict defect occurrences using the Mel-Frequency Cepstral Coefficients (MFCCs) of the lasermaterial interaction sound. The CNN model is compared to various classic machine learning models trained on the denoised acoustic dataset and raw acoustic dataset. The validation results shows that the CNN model trained on the denoised dataset outperforms others with the highest overall accuracy (89%), keyhole pore prediction accuracy (93%), and AUC-ROC score (98%). Furthermore, the trained CNN model can be deployed into an in-house developed software platform for online quality monitoring. The proposed strategy is the first study to use acoustic signals with deep learning for insitu defect detection in LDED process.

ADARES: Adaptive Resource Management for Virtual Machines

Virtual execution environments allow for consolidation of multiple applications onto the same physical server, thereby enabling more efficient use of server resources. However, users often statically configure the resources of virtual machines through guesswork, resulting in either insufficient resource allocations that hinder VM performance, or excessive allocations that waste precious data center resources. In this paper, we first characterize real-world resource allocation and utilization of VMs through the analysis of an extensive dataset, consisting of more than 250k VMs from over 3.6k private enterprise clusters. Our large-scale analysis confirms that VMs are often misconfigured, either overprovisioned or underprovisioned, and that this problem is pervasive across a wide range of private clusters. We then propose ADARES, an adaptive system that dynamically adjusts VM resources using machine learning techniques. In particular, ADARES leverages the contextual bandits framework to effectively manage the adaptations. Our system exploits easily collectible data, at the cluster, node, and VM levels, to make more sensible allocation decisions, and uses transfer learning to safely explore the configurations space and speed up training. Our empirical evaluation shows that ADARES can significantly improve system utilization without sacrificing performance. For instance, when compared to threshold and prediction-based baselines, it achieves more predictable VM-level performance and also reduces the amount of virtual CPUs and memory provisioned by up to 35% and 60% respectively for synthetic workloads on real clusters.