Lequn Chen

Lequn Chen, Zihao Ye, Yongji Wu et al. · uw

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 .

Yilong Zhao, Chien-Yu Lin, Kan Zhu et al. · uw

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 in the serving context. Atom improves end-to-end throughput (token/s) by up to $7.7\times$ compared to the FP16 and by $2.5\times$ compared to INT8 quantization, while maintaining the same latency target.

Nandor Licker, Kevin Hu, Vladimir Zaytsev et al.

Emerging Large Language Model (LLM) system patterns, such as disaggregated inference, Mixture-of-Experts (MoE) routing, and asynchronous reinforcement fine-tuning, require flexible point-to-point communication beyond simple collectives. Existing implementations are locked to specific Network Interface Controllers (NICs), hindering integration into inference engines and portability across hardware providers. We present fabric-lib, which bridges the functionality of common NICs to expose a uniform interface. fabric-lib exposes one-sided WriteImm operations with a ImmCounter primitive for completion notification, without ordering assumptions of network transport, transparently managing multiple NICs per GPU. We demonstrate peak throughput of 400 Gbps on both NVIDIA ConnectX-7 and AWS Elastic Fabric Adapter (EFA). We showcase fabric-lib through three production systems: (1) KvCache transfer for disaggregated inference with dynamic scaling, (2) RL weight updates achieving 1.3 seconds for trillion-parameter models, and (3) MoE dispatch/combine implementation exceeding DeepEP decode latency on ConnectX-7, with the first viable latencies on EFA. We demonstrate that our portable point-to-point communication complements collectives while avoiding lock-in. fabric-lib is open-sourced at https://github.com/perplexityai/pplx-garden/

Lequn Chen, Weixin Deng, Anirudh Canumalla et al.

Having large batch sizes is one of the most critical aspects of increasing the accelerator efficiency and the performance of DNN model inference. However, existing model serving systems cannot achieve adequate batch sizes while meeting latency objectives as these systems eagerly dispatch requests to accelerators to minimize the accelerator idle time. We propose Symphony, a DNN serving system that explores deferred batch scheduling to optimize system efficiency and throughput. Further, unlike other prior systems, Symphony's GPU usage is load-proportional: it consolidates workloads on the appropriate number of GPUs and works smoothly with cluster auto-scaling tools. Symphony consists of two core design points. First, Symphony defines a schedulable window in which a batch of inference requests can be dispatched. This window is computed in order to improve accelerator efficiency while meeting the request's SLO. Second, Symphony implements a scalable, low-latency, fine-grained coordination scheme across accelerators to dispatch and execute requests in the schedulable window. Through extensive scheduler-only benchmarks, we demonstrate that Symphony can schedule millions of requests per second and coordinate thousands of GPUs while also enabling robust autoscaling that adapts to workload changes. Symphony outperforms prior systems by achieving 5x higher goodput when given the same number of GPUs and 60% reduction in GPUs when given the same workload.

Jiarui Xie, Mutahar Safdar, Lequn Chen et al.

Various machine learning (ML)-based in-situ monitoring systems have been developed to detect anomalies and defects in laser additive manufacturing (LAM) processes. While multimodal fusion, which integrates data from visual, audio, and other modalities, can improve monitoring performance, it also increases hardware, computational, and operational costs. This paper introduces a cross-modality knowledge transfer (CMKT) methodology for LAM in-situ monitoring, which transfers knowledge from a source modality to a target modality. CMKT enhances the representativeness of the features extracted from the target modality, allowing the removal of source modality sensors during prediction. This paper proposes three CMKT methods: semantic alignment, fully supervised mapping, and semi-supervised mapping. The semantic alignment method establishes a shared encoded space between modalities to facilitate knowledge transfer. It employs a semantic alignment loss to align the distributions of identical groups (e.g., visual and audio defective groups) and a separation loss to distinguish different groups (e.g., visual defective and audio defect-free groups). The two mapping methods transfer knowledge by deriving features from one modality to another using fully supervised and semi-supervised learning approaches. In a case study for LAM in-situ defect detection, the proposed CMKT methods were compared with multimodal audio-visual fusion. The semantic alignment method achieved an accuracy of 98.6% while removing the audio modality during the prediction phase, which is comparable to the 98.2% accuracy obtained through multimodal fusion. Using explainable artificial intelligence, we discovered that semantic alignment CMKT can extract more representative features while reducing noise by leveraging the inherent correlations between modalities.

Muhammad Tayyab Khan, Wenhe Feng, Lequn Chen et al.

The integration of Computer-Aided Design (CAD), Computer-Aided Process Planning (CAPP), and Computer-Aided Manufacturing (CAM) plays a crucial role in modern manufacturing, facilitating seamless transitions from digital designs to physical products. However, a significant challenge within this integration is the Automatic Feature Recognition (AFR) of CAD models, especially in the context of hybrid manufacturing that combines subtractive and additive manufacturing processes. Traditional AFR methods, focused mainly on the identification of subtractive (machined) features including holes, fillets, chamfers, pockets, and slots, fail to recognize features pertinent to additive manufacturing. Furthermore, the traditional methods fall short in accurately extracting geometric dimensions and orientations, which are also key factors for effective manufacturing process planning. This paper presents a novel approach for creating a synthetic CAD dataset that encompasses features relevant to both additive and subtractive machining through Python Open Cascade. The Hierarchical Graph Convolutional Neural Network (HGCNN) model is implemented to accurately identify the composite additive-subtractive features within the synthetic CAD dataset. The key novelty and contribution of the proposed methodology lie in its ability to recognize a wide range of manufacturing features, and precisely extracting their dimensions, orientations, and stock sizes. The proposed model demonstrates remarkable feature recognition accuracy exceeding 97% and a dimension extraction accuracy of 100% for identified features. Therefore, the proposed methodology enhances the integration of CAD, CAPP, and CAM within hybrid manufacturing by providing precise feature recognition and dimension extraction. It facilitates improved manufacturing process planning, by enabling more informed decision-making.

Muhammad Tayyab Khan, Lequn Chen, Ye Han Ng et al.

Geometric Dimensioning and Tolerancing (GD&T) plays a critical role in manufacturing by defining acceptable variations in part features to ensure component quality and functionality. However, extracting GD&T information from 2D engineering drawings is a time-consuming and labor-intensive task, often relying on manual efforts or semi-automated tools. To address these challenges, this study proposes an automated and computationally efficient GD&T extraction method by fine-tuning Florence-2, an open-source vision-language model (VLM). The model is trained on a dataset of 400 drawings with ground truth annotations provided by domain experts. For comparison, two state-of-the-art closed-source VLMs, GPT-4o and Claude-3.5-Sonnet, are evaluated on the same dataset. All models are assessed using precision, recall, F1-score, and hallucination metrics. Due to the computational cost and impracticality of fine-tuning large closed-source VLMs for domain-specific tasks, GPT-4o and Claude-3.5-Sonnet are evaluated in a zero-shot setting. In contrast, Florence-2, a smaller model with 0.23 billion parameters, is optimized through full-parameter fine-tuning across three distinct experiments, each utilizing datasets augmented to different levels. The results show that Florence-2 achieves a 29.95% increase in precision, a 37.75% increase in recall, a 52.40% improvement in F1-score, and a 43.15% reduction in hallucination rate compared to the best-performing closed-source model. These findings highlight the effectiveness of fine-tuning smaller, open-source VLMs like Florence-2, offering a practical and efficient solution for automated GD&T extraction to support downstream manufacturing tasks.

Muhammad Tayyab Khan, Lequn Chen, Zane Yong et al.

Efficient and accurate extraction of key information from 2D engineering drawings is essential for advancing digital manufacturing workflows. Such information includes geometric dimensioning and tolerancing (GD&T), measures, material specifications, and textual annotations. Manual extraction is slow and labor-intensive, while generic OCR models often fail due to complex layouts, engineering symbols, and rotated text, leading to incomplete and unreliable outputs. These limitations result in incomplete and unreliable outputs. To address these challenges, we propose a hybrid vision-language framework that integrates a rotation-aware object detection model (YOLOv11-obb) with a transformer-based vision-language parser. Our structured pipeline applies YOLOv11-OBB to localize annotations and extract oriented bounding box (OBB) patches, which are then parsed into structured outputs using a fine-tuned, lightweight vision-language model (VLM). We curate a dataset of 1,367 2D mechanical drawings annotated across nine key categories. YOLOv11-OBB is trained on this dataset to detect OBBs and extract annotation patches. These are parsed using two open-source VLMs: Donut and Florence-2. Both models are lightweight and well-suited for specialized industrial tasks under limited computational overhead. Following fine-tuning of both models on the curated dataset of image patches paired with structured annotation labels, a comparative experiment is conducted to evaluate parsing performance across four key metrics. Donut outperforms Florence-2, achieving 88.5% precision, 99.2% recall, and a 93.5% F1-score, with a hallucination rate of 11.5%. Finally, a case study demonstrates how the extracted structured information supports downstream manufacturing tasks such as process and tool selection, showcasing the practical utility of the proposed framework in modernizing 2D drawing interpretation.

Zihao Ye, Lequn Chen, Ruihang Lai et al. · openai, uw

Transformers, driven by attention mechanisms, form the foundation of large language models (LLMs). As these models scale up, efficient GPU attention kernels become essential for high-throughput and low-latency inference. Diverse LLM applications demand flexible and high-performance attention solutions. We present FlashInfer: a customizable and efficient attention engine for LLM serving. FlashInfer tackles KV-cache storage heterogeneity using block-sparse format and composable formats to optimize memory access and reduce redundancy. It also offers a customizable attention template, enabling adaptation to various settings through Just-In-Time (JIT) compilation. Additionally, FlashInfer's load-balanced scheduling algorithm adjusts to dynamism of user requests while maintaining compatibility with CUDAGraph which requires static configuration. FlashInfer have been integrated into leading LLM serving frameworks like SGLang, vLLM and MLC-Engine. Comprehensive kernel-level and end-to-end evaluations demonstrate FlashInfer's ability to significantly boost kernel performance across diverse inference scenarios: compared to state-of-the-art LLM serving solutions, FlashInfer achieve 29-69% inter-token-latency reduction compared to compiler backends for LLM serving benchmark, 28-30% latency reduction for long-context inference, and 13-17% speedup for LLM serving with parallel generation.

Yao Lu, Song Bian, Lequn Chen et al.

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.

Muhammad Tayyab Khan, Zane Yong, Lequn Chen et al.

Accurate extraction of key information from 2D engineering drawings is crucial for high-precision manufacturing. Manual extraction is slow and labor-intensive, while traditional Optical Character Recognition (OCR) techniques often struggle with complex layouts and overlapping symbols, resulting in unstructured outputs. To address these challenges, this paper proposes a novel hybrid deep learning framework for structured information extraction by integrating an Oriented Bounding Box (OBB) detection model with a transformer-based document parsing model (Donut). An in-house annotated dataset is used to train YOLOv11 for detecting nine key categories: Geometric Dimensioning and Tolerancing (GD&T), General Tolerances, Measures, Materials, Notes, Radii, Surface Roughness, Threads, and Title Blocks. Detected OBBs are cropped into images and labeled to fine-tune Donut for structured JSON output. Fine-tuning strategies include a single model trained across all categories and category-specific models. Results show that the single model consistently outperforms category-specific ones across all evaluation metrics, achieving higher precision (94.77% for GD&T), recall (100% for most categories), and F1 score (97.3%), while reducing hallucinations (5.23%). The proposed framework improves accuracy, reduces manual effort, and supports scalable deployment in precision-driven industries.

Muhammad Tayyab Khan, Zane Yong, Lequn Chen et al.

Engineering drawings are fundamental to manufacturing communication, serving as the primary medium for conveying design intent, tolerances, and production details. However, interpreting complex multi-view drawings with dense annotations remains challenging using manual methods, generic optical character recognition (OCR) systems, or traditional deep learning approaches, due to varied layouts, orientations, and mixed symbolic-textual content. To address these challenges, this paper proposes a three-stage hybrid framework for the automated interpretation of 2D multi-view engineering drawings using modern detection and vision language models (VLMs). In the first stage, YOLOv11-det performs layout segmentation to localize key regions such as views, title blocks, and notes. The second stage uses YOLOv11-obb for orientation-aware, fine-grained detection of annotations, including measures, GD&T symbols, and surface roughness indicators. The third stage employs two Donut-based, OCR-free VLMs for semantic content parsing: the Alphabetical VLM extracts textual and categorical information from title blocks and notes, while the Numerical VLM interprets quantitative data such as measures, GD&T frames, and surface roughness. Two specialized datasets were developed to ensure robustness and generalization: 1,000 drawings for layout detection and 1,406 for annotation-level training. The Alphabetical VLM achieved an overall F1 score of 0.672, while the Numerical VLM reached 0.963, demonstrating strong performance in textual and quantitative interpretation, respectively. The unified JSON output enables seamless integration with CAD and manufacturing databases, providing a scalable solution for intelligent engineering drawing analysis.

Muhammad Tayyab Khan, Lequn Chen, Wenhe Feng et al.

Metal additive manufacturing (AM) involves complex interdependencies among processes, materials, feedstock, and post-processing steps. However, the underlying relationships and domain knowledge remain fragmented across literature and static databases that often require expert-level queries, limiting their applicability in design and planning. To address these limitations, we develop a novel and structured knowledge graph (KG), representing 53 distinct metals and alloys across seven material categories, nine AM processes, four feedstock types, and corresponding post-processing requirements. A large language model (LLM) interface, guided by a few-shot prompting strategy, enables natural language querying without the need for formal query syntax. The system supports a range of tasks, including compatibility evaluation, constraint-based filtering, and design for AM (DfAM) guidance. User queries in natural language are normalized, translated into Cypher, and executed on the KG, with results returned in a structured format. This work introduces the first interactive system that connects a domain-specific metal AM KG with an LLM interface, delivering accessible and explainable decision support for engineers and promoting human-centered tools in manufacturing knowledge systems.

Ignacio Cano, Lequn Chen, Pedro Fonseca et al.

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.

Kai Mast, Lequn Chen, Emin Gün Sirer

Many applications require the immutable and consistent sharing of data across organizational boundaries. Because conventional datastores cannot provide this functionality, blockchains have been proposed as one possible solution. Yet public blockchains are energy inefficient, hard to scale and suffer from limited throughput and high latencies, while permissioned blockchains depend on specially designated nodes, potentially leak meta-information, and also suffer from scale and performance bottlenecks. This paper presents CreDB, a datastore that provides blockchain-like guarantees of integrity using trusted execution environments. CreDB employs four novel mechanisms to support a new class of applications. First, it creates a permanent record of every transaction, known as a witness, that clients can then use not only to audit the database but to prove to third parties that desired actions took place. Second, it associates with every object an inseparable and inviolable policy, which not only performs access control but enables the datastore to implement state machines whose behavior is amenable to analysis. Third, timeline inspection allows authorized parties to inspect and reason about the history of changes made to the data. Finally, CreDB provides a protected function evaluation mechanism that allows integrity-protected computation over private data. The paper describes these mechanisms, and the applications they collectively enable, in detail. We have fully implemented a prototype of CreDB on Intel SGX. Evaluation shows that CreDB can serve as a drop-in replacement for other NoSQL stores, such as MongoDB while providing stronger integrity guarantees.