Arvind Krishnamurthy

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.

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.

Thierry Moreau, Tianqi Chen, Luis Vega et al.

Specialized Deep Learning (DL) acceleration stacks, designed for a specific set of frameworks, model architectures, operators, and data types, offer the allure of high performance while sacrificing flexibility. Changes in algorithms, models, operators, or numerical systems threaten the viability of specialized hardware accelerators. We propose VTA, a programmable deep learning architecture template designed to be extensible in the face of evolving workloads. VTA achieves this flexibility via a parametrizable architecture, two-level ISA, and a JIT compiler. The two-level ISA is based on (1) a task-ISA that explicitly orchestrates concurrent compute and memory tasks and (2) a microcode-ISA which implements a wide variety of operators with single-cycle tensor-tensor operations. Next, we propose a runtime system equipped with a JIT compiler for flexible code-generation and heterogeneous execution that enables effective use of the VTA architecture. VTA is integrated and open-sourced into Apache TVM, a state-of-the-art deep learning compilation stack that provides flexibility for diverse models and divergent hardware backends. We propose a flow that performs design space exploration to generate a customized hardware architecture and software operator library that can be leveraged by mainstream learning frameworks. We demonstrate our approach by deploying optimized deep learning models used for object classification and style transfer on edge-class FPGAs.

Tianqi Chen, Thierry Moreau, Ziheng Jiang et al.

There is an increasing need to bring machine learning to a wide diversity of hardware devices. Current frameworks rely on vendor-specific operator libraries and optimize for a narrow range of server-class GPUs. Deploying workloads to new platforms -- such as mobile phones, embedded devices, and accelerators (e.g., FPGAs, ASICs) -- requires significant manual effort. We propose TVM, a compiler that exposes graph-level and operator-level optimizations to provide performance portability to deep learning workloads across diverse hardware back-ends. TVM solves optimization challenges specific to deep learning, such as high-level operator fusion, mapping to arbitrary hardware primitives, and memory latency hiding. It also automates optimization of low-level programs to hardware characteristics by employing a novel, learning-based cost modeling method for rapid exploration of code optimizations. Experimental results show that TVM delivers performance across hardware back-ends that are competitive with state-of-the-art, hand-tuned libraries for low-power CPU, mobile GPU, and server-class GPUs. We also demonstrate TVM's ability to target new accelerator back-ends, such as the FPGA-based generic deep learning accelerator. The system is open sourced and in production use inside several major companies.

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.

Kan Zhu, Tian Tang, Qinyu Xu et al.

Long-context models are essential for many applications but face inefficiencies in loading large KV caches during decoding. Prior methods enforce fixed token budgets for sparse attention, assuming a set number of tokens can approximate full attention. However, these methods overlook variations in the importance of attention across heads, layers, and contexts. To address these limitations, we propose Tactic, a sparsity-adaptive and calibration-free sparse attention mechanism that dynamically selects tokens based on their cumulative attention scores rather than a fixed token budget. By setting a target fraction of total attention scores, Tactic ensures that token selection naturally adapts to variations in attention sparsity. To efficiently approximate this selection, Tactic leverages clustering-based sorting and distribution fitting, allowing it to accurately estimate token importance with minimal computational overhead. We show that Tactic outperforms existing sparse attention algorithms, achieving superior accuracy and up to 7.29x decode attention speedup. This improvement translates to an overall 1.58x end-to-end inference speedup, making Tactic a practical and effective solution for long-context LLM inference in accuracy-sensitive applications.

Yifan Yu, Yu Gan, Nikhil Sarda et al.

Large language models (LLMs) have excelled in various applications, yet serving them at scale is challenging due to their substantial resource demands and high latency. Our real-world studies reveal that over 70% of user requests to LLMs have semantically similar counterparts, suggesting the potential for knowledge transfer among requests. However, naively caching and reusing past responses leads to a big quality drop. In this paper, we introduce IC-Cache, a caching system that enables live LLM capability augmentation to improve serving efficiency: by leveraging historical request-response pairs from larger models as in-context examples, IC-Cache empowers small LLMs to imitate and even exceed the compositional abilities (e.g., reasoning) of their larger counterparts, enabling selective offloading of requests to reduce cost and latency. Achieving this live augmentation at scale introduces intricate trade-offs between response quality, latency, and system throughput. For a new request, IC-Cache efficiently selects similar, high-utility examples to prepend them to the new request's input. At scale, it adaptively routes requests across LLMs of varying capabilities, accounting for response quality and serving loads. IC-Cache employs a cost-aware cache replay mechanism that refines example quality offline to maximize online cache utility and efficiency. Evaluations on millions of realistic requests demonstrate that IC-Cache improves LLM serving throughput by 1.4-5.9x and reduces latency by 28-71% without hurting response quality.

Yongji Wu, Xueshen Liu, Haizhong Zheng et al.

Reinforcement learning (RL) has become essential for unlocking advanced reasoning capabilities in large language models (LLMs). RL workflows involve interleaving rollout and training stages with fundamentally different resource requirements. Rollout typically dominates overall execution time, yet scales efficiently through multiple independent instances. In contrast, training requires tightly-coupled GPUs with full-mesh communication. Existing RL frameworks fall into two categories: co-located and disaggregated architectures. Co-located ones fail to address this resource tension by forcing both stages to share the same GPUs. Disaggregated architectures, without modifications of well-established RL algorithms, suffer from resource under-utilization. Meanwhile, preemptible GPU resources, i.e., spot instances on public clouds and spare capacity in production clusters, present significant cost-saving opportunities for accelerating RL workflows, if efficiently harvested for rollout. In this paper, we present RLBoost, a systematic solution for cost-efficient RL training that harvests preemptible GPU resources. Our key insight is that rollout's stateless and embarrassingly parallel nature aligns perfectly with preemptible and often fragmented resources. To efficiently utilize these resources despite frequent and unpredictable availability changes, RLBoost adopts a hybrid architecture with three key techniques: (1) adaptive rollout offload to dynamically adjust workloads on the reserved (on-demand) cluster, (2) pull-based weight transfer that quickly provisions newly available instances, and (3) token-level response collection and migration for efficient preemption handling and continuous load balancing. Extensive experiments show RLBoost increases training throughput by 1.51x-1.97x while improving cost efficiency by 28%-49% compared to using only on-demand GPU resources.

Fanjiang Ye, Zepeng Zhao, Yi Mu et al.

Diffusion models have recently achieved remarkable success in generative tasks (e.g., image and video generation), and the demand for high-quality content (e.g., 2K/4K videos) is rapidly increasing across various domains. However, generating ultra-high-resolution videos on existing standard-resolution (e.g., 720p) platforms remains challenging due to the excessive re-training requirements and prohibitively high computational and memory costs. To this end, we introduce SuperGen, an efficient tile-based framework for ultra-high-resolution video generation. SuperGen features a novel training-free algorithmic innovation with tiling to successfully support a wide range of resolutions without additional training efforts while significantly reducing both memory footprint and computational complexity. Moreover, SuperGen incorporates a tile-tailored, adaptive, region-aware caching strategy that accelerates video generation by exploiting redundancy across denoising steps and spatial regions. SuperGen also integrates cache-guided, communication-minimized tile parallelism for enhanced throughput and minimized latency. Evaluations demonstrate that SuperGen harvests the maximum performance gains while achieving high output quality across various benchmarks.

Kan Zhu, Haiyang Shi, Le Xu et al.

Advances in Large Language Models (LLMs) have led to a surge of LLM-powered applications. These applications have diverse token-generation latency requirements. As a result, simply classifying workloads as latency-sensitive (LS) or best-effort (BE) overlooks the nuances within the latency-sensitive category and results in suboptimal user experiences and scheduling opportunities. However, efficiently serving requests with multiple SLO requirements poses significant challenges. First, all requests within a batch generate new tokens simultaneously, which can misalign them with their distinct SLO requirements. Moreover, while existing systems focus on auto-scaling for handling various overall request rates, the diversity of SLOs necessitates fine-grained auto-scaling among these SLO tiers. Finally, unlike LS/BE scenarios, where BE requests can be aborted at any time to ensure the SLO attainment of LS requests, those with different latency-sensitive SLOs cannot tolerate prolonged delays, and tail latency must be controlled. To tackle these challenges, we propose PolyServe, a novel multi-SLO scheduling policy at scale that maintains high SLO attainment while maximizing throughput. PolyServe first groups requests into multiple bins based on their per-token latency requirement, then schedules each bin to a subset of the server fleet. PolyServe routes requests to the highest-load but still SLO-attainable server to create a load gradient that facilitates auto-scaling. To increase utilization, PolyServe permits looser-SLO requests to share tighter-SLO instances when their own servers are saturated. PolyServe uses profiling data to guide scheduling decisions and manage tail latency through request-wait-time-aware scheduling, dynamic chunking, and continuous chunked prefill prediction. PolyServe achieves 1.23x goodput gain compared to existing policies, achieving up to 92.5% of optimal goodput.

Liangyu Zhao, Saeed Maleki, Yuanhong Wang et al.

As modern DNN models grow ever larger, collective communications between the accelerators (allreduce, etc.) emerge as a significant performance bottleneck. Designing efficient communication schedules is challenging, given today's heterogeneous and diverse network fabrics. We present ForestColl, a tool that generates throughput-optimal schedules for any network topology. ForestColl constructs broadcast/aggregation spanning trees as the communication schedule, achieving theoretical optimality. Its schedule generation runs in polynomial time and is highly scalable. ForestColl supports any network fabric, including both switching fabrics and direct accelerator connections. We evaluated ForestColl on AMD MI250 and NVIDIA DGX A100 & H100 clusters. ForestColl showed significant improvements over the vendors' own optimized communication libraries across various settings and in LLM training. ForestColl also outperformed other state-of-the-art schedule generation techniques with both more efficient generated schedules and substantially faster generation speed.

Liangyu Zhao, Arvind Krishnamurthy

We present a strongly polynomial-time algorithm to generate bandwidth optimal allgather/reduce-scatter on any network topology, with or without switches. Our algorithm constructs pipeline schedules achieving provably the best possible bandwidth performance on a given topology. To provide a universal solution, we model the network topology as a directed graph with heterogeneous link capacities and switches directly as vertices in the graph representation. The algorithm is strongly polynomial-time with respect to the topology size. This work heavily relies on previous graph theory work on edge-disjoint spanning trees and edge splitting. While we focus on allgather, the methods in this paper can be easily extended to generate schedules for reduce, broadcast, reduce-scatter, and allreduce.

Liangyu Zhao, Siddharth Pal, Tapan Chugh et al.

We consider the problem of distilling efficient network topologies for collective communications. We provide an algorithmic framework for constructing direct-connect topologies optimized for the latency vs. bandwidth trade-off associated with the workload. Our approach synthesizes many different topologies and schedules for a given cluster size and degree and then identifies the appropriate topology and schedule for a given workload. Our algorithms start from small, optimal base topologies and associated communication schedules and use techniques that can be iteratively applied to derive much larger topologies and schedules. Additionally, we incorporate well-studied large-scale graph topologies into our algorithmic framework by producing efficient collective schedules for them using a novel polynomial-time algorithm. Our evaluation uses multiple testbeds and large-scale simulations to demonstrate significant performance benefits from our derived topologies and schedules.

Liang Luo, Jacob Nelson, Arvind Krishnamurthy et al.

ML workloads are becoming increasingly popular in the cloud. Good cloud training performance is contingent on efficient parameter exchange among VMs. We find that Collectives, the widely used distributed communication algorithms, cannot perform optimally out of the box due to the hierarchical topology of datacenter networks and multi-tenancy nature of the cloudenvironment.In this paper, we present Cloud Collectives , a prototype that accelerates collectives by reordering theranks of participating VMs such that the communication pattern dictated by the selected collectives operation best exploits the locality in the network.Collectives is non-intrusive, requires no code changes nor rebuild of an existing application, and runs without support from cloud providers. Our preliminary application of Cloud Collectives on allreduce operations in public clouds results in a speedup of up to 3.7x in multiple microbenchmarks and 1.3x in real-world workloads of distributed training of deep neural networks and gradient boosted decision trees using state-of-the-art frameworks.

Yuchen Jin, Tianyi Zhou, Liangyu Zhao et al.

The learning rate (LR) schedule is one of the most important hyper-parameters needing careful tuning in training DNNs. However, it is also one of the least automated parts of machine learning systems and usually costs significant manual effort and computing. Though there are pre-defined LR schedules and optimizers with adaptive LR, they introduce new hyperparameters that need to be tuned separately for different tasks/datasets. In this paper, we consider the question: Can we automatically tune the LR over the course of training without human involvement? We propose an efficient method, AutoLRS, which automatically optimizes the LR for each training stage by modeling training dynamics. AutoLRS aims to find an LR applied to every $τ$ steps that minimizes the resulted validation loss. We solve this black-box optimization on the fly by Bayesian optimization (BO). However, collecting training instances for BO requires a system to evaluate each LR queried by BO's acquisition function for $τ$ steps, which is prohibitively expensive in practice. Instead, we apply each candidate LR for only $τ'\llτ$ steps and train an exponential model to predict the validation loss after $τ$ steps. This mutual-training process between BO and the loss-prediction model allows us to limit the training steps invested in the BO search. We demonstrate the advantages and the generality of AutoLRS through extensive experiments of training DNNs for tasks from diverse domains using different optimizers. The LR schedules auto-generated by AutoLRS lead to a speedup of $1.22\times$, $1.43\times$, and $1.5\times$ when training ResNet-50, Transformer, and BERT, respectively, compared to the LR schedules in their original papers, and an average speedup of $1.31\times$ over state-of-the-art heavily-tuned LR schedules.

Adriana Szekeres, Irene Zhang, Katelin Bailey et al.

Today's mobile devices sense, collect, and store huge amounts of personal information, which users share with family and friends through a wide range of applications. Once users give applications access to their data, they must implicitly trust that the apps correctly maintain data privacy. As we know from both experience and all-too-frequent press articles, that trust is often misplaced. While users do not trust applications, they do trust their mobile devices and operating systems. Unfortunately, sharing applications are not limited to mobile clients but must also run on cloud services to share data between users. In this paper, we leverage the trust that users have in their mobile OSes to vet cloud services. To do so, we define a new Secure Application Flow Enforcement (SAFE) framework, which requires cloud services to attest to a system stack that will enforce policies provided by the mobile OS for user data. We implement a mobile OS that enforces SAFE policies on unmodified mobile apps and two systems for enforcing policies on untrusted cloud services. Using these prototypes, we demonstrate that it is possible to enforce existing user privacy policies on unmodified applications.

Raymond Cheng, William Scott, Elisaweta Masserova et al.

Talek is a private group messaging system that sends messages through potentially untrustworthy servers, while hiding both data content and the communication patterns among its users. Talek explores a new point in the design space of private messaging; it guarantees access sequence indistinguishability, which is among the strongest guarantees in the space, while assuming an anytrust threat model, which is only slightly weaker than the strongest threat model currently found in related work. Our results suggest that this is a pragmatic point in the design space, since it supports strong privacy and good performance: we demonstrate a 3-server Talek cluster that achieves throughput of 9,433 messages/second for 32,000 active users with 1.7-second end-to-end latency. To achieve its security goals without coordination between clients, Talek relies on information-theoretic private information retrieval. To achieve good performance and minimize server-side storage, Talek introduces new techniques and optimizations that may be of independent interest, e.g., a novel use of blocked cuckoo hashing and support for private notifications. The latter provide a private, efficient mechanism for users to learn, without polling, which logs have new messages.

Amedeo Sapio, Marco Canini, Chen-Yu Ho et al.

Training machine learning models in parallel is an increasingly important workload. We accelerate distributed parallel training by designing a communication primitive that uses a programmable switch dataplane to execute a key step of the training process. Our approach, SwitchML, reduces the volume of exchanged data by aggregating the model updates from multiple workers in the network. We co-design the switch processing with the end-host protocols and ML frameworks to provide an efficient solution that speeds up training by up to 5.5$\times$ for a number of real-world benchmark models.

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.

Tianqi Chen, Lianmin Zheng, Eddie Yan et al.

We introduce a learning-based framework to optimize tensor programs for deep learning workloads. Efficient implementations of tensor operators, such as matrix multiplication and high dimensional convolution, are key enablers of effective deep learning systems. However, existing systems rely on manually optimized libraries such as cuDNN where only a narrow range of server class GPUs are well-supported. The reliance on hardware-specific operator libraries limits the applicability of high-level graph optimizations and incurs significant engineering costs when deploying to new hardware targets. We use learning to remove this engineering burden. We learn domain-specific statistical cost models to guide the search of tensor operator implementations over billions of possible program variants. We further accelerate the search by effective model transfer across workloads. Experimental results show that our framework delivers performance competitive with state-of-the-art hand-tuned libraries for low-power CPU, mobile GPU, and server-class GPU.

Liang Luo, Jacob Nelson, Luis Ceze et al.

Distributed deep neural network (DDNN) training constitutes an increasingly important workload that frequently runs in the cloud. Larger DNN models and faster compute engines are shifting DDNN training bottlenecks from computation to communication. This paper characterizes DDNN training to precisely pinpoint these bottlenecks. We found that timely training requires high performance parameter servers (PSs) with optimized network stacks and gradient processing pipelines, as well as server and network hardware with balanced computation and communication resources. We therefore propose PHub, a high performance multi-tenant, rack-scale PS design. PHub co-designs the PS software and hardware to accelerate rack-level and hierarchical cross-rack parameter exchange, with an API compatible with many DDNN training frameworks. PHub provides a performance improvement of up to 2.7x compared to state-of-the-art distributed training techniques for cloud-based ImageNet workloads, with 25% better throughput per dollar.

Haichen Shen, Seungyeop Han, Matthai Philipose et al.

Recent advances have enabled "oracle" classifiers that can classify across many classes and input distributions with high accuracy without retraining. However, these classifiers are relatively heavyweight, so that applying them to classify video is costly. We show that day-to-day video exhibits highly skewed class distributions over the short term, and that these distributions can be classified by much simpler models. We formulate the problem of detecting the short-term skews online and exploiting models based on it as a new sequential decision making problem dubbed the Online Bandit Problem, and present a new algorithm to solve it. When applied to recognizing faces in TV shows and movies, we realize end-to-end classification speedups of 2.4-7.8x/2.6-11.2x (on GPU/CPU) relative to a state-of-the-art convolutional neural network, at competitive accuracy.