Alexey Tumanov

Amey Agrawal, Anmol Agarwal, Nitin Kedia et al. · gatech

Serving large language models (LLMs) in production can incur substantial costs, which has prompted recent advances in inference system optimizations. Today, these systems are evaluated against conventional latency and throughput metrics (eg. TTFT, TBT, Normalised Latency and TPOT). However, these metrics fail to fully capture the nuances of LLM inference, leading to an incomplete assessment of user-facing performance crucial for real-time applications such as chat and translation. In this paper, we first identify the pitfalls of current performance metrics in evaluating LLM inference systems. We then propose Etalon, a comprehensive performance evaluation framework that includes fluidity-index -- a novel metric designed to reflect the intricacies of the LLM inference process and its impact on real-time user experience. Finally, we evaluate various existing open-source platforms and model-as-a-service offerings using Etalon, discussing their strengths and weaknesses. Etalon is available at https://github.com/project-etalon/etalon.

Debopam Sanyal, Jui-Tse Hung, Manav Agrawal et al.

Model-serving systems have become increasingly popular, especially in real-time web applications. In such systems, users send queries to the server and specify the desired performance metrics (e.g., desired accuracy, latency). The server maintains a set of models (model zoo) in the back-end and serves the queries based on the specified metrics. This paper examines the security, specifically robustness against model extraction attacks, of such systems. Existing black-box attacks assume a single model can be repeatedly selected for serving inference requests. Modern inference serving systems break this assumption. Thus, they cannot be directly applied to extract a victim model, as models are hidden behind a layer of abstraction exposed by the serving system. An attacker can no longer identify which model she is interacting with. To this end, we first propose a query-efficient fingerprinting algorithm to enable the attacker to trigger any desired model consistently. We show that by using our fingerprinting algorithm, model extraction can have fidelity and accuracy scores within $1\%$ of the scores obtained when attacking a single, explicitly specified model, as well as up to $14.6\%$ gain in accuracy and up to $7.7\%$ gain in fidelity compared to the naive attack. Second, we counter the proposed attack with a noise-based defense mechanism that thwarts fingerprinting by adding noise to the specified performance metrics. The proposed defense strategy reduces the attack's accuracy and fidelity by up to $9.8\%$ and $4.8\%$, respectively (on medium-sized model extraction). Third, we show that the proposed defense induces a fundamental trade-off between the level of protection and system goodput, achieving configurable and significant victim model extraction protection while maintaining acceptable goodput ($>80\%$). We implement the proposed defense in a real system with plans to open source.

Debopam Sanyal, Anantharaman Iyer, Alind Khare et al.

Given the wide range of deployment targets, flexible model selection is essential for optimizing performance within a given compute budget. Recent work demonstrates that stitching pretrained models within a model family enables cost-effective interpolation of the accuracy-efficiency tradeoff space. Stitching transforms intermediate activations from one pretrained model into another, producing a new interpolated stitched network. Such networks provide a pool of deployment options along the accuracy-efficiency spectrum. However, existing stitching approaches often yield suboptimal tradeoffs and lack generalizability, as they primarily rely on heuristics to select stitch configurations. We argue that constructing improved accuracy-efficiency tradeoffs requires explicitly capturing and leveraging the similarity between pretrained models being stitched. To this end, we introduce KLAS, a novel stitch selection framework that automates and generalizes stitch selection across model families by leveraging KL divergence between intermediate representations. KLAS identifies the most promising binary stitches from the $O(k^2n^2)$ possibilities for $k$ pretrained models of depth $n$. Through comprehensive experiments, we demonstrate that KLAS improves the accuracy-efficiency curve of stitched models at the same finetuning cost as baselines. KLAS achieves up to $1.21\%$ higher ImageNet-1K top-1 accuracy at the same computational cost, or maintains accuracy with a $1.33\times$ reduction in FLOPs.

Amey Agrawal, Haoran Qiu, Junda Chen et al. · gatech

Deploying million-token Large Language Models (LLMs) is challenging because production workloads are highly heterogeneous, mixing short queries and long documents. This heterogeneity, combined with the quadratic complexity of attention, creates severe convoy effects where long-running requests stall short, interactive ones, degrading system responsiveness. We present Medha, a serving system that eliminates these convoys by introducing fine-grained, preemptive scheduling to LLM inference. Medha makes preemption practical with a co-designed set of mechanisms -- including Adaptive Chunking and Stream Pipeline Parallel that overcome the perceived inefficiencies and scaling challenges of chunking. Additionally, we present a new parallelism strategy KV-Cache Parallelism to reduce the decode latency and afford interactivity despite very long context. These mechanisms are orchestrated by a Length-Aware Relative Slack (LARS) scheduler, a deadline and heterogeneity-aware scheduling policy that prevents both the convoy effect and the starvation that plagues simpler policies. Under a heterogeneous workload, Medha improves throughput by 5.7x while reducing median and 99th percentile latency by 30x and 174x, respectively, compared to state-of-the-art non-preemptive systems.

Amey Agrawal, Sameer Reddy, Satwik Bhattamishra et al. · gatech

With the increase in the scale of Deep Learning (DL) training workloads in terms of compute resources and time consumption, the likelihood of encountering in-training failures rises substantially, leading to lost work and resource wastage. Such failures are typically offset by a checkpointing mechanism, which comes at the cost of storage and network bandwidth overhead. State-of-the-art approaches involve lossy model compression mechanisms, which induce a tradeoff between the resulting model quality (accuracy) and compression ratio. Delta compression is then used to further reduce the overhead by only storing the difference between consecutive checkpoints. We make a key enabling observation that the sensitivity of model weights to compression varies during training, and different weights benefit from different quantization levels (ranging from retaining full precision to pruning). We propose (1) a non-uniform quantization scheme that leverages this variation, (2) an efficient search mechanism that dynamically finds the best quantization configurations, and (3) a quantization-aware delta compression mechanism that rearranges weights to minimize checkpoint differences, thereby maximizing compression. We instantiate these contributions in Inshrinkerator - a framework for DL workload checkpoint compression. Our experiments show that Inshrinkerator consistently achieves a better tradeoff between accuracy and compression ratios compared to prior works, enabling a compression ratio up to 39x and withstanding up to 10 restores with negligible accuracy impact for fault-tolerant training. Inshrinkerator achieves at least an order of magnitude reduction in checkpoint storage overhead for training failure recovery as well as transfer learning use cases without any loss of accuracy.

Yanbo Xu, Alind Khare, Glenn Matlin et al. · gatech

Machine Learning (ML) research has focused on maximizing the accuracy of predictive tasks. ML models, however, are increasingly more complex, resource intensive, and costlier to deploy in resource-constrained environments. These issues are exacerbated for prediction tasks with sequential classification on progressively transitioned stages with ''happens-before'' relation between them.We argue that it is possible to ''unfold'' a monolithic single multi-class classifier, typically trained for all stages using all data, into a series of single-stage classifiers. Each single-stage classifier can be cascaded gradually from cheaper to more expensive binary classifiers that are trained using only the necessary data modalities or features required for that stage. UnfoldML is a cost-aware and uncertainty-based dynamic 2D prediction pipeline for multi-stage classification that enables (1) navigation of the accuracy/cost tradeoff space, (2) reducing the spatio-temporal cost of inference by orders of magnitude, and (3) early prediction on proceeding stages. UnfoldML achieves orders of magnitude better cost in clinical settings, while detecting multi-stage disease development in real time. It achieves within 0.1% accuracy from the highest-performing multi-class baseline, while saving close to 20X on spatio-temporal cost of inference and earlier (3.5hrs) disease onset prediction. We also show that UnfoldML generalizes to image classification, where it can predict different level of labels (from coarse to fine) given different level of abstractions of a image, saving close to 5X cost with as little as 0.4% accuracy reduction.

Payman Behnam, Jianming Tong, Alind Khare et al.

A growing number of applications depend on Machine Learning (ML) functionality and benefits from both higher quality ML predictions and better timeliness (latency) at the same time. A growing body of research in computer architecture, ML, and systems software literature focuses on reaching better latency-accuracy tradeoffs for ML models. Efforts include compression, quantization, pruning, early-exit models, mixed DNN precision, as well as ML inference accelerator designs that minimize latency and energy, while preserving delivered accuracy. All of them, however, yield improvements for a single static point in the latency-accuracy tradeoff space. We make a case for applications that operate in dynamically changing deployment scenarios, where no single static point is optimal. We draw on a recently proposed weight-shared SuperNet mechanism to enable serving a stream of queries that uses (activates) different SubNets within this weight-shared construct. This creates an opportunity to exploit the inherent temporal locality with our proposed SubGraph Stationary (SGS) optimization. We take a hardware-software co-design approach with a real implementation of SGS in SushiAccel and the implementation of a software scheduler SushiSched controlling which SubNets to serve and what to cache in real-time. Combined, they are vertically integrated into SUSHI-an inference serving stack. For the stream of queries, SUSHI yields up to 25% improvement in latency, 0.98% increase in served accuracy. SUSHI can achieve up to 78.7% off-chip energy savings.

Sachit Kuhar, Alexey Tumanov, Judy Hoffman

Efficient inference of Deep Neural Networks (DNNs) is essential to making AI ubiquitous. Two important algorithmic techniques have shown promise for enabling efficient inference - sparsity and binarization. These techniques translate into weight sparsity and weight repetition at the hardware-software level enabling the deployment of DNNs with critically low power and latency requirements. We propose a new method called signed-binary networks to improve efficiency further (by exploiting both weight sparsity and weight repetition together) while maintaining similar accuracy. Our method achieves comparable accuracy on ImageNet and CIFAR10 datasets with binary and can lead to 69% sparsity. We observe real speedup when deploying these models on general-purpose devices and show that this high percentage of unstructured sparsity can lead to a further reduction in energy consumption on ASICs.

Amey Agrawal, Nitin Kedia, Ashish Panwar et al. · gatech

Each LLM serving request goes through two phases. The first is prefill which processes the entire input prompt and produces the first output token and the second is decode which generates the rest of output tokens, one-at-a-time. Prefill iterations have high latency but saturate GPU compute due to parallel processing of the input prompt. In contrast, decode iterations have low latency but also low compute utilization because a decode iteration processes only a single token per request. This makes batching highly effective for decodes and consequently for overall throughput. However, batching multiple requests leads to an interleaving of prefill and decode iterations which makes it challenging to achieve both high throughput and low latency. We introduce an efficient LLM inference scheduler, Sarathi-Serve, to address this throughput-latency tradeoff. Sarathi-Serve introduces chunked-prefills which splits a prefill request into near equal sized chunks and creates stall-free schedules that adds new requests in a batch without pausing ongoing decodes. Stall-free scheduling unlocks the opportunity to improve throughput with large batch sizes while minimizing the effect of batching on latency. Furthermore, uniform batches in Sarathi-Serve ameliorate the imbalance between iterations resulting in minimal pipeline bubbles. Our techniques yield significant improvements in inference performance across models and hardware under tail latency constraints. For Mistral-7B on single A100 GPUs, we achieve 2.6x higher serving capacity and up to 3.7x higher serving capacity for the Yi-34B model on two A100 GPUs as compared to vLLM. When used with pipeline parallelism on Falcon-180B, Sarathi-Serve provides up to 5.6x gain in the end-to-end serving capacity. The source code for Sarathi-Serve is available at https://github.com/microsoft/sarathi-serve.

Alind Khare, Animesh Agrawal, Aditya Annavajjala et al.

Neural Architecture Search (NAS) for Federated Learning (FL) is an emerging field. It automates the design and training of Deep Neural Networks (DNNs) when data cannot be centralized due to privacy, communication costs, or regulatory restrictions. Recent federated NAS methods not only reduce manual effort but also help achieve higher accuracy than traditional FL methods like FedAvg. Despite the success, existing federated NAS methods still fall short in satisfying diverse deployment targets common in on-device inference like hardware, latency budgets, or variable battery levels. Most federated NAS methods search for only a limited range of neuro-architectural patterns, repeat them in a DNN, thereby restricting achievable performance. Moreover, these methods incur prohibitive training costs to satisfy deployment targets. They perform the training and search of DNN architectures repeatedly for each case. SuperFedNAS addresses these challenges by decoupling the training and search in federated NAS. SuperFedNAS co-trains a large number of diverse DNN architectures contained inside one supernet in the FL setting. Post-training, clients perform NAS locally to find specialized DNNs by extracting different parts of the trained supernet with no additional training. SuperFedNAS takes O(1) (instead of O(N)) cost to find specialized DNN architectures in FL for any N deployment targets. As part of SuperFedNAS, we introduce MaxNet - a novel FL training algorithm that performs multi-objective federated optimization of a large number of DNN architectures ($\approx 5*10^8$) under different client data distributions. Overall, SuperFedNAS achieves upto 37.7% higher accuracy for the same MACs or upto 8.13x reduction in MACs for the same accuracy than existing federated NAS methods.

Hugo Latapie, Shan Yu, Patrick Hammer et al.

Traditional computer vision models often necessitate extensive data acquisition, annotation, and validation. These models frequently struggle in real-world applications, resulting in high false positive and negative rates, and exhibit poor adaptability to new scenarios, often requiring costly retraining. To address these issues, we present Ethosight, a flexible and adaptable zero-shot video analytics system. Ethosight begins from a clean slate based on user-defined video analytics, specified through natural language or keywords, and leverages joint embedding models and reasoning mechanisms informed by ontologies such as WordNet and ConceptNet. Ethosight operates effectively on low-cost edge devices and supports enhanced runtime adaptation, thereby offering a new approach to continuous learning without catastrophic forgetting. We provide empirical validation of Ethosight's promising effectiveness across diverse and complex use cases, while highlighting areas for further improvement. A significant contribution of this work is the release of all source code and datasets to enable full reproducibility and to foster further innovation in both the research and commercial domains.

Amey Agrawal, Nitin Kedia, Jayashree Mohan et al. · gatech

Optimizing the deployment of Large language models (LLMs) is expensive today since it requires experimentally running an application workload against an LLM implementation while exploring large configuration space formed by system knobs such as parallelization strategies, batching techniques, and scheduling policies. To address this challenge, we present Vidur - a large-scale, high-fidelity, easily-extensible simulation framework for LLM inference performance. Vidur models the performance of LLM operators using a combination of experimental profiling and predictive modeling, and evaluates the end-to-end inference performance for different workloads by estimating several metrics of interest such as latency and throughput. We validate the fidelity of Vidur on several LLMs and show that it estimates inference latency with less than 9% error across the range. Further, we present Vidur-Search, a configuration search tool that helps optimize LLM deployment. Vidur-Search uses Vidur to automatically identify the most cost-effective deployment configuration that meets application performance constraints. For example, Vidur-Search finds the best deployment configuration for LLaMA2-70B in one hour on a CPU machine, in contrast to a deployment-based exploration which would require 42K GPU hours - costing ~218K dollars. Source code for Vidur is available at https://github.com/microsoft/vidur.

Amey Agrawal, Mayank Yadav, Sukrit Kumar et al. · gatech

Deploying LLMs efficiently requires testing hundreds of serving configurations, but evaluating each one on a GPU cluster takes hours and costs thousands of dollars. Discrete-event simulators are faster and cheaper, but they require re-implementing the serving system's control logic -- a burden that compounds as frameworks evolve. We present Revati, a time-warp emulator that enables performance modeling by directly executing real serving system code at simulation-like speed. The system intercepts CUDA API calls to virtualize device management, allowing serving frameworks to run without physical GPUs. Instead of executing GPU kernels, it performs time jumps -- fast-forwarding virtual time by predicted kernel durations. We propose a coordination protocol that synchronizes these jumps across distributed processes while preserving causality. On vLLM and SGLang, Revati achieves less than 5% prediction error across multiple models and parallelism configurations, while running 5-17x faster than real GPU execution.

Kaustubh Ponkshe, Raghav Singhal, Eduard Gorbunov et al.

Low-rank adapters have become standard for efficiently fine-tuning large language models, but they often fall short of achieving the performance of full fine-tuning. We propose a method, LoRA Silver Bullet or LoRA-SB, that approximates full fine-tuning within low-rank subspaces using a carefully designed initialization strategy. We theoretically demonstrate that the architecture of LoRA-XS, which inserts a learnable r x r matrix between B and A while keeping other matrices fixed, provides the precise conditions needed for this approximation. We leverage its constrained update space to achieve optimal scaling for high-rank gradient updates while removing the need for scaling factor tuning. We prove that our initialization offers an optimal low-rank approximation of the initial gradient and preserves update directions throughout training. Extensive experiments across mathematical reasoning, commonsense reasoning, and language understanding tasks demonstrate that our approach exceeds the performance of LoRA (and baselines) while using 27-90 times fewer learnable parameters, and comprehensively outperforms LoRA-XS. Our findings establish that it is possible to simulate full fine-tuning in low-rank subspaces, and achieve significant parameter efficiency gains without sacrificing performance. Our code is publicly available at: https://github.com/CERT-Lab/lora-sb.

Anshul Ahluwalia, Payman Behnam, Rohit Das et al.

Recent advancements in Graph Neural Networks (GNNs) have led to increased model sizes to enhance their capacity and accuracy. Such large models incur high memory usage, latency, and computational costs, thereby restricting their inference deployment. GNN compression techniques compress large GNNs into smaller ones with negligible accuracy loss. One of the most promising compression techniques is knowledge distillation (KD). However, most KD approaches for GNNs only consider the outputs of the last layers and do not consider the outputs of the intermediate layers of the GNNs. The intermediate layers may contain important inductive biases indicated by the graph structure and embeddings. Ignoring these layers may lead to a high accuracy drop, especially when the compression ratio is high. To address these shortcomings, we propose a novel KD approach for GNN compression that we call Structure and Embedding Distillation with Attention (STRIDE). STRIDE utilizes attention to identify important intermediate teacher-student layer pairs and focuses on using those pairs to align graph structure and node embeddings. We evaluate STRIDE on several datasets, such as OGBN-Mag and OGBN-Arxiv, using different model architectures, including GCNIIs, RGCNs, and GraphSAGE. On average, STRIDE achieves a 2.13% increase in accuracy with a 32.3X compression ratio on OGBN-Mag, a large graph dataset, compared to state-of-the-art approaches. On smaller datasets (e.g., Pubmed), STRIDE achieves up to a 141X compression ratio with the same accuracy as state-of-the-art approaches. These results highlight the effectiveness of focusing on intermediate-layer knowledge to obtain compact, accurate, and practical GNN models.

Payman Behnam, Yaosheng Fu, Ritchie Zhao et al.

Transformer-based Large Language Models rely critically on the KV cache to efficiently handle extended contexts during the decode phase. Yet, the size of the KV cache grows proportionally with the input length, burdening both memory bandwidth and capacity as decoding progresses. To address this challenge, we present RocketKV, a training-free KV cache compression strategy containing two consecutive stages. In the first stage, it performs coarse-grain permanent KV cache eviction on the input sequence tokens. In the second stage, it adopts a hybrid sparse attention method to conduct fine-grain top-k sparse attention, approximating the attention scores by leveraging both head and sequence dimensionality reductions. We show that RocketKV provides a compression ratio of up to 400$\times$, end-to-end speedup of up to 3.7$\times$ as well as peak memory reduction of up to 32.6% in the decode phase on an NVIDIA A100 GPU compared to the full KV cache baseline, while achieving negligible accuracy loss on a variety of long-context tasks. We also propose a variant of RocketKV for multi-turn scenarios, which consistently outperforms other existing methods and achieves accuracy nearly on par with an oracle top-k attention scheme. The source code is available here: https://github.com/NVlabs/RocketKV.

Aditya Annavajjala, Alind Khare, Animesh Agrawal et al.

CNNs are increasingly deployed across different hardware, dynamic environments, and low-power embedded devices. This has led to the design and training of CNN architectures with the goal of maximizing accuracy subject to such variable deployment constraints. As the number of deployment scenarios grows, there is a need to find scalable solutions to design and train specialized CNNs. Once-for-all training has emerged as a scalable approach that jointly co-trains many models (subnets) at once with a constant training cost and finds specialized CNNs later. The scalability is achieved by training the full model and simultaneously reducing it to smaller subnets that share model weights (weight-shared shrinking). However, existing once-for-all training approaches incur huge training costs reaching 1200 GPU hours. We argue this is because they either start the process of shrinking the full model too early or too late. Hence, we propose Delayed $ε$-Shrinking (D$ε$pS) that starts the process of shrinking the full model when it is partially trained (~50%) which leads to training cost improvement and better in-place knowledge distillation to smaller models. The proposed approach also consists of novel heuristics that dynamically adjust subnet learning rates incrementally (E), leading to improved weight-shared knowledge distillation from larger to smaller subnets as well. As a result, DEpS outperforms state-of-the-art once-for-all training techniques across different datasets including CIFAR10/100, ImageNet-100, and ImageNet-1k on accuracy and cost. It achieves 1.83% higher ImageNet-1k top1 accuracy or the same accuracy with 1.3x reduction in FLOPs and 2.5x drop in training cost (GPU*hrs)

Alind Khare, Dhruv Garg, Sukrit Kalra et al.

The increasing deployment of ML models on the critical path of production applications in both datacenter and the edge requires ML inference serving systems to serve these models under unpredictable and bursty request arrival rates. Serving models under such conditions requires these systems to strike a careful balance between the latency and accuracy requirements of the application and the overall efficiency of utilization of scarce resources. State-of-the-art systems resolve this tension by either choosing a static point in the latency-accuracy tradeoff space to serve all requests or load specific models on the critical path of request serving. In this work, we instead resolve this tension by simultaneously serving the entire-range of models spanning the latency-accuracy tradeoff space. Our novel mechanism, SubNetAct, achieves this by carefully inserting specialized operators in weight-shared SuperNetworks. These operators enable SubNetAct to dynamically route requests through the network to meet a latency and accuracy target. SubNetAct requires upto 2.6x lower memory to serve a vastly-higher number of models than prior state-of-the-art. In addition, SubNetAct's near-instantaneous actuation of models unlocks the design space of fine-grained, reactive scheduling policies. We explore the design of one such extremely effective policy, SlackFit and instantiate both SubNetAct and SlackFit in a real system, SuperServe. SuperServe achieves 4.67% higher accuracy for the same SLO attainment and 2.85x higher SLO attainment for the same accuracy on a trace derived from the real-world Microsoft Azure Functions workload and yields the best trade-offs on a wide range of extremely-bursty synthetic traces automatically.

Amey Agrawal, Nitin Kedia, Anmol Agarwal et al. · gatech

The rapid evolution of Large Language Model (LLM) inference systems has yielded significant efficiency improvements. However, our systematic analysis reveals that current evaluation methodologies frequently exhibit fundamental flaws, often manifesting as common evaluation anti-patterns that obscure true performance characteristics and impede scientific progress. Through a comprehensive examination of recent systems, we identify recurring anti-patterns across three key dimensions: Baseline Fairness, Evaluation Setup, and Metric Design. These anti-patterns are uniquely problematic for LLM inference due to its dual-phase nature combining distinct prefill and decode operations, its handling of highly heterogeneous workloads, and its strict temporal requirements for interactive use. We demonstrate how common anti-patterns -- such as inadequate baseline comparisons that conflate engineering effort with algorithmic novelty, workload selections that fail to represent production scenarios, and metric normalizations that hide substantial performance variability like generation stalls-lead to misleading conclusions. To address these challenges, we provide a comprehensive checklist derived from our analysis, establishing a framework for recognizing and avoiding these anti-patterns in favor of robust LLM inference evaluation. To demonstrate the practical application of our framework, we present a case study analyzing speculative decoding, a technique whose bursty, non-uniform token generation is easily misinterpreted when evaluated using approaches characteristic of these anti-patterns. Our work establishes a rigorous foundation for evaluation methodology, enabling meaningful comparisons, ensuring reproducible results, and ultimately accelerating genuine progress in LLM inference systems by moving beyond common anti-patterns to align evaluation with real-world requirements.

Srihas Yarlagadda, Amey Agrawal, Elton Pinto et al. · gatech

Training large foundation models costs hundreds of millions of dollars, making deployment optimization critical. Current approaches require machine learning engineers to manually craft training recipes through error-prone trial-and-error on expensive compute clusters. To enable efficient exploration of training configurations, researchers have developed performance modeling systems. However, these systems force users to translate their workloads into custom specification languages, introducing a fundamental semantic gap between the actual workload and its representation. This gap creates an inherent tradeoff: systems must either support a narrow set of workloads to maintain usability, require complex specifications that limit practical adoption, or compromise prediction accuracy with simplified performance models. We present Maya, a performance modeling system that eliminates these tradeoffs through transparent device emulation. By operating at the narrow interface between training frameworks and accelerator devices, Maya can capture complete workload behavior without requiring code modifications or translations. Maya intercepts device API calls from unmodified training code to directly observe low-level operations, enabling accurate performance prediction while maintaining both ease of use and generality. Our evaluation shows Maya achieves less than 5% prediction error across diverse models and optimization strategies, identifying configurations that reduce training costs by up to 56% compared to existing approaches.

Sachit Kuhar, Yash Jain, Alexey Tumanov · gatech, microsoft-research

Efficient inference of Deep Neural Networks (DNNs) on resource-constrained edge devices is essential. Quantization and sparsity are key techniques that translate to repetition and sparsity within tensors at the hardware-software interface. This paper introduces the concept of repetition-sparsity trade-off that helps explain computational efficiency during inference. We propose PLUM, a unified co-design framework that integrates DNN inference systems and quantization (forward and backward pass) to leverage the repetition-sparsity trade-off to improve inference efficiency. Our results demonstrate that PLUM's quantization method is more accurate than binary quantization with the same number of non-zero weights. Detailed analysis indicates that signed binarization generates a smaller distribution of effectual (non-zero) parameters nested within a larger distribution of total parameters of latent full-precision weights for a DNN block. Finally, the proposed PLUM framework achieves a 26% speedup on real hardware, doubles energy efficiency, and reduces density by 2.8x compared to binary methods while retaining top-1 accuracy when compared to prior-art methods for ResNets on ImageNet (by achieving 66.2% top-1 accuracy), presenting an alternative solution for deploying efficient models in resource-limited environments.

Manas Sahni, Shreya Varshini, Alind Khare et al.

The emergence of CNNs in mainstream deployment has necessitated methods to design and train efficient architectures tailored to maximize the accuracy under diverse hardware & latency constraints. To scale these resource-intensive tasks with an increasing number of deployment targets, Once-For-All (OFA) proposed an approach to jointly train several models at once with a constant training cost. However, this cost remains as high as 40-50 GPU days and also suffers from a combinatorial explosion of sub-optimal model configurations. We seek to reduce this search space -- and hence the training budget -- by constraining search to models close to the accuracy-latency Pareto frontier. We incorporate insights of compound relationships between model dimensions to build CompOFA, a design space smaller by several orders of magnitude. Through experiments on ImageNet, we demonstrate that even with simple heuristics we can achieve a 2x reduction in training time and 216x speedup in model search/extraction time compared to the state of the art, without loss of Pareto optimality! We also show that this smaller design space is dense enough to support equally accurate models for a similar diversity of hardware and latency targets, while also reducing the complexity of the training and subsequent extraction algorithms.

Shenda Hong, Yanbo Xu, Alind Khare et al.

Deep learning models have achieved expert-level performance in healthcare with an exclusive focus on training accurate models. However, in many clinical environments such as intensive care unit (ICU), real-time model serving is equally if not more important than accuracy, because in ICU patient care is simultaneously more urgent and more expensive. Clinical decisions and their timeliness, therefore, directly affect both the patient outcome and the cost of care. To make timely decisions, we argue the underlying serving system must be latency-aware. To compound the challenge, health analytic applications often require a combination of models instead of a single model, to better specialize individual models for different targets, multi-modal data, different prediction windows, and potentially personalized predictions. To address these challenges, we propose HOLMES-an online model ensemble serving framework for healthcare applications. HOLMES dynamically identifies the best performing set of models to ensemble for highest accuracy, while also satisfying sub-second latency constraints on end-to-end prediction. We demonstrate that HOLMES is able to navigate the accuracy/latency tradeoff efficiently, compose the ensemble, and serve the model ensemble pipeline, scaling to simultaneously streaming data from 100 patients, each producing waveform data at 250~Hz. HOLMES outperforms the conventional offline batch-processed inference for the same clinical task in terms of accuracy and latency (by order of magnitude). HOLMES is tested on risk prediction task on pediatric cardio ICU data with above 95% prediction accuracy and sub-second latency on 64-bed simulation.

Richard Liaw, Romil Bhardwaj, Lisa Dunlap et al.

Prior research in resource scheduling for machine learning training workloads has largely focused on minimizing job completion times. Commonly, these model training workloads collectively search over a large number of parameter values that control the learning process in a hyperparameter search. It is preferable to identify and maximally provision the best-performing hyperparameter configuration (trial) to achieve the highest accuracy result as soon as possible. To optimally trade-off evaluating multiple configurations and training the most promising ones by a fixed deadline, we design and build HyperSched -- a dynamic application-level resource scheduler to track, identify, and preferentially allocate resources to the best performing trials to maximize accuracy by the deadline. HyperSched leverages three properties of a hyperparameter search workload over-looked in prior work - trial disposability, progressively identifiable rankings among different configurations, and space-time constraints - to outperform standard hyperparameter search algorithms across a variety of benchmarks.

Paras Jain, Xiangxi Mo, Ajay Jain et al.

Current trends in Machine Learning~(ML) inference on hardware accelerated devices (e.g., GPUs, TPUs) point to alarmingly low utilization. As ML inference is increasingly time-bounded by tight latency SLOs, increasing data parallelism is not an option. The need for better efficiency motivates GPU multiplexing. Furthermore, existing GPU programming abstractions force programmers to micro-manage GPU resources in an early-binding, context-free fashion. We propose a VLIW-inspired Out-of-Order (OoO) Just-in-Time (JIT) compiler that coalesces and reorders execution kernels at runtime for throughput-optimal device utilization while satisfying latency SLOs. We quantify the inefficiencies of space-only and time-only multiplexing alternatives and demonstrate an achievable 7.7x opportunity gap through spatial coalescing.

Philipp Moritz, Robert Nishihara, Stephanie Wang et al.

The next generation of AI applications will continuously interact with the environment and learn from these interactions. These applications impose new and demanding systems requirements, both in terms of performance and flexibility. In this paper, we consider these requirements and present Ray---a distributed system to address them. Ray implements a unified interface that can express both task-parallel and actor-based computations, supported by a single dynamic execution engine. To meet the performance requirements, Ray employs a distributed scheduler and a distributed and fault-tolerant store to manage the system's control state. In our experiments, we demonstrate scaling beyond 1.8 million tasks per second and better performance than existing specialized systems for several challenging reinforcement learning applications.

Xin Wang, Yujia Luo, Daniel Crankshaw et al.

Advances in deep learning have led to substantial increases in prediction accuracy but have been accompanied by increases in the cost of rendering predictions. We conjecture that fora majority of real-world inputs, the recent advances in deep learning have created models that effectively "overthink" on simple inputs. In this paper, we revisit the classic question of building model cascades that primarily leverage class asymmetry to reduce cost. We introduce the "I Don't Know"(IDK) prediction cascades framework, a general framework to systematically compose a set of pre-trained models to accelerate inference without a loss in prediction accuracy. We propose two search based methods for constructing cascades as well as a new cost-aware objective within this framework. The proposed IDK cascade framework can be easily adopted in the existing model serving systems without additional model re-training. We evaluate the proposed techniques on a range of benchmarks to demonstrate the effectiveness of the proposed framework.

Robert Nishihara, Philipp Moritz, Stephanie Wang et al.

Machine learning applications are increasingly deployed not only to serve predictions using static models, but also as tightly-integrated components of feedback loops involving dynamic, real-time decision making. These applications pose a new set of requirements, none of which are difficult to achieve in isolation, but the combination of which creates a challenge for existing distributed execution frameworks: computation with millisecond latency at high throughput, adaptive construction of arbitrary task graphs, and execution of heterogeneous kernels over diverse sets of resources. We assert that a new distributed execution framework is needed for such ML applications and propose a candidate approach with a proof-of-concept architecture that achieves a 63x performance improvement over a state-of-the-art execution framework for a representative application.