arXiv subjects

Lianmin Zheng

Publications and source records attributed to Lianmin Zheng.

At least 37 records · Page 2Linked to original sources

arXiv2023

On Optimal Caching and Model Multiplexing for Large Model Inference

Large Language Models (LLMs) and other large foundation models have achieved noteworthy success, but their size exacerbates existing resource consumption and latency challenges. In particular, the large-scale deployment of these models is hindered by the significant resource requirements during inference. In this paper, we study two approaches for mitigating these challenges: employing a cache to store previous queries and learning a model multiplexer to choose from an ensemble of models for query processing. Theoretically, we provide an optimal algorithm for jointly optimizing both approaches to reduce the inference cost in both offline and online tabular settings. By combining a caching algorithm, namely Greedy Dual Size with Frequency (GDSF) or Least Expected Cost (LEC), with a model multiplexer, we achieve optimal rates in both offline and online settings. Empirically, simulations show that the combination of our caching and model multiplexing algorithms greatly improves over the baselines, with up to $50\times$ improvement over the baseline when the ratio between the maximum cost and minimum cost is $100$. Experiments on real datasets show a $4.3\times$ improvement in FLOPs over the baseline when the ratio for FLOPs is $10$, and a $1.8\times$ improvement in latency when the ratio for average latency is $1.85$.

cs.LG↗

arXiv2023

AlpaServe: Statistical Multiplexing with Model Parallelism for Deep Learning Serving

Model parallelism is conventionally viewed as a method to scale a single large deep learning model beyond the memory limits of a single device. In this paper, we demonstrate that model parallelism can be additionally used for the statistical multiplexing of multiple devices when serving multiple models, even when a single model can fit into a single device. Our work reveals a fundamental trade-off between the overhead introduced by model parallelism and the opportunity to exploit statistical multiplexing to reduce serving latency in the presence of bursty workloads. We explore the new trade-off space and present a novel serving system, AlpaServe, that determines an efficient strategy for placing and parallelizing collections of large deep learning models across a distributed cluster. Evaluation results on production workloads show that AlpaServe can process requests at up to 10x higher rates or 6x more burstiness while staying within latency constraints for more than 99% of requests.

cs.LG↗

arXiv2023

FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU

The high computational and memory requirements of large language model (LLM) inference make it feasible only with multiple high-end accelerators. Motivated by the emerging demand for latency-insensitive tasks with batched processing, this paper initiates the study of high-throughput LLM inference using limited resources, such as a single commodity GPU. We present FlexGen, a high-throughput generation engine for running LLMs with limited GPU memory. FlexGen can be flexibly configured under various hardware resource constraints by aggregating memory and computation from the GPU, CPU, and disk. By solving a linear programming problem, it searches for efficient patterns to store and access tensors. FlexGen further compresses the weights and the attention cache to 4 bits with negligible accuracy loss. These techniques enable FlexGen to have a larger space of batch size choices and thus significantly increase maximum throughput. As a result, when running OPT-175B on a single 16GB GPU, FlexGen achieves significantly higher throughput compared to state-of-the-art offloading systems, reaching a generation throughput of 1 token/s for the first time with an effective batch size of 144. On the HELM benchmark, FlexGen can benchmark a 30B model with a 16GB GPU on 7 representative sub-scenarios in 21 hours. The code is available at https://github.com/FMInference/FlexGen

cs.LG↗

arXiv2022

TensorIR: An Abstraction for Automatic Tensorized Program Optimization

Deploying deep learning models on various devices has become an important topic. The wave of hardware specialization brings a diverse set of acceleration primitives for multi-dimensional tensor computations. These new acceleration primitives, along with the emerging machine learning models, bring tremendous engineering challenges. In this paper, we present TensorIR, a compiler abstraction for optimizing programs with these tensor computation primitives. TensorIR generalizes the loop nest representation used in existing machine learning compilers to bring tensor computation as the first-class citizen. Finally, we build an end-to-end framework on top of our abstraction to automatically optimize deep learning models for given tensor computation primitives. Experimental results show that TensorIR compilation automatically uses the tensor computation primitives for given hardware backends and delivers performance that is competitive to state-of-art hand-optimized systems across platforms.

cs.LG↗

arXiv2022

GACT: Activation Compressed Training for Generic Network Architectures

Training large neural network (NN) models requires extensive memory resources, and Activation Compressed Training (ACT) is a promising approach to reduce training memory footprint. This paper presents GACT, an ACT framework to support a broad range of machine learning tasks for generic NN architectures with limited domain knowledge. By analyzing a linearized version of ACT's approximate gradient, we prove the convergence of GACT without prior knowledge on operator type or model architecture. To make training stable, we propose an algorithm that decides the compression ratio for each tensor by estimating its impact on the gradient at run time. We implement GACT as a PyTorch library that readily applies to any NN architecture. GACT reduces the activation memory for convolutional NNs, transformers, and graph NNs by up to 8.1x, enabling training with a 4.2x to 24.7x larger batch size, with negligible accuracy loss. We implement GACT as a PyTorch library at https://github.com/LiuXiaoxuanPKU/GACT-ICML.

cs.LG↗

arXiv2022

NumS: Scalable Array Programming for the Cloud

Scientists increasingly rely on Python tools to perform scalable distributed memory array operations using rich, NumPy-like expressions. However, many of these tools rely on dynamic schedulers optimized for abstract task graphs, which often encounter memory and network bandwidth-related bottlenecks due to sub-optimal data and operator placement decisions. Tools built on the message passing interface (MPI), such as ScaLAPACK and SLATE, have better scaling properties, but these solutions require specialized knowledge to use. In this work, we present NumS, an array programming library which optimizes NumPy-like expressions on task-based distributed systems. This is achieved through a novel scheduler called Load Simulated Hierarchical Scheduling (LSHS). LSHS is a local search method which optimizes operator placement by minimizing maximum memory and network load on any given node within a distributed system. Coupled with a heuristic for load balanced data layouts, our approach is capable of attaining communication lower bounds on some common numerical operations, and our empirical study shows that LSHS enhances performance on Ray by decreasing network load by a factor of 2x, requiring 4x less memory, and reducing execution time by 10x on the logistic regression problem. On terabyte-scale data, NumS achieves competitive performance to SLATE on DGEMM, up to 20x speedup over Dask on a key operation for tensor factorization, and a 2x speedup on logistic regression compared to Dask ML and Spark's MLlib.

cs.DC↗

arXiv2022

Alpa: Automating Inter- and Intra-Operator Parallelism for Distributed Deep Learning

Alpa automates model-parallel training of large deep learning (DL) models by generating execution plans that unify data, operator, and pipeline parallelism. Existing model-parallel training systems either require users to manually create a parallelization plan or automatically generate one from a limited space of model parallelism configurations. They do not suffice to scale out complex DL models on distributed compute devices. Alpa distributes the training of large DL models by viewing parallelisms as two hierarchical levels: inter-operator and intra-operator parallelisms. Based on it, Alpa constructs a new hierarchical space for massive model-parallel execution plans. Alpa designs a number of compilation passes to automatically derive efficient parallel execution plans at each parallelism level. Alpa implements an efficient runtime to orchestrate the two-level parallel execution on distributed compute devices. Our evaluation shows Alpa generates parallelization plans that match or outperform hand-tuned model-parallel training systems even on models they are designed for. Unlike specialized systems, Alpa also generalizes to models with heterogeneous architectures and models without manually-designed plans. Alpa's source code is publicly available at https://github.com/alpa-projects/alpa

cs.LG↗

arXiv2021

ActNN: Reducing Training Memory Footprint via 2-Bit Activation Compressed Training

The increasing size of neural network models has been critical for improvements in their accuracy, but device memory is not growing at the same rate. This creates fundamental challenges for training neural networks within limited memory environments. In this work, we propose ActNN, a memory-efficient training framework that stores randomly quantized activations for back propagation. We prove the convergence of ActNN for general network architectures, and we characterize the impact of quantization on the convergence via an exact expression for the gradient variance. Using our theory, we propose novel mixed-precision quantization strategies that exploit the activation's heterogeneity across feature dimensions, samples, and layers. These techniques can be readily applied to existing dynamic graph frameworks, such as PyTorch, simply by substituting the layers. We evaluate ActNN on mainstream computer vision models for classification, detection, and segmentation tasks. On all these tasks, ActNN compresses the activation to 2 bits on average, with negligible accuracy loss. ActNN reduces the memory footprint of the activation by 12x, and it enables training with a 6.6x to 14x larger batch size.

cs.LG↗

arXiv2021

Eliminating uncertainty of thermal emittance measurement in solenoid scans due to rf and solenoid fields overlap

The solenoid scan is one of the most common methods for the in-situ measurement of the thermal emittance of a photocathode in an rf photoinjector. The fringe field of the solenoid overlaps with the gun rf field in quite a number of photoinjectors, which makes accurate knowledge of the transfer matrix challenging, thus increases the measurement uncertainty of the thermal emittance. This paper summarizes two methods that have been used to solve the overlap issue and explains their deficiencies. Furthermore, we provide a new method to eliminate the measurement error due to the overlap issue in solenoid scans. The new method is systematically demonstrated using theoretical derivations, beam dynamics simulations, and experimental data based on the photoinjector configurations from three different groups, proving that the measurement error with the new method is very small and can be ignored in most of the photoinjector configurations.

physics.acc-ph↗

arXiv2020

Rapid thermal emittance and quantum efficiency mapping of a cesium telluride cathode in an rf photoinjector using multiple laser beamlets

Thermal emittance and quantum efficiency (QE) are key figures of merit of photocathodes, and their uniformity is critical to high-performance photoinjectors. Several QE mapping technologies have been successfully developed; however, there is still a dearth of information on thermal emittance maps. This is because of the extremely time-consuming procedure to gather measurements by scanning a small beam across the cathode with fine steps. To simplify the mapping procedure, and to reduce the time required to take measurements, we propose a new method that requires only a single scan of the solenoid current to simultaneously obtain thermal emittance and QE distribution by using a pattern beam with multiple beamlets. In this paper, its feasibility has been confirmed by both beam dynamics simulation and theoretical analysis. The method has been successfully demonstrated in a proof-of-principle experiment using an L-band radiofrequency photoinjector with a cesium telluride cathode. In the experiment, seven beamlets were generated from a microlens array system and their corresponding thermal emittance and QE varied from 0.93 to 1.14 $μ$m/mm and from 4.6 to 8.7%, respectively. We also discuss the limitations and future improvements of the method in this paper.

physics.acc-ph↗

arXiv2019

A Unified Optimization Approach for CNN Model Inference on Integrated GPUs

Modern deep learning applications urge to push the model inference taking place at the edge devices for multiple reasons such as achieving shorter latency, relieving the burden of the network connecting to the cloud, and protecting user privacy. The Convolutional Neural Network (\emph{CNN}) is one of the most widely used model family in the applications. Given the high computational complexity of the CNN models, it is favorable to execute them on the integrated GPUs at the edge devices, which are ubiquitous and have more power and better energy efficiency than the accompanying CPUs. However, programming on integrated GPUs efficiently is challenging due to the variety of their architectures and programming interfaces. This paper proposes an end-to-end solution to execute CNN model inference on the integrated GPUs at the edge, which uses a unified IR to represent and optimize vision-specific operators on integrated GPUs from multiple vendors, as well as leverages machine learning-based scheduling search schemes to optimize computationally-intensive operators like convolution. Our solution even provides a fallback mechanism for operators not suitable or convenient to run on GPUs. The evaluation results suggest that compared to state-of-the-art solutions backed up by the vendor-provided high-performance libraries on Intel Graphics, ARM Mali GPU, and Nvidia integrated Maxwell GPU, our solution achieves similar, or even better (up to 1.62$\times$), performance on a number of popular image classification and object detection models. In addition, our solution has a wider model coverage and is more flexible to embrace new models. Our solution has been adopted in production services in AWS and is open-sourced.

cs.DC↗

arXiv2019

Development and high-power testing of an X-band dielectric-loaded power extractor

Dielectric loaded structures are promising candidates for use in the structure wakefield acceleration (SWFA) technique, for both the collinear wakefield and the two-beam acceleration (CWA and TBA respectively) approaches, due to their low fabrication cost, low rf losses, and the potential to withstand high gradient. A short pulse (<=20 ns) TBA program is under development at the Argonne Wakefield Accelerator (AWA) facility where dielectric loaded structures are being used for both the power extractor/transfer structure (PETS) and the accelerator. In this study, an X-band 11.7 GHz dielectric PETS was developed and tested at the AWA facility to demonstrate high power wakefield generation. The PETS was driven by a train of eight electron bunches separated by 769.2 ps (9 times of the X-band rf period) in order to achieve coherent wakefield superposition. A total train charge of 360 nC was passed through the PETS structure to generate ~200 MW, ~3 ns flat-top rf pulses without rf breakdown. A future experiment is being planned to increase the generated rf power to approximately ~1 GW by optimizing the structure design and improving the drive beam quality.

physics.acc-ph↗

arXiv2019

A Hardware-Software Blueprint for Flexible Deep Learning Specialization

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.

cs.LG↗

arXiv2019

Experimental demonstration of the correction of coupled transverse dynamics aberration in an rf photoinjector

The production of electron bunches with low transverse emittance approaches the thermal emittance of the photocathode as various aberrations are corrected. Recently, the coupled transverse dynamics aberration was theoretically identified as a significant source of emittance growth and a corrector magnet was proposed for its elimination [D.H. Dowell, F. Zhou, and J. Schmerge, PRAB 21, 010101 (2018)]. This aberration arises when the beam acquires an asymmetric distribution that is then rotated with respect to the transverse reference axis thus introducing a correlation in the vertical and horizontal planes. The asymmetry is introduced by a weak quadrupole field in the rf gun or emittance compensation solenoid and the rotation is caused by the solenoid. This Letter presents an experimental study of the coupled transverse dynamics aberration in an rf photoinjector and demonstrates its elimination by a quadrupole corrector consisting of a normal and a skew quadrupole. The experimental results agree well with theoretical predictions and numerical simulations. The study also demonstrates the emittance of a low charge beam can be preserved during transportation at its thermal value, which was 1.05 mm mrad/mm, for the cesium telluride photocathode and 248 nm UV laser used.

physics.acc-ph↗

arXiv2019

Learning to Optimize Tensor Programs

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.

cs.LG↗

arXiv2018

Overestimation of thermal emittance in solenoid scans due to coupled transverse motion

The solenoid scan is a widely used method for the in-situ measurement of the thermal emittance in a photocathode gun. The popularity of this method is due to its simplicity and convenience since all rf photocathode guns are equipped with an emittance compensation solenoid. This paper shows that the solenoid scan measurement overestimates the thermal emittance in the ordinary measurement configuration due to a weak quadrupole field (present in either the rf gun or gun solenoid) followed by a rotation in the solenoid. This coupled transverse dynamics aberration introduces a correlation between the beam's horizontal and vertical motion leading to an increase in the measured 2D transverse emittance, thus the overestimation of the thermal emittance. This effect was systematically studied using both analytic expressions and numerical simulations. These studies were experimentally verified using an L-band 1.6-cell rf photocathode gun with a cesium telluride cathode, which shows a thermal emittance overestimation of 35% with a rms laser spot size of 2.7 mm. The paper concludes by showing that the accuracy of the solenoid scan can be improved by using a quadrupole magnet corrector, consisting of a pair of normal and skew quadrupole magnets.

physics.acc-ph↗

arXiv2018

TVM: An Automated End-to-End Optimizing Compiler for Deep Learning

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.

cs.LG↗

arXiv2018

Size-to-depth: A New Perspective for Single Image Depth Estimation

In this paper we consider the problem of single monocular image depth estimation. It is a challenging problem due to its ill-posedness nature and has found wide application in industry. Previous efforts belongs roughly to two families: learning-based method and interactive method. Learning-based method, in which deep convolutional neural network (CNN) is widely used, can achieve good result. But they suffer low generalization ability and typically perform poorly for unfamiliar scenes. Besides, data-hungry nature for such method makes data aquisition expensive and time-consuming. Interactive method requires human annotation of depth which, however, is errorneous and of large variance. To overcome these problems, we propose a new perspective for single monocular image depth estimation problem: size to depth. Our method require sparse label for real-world size of object rather than raw depth. A Coarse depth map is then inferred following geometric relationships according to size labels. Then we refine the depth map by doing energy function optimization on conditional random field(CRF). We experimentally demonstrate that our method outperforms traditional depth-labeling methods and can produce satisfactory depth maps.

cs.CV↗