Shigang Li

Wenqi Jiang, Shigang Li, Yu Zhu et al. · amazon-science

Vector search has emerged as the foundation for large-scale information retrieval and machine learning systems, with search engines like Google and Bing processing tens of thousands of queries per second on petabyte-scale document datasets by evaluating vector similarities between encoded query texts and web documents. As performance demands for vector search systems surge, accelerated hardware offers a promising solution in the post-Moore's Law era. We introduce \textit{FANNS}, an end-to-end and scalable vector search framework on FPGAs. Given a user-provided recall requirement on a dataset and a hardware resource budget, \textit{FANNS} automatically co-designs hardware and algorithm, subsequently generating the corresponding accelerator. The framework also supports scale-out by incorporating a hardware TCP/IP stack in the accelerator. \textit{FANNS} attains up to 23.0$\times$ and 37.2$\times$ speedup compared to FPGA and CPU baselines, respectively, and demonstrates superior scalability to GPUs, achieving 5.5$\times$ and 7.6$\times$ speedup in median and 95\textsuperscript{th} percentile (P95) latency within an eight-accelerator configuration. The remarkable performance of \textit{FANNS} lays a robust groundwork for future FPGA integration in data centers and AI supercomputers.

Shigang Li, Kazuki Osawa, Torsten Hoefler

The exponentially growing model size drives the continued success of deep learning, but it brings prohibitive computation and memory cost. From the algorithm perspective, model sparsification and quantization have been studied to alleviate the problem. From the architecture perspective, hardware vendors provide Tensor cores for acceleration. However, it is very challenging to gain practical speedups from sparse, low-precision matrix operations on Tensor cores, because of the strict requirements for data layout and lack of support for efficiently manipulating the low-precision integers. We propose Magicube, a high-performance sparse-matrix library for low-precision integers on Tensor cores. Magicube supports SpMM and SDDMM, two major sparse operations in deep learning with mixed precision. Experimental results on an NVIDIA A100 GPU show that Magicube achieves on average 1.44x (up to 2.37x) speedup over the vendor-optimized library for sparse kernels, and 1.43x speedup over the state-of-the-art with a comparable accuracy for end-to-end sparse Transformer inference.

Torsten Hoefler, Tommaso Bonato, Daniele De Sensi et al.

Numerous microarchitectural optimizations unlocked tremendous processing power for deep neural networks that in turn fueled the AI revolution. With the exhaustion of such optimizations, the growth of modern AI is now gated by the performance of training systems, especially their data movement. Instead of focusing on single accelerators, we investigate data-movement characteristics of large-scale training at full system scale. Based on our workload analysis, we design HammingMesh, a novel network topology that provides high bandwidth at low cost with high job scheduling flexibility. Specifically, HammingMesh can support full bandwidth and isolation to deep learning training jobs with two dimensions of parallelism. Furthermore, it also supports high global bandwidth for generic traffic. Thus, HammingMesh will power future large-scale deep learning systems with extreme bandwidth requirements.

Kazuki Osawa, Shigang Li, Torsten Hoefler

Pipeline parallelism enables efficient training of Large Language Models (LLMs) on large-scale distributed accelerator clusters. Yet, pipeline bubbles during startup and tear-down reduce the utilization of accelerators. Although efficient pipeline schemes with micro-batching and bidirectional pipelines have been proposed to maximize utilization, a significant number of bubbles cannot be filled using synchronous forward and backward passes. To address this problem, we suggest that extra work be assigned to the bubbles to gain auxiliary benefits in LLM training. As an example in this direction, we propose PipeFisher, which assigns the work of K-FAC, a second-order optimization method based on the Fisher information matrix, to the bubbles to accelerate convergence. In Phase 1 pretraining of BERT-Base and -Large models, PipeFisher reduces the (simulated) training time to 50-75% compared to training with a first-order optimizer by greatly improving the accelerator utilization and benefiting from the improved convergence by K-FAC.

Baodong Wu, Lei Xia, Qingping Li et al.

Large language models (LLMs) with hundreds of billions or trillions of parameters, represented by chatGPT, have achieved profound impact on various fields. However, training LLMs with super-large-scale parameters requires large high-performance GPU clusters and long training periods lasting for months. Due to the inevitable hardware and software failures in large-scale clusters, maintaining uninterrupted and long-duration training is extremely challenging. As a result, A substantial amount of training time is devoted to task checkpoint saving and loading, task rescheduling and restart, and task manual anomaly checks, which greatly harms the overall training efficiency. To address these issues, we propose TRANSOM, a novel fault-tolerant LLM training system. In this work, we design three key subsystems: the training pipeline automatic fault tolerance and recovery mechanism named Transom Operator and Launcher (TOL), the training task multi-dimensional metric automatic anomaly detection system named Transom Eagle Eye (TEE), and the training checkpoint asynchronous access automatic fault tolerance and recovery technology named Transom Checkpoint Engine (TCE). Here, TOL manages the lifecycle of training tasks, while TEE is responsible for task monitoring and anomaly reporting. TEE detects training anomalies and reports them to TOL, who automatically enters the fault tolerance strategy to eliminate abnormal nodes and restart the training task. And the asynchronous checkpoint saving and loading functionality provided by TCE greatly shorten the fault tolerance overhead. The experimental results indicate that TRANSOM significantly enhances the efficiency of large-scale LLM training on clusters. Specifically, the pre-training time for GPT3-175B has been reduced by 28%, while checkpoint saving and loading performance have improved by a factor of 20.

Heyu Chen, Jianfeng Li, Shigang Li

Nowadays, panoramic images can be easily obtained by panoramic cameras. However, when the panoramic camera orientation is tilted, a non-upright panoramic image will be captured. Existing upright adjustment models focus on how to estimate more accurate camera orientation, and attribute image reconstruction to offline or post-processing tasks. To this end, we propose an online end-to-end network for upright adjustment. Our network is designed to reconstruct the image while finding the angle. Our network consists of three modules: orientation estimation, LUT online generation, and upright reconstruction. Direction estimation estimates the tilt angle of the panoramic image. Then, a converter block with upsampling function is designed to generate angle to LUT. This module can output corresponding online LUT for different input angles. Finally, a lightweight generative adversarial network (GAN) aims to generate upright images from shallow features. The experimental results show that in terms of angles, we have improved the accuracy of small angle errors. In terms of image reconstruction, In image reconstruction, we have achieved the first real-time online upright reconstruction of panoramic images using deep learning networks.

Yanxin Peng, Qingping Li, Baodong Wu et al.

As large language models (LLMs) continue to grow in size and complexity, efficient checkpoint saving\&loading has become crucial for managing storage, memory usage, and fault tolerance in LLM training. The current works do not comprehensively take into account the optimization of these several aspects. This paper proposes a novel checkpoint sparsification and quantization method that adapts dynamically to different training stages and model architectures. We present a comprehensive analysis of existing lossy and lossless compression techniques, identify current limitations, and introduce our adaptive approach that balances compression ratio, speed, and precision impact throughout the training process. Experiments on different sizes of LLMs demonstrate that our bitmask-based sparsification method achieves 16x compression ratio without compromising model accuracy. Additionally, the cluster-based quantization method achieves 2x compression ratio with little precision loss.

Jinliang Shi, Shigang Li, Youxuan Xu et al.

Sparse Matrix-matrix Multiplication (SpMM) and Sampled Dense-dense Matrix Multiplication (SDDMM) are important sparse operators in scientific computing and deep learning. Tensor Core Units (TCUs) enhance modern accelerators with superior computing power, which is promising to boost the performance of matrix operators to a higher level. However, due to the irregularity of unstructured sparse data, it is difficult to deliver practical speedups on TCUs. To this end, we propose FlashSparse, a novel approach to bridge the gap between sparse workloads and the TCU architecture. Specifically, FlashSparse minimizes the sparse granularity for SpMM and SDDMM on TCUs through a novel swap-and-transpose matrix multiplication strategy. Benefiting from the minimum sparse granularity, the computation redundancy is remarkably reduced while the computing power of TCUs is fully utilized. Besides, FlashSparse is equipped with a memory-efficient thread mapping strategy for coalesced data access and a sparse matrix storage format to save memory footprint. Extensive experimental results on H100 and RTX 4090 GPUs show that FlashSparse sets a new state-of-the-art for sparse matrix multiplications (geometric mean 5.5x speedup over DTC-SpMM and 3.22x speedup over RoDe).

Jinliang Shi, Shigang Li, Youxuan Xu et al.

Sparse matrix multiplication operators (i.e., SpMM and SDDMM) are widely used in deep learning and scientific computing. Modern accelerators are commonly equipped with Tensor cores and CUDA cores to accelerate sparse operators. The former brings superior computing power but only for structured matrix multiplication, while the latter has relatively lower performance but with higher programming flexibility. In this work, we discover that utilizing one resource alone leads to inferior performance for sparse matrix multiplication, due to their respective limitations. To this end, we propose Libra, a systematic approach that enables synergistic computation between CUDA and Tensor cores to achieve the best performance for sparse matrix multiplication. Specifically, we propose a 2D-aware workload distribution strategy to find out the sweet point of task mapping for different sparse operators, leveraging both the high performance of Tensor cores and the low computational redundancy on CUDA cores. In addition, Libra incorporates systematic optimizations for heterogeneous computing, including hybrid load-balancing, finely optimized kernel implementations, and GPU-accelerated preprocessing. Extensive experimental results on H100 and RTX 4090 GPUs show that Libra outperforms the state-of-the-art by on average 3.1x (up to 9.23x) over DTC-SpMM and 2.9x (up to 3.9x) for end-to-end GNN applications. Libra opens up a new perspective for sparse operator acceleration by fully exploiting the heterogeneous computing resources on GPUs.

Jingguo Liu, Han Yu, Shigang Li et al.

Due to the current lack of large-scale datasets at the million-scale level, tasks involving panoramic images predominantly rely on existing two-dimensional pre-trained image benchmark models as backbone networks. However, these networks are not equipped to recognize the distortions and discontinuities inherent in panoramic images, which adversely affects their performance in such tasks. In this paper, we introduce a novel spherical sampling method for panoramic images that enables the direct utilization of existing pre-trained models developed for two-dimensional images. Our method employs spherical discrete sampling based on the weights of the pre-trained models, effectively mitigating distortions while achieving favorable initial training values. Additionally, we apply the proposed sampling method to panoramic image segmentation, utilizing features obtained from the spherical model as masks for specific channel attentions, which yields commendable results on commonly used indoor datasets, Stanford2D3D.

Kazuki Osawa, Satoki Ishikawa, Rio Yokota et al.

Gradient preconditioning is a key technique to integrate the second-order information into gradients for improving and extending gradient-based learning algorithms. In deep learning, stochasticity, nonconvexity, and high dimensionality lead to a wide variety of gradient preconditioning methods, with implementation complexity and inconsistent performance and feasibility. We propose the Automatic Second-order Differentiation Library (ASDL), an extension library for PyTorch, which offers various implementations and a plug-and-play unified interface for gradient preconditioning. ASDL enables the study and structured comparison of a range of gradient preconditioning methods.

Shigang Li, Torsten Hoefler

Communication overhead is one of the major obstacles to train large deep learning models at scale. Gradient sparsification is a promising technique to reduce the communication volume. However, it is very challenging to obtain real performance improvement because of (1) the difficulty of achieving an scalable and efficient sparse allreduce algorithm and (2) the sparsification overhead. This paper proposes O$k$-Top$k$, a scheme for distributed training with sparse gradients. O$k$-Top$k$ integrates a novel sparse allreduce algorithm (less than 6$k$ communication volume which is asymptotically optimal) with the decentralized parallel Stochastic Gradient Descent (SGD) optimizer, and its convergence is proved. To reduce the sparsification overhead, O$k$-Top$k$ efficiently selects the top-$k$ gradient values according to an estimated threshold. Evaluations are conducted on the Piz Daint supercomputer with neural network models from different deep learning domains. Empirical results show that O$k$-Top$k$ achieves similar model accuracy to dense allreduce. Compared with the optimized dense and the state-of-the-art sparse allreduces, O$k$-Top$k$ is more scalable and significantly improves training throughput (e.g., 3.29x-12.95x improvement for BERT on 256 GPUs).

Oliver Rausch, Tal Ben-Nun, Nikoli Dryden et al.

Rapid progress in deep learning is leading to a diverse set of quickly changing models, with a dramatically growing demand for compute. However, as frameworks specialize performance optimization to patterns in popular networks, they implicitly constrain novel and diverse models that drive progress in research. We empower deep learning researchers by defining a flexible and user-customizable pipeline for optimizing training of arbitrary deep neural networks, based on data movement minimization. The pipeline begins with standard networks in PyTorch or ONNX and transforms computation through progressive lowering. We define four levels of general-purpose transformations, from local intra-operator optimizations to global data movement reduction. These operate on a data-centric graph intermediate representation that expresses computation and data movement at all levels of abstraction, including expanding basic operators such as convolutions to their underlying computations. Central to the design is the interactive and introspectable nature of the pipeline. Every part is extensible through a Python API, and can be tuned interactively using a GUI. We demonstrate competitive performance or speedups on ten different networks, with interactive optimizations discovering new opportunities in EfficientNet.

Shigang Li, Torsten Hoefler

Training large deep learning models at scale is very challenging. This paper proposes Chimera, a novel pipeline parallelism scheme which combines bidirectional pipelines for efficiently training large-scale models. Chimera is a synchronous approach and therefore no loss of accuracy, which is more convergence-friendly than asynchronous approaches. Compared with the latest synchronous pipeline approach, Chimera reduces the number of bubbles by up to 50%; benefiting from the sophisticated scheduling of bidirectional pipelines, Chimera has a more balanced activation memory consumption. Evaluations are conducted on Transformer based language models. For a GPT-2 model with 1.3 billion parameters running on 2,048 GPU nodes of the Piz Daint supercomputer, Chimera improves the training throughput by 1.16x-2.34x over the state-of-the-art synchronous and asynchronous pipeline approaches.

Andrei Ivanov, Nikoli Dryden, Tal Ben-Nun et al.

Transformers are one of the most important machine learning workloads today. Training one is a very compute-intensive task, often taking days or weeks, and significant attention has been given to optimizing transformers. Despite this, existing implementations do not efficiently utilize GPUs. We find that data movement is the key bottleneck when training. Due to Amdahl's Law and massive improvements in compute performance, training has now become memory-bound. Further, existing frameworks use suboptimal data layouts. Using these insights, we present a recipe for globally optimizing data movement in transformers. We reduce data movement by up to 22.91% and overall achieve a 1.30x performance improvement over state-of-the-art frameworks when training a BERT encoder layer and 1.19x for the entire BERT. Our approach is applicable more broadly to optimizing deep neural networks, and offers insight into how to tackle emerging performance bottlenecks.

Peter Grönquist, Chengyuan Yao, Tal Ben-Nun et al.

Quantifying uncertainty in weather forecasts is critical, especially for predicting extreme weather events. This is typically accomplished with ensemble prediction systems, which consist of many perturbed numerical weather simulations, or trajectories, run in parallel. These systems are associated with a high computational cost and often involve statistical post-processing steps to inexpensively improve their raw prediction qualities. We propose a mixed model that uses only a subset of the original weather trajectories combined with a post-processing step using deep neural networks. These enable the model to account for non-linear relationships that are not captured by current numerical models or post-processing methods. Applied to global data, our mixed models achieve a relative improvement in ensemble forecast skill (CRPS) of over 14%. Furthermore, we demonstrate that the improvement is larger for extreme weather events on select case studies. We also show that our post-processing can use fewer trajectories to achieve comparable results to the full ensemble. By using fewer trajectories, the computational costs of an ensemble prediction system can be reduced, allowing it to run at higher resolution and produce more accurate forecasts.

Peter Grönquist, Tal Ben-Nun, Nikoli Dryden et al.

Modern weather forecast models perform uncertainty quantification using ensemble prediction systems, which collect nonparametric statistics based on multiple perturbed simulations. To provide accurate estimation, dozens of such computationally intensive simulations must be run. We show that deep neural networks can be used on a small set of numerical weather simulations to estimate the spread of a weather forecast, significantly reducing computational cost. To train the system, we both modify the 3D U-Net architecture and explore models that incorporate temporal data. Our models serve as a starting point to improve uncertainty quantification in current real-time weather forecasting systems, which is vital for predicting extreme events.

Giorgi Nadiradze, Amirmojtaba Sabour, Peter Davies et al.

Decentralized optimization is emerging as a viable alternative for scalable distributed machine learning, but also introduces new challenges in terms of synchronization costs. To this end, several communication-reduction techniques, such as non-blocking communication, quantization, and local steps, have been explored in the decentralized setting. Due to the complexity of analyzing optimization in such a relaxed setting, this line of work often assumes \emph{global} communication rounds, which require additional synchronization. In this paper, we consider decentralized optimization in the simpler, but harder to analyze, \emph{asynchronous gossip} model, in which communication occurs in discrete, randomly chosen pairings among nodes. Perhaps surprisingly, we show that a variant of SGD called \emph{SwarmSGD} still converges in this setting, even if \emph{non-blocking communication}, \emph{quantization}, and \emph{local steps} are all applied \emph{in conjunction}, and even if the node data distributions and underlying graph topology are both \emph{heterogenous}. Our analysis is based on a new connection with multi-dimensional load-balancing processes. We implement this algorithm and deploy it in a super-computing environment, showing that it can outperform previous decentralized methods in terms of end-to-end training time, and that it can even rival carefully-tuned large-batch SGD for certain tasks.

Shigang Li, Tal Ben-Nun, Salvatore Di Girolamo et al.

Load imbalance pervasively exists in distributed deep learning training systems, either caused by the inherent imbalance in learned tasks or by the system itself. Traditional synchronous Stochastic Gradient Descent (SGD) achieves good accuracy for a wide variety of tasks, but relies on global synchronization to accumulate the gradients at every training step. In this paper, we propose eager-SGD, which relaxes the global synchronization for decentralized accumulation. To implement eager-SGD, we propose to use two partial collectives: solo and majority. With solo allreduce, the faster processes contribute their gradients eagerly without waiting for the slower processes, whereas with majority allreduce, at least half of the participants must contribute gradients before continuing, all without using a central parameter server. We theoretically prove the convergence of the algorithms and describe the partial collectives in detail. Experimental results on load-imbalanced environments (CIFAR-10, ImageNet, and UCF101 datasets) show that eager-SGD achieves 1.27x speedup over the state-of-the-art synchronous SGD, without losing accuracy.

Cheng Daning, Zhang Hanping, Xia Fen et al.

Automl is the key technology for machine learning problem. Current state of art hyperparameter optimization methods are based on traditional black-box optimization methods like SMBO (SMAC, TPE). The objective function of black-box optimization is non-smooth, or time-consuming to evaluate, or in some way noisy. Recent years, many researchers offered the work about the properties of hyperparameters. However, traditional hyperparameter optimization methods do not take those information into consideration. In this paper, we use gradient information and machine learning model analysis information to accelerate traditional hyperparameter optimization methods SMBO. In our L2 norm experiments, our method yielded state-of-the-art performance, and in many cases outperformed the previous best configuration approach.

Cheng Daning, Xia Fen, Li Shigang et al.

In AI research and industry, machine learning is the most widely used tool. One of the most important machine learning algorithms is Gradient Boosting Decision Tree, i.e. GBDT whose training process needs considerable computational resources and time. To shorten GBDT training time, many works tried to apply GBDT on Parameter Server. However, those GBDT algorithms are synchronous parallel algorithms which fail to make full use of Parameter Server. In this paper, we examine the possibility of using asynchronous parallel methods to train GBDT model and name this algorithm as asynch-SGBDT (asynchronous parallel stochastic gradient boosting decision tree). Our theoretical and experimental results indicate that the scalability of asynch-SGBDT is influenced by the sample diversity of datasets, sampling rate, step length and the setting of GBDT tree. Experimental results also show asynch-SGBDT training process reaches a linear speedup in asynchronous parallel manner when datasets and GBDT trees meet high scalability requirements.

Cheng Daning, Li Shigang, Zhang Yunquan

Stochastic gradient descent (SGD) is a popular stochastic optimization method in machine learning. Traditional parallel SGD algorithms, e.g., SimuParallel SGD, often require all nodes to have the same performance or to consume equal quantities of data. However, these requirements are difficult to satisfy when the parallel SGD algorithms run in a heterogeneous computing environment; low-performance nodes will exert a negative influence on the final result. In this paper, we propose an algorithm called weighted parallel SGD (WP-SGD). WP-SGD combines weighted model parameters from different nodes in the system to produce the final output. WP-SGD makes use of the reduction in standard deviation to compensate for the loss from the inconsistency in performance of nodes in the cluster, which means that WP-SGD does not require that all nodes consume equal quantities of data. We also analyze the theoretical feasibility of running two other parallel SGD algorithms combined with WP-SGD in a heterogeneous environment. The experimental results show that WP-SGD significantly outperforms the traditional parallel SGD algorithms on distributed training systems with an unbalanced workload.