Zero Redundancy Optimizer (ZeRO) has been used to train a wide range of large language models on massive GPU clusters due to its ease of use, efficiency, and good scalability. However, when training on low-bandwidth clusters, and/or when small batch size per GPU is used, ZeRO’s effective throughput is limited due to communication overheads. To alleviate this limitation, this paper introduces ZeRO++ composing of three communication volume reduction techniques (lowprecision all-gather, data remapping, and low-precision gradient averaging) to significantly reduce the communication volume up to 4x that enables up to 2.16x better throughput at 384 GPU scale. Our results also show ZeRO++ can speedup the RLHF by 3.3x compared to vanilla ZeRO. To verify the convergence of ZeRO++, we test up to 13B model for pretraining with 8/6-bits all gather and up to 30B model for finetuning with 4/2-bits all gather, and demonstrate on-par accuracy as original ZeRO (aka standard training). As a byproduct, the model trained with ZeRO++ is naturally weight-quantized, which can be directly used for inference without post-training quantization or quantization-aware training.

Large scale training requires massive parallelism to finish the training within a reasonable amount of time. To support massive parallelism, large batch training is the key enabler but often at the cost of generalization performance. Existing works explore adaptive batching or hand-tuned static large batching, in order to strike a balance between the computational efficiency and the performance. However, these methods can provide only coarse-grained adaption (e.g., at a epoch level) due to the intrinsic expensive calculation or hand tuning requirements. In this paper, we propose a fully automated and lightweight adaptive batching methodology to enable fine-grained batch size adaption (e.g., at a mini-batch level) that can achieve state-of-the-art performance with record breaking batch sizes. The core component of our method is a lightweight yet efficient representation of the critical gradient noise information. We open-source the proposed methodology by providing a plugin tool that supports mainstream machine learning frameworks. Extensive evaluations on popular benchmarks (e.g., CIFAR10, ImageNet, and BERT-Large) demonstrate that the proposed methodology outperforms state-of-the-art methodologies using adaptive batching approaches or hand-tuned static strategies in both performance and batch size. Particularly, we achieve a new state-of-the-art batch size of 78k in BERT-Large pretraining with SQuAD score 90.69 compared to 90.58 reported in previous state-of-the-art with 59k batch size.

Machine learning (ML) has been embedded in many Internet of Things (IoT) applications (e.g., smart home and autonomous driving). Yet it is often infeasible to deploy ML models on IoT devices due to resource limitation. Thus, deploying trained ML models in the cloud and providing inference services to IoT devices becomes a plausible solution. To provide low-latency ML serving to massive IoT devices, a natural and promising approach is to use parallelism in computation. However, existing ML systems (e.g., Tensorflow) and cloud ML-serving platforms (e.g., SageMaker) are service-level-objective (SLO) agnostic and rely on users to manually configure the parallelism at both request and operation levels. To address this challenge, we propose a region-based reinforcement learning (RRL)-based scheduling framework for ML serving in IoT applications that can efficiently identify optimal configurations under dynamic workloads. A key observation is that the system performance under similar configurations in a region can be accurately estimated by using the system performance under one of these configurations due to their correlation. We theoretically show that the RRL approach can achieve fast convergence speed at the cost of performance loss. To improve the performance, we propose an adaptive RRL algorithm based on Bayesian optimization to balance the convergence speed and the optimality. The proposed framework is prototyped and evaluated on the Tensorflow Serving system. Extensive experimental results show that the proposed approach can outperform state-of-the-art approaches by finding near-optimal solutions over eight times faster while reducing inference latency up to 88.9% and reducing SLO violation up to 91.6%.

The success of machine learning has prospered Machine-Learning-as-a-Service (MLaaS) - deploying trained machine learning (ML) models in cloud to provide low latency inference services at scale. To meet latency Service-Level-Objective (SLO), judicious parallelization at both request and operation levels is utterly important. However, existing ML systems (e.g., Tensorflow) and cloud ML serving platforms (e.g., SageMaker) are SLO-agnostic and rely on users to manually configure the parallelism. To provide low latency ML serving, this paper proposes a swift machine learning serving scheduling framework with a novel Region-based Reinforcement Learning (RRL) approach. RRL can efficiently identify the optimal parallelism configuration under different workloads by estimating performance of similar configurations with that of the known ones. We both theoretically and experimentally show that the RRL approach can outperform state-of-the-art approaches by finding near optimal solutions over 8 times faster while reducing inference latency up to 79.0% and reducing SLO violation up to 49.9%.