Daiyaan Arfeen

Daiyaan Arfeen, Zhen Zhang, Xinwei Fu et al.

Training Deep Neural Networks (DNNs) with billions of parameters generally involves pipeline-parallel (PP) execution. Unfortunately, PP model training can use GPUs inefficiently, especially at large scale, due to idle GPU time caused by pipeline bubbles, which are often 15-30% and can exceed 60% of the training job's GPU allocation. To improve the GPU utilization of PP model training, this paper describes PipeFill, which fills pipeline bubbles with execution of other pending jobs. By leveraging bubble GPU time, PipeFill reduces the GPU utilization sacrifice associated with scaling-up of large-model training. To context-switch between fill jobs and the main training job with minimal overhead to the main job, and maximize fill job efficiency, PipeFill carefully fits fill job work to measured bubble durations and GPU memory availability, introduces explicit pipeline-bubble instructions, and orchestrates placement and execution of fill jobs in pipeline bubbles. Experiments show that PipeFill can increase overall utilization by up to 63% for GPUs used in large-scale LLM training, with <2% slowdown of the training job, and 5-15% even for low-scale LLM training. For large-scale LLM training on 8K GPUs, the 63% increase translates to up to 2.6K additional GPUs worth of work completed.

Xupeng Miao, Gabriele Oliaro, Zhihao Zhang et al.

This paper introduces SpecInfer, a system that accelerates generative large language model (LLM) serving with tree-based speculative inference and verification. The key idea behind SpecInfer is leveraging small speculative models to predict the LLM's outputs; the predictions are organized as a token tree, whose nodes each represent a candidate token sequence. The correctness of all candidate token sequences represented by a token tree is verified against the LLM in parallel using a novel tree-based parallel decoding mechanism. SpecInfer uses an LLM as a token tree verifier instead of an incremental decoder, which significantly reduces the end-to-end latency and computational requirement for serving generative LLMs while provably preserving model quality. Our evaluation shows that SpecInfer outperforms existing LLM serving systems by 1.5-2.8x for distributed LLM inference and by 2.6-3.5x for offloading-based LLM inference, while preserving the same generative performance. SpecInfer is publicly available at https://github.com/flexflow/FlexFlow/

Daiyaan Arfeen, Dheevatsa Mudigere, Ankit More et al.

LLM training is scaled up to 10Ks of GPUs by a mix of data-(DP) and model-parallel (MP) execution. Critical to achieving efficiency is tensor-parallel (TP; a form of MP) execution within tightly-coupled subsets of GPUs, referred to as a scale-up domain, and the larger the scale-up domain the better the performance. New datacenter architectures are emerging with more GPUs able to be tightly-coupled in a scale-up domain, such as moving from 8 GPUs to 72 GPUs connected via NVLink. Unfortunately, larger scale-up domains increase the blast-radius of failures, with a failure of single GPU potentially impacting TP execution on the full scale-up domain, which can degrade overall LLM training throughput dramatically. With as few as 0.1% of GPUs being in a failed state, a high TP-degree job can experience nearly 10% reduction in LLM training throughput. We propose nonuniform-tensor-parallelism (NTP) to mitigate this amplified impact of GPU failures. In NTP, a DP replica that experiences GPU failures operates at a reduced TP degree, contributing throughput equal to the percentage of still-functional GPUs. We also propose a rack-design with improved electrical and thermal capabilities in order to sustain power-boosting of scale-up domains that have experienced failures; combined with NTP, this can allow the DP replica with the reduced TP degree (i.e., with failed GPUs) to keep up with the others, thereby achieving near-zero throughput loss for large-scale LLM training.

Byungsoo Jeon, Mengdi Wu, Shiyi Cao et al.

Deep neural networks (DNNs) continue to grow rapidly in size, making them infeasible to train on a single device. Pipeline parallelism is commonly used in existing DNN systems to support large-scale DNN training by partitioning a DNN into multiple stages, which concurrently perform DNN training for different micro-batches in a pipeline fashion. However, existing pipeline-parallel approaches only consider sequential pipeline stages and thus ignore the topology of a DNN, resulting in missed model-parallel opportunities. This paper presents graph pipeline parallelism (GPP), a new pipeline-parallel scheme that partitions a DNN into pipeline stages whose dependencies are identified by a directed acyclic graph. GPP generalizes existing sequential pipeline parallelism and preserves the inherent topology of a DNN to enable concurrent execution of computationally-independent operators, resulting in reduced memory requirement and improved GPU performance. In addition, we develop GraphPipe, a distributed system that exploits GPP strategies to enable performant and scalable DNN training. GraphPipe partitions a DNN into a graph of stages, optimizes micro-batch schedules for these stages, and parallelizes DNN training using the discovered GPP strategies. Evaluation on a variety of DNNs shows that GraphPipe outperforms existing pipeline-parallel systems such as PipeDream and Piper by up to 1.6X. GraphPipe also reduces the search time by 9-21X compared to PipeDream and Piper.

Zhen Dong, Zhewei Yao, Yaohui Cai et al.

Quantization is an effective method for reducing memory footprint and inference time of Neural Networks, e.g., for efficient inference in the cloud, especially at the edge. However, ultra low precision quantization could lead to significant degradation in model generalization. A promising method to address this is to perform mixed-precision quantization, where more sensitive layers are kept at higher precision. However, the search space for a mixed-precision quantization is exponential in the number of layers. Recent work has proposed HAWQ, a novel Hessian based framework, with the aim of reducing this exponential search space by using second-order information. While promising, this prior work has three major limitations: (i) HAWQV1 only uses the top Hessian eigenvalue as a measure of sensitivity and do not consider the rest of the Hessian spectrum; (ii) HAWQV1 approach only provides relative sensitivity of different layers and therefore requires a manual selection of the mixed-precision setting; and (iii) HAWQV1 does not consider mixed-precision activation quantization. Here, we present HAWQV2 which addresses these shortcomings. For (i), we perform a theoretical analysis showing that a better sensitivity metric is to compute the average of all of the Hessian eigenvalues. For (ii), we develop a Pareto frontier based method for selecting the exact bit precision of different layers without any manual selection. For (iii), we extend the Hessian analysis to mixed-precision activation quantization. We have found this to be very beneficial for object detection. We show that HAWQV2 achieves new state-of-the-art results for a wide range of tasks.

Daiyaan Arfeen, Jesse Zhang

We propose the use of unsupervised learning to train projection networks that project onto the latent space of an already trained generator. We apply our method to a trained StyleGAN, and use our projection network to perform image super-resolution and clustering of images into semantically identifiable groups.

Zhewei Yao, Amir Gholami, Daiyaan Arfeen et al.

The most straightforward method to accelerate Stochastic Gradient Descent (SGD) computation is to distribute the randomly selected batch of inputs over multiple processors. To keep the distributed processors fully utilized requires commensurately growing the batch size. However, large batch training often leads to poorer generalization. A recently proposed solution for this problem is to use adaptive batch sizes in SGD. In this case, one starts with a small number of processes and scales the processes as training progresses. Two major challenges with this approach are (i) that dynamically resizing the cluster can add non-trivial overhead, in part since it is currently not supported, and (ii) that the overall speed up is limited by the initial phase with smaller batches. In this work, we address both challenges by developing a new adaptive batch size framework, with autoscaling based on the Ray framework. This allows very efficient elastic scaling with negligible resizing overhead (0.32\% of time for ResNet18 ImageNet training). Furthermore, we propose a new adaptive batch size training scheme using second order methods and adversarial training. These enable increasing batch sizes earlier during training, which leads to better training time. We extensively evaluate our method on Cifar-10/100, SVHN, TinyImageNet, and ImageNet datasets, using multiple neural networks, including ResNets and smaller networks such as SqueezeNext. Our method exceeds the performance of existing solutions in terms of both accuracy and the number of SGD iterations (up to 1\% and $5\times$, respectively). Importantly, this is achieved without any additional hyper-parameter tuning to tailor our method in any of these experiments.