اشترك بالحزمة الذهبية واحصل على وصول غير محدود شمرا أكاديميا
تسجيل مستخدم جديدSequence Parallelism: Making 4D Parallelism Possible
99
0
0.0
(
0
)
اسأل ChatGPT حول البحث
ﻻ يوجد ملخص باللغة العربية
Within Transformer, self-attention is the key module to learn powerful context-aware representations. However, self-attention suffers from quadratic memory requirements with respect to the sequence length, which limits us to process longer sequence on GPU. In this work, we propose sequence parallelism, a memory efficient parallelism method to help us break input sequence length limitation and train with longer sequence on GPUs. Compared with existing parallelism, our approach no longer requires a single device to hold the whole sequence. Specifically, we split the input sequence into multiple chunks and feed each chunk into its corresponding device (i.e. GPU). To compute the attention output, we communicate attention embeddings among GPUs. Inspired by ring all-reduce, we integrated ring-style communication with self-attention calculation and proposed Ring Self-Attention (RSA). Our implementation is fully based on PyTorch. Without extra compiler or library changes, our approach is compatible with data parallelism and pipeline parallelism. Experiments show that sequence parallelism performs well when scaling with batch size and sequence length. Compared with tensor parallelism, our approach achieved $13.7times$ and $3.0times$ maximum batch size and sequence length respectively when scaling up to 64 NVIDIA P100 GPUs. We plan to integrate our sequence parallelism with data, pipeline and tensor parallelism to further train large-scale models with 4D parallelism in our future work.
قيم البحث
اقرأ أيضاً
In this paper, we consider hybrid parallelism -- a paradigm that employs both Data Parallelism (DP) and Model Parallelism (MP) -- to scale distributed training of large recommendation models. We propose a compression framework called Dynamic Communic
ation Thresholding (DCT) for communication-efficient hybrid training. DCT filters the entities to be communicated across the network through a simple hard-thresholding function, allowing only the most relevant information to pass through. For communication efficient DP, DCT compresses the parameter gradients sent to the parameter server during model synchronization. The threshold is updated only once every few thousand iterations to reduce the computational overhead of compression. For communication efficient MP, DCT incorporates a novel technique to compress the activations and gradients sent across the network during the forward and backward propagation, respectively. This is done by identifying and updating only the most relevant neurons of the neural network for each training sample in the data. We evaluate DCT on publicly available natural language processing and recommender models and datasets, as well as recommendation systems used in production at Facebook. DCT reduces communication by at least $100times$ and $20times$ during DP and MP, respectively. The algorithm has been deployed in production, and it improves end-to-end training time for a state-of-the-art industrial recommender model by 37%, without any loss in performance.
The Graph Augmented Multi-layer Perceptron (GA-MLP) model is an attractive alternative to Graph Neural Networks (GNNs). This is because it is resistant to the over-smoothing problem, and deeper GA-MLP models yield better performance. GA-MLP models ar
e traditionally optimized by the Stochastic Gradient Descent (SGD). However, SGD suffers from the layer dependency problem, which prevents the gradients of different layers of GA-MLP models from being calculated in parallel. In this paper, we propose a parallel deep learning Alternating Direction Method of Multipliers (pdADMM) framework to achieve model parallelism: parameters in each layer of GA-MLP models can be updated in parallel. The extended pdADMM-Q algorithm reduces communication cost by utilizing the quantization technique. Theoretical convergence to a critical point of the pdADMM algorithm and the pdADMM-Q algorithm is provided with a sublinear convergence rate $o(1/k)$. Extensive experiments in six benchmark datasets demonstrate that the pdADMM can lead to high speedup, and outperforms all the existing state-of-the-art comparison methods.
The number of parameters in state of the art neural networks has drastically increased in recent years. This surge of interest in large scale neural networks has motivated the development of new distributed training strategies enabling such models. O
ne such strategy is model-parallel distributed training. Unfortunately, model-parallelism suffers from poor resource utilisation, which leads to wasted resources. In this work, we improve upon recent developments in an idealised model-parallel optimisation setting: local learning. Motivated by poor resource utilisation, we introduce a class of intermediary strategies between local and global learning referred to as interlocking backpropagation. These strategies preserve many of the compute-efficiency advantages of local optimisation, while recovering much of the task performance achieved by global optimisation. We assess our strategies on both image classification ResNets and Transformer language models, finding that our strategy consistently out-performs local learning in terms of task performance, and out-performs global learning in training efficiency.
Deep Learning has attracted considerable attention across multiple application domains, including computer vision, signal processing and natural language processing. Although quite a few single node deep learning frameworks exist, such as tensorflow,
pytorch and keras, we still lack a complete processing structure that can accommodate large scale data processing, version control, and deployment, all while staying agnostic of any specific single node framework. To bridge this gap, this paper proposes a new, higher level framework, i.e. Nemesyst, which uses databases along with model sequentialisation to allow processes to be fed unique and transformed data at the point of need. This facilitates near real-time application and makes models available for further training or use at any node that has access to the database simultaneously. Nemesyst is well suited as an application framework for internet of things aggregated control systems, deploying deep learning techniques to optimise individual machines in massive networks. To demonstrate this framework, we adopted a case study in a novel domain; deploying deep learning to optimise the high speed control of electrical power consumed by a massive internet of things network of retail refrigeration systems in proportion to load available on the UK National Grid (a demand side response). The case study demonstrated for the first time in such a setting how deep learning models, such as Recurrent Neural Networks (vanilla and Long-Short-Term Memory) and Generative Adversarial Networks paired with Nemesyst, achieve compelling performance, whilst still being malleable to future adjustments as both the data and requirements inevitably change over time.
Machine learning frameworks adopt iterative optimizers to train neural networks. Conventional eager execution separates the updating of trainable parameters from forward and backward computations. However, this approach introduces nontrivial training
time overhead due to the lack of data locality and computation parallelism. In this work, we propose to fuse the optimizer with forward or backward computation to better leverage locality and parallelism during training. By reordering the forward computation, gradient calculation, and parameter updating, our proposed method improves the efficiency of iterative optimizers. Experimental results demonstrate that we can achieve an up to 20% training time reduction on various configurations. Since our methods do not alter the optimizer algorithm, they can be used as a general plug-in technique to the training process.
الأسئلة المقترحة
سجل دخول لتتمكن من نشر تعليقات
التعليقات
جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
