اشترك بالحزمة الذهبية واحصل على وصول غير محدود شمرا أكاديميا
تسجيل مستخدم جديدDAPPLE: A Pipelined Data Parallel Approach for Training Large Models
55
0
0.0
(
0
)
اسأل ChatGPT حول البحث
ﻻ يوجد ملخص باللغة العربية
It is a challenging task to train large DNN models on sophisticated GPU platforms with diversified interconnect capabilities. Recently, pipelined training has been proposed as an effective approach for improving device utilization. However, there are still several tricky issues to address: improving computing efficiency while ensuring convergence, and reducing memory usage without incurring additional computing costs. We propose DAPPLE, a synchronous training framework which combines data parallelism and pipeline parallelism for large DNN models. It features a novel parallelization strategy planner to solve the partition and placement problems, and explores the optimal hybrid strategy of data and pipeline parallelism. We also propose a new runtime scheduling algorithm to reduce device memory usage, which is orthogonal to re-computation approach and does not come at the expense of training throughput. Experiments show that DAPPLE planner consistently outperforms strategies generated by PipeDreams planner by up to 3.23x under synchronous training scenarios, and DAPPLE runtime outperforms GPipe by 1.6x speedup of training throughput and reduces the memory consumption of 12% at the same time.
قيم البحث
اقرأ أيضاً
This paper presents the design, implementation, and evaluation of the PyTorch distributed data parallel module. PyTorch is a widely-adopted scientific computing package used in deep learning research and applications. Recent advances in deep learning
argue for the value of large datasets and large models, which necessitates the ability to scale out model training to more computational resources. Data parallelism has emerged as a popular solution for distributed training thanks to its straightforward principle and broad applicability. In general, the technique of distributed data parallelism replicates the model on every computational resource to generate gradients independently and then communicates those gradients at each iteration to keep model replicas consistent. Despite the conceptual simplicity of the technique, the subtle dependencies between computation and communication make it non-trivial to optimize the distributed training efficiency. As of v1.5, PyTorch natively provides several techniques to accelerate distributed data parallel, including bucketing gradients, overlapping computation with communication, and skipping gradient synchronization. Evaluations show that, when configured appropriately, the PyTorch distributed data parallel module attains near-linear scalability using 256 GPUs.
To achieve reliability in distributed storage systems, data has usually been replicated across different nodes. However the increasing volume of data to be stored has motivated the introduction of erasure codes, a storage efficient alternative to rep
lication, particularly suited for archival in data centers, where old datasets (rarely accessed) can be erasure encoded, while replicas are maintained only for the latest data. Many recent works consider the design of new storage-centric erasure codes for improved repairability. In contrast, this paper addresses the migration from replication to encoding: traditionally erasure coding is an atomic operation in that a single node with the whole object encodes and uploads all the encoded pieces. Although large datasets can be concurrently archived by distributing individual object encodings among different nodes, the network and computing capacity of individual nodes constrain the archival process due to such atomicity. We propose a new pipelined coding strategy that distributes the network and computing load of single-object encodings among different nodes, which also speeds up multiple object archival. We further present RapidRAID codes, an explicit family of pipelined erasure codes which provides fast archival without compromising either data reliability or storage overheads. Finally, we provide a real implementation of RapidRAID codes and benchmark its performance using both a cluster of 50 nodes and a set of Amazon EC2 instances. Experiments show that RapidRAID codes reduce a single objects coding time by up to 90%, while when multiple objects are encoded concurrently, the reduction is up to 20%.
It has become common to publish large (billion parameter) language models that have been trained on private datasets. This paper demonstrates that in such settings, an adversary can perform a training data extraction attack to recover individual trai
ning examples by querying the language model. We demonstrate our attack on GPT-2, a language model trained on scrapes of the public Internet, and are able to extract hundreds of verbatim text sequences from the models training data. These extracted examples include (public) personally identifiable information (names, phone numbers, and email addresses), IRC conversations, code, and 128-bit UUIDs. Our attack is possible even though each of the above sequences are included in just one document in the training data. We comprehensively evaluate our extraction attack to understand the factors that contribute to its success. Worryingly, we find that larger models are more vulnerable than smaller models. We conclude by drawing lessons and discussing possible safeguards for training large language models.
Data-intensive applications are becoming commonplace in all science disciplines. They are comprised of a rich set of sub-domains such as data engineering, deep learning, and machine learning. These applications are built around efficient data abstrac
tions and operators that suit the applications of different domains. Often lack of a clear definition of data structures and operators in the field has led to other implementations that do not work well together. The HPTMT architecture that we proposed recently, identifies a set of data structures, operators, and an execution model for creating rich data applications that links all aspects of data engineering and data science together efficiently. This paper elaborates and illustrates this architecture using an end-to-end application with deep learning and data engineering parts working together.
Recently, mT5 - a massively multilingual version of T5 - leveraged a unified text-to-text format to attain state-of-the-art results on a wide variety of multilingual NLP tasks. In this paper, we investigate the impact of incorporating parallel data i
nto mT5 pre-training. We find that multi-tasking language modeling with objectives such as machine translation during pre-training is a straightforward way to improve performance on downstream multilingual and cross-lingual tasks. However, the gains start to diminish as the model capacity increases, suggesting that parallel data might not be as essential for larger models. At the same time, even at larger model sizes, we find that pre-training with parallel data still provides benefits in the limited labelled data regime.
سجل دخول لتتمكن من نشر تعليقات
التعليقات
جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
