Timezone: »

1-bit Adam: Communication Efficient Large-Scale Training with Adam's Convergence Speed
Hanlin Tang · Shaoduo Gan · Ammar Ahmad Awan · Samyam Rajbhandari · Conglong Li · Xiangru Lian · Ji Liu · Ce Zhang · Yuxiong He

Tue Jul 20 07:35 PM -- 07:40 PM (PDT) @

Scalable training of large models (like BERT and GPT-3) requires careful optimization rooted in model design, architecture, and system capabilities. From a system standpoint, communication has become a major bottleneck, especially on commodity systems with standard TCP interconnects that offer limited network bandwidth. Communication compression is an important technique to reduce training time on such systems. One of the most effective ways to compress communication is via error compensation compression, which offers robust convergence speed, even under 1-bit compression. However, state-of-the-art error compensation techniques only work with basic optimizers like SGD and momentum SGD, which are linearly dependent on the gradients. They do not work with non-linear gradient-based optimizers like Adam, which offer state-of-the-art convergence efficiency and accuracy for models like BERT. In this paper, we propose 1-bit Adam that reduces the communication volume by up to 5x, offers much better scalability, and provides the same convergence speed as uncompressed Adam. Our key finding is that Adam's variance becomes stable (after a warmup phase) and can be used as a fixed precondition for the rest of the training (compression phase). We performed experiments on up to 256 GPUs and show that 1-bit Adam enables up to 3.3x higher throughput for BERT-Large pre-training and up to 2.9x higher throughput for SQuAD fine-tuning. In addition, we provide theoretical analysis for 1-bit Adam.

Author Information

Hanlin Tang (University of Rochester)
Shaoduo Gan (ETH Zurich)
Ammar Ahmad Awan (Microsoft)
Samyam Rajbhandari (Microsoft)
Conglong Li (Microsoft)
Xiangru Lian (Kwai Inc.)
Ji Liu (Kwai Seattle AI lab, University of Rochester)

Ji Liu is an Assistant Professor in Computer Science, Electrical and Computer Engineering, and Goergen Institute for Data Science at University of Rochester (UR). He received his Ph.D. in Computer Science from University of Wisconsin-Madison. His research interests focus on distributed optimization and machine learning. He also has rich experiences in various data analytics applications in healthcare, bioinformatics, social network, computer vision, etc. His recent research focus is on asynchronous parallel optimization, sparse learning (compressed sensing) theory and algorithm, structural model estimation, online learning, abnormal event detection, feature / pattern extraction, etc. He published more than 40 papers in top CS journals and conferences including JMLR, SIOPT, TPAMI, TIP, TKDD, NIPS, ICML, UAI, SIGKDD, ICCV, CVPR, ECCV, AAAI, IJCAI, ACM MM, etc. He won the award of Best Paper honorable mention at SIGKDD 2010 and the award of Best Student Paper award at UAI 2015.

Ce Zhang (ETH Zurich)
Yuxiong He (Microsoft)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors