paper-with-me

홈 › Papers

PipeFisher: Efficient Training of Large Language Models Using Pipelining and Fisher Information Matrices

2022-11-25 · Kazuki Osawa, Shigang Li, Torsten Hoefler

Pipeline parallelism enables efficient training of Large Language Models (LLMs) on large-scale distributed accelerator clusters. Yet, pipeline bubbles during startup and tear-down reduce the utilization of accelerators. Although efficient pipeline schemes with micro-batching and bidirectional pipelines have been proposed to maximize utilization, a significant number of bubbles cannot be filled using synchronous forward and backward passes. To address this problem, we suggest that extra work be assigned to the bubbles to gain auxiliary benefits in LLM training. As an example in this direction, we propose PipeFisher, which assigns the work of K-FAC, a second-order optimization method based on the Fisher information matrix, to the bubbles to accelerate convergence. In Phase 1 pretraining of BERT-Base and -Large models, PipeFisher reduces the (simulated) training time to 50-75% compared to training with a first-order optimizer by greatly improving the accelerator utilization and benefiting from the improved convergence by K-FAC.

📄 PDF Abstract BibTeX arXiv:2211.14133

Code (1)

kazukiosawa/pipe-fisher 공식 구현 pytorch

Similar Papers 제목 키워드 기반

NestPipe: Large-Scale Recommendation Training on 1,500+ Accelerators via Nested Pipelining

2026-04-08 · Zhida Jiang, Zhaolong Xing, Huichao Chai, Tianxing Sun 외 arxiv

Modern recommendation models have increased to trillions of parameters. As cluster scales expand to O(1k), distributed training bottlenecks shift from computation and memory to data movement, especially lookup and commun…

SpikePipe: Accelerated Training of Spiking Neural Networks via Inter-Layer Pipelining and Multiprocessor Scheduling

2024-06-11 · Sai Sanjeet, Bibhu Datta Sahoo, Keshab K. Parhi

Spiking Neural Networks (SNNs) have gained popularity due to their high energy efficiency. Prior works have proposed various methods for training SNNs, including backpropagation-based methods. Training SNNs is computatio…

Scheduling

CTA-Pipelining: A Latency-Oriented Spatial Scaling Method for Multi-GPU Systems

2026-07-08 · Tingkai Liu, Muralidhar Andoorveedu, Sanjoy Das, Sanjay Patel 외 arxiv

The evolution of compute infrastructure has transformed multi-GPU systems into tightly integrated shared-memory structures. However, current software still mostly treats these coherent interconnects simply as high-speed …

Group-based Interleaved Pipeline Parallelism for Large-scale DNN Training

2021-09-29 · ICLR 2022 4 · Pengcheng Yang, XiaoMing Zhang, Wenpeng Zhang, Ming Yang 외

The recent trend of using large-scale deep neural networks (DNN) to boost performance has propelled the development of the parallel pipelining technique for efficient DNN training, which has resulted in the development o…

SkipPipe: Partial and Reordered Pipelining Framework for Training LLMs in Heterogeneous Networks

2025-02-27 · Nikolay Blagoev, Lydia Yiyu Chen, Oğuzhan Ersoy

Data and pipeline parallelism are ubiquitous for training of Large Language Models (LLM) on distributed nodes. Driven by the need for cost-effective training, recent work explores efficient communication arrangement for …

Scheduling