ProTrain: Efficient LLM Training via Automatic Memory Management

Hanmei Yang, Jin Zhou, Yao Fu, Xiaoqun Wang, Ramine Roane, Hui Guan, Tongping Liu

Keywords: ML System, Memory Optimization, Data Parallelism, ZeRO, Gradient Checkpointing, Tensor Offloading

TL;DR: a high-throughput training framework for LLMs that intelligently balance memory usage and performance via automatic memory management

Abstract:

Training billion-scale large language models (LLMs) with just a few consumer-grade graphics cards is key to democratizing LLM access. However, existing frameworks often depend on manual tuning of memory management settings, leading to inefficient hardware utilization and suboptimal performance. This paper introduces ProTrain, a novel training system that automatically tailors memory management policies to the model architecture and underlying hardware resources, eliminating the need for manual intervention. ProTrain features (1) automated memory management that abstracts complex memory management strategies into a few tunable configuration parameters and searches for optimal parameter settings using cost models and (2) a runtime profiler that provides precise estimates of latency, memory usage, and I/O bandwidth to build high-fidelity cost models.
ProTrain does not change the training algorithm and thus does not compromise accuracy. Experiments show that ProTrain improves training throughput by 1.43$\times$ to 2.71$\times$ compared to the state-of-the-art training systems.

Primary Area: infrastructure, software libraries, hardware, systems, etc.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5203