OmniBal: Towards Fast Instruction-Tuning for Vision-Language Models via Omniverse Computation Balance

TL;DR

OmniBal introduces a balanced training framework for vision-language models that optimizes data, model, and memory distribution, significantly accelerating training speed and improving efficiency across diverse models and datasets.

Contribution

This work presents a novel omniverse balanced training approach that rebalances computation in data, model, and memory aspects for vision-language models.

Findings

Achieves approximately 1.8× training speed-up compared to baseline.

Effectively balances computation load across devices in heterogeneous vision-language models.

Demonstrates generalizability across various models and datasets.

Abstract

Vision-language instruction-tuning models have recently achieved significant performance improvements. In this work, we discover that large-scale 3D parallel training on those models leads to an imbalanced computation load across different devices. The vision and language parts are inherently heterogeneous: their data distribution and model architecture differ significantly, which affects distributed training efficiency. To address this issue, we rebalance the computational load from data, model, and memory perspectives, achieving more balanced computation across devices. Specifically, for the data, instances are grouped into new balanced mini-batches within and across devices. A search-based method is employed for the model to achieve a more balanced partitioning. For memory optimization, we adaptively adjust the re-computation strategy for each partition to utilize the available memory fully. These three perspectives are not independent but are closely connected, forming an omniverse balanced training framework. Extensive experiments are conducted to validate the effectiveness of our method. Compared with the open-source training code of InternVL-Chat, training time is reduced greatly, achieving about 1.8$\times$ speed-up. Our method's efficacy and generalizability are further validated across various models and datasets. Codes will be released at https://github.com/ModelTC/OmniBal.