Multimodal Large Language Models (MLLMs) have achieved remarkable advances by integrating text, image, and audio understanding within a unified architecture. However, existing distributed training frameworks remain fundamentally data-blind: they parallelize computation without accounting for variations in input data characteristics. This data unawareness leads to severe computation skew across stages and microbatches, where heterogeneous multimodal inputs incur different processing costs. Consequently, GPU resources are unevenly utilized, synchronization delays accumulate, and overall training efficiency degrades. To address this limitation, we present DFLOP, a data-driven framework for multimodal LLM training pipeline optimization. DFLOP continuously profiles runtime behavior to capture data-induced computation variance and employs predictive scheduling to balance workloads across stages and microbatches. By coupling data characteristics with execution planning, DFLOP substantially improves GPU utilization and throughput. Extensive experiments on large-scale multimodal benchmarks show that DFLOP achieves up to 3.6x faster training compared to state-of-the-art distributed training frameworks.
@article{arxiv.2603.25120,
title = {DFLOP: A Data-driven Framework for Multimodal LLM Training Pipeline Optimization},
author = {Hyeonjun An and Sihyun Kim and Chaerim Lim and Hyunjoon Kim and Rathijit Sen and Sangmin Jung and Hyeonsoo Lee and Dongwook Kim and Takki Yu and Jinkyu Jeong and Youngsok Kim and Kwanghyun Park},
journal= {arXiv preprint arXiv:2603.25120},
year = {2026}
}
Comments
Accepted to Proceedings of the ACM on Management of Data (SIGMOD 2026)