DFLOP: A Data-driven Framework for Multimodal LLM Training Pipeline Optimization

TL;DR

DFLOP is a data-driven framework that optimizes multimodal LLM training by profiling data characteristics and balancing workloads, significantly improving GPU utilization and training speed.

Contribution

It introduces a novel data-aware scheduling approach for multimodal LLM training, addressing computation skew caused by heterogeneous input data.

Findings

DFLOP achieves up to 3.6x faster training compared to existing frameworks.

It effectively balances workloads across stages and microbatches.

The framework improves GPU utilization and overall training efficiency.

Abstract

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.