Multimodal Large Language Models (MLLMs) are becoming an important workload for large-scale AI training. However, training MLLMs efficiently is more challenging than training decoder-only LLMs. In practice, even with more powerful hardware and larger training clusters, MLLM training can still suffer from lower MFU and higher engineering cost.
A key reason is that MLLMs are not just larger language models. They usually contain heterogeneous model components, such as modality encoders and LLM backbones. These components have very different computation patterns and memory footprints. As a result, a training strategy that works well for homogeneous LLM layers may not directly fit the full MLLM structure.
Another challenge comes from the input data itself. Compared with text-only training, multimodal training needs to process images, videos, and other modalities. These inputs can be much heavier and more irregular than text. For example, a high-resolution image may generate a large number of visual tokens, while a video may contain many such frames. At the same time, images may have different resolutions, and videos may have different lengths. This makes the workload vary significantly across samples and microbatches.
These two properties introduce new difficulties for pipeline-based training. Pipeline parallelism relies on relatively balanced and predictable stage execution. However, in MLLM training, heterogeneous model components and dynamic multimodal inputs can easily break this assumption, leading to pipeline imbalance, idle time, and reduced training efficiency.
This issue aims to highlight the need for better framework-level support for MLLM training workloads. In particular, the framework should consider the following challenges:
Heterogeneous model components with different compute and memory requirements.
Dynamic multimodal inputs with varying resolutions, token counts, and sequence lengths.
Pipeline imbalance caused by both model-side heterogeneity and data-side workload variation.
Improving support for these challenges is important for making large-scale MLLM training more efficient, predictable, and easier to optimize in production environments.
Multimodal Large Language Models (MLLMs) are becoming an important workload for large-scale AI training. However, training MLLMs efficiently is more challenging than training decoder-only LLMs. In practice, even with more powerful hardware and larger training clusters, MLLM training can still suffer from lower MFU and higher engineering cost.
A key reason is that MLLMs are not just larger language models. They usually contain heterogeneous model components, such as modality encoders and LLM backbones. These components have very different computation patterns and memory footprints. As a result, a training strategy that works well for homogeneous LLM layers may not directly fit the full MLLM structure.
Another challenge comes from the input data itself. Compared with text-only training, multimodal training needs to process images, videos, and other modalities. These inputs can be much heavier and more irregular than text. For example, a high-resolution image may generate a large number of visual tokens, while a video may contain many such frames. At the same time, images may have different resolutions, and videos may have different lengths. This makes the workload vary significantly across samples and microbatches.
These two properties introduce new difficulties for pipeline-based training. Pipeline parallelism relies on relatively balanced and predictable stage execution. However, in MLLM training, heterogeneous model components and dynamic multimodal inputs can easily break this assumption, leading to pipeline imbalance, idle time, and reduced training efficiency.
This issue aims to highlight the need for better framework-level support for MLLM training workloads. In particular, the framework should consider the following challenges:
Improving support for these challenges is important for making large-scale MLLM training more efficient, predictable, and easier to optimize in production environments.