Inference Compute-Optimal Video Vision Language Models
This work addresses resource-efficient model configuration for video vision language tasks, providing practical guidance for selecting scaling factors under fixed compute budgets, but it is incremental as it builds on prior efficiency and performance optimizations.
The paper tackles the problem of optimally allocating inference compute across language model size, frame count, and visual tokens per frame in video vision language models, finding that task performance depends on these scaling factors and finetuning data size, with changes in data shifting the compute-optimal frontier.
This work investigates the optimal allocation of inference compute across three key scaling factors in video vision language models: language model size, frame count, and the number of visual tokens per frame. While prior works typically focuses on optimizing model efficiency or improving performance without considering resource constraints, we instead identify optimal model configuration under fixed inference compute budgets. We conduct large-scale training sweeps and careful parametric modeling of task performance to identify the inference compute-optimal frontier. Our experiments reveal how task performance depends on scaling factors and finetuning data size, as well as how changes in data size shift the compute-optimal frontier. These findings translate to practical tips for selecting these scaling factors.