LLaVA-Gemma: Accelerating Multimodal Foundation Models with a Compact Language Model