MoE-LLaVA

• Supported higher resolution input using google/siglip-so400m-patch14-384 as the vision encoder for a more detailed visual understanding.
• Changed capacity_factor to 1.5 to support stronger MoE-LLaVA.
• Added the results of MME benchmark and evaluation pipeline.
• Improved docs.
• Fixed typos.

We hope that community researchers can pay attention to the fact that large vision-language models can also be sparsified and even perform better.