Add Sglang docs

Prevent failing if tensors are of different shape

Fix multiple GPU inference.

[Fix] Add image_sizes for cli.py for LLaVA-v1.6

Pass image_size to model.generate() in cli.py