Support LLaVA-UHD

[choyakawa]

https://github.com/thunlp/LLaVA-UHD

This method is seemingly on par with or better than LLaVA 1.6 Next, however they opensourced the training code for reproduction.

LLM analysis from Gemini 1.5 pro:

Feature	LLaVA-UHD-13B	LLaVA-NeXT-7B	LLaVA-NeXT-13B	LLaVA-NeXT-34B	LLaVA 1.5-13B
VQAv2	81.7	81.8 (Vicuna) / 82.2 (Mistral)	82.8	83.7	80
GQA	65.2	64.2 (Vicuna) / 64.8 (Mistral)	65.4	67.1	63.3
TextVQA	67.7	64.9 (Vicuna) / 65.7 (Mistral)	67.1	69.5	61.3
ScienceQA	72	70.1 (Vicuna) / 72.8 (Mistral)	73.6	81.8	71.6
VizWiz	56.1	57.6 (Vicuna) / 60.0 (Mistral)	60.5	63.8	53.6
MMU (val)	36.4	35.8 (Vicuna) / 35.3 (Mistral)	36.2	51.1	36.4
MMU (test)	33.6	-	-	44.7	33.6
MME	1535	1519 (Vicuna) / 1498 (Mistral)	1575	1631	1531
POPE	89.1	86.5 (Vicuna) / 86.7 (Mistral)	86.2	87.7	85.9

Observations:

• LLaVA-UHD generally performs better than LLaVA 1.5 across all metrics.
• LLaVA-NeXT series shows comparable performance to LLaVA-UHD on most tasks, with slight variations depending on the specific model (Vicuna or Mistral).
• LLaVA-NeXT-34B stands out with significantly higher performance on ScienceQA and MMU tasks.

Originally posted by @choyakawa in thunlp/LLaVA-UHD#1 (comment)

[choyakawa]

Moreover, the model can be efficiently trained in academic settings, within 23 hours on 8 A100 GPUs (vs. 26 hours of LLaVA-1.5).

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.