add InternVL2-8B video model

[chchenhui]

This PR contributes InternVL2 video models. InternVL2 1B, 2B and 8B has passed the local test. For 4B model, there exists a bug in position ids of Phi 3 in the source code modeling_phi3.py when transformers package is the latest. See OpenGVLab/InternVL#405 for more details.

[jdf-prog]

Hi, in the get_info() function of the adapter, can you change it to video or image? We don't support other keys like raw_video. Also, it will be great if you can also include some information about the model in the get_info() function. After fixing these, I think you code should be merged.

[jdf-prog]

Hi @chchenhui , I saw there is a modification in unit_test.py. This is not allowed. And what's the reason of changing the param['prompt'] into param['internvl2_prompt']? This params is the interface we have previously defined.
A few things to do:

• you need to change the param['internvl2_prompt'] back to param['prompt'].
• the 'video in the params is the encoded video, you should decode it to get the local path. please see [model_dummy_video.py`](https://github.com/WildVision-AI/LMM-Engines/blob/main/lmm_engines/huggingface/model/dummy_video_model.py#L80) for example.

[chchenhui]

Hi @jdf-prog , previously the generated content was very short when using the default prompt text. That's why I tried other prompt text presented in internVL2 demo. I have changed to the default prompt now.

[jdf-prog]

Hi @jdf-prog , previously the generated content was very short when using the default prompt text. That's why I tried other prompt text presented in internVL2 demo. I have changed to the default prompt now.

I see. No worries, just to remind. Feel free to ping me once you finished the debuging. And thanks for the time for contribution!

[chchenhui]

hi @jdf-prog , please check the updates. The code has passed local testing and I have resolved the conflicts with the base branch.

[jdf-prog]

Thanks for the contribution, the code has already been merged.