examples : support MiniCPM-V-2

[guoQiNing]

Hi, this is an attempt to add support for MiniCPM-V-2.

This code is based on the llama.cpp repository at commit 7841fc72, and PR #6919 is a helpful reference.

I also notice the recent updates about libmtmd and mtmd-cli, and I'll continue adapting this code for mtmd compatibility soon, to help its merge into the latest llama.cpp.

Thank you for your consideration.

[ngxson]

If you can provide me a concise list of differences between minicpm-v-2 and the original minicpm-v, I can help doing this faster

[ngxson]

But wait, are you sure you are adding support for a new model?

To remind, we already support minicpmv-2.5, 2.6 and minicpm-o, so maybe there is no more work needed

[guoQiNing]

But wait, are you sure you are adding support for a new model?
To remind, we already support minicpmv-2.5, 2.6 and minicpm-o, so maybe there is no more work needed

Yes, I've noted the existing support for recent MiniCPM-V series models in llama.cpp.

However, MiniCPM-V 2.0 is a capable 2B parameter model, while the subsequent ones (minicpmv-2.5, 2.6 and minicpm-o) are all around 8B. And Its visual model architecture also has significant differences.

When using MiniCPM-V 2.0, I noticed that the related PR #6919 has not been merged and has bug in GPU environments. This PR is an attempt to provide fix and update.

I will try to provide a concise list of differences between MiniCPM-V 2.0 and the original MiniCPM-V soon.

Thank you so much for your reply!