Fix optimum-cli command for VLM example in README

[helena-intel]

With the existing command users get an error: Channel size 4304 should be divisible by size of group 128.

[helena-intel]

@AlexKoff88 @nikita-savelyevv on Xeon the method with --dataset, both with and without num-samples, gives an empty result with the README example. With --group-size 16 I do get a good result. Also tried to test on laptop yesterday, but there the model export did not work (unrelated to this).

[ilya-lavrenov]

build_jenkins

[helena-intel]

For now the commands suggested by Alexander and Nikita don't work with GenAI (see comments and linked issue). There's a ticket about that. Can anyone suggest another model that works well? Smaller is better. Otherwise I'll remove the INT4 suggestion for now, until the issue is fixed in GenAI.

[nikita-savelyevv]

For now the commands suggested by Alexander and Nikita don't work with GenAI (see comments and linked issue). There's a ticket about that. Can anyone suggest another model that works well? Smaller is better. Otherwise I'll remove the INT4 suggestion for now, until the issue is fixed in GenAI.

OpenGVLab/InternVL2-1B is quite small and works with VLMPipeline. But I would recommend to wait for the fix huggingface/optimum-intel#1058 to be merged.

[helena-intel]

OpenGVLab/InternVL2-1B is quite small and works with VLMPipeline. But I would recommend to wait for the fix huggingface/optimum-intel#1058 to be merged.

Thank you! That's a much smaller model, much more user friendly, especially on AI PC. I tested and it works well out of the box with GenAI with the example from the README (after installing einops and timm). I'll update the PR with this. If someone disagrees, let me know.

The fix in optimum-intel is not enough for this because then exporting will work, but inference with GenAI doesn't. When that is fixed in an upcoming GenAI release we could change back to MiniCPM - though I see no downside for having smaller models in README examples.

[helena-intel]

Changed README to use OpenGVLab/InternVL2-1B with INT4. With FP16/FP32 model export, inference with the README example is empty. I added it to the ticket about empty output with MiniCPM. For now, this model with INT4 works, is quick to download and convert, and it's just for an inference example in the README, INT4 should be fine.

[ilya-lavrenov]

build_jenkins