[Android] Support LLaVA and Phi-V

[davidlightmysterion]

This PR introduces the support of LLaVA and Phi-V on android device.

Not thoroughly tested, but it works on my device. (Android 14.0)

checkpoints:

• https://huggingface.co/davidlightmysterion/llava-1.5-7b-hf-q4f16_1-MLC
• https://huggingface.co/mlc-ai/Phi-3.5-vision-instruct-q4f16_1-MLC

[Mawriyo]

This is awesome work. @davidlightmysterion Can I recommend instead of hardcoding the image size using

std::vector<int64_t> new_shape = {1, 336, 336, 3};

Try something like this?

std::vector<int64_t> new_shape = {1, image_size, image_size, 3};

Looking at gemma 3 vision llm's they use different input image sizes. (896 x 896)

That way you no longer are limited to just these models for future use.

Remember this a recommendation and this is also just my first glance. I am going to dig deeper to see if this is actually what you would want to do or maybe you have a reason to hardcode.

Open to dialog for others in the community!

[davidlightmysterion]

This is awesome work. @davidlightmysterion Can I recommend instead of hardcoding the image size using

std::vector<int64_t> new_shape = {1, 336, 336, 3};

Try something like this?

std::vector<int64_t> new_shape = {1, image_size, image_size, 3};

Looking at gemma 3 vision llm's they use different input image sizes. (896 x 896)

That way you no longer are limited to just these models for future use.

Remember this a recommendation and this is also just my first glance. I am going to dig deeper to see if this is actually what you would want to do or maybe you have a reason to hardcode.

Open to dialog for others in the community!

Sure, thanks for the advice. I'll fix it later.