Very slow performance of mtmd model on Android devices

[kimminsu38oo]

Hello,

I tested the LLaVA 1.5 model on my mobile device (Galaxy S24 Ultra), and unfortunately the performance was extremely slow, to the point where it is barely usable.
https://github.com/ggml-org/llama.cpp/blob/master/docs/multimodal/llava.md#llava-15

Below are the profiling results from my test:

llama_perf_context_print:        load time =     793.34 ms
llama_perf_context_print: prompt eval time =   59422.18 ms /   634 tokens (   93.73 ms per token,    10.67 tokens per second)
llama_perf_context_print:        eval time =     173.36 ms /     1 runs   (  173.36 ms per token,     5.77 tokens per second)
llama_perf_context_print:       total time =   60802.96 ms /   635 tokens

image slice encoded in 11742 ms
decoding image batch 1/1, n_tokens_batch = 576
image decoded (batch 1/1) in 43292 ms

image decoding stage alone takes over 40 seconds..

I would like to ask:

1. Does llama.cpp actively support development for accelerating mtmd inference on mobile devices, or is this currently not a major focus?
2. Are there any plans to add features or techniques to accelerate mtmd inference, such as visual token pruning?

Thank you very much for your work on llama.cpp, and I would really appreciate any insights or guidance on this issue.