Multimodal inference at native resolution with Qwen3-VL

[vapetrov]

I'm running commit e5fc02c and using the model unsloth/Qwen3-VL-8B-Instruct-Q6_K.
One of Qwen3-VL's key features is the ability to process images at their native resolution.
With ik_llama.cpp I get exactly 768 tokens regardless of the input image size.
With llama.cpp I get between 3,000 - 4,000 tokens depending on image size.
With transformers I get roughly 1,000 tokens/ megapixel.

Is there any way to get ik_llama to use the native image size, or at least increase the number of tokens to something similar to the upstream llama.cpp?

[ikawrakow]

IIRC, the ability to handle different resolutions got added to llama.cpp only very recently. We haven't come around to port these changes to ik_llama.cpp, but it will eventually happen.