LLAVA1.6 performance huge drop after export/import using ModelFile

[vai-minzhou]

What is the issue?

Using ModelFile to change the temperature of LLAVA 1.6 or importing LLAVA from a GGUF file significantly reduces performance. This issue likely arises because the import operation treats the LLAVA 1.6 model as LLAVA 1.5 or earlier, ignoring the image-splitting into four parts. Please address this bug.

OS

Linux

GPU

Nvidia

CPU

AMD

Ollama version

Any version