OCR slows down with ollama-vision models

[ksof1]

Describe the bug
I use Ollama running minicpm to OCR image-based subtitles into text-based via SubtitleEdit 4.0.12. It starts with taking 0.5s to read the image and output text gradually slowing down. After ~40minutes of running it slowed down to 12s to process a single image. Pausing the OCR and restarting ollama-server reset the slowdown cycle and the process of slowdown started again from 0.5s

GPU Utilization also drops from initial ~80% to ~8% with momentary spikes to ~20% when the output is created.

How to reproduce
Steps to reproduce the error:

1. Load the model (minicpm-v:8b)
2. Input image-based subtitles into Subtitle edit
3. Select OCR method Ollama-vision, select minicpm-v:8b
4. Start OCR

Screenshots

Using latest nightly build of portable ipex-llm-ollama

[ksof1]

Appears to be fixed in latest nightly build.