Describe the Issue
Recently I've gotten into making use of an mmproj for a Qwen 3.5 finetune I run. I had no problems with image analysis before, but once I updated KoboldCPP to 118.1, I have gotten two system hangs and one outright system crash while attempting to have it analyze an image. If it freezes, the entire system locks, forcing a hard reset.
It doesn't do that all the time with every image, but it does that some of the time, and there seems to be no pattern to what triggers it. I have had no issues before now with it and my system is usually rock solid (weeks of uptime), so something tells me something went buggy with the latest update in a way it did not like. I'm going to try rolling back to the previous version to see if it continues there and will report back.
It's definitely not running out of RAM either - I've got a GeForce 4080 Super, the quant I have leaves about 1 GB of VRAM free, and I'm actually offloading image processing onto the CPU to maximize token generation speed. Main PC itself has 64 GB of RAM.
Additional Information:
Win11, Ryzen 5900X, GeForce 4080 Super, KoboldCpp 118.1. Seemed to not happen on 117.1.
Describe the Issue
Recently I've gotten into making use of an mmproj for a Qwen 3.5 finetune I run. I had no problems with image analysis before, but once I updated KoboldCPP to 118.1, I have gotten two system hangs and one outright system crash while attempting to have it analyze an image. If it freezes, the entire system locks, forcing a hard reset.
It doesn't do that all the time with every image, but it does that some of the time, and there seems to be no pattern to what triggers it. I have had no issues before now with it and my system is usually rock solid (weeks of uptime), so something tells me something went buggy with the latest update in a way it did not like. I'm going to try rolling back to the previous version to see if it continues there and will report back.
It's definitely not running out of RAM either - I've got a GeForce 4080 Super, the quant I have leaves about 1 GB of VRAM free, and I'm actually offloading image processing onto the CPU to maximize token generation speed. Main PC itself has 64 GB of RAM.
Additional Information:
Win11, Ryzen 5900X, GeForce 4080 Super, KoboldCpp 118.1. Seemed to not happen on 117.1.