The model I'm using seems dumber after updating

[Alice-Cheshire]

Describe the Issue
I updated to 1.118.1 a few days ago, and the model I've been using in SillyTavern seems to have become far dumber. Like to the point where a character will tell my character something, I'll acknowledge it, and then that character reacts with surprise that I know what they literally just told me. This was never an issue before, and it's happening with character cards I've used on numerous occasions. I've changed absolutely nothing with SillyTavern or with the configuration I'm using in KoboldCPP. I only updated the program.

Additional Information:
KoboldCPP: No Cuda build, version 1.118.1
GPU: AMD RX 6650 XT (Using the Vulkan driver for KoboldCPP)
CPU: Intel 10600K
OS: 64-bit Windows 10 Pro build 19045.5247
Model: Q6 RP Spectrum 24B

[henk717]

We need to know which version it last worked fine on. Upstrean llamacpp does a lot of changes and the other user that reported this refused to test thid so we have no ability to track down the regression without knowing at what version this starts.

[Alice-Cheshire]

Looks like the previous version that I updated from was 1.116. (No third number.) Thankfully, I always keep the previous version around when I update for when there's issues like this.

[henk717]

That is a lot closer than the 1.111 from the last report.
Does it also happrn on 1.117?

[Alice-Cheshire]

I haven't tested it thoroughly yet, but a preliminary test looks like it doesn't have the issue on 1.117.1. I branched one of the chats in SillyTavern at the exact moment that it went completely stupid on 1.118.1, and after a few regenerations, it has yet to act the way it did on 1.118.

Edit: Just for clarification, the point where I branched it was quite consistent about acting dumb in 1.118.1. So far after six regenerations, it hasn't had the same issue even a single time.