Context shift constantly failing?

#3
by Hardeh - opened

Is that only me, or with Q8_0, context shift randomly failing, causing full reprocess of context? Looks like i have this issue only with this model, other nemo-based models are fine. Is there some problem with tokenizer maybe?
I'm using koboldcpp_rocm, and tried Alpaca, Mistral and ChatML format - it's still the same. It can be fine for a few swipes, or even for a few dozen swipes, but then it fails randomly and reprocess everything again.

This model is composed of 3 fine tuned models. It is possible one or more is causing the issue and/or the way they were merged.
This model is design for mistral instruct template ; however it should also work with alpaca, and may work ok with chatml.

The other option: disable context shift.
"Context shift" is known to cause issues.

I see. Well, that's a shame. I have to use context shift anyway, because my GPU can hold up to 32k, and some chats are longer. Thanks anyway.

Hardeh changed discussion status to closed

Sign up or log in to comment