Bug: Gemma2 Context switching forgets original input #8251
Labels
bug-unconfirmed
medium severity
Used to report medium severity bugs in llama.cpp (e.g. Malfunctioning Features but still useable)
stale
What happened?
If I have a prompt like the following "<start_of_turn>user\nProductGroup: Anvil<end_of_turn><start_of_turn>user\nCan you give me the 25 most important characteristics for the previous named Productgroup? Respond in the following Json Format : [{'Characteristic':string,'Explanation':string,'ExampleValues':[string]}<end_of_turn><start_of_turn>model"
Then it starts out well, but if you set the context low (IE 512) then it starts breaking up output after context switching.
It seems like it has forgotten the Initial ProductGroup, and it just continuous outputting based on the examples from the JSON which don't mention the specific product group, so it just outputs for random products.
I don't know if this is just how it is supposed to work, or that it would be possible to add the original prompt in front when context switching.
I basically noticed this because of the server with parallel which splits the context size so you run into this much quicker, the quick-fix version is to simply up the context window
Name and Version
./llama-cli --version
version: 3281 (023b880)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
What operating system are you seeing the problem on?
Linux
Relevant log output
No response
The text was updated successfully, but these errors were encountered: