Repository navigation
MSYS2/MinGW64 Text generation continues even when reverse prompts are specified. #941
Description
llama.cpp often continues to generate tokens even I specify the necessary reverse prompts: '### Human:' or 'Character:'
I am running in interactive mode, sometimes I get back the control, but often it takes one or more turns where llama.cpp skips to detect that the reverse prompt strings are in the token stream and hand back control.
Impact: bad chat experience.
Environment: MSYS2/Mingw64. Latest llama.cpp version from master.
Interestingly I don't experience the same problem if I compile with VS2019.
Model used: Vicuna 13B:
ggml-vicuna-13B-1.1-q4_0.bin
Command line:
./main.exe -i -f prompts/chat-with-samantha.txt -m ggml-vicuna-13B-1.1-q4_0.bin -r 'You:' -r '### Human:' -r 'Stan:' --color -t 8 --keep -1 -c 2048 -n -1
Tried specifying the reverse prompt both with single and double quotes - same result.
Activity
[-]Text generation continues even when reverse prompts are specified.[/-][+]MSYS2/MinGW64 Text generation continues even when reverse prompts are specified.[/+]taka-tuos commentedon Apr 15, 2023
I ran into the same problem, but using '###' for the reverse prompt stopped the output correctly. However, '### ' (adding a space at the end) did not stop it, so I think there is a high possibility that there is a problem with handling spaces.
batmanonline commentedon Apr 15, 2023
Possible duplicate of issues #990 #941
github-actions commentedon Apr 11, 2024
This issue was closed because it has been inactive for 14 days since being marked as stale.
DeepSeek: enable option to merge Q and K tensors (ggml-org#941)
server: a handoff can record its cells as an in-place checkpoint, so …
kv-share: a backend that cannot export its KV degrades, instead of ki…