NOTE: Oobabooga supports kobold api and llama natively now.
vvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
Hoperator:
Hoperator is a primitive server which emulates enough of Kobold API to connect Kobold Client(UI1) to Oobabooga.
Usage:
- Start Oobabooga in no-stream mode
$ cd text-generation-webui
$ python server.py --model llama-13b --load-in-4bit --no-stream
NOTE: --no-stream is important. Hoperator relies on it.
- Start Hoperator
$ cd hoperator
# One of:
$ python kai_listener.py --generator oobabooga
$ python kai_listener.py --generator llama.cpp
- Start kobold AI client
$ cd KoboldAI-Client
$ ./play.sh
Select AI/Online Services/KoboldAI enter http://localhost:11111
as API server
NOTE: "http://" is important. "localhost:11111" will not work.
- Enjoy
CAVEATS(aka WORKSFORME/WONTFIX):
- API introduces the model as
hoperator/oobabooga
. Kobold has no idea what tokenizer to use and use something default. - Not all parameters are passed from the KAI to Oobabooga
- If Oobabooga runs out of memory or encounter any other error, it will not be reported to Kobold. Change settings and restart KAI client.
- There is no way to change ports for hoperator or Oobabooga other than through editing the source code.