Feat: Support vLLM with OpenAI compatible chat API
samos123 opened this issue · comments
Currently vLLM will throw the following error when Verba send a chat completion request:
✘ WebSocket Error: Invalid response object from API:
'{"object":"error","message":"Conversation roles must alternate
user/assistant/user/assistant/...","type":"BadRequestError","param":null,"code":400}'
(HTTP response code was 400)
I believe this is due to Verba not correctly alternating between user/assistant/user as the error message indicates.
Steps to reproduce:
- Deploy vLLM with openai compatible endpoint and mistral v2 7b
- Set
OPENAI_BASE_URL=http://vlm-end-point/v1
Maybe we should fix it on the vLLM side by doing something like this: vllm-project/vllm#2112 (comment)
Closing this since I was able to fix it by using a custom chat template for Mistral with vLLM