presence_penalty parameter results in Internal Server Error
in progress
M.R.
in progress
This post was marked as
planned
M.R.
Can you give more details, is it jsut one of the models are you using vLLM or openai endpoint?
C
Cysio
M.R. I'm using the https://api.parasail.io/v1 in the Text Completion mode through KoboldAI Lite (the OpenAI Compatible API option). I've pulled out the request that causes the issue, after removing presence_penalty it works:
{"max_tokens":200,"model":"parasail-deepseek-r1-0528","temperature":0.7,"top_p":0.6,"stop":["<|end▁of▁sentence|><|User|>","<|end▁of▁sentence|><|Assistant|>"],"presence_penalty":0,"prompt":"<|end▁of▁sentence|><|User|>test<|end▁of▁sentence|><|Assistant|>"}
R
R. K.
Hi Cysio, thank you for sharing the request details. I'm trying to identify these requests on our side. Would it be possible to send a couple more of these requests and provide us with these values?
- Timestamp of the requests
- User Agent you are using (if known)
- Name of the model
- temperature value
- top_p value
- presesence_penalty value
C
Cysio
R. K. Sample request at Fri, 13 Jun 2025 16:48:29 GMT
UA curl/8.14.1
Model parasail-deepseek-r1-0528
Temperature 0.7
top_p 0.6
presence_penalty 0
M.R.
Cysio: Thansk we are continuing to look into this
R
R. K.
Hi Cysio, thanks once more. I think I got the bug confirmed and applied a fix. Could you maybe retry to see if your request succeeds?
C
Cysio
R. K. Yes, it works both through curl and KoboldAI Lite
R
R. K.
Cysio perfect. Thanks for the quick reply!