Min P generation parameter

[LawrenceGrigoryan]

Feature request

Hey! I've noticed that min_p parameter is now available in transformers but haven't found it in TGI. Are you planning to add it in the nearest future?

The parameter itself is very well described here

I have implemented it in TGI for myself locally and I would be glad to contribute by submitting a PR :)

[avacaondata]

How have you implemented this locally? I'm seeking to do the same , would appreciate some guidance :) @LawrenceGrigoryan

[LawrenceGrigoryan]

@avacaondata
so basically you need to add MinPLogitsWarper to /server/text_generation_server/utils/logits_process.py and the min_p parameter to NextTokenChooser class in /server/text_generation_server/utils/tokens.py

Then go over a ton of files and add the min_p parameter wherever needed. Here I would suggest you searching for one of the existing generation params like top_p and adding the missing lines of code for the new parameter min_p

After you done that, just build the image with the default Dockerfile and use it as you usually do

If you have any further questions, feel free to ask :)