Added logit_bias support

[Vokturz]

This PR brings the logit_bias functionality, already present in llama.cpp, which allows users to adjust the likelihood of token occurrences in generated text.

For example, for the prompt "Once upon a " we have:

1. Without logit_bias

   curl -X POST http://localhost:11434/api/generate -d '{
      "model": "llama2:7b",
      "prompt": "Once upon a ",
      "stream": false,
      "options": {"seed": 0, "num_predict" : 10}
      }'
   {"model":"llama2:7b", ... , "response":"time, in a far-off land, there", ...}
   

2. With logit_bias removing "time" and "Time" words:

   curl -X POST http://localhost:11434/api/generate -d '{
      "model": "llama2:7b",
      "prompt": "Once upon a ",
      "stream": false,
      "options": {"seed": 0, "num_predict" : 10,
                  "logit_bias": [[2230,false],[931,false],[2481,false],[5974,false]]}
      }'
   {"model":"llama2:7b", ... , "response":"Once upon a...! What a great prompt!", ...}
   

3. With logit_bias removing "time" and "Time" words, and increasing "night" likelihood :

   curl -X POST http://localhost:11434/api/generate -d '{
      "model": "llama2:7b",
      "prompt": "Once upon a ",
      "stream": false,
      "options": {"seed": 0, "num_predict" : 10,
                  "logit_bias": [[2230,false],[931,false],[2481,false],[5974,false],[11147, 15]]}
      }'
   {"model":"llama2:7b", ... , "response":"night, in a land far away, there lived"", ...}
   

[nathanpbell]

Hi Voktruz, any chance of updating this for the latest main?

[K0IN]

any updates on this?