Katwinkl3/support other providers

[katwinkl3]

Key Info

• Implementation plan: link
• Priority: normal

What changed?

• added deepseek
• added model prefix for gemini and deepseek models (refered to https://docs.litellm.ai/docs/providers/deepseek)
• changed use_logprobs's logic

What do you want the reviewer(s) to focus on?

• For use_logprobs' change, im not sure if MODELS_WITH_LOGPROBS list is comprehensive, and if this logic should be added anywhere else in the pipeline.
• Also, I noticed that there are 2 ways to include the model param, 1 in completion_params (from openai_args) which is used in ReferenceCompletionTemplate and 2 in ConfigInput (which seems like it should be used everywhere else). If a model is only added for 1, the other pipeline process will use the default gpt model. Is this the expected behavior?
• I also tried using different models for 1 and 2 (both gemini), but the gpt default model was still used somewhere in the pipeline, but i havent looked into it

---

Checklist

• Did you link the GitHub issue?
• Did you follow deployment steps or bump the version if needed?
• Did you add/update tests?
• What QA did you do?
  • Tested on my own jupyter notebook

[kelsey-wong]

Also, I noticed that there are 2 ways to include the model param, 1 in completion_params (from openai_args) which is used in ReferenceCompletionTemplate and 2 in ConfigInput (which seems like it should be used everywhere else). If a model is only added for 1, the other pipeline process will use the default gpt model. Is this the expected behavior?

passing the model through ConfigInput is leftover from before I refactored to the openai params spec for the inference function, so I'll remove that. the model should only be set through the params, not the config input