This is to bring consistency with apply_model_params_to_body_openai. Both now use a mapping dictionary then call and return apply_model_params_to_body directly.
We copy the params from from the original key to the new key, then delete it. This is to ensure Ollama only gets valid options.
(Add a comment as well)
1) Ollama natively supports frequency_penalty, so this is not necessary.
2) Repeat_penalty is being added to Open WebUI in PR #10016, allowing Ollama users to pick which penalty methods they want.
1) Ollama supports sending the system prompt as a parameter, not as an option. (See https://github.com/ollama/ollama/blob/main/docs/api.md#request-8) However, it is in the options dictionary and needs moved to the payload dictionary.
2) After moving the system parameter from ollama_options to ollama_payload, delete it from ollama_options. This is to prevent Ollama throwing a warning about invalid options.
1) max_tokens was being looked for in openai_payload, but is present in openai_payload['options'], so is never found.
2) After copying the value for max_tokens to num_predict, delete max_tokens from the dictionary. This is to prevent Ollama throwing a warning about invalid option (max_tokens)
1) This may be legacy code?
2) All three of these parameters, temperature, top_p and seed, are found in openai_payload["options"], not openai_payload. They do not need remapped any longer.
1) max_completion_tokens is being looked for in openai_payload, but would be located in openai_payload['options'], so is never found. (This applies to the prior two commits as well).
2) max_completion_tokens is not sent from the frontend, only max_tokens. It does not appear in AdvancedParams.svelte.
2b) Openai.py does use max_completion_tokens, but for o1,o3 models and converts it from max_tokens.
1) Ollama natively supports frequency_penalty.
2) repeat_penaltywas added to Open Webui in PR #10016 and is not merged to main yet at this time. Once both changes go live, Ollama users can freely choose between frequency/presence penalty, or repeat penalty, as they choose.
- Add Presence Penalty, which I believe most providers have.
- Add repeat_penalty, an Ollama / Llama.cpp only parameter
- Widen temperature range to 2 (supported by many providers and similar front ends)
- Widen Top K range (supported by many providers and similar front ends)
- Adjust description of freqency_penalty in line with OpenAI documentation
- Add presence_penalty with an OpenAI documentation based description
- Add repeat_penalty with an Ollama documentation based description
ferret99gt