More information about that : Ollama support is still not fully implemented by Litellm. see this Issue.
When providing api_base key to completion() it is used to generate text with the right ollama server but some function might not get the api_base argument and fallback to localhost. So we can have that sort of log when setting the logger :

INFO:httpx:HTTP Request: POST http://XXX.XXX.XXX.XXX:11434/api/generate "HTTP/1.1 200 OK" (Right call to remote server)
00:26:08 - LiteLLM:INFO: cost_calculator.py:656 - selected model name for cost calculation: ollama/qwen3:8b
INFO:LiteLLM:selected model name for cost calculation: ollama/qwen3:8b
INFO:httpx:HTTP Request: POST http://localhost:11434/api/show "HTTP/1.1 200 OK"

As we can see there is still an http call to localhost, so the cost calculation (which is 0) is provided by the local ollama. I'm not sure if it affects other functionality.

Anyway the only way that it affects us is : if you have an ollama server locally and a remote ollama server, you might get an http request to your local ollama server. if we want to fix this on our side we could add

if api_base:
  os.environ["OLLAMA_API_BASE"] = api_base

In the LiteLLMModel init But might be better to just wait for a fix inside LiteLLM.

TL;DR: LiteLLM does not fully support Ollama, but the PR is still valid. 👋