mirror of
https://github.com/BerriAI/litellm.git
synced 2026-09-07 08:26:10 +00:00
84dd96b2e4
classifier_llm_config gains reasoning_effort, sent on the classifier's own acompletion and mirrored into the logged request body. Classification runs on every request and is short, so a reasoning classifier model is usually worth holding at a cheap level. Setting the effort on the deployment instead moves it for every request that deployment serves, which is wrong whenever the classifier model also serves normal traffic. The level is filtered through the Router's existing target capability filter before either consumer is built, so a level the classifier model refuses reaches neither the provider call nor the spend-log request body. That filter is now public as params_the_target_accepts, since the classifier path is a third caller alongside the two tier-overlay sites, and it already owns the fail-open rules and the guard against capability lookups that authenticate. The dashboard carries the field through both payload builders, since each rebuilds classifier_llm_config from named keys and would otherwise wipe a value set in config.yaml on the next edit. |
||
|---|---|---|
| .. | ||
| litellm-dashboard | ||
| Dockerfile | ||
| nginx.conf | ||