Skip to main content

ADR-0103: vLLM and LiteLLM roles, LiteLLM pinned to 1.106.5

Context​

Two components are often confused: the inference engine and the gateway.

Decision​

vLLM is the local inference engine, LiteLLM the unified gateway. LiteLLM is pinned to v1.106.5 by digest, same version for the proxy image and the optional Python extra eai[litellm].

Consequences​

A LiteLLM upgrade changes catalogue, image and extra together.

Evidence​

  • RULES.md (section 2, point 3)
  • .navin/continuity/DECISIONS.md (LiteLLM v1.106.5)
  • navin-manager/catalog/services/litellm.yaml
  • services/control-plane/tests/test_platform_litellm_version.py
  • services/control-plane/src/eai/integrations/vllm.py