Vibeleaderboard
← All Intel
Intel / repo

litellm v1.108.2 — 1.108.2 (LTS)

Source
github.com
Author
litellm-release[bot]
Date

Context

LiteLLM is an open-source library and proxy server that lets applications call many AI model providers through one interface, and version 1.108.2 is a patch on its long-term support (LTS) line. LTS generally marks releases maintained with small fixes for operators who would rather not follow every feature release, and the notes say the `lts` Docker image tag always points to the newest of them. This patch contains 12 pull requests, 11 fixes and one behavior change, and according to the notes no action is needed to upgrade.

The fixes most likely to matter involve credentials and data exposure. One stops deployments under the `openai/` prefix that have an empty `api_key` from sending the value of `OPENAI_ADMIN_KEY` as their bearer , the credential attached to each request; the notes do not say which earlier versions were affected. Another stops audio transcription errors from echoing the provider's full reply, transcript included, to clients and into logs. A third fixes spend records being dropped, and valid keys being refused with a 401 error, while Postgres restarts or shuts down. The single behavior change concerns hosted tools, which use the Model Context Protocol for connecting models to tools: when a translated route cannot run one, LiteLLM now returns a clear 400 error, or drops the tool if `drop_params` is set. The remaining fixes are provider- and feature-specific, covering AWS Bedrock, Vertex AI and Gemini text-to-speech, chat completions, the Responses API and routing.

The notes also give a `cosign verify` command for confirming that an image was signed by the project before running it. It checks against the public key from commit `e82a3c6`, the replacement signing key LiteLLM moved to in 1.106.3. Asked about that rotation, a maintainer said it was not routine and that the team had reason to believe the old key was exposed when credentials were stolen in the October 21 compromise of the project's release pipeline. The same command applies to the `litellm-database` and `litellm-non_root` images and to the `docker.io/litellm/*` images.

Key takeaways · AI-distilled
  • The empty-key bug was a credential mix-up: openai/ deployments configured with an empty api_key sent the OPENAI_ADMIN_KEY value as the bearer token. Release 1.108.2 fixes it (#50218).
  • Audio transcription errors had been echoing the full provider reply, transcript included, to clients and to logs. This release stops that (#50325).
  • While Postgres restarted or shut down, LiteLLM dropped spend rows and returned 401 for valid keys. Both are fixed, so cost tracking and auth hold up through database restarts (#50132).
  • With turn_off_message_logging on, prompt-caching routing did not pin follow-up requests to one deployment. The patch restores that pinning (#50276).
  • Release images can be verified with cosign against the public key in the repository, and the lts image tag always points to the newest LTS patch.
Terms in this piece · Glossary
  • token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
  • MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
Recommended reads
Comments

Checking sign-in…

Loading comments…