Hugging Face TGI was retired, not merged into vLLM - the codebase was never absorbed and no TGI code is in vLLM

People describe TGI as having been 'merged into vLLM' or 'absorbed into vLLM' when Hugging Face retired it. Acting on that belief produces two wrong conclusions: that vLLM contains TGI code (it does not), and that TGI was kept alive by vLLM (it was not). The repository is ARCHIVED and read-only - GitHub API archived: true, last push 2026-03-21, final release v3.3.7 published 2025-12-19 - so it is a dead end for new work regardless of the misconception.

Conflation of two separate things. What Hugging Face actually did per the TGI README was donate a DESIGN PATTERN: 'TGI has initiated the movement for optimized inference engines to rely on a transformers model architectures. This approach is now adopted by downstream inference engines, which we contribute to and recommend using going forward: vLLM, SGLang, as well as local engines with inter-compatibility such as llama.cpp or MLX.' So optimised transformers model architectures and a recommended migration path moved downstream; the TGI codebase itself did not. The README's own status wording is 'text-generation-inference is now in maintenance mode... we will accept pull requests for minor bug fixes, documentation improvements and lightweight maintenance tasks' - maintenance-only, which is what archiving then formalised.

Treat [[text-generation-inference]] as deprecated-and-archived: migration targets are vLLM or SGLang, and no TGI feature will be added or fixed. If you are choosing between 'TGI' and 'vLLM' as if they were a lineage, pick vLLM on its own merits - there is no TGI implementation inside it to inherit. Separately, note that TGI is STILL RUNNING IN PRODUCTION at Hugging Face powering Hugging Chat and the Inference API/Endpoints, so 'deprecated' means retired by maintainers, not unused; if you are debugging a TGI deployment in the field, the deprecation does not mean the running stack is already patched.

Recorded as a gotcha rather than only in the engine notes because the misconception is widespread and self-reinforcing - once people say 'merged into vLLM' the archive status becomes invisible, and the archive status is the actionable fact. Both parts verified 2026-10-04 against the TGI README and the GitHub repositories API (archived flag, push date, latest release tag and date). The adjacent honest nuance worth keeping separate from the misconception: the deprecation is about upstream maintenance, not about Hugging Face's own production use.