Source: vLLM docs: Online Serving
Overview of vLLM's HTTP server: the endpoint families it exposes and their caveats.
Contents
- What: official documentation (latest, read 2026-11-14)
- Author: vLLM project
- Date: 2026-11-14 (accessed)
- Original: vLLM docs: Online Serving
Summary
Index of what vllm serve <model> exposes: OpenAI-style routes, plus Cohere-style rerank/embed, an Anthropic Messages route, tokenizer utilities and pooling/scoring endpoints, with notes on chat templates and unsafe dev endpoints.
Key claims
- OpenAI-style routes:
/v1/completions,/v1/chat/completions,/v1/responses,/v1/embeddings,/v1/audio/transcriptionsand/v1/audio/translations,/v1/realtime. - Non-OpenAI routes:
/tokenize,/detokenize,/pooling,/score,/rerank(incl. Cohere-compatible/v2/rerank),/classify,/v2/embed, Anthropic-style/v1/messages. - Chat needs a chat template, from the tokenizer config or
--chat-template; without one, chat requests fail. Content format is auto-detected, overridable with--chat-template-content-format. - Completions ignores
suffix; Chat ignoresuser;parallel_tool_calls=truedoes not guarantee multiple calls. - Dev endpoints (
VLLM_SERVER_DEV_MODE=1) are not for production; runtime LoRA loading is for local development only.