Source: vLLM Structured Outputs Docs

vLLM documentation for grammar-constrained generation: the structured_outputs parameters (choice, regex, json, grammar, structural_tag), the xgrammar/guidance backends, and reasoning-parser interaction.

Contents
  • What: official documentation page
  • Author/publisher: vLLM project
  • Date: living document (latest); read 2026-11-14
  • Original: Structured Outputs

Summary

vLLM exposes constrained decoding on its OpenAI-compatible server and offline API. Requests carry a structured_outputs object (in extra_body online, StructuredOutputsParams offline) with exactly one constraint kind. Backends are pluggable and chosen automatically by default.

Key claims

  • Constraint kinds: choice (one of a list), regex, json (schema or Pydantic-derived), grammar (EBNF CFG), structural_tag (JSON schema applied inside tagged spans of otherwise free text). response_format with a JSON schema also works.
  • The old top-level guided_json, guided_regex, guided_choice, guided_grammar, guided_whitespace_pattern fields and guided_decoding_backend were removed in v0.12.0.
  • Backends: xgrammar and guidance (llguidance), selected by auto per request or forced via --structured-outputs-config.backend. Both use Rust-style regex syntax. lm-format-enforcer (Python re syntax) is deprecated from v0.33.0 and slated for removal in v0.35.0.
  • With --reasoning-parser (e.g. deepseek_r1) the constraint applies to the final content, and reasoning is returned separately; for some models constraint-in-reasoning must be enabled explicitly (enable_in_reasoning).
  • Grammars that reference boundary tokens like </think> must use token-level elements, because a text literal can match the same spelling with different token IDs.

Feeds

Constrained decoding

Suggest changes →