Skip to content

vllm.parser.engine

Streaming parser engine framework for tool call and reasoning extraction.

Instead of hand-rolling a parser for every model's tool-call / reasoning format, each format is declared as a ParserEngineConfig (terminals, states, and transitions) and a shared incremental engine handles streaming, ambiguity buffering, token-ID mapping, and delta computation.

Modules:

  • adapters –

    Adapters that expose :class:ParserEngine through the legacy

  • events –

    Semantic event types emitted by the streaming parser engine.

  • incremental_lexer –

    Incremental text lexer that converts text chunks into terminal

  • parser_engine –

    Parser engine base that handles both reasoning and tool call

  • parser_engine_config –

    Declarative configuration for model tool-call and reasoning formats.

  • registered_adapters –

    Concrete adapter classes for each registered parser engine.

  • streaming_parser_engine –

    Streaming parser engine that orchestrates token ID scanning,

  • token_id_scanner –

    Scan delta token IDs for special tokens and split the stream into