vllm-reasoning-parsers

What problem does it solve?

The vLLM reasoning parser framework provides a centralized reference for implementing, validating, and debugging the split between model reasoning (CoT) and final content across all supported parsers, wrappers, and plugins.

Core Features & Use Cases

• Comprehensive registry of 25+ reasoning parsers with guidance on start/end tokens, streaming vs non-streaming behavior, and tokenization quirks.
• Guidance for building custom parsers and plugins, testing against non-streaming and streaming paths, and handling tool calls and structured outputs.
• Use cases include integrating with vLLM chat endpoints, debugging reasoning/content routing, and extending support for new models (DeepSeek, Qwen3, Granite, Harmony, etc.) through delegated wrappers or custom parsers.

Quick Start

Observe and test end-to-end reasoning parsing by sending a chat request to a vLLM server configured with --reasoning-parser vllm-reasoning-parsers and inspect the reasoning and content fields.