vllm-chat-templates

What problem does it solve?

Diagnoses and triages vLLM chat-template problems by separating template-side issues from parser failures, guiding operators to identify whether the symptom originates in the template, the reasoning parser, or the tool parser, and pointing to known fixes and recovery steps.

Core Features & Use Cases

• Layered triage guidance for the three core shells: chat template (Jinja), reasoning parser, and tool parser, with cross-layer diagnostics.
• Per-family familiarity with flags, known bugs, workarounds, and up-to-date guidelines for model families like Qwen, DeepSeek, GPT-OSS, Kimi-K2, Llama-4, Mistral, Gemma, Phi-4, and GLM.
• Production-oriented debugging workflow including enabling prompt logging, bypassing vLLM, capturing raw outputs, pinning versions, and using debugging references to accelerate problem resolution.

Quick Start

Start by logging the actual prompt served by vLLM and compare byte-for-byte against the model card example to determine whether the issue originates from the template, the parser, or forwarded kwargs.