Principle:Mit han lab Llm awq Conversation Template Management
| Knowledge Sources | |
|---|---|
| Domains | NLP, Prompt_Engineering |
| Last Updated | 2026-02-15 00:00 GMT |
Overview
Principle of managing conversation prompt templates that format user/assistant message histories into model-specific prompt strings.
Description
Different LLMs expect different prompt formats (ChatML, LLAMA2, Vicuna, MPT, etc.). Conversation template management provides a registry pattern where named templates define the system prompt, role labels, separator tokens, and stop criteria. A conversation object accumulates messages and produces a formatted prompt string via get_prompt(). This abstraction decouples the application logic from model-specific formatting requirements.
Usage
Apply this principle whenever an application needs to support multiple chat model backends with different prompt formats. The template pattern ensures correct formatting without hardcoding model-specific logic.
Theoretical Basis
The core pattern is a Strategy design: each separator style defines a different formatting algorithm, selected at runtime based on the model. The Registry pattern (global dict of templates) enables extensibility.