Principle:Predibase Lorax Chat Template Rendering

Knowledge Sources	HuggingFace Chat Templates Jinja2 Templates
Domains	NLP, Prompt_Engineering
Last Updated	2026-02-08 02:00 GMT

Overview

A prompt construction mechanism that converts structured chat messages (system/user/assistant turns) into a flat token string using model-specific Jinja2 chat templates.

Description

Chat Template Rendering bridges the gap between the high-level OpenAI message format and the raw text input that language models consume. Different models use different prompt formats (e.g., Llama uses [INST]...[/INST], Mistral uses [INST]...[/INST], ChatML uses <|im_start|>...<|im_end|>).

The rendering process:

1. Parse structured messages (role + content pairs)
2. Load the model's Jinja2 chat template from the tokenizer config
3. Render messages through the template engine (minijinja in Rust)
4. Optionally inject tool definitions and system guidelines

Usage

Applied automatically when using the /v1/chat/completions endpoint. The template is loaded from the model's tokenizer_config.json and applied to each request's message array.

Theoretical Basis

Pseudo-code:

# Chat template rendering
template = load_jinja_template(tokenizer_config)
messages = [
    {"role": "system", "content": "You are helpful."},
    {"role": "user", "content": "Hello"},
]
prompt = template.render(messages=messages, add_generation_prompt=True)
# Result: "<s>[INST] You are helpful.\n\nHello [/INST]"

Related Pages

Implemented By

• Implementation:Predibase_Lorax_Prepare_Chat_Input