Module chat 

Chat formatting for each model family (mirrors their chat templates), including tool calling.

Structs

Gemma3Chat

Gemma 3: <start_of_turn> turns, no system role (the system prompt opens the first user turn), no thinking and no tool tokens.

GemmaChat

Llama3Chat

Llama 3.x instruct: header-delimited turns; the template always opens with a system turn carrying the knowledge cutoff and today’s date (“25 Oct 2026”, as strftime_now("%d %b %Y")).

Qwen2Chat

Qwen2 / Qwen2.5 ChatML: like Qwen3 without thinking, and the template’s default system prompt when none is given.

QwenChat

Qwen ChatML templates. Qwen3.5 / 3.8: thinking is on by default in the model, so a non-thinking request closes the think block in the generation prompt, and a thinking one opens it; Qwen3.8’s also adds a reasoning-effort instruction to the system turn in thinking mode (and a small model given that instruction anyway can deliberate about it endlessly). Qwen3: the model opens its own think block, and tools are declared and called in JSON.

Constants

EOT

Statics

QWEN

Qwen3.5, Qwen3.8 (with the reasoning-effort instruction) and Qwen3 formats.

QWEN3

QWEN_EFFORT

Traits

ChatFormat

Chat formatting for a model family.

Functions

conversation_start

Text for the start of a conversation: <bos> plus the optional system turn.

end_of_model_turn

Text to append after the model has finished its reply (before the next user turn).

user_turn

A user turn followed by the generation prompt for the model.

with_turn_end

eos plus the chat format’s end-of-turn token, looked up in the tokenizer.