use std::sync::Arc;
use bunsen::{
errors::BunsenResult,
kits::{
language::generation::nanochat::NanoChatGpt,
support::{
construct::LmConstruct,
model_config::GenerationPolicy,
tokens::{
BunsenTokenizer,
Tokenizer,
},
},
},
};
use crate::chat::{
ChatLm,
ChatMessage,
ChatParts,
ChatRenderer,
LmChat,
NanoChatLayout,
backend::missing,
};
pub type NanoChatChat = LmChat<NanoChatGpt>;
impl ChatLm for NanoChatGpt {
fn chat_parts(bundle: &LmConstruct<Self>) -> BunsenResult<ChatParts> {
let path = bundle.tokenizer().ok_or_else(|| missing("tokenizer.pkl"))?;
let tokenizer: Arc<dyn Tokenizer> =
Arc::new(BunsenTokenizer::from_tiktoken_pickle_file(path)?);
let layout = NanoChatLayout::from_tokenizer(tokenizer.as_ref())?;
let generation = GenerationPolicy {
stop_ids: layout.stop_ids(),
..bundle.policy.generation.clone()
};
Ok(ChatParts {
tokenizer,
renderer: Arc::new(layout),
generation,
})
}
}
impl ChatRenderer for NanoChatLayout {
fn render_prompt(
&self,
tokenizer: &dyn Tokenizer,
messages: &[ChatMessage],
_enable_thinking: Option<bool>,
) -> BunsenResult<Vec<i64>> {
self.render(tokenizer, messages)
}
}