use burn::{
Tensor,
config::Config,
prelude::{
Int,
TensorData,
},
};
use super::blocks::Qwen3Asr;
pub const QWEN3_ASR_EOS: [usize; 2] = [151_643, 151_645];
#[derive(Config, Debug)]
pub struct Qwen3AsrGreedyDecoderConfig {
#[config(default = "512")]
pub max_new_tokens: usize,
#[config(default = "QWEN3_ASR_EOS.to_vec()")]
pub eos: Vec<usize>,
}
impl Qwen3AsrGreedyDecoderConfig {
pub fn init(&self) -> Qwen3AsrGreedyDecoder {
Qwen3AsrGreedyDecoder {
config: self.clone(),
}
}
}
#[derive(Clone, Debug)]
pub struct Qwen3AsrGreedyDecoder {
config: Qwen3AsrGreedyDecoderConfig,
}
impl Qwen3AsrGreedyDecoder {
pub fn config(&self) -> &Qwen3AsrGreedyDecoderConfig {
&self.config
}
pub fn decode(
&self,
model: &Qwen3Asr,
embeds: Tensor<3>,
) -> Vec<usize> {
let lm = &model.language_model;
let device = embeds.device();
let seq = embeds.dims()[1];
let total = seq + self.config.max_new_tokens;
let r_emb = lm.rotary(total, &device);
let mut cache = lm.new_cache(1, total);
let argmax_last = |hidden: Tensor<3>| -> usize {
let n = hidden.dims()[1];
lm.logits(hidden.slice_dim(1, n - 1..n))
.argmax(2)
.into_scalar::<i64>() as usize
};
let mut token = argmax_last(lm.forward_hidden(embeds, &r_emb, Some(&mut cache)));
let mut out = vec![token];
while !self.config.eos.contains(&token) && out.len() < self.config.max_new_tokens {
let ids =
Tensor::<2, Int>::from_data(TensorData::new(vec![token as i64], [1, 1]), &device);
token = argmax_last(lm.forward_hidden(lm.embed(ids), &r_emb, Some(&mut cache)));
out.push(token);
}
out
}
}