use burn::Tensor;
use super::{
LmAttentionOnlyKind,
LmDecodeKind,
LmDecoderMeta,
LmDecoderModel,
LmHybridKind,
LmKvForwardHidden,
LmUncachedForwardHidden,
};
use crate::{
blocks::transformers::embedding::RotarySet,
ops::transformers::attention::{
KVCache,
KVCacheConfig,
KVCacheModel,
},
};
pub trait LmDecoderHolder {
type Kind: LmDecodeKind;
fn lm(&self) -> &LmDecoderModel<Self::Kind>;
fn lm_mut(&mut self) -> &mut LmDecoderModel<Self::Kind>;
fn from_lm(lm: LmDecoderModel<Self::Kind>) -> Self;
}
pub trait LmDecoderWrapper: LmDecoderHolder {}
impl<W: LmDecoderWrapper<Kind = LmAttentionOnlyKind>> LmKvForwardHidden for W {
fn forward_hidden(
&self,
embeds: Tensor<3>,
r_emb: &RotarySet,
kv_cache: Option<&mut KVCache>,
) -> Tensor<3> {
self.lm().forward_hidden(embeds, r_emb, kv_cache)
}
}
impl<W: LmDecoderWrapper<Kind = LmHybridKind>> LmUncachedForwardHidden for W {
fn forward_hidden(
&self,
embeds: Tensor<3>,
r_emb: &RotarySet,
) -> Tensor<3> {
self.lm().forward_hidden(embeds, r_emb)
}
}
impl<W: LmDecoderHolder> LmDecoderMeta for W {
fn n_vocab(&self) -> usize {
self.lm().n_vocab()
}
fn n_logits(&self) -> usize {
self.lm().n_logits()
}
fn d_model(&self) -> usize {
self.lm().d_model()
}
fn n_layers(&self) -> usize {
self.lm().n_layers()
}
fn n_kv_heads(&self) -> usize {
self.lm().n_kv_heads()
}
fn head_dim(&self) -> usize {
self.lm().head_dim()
}
}
impl<W: LmDecoderHolder<Kind = LmAttentionOnlyKind>> KVCacheModel for W {
fn kv_cache_config(&self) -> KVCacheConfig {
self.lm().kv_cache_config()
}
}