A high-performance inference engine for AI models
Embedding hot fix
LuckyIYI24561e57d6d2
diff --git a/crates/uzu/src/encodable_block/decoder.rs b/crates/uzu/src/encodable_block/decoder.rsindex 305d89a84..8ee3c352b 100644--- a/crates/uzu/src/encodable_block/decoder.rs+++ b/crates/uzu/src/encodable_block/decoder.rs@@ -37,17 +37,29 @@ impl<B: Backend> Decoder<B> {decoder_config: Rc<DecoderConfig>,root_weight_loader: &ParameterTree<B::Context>,) -> Self {- Self::new_with_embedding_and_readout_subtrees(- context,- decoder_config,- root_weight_loader,- "transformer",- "embedding",- "embedding",+ let embedding_weight_loader = root_weight_loader.subtree("embedding").expect("Failed to get embedding subtree");++ let embed = Embedding::new(+ context.as_ref(),+ decoder_config.vocab_size as u32,+ decoder_config.model_dim as u32,+ &decoder_config.embedding_config,+ &embedding_weight_loader,)+ .expect("Failed to create embedding");++ let (layers, norm) =+ Self::build_transformer_layers_and_norm(context, decoder_config, root_weight_loader, "transformer");++ Self {+ embed,+ layers,+ norm,+ }}/// Used by models whose token lookup weights and logits readout weights+ #[cfg(all(feature = "audio-runtime", metal_backend))]pub fn new_with_embedding_and_readout_subtrees(context: Rc<B::Context>,decoder_config: Rc<DecoderConfig>,diff --git a/crates/uzu/src/encodable_block/embedding.rs b/crates/uzu/src/encodable_block/embedding.rsindex 92c68ed77..0ea8075ed 100644--- a/crates/uzu/src/encodable_block/embedding.rs+++ b/crates/uzu/src/encodable_block/embedding.rs@@ -120,6 +120,7 @@ const SHARED_TREE_UNTIED_KEYS: UntiedWeightKeys = UntiedWeightKeys {readout_biases: "output_biases",};+#[cfg(all(feature = "audio-runtime", metal_backend))]const SPLIT_TREE_UNTIED_KEYS: UntiedWeightKeys = UntiedWeightKeys {lookup_weights: "weights",lookup_scales: "scales",@@ -168,6 +169,7 @@ impl<B: Backend> Embedding<B> {)}+ #[cfg(all(feature = "audio-runtime", metal_backend))]pub(crate) fn new_with_lookup_and_readout_trees(context: &B::Context,vocab_size: u32,