diff --git a/candle-holder-models/src/models/llama/modeling.rs b/candle-holder-models/src/models/llama/modeling.rs
index 8c7814e..151e423 100644
--- a/candle-holder-models/src/models/llama/modeling.rs
+++ b/candle-holder-models/src/models/llama/modeling.rs
@@ -3,8 +3,8 @@ use std::sync::Arc;
 use candle_core::{DType, Device, Module, Tensor};
 use candle_holder::{Error, Result};
 use candle_nn::{
-    embedding, init::DEFAULT_KAIMING_NORMAL, linear_no_bias, ops::softmax_last_dim, rms_norm,
-    rotary_emb::rope, Dropout, Embedding, Linear, RmsNorm, VarBuilder,
+    embedding, linear_no_bias, ops::softmax_last_dim, rms_norm, rotary_emb::rope, Dropout,
+    Embedding, Linear, RmsNorm, VarBuilder,
 };
 
 use crate::{
@@ -417,12 +417,11 @@ pub struct LlamaForCausalLM {
 impl LlamaForCausalLM {
     fn load_lm_head(vb: VarBuilder, config: &LlamaConfig) -> Result<Linear> {
         let lm_head = if config.tie_word_embeddings.unwrap_or(false) {
-            let init_ws = DEFAULT_KAIMING_NORMAL;
-            let ws = vb
-                .pp("model.embed_tokens")
-                .get_with_hints((config.vocab_size, config.hidden_size), "weight", init_ws)?
-                .t()?;
-            Linear::new(ws, None)
+            linear_no_bias(
+                config.hidden_size,
+                config.vocab_size,
+                vb.pp("model.embed_tokens"),
+            )?
         } else {
             linear_no_bias(config.hidden_size, config.vocab_size, vb.pp("lm_head"))?
         };