diff --git a/llama.cpp b/llama.cpp
index 85fa4ad76..22ff38012 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -2681,10 +2681,10 @@ static void llm_load_hparams(
         // gpt-j n_rot = rotary_dim
     }
 
-    hparams.n_embd_head_k = hparams.n_embd / hparams.n_head_kv;
+    hparams.n_embd_head_k = hparams.n_embd / hparams.n_head;
     ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k, false);
 
-    hparams.n_embd_head_v = hparams.n_embd / hparams.n_head_kv;
+    hparams.n_embd_head_v = hparams.n_embd / hparams.n_head;
     ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false);
 
     // arch-specific KVs