diff --git a/llama.cpp b/llama.cpp
index e46147b17..512805980 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -790,6 +790,7 @@ static bool llama_eval_internal(
     memcpy(embd->data, tokens, N*ggml_element_size(embd));
 
     struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd);
+    struct ggml_tensor * scale = ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head));
 
     for (int il = 0; il < n_layer; ++il) {
         struct ggml_tensor * inpSA = inpL;
@@ -850,7 +851,7 @@ static bool llama_eval_internal(
             struct ggml_tensor * KQ_scaled =
                 ggml_scale_inplace(ctx0,
                         KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
+                        scale);
 
             // KQ_masked = mask_past(KQ_scaled)
             struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);