fix

Author: orca-zhang

src/llama.cpp

@@ -605,8 +605,8 @@ static struct ggml_tensor * llm_build_kqv(
 
         if (n_embd_head_v < n_embd_head_k) {
             cur = ggml_cont(ctx, cur);
-            cur = ggml_cont(ctx, ggml_view_3d(ctx, cur, n_embd_head_v, n_head, n_tokens,
-                               ggml_element_size(cur) * n_embd_head_v_out,
+            cur = ggml_cont(ctx, ggml_view_3d(ctx, cur, n_head, n_embd_head_v, n_tokens,
+                               ggml_element_size(cur) * n_head,
                                ggml_element_size(cur) * n_embd_head_v_out * n_head,
                                0));
         }