fix

Author: orca-zhang

src/llama.cpp

@@ -603,7 +603,7 @@ static struct ggml_tensor * llm_build_kqv(
         ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
 
         if (n_embd_head_v < n_embd_head_k) {
-            cur = ggml_reshape_3d(ctx, cur, n_embd_head_v, n_head, n_tokens);
+            cur = ggml_reshape_3d(ctx, cur, n_embd_head_v_out, n_head, n_tokens);
             cur = ggml_cont(ctx, ggml_view_3d(ctx, cur, n_embd_head_v, n_head, n_tokens,
                             ggml_row_size(cur->type, n_embd_head_v_out),
                             ggml_row_size(cur->type, n_embd_head_v_out * n_head),