diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu
index b2668113..d5654df5 100644
--- a/ggml/src/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda.cu
@@ -4283,7 +4283,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
     [[maybe_unused]] const bool integrated = false; //ggml_cuda_info().devices[cuda_ctx->device].integrated;
 
 #ifdef USE_CUDA_GRAPH
-    auto graph = ggml_cuda_get_graph(*cuda_ctx, ggml_cuda_graph_get_key(cgraph));
+    auto graph = use_cuda_graph ? ggml_cuda_get_graph(*cuda_ctx, ggml_cuda_graph_get_key(cgraph)) : nullptr;
 #endif
 
     //printf("======================== %s: graph with %d nodes on device %d. time = %ld\n", __func__, cgraph->n_nodes, cuda_ctx->device, ggml_time_us());