diff --git a/CMakeLists.txt b/CMakeLists.txt
index 4e91fe28d..35eb16910 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -142,6 +142,11 @@ if (LLAMA_CUBLAS)
 
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cuda_driver)
 
+    # CUDA graphs are independent of whether GPU architectures were supplied.
+    if(CUDAToolkit_VERSION VERSION_GREATER 12)
+        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+    endif()
+
     if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
         # 50 == lowest CUDA 12 standard
         # 60 == f16 CUDA intrinsics
@@ -151,11 +156,9 @@ if (LLAMA_CUBLAS)
         # 80 == Ampere, asynchronous data loading, faster tensor core instructions
         message("CUDA Toolkit Version: ${CUDAToolkit_VERSION}")
         if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 13)
-                add_compile_definitions(GGML_CUDA_USE_GRAPHS) #try enable cuda graphs on cu12 build
                 add_compile_definitions(KCPP_LIMIT_CUDA_MAX_ARCH=860)
                 set(CMAKE_CUDA_ARCHITECTURES "75-virtual;80-virtual;86-virtual") # lowest CUDA 13 standard
         elseif(CUDAToolkit_VERSION VERSION_GREATER 12)
-                add_compile_definitions(GGML_CUDA_USE_GRAPHS) #try enable cuda graphs on cu12 build
                 add_compile_definitions(KCPP_LIMIT_CUDA_MAX_ARCH=800)
                 set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;75-virtual;80-virtual") # lowest CUDA 12 standard + lowest for integer intrinsics
         else()
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index c1f337dd4..e3b802d88 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -2705,6 +2705,9 @@ static bool ggml_cuda_graph_debug_enabled() {
 static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
 
     bool use_cuda_graph = true;
+    // Opt-in ADD capture: preserve the existing restriction by default.
+    // This does not bypass other capture compatibility checks.
+    static const bool allow_add_capture = ggml_cuda_env_flag("GGML_CUDA_GRAPH_ALLOW_ADD");
     // Loop over nodes in GGML graph to obtain info needed for CUDA graph
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
@@ -2722,13 +2725,18 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
                 // ref: https://github.com/ggml-org/llama.cpp/pull/18958
                 use_cuda_graph = false;
                 if (ggml_cuda_graph_debug_enabled()) {
-                    GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
+                    GGML_LOG_DEBUG("%s: CUDA%d disabling CUDA graphs at MUL_MAT_ID node '%s'\n",
+                            __func__, ggml_cuda_get_device(), node->name);
                 }
             }
         }
 
-        if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {
+        if (!allow_add_capture && node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {
             use_cuda_graph = false;
+            if (ggml_cuda_graph_debug_enabled()) {
+                GGML_LOG_DEBUG("%s: CUDA%d disabling CUDA graphs at ADD node '%s', src1 ne[1]=%lld\n",
+                        __func__, ggml_cuda_get_device(), node->name, (long long) node->src[1]->ne[1]);
+            }
         }
 
         if (!use_cuda_graph) {