ggml : add ggml_build_forward_select (#18550)

* ggml : add ggml_build_forward_select

* cuda : adapt CUDA graph compat to new feature

* vulkan : update logic to handle command buffer closing

* ggml : check compute for fusion

* ggml : add comment

ggml/src/ggml-cuda/common.cuh

@@ -1123,6 +1123,7 @@ struct ggml_tensor_extra_gpu {
 struct ggml_cuda_graph_node_properties {
     void * node_address;
     ggml_op node_op;
+    int32_t flags;
     int64_t ne[GGML_MAX_DIMS];
     size_t nb[GGML_MAX_DIMS];
     void * src_address[GGML_MAX_SRC];

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2918,6 +2918,7 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
 static void ggml_cuda_graph_node_set_properties(ggml_cuda_graph_node_properties * props, ggml_tensor * node) {
     props->node_address = node->data;
     props->node_op = node->op;
+    props->flags = node->flags;
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
         props->ne[i] = node->ne[i];
         props->nb[i] = node->nb[i];
@@ -2961,6 +2962,10 @@ static bool ggml_cuda_graph_node_properties_match(ggml_tensor * node, ggml_cuda_
         return false;
     }
 
+    if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) != (props->flags & GGML_TENSOR_FLAG_COMPUTE)) {
+        return false;
+    }
+
     return true;
 }
 
@@ -3378,6 +3383,9 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                     continue;
                 }
 
+                if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+                    continue;
+                }
 
                 // start of fusion operations
                 static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);