perf: gate planner budget cap on stream_layers_enabled

Author: fszontagh

src/ggml_extend.hpp

@@ -2470,17 +2470,21 @@ struct GGMLRunner {
             *effective_budget_out = effective_budget;
         }
 
-        // When the model dwarfs the budget, cap the planner at a quarter so
-        // it builds smaller merged segments and chunk-K can fit alongside.
-        // Otherwise leave the planner free to merge into one large segment.
-        size_t total_params_bytes = 0;
-        for (const ggml_tensor* t : params_tensor_set_) {
-            if (t != nullptr) {
-                total_params_bytes += ggml_nbytes(t);
+        // When streaming and the model dwarfs the budget, cap the planner at
+        // a quarter so it builds smaller merged segments and chunk-K can fit
+        // alongside. Without streaming the cap only adds dispatch overhead.
+        size_t planner_budget = effective_budget;
+        if (stream_layers_enabled) {
+            size_t total_params_bytes = 0;
+            for (const ggml_tensor* t : params_tensor_set_) {
+                if (t != nullptr) {
+                    total_params_bytes += ggml_nbytes(t);
+                }
+            }
+            if (total_params_bytes * 4 > effective_budget * 3) {
+                planner_budget = effective_budget / 4;
             }
         }
-        const size_t planner_budget =
-            (total_params_bytes * 4 > effective_budget * 3) ? effective_budget / 4 : effective_budget;
 
         *plan_out = sd::ggml_graph_cut::resolve_plan(runtime_backend,
                                                      gf,