server: fix router args not being forwarded to child instances (#24760)

diff --git a/tools/server/server-models.cpp b/tools/server/server-models.cpp
index ff9a0df12f4bf3a25ad0b451c15db1dcb8997e01..cc7184165d33a1c390c1c09db6084c9c011e44b2 100644 (file)
--- a/tools/server/server-models.cpp
+++ b/tools/server/server-models.cpp
@@ -351,6 +351,12 @@ void server_models::load_models() {
         source_map[name] = SERVER_MODEL_SOURCE_PRESET;
     }
 
+    // overlay router's own CLI args on top of every model preset so that
+    // e.g. `llama-server --temp 0` is honoured by all child processes
+    for (auto & [name, preset] : final_presets) {
+        preset.merge(base_preset);
+    }
+
     auto get_source = [&](const std::string & name) {
         return source_map.count(name) ? source_map.at(name) : SERVER_MODEL_SOURCE_PRESET;
     };