server : report the trained context in the models listing

update_caps already resolves the model file offline to read its
modalities, so it now reads the trained context from the same GGUF
metadata, and GET /models reports it as context_length when it is
known. A router listing then carries the context without any Hub
request, which lets the UI sort and filter by it offline.

Assisted-by: pi:zai-org/GLM-5.3-Flash

tools/server/server-models.cpp

@@ -585,6 +585,11 @@ void server_model_meta::update_caps(const common_params & base) {
 
     // offline discovery cannot see video; a loaded model reports it
     architecture = server_model_architecture_json(inp_image, inp_audio, false, output_modalities);
+
+    // the trained context, read from the GGUF metadata
+    if (!params.model.path.empty()) {
+        n_ctx_train = common_get_gguf_n_ctx_train(params.model.path);
+    }
 }
 
 //
@@ -2125,6 +2130,10 @@ void server_models_routes::init_routes() {
                 // TODO: add other fields, may require reading GGUF metadata
             };
 
+            if (meta.n_ctx_train > 0) {
+                model_info["context_length"] = meta.n_ctx_train;
+            }
+
             // merge with loaded_info from the child process if available
             if (meta.is_running()) {
                 for (auto it = meta.loaded_info.begin(); it != meta.loaded_info.end(); ++it) {

tools/server/server-models.h

@@ -86,6 +86,7 @@ struct server_model_meta {
     int stop_timeout = 0; // seconds to wait before force-killing the model instance during shutdown
     bool hidden = false; // hidden from GET /models, but still accept if requested
     json architecture = server_model_architecture_json(false, false, false, {"text"});
+    uint32_t n_ctx_train = 0; // trained context, read from the GGUF metadata; 0 when unknown
 
     bool is_ready() const {
         return status == SERVER_MODEL_STATUS_LOADED;