config: the readback and the resolving view say what they are (#35027)

rust/sglang-server/src/api_server/common.rs

@@ -70,10 +70,18 @@ async fn await_control_result(
 
 /// `GET /get_model_info` (+ `/model_info` alias) — static model metadata from
 /// `server_args` (no scheduler round-trip); `is_generation` always true.
+///
+/// Under `SGLANG_RUST_SERVER=1` this is the only `/model_info` a client can
+/// reach — `launch_server` never mounts the Python app — so it answers the same
+/// keys. It answers them from the launch blob, which is the whole of this
+/// server's config knowledge: `server_args` is parsed once at boot and held
+/// behind an `Arc`, and no route mounted here changes weights or parsers, so
+/// the launch values are also the current ones.
 async fn model_info(State(state): State<AppState>) -> Response {
     let sa = &state.server_args;
     let body = serde_json::json!({
         "model_path": sa.model_path,
+        "served_model_name": sa.served_model_name,
         "tokenizer_path": sa.tokenizer_path,
         "is_generation": true,
         // Python's `TokenizerManager` merges this into every request
@@ -81,7 +89,14 @@ async fn model_info(State(state): State<AppState>) -> Response {
         // `RustServer.launch` REFUSES to start when it is set. It can therefore
         // only be null here — echoing it keeps the field's shape.
         "preferred_sampling_params": sa.preferred_sampling_params,
-        "weight_version": serde_json::Value::Null,
+        // Python answers this through `config_value`, so a control-plane write
+        // moves it there; here it is the launch value.
+        "weight_version": sa.weight_version,
+        "load_format": sa.load_format,
+        // `auto` never reaches the blob: `resolve_auto_parsers` writes the
+        // selected parser into `server_args` before the scheduler forks.
+        "reasoning_parser": sa.reasoning_parser,
+        "tool_call_parser": sa.tool_call_parser,
     });
     (
         StatusCode::OK,

rust/sglang-server/src/runtime/config.rs

@@ -72,6 +72,18 @@ pub struct ServerArgs {
     /// HF revision, used only when `tokenizer_path` is a repo id. `None` → main.
     #[serde(default)]
     pub revision: Option<String>,
+    /// Weight format selected by `--load-format`, reported by `/get_model_info`.
+    /// The blob carries the post-`__post_init__` value (`auto` is already
+    /// narrowed to `gguf` / `mistral` / `runai_streamer` / `remote` where the
+    /// checkpoint demands it). Not consumed for loading -- the scheduler owns
+    /// that; `None` only when the blob omits the key.
+    #[serde(default)]
+    pub load_format: Option<String>,
+    /// Operator-supplied weight version, reported by `/model_info`. Defaults to
+    /// `"default"` on the Python side, so it is present in every blob; `None`
+    /// only when the blob omits the key.
+    #[serde(default)]
+    pub weight_version: Option<String>,
     /// HTTP bind address (see [`Self::bind`]).
     #[serde(default = "default_host")]
     pub host: String,