[router] fix get load response parsing (#11213)

sgl-router/py_test/fixtures/mock_worker.py

@@ -141,7 +141,17 @@ def create_app(args: argparse.Namespace) -> FastAPI:
     @app.get("/get_load")
     async def get_load(request: Request):
         check_api_key(request)
-        return JSONResponse({"load": _inflight})
+        # Return format matching real workers: array of load info per DP rank
+        return JSONResponse(
+            [
+                {
+                    "dp_rank": 0,
+                    "num_reqs": _inflight,
+                    "num_waiting_reqs": 0,
+                    "num_tokens": _inflight,
+                }
+            ]
+        )
 
     def make_json_response(obj: dict, status_code: int = 200) -> JSONResponse:
         resp = JSONResponse(obj, status_code=status_code)

sgl-router/src/core/worker_manager.rs

@@ -1252,11 +1252,22 @@ impl WorkerManager {
             Ok(response) if response.status().is_success() => {
                 match response.json::<Value>().await {
                     Ok(json) => {
-                        if let Some(load) = json.get("load").and_then(|v| v.as_i64()) {
+                        // The /get_load endpoint returns an array of load info objects (one per DP rank)
-                            debug!("Worker {} load: {}", url, load);
+                        // Each object has: {dp_rank, num_reqs, num_waiting_reqs, num_tokens}
-                            Some(load as isize)
+                        if let Some(array) = json.as_array() {
+                            let total_tokens: i64 = array
+                                .iter()
+                                .filter_map(|entry| {
+                                    entry.get("num_tokens").and_then(|v| v.as_i64())
+                                })
+                                .sum();
+                            debug!("Worker {} load (total tokens): {}", url, total_tokens);
+                            Some(total_tokens as isize)
                         } else {
-                            warn!("Invalid load response from {}: {:?}", url, json);
+                            warn!(
+                                "Invalid load response from {}: expected array, got {:?}",
+                                url, json
+                            );
                             None
                         }
                     }