fix: 模型评测异步加载等待与多节点路由修复

- eval_runner 等待异步模型加载完成（InferenceSession.wait_until_loaded），
  修复 "model load failed: unknown"
- 评测算力节点选择：优先页面选择的节点 / 模型所在节点（_select_eval_node），
  多节点时不再派发到不可达节点导致连接超时
- 前端评测 GPU 选择改为节点感知（节点:GPU 复合值），透传 compute_node_id，
  并检查 startEval 结果展示真实错误
- 大模型评价（judge）使用模型记录的真实 API 模型名（api_model），
  避免用平台内部名调用 LLM API 导致 HTTP 400
- 新增后端节点选择与 compute wait_until_loaded 单元测试

Co-Authored-By: Claude <noreply@anthropic.com>

backend/app/api/v1/endpoints/platform.py

@@ -33,6 +33,22 @@ def _select_first_online_node(store: Any) -> dict[str, Any] | None:
     return None
 
 
+def _select_eval_node(store: Any, preferred_node_id: str | None = None) -> dict[str, Any] | None:
+    """Select the compute node for an eval job.
+
+    被评测模型是节点相关的（训练/合并产物只存在于对应算力节点），因此优先使用
+    页面选择的节点或模型所在节点；若该节点不可用则明确失败，绝不派发到其它
+    可能没有模型路径的节点（多算力节点场景下这是评测失败的主因）。
+    """
+    if preferred_node_id:
+        node = next((n for n in store.compute_nodes() if n.get("id") == preferred_node_id), None)
+        if node:
+            if node.get("enabled") and node.get("scheduler_status") == "online":
+                return node
+            return None
+    return _select_first_online_node(store)
+
+
 def _candidate_online_nodes(store: Any, preferred_node_id: str | None = None) -> list[dict[str, Any]]:
     nodes = [node for node in store.compute_nodes() if node.get("enabled") and node.get("scheduler_status") == "online"]
     if not preferred_node_id:
@@ -1382,13 +1398,16 @@ async def model_eval_start(payload: dict[str, Any] = Body(...)) -> dict[str, Any
     model_id = str(payload.get("model_id", ""))
     model_path = ""
     adapter_path = payload.get("adapter_path", "")
+    model_node_id = ""
     try:
         db_model = store.model(model_id)
         model_path = db_model.get("path", "")
+        model_node_id = db_model.get("compute_node_id") or ""
     except KeyError:
         # Try trained_models table (IDs prefixed with tm_)
         trained = next((m for m in store.trained_models() if m["id"] == model_id), None)
         if trained:
+            model_node_id = trained.get("compute_node_id") or ""
             merged_path = trained.get("merged_path", "")
             base_path = trained.get("base_model_path", "")
             if trained.get("merged") and merged_path:
@@ -1438,16 +1457,22 @@ async def model_eval_start(payload: dict[str, Any] = Body(...)) -> dict[str, Any
             eval_model_name = dim.get("eval_model", "")
             api_url = ""
             api_key = ""
+            api_model_name = ""
             if eval_model_name:
                 try:
                     eval_model = store.model(eval_model_name) if eval_model_name.startswith("m_") else store.model_by_name(eval_model_name)
-                    api_url = eval_model.get("api_url", "")
-                    api_key = eval_model.get("api_key", "")
+                    if isinstance(eval_model, dict):
+                        api_url = eval_model.get("api_url", "")
+                        api_key = eval_model.get("api_key", "")
+                        # 模型记录里的 model_name 是真实 API 模型名（如 deepseek-chat），
+                        # 优先传给评测器，避免用平台内部名称调用 LLM API
+                        api_model_name = eval_model.get("model_name") or ""
                 except (KeyError, Exception):
                     pass
             dimension_cfg = {
                 "type": dim.get("type", ""),
                 "eval_model": eval_model_name,
+                "api_model": api_model_name or eval_model_name,
                 "eval_method": dim.get("eval_method", ""),
                 "eval_prompt": dim.get("eval_prompt", ""),
                 "api_url": api_url,
@@ -1459,11 +1484,13 @@ async def model_eval_start(payload: dict[str, Any] = Body(...)) -> dict[str, Any
         except KeyError:
             pass
 
-    # 5. Select compute node
-    node = _select_first_online_node(store)
+    # 5. Select compute node: 优先页面选择的节点 / 模型所在节点，避免多节点时选错
+    preferred_node_id = payload.get("compute_node_id") or payload.get("node_id") or model_node_id
+    node = _select_eval_node(store, preferred_node_id)
     if not node:
-        store.update_eval_task(task["id"], {"status": "failed", "error": "no online compute node"})
-        return ok({"task_id": task["id"], "status": "failed", "error": "no online compute node"})
+        message = "no online compute node" if not preferred_node_id else f"model compute node not schedulable: {preferred_node_id}"
+        store.update_eval_task(task["id"], {"status": "failed", "error": message})
+        return ok({"task_id": task["id"], "status": "failed", "error": message})
 
     # 6. Build eval job payload
     output_dir = f"/data/yg-ft/outputs/{task['id']}"

backend/tests/test_compare_inference_async.py

@@ -100,6 +100,28 @@ def _patch_store(monkeypatch, store: FakeInferenceStore) -> None:
     monkeypatch.setattr(platform, "get_settings", lambda: SimpleNamespace(compute_mode="real"))
 
 
+def test_select_eval_node_prefers_model_node(monkeypatch) -> None:
+    from app.api.v1.endpoints.platform import _select_eval_node
+
+    store = FakeInferenceStore(nodes=[_node("n1"), _node("n2")])
+    # 指定模型所在节点时优先返回该节点
+    assert _select_eval_node(store, "n2")["id"] == "n2"
+    # 无指定节点时回退到第一个在线节点
+    assert _select_eval_node(store, None)["id"] == "n1"
+
+
+def test_select_eval_node_returns_none_when_model_node_offline(monkeypatch) -> None:
+    from app.api.v1.endpoints.platform import _select_eval_node
+
+    nodes = [_node("n1"), _node("n2")]
+    nodes[1]["enabled"] = False
+    store = FakeInferenceStore(nodes=nodes)
+    # 模型所在节点不可用 → 明确失败，不派发到其它节点
+    assert _select_eval_node(store, "n2") is None
+    # 无指定节点时仍回退第一个在线节点
+    assert _select_eval_node(store, None)["id"] == "n1"
+
+
 def test_model_compare_load_dispatches_and_returns_starting(monkeypatch) -> None:
     store = FakeInferenceStore(tasks=[_task("t1", node_id="n1")], nodes=[_node("n1")])
     _patch_store(monkeypatch, store)