feat: 模型推理异步加载与对话链路修复，同步基线

模型推理全异步化改造：
- 计算节点 InferenceSession 改为后台线程异步加载模型，load 立即返回，
  加载期间事件循环保持响应（/inference/status 与 /health 不阻塞）
- 后端模型加载改为异步派发 + 轮询对账器（reconcile_inference_loads），
  任务状态由 starting 自动推进到 ready/error，解决多节点启动超时
  （timeout of 120000ms exceeded）
- 推理删除/卸载改为任务感知 + 短超时，删除先删记录再 best-effort 卸载，
  不再被不可达节点阻塞；同节点新模型替换旧任务标记失效
- 流式对话透传 task_id/node_id 路由到真正加载模型的算力节点，
  useStreamChat 解析 SSE 错误帧以干净文案展示
- 对话历史按任务 id 本地持久化，退出重进可恢复；移除页脚提示文本
- 新增后端推理异步加载与计算节点异步状态机单元测试

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

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

@@ -33,6 +33,15 @@ def _select_first_online_node(store: Any) -> dict[str, Any] | None:
     return None
 
 
+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:
+        return nodes
+    preferred = [node for node in nodes if node.get("id") == preferred_node_id]
+    others = [node for node in nodes if node.get("id") != preferred_node_id]
+    return preferred + others
+
+
 def _build_messages_payload(payload: dict[str, Any]) -> dict[str, Any]:
     """Convert frontend inference payload to compute API messages format.
 
@@ -61,10 +70,47 @@ def _build_messages_payload(payload: dict[str, Any]) -> dict[str, Any]:
     }
 
 
+def _node_for_inference_payload(store: Any, payload: dict[str, Any]) -> dict[str, Any] | None:
+    node_id = payload.get("node_id") or payload.get("compute_node_id")
+    task_id = payload.get("task_id") or payload.get("compare_task_id")
+    if task_id and not node_id:
+        try:
+            task = store.compare_task(str(task_id))
+            load_status = task.get("load_status") or {}
+            if isinstance(load_status, str):
+                load_status = json.loads(load_status)
+            loaded_models = load_status.get("loaded_models") or []
+            ready_model = next((item for item in loaded_models if item.get("status") in {"ready", "running"} and item.get("node_id")), None)
+            if ready_model:
+                node_id = ready_model.get("node_id")
+        except Exception:
+            node_id = None
+    if node_id:
+        return next((node for node in store.compute_nodes() if node.get("id") == node_id), None)
+    return _select_first_online_node(store)
+
+
 async def _stream_chat_proxy(payload: dict[str, Any]) -> StreamingResponse:
     """Common SSE streaming proxy: convert payload → forward to compute node → stream back."""
     store = get_platform_store()
-    node = _select_first_online_node(store)
+    # 任务仍在加载中时，直接返回明确的加载中提示，避免转发到尚未就绪的节点
+    task_id = payload.get("task_id") or payload.get("compare_task_id")
+    if task_id:
+        try:
+            task = store.compare_task(str(task_id))
+            load_status = task.get("load_status") or {}
+            if isinstance(load_status, str):
+                load_status = json.loads(load_status)
+            items = load_status.get("loaded_models") or []
+            if items and not any(item.get("status") in {"ready", "running"} for item in items):
+                if any(item.get("status") == "starting" for item in items):
+                    return StreamingResponse(
+                        iter(['data: {"error": "模型加载中，请稍候再试"}\n\n']),
+                        media_type="text/event-stream",
+                    )
+        except Exception:  # noqa: BLE001 - fall through to normal routing on lookup errors
+            pass
+    node = _node_for_inference_payload(store, payload)
     if not node:
         return StreamingResponse(
             iter(['data: {"error": "no online compute node available for inference"}\n\n']),
@@ -729,12 +775,17 @@ async def merge_model(payload: dict[str, Any] = Body(...)) -> dict[str, Any]:
         None,
     )
     base_model_path = payload.get("base_model_path") or (trained_model and trained_model.get("base_model_path"))
-    adapter_path = payload.get("adapter_path") or payload.get("adapter_name_or_path") or (trained_model and trained_model.get("merged_path"))
+    adapter_path = (
+        payload.get("adapter_path")
+        or payload.get("adapter_name_or_path")
+        or (trained_model and (trained_model.get("artifact_dir") or trained_model.get("adapter_path") or trained_model.get("merged_path")))
+    )
     if not base_model_path:
         raise fail(400, "base_model_path is required")
     if not adapter_path:
         raise fail(400, "adapter_path is required")
-    node = store.schedule_node({**payload, "gpus": payload.get("gpus") or []})
+    requested_node_id = payload.get("requested_node_id") or payload.get("compute_node_id") or (trained_model and trained_model.get("compute_node_id"))
+    node = store.schedule_node({**payload, "requested_node_id": requested_node_id, "gpus": payload.get("gpus") or []})
     health = node.get("health_detail") or {}
     output_root = str(health.get("output_root") or f"{node['data_root'].rstrip('/')}/outputs")
     output_name = str(payload.get("output_model_name") or payload.get("merged_model_name") or f"{trained_model_id or 'model'}-merged")
@@ -753,11 +804,13 @@ async def merge_model(payload: dict[str, Any] = Body(...)) -> dict[str, Any]:
         "gpus": payload.get("gpus") or [],
         "trained_model_id": trained_model["id"] if trained_model else trained_model_id,
         "model_name": trained_model["name"] if trained_model else payload.get("model_name"),
+        "compute_node_id": node["id"],
+        "compute_node_code": node.get("code"),
     }
     if get_settings().compute_mode == "simulator":
         job = {"id": job_payload["id"], "status": "queued", "progress": 10, "command": [], "output_dir": output_dir}
     else:
-        client = ComputeNodeClient(node["api_base_url"])
+        client = ComputeNodeClient(node["api_base_url"], timeout=900)
         preview = await client.validate_job(job_payload)
         if not preview.get("valid", False):
             raise fail(409, "; ".join(preview.get("errors") or ["merge preflight failed"]))
@@ -1537,28 +1590,49 @@ async def model_compare_detail(task_id: str) -> dict[str, Any]:
         raise fail(404, "compare task not found")
 
 
-async def _unload_from_compute_node() -> dict[str, Any]:
+async def _unload_from_compute_node(store: Any, task: dict[str, Any] | None = None) -> dict[str, Any]:
-    """Best-effort unload the inference model from the first online compute node."""
+    """Best-effort unload the inference model from the node(s) that hold it.
-    store = get_platform_store()
+
-    # Clear all inference tracking — only one model can be loaded at a time
+    任务感知：优先卸载 ``task.load_status.loaded_models`` 中记录的节点；
-    for node in store.compute_nodes():
+    无任务时回退到平台记录的已加载推理的节点。每个节点使用短超时，
-        store.mark_inference_unloaded(node["id"])
+    保证卸载永远不会长时间阻塞调用方（例如删除操作）。
-    node = _select_first_online_node(store)
+    """
-    if not node:
+    node_ids: set[str] = set()
-        return {"unloaded": False, "error": "no online compute node"}
+    if task:
+        load_status = task.get("load_status") or {}
+        if isinstance(load_status, str):
             try:
-        client = ComputeNodeClient(node["api_base_url"])
+                load_status = json.loads(load_status)
-        result = await client._request("POST", "/inference/unload", json_data={})
+            except json.JSONDecodeError:
-        return result
+                load_status = {}
-    except Exception as exc:
+        node_ids = {item.get("node_id") for item in load_status.get("loaded_models") or [] if item.get("node_id")}
-        return {"unloaded": False, "error": str(exc)}
+    if not node_ids:
+        node_ids = {node["id"] for node in store.compute_nodes() if store.is_inference_loaded(node["id"])}
+    nodes = [node for node in store.compute_nodes() if node["id"] in node_ids]
+    results: list[dict[str, Any]] = []
+    for node in nodes:
+        try:
+            result = await ComputeNodeClient(node["api_base_url"]).inference_unload()
+            results.append({"node_id": node["id"], "node_code": node.get("code"), "success": True, "result": result})
+        except Exception as exc:  # noqa: BLE001 - best-effort unload must not raise
+            results.append({"node_id": node["id"], "node_code": node.get("code"), "success": False, "error": str(exc)})
+        finally:
+            store.mark_inference_unloaded(node["id"])
+    return {"unloaded": bool(results), "nodes": results}
 
 
 @router.delete("/model-compare/{task_id}")
 async def model_compare_delete(task_id: str) -> dict[str, Any]:
-    # 删除前先释放算力节点上的模型
+    # 先删记录（快），再 best-effort 释放算力节点上的模型——删除绝不被卸载阻塞
-    await _unload_from_compute_node()
+    try:
+        task = get_platform_store().compare_task(task_id)
+    except KeyError:
+        raise fail(404, "compare task not found")
     get_platform_store().delete_compare_task(task_id)
+    try:
+        await _unload_from_compute_node(get_platform_store(), task=task)
+    except Exception:  # noqa: BLE001 - deletion must succeed even if unload fails
+        pass
     return ok({"deleted": task_id})
 
 
@@ -1585,9 +1659,44 @@ async def model_compare_update_load_status(task_id: str, payload: dict[str, Any]
         raise fail(404, "compare task not found")
 
 
+def _invalidate_superseded_models(store: Any, task_id: str, loaded_models: list[dict[str, Any]]) -> None:
+    """同一计算节点同一时刻只能加载一个推理模型。
+
+    当新任务把模型派发到了某节点后，把其它任务中在该节点上 ready/running
+    的模型标记为已被替换，保持平台 DB 与计算节点实际状态一致。
+    """
+    taken_node_ids = {m.get("node_id") for m in loaded_models if m.get("node_id") and m.get("status") == "starting"}
+    if not taken_node_ids:
+        return
+    for other in store.compare_tasks():
+        if str(other.get("id")) == str(task_id):
+            continue
+        load_status = other.get("load_status") or {}
+        if isinstance(load_status, str):
+            try:
+                load_status = json.loads(load_status)
+            except json.JSONDecodeError:
+                load_status = {}
+        items = load_status.get("loaded_models") or []
+        changed = False
+        for item in items:
+            if item.get("node_id") in taken_node_ids and item.get("status") in {"ready", "running"}:
+                item["status"] = "error"
+                item["error"] = "模型已被其他推理任务替换"
+                changed = True
+        if changed:
+            new_status = "loaded" if any(i.get("status") in {"ready", "running"} for i in items) else "failed"
+            store.update_compare_task(other["id"], {"status": new_status, "load_status": {"loaded_models": items}})
+
+
 @router.post("/model-compare/{task_id}/load")
 async def model_compare_load(task_id: str) -> dict[str, Any]:
-    """真正加载模型到算力节点（不再使用假 PID/端口）。"""
+    """异步派发模型加载到算力节点，立即返回。
+
+    加载进度由轮询对账器（compute_poller → reconcile_inference_loads）推进：
+    任务项先以 status=starting 记录，对账器查询节点 /inference/status 后
+    推进到 ready/error。这里只负责把加载请求派发出去，绝不同步等待加载完成。
+    """
     try:
         store = get_platform_store()
         task = store.compare_task(task_id)
@@ -1597,15 +1706,14 @@ async def model_compare_load(task_id: str) -> dict[str, Any]:
                 models = json.loads(models)
             except json.JSONDecodeError:
                 models = []
-        # 选取在线算力节点
+        online_nodes = _candidate_online_nodes(store)
-        node = _select_first_online_node(store)
+        if not online_nodes:
-        if not node:
             return ok({"status": "failed", "error": "no online compute node"})
-        client = ComputeNodeClient(node["api_base_url"])
         loaded_models = []
         for item in models:
             if not isinstance(item, dict):
                 continue
+            preferred_node_id = item.get("node_id") or item.get("compute_node_id")
             model_path = item.get("model_path", "")
             if not model_path:
                 # 尝试从模型库获取路径
@@ -1614,28 +1722,46 @@ async def model_compare_load(task_id: str) -> dict[str, Any]:
                     db_model = store.model(model_id)
                     model_path = db_model.get("path", "")
                 except KeyError:
-                    pass
+                    trained_model = next((m for m in store.trained_models() if str(m.get("id")) == str(model_id)), None)
+                    if trained_model:
+                        model_path = trained_model.get("merged_path") or trained_model.get("artifact_dir") or ""
+                        preferred_node_id = preferred_node_id or trained_model.get("compute_node_id")
             if not model_path:
                 loaded_models.append({**item, "status": "error", "error": "model_path not found"})
                 continue
-            # 真正调用算力节点加载模型
             load_payload = {
                 "model_name_or_path": model_path,
                 "template": item.get("template", "qwen"),
             }
             if item.get("adapter_path"):
                 load_payload["adapter_name_or_path"] = item["adapter_path"]
+            if get_settings().compute_mode == "simulator":
+                loaded_models.append({**item, "status": "ready", "node_id": "", "node_name": ""})
+                continue
+            # 只派发：HTTP 响应成功即视为已接受（节点会异步加载），loaded 字段忽略
+            item_dispatched = False
+            errors = []
+            for node in _candidate_online_nodes(store, preferred_node_id):
                 try:
-                result = await client._request("POST", "/inference/load", json_data=load_payload)
+                    client = ComputeNodeClient(node["api_base_url"])
-                if result.get("loaded"):
+                    await client.inference_load(load_payload)
                     store.mark_inference_loaded(node["id"])
-                    loaded_models.append({**item, "status": "ready"})
+                    loaded_models.append({**item, "status": "starting", "node_id": node["id"], "node_name": node.get("name")})
+                    item_dispatched = True
+                    break
+                except Exception as exc:  # noqa: BLE001 - try next candidate node
+                    errors.append(f"{node.get('name') or node.get('code')}: {exc}")
+            if not item_dispatched:
+                loaded_models.append({**item, "status": "error", "error": "; ".join(errors) or "load dispatch failed"})
+        if any(m.get("status") == "starting" for m in loaded_models):
+            status = "starting"
+        elif any(m.get("status") == "error" for m in loaded_models):
+            status = "failed"
         else:
-                    loaded_models.append({**item, "status": "error", "error": result.get("error", "load failed")})
+            status = "loaded"
-            except Exception as exc:
-                loaded_models.append({**item, "status": "error", "error": str(exc)})
-        status = "loaded" if any(m.get("status") == "ready" for m in loaded_models) else "failed"
         updated = store.update_compare_task(task_id, {"status": status, "load_status": {"loaded_models": loaded_models}})
+        # 同一节点同一时刻只能有一个推理模型；新任务占用了节点后，把其它任务上该节点的模型标记为已被替换
+        _invalidate_superseded_models(store, task_id, loaded_models)
         return ok(updated)
     except KeyError:
         raise fail(404, "compare task not found")
@@ -1645,8 +1771,9 @@ async def model_compare_load(task_id: str) -> dict[str, Any]:
 async def model_compare_unload(task_id: str) -> dict[str, Any]:
     try:
         store = get_platform_store()
-        # 真正释放算力节点上的模型资源
+        task = store.compare_task(task_id)
-        unload_result = await _unload_from_compute_node()
+        # 任务感知卸载：只释放该任务实际加载到的节点，短超时快速返回
+        unload_result = await _unload_from_compute_node(store, task=task)
         updated = store.update_compare_task(task_id, {"status": "pending", "load_status": {"loaded_models": []}})
         return ok({"task": updated, "unload": unload_result})
     except KeyError:
@@ -1662,7 +1789,7 @@ async def model_compare_start_model(task_id: str, payload: dict[str, Any] = Body
 async def model_compare_chat_with_port(payload: dict[str, Any] = Body(...)) -> dict[str, Any]:
     """Proxy non-streaming chat to the compute node running the inference model."""
     store = get_platform_store()
-    node = _select_first_online_node(store)
+    node = _node_for_inference_payload(store, payload)
     if not node:
         return ok({"response": "no online compute node available for inference", "request": payload})
     try:
@@ -1717,8 +1844,9 @@ async def model_chat_local_preload(payload: dict[str, Any] = Body(...)) -> dict[
         return ok({"loaded": False, "error": "no online compute node"})
     try:
         client = ComputeNodeClient(node["api_base_url"])
-        result = await client._request("POST", "/inference/load", json_data=payload)
+        # 计算节点现在异步加载：HTTP 接受（loading/ready）即视为派发成功
-        if result.get("loaded"):
+        result = await client.inference_load(payload)
+        if result.get("loaded") or result.get("status") in {"loading", "ready"}:
             store.mark_inference_loaded(node["id"])
         return ok(result)
     except Exception as exc:
@@ -1729,18 +1857,19 @@ async def model_chat_local_preload(payload: dict[str, Any] = Body(...)) -> dict[
 async def model_chat_local_unload() -> dict[str, Any]:
     """Unload the inference model from the compute node."""
     store = get_platform_store()
-    # Clear all inference tracking
+    # 释放所有已加载推理的节点（短超时，best-effort）
+    results: list[dict[str, Any]] = []
     for n in store.compute_nodes():
-        store.mark_inference_unloaded(n["id"])
+        if not store.is_inference_loaded(n["id"]):
-    node = _select_first_online_node(store)
+            continue
-    if not node:
-        return ok({"unloaded": False, "error": "no online compute node"})
         try:
-        client = ComputeNodeClient(node["api_base_url"])
+            result = await ComputeNodeClient(n["api_base_url"]).inference_unload()
-        result = await client._request("POST", "/inference/unload", json_data={})
+            results.append({"node_id": n["id"], "success": True, "result": result})
-        return ok(result)
+        except Exception as exc:  # noqa: BLE001 - best-effort unload
-    except Exception as exc:
+            results.append({"node_id": n["id"], "success": False, "error": str(exc)})
-        return ok({"unloaded": False, "error": str(exc)})
+        finally:
+            store.mark_inference_unloaded(n["id"])
+    return ok({"unloaded": True, "nodes": results})
 
 
 @router.get("/model-chat/local/status")
@@ -1752,7 +1881,7 @@ async def model_chat_local_status() -> dict[str, Any]:
         return ok({"loaded": False, "error": "no online compute node"})
     try:
         client = ComputeNodeClient(node["api_base_url"])
-        result = await client._request("GET", "/inference/status")
+        result = await client.inference_status()
         return ok(result)
     except Exception as exc:
         return ok({"loaded": False, "error": str(exc)})
@@ -1770,8 +1899,9 @@ async def model_chat_trained_preload(payload: dict[str, Any] = Body(...)) -> dic
         return ok({"loaded": False, "error": "no online compute node"})
     try:
         client = ComputeNodeClient(node["api_base_url"])
-        result = await client._request("POST", "/inference/load", json_data=payload)
+        # 计算节点现在异步加载：HTTP 接受（loading/ready）即视为派发成功
-        if result.get("loaded"):
+        result = await client.inference_load(payload)
+        if result.get("loaded") or result.get("status") in {"loading", "ready"}:
             store.mark_inference_loaded(node["id"])
         return ok(result)
     except Exception as exc:

backend/app/db/platform_store.py

@@ -208,8 +208,12 @@ def parse_training_metric_line(line: str) -> dict[str, float] | None:
     if "loss" not in line and "learning_rate" not in line:
         return None
     result: dict[str, float] = {}
+    number_pattern = r"([-+]?(?:\d+(?:\.\d*)?|\.\d+)(?:[eE][-+]?\d+)?)"
+    step_match = re.search(rf"(?:^|[\s,{{])['\"]?step['\"]?\s*(?:=|:)\s*{number_pattern}", line, re.I)
+    if step_match:
+        result["step"] = float(step_match.group(1))
     for key in ["loss", "grad_norm", "learning_rate", "epoch"]:
-        match = re.search(rf"['\"]?{key}['\"]?\s*:\s*([-+]?\d+(?:\.\d+)?(?:[eE][-+]?\d+)?)", line)
+        match = re.search(rf"['\"]?{key}['\"]?\s*(?:=|:)\s*{number_pattern}", line, re.I)
         if match:
             result[key] = float(match.group(1))
     return result or None
@@ -452,6 +456,15 @@ class PlatformStore:
             )
             self._ensure_columns(conn, "gpus", {"last_seen_at": "TEXT"})
             self._ensure_columns(conn, "fine_tune_tasks", {"compute_job_id": "TEXT"})
+            self._ensure_columns(
+                conn,
+                "trained_models",
+                {
+                    "artifact_dir": "TEXT",
+                    "compute_node_id": "TEXT",
+                    "compute_node_name": "TEXT",
+                },
+            )
             self._ensure_columns(
                 conn,
                 "resource_replicas",
@@ -589,6 +602,15 @@ class PlatformStore:
         name = task.get("output_model_name") or f"{task['name']}-lora"
         exists = conn.execute("SELECT id FROM trained_models WHERE name=?", (name,)).fetchone()
         if exists:
+            conn.execute(
+                """
+                UPDATE trained_models
+                SET compute_node_id=COALESCE(compute_node_id, ?),
+                    compute_node_name=COALESCE(compute_node_name, ?)
+                WHERE id=?
+                """,
+                (task.get("compute_node_id"), task.get("compute_node_code") or task.get("compute_node_name"), exists["id"]),
+            )
             return
         model = conn.execute("SELECT path FROM models WHERE id=?", (task.get("base_model"),)).fetchone()
         output_dir = task.get("output_dir") or f"/data/yg-ft/outputs/{task['name']}"
@@ -596,8 +618,8 @@ class PlatformStore:
         conn.execute(
             """
             INSERT INTO trained_models
-            (id, name, train_methods, base_model_path, create_time, merged, merging, merged_path, artifact_dir)
+            (id, name, train_methods, base_model_path, create_time, merged, merging, merged_path, artifact_dir, compute_node_id, compute_node_name)
-            VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+            VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
             """,
             (
                 trained_model_id,
@@ -609,6 +631,8 @@ class PlatformStore:
                 0,
                 output_dir,
                 output_dir,
+                task.get("compute_node_id"),
+                task.get("compute_node_code") or task.get("compute_node_name"),
             ),
         )
         # Use real artifact data from compute node when available
@@ -875,7 +899,7 @@ class PlatformStore:
                 (
                     new_id("metric"),
                     task_id,
-                    line_number,
+                    int(metric.get("step") or line_number),
                     metric.get("epoch"),
                     metric.get("loss"),
                     metric.get("grad_norm"),
@@ -1342,15 +1366,25 @@ class PlatformStore:
         self.refresh_runtime_state()
         with self.connect() as conn:
             rows = conn.execute("SELECT * FROM trained_models ORDER BY create_time DESC").fetchall()
-            return [
+            items = []
-                {
+            for row in rows:
+                item = {
                     **dict(row),
                     "train_methods": json_loads(row["train_methods"], []),
                     "merged": bool(row["merged"]),
                     "merging": bool(row["merging"]),
                 }
-                for row in rows
+                if not item.get("compute_node_id"):
-            ]
+                    task_rows = conn.execute("SELECT payload FROM fine_tune_tasks ORDER BY create_time DESC").fetchall()
+                    for task in task_rows:
+                        task_payload = json_loads(task["payload"], {})
+                        output_name = task_payload.get("output_model_name") or f"{task_payload.get('name')}-lora"
+                        if output_name == item["name"]:
+                            item["compute_node_id"] = task_payload.get("compute_node_id")
+                            item["compute_node_name"] = task_payload.get("compute_node_code") or task_payload.get("compute_node_name")
+                            break
+                items.append(item)
+            return items
 
     def delete_trained_model(self, model_id: str) -> None:
         with self.connect() as conn:

backend/app/db/sql/001_platform_runtime.sql

@@ -33,7 +33,10 @@ CREATE TABLE IF NOT EXISTS trained_models (
     create_time TEXT NOT NULL,
     merged INTEGER NOT NULL DEFAULT 0,
     merging INTEGER NOT NULL DEFAULT 0,
-    merged_path TEXT
+    merged_path TEXT,
+    artifact_dir TEXT,
+    compute_node_id TEXT,
+    compute_node_name TEXT
 );
 
 CREATE TABLE IF NOT EXISTS model_lineage (

backend/app/modules/compute_gateway/client.py

@@ -33,6 +33,16 @@ def _unwrap_dict(payload: Any) -> dict[str, Any]:
     return payload if isinstance(payload, dict) else {}
 
 
+# Inference calls are intentionally short-timeout:
+#   - load dispatch only confirms the compute node accepted the request
+#     (the actual model load now runs asynchronously on the node).
+#   - status/unload must never block the platform for long when a node is
+#     unreachable but still marked online.
+INFERENCE_LOAD_TIMEOUT = httpx.Timeout(30, connect=10)
+INFERENCE_STATUS_TIMEOUT = httpx.Timeout(30, connect=5)
+INFERENCE_UNLOAD_TIMEOUT = httpx.Timeout(30, connect=5)
+
+
 class ComputeNodeClient:
     """Application-side client for one compute node.
 
@@ -182,10 +192,16 @@ class ComputeNodeClient:
             response.raise_for_status()
             return _unwrap_dict(response.json())
 
-    async def _request(self, method: str, path: str, json_data: dict[str, Any] | None = None) -> dict[str, Any]:
+    async def _request(
+        self,
+        method: str,
+        path: str,
+        json_data: dict[str, Any] | None = None,
+        timeout: float | None = None,
+    ) -> dict[str, Any]:
         """Generic request method for compute API endpoints."""
         url = _join_url(self.api_base_url, f"{self.route_prefix}{path}")
-        async with httpx.AsyncClient(timeout=300, headers=self.headers()) as client:
+        async with httpx.AsyncClient(timeout=timeout or 300, headers=self.headers()) as client:
             if method.upper() == "GET":
                 response = await client.get(url)
             else:
@@ -193,6 +209,19 @@ class ComputeNodeClient:
             response.raise_for_status()
             return _unwrap_dict(response.json())
 
+    # ── Inference helpers (short timeouts — see module constants) ──────────
+
+    async def inference_load(self, payload: dict[str, Any]) -> dict[str, Any]:
+        """Dispatch a model load. Returns as soon as the node accepts the
+        request; the node now loads asynchronously (status goes 'loading')."""
+        return await self._request("POST", "/inference/load", json_data=payload, timeout=INFERENCE_LOAD_TIMEOUT)
+
+    async def inference_status(self) -> dict[str, Any]:
+        return await self._request("GET", "/inference/status", timeout=INFERENCE_STATUS_TIMEOUT)
+
+    async def inference_unload(self) -> dict[str, Any]:
+        return await self._request("POST", "/inference/unload", json_data={}, timeout=INFERENCE_UNLOAD_TIMEOUT)
+
     async def upload_file(
         self,
         filename: str,

backend/app/modules/compute_gateway/sync.py

@@ -1,15 +1,101 @@
 from __future__ import annotations
 
+import json
+import time
 from typing import Any
 
 from app.db.platform_store import get_platform_store
 from app.modules.compute_gateway.client import ComputeNodeClient
 
+# starting 状态允许的最大轮询次数（约 40 * 3s ≈ 2 分钟），超过即判定节点不可达
+MAX_STARTING_ATTEMPTS = 40
+
 
 def _node_for_task(task: dict[str, Any]) -> dict[str, Any] | None:
     return next((node for node in get_platform_store().compute_nodes() if node["id"] == task.get("compute_node_id")), None)
 
 
+def _parse_inference_load_status(task: dict[str, Any]) -> tuple[list[dict[str, Any]], dict[str, Any]]:
+    load_status = task.get("load_status") or {}
+    if isinstance(load_status, str):
+        try:
+            load_status = json.loads(load_status)
+        except (json.JSONDecodeError, TypeError):
+            load_status = {}
+    return load_status.get("loaded_models") or [], load_status
+
+
+async def reconcile_inference_loads(store: Any) -> list[dict[str, Any]]:
+    """推进处于 starting 状态的推理加载。
+
+    模型加载已改为异步派发：/model-compare/{id}/load 立即返回，这里在每次
+    轮询时查询对应计算节点的 /inference/status，把任务从 starting 推进到
+    ready/error。使用短超时，单节点不可达不会阻塞整轮轮询。
+    """
+    reconciled: list[dict[str, Any]] = []
+    now = time.time()
+    for task in store.compare_tasks():
+        items, _ = _parse_inference_load_status(task)
+        if not any(item.get("status") == "starting" for item in items):
+            continue
+        # dirty 只要处理过任一 starting 项就置位：load_attempts / last_polled_at
+        # 必须落库，否则节点不可达时计数不会累积，封顶逻辑永远触发不了
+        dirty = False
+        for item in items:
+            if item.get("status") != "starting":
+                continue
+            # 节流：同一 item 每 3s 只查询一次
+            if now - float(item.get("last_polled_at") or 0) < 3:
+                continue
+            item["last_polled_at"] = now
+            item["load_attempts"] = int(item.get("load_attempts") or 0) + 1
+            dirty = True
+            node = next((n for n in store.compute_nodes() if n["id"] == item.get("node_id")), None)
+            if not node:
+                item["status"] = "error"
+                item["error"] = "compute node deleted"
+                store.mark_inference_unloaded(item.get("node_id") or "")
+                continue
+            if not node.get("enabled") or node.get("scheduler_status") != "online":
+                item["status"] = "error"
+                item["error"] = "compute node offline"
+                store.mark_inference_unloaded(node["id"])
+                continue
+            try:
+                status = await ComputeNodeClient(node["api_base_url"]).inference_status()
+            except Exception as exc:  # noqa: BLE001 - node unreachable; keep retrying until cap
+                if int(item.get("load_attempts") or 0) >= MAX_STARTING_ATTEMPTS:
+                    item["status"] = "error"
+                    item["error"] = f"compute node unreachable: {exc}"
+                    store.mark_inference_unloaded(node["id"])
+                continue
+            node_status = status.get("status")
+            if node_status == "ready":
+                item["status"] = "ready"
+                item.pop("error", None)
+                store.mark_inference_loaded(node["id"])
+            elif node_status == "error":
+                item["status"] = "error"
+                item["error"] = status.get("error") or "model load failed on compute node"
+                store.mark_inference_unloaded(node["id"])
+            elif node_status == "idle":
+                # 节点重启导致已加载模型丢失
+                item["status"] = "error"
+                item["error"] = "model disappeared from compute node (node may have restarted)"
+                store.mark_inference_unloaded(node["id"])
+            # node_status == "loading" -> 保持 starting，下轮再查
+        if dirty:
+            if any(i.get("status") in {"ready", "running"} for i in items):
+                new_status = "loaded"
+            elif any(i.get("status") == "starting" for i in items):
+                new_status = "starting"  # 仍在加载中，保持 starting
+            else:
+                new_status = "failed"
+            store.update_compare_task(task["id"], {"status": new_status, "load_status": {"loaded_models": items}})
+            reconciled.append({"task_id": task["id"], "status": new_status})
+    return reconciled
+
+
 async def fetch_eval_result_content(client: ComputeNodeClient, node: dict[str, Any], job: dict[str, Any]) -> dict[str, Any] | None:
     output_dir = job.get("output_dir")
     if not output_dir:
@@ -95,5 +181,13 @@ async def poll_compute_jobs_once() -> dict[str, Any]:
         except Exception as exc:  # noqa: BLE001
             failed.append({"eval_task_id": eval_task["id"], "error": str(exc)})
 
+    # ── Inference load reconciliation ─────────────────────────────────────
+    try:
+        inference_reconciled = await reconcile_inference_loads(store)
+    except Exception as exc:  # noqa: BLE001 - keep polling alive
+        failed.append({"inference_reconcile": str(exc)})
+        inference_reconciled = []
+
     return {"synced": len(synced) + len(standalone_synced) + eval_synced, "failed": failed,
-            "items": synced, "standalone": standalone_synced, "eval_synced": eval_synced}
+            "items": synced, "standalone": standalone_synced, "eval_synced": eval_synced,
+            "inference_reconciled": inference_reconciled}

backend/tests/test_compare_inference_async.py

@@ -0,0 +1,254 @@
+"""
+模型推理异步加载改造的单元测试。
+
+覆盖：
+- model_compare_load：异步派发，立即返回 starting + 节点信息（不等待加载完成）
+- model_compare_delete：先删记录，卸载失败也不阻塞删除
+- reconcile_inference_loads：starting -> ready/error/idle/不可达的状态迁移与封顶
+- _unload_from_compute_node：任务感知，只命中记录中的节点
+"""
+
+from __future__ import annotations
+
+import asyncio
+from types import SimpleNamespace
+from typing import Any
+
+from app.api.v1.endpoints.platform import model_compare_delete, model_compare_load
+import app.api.v1.endpoints.platform as platform
+from app.modules.compute_gateway.client import ComputeNodeClient
+from app.modules.compute_gateway.sync import MAX_STARTING_ATTEMPTS, reconcile_inference_loads
+
+
+class FakeInferenceStore:
+    """内存 store，仅实现推理加载/对账用到的接口。"""
+
+    def __init__(self, tasks: list[dict[str, Any]] | None = None, nodes: list[dict[str, Any]] | None = None) -> None:
+        self._tasks: dict[str, dict[str, Any]] = {t["id"]: dict(t) for t in (tasks or [])}
+        self._nodes = nodes or []
+        self._inference_nodes: set[str] = set()
+
+    def compare_task(self, task_id: str) -> dict[str, Any]:
+        if task_id not in self._tasks:
+            raise KeyError(task_id)
+        return dict(self._tasks[task_id])
+
+    def compare_tasks(self) -> list[dict[str, Any]]:
+        return [dict(t) for t in self._tasks.values()]
+
+    def update_compare_task(self, task_id: str, payload: dict[str, Any]) -> dict[str, Any]:
+        current = self._tasks[task_id]
+        merged = {**current, **payload, "id": task_id}
+        self._tasks[task_id] = merged
+        return dict(merged)
+
+    def delete_compare_task(self, task_id: str) -> None:
+        self._tasks.pop(task_id, None)
+
+    def compute_nodes(self) -> list[dict[str, Any]]:
+        return [dict(n) for n in self._nodes]
+
+    def model(self, model_id: str) -> dict[str, Any]:
+        raise KeyError(model_id)
+
+    def trained_models(self) -> list[dict[str, Any]]:
+        return []
+
+    def mark_inference_loaded(self, node_id: str) -> None:
+        self._inference_nodes.add(node_id)
+
+    def mark_inference_unloaded(self, node_id: str) -> None:
+        self._inference_nodes.discard(node_id)
+
+    def is_inference_loaded(self, node_id: str) -> bool:
+        return node_id in self._inference_nodes
+
+
+def _node(node_id: str, code: str = "") -> dict[str, Any]:
+    return {
+        "id": node_id,
+        "code": code or node_id,
+        "name": code or node_id,
+        "api_base_url": f"http://{code or node_id}:19100",
+        "enabled": True,
+        "scheduler_status": "online",
+    }
+
+
+def _task(task_id: str, *, node_id: str | None = None, load_status: dict[str, Any] | None = None) -> dict[str, Any]:
+    return {
+        "id": task_id,
+        "name": f"task-{task_id}",
+        "status": "pending",
+        "models": [
+            {"model_id": "m_1", "model_name": "qwen", "model_path": "/models/qwen", "node_id": node_id}
+        ],
+        "load_status": load_status or {"loaded_models": []},
+    }
+
+
+async def _fake_inference_load(self, payload: dict[str, Any]) -> dict[str, Any]:
+    return {"loaded": False, "status": "loading", "request_id": "req-1"}
+
+
+async def _fake_inference_unload(self) -> dict[str, Any]:
+    return {"unloaded": True, "status": "idle"}
+
+
+def _patch_store(monkeypatch, store: FakeInferenceStore) -> None:
+    monkeypatch.setattr(platform, "get_platform_store", lambda: store)
+    monkeypatch.setattr(platform, "get_settings", lambda: SimpleNamespace(compute_mode="real"))
+
+
+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)
+    monkeypatch.setattr(ComputeNodeClient, "inference_load", _fake_inference_load)
+
+    result = asyncio.run(model_compare_load("t1"))
+    assert result["code"] == 0
+    updated = result["data"]
+    assert updated["status"] == "starting"
+    items = updated["load_status"]["loaded_models"]
+    assert items[0]["status"] == "starting"
+    assert items[0]["node_id"] == "n1"
+    assert "n1" in store._inference_nodes
+
+
+def test_model_compare_load_marks_error_when_all_nodes_fail(monkeypatch) -> None:
+    store = FakeInferenceStore(tasks=[_task("t1", node_id="n1")], nodes=[_node("n1")])
+    _patch_store(monkeypatch, store)
+
+    async def _raise(self, payload: dict[str, Any]) -> dict[str, Any]:
+        raise RuntimeError("conn refused")
+
+    monkeypatch.setattr(ComputeNodeClient, "inference_load", _raise)
+
+    result = asyncio.run(model_compare_load("t1"))
+    updated = result["data"]
+    assert updated["status"] == "failed"
+    assert updated["load_status"]["loaded_models"][0]["status"] == "error"
+    assert "conn refused" in updated["load_status"]["loaded_models"][0]["error"]
+
+
+def test_model_compare_delete_removes_record_even_if_unload_raises(monkeypatch) -> None:
+    task = _task(
+        "t1",
+        node_id="n1",
+        load_status={"loaded_models": [{"model_id": "m_1", "status": "ready", "node_id": "n1"}]},
+    )
+    store = FakeInferenceStore(tasks=[task], nodes=[_node("n1")])
+    _patch_store(monkeypatch, store)
+
+    async def _raise(self) -> dict[str, Any]:
+        raise RuntimeError("unload boom")
+
+    monkeypatch.setattr(ComputeNodeClient, "inference_unload", _raise)
+
+    result = asyncio.run(model_compare_delete("t1"))
+    assert result["data"] == {"deleted": "t1"}
+    assert "t1" not in store._tasks
+    # finally 中仍清掉了节点标记
+    assert "n1" not in store._inference_nodes
+
+
+def test_unload_from_compute_node_only_hits_recorded_node(monkeypatch) -> None:
+    task = _task(
+        "t1",
+        load_status={"loaded_models": [{"model_id": "m_1", "status": "ready", "node_id": "n1"}]},
+    )
+    store = FakeInferenceStore(tasks=[task], nodes=[_node("n1"), _node("n2")])
+    _patch_store(monkeypatch, store)
+    monkeypatch.setattr(ComputeNodeClient, "inference_unload", _fake_inference_unload)
+
+    from app.api.v1.endpoints.platform import _unload_from_compute_node
+
+    result = asyncio.run(_unload_from_compute_node(store, task=task))
+    assert result["unloaded"] is True
+    # 只命中任务记录中的节点 n1，n2 未被卸载
+    assert [r["node_id"] for r in result["nodes"]] == ["n1"]
+    assert "n1" not in store._inference_nodes
+
+
+async def _status_ready(self) -> dict[str, Any]:
+    return {"loaded": True, "status": "ready", "model_name": "qwen"}
+
+
+def test_reconcile_transitions_starting_to_ready(monkeypatch) -> None:
+    task = _task(
+        "t1",
+        node_id="n1",
+        load_status={"loaded_models": [{"model_id": "m_1", "status": "starting", "node_id": "n1"}]},
+    )
+    store = FakeInferenceStore(tasks=[task], nodes=[_node("n1")])
+    monkeypatch.setattr(ComputeNodeClient, "inference_status", _status_ready)
+
+    reconciled = asyncio.run(reconcile_inference_loads(store))
+    assert reconciled == [{"task_id": "t1", "status": "loaded"}]
+    updated = store._tasks["t1"]
+    assert updated["status"] == "loaded"
+    assert updated["load_status"]["loaded_models"][0]["status"] == "ready"
+    assert "n1" in store._inference_nodes
+
+
+def test_reconcile_transitions_to_error_and_failed(monkeypatch) -> None:
+    async def _status_error(self) -> dict[str, Any]:
+        return {"loaded": False, "status": "error", "error": "CUDA out of memory"}
+
+    task = _task(
+        "t1",
+        node_id="n1",
+        load_status={"loaded_models": [{"model_id": "m_1", "status": "starting", "node_id": "n1"}]},
+    )
+    store = FakeInferenceStore(tasks=[task], nodes=[_node("n1")])
+    monkeypatch.setattr(ComputeNodeClient, "inference_status", _status_error)
+
+    reconciled = asyncio.run(reconcile_inference_loads(store))
+    assert reconciled == [{"task_id": "t1", "status": "failed"}]
+    item = store._tasks["t1"]["load_status"]["loaded_models"][0]
+    assert item["status"] == "error"
+    assert "CUDA out of memory" in item["error"]
+    assert "n1" not in store._inference_nodes
+
+
+def test_reconcile_idle_marks_model_disappeared(monkeypatch) -> None:
+    async def _status_idle(self) -> dict[str, Any]:
+        return {"loaded": False, "status": "idle"}
+
+    task = _task(
+        "t1",
+        node_id="n1",
+        load_status={"loaded_models": [{"model_id": "m_1", "status": "starting", "node_id": "n1"}]},
+    )
+    store = FakeInferenceStore(tasks=[task], nodes=[_node("n1")])
+    monkeypatch.setattr(ComputeNodeClient, "inference_status", _status_idle)
+
+    asyncio.run(reconcile_inference_loads(store))
+    item = store._tasks["t1"]["load_status"]["loaded_models"][0]
+    assert item["status"] == "error"
+    assert "disappeared" in item["error"]
+    assert store._tasks["t1"]["status"] == "failed"
+
+
+def test_reconcile_unreachable_node_flips_to_error_after_cap(monkeypatch) -> None:
+    async def _raise(self) -> dict[str, Any]:
+        raise RuntimeError("conn refused")
+
+    task = _task(
+        "t1",
+        node_id="n1",
+        load_status={"loaded_models": [{"model_id": "m_1", "status": "starting", "node_id": "n1"}]},
+    )
+    store = FakeInferenceStore(tasks=[task], nodes=[_node("n1")])
+    monkeypatch.setattr(ComputeNodeClient, "inference_status", _raise)
+
+    # 每次轮询前重置节流时间戳，逐次推进 load_attempts 到封顶
+    for _ in range(MAX_STARTING_ATTEMPTS):
+        item = store._tasks["t1"]["load_status"]["loaded_models"][0]
+        item["last_polled_at"] = 0
+        asyncio.run(reconcile_inference_loads(store))
+
+    item = store._tasks["t1"]["load_status"]["loaded_models"][0]
+    assert item["status"] == "error"
+    assert "unreachable" in item["error"]
+    assert store._tasks["t1"]["status"] == "failed"

compute/api/main.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import asyncio
 import json
 import os
 import math
@@ -717,7 +718,9 @@ def create_app() -> FastAPI:
     @app.post(f"{route_prefix}/inference/unload")
     async def inference_unload() -> dict[str, Any]:
         """Unload the currently loaded model and free GPU memory."""
-        return get_inference_session().unload()
+        # Teardown (gc.collect + cuda.empty_cache) can take a while; run it off
+        # the event loop so /health and /inference/status stay responsive.
+        return await asyncio.to_thread(get_inference_session().unload)
 
     @app.get(f"{route_prefix}/inference/status")
     async def inference_status() -> dict[str, Any]:
@@ -737,7 +740,10 @@ def create_app() -> FastAPI:
         messages = payload.get("messages") or []
         if not messages:
             raise HTTPException(status_code=400, detail="messages is required")
-        result = get_inference_session().chat(
+        # Generation is long-running; run it in a thread so the event loop keeps
+        # serving /inference/status and /health during inference.
+        result = await asyncio.to_thread(
+            get_inference_session().chat,
             messages=messages,
             temperature=float(payload.get("temperature", 0.95)),
             top_p=float(payload.get("top_p", 0.7)),

compute/engines/llama_factory/inference.py

@@ -2,114 +2,208 @@ from __future__ import annotations
 
 import threading
 import time
-from typing import Any
+import uuid
+from typing import Any, Iterator
 
 
 class InferenceSession:
-    """Manages a loaded model for inference with LLaMA-Factory ChatModel."""
+    """Manages a loaded model for inference with LLaMA-Factory ChatModel.
+
+    Model loading is asynchronous: ``load()`` spawns a background daemon thread
+    and returns immediately with ``status == "loading"``. ``info()`` (served by
+    ``/inference/status``) is always responsive, so the platform backend can
+    poll loading progress without being blocked by a minutes-long model load —
+    which previously froze the whole compute node event loop.
+
+    State machine: idle -> loading -> ready | error, ready -> idle (unload),
+    loading -> idle (cancelled). Long operations (ChatModel build, teardown,
+    generation) never run while holding ``_state_lock``; they either run in the
+    worker thread or under ``_chat_lock`` only.
+    """
 
     def __init__(self) -> None:
+        self._state_lock = threading.Lock()  # brief state transitions only
+        self._chat_lock = threading.Lock()  # serialize chat/teardown
+        self._status: str = "idle"
+        self._error: str = ""
+        self._request_id: str = ""
+        self._load_args: dict[str, Any] = {}
+        self._teardown_old = False  # load-while-ready: unload old before loading new
+        self._cancel_requested = False  # unload-while-loading: tear down after load finishes
+        self._load_thread: threading.Thread | None = None
         self._model: Any = None
         self._tokenizer: Any = None
         self._generating_args: dict[str, Any] = {}
         self._model_name: str = ""
         self._adapter_path: str = ""
-        self._lock = threading.Lock()
         self._loaded_at: float = 0.0
-        self._status: str = "idle"
 
     @property
     def status(self) -> str:
+        with self._state_lock:
             return self._status
 
-    @property
-    def model_name(self) -> str:
-        return self._model_name
-
-    @property
-    def adapter_path(self) -> str:
-        return self._adapter_path
-
-    @property
-    def loaded_at(self) -> float:
-        return self._loaded_at
-
     def info(self) -> dict[str, Any]:
+        with self._state_lock:
             return {
                 "loaded": self._status == "ready",
                 "status": self._status,
                 "model_name": self._model_name,
                 "adapter_path": self._adapter_path,
                 "loaded_at": self._loaded_at,
+                "request_id": self._request_id,
+                "error": self._error,
             }
 
-    def load(self, model_name_or_path, adapter_name_or_path="", template="qwen", infer_backend="huggingface", infer_dtype="auto", **kwargs):
+    def load(
-        with self._lock:
+        self,
+        model_name_or_path,
+        adapter_name_or_path="",
+        template="qwen",
+        infer_backend="huggingface",
+        infer_dtype="auto",
+        **kwargs,
+    ) -> dict[str, Any]:
+        with self._state_lock:
             if self._status == "loading":
-                return {"loaded": False, "error": "model is already loading"}
+                # A model is already loading — dedupe, reuse the same request id.
-            if self._status == "ready":
+                return {"loaded": False, "status": "loading", "request_id": self._request_id}
-                self.unload()
+            self._teardown_old = self._status == "ready"
             self._status = "loading"
+            self._error = ""
+            self._request_id = uuid.uuid4().hex[:12]
+            self._cancel_requested = False
+            self._load_args = {
+                "model_name_or_path": model_name_or_path,
+                "template": template,
+                "infer_backend": infer_backend,
+                "infer_dtype": infer_dtype,
+            }
+            if adapter_name_or_path:
+                self._load_args["adapter_name_or_path"] = adapter_name_or_path
+            self._load_args.update(kwargs)
             self._model_name = model_name_or_path
             self._adapter_path = adapter_name_or_path
+            self._load_thread = threading.Thread(target=self._load_worker, daemon=True)
+            self._load_thread.start()
+            return {"loaded": False, "status": "loading", "request_id": self._request_id}
+
+    def _load_worker(self) -> None:
+        """Build the ChatModel off the state lock so info() never blocks."""
+        model = None
+        tokenizer = None
+        generating_args: dict[str, Any] = {}
+        error = ""
         try:
+            if self._teardown_old:
+                self._release_model()
             from llamafactory.chat import ChatModel
             from llamafactory.hparams import get_infer_args
-                args = {"model_name_or_path": model_name_or_path, "template": template, "infer_backend": infer_backend, "infer_dtype": infer_dtype}
-                if adapter_name_or_path:
-                    args["adapter_name_or_path"] = adapter_name_or_path
-                args.update(kwargs)
-                infer_result = get_infer_args(args)
-                # ChatModel internally re-parses the args dict via get_infer_args,
-                # so pass the original args (not the parsed dataclass objects).
-                self._model = ChatModel(args)
-                self._tokenizer = getattr(self._model, 'tokenizer', None) or self._model.engine.tokenizer
-                # Extract generating_args (last element) for later use in chat()
-                generating_args = infer_result[-1]
-                if hasattr(generating_args, '__dataclass_fields__'):
-                    self._generating_args = {k: v for k, v in vars(generating_args).items()
-                                              if not k.startswith('_')}
-                else:
-                    self._generating_args = dict(generating_args)
-                self._loaded_at = time.time()
-                self._status = "ready"
-                return {"loaded": True, "status": "ready"}
-            except Exception as exc:
-                self._status = "error"
-                self._model = None
-                return {"loaded": False, "status": "error", "error": str(exc)}
 
-    def unload(self):
+            args = dict(self._load_args)
-        with self._lock:
+            infer_result = get_infer_args(args)
-            if self._model is not None:
+            model = ChatModel(args)
-                try:
+            tokenizer = getattr(model, "tokenizer", None) or model.engine.tokenizer
-                    del self._model
+            generating_args = infer_result[-1]
-                except Exception:
+            if hasattr(generating_args, "__dataclass_fields__"):
-                    pass
+                generating_args = {
+                    k: v for k, v in vars(generating_args).items() if not k.startswith("_")
+                }
+            else:
+                generating_args = dict(generating_args)
+        except Exception as exc:  # noqa: BLE001 - surface load failure via status
+            error = str(exc)
+        with self._state_lock:
+            if error:
                 self._model = None
                 self._tokenizer = None
+                self._status = "error"
+                self._error = error
+                return
+            if self._cancel_requested:
+                # Unload was requested while loading — drop the fresh model.
+                model = None
+                tokenizer = None
+                self._model = None
+                self._tokenizer = None
+                self._status = "idle"
+                return
+            self._model = model
+            self._tokenizer = tokenizer
+            self._generating_args = generating_args
+            self._loaded_at = time.time()
+            self._status = "ready"
+
+    def _release_model(self) -> None:
+        with self._chat_lock:
+            with self._state_lock:
+                self._status = "unloading"
+                model = self._model
+                self._model = None
+                self._tokenizer = None
+            if model is not None:
+                try:
+                    del model
+                except Exception:  # noqa: BLE001 - best-effort teardown
+                    pass
                 # 强制释放 PyTorch CUDA 缓存，真正归还 GPU 显存
                 try:
                     import gc
+
                     gc.collect()
                     import torch
+
                     if torch.cuda.is_available():
                         torch.cuda.empty_cache()
                         torch.cuda.synchronize()
-                except Exception:
+                except Exception:  # noqa: BLE001 - teardown must not raise
                     pass
+            with self._state_lock:
                 self._status = "idle"
                 self._model_name = ""
                 self._adapter_path = ""
                 self._loaded_at = 0.0
-            return {"unloaded": True}
+                self._error = ""
 
-    def chat(self, messages, temperature=0.95, top_p=0.7, max_new_tokens=1024, do_sample=True, **kwargs):
+    def unload(self) -> dict[str, Any]:
-        with self._lock:
+        with self._state_lock:
+            if self._status == "loading":
+                # Ask the worker to tear down right after the load finishes.
+                self._cancel_requested = True
+                return {"unloaded": False, "status": "cancelling", "request_id": self._request_id}
+            was_ready = self._status == "ready"
+        if was_ready:
+            self._release_model()
+        else:
+            with self._state_lock:
+                self._model = None
+                self._tokenizer = None
+                self._status = "idle"
+                self._model_name = ""
+                self._adapter_path = ""
+                self._loaded_at = 0.0
+                self._error = ""
+        return {"unloaded": True, "status": "idle"}
+
+    def chat(self, messages, temperature=0.95, top_p=0.7, max_new_tokens=1024, do_sample=True, **kwargs) -> dict[str, Any]:
+        with self._chat_lock:
+            with self._state_lock:
+                if self._status == "loading":
+                    return {
+                        "error": f"model is still loading (request_id={self._request_id}); please retry",
+                        "response": "",
+                    }
+                if self._status == "error":
+                    return {"error": f"model load failed: {self._error}", "response": ""}
                 if self._status != "ready" or self._model is None:
                     return {"error": "model not loaded", "response": ""}
             try:
-                generate_kwargs = {"temperature": temperature, "top_p": top_p, "max_new_tokens": max_new_tokens, "do_sample": do_sample}
+                generate_kwargs = {
+                    "temperature": temperature,
+                    "top_p": top_p,
+                    "max_new_tokens": max_new_tokens,
+                    "do_sample": do_sample,
+                }
                 generate_kwargs.update(kwargs)
                 system = next((m["content"] for m in messages if m["role"] == "system"), None)
                 user_messages = [m for m in messages if m["role"] != "system"]
@@ -118,11 +212,18 @@ class InferenceSession:
                     responses.append(response)
                 full_response = "".join(str(r) for r in responses)
                 return {"response": full_response}
-            except Exception as exc:
+            except Exception as exc:  # noqa: BLE001 - return generation error to caller
                 return {"error": str(exc), "response": ""}
 
-    def chat_stream(self, messages, **kwargs):
+    def chat_stream(self, messages, **kwargs) -> Iterator[str]:
-        with self._lock:
+        with self._chat_lock:
+            with self._state_lock:
+                if self._status == "loading":
+                    yield 'data: {"error": "model is still loading; please retry"}\n\n'
+                    return
+                if self._status == "error":
+                    yield 'data: {"error": "model load failed: ' + str(self._error) + '"}\n\n'
+                    return
                 if self._status != "ready" or self._model is None:
                     yield 'data: {"error": "model not loaded"}\n\n'
                     return
@@ -132,13 +233,14 @@ class InferenceSession:
                 user_messages = [m for m in messages if m["role"] != "system"]
                 for new_text in self._model.stream_chat(user_messages, system=system, **generate_kwargs):
                     yield new_text
-            except Exception as exc:
+            except Exception as exc:  # noqa: BLE001 - stream error as SSE event
                 yield 'data: {"error": "' + str(exc) + '"}\n\n'
 
 
 _inference_session = None
 
-def get_inference_session():
+
+def get_inference_session() -> InferenceSession:
     global _inference_session
     if _inference_session is None:
         _inference_session = InferenceSession()

compute/tests/test_inference_session.py

@@ -0,0 +1,127 @@
+from __future__ import annotations
+
+import sys
+import time
+import types
+from typing import Any
+
+import pytest
+
+from compute.engines.llama_factory.inference import InferenceSession
+
+# 模拟模型加载耗时，用于验证 load() 立即返回、info() 不阻塞
+LOAD_DELAY = 0.2
+
+
+class FakeChatModel:
+    def __init__(self, args: dict[str, Any]) -> None:
+        time.sleep(LOAD_DELAY)
+        self.tokenizer = object()
+        self.engine = types.SimpleNamespace(tokenizer=object())
+        self._output = "hello from model"
+
+    def stream_chat(self, *args, **kwargs):
+        for _ in range(1):
+            yield self._output
+
+
+class FailingChatModel:
+    def __init__(self, args: dict[str, Any]) -> None:
+        time.sleep(LOAD_DELAY)
+        raise RuntimeError("boom: fake load failure")
+
+
+def _get_infer_args(args: dict[str, Any]) -> list[Any]:
+    # 最后一个元素为 generating_args，worker 会转成 dict
+    return [None, None, {"temperature": 0.7}]
+
+
+def _install_llamafactory(monkeypatch, chat_model: type) -> None:
+    llmf = types.ModuleType("llamafactory")
+    chat_mod = types.ModuleType("llamafactory.chat")
+    hparams_mod = types.ModuleType("llamafactory.hparams")
+    chat_mod.ChatModel = chat_model
+    hparams_mod.get_infer_args = _get_infer_args
+    llmf.chat = chat_mod
+    llmf.hparams = hparams_mod
+    monkeypatch.setitem(sys.modules, "llamafactory", llmf)
+    monkeypatch.setitem(sys.modules, "llamafactory.chat", chat_mod)
+    monkeypatch.setitem(sys.modules, "llamafactory.hparams", hparams_mod)
+
+
+@pytest.fixture
+def stub_llamafactory(monkeypatch) -> None:
+    _install_llamafactory(monkeypatch, FakeChatModel)
+
+
+@pytest.fixture
+def stub_failing_llamafactory(monkeypatch) -> None:
+    _install_llamafactory(monkeypatch, FailingChatModel)
+
+
+def _wait_for_status(session: InferenceSession, status: str, timeout: float = 3.0) -> bool:
+    deadline = time.time() + timeout
+    while time.time() < deadline:
+        if session.info()["status"] == status:
+            return True
+        time.sleep(0.02)
+    return False
+
+
+def test_load_returns_immediately_then_ready(stub_llamafactory) -> None:
+    session = InferenceSession()
+    started = time.time()
+    result = session.load("/models/qwen")
+    assert result["status"] == "loading"
+    assert result["loaded"] is False
+    assert result["request_id"]
+    # 在慢加载完成前就返回，且 info() 加载期间可响应
+    assert time.time() - started < LOAD_DELAY
+    assert session.info()["status"] == "loading"
+    assert _wait_for_status(session, "ready")
+    info = session.info()
+    assert info["loaded"] is True
+    assert info["status"] == "ready"
+    assert info["model_name"] == "/models/qwen"
+
+
+def test_second_load_while_loading_deduped(stub_llamafactory) -> None:
+    session = InferenceSession()
+    r1 = session.load("/models/a")
+    r2 = session.load("/models/b")
+    assert r2["status"] == "loading"
+    assert r2["request_id"] == r1["request_id"]
+    assert _wait_for_status(session, "ready")
+    assert session.info()["status"] == "ready"
+
+
+def test_load_error_surfaces_in_status(stub_failing_llamafactory) -> None:
+    session = InferenceSession()
+    session.load("/models/bad")
+    assert _wait_for_status(session, "error")
+    assert "boom" in session.info()["error"]
+
+
+def test_unload_while_loading_cancels(stub_llamafactory) -> None:
+    session = InferenceSession()
+    session.load("/models/qwen")
+    result = session.unload()
+    assert result["status"] == "cancelling"
+    assert _wait_for_status(session, "idle")
+
+
+def test_chat_while_loading_returns_loading_error(stub_llamafactory) -> None:
+    session = InferenceSession()
+    session.load("/models/qwen")
+    out = session.chat([{"role": "user", "content": "hi"}])
+    assert "still loading" in (out.get("error") or "")
+    assert _wait_for_status(session, "ready")
+    out = session.chat([{"role": "user", "content": "hi"}])
+    assert out.get("response") == "hello from model"
+
+
+def test_chat_stream_while_loading_yields_error(stub_llamafactory) -> None:
+    session = InferenceSession()
+    session.load("/models/qwen")
+    chunks = list(session.chat_stream([{"role": "user", "content": "hi"}]))
+    assert any("still loading" in c for c in chunks)

docker/nginx.conf.template

@@ -14,8 +14,8 @@ server {
     proxy_set_header X-Real-IP $remote_addr;
     proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
     proxy_set_header X-Forwarded-Proto $scheme;
-    proxy_read_timeout 300s;
+    proxy_read_timeout 900s;
-    proxy_send_timeout 300s;
+    proxy_send_timeout 900s;
   }
 
   location = /modelTF {
@@ -25,8 +25,8 @@ server {
     proxy_set_header X-Real-IP $remote_addr;
     proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
     proxy_set_header X-Forwarded-Proto $scheme;
-    proxy_read_timeout 300s;
+    proxy_read_timeout 900s;
-    proxy_send_timeout 300s;
+    proxy_send_timeout 900s;
   }
 
   location ~* \.(?:js|css|png|jpg|jpeg|gif|ico|svg|woff|woff2|ttf)$ {

frontend/src/api/modules/compare.ts

@@ -1,6 +1,8 @@
 import { get, post, del } from '../request'
 import type { CompareTask, CompareModelRef } from '@/types'
 
+const INFERENCE_START_TIMEOUT_MS = 15 * 60 * 1000
+
 /** 推理/对比任务列表 */
 export const getCompareList = () => get<CompareTask[]>('/model-compare')
 
@@ -12,7 +14,7 @@ export const createCompare = (data: Partial<CompareTask>) =>
   post<{ id: string | number }>('/model-compare', data)
 
 /** 删除任务 */
-export const deleteCompare = (id: string | number) => del(`/model-compare/${id}`)
+export const deleteCompare = (id: string | number) => del(`/model-compare/${id}`, undefined, { timeout: 60_000 })
 
 /** 更新任务加载状态 */
 export const updateLoadStatus = (id: string | number, load_status: any) =>
@@ -34,7 +36,8 @@ export const stopModelByPid = (pid: number) =>
   post('/model-compare/stop-by-pid', { pid })
 
 /** 加载任务 */
-export const loadCompare = (id: string | number) => post(`/model-compare/${id}/load`)
+export const loadCompare = (id: string | number) =>
+  post(`/model-compare/${id}/load`, undefined, { timeout: INFERENCE_START_TIMEOUT_MS })
 
 /** 卸载任务 */
 export const unloadCompare = (id: string | number) => post(`/model-compare/${id}/unload`)
@@ -81,6 +84,10 @@ export const streamChatReal = (data: any): Promise<Response> => {
       temperature: data.temperature ?? 0.7,
       top_p: data.top_p ?? 0.95,
       max_tokens: data.max_tokens ?? 2048,
+      // 透传 task_id/node_id，让后端按 load_status 路由到真正加载了模型的算力节点，
+      // 避免在多节点时回退到“第一个在线节点”导致连接失败
+      task_id: data.task_id,
+      node_id: data.node_id,
     }),
   })
 }
@@ -94,8 +101,8 @@ export const batchChat = (data: any) => post('/model-chat/batch', data)
 /** 本地 transformers 模型对话 */
 export const localChat = (data: any) => post('/model-chat/local/chat', data)
 
-/** 预加载本地模型（模型加载耗时长，超时 5 分钟） */
+/** 预加载本地模型（模型加载耗时长，超时 15 分钟） */
-export const preloadLocalModel = (data: any) => post('/model-chat/local/preload', data, { timeout: 300000 })
+export const preloadLocalModel = (data: any) => post('/model-chat/local/preload', data, { timeout: INFERENCE_START_TIMEOUT_MS })
 
-/** 预加载已训练模型（超时 5 分钟） */
+/** 预加载已训练模型（超时 15 分钟） */
-export const preloadTrainedModel = (data: any) => post('/model-chat/trained/preload', data, { timeout: 300000 })
+export const preloadTrainedModel = (data: any) => post('/model-chat/trained/preload', data, { timeout: INFERENCE_START_TIMEOUT_MS })

frontend/src/api/modules/fineTune.ts

@@ -1,6 +1,16 @@
 import { get, post, put, del } from '../request'
 import type { FineTuneStartPayload, FineTuneTask, TrainingProgress, LogContent } from '@/types'
 
+export interface FineTuneMetricPoint {
+  step: number
+  epoch?: number | null
+  loss?: number | null
+  grad_norm?: number | null
+  learning_rate?: number | null
+  raw?: string
+  create_time?: string
+}
+
 export interface TrainingDiagnostic {
   level: string
   title: string
@@ -80,6 +90,10 @@ export const getFineTuneLogs = (
   params: { tail_lines?: number; offset?: number; limit?: number } = {},
 ) => get<LogContent & { job_id?: string; source?: string }>(`/fine-tune/${id}/logs`, params)
 
+/** 获取训练指标曲线数据 */
+export const getFineTuneMetrics = (id: string | number) =>
+  get<FineTuneMetricPoint[]>(`/fine-tune/${id}/metrics`)
+
 /** 启动 TensorBoard */
 export const startTensorboard = () => post('/fine-tune/tensorboard/start')
 

frontend/src/api/modules/model.ts

@@ -88,10 +88,14 @@ export const updateModelPurpose = (id: string | number, purpose: string) =>
 
 /** 合并 LoRA 权重 */
 export const mergeModel = (data: {
+  trained_model_id?: string | number
   model_name: string
   train_method: string
   base_model_path: string
-}) => post('/model-manage/merge', data)
+  adapter_path?: string
+  compute_node_id?: string
+  output_model_name?: string
+}) => post('/model-manage/merge', data, { timeout: 15 * 60 * 1000 })
 
 /** 导出已训练模型权重 */
 export const exportModelUrl = (modelName: string) =>

frontend/src/composables/useStreamChat.ts

@@ -44,6 +44,24 @@ export function useStreamChat() {
   })
   const loading = ref(false)
 
+  /** 从 SSE 帧中提取错误信息（后端/计算节点错误以 data: {"error": "..."} 形式下发） */
+  function extractSseError(buffer: string): string | null {
+    const trimmed = buffer.trim()
+    if (!trimmed.startsWith('data: ')) return null
+    const lines = trimmed.split(/\r?\n/)
+    for (let i = lines.length - 1; i >= 0; i--) {
+      const line = lines[i].trim()
+      if (!line.startsWith('data: ')) continue
+      try {
+        const obj = JSON.parse(line.slice(6))
+        if (obj && typeof obj.error === 'string' && obj.error) return obj.error
+      } catch {
+        /* 非 JSON 的 data 行忽略 */
+      }
+    }
+    return trimmed
+  }
+
   /** 从内容中解析 think 标签 */
   function parseContent(content: string) {
     const thinkRegex = /<think>([\s\S]*?)(<\/think>)?/g
@@ -121,6 +139,16 @@ export function useStreamChat() {
       }
 
       // 最终更新
+      // 若整段响应是 SSE 错误帧，提取 error 字段以干净文案展示
+      const sseError = extractSseError(buffer)
+      if (sseError) {
+        message.value.isThinking = false
+        message.value.isStreaming = false
+        message.value.done = true
+        message.value.error = sseError
+        message.value.displayContent = sseError
+        return
+      }
       const parsed = parseContent(buffer)
       message.value.thinkContent = parsed.think
       message.value.displayContent = parsed.display

frontend/src/types/index.ts

@@ -34,6 +34,10 @@ export interface TrainedModel {
   name: string
   train_methods?: TrainMethod[]
   base_model_path?: string
+  artifact_dir?: string
+  adapter_path?: string
+  compute_node_id?: string
+  compute_node_name?: string
   create_time?: string
   merged?: boolean
   merging?: boolean
@@ -212,6 +216,9 @@ export interface LoadedModel {
   status?: string
   pid?: number
   port?: number
+  node_id?: string
+  node_name?: string
+  error?: string
 }
 
 export interface CompareTask {
@@ -230,6 +237,8 @@ export interface CompareModelRef {
   model_name: string
   model_path: string
   gpu_id: number
+  node_id?: string
+  node_name?: string
   source?: string
   port?: number
 }

frontend/src/views/inference/InferenceChatView.vue

@@ -1,10 +1,10 @@
 <script setup lang="ts">
-import { ref, reactive, nextTick, onMounted, watch } from 'vue'
+import { ref, reactive, nextTick, onMounted, onUnmounted, watch } from 'vue'
 import { useRoute, useRouter } from 'vue-router'
 import { ElMessage } from 'element-plus'
 import MarkdownView from '@/components/MarkdownView.vue'
 import { useStreamChat } from '@/composables/useStreamChat'
-import { getCompare } from '@/api/modules/compare'
+import { getCompare, getLoadStatus } from '@/api/modules/compare'
 import type { CompareTask, LoadedModel } from '@/types'
 
 const route = useRoute()
@@ -37,6 +37,10 @@ const contentRef = ref<HTMLElement>()
 let activeAssistant: ChatMessage | null = null
 /** 设置面板抽屉 */
 const showSettings = ref(false)
+/** 模型仍在加载中（直接 URL 进入 chat 时兜底轮询就绪状态） */
+const taskLoading = ref(false)
+const taskError = ref('')
+let statusTimer: ReturnType<typeof setInterval> | null = null
 
 /** 获取任务信息，定位已启动的模型（mock 模式跳过） */
 async function loadTask() {
@@ -45,6 +49,14 @@ async function loadTask() {
     task.value = await getCompare(taskId)
     const models = parseLoadedModels(task.value)
     if (models[0]?.model_name) modelName.value = models[0].model_name
+    // 恢复本地保存的历史对话
+    restoreHistory()
+    // 模型仍在上次加载中：启动轮询等待就绪
+    if (models.some((m) => m.status === 'starting')) {
+      taskLoading.value = true
+      await pollTaskStatus()
+      statusTimer = setInterval(pollTaskStatus, 3000)
+    }
   } catch {
     // ignore
   }
@@ -60,6 +72,80 @@ function parseLoadedModels(t: CompareTask | null): LoadedModel[] {
   }
 }
 
+/** 对话历史本地持久化（按任务 id 存储，退出重进可恢复） */
+const STORAGE_PREFIX = 'ygft_chat_history_'
+
+function historyKey(id: string | number): string {
+  return `${STORAGE_PREFIX}${id}`
+}
+
+function saveHistory() {
+  if (isMock) return
+  try {
+    const snapshot = messages.value.map((m) => ({
+      role: m.role,
+      content: m.content,
+      think: m.think,
+      done: true,
+    }))
+    localStorage.setItem(historyKey(taskId), JSON.stringify(snapshot))
+  } catch {
+    // 存储失败忽略
+  }
+}
+
+function restoreHistory() {
+  if (isMock) return
+  try {
+    const raw = localStorage.getItem(historyKey(taskId))
+    if (!raw) return
+    const parsed = JSON.parse(raw)
+    if (Array.isArray(parsed)) {
+      messages.value = parsed.map((m) => ({
+        role: m.role === 'user' ? 'user' : 'assistant',
+        content: m.content || '',
+        think: m.think || '',
+        isThinking: false,
+        isStreaming: false,
+        done: true,
+      }))
+    }
+  } catch {
+    // 恢复失败忽略
+  }
+}
+
+/** 停止就绪状态轮询 */
+function stopStatusPolling() {
+  if (statusTimer) {
+    clearInterval(statusTimer)
+    statusTimer = null
+  }
+}
+
+/** 轮询任务加载状态：starting → ready/error */
+async function pollTaskStatus() {
+  try {
+    const st = await getLoadStatus(taskId)
+    const items = st.loaded_models || []
+    const anyReady = items.some((m) => m.status === 'ready' || m.status === 'running')
+    const anyError = items.some((m) => m.status === 'error')
+    if (anyReady) {
+      taskLoading.value = false
+      taskError.value = ''
+      stopStatusPolling()
+    } else if (anyError) {
+      taskLoading.value = false
+      taskError.value = items.find((m) => m.status === 'error')?.error || '模型加载失败'
+      stopStatusPolling()
+    } else {
+      taskLoading.value = true
+    }
+  } catch {
+    // 轮询失败忽略，下次再试
+  }
+}
+
 async function handleSend() {
   const question = inputQuestion.value.trim()
   if (!question || loading.value) return
@@ -76,6 +162,7 @@ async function handleSend() {
     done: false,
   })
   messages.value.push(assistantMsg)
+  saveHistory()
 
   inputQuestion.value = ''
   await nextTick()
@@ -85,6 +172,7 @@ async function handleSend() {
   // mock 模式：直接用假数据逐字填充
   if (isMock) {
     await mockReply(assistantMsg, question)
+    saveHistory()
     return
   }
 
@@ -94,6 +182,7 @@ async function handleSend() {
   await send(
     {
       model_path: route.query.model_path as string || '',
+      task_id: taskId,
       system_prompt: systemPrompt.value,
       user_question: question,
       temperature: temperature.value,
@@ -111,6 +200,7 @@ async function handleSend() {
   assistantMsg.done = true
   activeAssistant = null
   reset()
+  saveHistory()
   await nextTick()
   scrollToBottom()
 }
@@ -169,6 +259,11 @@ function handleNewChat() {
   activeAssistant = null
   messages.value = []
   reset()
+  try {
+    localStorage.removeItem(historyKey(taskId))
+  } catch {
+    // 忽略
+  }
 }
 
 /** 输入框自适应高度 */
@@ -185,6 +280,7 @@ function resetInputHeight() {
 }
 
 onMounted(loadTask)
+onUnmounted(stopStatusPolling)
 </script>
 
 <template>
@@ -252,6 +348,12 @@ onMounted(loadTask)
 
     <!-- 输入栏 -->
     <footer class="chat-input-container">
+      <div v-if="taskLoading" class="loading-hint">
+        <i class="fa fa-spinner fa-spin" style="margin-right: 6px" />模型加载中，就绪后即可对话...
+      </div>
+      <div v-else-if="taskError" class="loading-hint error">
+        <i class="fa fa-exclamation-circle" style="margin-right: 6px" />{{ taskError }}
+      </div>
       <div class="chat-input-inner">
         <button class="clear-btn" title="清空对话" @click="handleNewChat">
           <i class="fa fa-eraser" />
@@ -261,22 +363,21 @@ onMounted(loadTask)
             v-model="inputQuestion"
             class="input-box"
             rows="1"
-            :disabled="loading"
+            :disabled="loading || taskLoading"
             placeholder="给模型发送消息..."
             @keydown.enter.exact.prevent="handleSend"
             @input="autoResize"
           />
           <button
             class="send-btn"
-            :class="{ active: inputQuestion.trim() && !loading }"
+            :class="{ active: inputQuestion.trim() && !loading && !taskLoading }"
-            :disabled="!inputQuestion.trim() || loading"
+            :disabled="!inputQuestion.trim() || loading || taskLoading"
             @click="handleSend"
           >
             <i class="fa fa-arrow-up" />
           </button>
         </div>
       </div>
-      <div class="footer-hint">内容由 AI 生成，请仔细甄别。</div>
     </footer>
 
     <!-- 设置抽屉（系统提示词等） -->
@@ -703,9 +804,18 @@ onMounted(loadTask)
   }
 }
 
-.footer-hint {
+.loading-hint {
-  margin-top: 12px;
+  margin-bottom: 10px;
-  font-size: 12px;
+  padding: 6px 14px;
-  color: #9ca3af;
+  font-size: 13px;
+  color: #b45309;
+  background: #fef3c7;
+  border-radius: 8px;
+  text-align: center;
+
+  &.error {
+    color: #b91c1c;
+    background: #fee2e2;
+  }
 }
 </style>

frontend/src/views/inference/InferenceCreateView.vue

@@ -1,12 +1,12 @@
 <script setup lang="ts">
-import { ref, reactive, computed, onMounted } from 'vue'
+import { ref, reactive, computed, onMounted, watch } from 'vue'
 import { useRouter } from 'vue-router'
 import { ElMessage, type FormInstance, type FormRules } from 'element-plus'
 import PageCard from '@/components/PageCard.vue'
 import { getModelList, getTrainedModels } from '@/api/modules/model'
 import { getSystemInfo } from '@/api/modules/system'
 import { getComputeNodes, type ComputeNode } from '@/api/modules/compute'
-import { createCompare, preloadLocalModel, preloadTrainedModel } from '@/api/modules/compare'
+import { createCompare, loadCompare } from '@/api/modules/compare'
 import type { ModelItem, TrainedModel, GpuInfo } from '@/types'
 
 const router = useRouter()
@@ -27,6 +27,8 @@ interface SelectableModel {
   name: string
   source: 'database' | 'trained'
   model_path: string
+  compute_node_id?: string
+  compute_node_name?: string
   merged?: boolean
   merging?: boolean
   disabled?: boolean
@@ -51,6 +53,8 @@ const trainedOptions = computed<SelectableModel[]>(() =>
     name: m.name,
     source: 'trained',
     model_path: m.merged_path || m.base_model_path || '',
+    compute_node_id: m.compute_node_id,
+    compute_node_name: m.compute_node_name,
     merged: m.merged,
     merging: m.merging,
     disabled: m.merged === false,
@@ -80,7 +84,7 @@ const form = reactive({
   /** 选中的模型 key（单选） */
   model_key: '',
   /** 使用的 GPU */
-  gpu_id: 0,
+  gpu_key: '',
 })
 
 const rules: FormRules = {
@@ -90,6 +94,13 @@ const rules: FormRules = {
 
 /** 当前选中的模型对象 */
 const selectedModel = computed(() => modelMap.value[form.model_key])
+const selectedGpu = computed(() => idleGpus.value.find((g) => `${g.node_id || ''}:${g.id ?? 0}` === form.gpu_key))
+
+watch(selectedModel, (model) => {
+  if (!model?.compute_node_id) return
+  const gpu = idleGpus.value.find((item) => item.node_id === model.compute_node_id)
+  if (gpu) form.gpu_key = `${gpu.node_id || ''}:${gpu.id ?? 0}`
+})
 
 async function handleSubmit() {
   if (!formRef.value) return
@@ -101,62 +112,47 @@ async function handleSubmit() {
       return
     }
     submitting.value = true
-    startupStatus.value = '正在启动模型服务...'
+    startupStatus.value = '正在创建推理任务...'
     try {
-      // Step 1: 将模型加载到算力节点
+      if (!m.model_path) {
-      const preloadPayload = {
+        ElMessage.warning('当前模型未配置算力节点可访问路径，请先在模型管理中维护模型路径')
-        model_name_or_path: m.model_path,
-        model_name: m.name,
-        template: 'qwen',
-      }
-      let preloadResult: any
-      if (m.source === 'trained') {
-        preloadResult = await preloadTrainedModel(preloadPayload)
-      } else {
-        preloadResult = await preloadLocalModel(preloadPayload)
-      }
-
-      if (preloadResult && (preloadResult as any).error) {
-        ElMessage.warning(`模型加载失败：${(preloadResult as any).error}`)
-        submitting.value = false
-        startupStatus.value = ''
         return
       }
 
-      // Step 2: 创建推理任务记录
+      // Step 1: 创建推理任务记录
       const taskResult = await createCompare({
         name: form.name || m.name,
         description: form.description,
+        status: 'pending',
         models: [
           {
             model_id: String(m.id),
             model_name: m.name,
             model_path: m.model_path,
             source: m.source,
-            gpu_id: form.gpu_id,
+            gpu_id: selectedGpu.value?.id ?? 0,
+            node_id: selectedGpu.value?.node_id || m.compute_node_id,
+            node_name: selectedGpu.value?.node_name || m.compute_node_name,
           },
         ],
       })
       const taskId = taskResult?.id || 'unknown'
 
-      ElMessage.success('模型已启动')
+      // Step 2: 统一通过推理任务加载接口异步派发模型加载，状态会落到列表记录中。
-      router.push({
+      startupStatus.value = '正在启动模型服务，首次加载可能需要数分钟...'
-        path: `/model-inference/chat/${taskId}`,
+      const loadResult: any = await loadCompare(taskId)
-        query: {
+      if (loadResult?.status === 'failed' || loadResult?.error) {
-          model: m.name,
+        ElMessage.warning(`模型加载失败：${loadResult?.error || '请检查算力节点日志'}`)
-          source: m.source,
+        router.push('/model-inference')
-          model_path: m.model_path,
+        return
-        },
+      }
-      })
+
+      // 加载为异步派发，回到列表页可看到“启动中 → 已就绪”的状态流转
+      ElMessage.success('模型加载中，就绪后即可对话')
+      router.push('/model-inference')
     } catch (e: any) {
-      // 真实 API 失败时回退到 mock 模式（方便无算力节点的开发调试）
-      const m = selectedModel.value!
       const reason = e?.message || e?.toString() || '未知错误'
-      ElMessage.warning(`推理服务启动失败：${reason}，进入 mock 演示模式`)
+      ElMessage.warning(`推理服务启动失败：${reason}`)
-      router.push({
-        path: '/model-inference/chat/mock',
-        query: { model: m.name },
-      })
     } finally {
       submitting.value = false
       startupStatus.value = ''
@@ -181,7 +177,10 @@ async function loadData() {
     gpus.value = sys?.gpu || []
     computeNodes.value = nodes || []
     // 默认选中第一个空闲 GPU
-    if (idleGpus.value.length > 0) form.gpu_id = idleGpus.value[0].id ?? 0
+    if (idleGpus.value.length > 0) {
+      const firstGpu = idleGpus.value[0]
+      form.gpu_key = `${firstGpu.node_id || ''}:${firstGpu.id ?? 0}`
+    }
   } catch {
     // ignore
   }
@@ -229,12 +228,12 @@ onMounted(loadData)
       </el-form-item>
 
       <el-form-item label="GPU">
-        <el-select v-model="form.gpu_id" style="width: 400px">
+        <el-select v-model="form.gpu_key" style="width: 400px">
           <el-option
             v-for="g in idleGpus"
-            :key="g.id ?? 0"
+            :key="`${g.node_id || ''}:${g.id ?? 0}`"
-            :label="`${g.name} (GPU${g.id ?? 0}) [空闲]`"
+            :label="`${g.node_name || g.node_code || '算力节点'} / ${g.name} (GPU${g.id ?? 0}) [空闲]`"
-            :value="g.id ?? 0"
+            :value="`${g.node_id || ''}:${g.id ?? 0}`"
           />
         </el-select>
       </el-form-item>

frontend/src/views/inference/InferenceListView.vue

@@ -92,11 +92,9 @@ async function handleUnload(row: any) {
   loadData()
 }
 
-/** 删除（先释放算力节点再删除记录） */
+/** 删除（后端删除内部会 best-effort 释放算力节点，这里直接删记录） */
 async function handleDelete(row: any) {
   await ElMessageBox.confirm('确定要删除该推理记录吗？将先释放算力节点再删除。', '确认删除', { type: 'warning' })
-  // 先释放算力节点上的模型
-  await unloadCompare(row.id).catch(() => {})
   await deleteCompare(row.id)
   dataList.value = dataList.value.filter((item) => item.id !== row.id)
   await loadData(true)

frontend/src/views/model/MergeWeightsView.vue

@@ -18,9 +18,12 @@ const trainedModels = ref<TrainedModel[]>([])
 const currentModel = computed(() => trainedModels.value.find((m) => m.name === modelName.value))
 
 const form = reactive({
+  trained_model_id: '',
   model_name: modelName.value,
   train_method: method.value,
   base_model_path: '',
+  adapter_path: '',
+  compute_node_id: '',
 })
 
 async function loadModel() {
@@ -28,23 +31,30 @@ async function loadModel() {
     const res = await getTrainedModels()
     trainedModels.value = res?.models || []
     const target = trainedModels.value.find((m) => m.name === modelName.value)
+    form.trained_model_id = target?.id == null ? '' : String(target.id)
     form.base_model_path = target?.base_model_path || ''
+    form.adapter_path = target?.artifact_dir || target?.adapter_path || target?.merged_path || ''
+    form.compute_node_id = target?.compute_node_id || ''
   } catch {
     // ignore
   }
 }
 
 async function handleMerge() {
-  if (!form.model_name || !form.base_model_path) {
+  if (!form.model_name || !form.base_model_path || !form.adapter_path) {
     ElMessage.warning('缺少模型信息')
     return
   }
   merging.value = true
   try {
     await mergeModel({
+      trained_model_id: form.trained_model_id || form.model_name,
       model_name: form.model_name,
       train_method: form.train_method,
       base_model_path: form.base_model_path,
+      adapter_path: form.adapter_path,
+      compute_node_id: form.compute_node_id,
+      output_model_name: `${form.model_name}-merged`,
     })
     ElMessage.success('合并成功')
     router.push('/model-manage')
@@ -82,6 +92,9 @@ onMounted(loadModel)
       <el-form-item label="基座模型路径">
         <el-input v-model="form.base_model_path" placeholder="基座模型路径" />
       </el-form-item>
+      <el-form-item label="Adapter 路径">
+        <el-input v-model="form.adapter_path" placeholder="LoRA Adapter 权重目录" />
+      </el-form-item>
 
       <el-form-item>
         <el-button type="primary" :loading="merging" @click="handleMerge">

frontend/src/views/system/TrainingLogView.vue

@@ -8,13 +8,14 @@ import TrainingTaskOverview from './training-log/TrainingTaskOverview.vue'
 import { usePolling } from '@/composables/usePolling'
 import '@/plugins/echarts-training-log'
 import { useModelsStore } from '@/stores/models'
-import { getFineTune, getFineTuneDiagnostics, getFineTuneLogs, type TrainingDiagnostic } from '@/api/modules/fineTune'
+import { getFineTune, getFineTuneDiagnostics, getFineTuneLogs, getFineTuneMetrics, type TrainingDiagnostic } from '@/api/modules/fineTune'
 import { getTrainingLogFiles, getTrainingLogContent } from '@/api/modules/log'
 import { getDataset } from '@/api/modules/dataset'
 import { getSystemInfo } from '@/api/modules/system'
 import { TRAIN_TYPE_MAP, TRAIN_METHOD_MAP } from '@/constants'
 import {
   buildMetricChartOption,
+  metricsFromApi,
   parseTrainingLog,
   resolveTrainingLogFile,
 } from './training-log/trainingLogModel'
@@ -42,6 +43,7 @@ const loading = ref(true)
 
 // 训练指标数据（ECharts 接收 number[]，下标即 step）
 const metricData = reactive({
+  steps: [] as number[],
   loss: [] as number[],
   gradNorm: [] as number[],
   lr: [] as number[],
@@ -62,9 +64,9 @@ const gpuExpanded = ref(false)
 let refreshInFlight = false
 
 /** 三个曲线的 ECharts 配置（响应式，数据变化自动重绘） */
-const lossChartOption = computed(() => buildMetricChartOption('Loss', metricData.loss, '#4f46e5'))
+const lossChartOption = computed(() => buildMetricChartOption('Loss', metricData.loss, metricData.steps, '#4f46e5'))
-const gradChartOption = computed(() => buildMetricChartOption('Grad Norm', metricData.gradNorm, '#3b82f6'))
+const gradChartOption = computed(() => buildMetricChartOption('Grad Norm', metricData.gradNorm, metricData.steps, '#3b82f6'))
-const lrChartOption = computed(() => buildMetricChartOption('Learning Rate', metricData.lr, '#14b8a6', true))
+const lrChartOption = computed(() => buildMetricChartOption('Learning Rate', metricData.lr, metricData.steps, '#14b8a6', true))
 const baseModelName = computed(() => task.value?.base_model != null
   ? modelsStore.getModelName(task.value.base_model)
   : '未配置')
@@ -77,10 +79,10 @@ const trainingMethodName = computed(() => task.value?.train_method
 const taskGpuLabel = computed(() => task.value?.gpus?.length
   ? task.value.gpus.map((gpuId) => `GPU ${gpuId}`).join('、')
   : '未配置')
-const latestLoss = computed(() => metricData.loss[metricData.loss.length - 1])
+const latestLoss = computed(() => lastFinite(metricData.loss))
-const latestGradNorm = computed(() => metricData.gradNorm[metricData.gradNorm.length - 1])
+const latestGradNorm = computed(() => lastFinite(metricData.gradNorm))
-const latestLearningRate = computed(() => metricData.lr[metricData.lr.length - 1])
+const latestLearningRate = computed(() => lastFinite(metricData.lr))
-const latestEpoch = computed(() => metricData.epoch[metricData.epoch.length - 1])
+const latestEpoch = computed(() => lastFinite(metricData.epoch))
 const logLineCount = computed(() => logContent.value ? logContent.value.split(/\r?\n/).length : 0)
 const taskGpuItems = computed<TaskGpuItem[]>(() => (task.value?.gpus ?? []).map((gpuId) => {
   const index = Number(gpuId)
@@ -123,7 +125,7 @@ const gpuRefreshState = computed(() => {
   })
   return gpuLoadError.value
     ? `更新失败 · 最后更新 ${updateTime}`
-    : `${updateTime} 更新 · 每 5 秒刷新`
+    : `${updateTime} 更新 · 每 3 秒刷新`
 })
 
 function formatMetric(value?: number, scientific = false) {
@@ -131,6 +133,13 @@ function formatMetric(value?: number, scientific = false) {
   return scientific ? value.toExponential(2) : value.toFixed(4).replace(/0+$/, '').replace(/\.$/, '')
 }
 
+function lastFinite(values: number[]) {
+  for (let index = values.length - 1; index >= 0; index -= 1) {
+    if (Number.isFinite(values[index])) return values[index]
+  }
+  return undefined
+}
+
 function safePercent(value?: number) {
   return Math.round(Math.min(100, Math.max(0, Number(value || 0))))
 }
@@ -216,6 +225,7 @@ const isLoraMethod = computed(() =>
 function applyLogContent(content: string) {
   const parsed = parseTrainingLog(content)
   logContent.value = content
+  metricData.steps = parsed.metrics.steps
   metricData.loss = parsed.metrics.loss
   metricData.gradNorm = parsed.metrics.gradNorm
   metricData.lr = parsed.metrics.lr
@@ -223,6 +233,26 @@ function applyLogContent(content: string) {
   Object.assign(summary, parsed.summary)
 }
 
+function applyMetricData(metrics = { steps: [] as number[], loss: [] as number[], gradNorm: [] as number[], lr: [] as number[], epoch: [] as number[] }) {
+  metricData.steps = metrics.steps
+  metricData.loss = metrics.loss
+  metricData.gradNorm = metrics.gradNorm
+  metricData.lr = metrics.lr
+  metricData.epoch = metrics.epoch
+}
+
+async function loadMetrics(currentTask: FineTuneTask) {
+  try {
+    const points = await getFineTuneMetrics(currentTask.id)
+    const parsed = metricsFromApi(points || [])
+    if (parsed.loss.length || parsed.gradNorm.length || parsed.lr.length) {
+      applyMetricData(parsed)
+    }
+  } catch {
+    // 日志解析结果会作为兜底曲线数据。
+  }
+}
+
 async function loadLog(currentTask: FineTuneTask) {
   try {
     const runtime = await getFineTuneLogs(currentTask.id, { tail_lines: 800 })
@@ -277,6 +307,7 @@ async function refreshAll() {
       ? loadDataset(currentTask.train_dataset_id)
       : Promise.resolve()
     await Promise.all([datasetPromise, loadLog(currentTask), loadGpuStatus(), loadDiagnostics(currentTask)])
+    await loadMetrics(currentTask)
   } finally {
     loading.value = false
     refreshInFlight = false
@@ -523,7 +554,7 @@ onMounted(async () => {
 
     <!-- 训练曲线 -->
     <PageCard class="metrics-panel" title="训练曲线" subtitle="持续监控模型收敛情况与学习率变化">
-      <template #extra><span class="refresh-state">每 5 秒刷新</span></template>
+      <template #extra><span class="refresh-state">每 3 秒刷新</span></template>
       <div class="chart-list" aria-label="训练指标曲线">
         <section class="chart-section">
           <div class="chart-section-header">
@@ -560,7 +591,7 @@ onMounted(async () => {
 
     <!-- 原始日志 -->
     <PageCard class="log-card" title="训练日志" subtitle="查看训练任务的原始运行输出">
-      <template #extra><span class="log-meta">{{ logLineCount }} 行 · 每 5 秒刷新</span></template>
+      <template #extra><span class="log-meta">{{ logLineCount }} 行 · 每 3 秒刷新</span></template>
       <pre class="log-pre">{{ logContent || '暂无日志' }}</pre>
     </PageCard>
     </template>

frontend/src/views/system/training-log/trainingLogModel.ts

@@ -1,7 +1,9 @@
 import type { EChartsOption } from 'echarts'
 import type { FineTuneTask, TrainingLogFile } from '@/types'
+import type { FineTuneMetricPoint } from '@/api/modules/fineTune'
 
 export interface TrainingMetricData {
+  steps: number[]
   loss: number[]
   gradNorm: number[]
   lr: number[]
@@ -26,7 +28,7 @@ function escapeRegExp(value: string) {
 }
 
 function extractNumber(source: string, key: string) {
-  const match = source.match(new RegExp(`['"]?${escapeRegExp(key)}['"]?\\s*:\\s*(${NUMBER_SOURCE})`, 'i'))
+  const match = source.match(new RegExp(`['"]?${escapeRegExp(key)}['"]?\\s*(?:=|:)\\s*(${NUMBER_SOURCE})`, 'i'))
   return match ? Number(match[1]) : undefined
 }
 
@@ -57,24 +59,44 @@ export function resolveTrainingLogFile(
 
 /** 解析日志中的逐步指标。字段顺序和常见数值格式均不受限制。 */
 export function parseTrainingMetrics(text: string): TrainingMetricData {
-  const metrics: TrainingMetricData = { loss: [], gradNorm: [], lr: [], epoch: [] }
+  const metrics: TrainingMetricData = { steps: [], loss: [], gradNorm: [], lr: [], epoch: [] }
-  const blocks = text.match(/\{[^{}\r\n]*\}/g) || []
+  const candidates = text
+    .split(/\r?\n/)
+    .flatMap((line) => {
+      const blocks = line.match(/\{[^{}\r\n]*\}/g)
+      return blocks?.length ? blocks.map((block) => `${line} ${block}`) : [line]
+    })
 
-  for (const block of blocks) {
+  for (const [index, line] of candidates.entries()) {
-    const loss = extractNumber(block, 'loss')
+    const loss = extractNumber(line, 'loss')
-    const gradNorm = extractNumber(block, 'grad_norm')
+    const gradNorm = extractNumber(line, 'grad_norm')
-    const learningRate = extractNumber(block, 'learning_rate')
+    const learningRate = extractNumber(line, 'learning_rate')
-    const epoch = extractNumber(block, 'epoch')
+    const epoch = extractNumber(line, 'epoch')
-    if (loss == null || gradNorm == null || learningRate == null) continue
+    if (loss == null && gradNorm == null && learningRate == null) continue
-    metrics.loss.push(loss)
+    metrics.steps.push(extractNumber(line, 'step') ?? metrics.steps.length + index + 1)
-    metrics.gradNorm.push(gradNorm)
+    metrics.loss.push(loss ?? Number.NaN)
-    metrics.lr.push(learningRate)
+    metrics.gradNorm.push(gradNorm ?? Number.NaN)
-    if (epoch != null) metrics.epoch.push(epoch)
+    metrics.lr.push(learningRate ?? Number.NaN)
+    metrics.epoch.push(epoch ?? Number.NaN)
   }
 
   return metrics
 }
 
+export function metricsFromApi(points: FineTuneMetricPoint[]): TrainingMetricData {
+  const metrics: TrainingMetricData = { steps: [], loss: [], gradNorm: [], lr: [], epoch: [] }
+  for (const [index, point] of points.entries()) {
+    const hasMetric = point.loss != null || point.grad_norm != null || point.learning_rate != null
+    if (!hasMetric) continue
+    metrics.steps.push(Number(point.step || index + 1))
+    metrics.loss.push(point.loss == null ? Number.NaN : Number(point.loss))
+    metrics.gradNorm.push(point.grad_norm == null ? Number.NaN : Number(point.grad_norm))
+    metrics.lr.push(point.learning_rate == null ? Number.NaN : Number(point.learning_rate))
+    metrics.epoch.push(point.epoch == null ? Number.NaN : Number(point.epoch))
+  }
+  return metrics
+}
+
 /** 每次都返回新对象，日志截断或切换时不会残留上一轮汇总。 */
 export function parseTrainingSummary(text: string): TrainingSummary {
   const emptySummary: TrainingSummary = { epoch: '', trainLoss: '', runtime: '' }
@@ -102,11 +124,23 @@ export function parseTrainingLog(text: string): ParsedTrainingLog {
 export function buildMetricChartOption(
   label: string,
   data: number[],
+  steps: number[],
   color: string,
   logScale = false,
 ): EChartsOption {
+  const visibleData = data.map((value) => (Number.isFinite(value) ? value : null))
   return {
     grid: { top: 24, right: 20, bottom: 56, left: 56 },
+    graphic: visibleData.some((value) => value != null)
+      ? []
+      : [
+          {
+            type: 'text',
+            left: 'center',
+            top: 'middle',
+            style: { text: '暂无训练指标数据', fill: '#94a3b8', fontSize: 13 },
+          },
+        ],
     tooltip: {
       trigger: 'axis',
       axisPointer: { type: 'cross' },
@@ -116,6 +150,7 @@ export function buildMetricChartOption(
     },
     xAxis: {
       type: 'category',
+      data: steps.map((step, index) => (Number.isFinite(step) ? String(step) : String(index + 1))),
       boundaryGap: false,
       name: 'Step',
       nameTextStyle: { color: '#94a3b8', fontSize: 11 },
@@ -142,7 +177,7 @@ export function buildMetricChartOption(
       {
         name: label,
         type: 'line',
-        data,
+        data: visibleData,
         smooth: true,
         symbol: 'none',
         lineStyle: { width: 2, color },