更新前端看板

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

@@ -1,7 +1,6 @@
 from fastapi import APIRouter
 
 from app.core.logging import get_logger
-from app.db.platform_store import get_platform_store
 
 router = APIRouter()
 logger = get_logger(__name__)
@@ -10,5 +9,9 @@ logger = get_logger(__name__)
 @router.get("/health")
 async def health_check() -> dict[str, object]:
     logger.info("health check requested")
-    return {"code": 0, "message": "ok", "data": get_platform_store().health_metrics()}
+    return {
+        "code": 0,
+        "message": "ok",
+        "data": {"cpu_percent": 0.0, "memory_percent": 0.0, "disk_percent": 0.0},
+    }
 

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

@@ -15,7 +15,7 @@ from app.core.auth import filter_accessible_resource_ids, get_current_user, has_
 from app.core.config import get_settings
 from app.db.platform_store import get_platform_store
 from app.modules.compute_gateway.client import ComputeNodeClient
-from app.modules.compute_gateway.sync import poll_compute_jobs_once
+from app.modules.compute_gateway.sync import fetch_eval_result_content, poll_compute_jobs_once
 
 router = APIRouter()
 
@@ -33,6 +33,62 @@ def _select_first_online_node(store: Any) -> dict[str, Any] | None:
     return None
 
 
+def _build_messages_payload(payload: dict[str, Any]) -> dict[str, Any]:
+    """Convert frontend inference payload to compute API messages format.
+
+    Accepts both:
+      - OpenAI-style: {messages: [{role, content}, ...], temperature, ...}
+      - Frontend-style: {user_question, system_prompt, temperature, ...}
+    """
+    if payload.get("messages"):
+        messages = payload["messages"]
+        # messages already in OpenAI format; pass through with optional system prompt
+        if payload.get("system_prompt") and not any(m.get("role") == "system" for m in messages):
+            messages = [{"role": "system", "content": payload["system_prompt"]}] + list(messages)
+    else:
+        messages = []
+        if payload.get("system_prompt"):
+            messages.append({"role": "system", "content": payload["system_prompt"]})
+        question = payload.get("user_question") or payload.get("question") or ""
+        if question:
+            messages.append({"role": "user", "content": question})
+    return {
+        "messages": messages,
+        "temperature": float(payload.get("temperature", 0.7)),
+        "top_p": float(payload.get("top_p", 0.95)),
+        "max_new_tokens": int(payload.get("max_tokens", 2048)),
+        "do_sample": bool(payload.get("do_sample", True)),
+    }
+
+
+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)
+    if not node:
+        return StreamingResponse(
+            iter(['data: {"error": "no online compute node available for inference"}\n\n']),
+            media_type="text/event-stream",
+        )
+    client = ComputeNodeClient(node["api_base_url"])
+    compute_payload = _build_messages_payload(payload)
+
+    async def stream_proxy():
+        async with httpx.AsyncClient(timeout=300) as http:
+            url = f"{node['api_base_url'].rstrip('/')}{client.route_prefix}/inference/chat/stream"
+            try:
+                async with http.stream("POST", url, json=compute_payload, headers=client.headers()) as resp:
+                    if resp.status_code >= 400:
+                        yield f'data: {{"error": "compute node returned {resp.status_code}"}}\n\n'.encode()
+                        return
+                    async for chunk in resp.aiter_bytes():
+                        yield chunk
+            except Exception as exc:
+                yield f'data: {{"error": "stream proxy failed: {exc}"}}\n\n'.encode()
+
+    return StreamingResponse(stream_proxy(), media_type="text/event-stream")
+
+
 def fail(status_code: int, message: str) -> HTTPException:
     return HTTPException(status_code=status_code, detail={"code": status_code, "message": message, "data": None})
 
@@ -856,6 +912,7 @@ async def upload_dataset_files(
 ) -> dict[str, Any]:
     created: list[dict[str, Any]] = []
     compute_sync: list[dict[str, Any]] = []
+    pending_sync: list[tuple[str, str, bytes]] = []
     store = get_platform_store()
     try:
         store.dataset(dataset_id)
@@ -867,16 +924,18 @@ async def upload_dataset_files(
             content = raw.decode("utf-8", errors="replace")
             created_file = store.add_dataset_file(conn, dataset_id, file.filename or "upload.jsonl", content)
             created.append(created_file)
-            if sync_to_compute:
-                compute_sync.extend(
-                    await _sync_dataset_file_to_compute_nodes(
-                        store,
-                        dataset_id,
-                        created_file["id"],
-                        created_file["name"],
-                        raw,
-                    )
+            pending_sync.append((created_file["id"], created_file["name"], raw))
+    if sync_to_compute:
+        for file_id, file_name, raw in pending_sync:
+            compute_sync.extend(
+                await _sync_dataset_file_to_compute_nodes(
+                    store,
+                    dataset_id,
+                    file_id,
+                    file_name,
+                    raw,
                 )
+            )
     return ok({"files": created, "compute_sync": compute_sync})
 
 
@@ -1212,7 +1271,23 @@ async def model_eval_list(current_user: dict = Depends(get_current_user)) -> dic
 @router.get("/model-eval/{task_id}")
 async def model_eval_detail(task_id: str, current_user: dict = Depends(get_current_user)) -> dict[str, Any]:
     try:
-        task = get_platform_store().eval_task(task_id)
+        store = get_platform_store()
+        task = store.eval_task(task_id)
+        if task.get("compute_job_id") and task.get("compute_node_id") and task.get("status") in {"queued", "running", "completed"}:
+            node = next(
+                (n for n in store.compute_nodes() if n["id"] == task.get("compute_node_id")),
+                None,
+            )
+            if node:
+                try:
+                    client = ComputeNodeClient(node["api_base_url"])
+                    job = await client.get_job(task["compute_job_id"])
+                    result_content = None
+                    if job.get("status") == "completed" and not task.get("samples"):
+                        result_content = await fetch_eval_result_content(client, node, job)
+                    task = store.apply_eval_job_result(task_id, job, result_content)
+                except Exception:
+                    pass
     except KeyError:
         raise fail(404, "eval task not found")
     if not has_resource_access("eval", task_id, current_user, "read"):
@@ -1222,8 +1297,149 @@ async def model_eval_detail(task_id: str, current_user: dict = Depends(get_curre
 
 @router.post("/model-eval/start")
 async def model_eval_start(payload: dict[str, Any] = Body(...)) -> dict[str, Any]:
-    task = get_platform_store().create_eval_task(payload)
-    return ok({"task_id": task["id"], **task})
+    """Start an evaluation task: submit eval job to compute node."""
+    store = get_platform_store()
+    # 1. Create eval task record
+    task = store.create_eval_task({**payload, "status": "pending"})
+
+    # 2. Resolve model path (supports both regular models and trained models)
+    model_id = str(payload.get("model_id", ""))
+    model_path = ""
+    adapter_path = payload.get("adapter_path", "")
+    try:
+        db_model = store.model(model_id)
+        model_path = db_model.get("path", "")
+    except KeyError:
+        # Try trained_models table (IDs prefixed with tm_)
+        trained = next((m for m in store.trained_models() if m["id"] == model_id), None)
+        if trained:
+            merged_path = trained.get("merged_path", "")
+            base_path = trained.get("base_model_path", "")
+            if trained.get("merged") and merged_path:
+                # Merged model: use merged_path as model, no adapter needed
+                model_path = merged_path
+            elif base_path:
+                # Unmerged: use base model + adapter checkpoint
+                model_path = base_path
+                if merged_path:
+                    adapter_path = merged_path
+            else:
+                model_path = merged_path or base_path
+    if not model_path:
+        store.update_eval_task(task["id"], {"status": "failed", "error": "model not found or no path"})
+        return ok({"task_id": task["id"], "status": "failed", "error": "model not found or no path"})
+
+    # 3. Resolve dataset file
+    dataset_id = str(payload.get("dataset_id", ""))
+    dataset_path = ""
+    try:
+        ds_files = store.training_dataset_files(dataset_id)
+        if ds_files:
+            dataset_path = ds_files[0].get("local_path") or ds_files[0].get("name", "")
+    except Exception:
+        pass
+    if not dataset_path:
+        # Try to get file content and sync to compute
+        try:
+            ds = store.dataset(dataset_id)
+            for f in ds.get("files", []):
+                if f.get("content"):
+                    dataset_path = f.get("name", f"dataset_{dataset_id}.jsonl")
+                    break
+        except KeyError:
+            pass
+    if not dataset_path:
+        store.update_eval_task(task["id"], {"status": "failed", "error": "dataset not found or no files"})
+        return ok({"task_id": task["id"], "status": "failed", "error": "dataset not found or no files"})
+
+    # 4. Resolve dimension config
+    dimension_id = str(payload.get("dimension_id", ""))
+    dimension_cfg: dict[str, Any] = {}
+    if dimension_id:
+        try:
+            dim = store.dimension(dimension_id)
+            # Resolve eval model API config
+            eval_model_name = dim.get("eval_model", "")
+            api_url = ""
+            api_key = ""
+            if eval_model_name:
+                try:
+                    eval_model = store.model(eval_model_name) if eval_model_name.startswith("m_") else store.model_by_name(eval_model_name)
+                    api_url = eval_model.get("api_url", "")
+                    api_key = eval_model.get("api_key", "")
+                except (KeyError, Exception):
+                    pass
+            dimension_cfg = {
+                "type": dim.get("type", ""),
+                "eval_model": eval_model_name,
+                "eval_method": dim.get("eval_method", ""),
+                "eval_prompt": dim.get("eval_prompt", ""),
+                "api_url": api_url,
+                "api_key": api_key,
+                "score_min": dim.get("score_min", 0),
+                "score_max": dim.get("score_max", 5),
+                "pass_threshold": dim.get("pass_threshold", 3),
+            }
+        except KeyError:
+            pass
+
+    # 5. Select compute node
+    node = _select_first_online_node(store)
+    if not node:
+        store.update_eval_task(task["id"], {"status": "failed", "error": "no online compute node"})
+        return ok({"task_id": task["id"], "status": "failed", "error": "no online compute node"})
+
+    # 6. Build eval job payload
+    output_dir = f"/data/yg-ft/outputs/{task['id']}"
+    job_payload = {
+        "id": f"eval_{task['id']}",
+        "name": task.get("eval_task_name", task["id"]),
+        "engine": "eval",
+        "model_name_or_path": model_path,
+        "adapter_name_or_path": adapter_path,
+        "template": payload.get("template", "qwen"),
+        "dataset_path": dataset_path,
+        "output_dir": output_dir,
+        "basic_metrics": payload.get("basic_metrics", {}),
+        "dimension": dimension_cfg,
+        "gpus": [int(payload.get("gpu_id", 0))],
+        "temperature": payload.get("temperature", 0.1),
+        "max_new_tokens": payload.get("max_new_tokens", 512),
+        "compute_node_id": node["id"],
+    }
+
+    # 7. Submit to compute node via create_job (uses engine="eval" path)
+    try:
+        client = ComputeNodeClient(node["api_base_url"])
+        # Sync dataset file to compute node if needed
+        if not dataset_path.startswith("/"):
+            try:
+                ds_files = store.training_dataset_files(dataset_id)
+                if ds_files and ds_files[0].get("content"):
+                    upload_result = await client.upload_file(
+                        ds_files[0].get("name", "eval_data.jsonl"),
+                        ds_files[0]["content"].encode("utf-8"),
+                        f"datasets/{dataset_id}/{ds_files[0].get('name', 'eval_data.jsonl')}",
+                        resource_type="dataset",
+                        resource_id=dataset_id,
+                    )
+                    job_payload["dataset_path"] = upload_result.get("local_path", dataset_path)
+            except Exception:
+                pass
+
+        job = await client.create_job(job_payload)
+        store.update_eval_task(task["id"], {
+            "status": "running",
+            "compute_job_id": job.get("id"),
+            "compute_node_id": node["id"],
+            "output_dir": output_dir,
+        })
+        if job.get("status") in {"queued", "running"}:
+            store.mark_inference_loaded(node["id"])
+        return ok({"task_id": task["id"], "status": "running", "job": job})
+    except Exception as exc:
+        store.update_eval_task(task["id"], {"status": "failed", "error": str(exc)})
+        return ok({"task_id": task["id"], "status": "failed", "error": str(exc)})
 
 
 @router.delete("/model-eval/{task_id}")
@@ -1298,8 +1514,27 @@ 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]:
+    """Best-effort unload the inference model from the first online compute node."""
+    store = get_platform_store()
+    # Clear all inference tracking — only one model can be loaded at a time
+    for node in store.compute_nodes():
+        store.mark_inference_unloaded(node["id"])
+    node = _select_first_online_node(store)
+    if not node:
+        return {"unloaded": False, "error": "no online compute node"}
+    try:
+        client = ComputeNodeClient(node["api_base_url"])
+        result = await client._request("POST", "/inference/unload", json_data={})
+        return result
+    except Exception as exc:
+        return {"unloaded": False, "error": str(exc)}
+
+
 @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
+    await _unload_from_compute_node()
     get_platform_store().delete_compare_task(task_id)
     return ok({"deleted": task_id})
 
@@ -1329,26 +1564,56 @@ async def model_compare_update_load_status(task_id: str, payload: dict[str, Any]
 
 @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）。"""
     try:
-        task = get_platform_store().compare_task(task_id)
+        store = get_platform_store()
+        task = store.compare_task(task_id)
         models = task.get("models") or []
         if isinstance(models, str):
             try:
                 models = json.loads(models)
             except json.JSONDecodeError:
                 models = []
-        loaded_models = [
-            {
-                "model_id": item.get("model_id"),
-                "model_name": item.get("model_name"),
-                "status": "ready",
-                "pid": 45000 + index,
-                "port": item.get("port") or 18000 + index,
+        # 选取在线算力节点
+        node = _select_first_online_node(store)
+        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
+            model_path = item.get("model_path", "")
+            if not model_path:
+                # 尝试从模型库获取路径
+                model_id = item.get("model_id", "")
+                try:
+                    db_model = store.model(model_id)
+                    model_path = db_model.get("path", "")
+                except KeyError:
+                    pass
+            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"),
             }
-            for index, item in enumerate(models)
-            if isinstance(item, dict)
-        ]
-        return ok(get_platform_store().update_compare_task(task_id, {"status": "loaded", "load_status": {"loaded_models": loaded_models}}))
+            if item.get("adapter_path"):
+                load_payload["adapter_name_or_path"] = item["adapter_path"]
+            try:
+                result = await client._request("POST", "/inference/load", json_data=load_payload)
+                if result.get("loaded"):
+                    store.mark_inference_loaded(node["id"])
+                    loaded_models.append({**item, "status": "ready"})
+                else:
+                    loaded_models.append({**item, "status": "error", "error": result.get("error", "load failed")})
+            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}})
+        return ok(updated)
     except KeyError:
         raise fail(404, "compare task not found")
 
@@ -1356,7 +1621,11 @@ async def model_compare_load(task_id: str) -> dict[str, Any]:
 @router.post("/model-compare/{task_id}/unload")
 async def model_compare_unload(task_id: str) -> dict[str, Any]:
     try:
-        return ok(get_platform_store().update_compare_task(task_id, {"status": "pending", "load_status": {"loaded_models": []}}))
+        store = get_platform_store()
+        # 真正释放算力节点上的模型资源
+        unload_result = await _unload_from_compute_node()
+        updated = store.update_compare_task(task_id, {"status": "pending", "load_status": {"loaded_models": []}})
+        return ok({"task": updated, "unload": unload_result})
     except KeyError:
         raise fail(404, "compare task not found")
 
@@ -1368,18 +1637,23 @@ async def model_compare_start_model(task_id: str, payload: dict[str, Any] = Body
 
 @router.post("/model-compare/chat-with-port")
 async def model_compare_chat_with_port(payload: dict[str, Any] = Body(...)) -> dict[str, Any]:
-    question = ""
-    for message in payload.get("messages") or []:
-        if message.get("role") == "user":
-            question = str(message.get("content") or "")
-    content = f"当前后端已收到推理请求：{question[:120]}"
-    return ok({"response": content, "content": content})
+    """Proxy non-streaming chat to the compute node running the inference model."""
+    store = get_platform_store()
+    node = _select_first_online_node(store)
+    if not node:
+        return ok({"response": "no online compute node available for inference", "request": payload})
+    try:
+        client = ComputeNodeClient(node["api_base_url"])
+        result = await client._request("POST", "/inference/chat", json_data=_build_messages_payload(payload))
+        return ok(result)
+    except Exception as exc:
+        return ok({"response": f"inference failed: {exc}", "request": payload})
 
 
 @router.post("/model-compare/stream-chat")
-async def model_compare_stream_chat(payload: dict[str, Any] = Body(...)) -> dict[str, Any]:
-    question = payload.get("user_question") or payload.get("question") or ""
-    return ok({"response": f"当前后端已收到流式推理请求：{str(question)[:120]}"})
+async def model_compare_stream_chat(payload: dict[str, Any] = Body(...)) -> StreamingResponse:
+    """Stream chat from the compute node (SSE proxy)."""
+    return await _stream_chat_proxy(payload)
 
 
 @router.post("/model-chat/batch")
@@ -1396,7 +1670,7 @@ async def model_chat_local(payload: dict[str, Any] = Body(...)) -> dict[str, Any
         return ok({"response": "no online compute node available for inference", "request": payload})
     try:
         client = ComputeNodeClient(node["api_base_url"])
-        result = await client._request("POST", "/inference/chat", json_data=payload)
+        result = await client._request("POST", "/inference/chat", json_data=_build_messages_payload(payload))
         return ok(result)
     except Exception as exc:
         return ok({"response": f"inference failed: {exc}", "request": payload})
@@ -1405,28 +1679,15 @@ async def model_chat_local(payload: dict[str, Any] = Body(...)) -> dict[str, Any
 @router.post("/model-chat/local/chat/stream")
 async def model_chat_local_stream(payload: dict[str, Any] = Body(...)) -> StreamingResponse:
     """Stream chat from the compute node."""
-    store = get_platform_store()
-    node = _select_first_online_node(store)
-    if not node:
-        return StreamingResponse(
-            iter(['data: {"error": "no online compute node"}\n\n']),
-            media_type="text/event-stream",
-        )
-    client = ComputeNodeClient(node["api_base_url"])
-
-    async def stream_proxy():
-        async with httpx.AsyncClient(timeout=300) as http:
-            url = f"{node['api_base_url'].rstrip('/')}/modelTF/inference/chat/stream"
-            async with http.stream("POST", url, json=payload, headers=client.headers()) as resp:
-                async for chunk in resp.aiter_bytes():
-                    yield chunk
-
-    return StreamingResponse(stream_proxy(), media_type="text/event-stream")
+    return await _stream_chat_proxy(payload)
 
 
 @router.post("/model-chat/local/preload")
 async def model_chat_local_preload(payload: dict[str, Any] = Body(...)) -> dict[str, Any]:
     """Load a model on the compute node for inference."""
+    model_path = (payload.get("model_name_or_path") or "").strip()
+    if not model_path:
+        return ok({"loaded": False, "error": "model_name_or_path is required"})
     store = get_platform_store()
     node = _select_first_online_node(store)
     if not node:
@@ -1434,6 +1695,8 @@ async def model_chat_local_preload(payload: dict[str, Any] = Body(...)) -> dict[
     try:
         client = ComputeNodeClient(node["api_base_url"])
         result = await client._request("POST", "/inference/load", json_data=payload)
+        if result.get("loaded"):
+            store.mark_inference_loaded(node["id"])
         return ok(result)
     except Exception as exc:
         return ok({"loaded": False, "error": str(exc)})
@@ -1443,6 +1706,9 @@ 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
+    for n in store.compute_nodes():
+        store.mark_inference_unloaded(n["id"])
     node = _select_first_online_node(store)
     if not node:
         return ok({"unloaded": False, "error": "no online compute node"})
@@ -1472,6 +1738,9 @@ async def model_chat_local_status() -> dict[str, Any]:
 @router.post("/model-chat/trained/preload")
 async def model_chat_trained_preload(payload: dict[str, Any] = Body(...)) -> dict[str, Any]:
     """Load a trained model (base + adapter) on the compute node for inference."""
+    model_path = (payload.get("model_name_or_path") or "").strip()
+    if not model_path:
+        return ok({"loaded": False, "error": "model_name_or_path is required"})
     store = get_platform_store()
     node = _select_first_online_node(store)
     if not node:
@@ -1479,6 +1748,8 @@ async def model_chat_trained_preload(payload: dict[str, Any] = Body(...)) -> dic
     try:
         client = ComputeNodeClient(node["api_base_url"])
         result = await client._request("POST", "/inference/load", json_data=payload)
+        if result.get("loaded"):
+            store.mark_inference_loaded(node["id"])
         return ok(result)
     except Exception as exc:
         return ok({"loaded": False, "error": str(exc)})
@@ -1517,6 +1788,16 @@ async def update_compute_node(node_id: str, payload: dict[str, Any] = Body(...))
         raise fail(400, str(exc))
 
 
+@router.delete("/compute/nodes/{node_id}")
+async def delete_compute_node(node_id: str) -> dict[str, Any]:
+    try:
+        return ok(get_platform_store().delete_compute_node(node_id))
+    except KeyError:
+        raise fail(404, "compute node not found")
+    except ValueError as exc:
+        raise fail(400, str(exc))
+
+
 @router.post("/compute/nodes/{node_id}/test-connection")
 async def test_compute_node(node_id: str) -> dict[str, Any]:
     store = get_platform_store()

backend/app/db/platform_store.py

@@ -88,6 +88,22 @@ def parse_size_bytes(value: Any) -> int:
     return max(0, round(amount * _SIZE_UNIT_BYTES[unit]))
 
 
+def count_dataset_records(content: str) -> int:
+    text = (content or "").strip()
+    if not text:
+        return 0
+
+    try:
+        value = json.loads(text)
+        if isinstance(value, list):
+            return len(value)
+        return 1
+    except (TypeError, ValueError, json.JSONDecodeError):
+        pass
+
+    return len([line for line in text.splitlines() if line.strip()])
+
+
 def version_number(value: Any, default: int = 0) -> int:
     try:
         number = int(value)
@@ -296,6 +312,7 @@ class PlatformStore:
         # request (notably expensive against the remote PostgreSQL instance).
         # TCP keepalive 让操作系统持续保活连接，抵抗远程库空闲静默断连。
         pool_kwargs = {
+            "connect_timeout": 5,
             "keepalives": 1,
             "keepalives_idle": 10,
             "keepalives_interval": 5,
@@ -320,6 +337,19 @@ class PlatformStore:
         self._pool.open()
         self.ensure_schema()
         self.ensure_seed_data()
+        # Track which compute nodes have an active inference model loaded
+        self._inference_nodes: set[str] = set()
+
+    # ── inference node tracking ────────────────────────────────────
+
+    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
 
     @contextmanager
     def connect(self) -> Iterator["PgConnection"]:
@@ -1298,6 +1328,7 @@ class PlatformStore:
             file_size_bytes = int(file_row.get("size_bytes") or 0)
             if file_size_bytes <= 0:
                 file_size_bytes = parse_size_bytes(file_row.get("size"))
+            file_record_count = int(file_row.get("record_count") or 0)
             decoded_files.append(
                 {
                     "id": file_row["id"],
@@ -1306,7 +1337,7 @@ class PlatformStore:
                     "size_bytes": file_size_bytes,
                     **dataset_file_version_summary(file_row),
                     "create_time": file_row["create_time"],
-                    "record_count": int(file_row.get("record_count") or 0),
+                    "record_count": file_record_count,
                     "split": metadata.get("file_split"),
                 }
             )
@@ -1324,6 +1355,9 @@ class PlatformStore:
             total_size_bytes = int(row.get("size_bytes") or 0)
             if total_size_bytes <= 0:
                 total_size_bytes = parse_size_bytes(row.get("size"))
+        total_record_count = sum(int(item.get("record_count") or 0) for item in decoded_files)
+        if not decoded_files:
+            total_record_count = int(row.get("record_count") or row.get("count") or 0)
         current_version_nos = sorted(
             {
                 int(item["current_version_no"])
@@ -1333,6 +1367,8 @@ class PlatformStore:
         )
         return {
             **dict(row),
+            "count": total_record_count,
+            "record_count": total_record_count,
             "size_bytes": total_size_bytes,
             "current_version_no": (
                 current_version_nos[0] if len(current_version_nos) == 1 else None
@@ -1407,7 +1443,7 @@ class PlatformStore:
         version_id = f"{file_id}_v1"
         size_bytes = len(content.encode("utf-8"))
         size = f"{size_bytes} B"
-        record_count = len([line for line in content.splitlines() if line.strip()])
+        record_count = count_dataset_records(content)
         version = {
             "id": version_id,
             "version": 1,
@@ -1440,10 +1476,18 @@ class PlatformStore:
         )
         conn.execute(
             """UPDATE datasets
-               SET count=count+?, record_count=record_count+?,
-                   size_bytes=size_bytes+?, size=((size_bytes+?)::text || ' B')
+               SET count=stats.record_count,
+                   record_count=stats.record_count,
+                   size_bytes=stats.size_bytes,
+                   size=(stats.size_bytes::text || ' B')
+               FROM (
+                   SELECT COALESCE(SUM(record_count), 0) AS record_count,
+                          COALESCE(SUM(size_bytes), 0) AS size_bytes
+                   FROM dataset_files
+                   WHERE dataset_id=?
+               ) stats
                WHERE id=?""",
-            (record_count, record_count, size_bytes, size_bytes, dataset_id),
+            (dataset_id, dataset_id),
         )
         return {
             "id": file_id,
@@ -1548,19 +1592,57 @@ class PlatformStore:
             row = conn.execute("SELECT * FROM dataset_files WHERE id=?", (file_id,)).fetchone()
             if not row:
                 raise KeyError(file_id)
+            content = payload.get("content", "")
+            size_bytes = len(content.encode("utf-8"))
+            record_count = count_dataset_records(content)
             versions = json_loads(row["versions"], [])
             version = {
                 "id": f"{file_id}_v{len(versions) + 1}",
                 "version": len(versions) + 1,
+                "version_no": len(versions) + 1,
                 "create_time": utcnow(),
                 "description": payload.get("description", "online edit"),
+                "size_bytes": size_bytes,
+                "record_count": record_count,
             }
             versions.append(version)
             conn.execute(
-                "UPDATE dataset_files SET content=?, active_version_id=?, versions=? WHERE id=?",
-                (payload.get("content", ""), version["id"], json_dumps(versions), file_id),
+                """
+                UPDATE dataset_files
+                SET content=?, active_version_id=?, current_version_id=?, versions=?,
+                    size_bytes=?, size=?, record_count=?, version_no=?
+                WHERE id=?
+                """,
+                (
+                    content,
+                    version["id"],
+                    version["id"],
+                    json_dumps(versions),
+                    size_bytes,
+                    f"{size_bytes} B",
+                    record_count,
+                    version["version_no"],
+                    file_id,
+                ),
             )
-            return {"version": version, "content": payload.get("content", "")}
+            conn.execute(
+                """UPDATE datasets
+                   SET count=stats.record_count,
+                       record_count=stats.record_count,
+                       size_bytes=stats.size_bytes,
+                       size=(stats.size_bytes::text || ' B')
+                   FROM (
+                       SELECT dataset_id,
+                              COALESCE(SUM(record_count), 0) AS record_count,
+                              COALESCE(SUM(size_bytes), 0) AS size_bytes
+                       FROM dataset_files
+                       WHERE dataset_id=(SELECT dataset_id FROM dataset_files WHERE id=?)
+                       GROUP BY dataset_id
+                   ) stats
+                   WHERE datasets.id=stats.dataset_id""",
+                (file_id,),
+            )
+            return {"version": version, "content": content}
 
     def activate_file_version(self, file_id: str, version_id: str) -> dict[str, Any]:
         with self.connect() as conn:
@@ -2114,10 +2196,63 @@ class PlatformStore:
             )
         return self.eval_task(task_id)
 
+    def update_eval_task(self, task_id: str, updates: dict[str, Any]) -> dict[str, Any]:
+        """Update fields in an eval task's payload without replacing the whole record."""
+        task = self.eval_task(task_id)
+        merged = {**task, **updates}
+        with self.connect() as conn:
+            conn.execute(
+                "UPDATE eval_tasks SET payload=?, status=? WHERE id=?",
+                (json_dumps(merged), merged.get("status", task.get("status", "pending")), task_id),
+            )
+        return self.eval_task(task_id)
+
     def delete_eval_task(self, task_id: str) -> None:
         with self.connect() as conn:
             conn.execute("DELETE FROM eval_tasks WHERE id=?", (task_id,))
 
+    def running_eval_tasks(self) -> list[dict[str, Any]]:
+        """Return eval tasks that have been submitted to a compute node and are still running."""
+        return [
+            task for task in self.eval_tasks()
+            if task.get("compute_job_id") and task.get("status") in {"queued", "running"}
+        ]
+
+    def apply_eval_job_result(self, task_id: str, job: dict[str, Any], result_content: dict[str, Any] | None = None) -> dict[str, Any]:
+        """Sync a compute job status/result back to an eval task."""
+        task = self.eval_task(task_id)
+        job_status = str(job.get("status", ""))
+        status_map = {"queued": "running", "running": "running", "completed": "completed",
+                       "failed": "failed", "stopped": "stopped"}
+        new_status = status_map.get(job_status, job_status or task.get("status", "pending"))
+        updates: dict[str, Any] = {
+            "status": new_status,
+            "progress": int(job.get("progress", 0)),
+            "output_dir": job.get("output_dir", task.get("output_dir", "")),
+        }
+        # On completion, populate results from eval_results.json content
+        if new_status == "completed" and result_content:
+            updates.update({
+                "overall_score": result_content.get("overall_score", 0),
+                "overall_score_max": result_content.get("overall_score_max", 100),
+                "overall_evaluation": result_content.get("overall_evaluation", ""),
+                "improvement_suggestions": result_content.get("improvement_suggestions", []),
+                "dimension_summary": result_content.get("dimension_summary", []),
+                "samples": result_content.get("samples", []),
+                "sample_count": result_content.get("sample_count", 0),
+                "completed_count": result_content.get("completed_count", 0),
+                "passed_count": result_content.get("passed_count", 0),
+                "basic_metrics": result_content.get("basic_metrics", {}),
+                "score": result_content.get("overall_score", 0),
+                "completed_time": utcnow(),
+            })
+        elif new_status in {"failed", "stopped"}:
+            updates.update({
+                "error": job.get("error") or task.get("error") or "",
+                "completed_time": utcnow(),
+            })
+        return self.update_eval_task(task_id, updates)
+
     def dimensions(self) -> list[dict[str, Any]]:
         with self.connect() as conn:
             rows = conn.execute("SELECT * FROM eval_dimensions ORDER BY create_time DESC").fetchall()
@@ -2284,6 +2419,15 @@ class PlatformStore:
         ).fetchall()
         return {int(row["gpu_index"]) for row in rows}
 
+    def _node_gpu_indexes(self, conn: PgConnection, node: dict[str, Any]) -> set[int]:
+        rows = conn.execute("SELECT gpu_index FROM gpus WHERE node_id=?", (node["id"],)).fetchall()
+        if rows:
+            return {int(row["gpu_index"]) for row in rows}
+        return set(range(max(0, int(node.get("gpu_count") or 0))))
+
+    def _node_capacity(self, node: dict[str, Any]) -> int:
+        return max(1, int(node.get("max_parallel_jobs") or 1), int(node.get("gpu_count") or 0))
+
     def _schedule_node_locked(self, conn: PgConnection, payload: dict[str, Any]) -> dict[str, Any]:
         requested = payload.get("requested_node_id") or payload.get("compute_node_id")
         requested_gpus = [int(item) for item in payload.get("gpus") or []]
@@ -2291,13 +2435,15 @@ class PlatformStore:
         candidates = [
             n
             for n in nodes
-            if n["enabled"] and n["scheduler_status"] == "online" and n["current_running_jobs"] < n["max_parallel_jobs"]
+            if n["enabled"] and n["scheduler_status"] == "online" and n["current_running_jobs"] < self._node_capacity(n)
         ]
         if requested_gpus:
+            requested_gpu_set = set(requested_gpus)
             candidates = [
                 node
                 for node in candidates
-                if not set(requested_gpus).intersection(self._active_gpu_indexes(conn, node["id"]))
+                if requested_gpu_set.issubset(self._node_gpu_indexes(conn, node))
+                and not requested_gpu_set.intersection(self._active_gpu_indexes(conn, node["id"]))
             ]
         if requested:
             selected = next((n for n in candidates if n["id"] == requested), None)
@@ -2312,8 +2458,8 @@ class PlatformStore:
                     reason = "disabled"
                 elif node["scheduler_status"] != "online":
                     reason = f"status={node['scheduler_status']}"
-                elif node["current_running_jobs"] >= node["max_parallel_jobs"]:
-                    reason = f"capacity full {node['current_running_jobs']}/{node['max_parallel_jobs']}"
+                elif node["current_running_jobs"] >= self._node_capacity(node):
+                    reason = f"capacity full {node['current_running_jobs']}/{self._node_capacity(node)}"
                 else:
                     reason = "not selected"
                 reasons.append(f"{node['code']}({reason})")
@@ -2664,6 +2810,24 @@ class PlatformStore:
             )
         return next(node for node in self.compute_nodes() if node["id"] == node_id)
 
+    def delete_compute_node(self, node_id: str) -> dict[str, Any]:
+        with self.connect() as conn:
+            node = conn.execute("SELECT * FROM compute_nodes WHERE id=?", (node_id,)).fetchone()
+            if not node:
+                raise KeyError(node_id)
+            active = conn.execute(
+                """
+                SELECT COUNT(*) AS cnt
+                FROM fine_tune_tasks
+                WHERE compute_node_id=? AND status IN ('syncing','queued','running')
+                """,
+                (node_id,),
+            ).fetchone()
+            if active and int(active["cnt"] or 0) > 0:
+                raise ValueError("compute node has active training tasks")
+            conn.execute("DELETE FROM compute_nodes WHERE id=?", (node_id,))
+        return {"deleted": node_id}
+
     def update_compute_node_health(self, node_id: str, health: dict[str, Any], success: bool, error: str | None = None) -> dict[str, Any]:
         current = next((n for n in self.compute_nodes() if n["id"] == node_id), None)
         if not current:
@@ -2749,6 +2913,11 @@ class PlatformStore:
                 )
                 busy = task is not None and task.get("status") == "running"
                 reserved = task is not None and task.get("status") in {"syncing", "queued"}
+                # Also mark GPU as busy if an inference model is loaded on this node
+                inference_busy = self.is_inference_loaded(row["node_id"])
+                if inference_busy and not busy:
+                    busy = True
+                    reserved = False
                 memory_used = round(row["memory_total_gb"] * (0.72 if busy else 0.18 if reserved else 0.04), 1)
                 gpu_percent = 86 if busy else 22 if reserved else 3
                 memory_total = float(row["memory_total_gb"] or 0)
@@ -2827,11 +2996,12 @@ class PlatformStore:
         }
 
     def health_metrics(self) -> dict[str, float]:
-        info = self.system_info()
+        # Health checks must stay lightweight. The Docker healthcheck and page
+        # refresh probes should not wait on dashboard/GPU/database aggregation.
         return {
-            "cpu_percent": info["cpu"]["percent"],
-            "memory_percent": info["memory"]["percent"],
-            "disk_percent": info["disk"]["percent"],
+            "cpu_percent": 0.0,
+            "memory_percent": 0.0,
+            "disk_percent": 0.0,
         }
 
     def queue(self) -> list[dict[str, Any]]:

backend/app/modules/compute_gateway/client.py

@@ -207,7 +207,8 @@ class ComputeNodeClient:
             "resource_id": resource_id or "",
         }
         files = {"file": (filename, content)}
-        async with httpx.AsyncClient(timeout=max(self.timeout, 60), headers=self.headers()) as client:
+        timeout = httpx.Timeout(max(self.timeout, 60), connect=self.timeout)
+        async with httpx.AsyncClient(timeout=timeout, headers=self.headers()) as client:
             response = await client.post(
                 _join_url(self.api_base_url, f"{self.route_prefix}/compute/files/upload"),
                 data=data,

backend/app/modules/compute_gateway/sync.py

@@ -10,6 +10,24 @@ 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)
 
 
+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:
+        return None
+    full_path = f"{str(output_dir).rstrip('/')}/eval_results.json"
+    data_root = "/data/yg-ft/"
+    if full_path.startswith(data_root):
+        full_path = full_path[len(data_root):]
+    rel_path = full_path.lstrip("/")
+    import httpx
+    url = f"{node['api_base_url'].rstrip('/')}/modelTF/compute/files/read"
+    async with httpx.AsyncClient(timeout=30, headers=client.headers()) as http:
+        response = await http.get(url, params={"path": rel_path})
+        response.raise_for_status()
+        payload = response.json()
+        return payload if isinstance(payload, dict) else None
+
+
 async def poll_compute_jobs_once() -> dict[str, Any]:
     store = get_platform_store()
     synced: list[dict[str, Any]] = []
@@ -48,4 +66,34 @@ async def poll_compute_jobs_once() -> dict[str, Any]:
             standalone_synced.append(store.sync_model_merge_job(record["id"], job))
         except Exception as exc:  # noqa: BLE001 - keep polling other jobs
             failed.append({"job_id": record["id"], "error": str(exc)})
-    return {"synced": len(synced) + len(standalone_synced), "failed": failed, "items": synced, "standalone": standalone_synced}
+
+    # ── Eval job sync ────────────────────────────────────────────────
+    eval_synced = 0
+    for eval_task in store.running_eval_tasks():
+        node = next(
+            (item for item in store.compute_nodes() if item["id"] == eval_task.get("compute_node_id")),
+            None,
+        )
+        if not node:
+            failed.append({"eval_task_id": eval_task["id"], "error": "compute node not found"})
+            continue
+        try:
+            client = ComputeNodeClient(node["api_base_url"])
+            job = await client.get_job(eval_task["compute_job_id"])
+            result_content = None
+            # Try to read eval_results.json from the job output directory
+            if job.get("status") == "completed" and job.get("output_dir"):
+                try:
+                    result_content = await fetch_eval_result_content(client, node, job)
+                except Exception:
+                    pass
+            store.apply_eval_job_result(eval_task["id"], job, result_content)
+            # If job completed, un-mark inference loaded
+            if job.get("status") in {"completed", "failed", "stopped"}:
+                store.mark_inference_unloaded(node["id"])
+            eval_synced += 1
+        except Exception as exc:  # noqa: BLE001
+            failed.append({"eval_task_id": eval_task["id"], "error": str(exc)})
+
+    return {"synced": len(synced) + len(standalone_synced) + eval_synced, "failed": failed,
+            "items": synced, "standalone": standalone_synced, "eval_synced": eval_synced}

backend/pyproject.toml

@@ -10,6 +10,7 @@ dependencies = [
   "pydantic>=2.7.0",
   "sqlalchemy>=2.0.30",
   "psycopg[binary]>=3.2.1",
+  "psycopg-pool>=3.2.1",
   "alembic>=1.13.1",
   "redis>=5.0.4",
   "httpx>=0.27.0",

backend/requirements.txt

@@ -4,6 +4,7 @@ python-multipart>=0.0.9
 pydantic>=2.7.0
 sqlalchemy>=2.0.30
 psycopg[binary]>=3.2.1
+psycopg-pool>=3.2.1
 alembic>=1.13.1
 redis>=5.0.4
 httpx>=0.27.0

compute/api/main.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import json
 import os
 import math
 import hashlib
@@ -449,6 +450,29 @@ def create_app() -> FastAPI:
                 accelerator_errors, accelerator_warnings, accelerator = _validate_training_accelerator(payload)
                 errors.extend(accelerator_errors)
                 warnings.extend(accelerator_warnings)
+        elif engine == "eval":
+            # Eval engine: validate model path and dataset path
+            if not payload.get("model_name_or_path"):
+                errors.append("model_name_or_path is required for eval")
+            else:
+                path_checks.append(_check_path_item({
+                    "name": "model_name_or_path",
+                    "path": payload.get("model_name_or_path", ""),
+                    "type": "any",
+                    "required": True,
+                }))
+            if payload.get("dataset_path"):
+                path_checks.append(_check_path_item({
+                    "name": "dataset_path",
+                    "path": payload.get("dataset_path", ""),
+                    "type": "file",
+                    "required": True,
+                }))
+            else:
+                errors.append("dataset_path is required for eval")
+            if shutil.which("python") is None:
+                errors.append("python runtime not found")
+
         elif engine == "smoke":
             warnings.append("smoke engine skips model and dataset path checks")
 
@@ -688,8 +712,6 @@ def create_app() -> FastAPI:
             infer_backend=payload.get("infer_backend", "huggingface"),
             infer_dtype=payload.get("infer_dtype", "auto"),
         )
-        if not result.get("loaded"):
-            raise HTTPException(status_code=500, detail=result.get("error", "model load failed"))
         return result
 
     @app.post(f"{route_prefix}/inference/unload")
@@ -813,6 +835,22 @@ def create_app() -> FastAPI:
             "checksum_sha256": checksum,
         }
 
+    @app.get(f"{route_prefix}/compute/files/read")
+    async def read_file(path: str = Query(...)) -> JSONResponse:
+        """Read a text file from within YG_FT_DATA_ROOT.  Used by the backend
+        to fetch eval results and other job outputs."""
+        data_root = Path(os.getenv("YG_FT_DATA_ROOT", "/data/yg-ft"))
+        target = (data_root / path.lstrip("/\\")).resolve()
+        if not _path_inside(data_root, target):
+            raise HTTPException(status_code=400, detail="path must stay inside YG_FT_DATA_ROOT")
+        if not target.is_file():
+            raise HTTPException(status_code=404, detail="file not found")
+        try:
+            content = target.read_text(encoding="utf-8")
+            return JSONResponse(json.loads(content) if content.strip().startswith("{") else {"content": content})
+        except Exception as exc:
+            raise HTTPException(status_code=500, detail=str(exc))
+
     @app.get(f"{route_prefix}/compute/files/{{file_id}}/download")
     async def download_file(file_id: str) -> FileResponse:
         upload_root = Path(os.getenv("YG_FT_DATA_ROOT", "/data/yg-ft")) / "uploads"

compute/engines/llama_factory/adapter.py

@@ -204,6 +204,31 @@ def build_command(config: dict[str, Any], llama_factory_home: str = "/app/LLaMA-
             command.extend(["--quantization_bit", str(quantization_bit)])
         return LlamaFactoryCommand(command=command, work_dir=str(Path(llama_factory_home)), env={})
 
+    if engine == "eval":
+        output_dir = config.get("output_dir") or f"/data/yg-ft/outputs/{config.get('name', 'eval-job')}"
+        eval_config_path = str(Path(output_dir) / "eval_config.json")
+        eval_config = {
+            "model_name_or_path": config.get("model_name_or_path", ""),
+            "adapter_name_or_path": config.get("adapter_name_or_path", ""),
+            "template": config.get("template", "qwen"),
+            "dataset_path": config.get("dataset_path", ""),
+            "output_dir": output_dir,
+            "basic_metrics": config.get("basic_metrics", {}),
+            "dimension": config.get("dimension", {}),
+            "temperature": config.get("temperature", 0.1),
+            "top_p": config.get("top_p", 0.95),
+            "max_new_tokens": config.get("max_new_tokens", 512),
+            "infer_backend": config.get("infer_backend", "huggingface"),
+            "infer_dtype": config.get("infer_dtype", "auto"),
+        }
+        Path(output_dir).mkdir(parents=True, exist_ok=True)
+        Path(eval_config_path).write_text(json.dumps(eval_config, ensure_ascii=False, indent=2), encoding="utf-8")
+        return LlamaFactoryCommand(
+            command=["python", "-u", "-m", "compute.engines.llama_factory.eval_runner", "--config", eval_config_path],
+            work_dir="/app",
+            env={},
+        )
+
     errors = validate_config(config)
     if errors:
         raise ValueError("; ".join(errors))

compute/engines/llama_factory/eval_runner.py

@@ -0,0 +1,480 @@
+"""
+Evaluation runner — executes model evaluation as a subprocess job.
+
+Usage:
+    python -m compute.engines.llama_factory.eval_runner --config <config_json_path>
+
+The config JSON is written by the compute API before spawning this subprocess.
+Results are written to ``output_dir/eval_results.json`` and progress is printed
+to stdout (captured as job logs).
+"""
+
+from __future__ import annotations
+
+import json
+import math
+import re
+import sys
+import time
+from difflib import SequenceMatcher
+from pathlib import Path
+from typing import Any
+
+
+def _load_dataset(path: str) -> list[dict[str, Any]]:
+    """Load a JSON or JSONL dataset file.
+
+    Supports common field names used across the platform:
+    * ``instruction`` + ``input`` + ``output``  (Alpaca-style)
+    * ``question`` + ``answer``
+    * ``messages``  (ShareGPT-style – the last assistant message is treated as reference)
+    """
+    file_path = Path(path)
+    text = file_path.read_text(encoding="utf-8", errors="replace").strip()
+    if not text:
+        return []
+    if file_path.suffix.lower() == ".json":
+        value = json.loads(text)
+        if isinstance(value, list):
+            return [item for item in value if isinstance(item, dict)]
+        return [value] if isinstance(value, dict) else []
+
+    samples: list[dict[str, Any]] = []
+    for line in text.splitlines():
+        line = line.strip()
+        if not line:
+            continue
+        try:
+            obj = json.loads(line)
+        except json.JSONDecodeError:
+            continue
+        if isinstance(obj, dict):
+            samples.append(obj)
+    return samples
+
+
+def _sample_question(sample: dict[str, Any]) -> str:
+    """Extract the user-facing question / instruction from a sample."""
+    if sample.get("instruction"):
+        text = sample["instruction"]
+        if sample.get("input"):
+            text += "\n" + sample["input"]
+        return text
+    if sample.get("question"):
+        return sample["question"]
+    # ShareGPT-style: use the last user message as question
+    messages = sample.get("messages") or []
+    user_msgs = [m["content"] for m in messages if m.get("role") == "user"]
+    return user_msgs[-1] if user_msgs else ""
+
+
+def _sample_reference(sample: dict[str, Any]) -> str:
+    """Extract the reference answer from a sample."""
+    if sample.get("output"):
+        return sample["output"]
+    if sample.get("answer"):
+        return sample["answer"]
+    messages = sample.get("messages") or []
+    assistant_msgs = [m["content"] for m in messages if m.get("role") == "assistant"]
+    return assistant_msgs[-1] if assistant_msgs else ""
+
+
+# ---------------------------------------------------------------------------
+# Basic metrics
+# ---------------------------------------------------------------------------
+
+def _compute_bleu(references: list[str], predictions: list[str], ngram: int = 4) -> dict[str, Any]:
+    """Compute BLEU score via sacrebleu (corpus-level)."""
+    try:
+        from sacrebleu.metrics import BLEU
+    except ImportError:
+        return {"enabled": False, "error": "sacrebleu not installed", "score": 0}
+    bleu = BLEU(max_ngram_order=ngram)
+    # sacrebleu expects list-of-strings; we have one reference per prediction
+    score = bleu.corpus_score(predictions, [references])
+    return {
+        "enabled": True,
+        "score": round(score.score, 2),
+        "bleu": round(score.score, 2),
+    }
+
+
+def _compute_rouge(references: list[str], predictions: list[str], methods: list[str] | None = None) -> dict[str, Any]:
+    """Compute ROUGE scores via rouge-score."""
+    try:
+        from rouge_score import rouge_scorer
+    except ImportError:
+        return {"enabled": False, "error": "rouge-score not installed", "score": 0}
+    methods = methods or ["rouge1", "rouge2", "rougeL"]
+    # Normalize: map "rouge_1"/"rouge1" → "rouge1", "rouge_l"/"rougeL" → "rougeL"
+    _rouge_aliases = {"rouge_1": "rouge1", "rouge_2": "rouge2", "rouge_l": "rougeL"}
+    methods = [_rouge_aliases.get(m, m.replace("_", "")) for m in methods]
+    scorer = rouge_scorer.RougeScorer(methods, use_stemmer=True)
+    totals: dict[str, float] = {}
+    n = max(len(predictions), 1)
+    for ref, pred in zip(references, predictions):
+        result = scorer.score(ref, pred)
+        for key in methods:
+            totals[key] = totals.get(key, 0) + result[key].fmeasure
+    avg = {k: round(v / n, 4) for k, v in totals.items()}
+    return {"enabled": True, "score": round(avg.get("rougeL", avg.get("rouge1", 0)) * 100, 2), **avg}
+
+
+def _compute_cosine(references: list[str], predictions: list[str]) -> dict[str, Any]:
+    """Compute average cosine similarity via sklearn."""
+    try:
+        from sklearn.feature_extraction.text import TfidfVectorizer
+        from sklearn.metrics.pairwise import cosine_similarity
+    except ImportError:
+        return {"enabled": False, "error": "scikit-learn not installed", "score": 0}
+    try:
+        vectorizer = TfidfVectorizer()
+        tfidf = vectorizer.fit_transform(references + predictions)
+        n = len(references)
+        ref_vec = tfidf[:n]
+        pred_vec = tfidf[n:]
+        sims = cosine_similarity(ref_vec, pred_vec).diagonal()
+        return {"enabled": True, "score": round(float(sims.mean()) * 100, 2)}
+    except ValueError:
+        return {"enabled": True, "score": 0, "error": "insufficient text for vectorization"}
+
+
+def _normalize_text(value: str) -> str:
+    return re.sub(r"\s+", " ", str(value or "").strip().lower())
+
+
+def _compute_exact_match(references: list[str], predictions: list[str]) -> dict[str, Any]:
+    total = len(predictions)
+    if not total:
+        return {"enabled": True, "score": 0, "matched": 0, "total": 0}
+    matched = sum(
+        1
+        for ref, pred in zip(references, predictions)
+        if _normalize_text(ref) == _normalize_text(pred)
+    )
+    return {"enabled": True, "score": round(matched / total * 100, 2), "matched": matched, "total": total}
+
+
+def _compute_text_similarity(references: list[str], predictions: list[str]) -> dict[str, Any]:
+    if not predictions:
+        return {"enabled": True, "score": 0}
+    scores = [
+        SequenceMatcher(None, _normalize_text(ref), _normalize_text(pred)).ratio()
+        for ref, pred in zip(references, predictions)
+    ]
+    return {"enabled": True, "score": round(sum(scores) / max(len(scores), 1) * 100, 2)}
+
+
+# ---------------------------------------------------------------------------
+# LLM Judge
+# ---------------------------------------------------------------------------
+
+def _judge_sample(
+    question: str,
+    reference: str,
+    prediction: str,
+    config: dict[str, Any],
+) -> dict[str, Any]:
+    """Call an OpenAI-compatible LLM to judge a single sample.
+
+    Returns a dict with keys:
+        score, max_score, passed, judgement, evaluation_reason, error_type
+    """
+    api_url = (config.get("api_url") or "").strip().rstrip("/")
+    api_key = (config.get("api_key") or "").strip()
+    eval_model = (config.get("eval_model") or "").strip()
+    eval_prompt = (config.get("eval_prompt") or "").strip()
+    score_min = float(config.get("score_min", 0))
+    score_max = float(config.get("score_max", 5))
+    pass_threshold = float(config.get("pass_threshold", 3))
+
+    if not api_url or not eval_model:
+        return {"score": 0, "max_score": score_max, "passed": False, "judgement": "未配置",
+                "evaluation_reason": "未配置评测模型", "error_type": "其他"}
+
+    system_msg = (
+        eval_prompt
+        or "你是一个专业的评测专家。请根据参考答-案对被测模型的输出进行评分。"
+    )
+    user_msg = (
+        f"## 问题\n{question}\n\n"
+        f"## 参考答案\n{reference}\n\n"
+        f"## 模型输出\n{prediction}\n\n"
+        f"请给出 {score_min}-{score_max} 分的评分，并说明理由。"
+    )
+
+    try:
+        import urllib.request
+        import urllib.error
+
+        body = json.dumps({
+            "model": eval_model,
+            "messages": [
+                {"role": "system", "content": system_msg},
+                {"role": "user", "content": user_msg},
+            ],
+            "temperature": 0.3,
+            "max_tokens": 512,
+        }).encode("utf-8")
+
+        req = urllib.request.Request(
+            f"{api_url}/v1/chat/completions",
+            data=body,
+            headers={
+                "Content-Type": "application/json",
+                "Authorization": f"Bearer {api_key}",
+            },
+        )
+        resp = urllib.request.urlopen(req, timeout=120)
+        data = json.loads(resp.read().decode("utf-8"))
+        reply = data["choices"][0]["message"]["content"]
+    except Exception as exc:
+        return {"score": 0, "max_score": score_max, "passed": False,
+                "judgement": "错误", "evaluation_reason": f"评测模型调用失败: {exc}",
+                "error_type": "其他"}
+
+    # Parse score from reply — look for patterns like "4分" or "Score: 4"
+    score = 0
+    import re
+    score_patterns = [
+        r'(?:得分|分数|评分|score)[^\d]*(\d+(?:\.\d+)?)',
+        r'(\d+(?:\.\d+)?)\s*分',
+        r'(\d+(?:\.\d+)?)\s*/\s*\d+',
+    ]
+    for pat in score_patterns:
+        m = re.search(pat, reply, re.IGNORECASE)
+        if m:
+            try:
+                score = float(m.group(1))
+            except ValueError:
+                continue
+            break
+    score = max(score_min, min(score_max, score))
+    passed = score >= pass_threshold
+
+    # Determine judgement label
+    if score >= pass_threshold + 1:
+        judgement = "正确"
+    elif score >= pass_threshold:
+        judgement = "部分正确"
+    else:
+        judgement = "错误"
+
+    # Guess error type from reply
+    reply_lower = reply.lower()
+    if any(w in reply_lower for w in ["幻觉", "hallucination", "编造"]):
+        error_type = "幻觉"
+    elif any(w in reply_lower for w in ["不完整", "incomplete", "遗漏"]):
+        error_type = "不完整"
+    elif any(w in reply_lower for w in ["格式", "format"]):
+        error_type = "格式偏差"
+    elif any(w in reply_lower for w in ["混淆", "confusion", "错误"]):
+        error_type = "混淆"
+    else:
+        error_type = "其他"
+
+    return {
+        "score": score,
+        "max_score": score_max,
+        "passed": passed,
+        "judgement": judgement,
+        "evaluation_reason": reply[:2000],
+        "error_type": error_type,
+    }
+
+
+# ---------------------------------------------------------------------------
+# Main entry point
+# ---------------------------------------------------------------------------
+
+def run_eval(config: dict[str, Any]) -> dict[str, Any]:
+    """Execute a full evaluation run.  Returns the result dict (also written to file)."""
+    model_path = config["model_name_or_path"]
+    adapter_path = config.get("adapter_name_or_path", "")
+    template = config.get("template", "qwen")
+    dataset_path = config["dataset_path"]
+    output_dir = Path(config["output_dir"])
+    output_dir.mkdir(parents=True, exist_ok=True)
+    basic_cfg = config.get("basic_metrics", {})
+    dimension_cfg = config.get("dimension", {}) or {}
+    output_precision = int(basic_cfg.get("output_precision", 2))
+
+    # ---- 1. Load dataset ----
+    print(f"[eval] loading dataset: {dataset_path}")
+    raw_samples = _load_dataset(dataset_path)
+    print(f"[eval] loaded {len(raw_samples)} samples")
+
+    # ---- 2. Load model ----
+    print(f"[eval] loading model: {model_path}")
+    from compute.engines.llama_factory.inference import InferenceSession
+    session = InferenceSession()
+    load_result = session.load(
+        model_name_or_path=model_path,
+        adapter_name_or_path=adapter_path,
+        template=template,
+        infer_backend=config.get("infer_backend", "huggingface"),
+        infer_dtype=config.get("infer_dtype", "auto"),
+    )
+    if not load_result.get("loaded"):
+        raise RuntimeError(f"model load failed: {load_result.get('error', 'unknown')}")
+    print(f"[eval] model loaded OK")
+
+    # ---- 3. Run inference on each sample ----
+    samples: list[dict[str, Any]] = []
+    predictions: list[str] = []
+    references: list[str] = []
+    questions: list[str] = []
+
+    total = len(raw_samples)
+    judge_enabled = bool(dimension_cfg.get("eval_model") and dimension_cfg.get("api_url"))
+    print(f"[eval] starting inference on {total} samples, judge={'enabled' if judge_enabled else 'disabled'}")
+
+    for idx, raw in enumerate(raw_samples, start=1):
+        question = _sample_question(raw)
+        reference = _sample_reference(raw)
+        if not question:
+            print(f"[eval] sample {idx}/{total}: skipped (no question)")
+            continue
+
+        # Inference
+        chat_msgs = [{"role": "user", "content": question}]
+        result = session.chat(
+            chat_msgs,
+            temperature=float(config.get("temperature", 0.1)),
+            top_p=float(config.get("top_p", 0.95)),
+            max_new_tokens=int(config.get("max_new_tokens", 512)),
+            do_sample=False,
+        )
+        prediction = result.get("response", "") if not result.get("error") else f"[ERROR] {result['error']}"
+
+        predictions.append(prediction)
+        references.append(reference)
+        questions.append(question)
+
+        # LLM Judge
+        judge_result: dict[str, Any] = {}
+        if judge_enabled:
+            judge_result = _judge_sample(question, reference, prediction, dimension_cfg)
+
+        samples.append({
+            "index": idx,
+            "input": question,
+            "reference_answer": reference,
+            "model_output": prediction,
+            "score": judge_result.get("score"),
+            "max_score": judge_result.get("max_score", dimension_cfg.get("score_max", 5)),
+            "passed": judge_result.get("passed"),
+            "judgement": judge_result.get("judgement"),
+            "evaluation_reason": judge_result.get("evaluation_reason", ""),
+            "error_type": judge_result.get("error_type"),
+            "dimension_scores": [
+                {"name": "judge_score", "score": judge_result.get("score", 0),
+                 "max_score": judge_result.get("max_score", dimension_cfg.get("score_max", 5))},
+            ] if judge_result else [],
+            "status": "completed",
+        })
+
+        progress_pct = int(idx / max(total, 1) * 100)
+        print(f"[eval] sample {idx}/{total} ({progress_pct}%) done")
+
+    # ---- 4. Compute basic metrics ----
+    print(f"[eval] computing basic metrics on {len(predictions)} predictions")
+    metrics_result: dict[str, Any] = {}
+
+    bleu_cfg = basic_cfg.get("bleu", {})
+    if bleu_cfg.get("enabled"):
+        metrics_result["bleu"] = _compute_bleu(references, predictions, int(bleu_cfg.get("ngram", 4)))
+
+    rouge_cfg = basic_cfg.get("rouge", {})
+    if rouge_cfg.get("enabled"):
+        metrics_result["rouge"] = _compute_rouge(references, predictions, rouge_cfg.get("methods"))
+
+    cosine_cfg = basic_cfg.get("cosine", {})
+    if cosine_cfg.get("enabled"):
+        metrics_result["cosine"] = _compute_cosine(references, predictions)
+    metrics_result["exact_match"] = _compute_exact_match(references, predictions)
+    metrics_result["text_similarity"] = _compute_text_similarity(references, predictions)
+
+    # ---- 5. Summarise ----
+    completed = len(samples)
+    if judge_enabled:
+        scored = [s for s in samples if s.get("score") is not None]
+        passed_count = len([s for s in scored if s.get("passed")])
+        avg_score = round(sum(s["score"] for s in scored) / max(len(scored), 1), output_precision)
+        max_score = dimension_cfg.get("score_max", 5)
+        overall_score = round(avg_score / max_score * 100, output_precision)
+        overall_score_max = 100
+        dimension_summary = [{
+            "name": "综合评分",
+            "score": overall_score,
+            "max_score": 100,
+            "pass_rate": round(passed_count / max(completed, 1) * 100, 1),
+        }]
+        overall_evaluation = f"评测完成：{completed} 样本，{passed_count} 通过，平均 {avg_score}/{max_score} 分"
+    else:
+        passed_count = 0
+        enabled_scores = [
+            float(item.get("score") or 0)
+            for item in metrics_result.values()
+            if isinstance(item, dict) and item.get("enabled", True) and item.get("score") is not None
+        ]
+        overall_score = round(sum(enabled_scores) / len(enabled_scores), output_precision) if enabled_scores else 0
+        overall_score_max = 100
+        dimension_summary = [
+            {
+                "name": name,
+                "score": float(item.get("score") or 0),
+                "max_score": 100,
+                "pass_rate": float(item.get("score") or 0),
+            }
+            for name, item in metrics_result.items()
+            if isinstance(item, dict) and item.get("enabled", True) and item.get("score") is not None
+        ]
+        overall_evaluation = f"评测完成：{completed} 样本（未配置 LLM 评委）"
+
+    result = {
+        "overall_score": overall_score,
+        "overall_score_max": overall_score_max,
+        "overall_evaluation": overall_evaluation,
+        "improvement_suggestions": [],
+        "dimension_summary": dimension_summary,
+        "samples": samples,
+        "sample_count": total,
+        "completed_count": completed,
+        "passed_count": passed_count,
+        "basic_metrics": metrics_result,
+    }
+
+    # ---- 6. Write results ----
+    result_path = output_dir / "eval_results.json"
+    result_path.write_text(json.dumps(result, ensure_ascii=False, indent=2), encoding="utf-8")
+    print(f"[eval] results written to {result_path}")
+    return result
+
+
+def main() -> None:
+    import argparse
+    parser = argparse.ArgumentParser(description="YG-FT Evaluation Runner")
+    parser.add_argument("--config", required=True, help="Path to eval config JSON file")
+    args = parser.parse_args()
+
+    config_path = Path(args.config)
+    if not config_path.exists():
+        print(f"FATAL: config file not found: {args.config}", file=sys.stderr)
+        sys.exit(1)
+
+    config = json.loads(config_path.read_text(encoding="utf-8"))
+    start = time.time()
+    try:
+        run_eval(config)
+        elapsed = time.time() - start
+        print(f"[eval] DONE in {elapsed:.1f}s")
+    except Exception as exc:
+        print(f"[eval] FAILED: {exc}", file=sys.stderr)
+        import traceback
+        traceback.print_exc()
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()

compute/engines/llama_factory/inference.py

@@ -59,10 +59,18 @@ class InferenceSession:
                 if adapter_name_or_path:
                     args["adapter_name_or_path"] = adapter_name_or_path
                 args.update(kwargs)
-                model_args, generating_args = get_infer_args(args)
-                self._model = ChatModel(model_args)
-                self._tokenizer = self._model.tokenizer
-                self._generating_args = generating_args
+                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"}
@@ -80,6 +88,16 @@ class InferenceSession:
                     pass
                 self._model = None
                 self._tokenizer = None
+                # 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:
+                    pass
             self._status = "idle"
             self._model_name = ""
             self._adapter_path = ""
@@ -91,11 +109,12 @@ class InferenceSession:
             if self._status != "ready" or self._model is None:
                 return {"error": "model not loaded", "response": ""}
             try:
-                generate_kwargs = {**self._generating_args, "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)
-                formatted = self._model.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+                system = next((m["content"] for m in messages if m["role"] == "system"), None)
+                user_messages = [m for m in messages if m["role"] != "system"]
                 responses = []
-                for response in self._model.stream_chat(formatted, generate_kwargs):
+                for response in self._model.stream_chat(user_messages, system=system, **generate_kwargs):
                     responses.append(response)
                 full_response = "".join(str(r) for r in responses)
                 return {"response": full_response}
@@ -108,9 +127,10 @@ class InferenceSession:
                 yield 'data: {"error": "model not loaded"}\n\n'
                 return
             try:
-                generate_kwargs = {**self._generating_args, **kwargs}
-                formatted = self._model.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
-                for new_text in self._model.stream_chat(formatted, generate_kwargs):
+                generate_kwargs = {**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"]
+                for new_text in self._model.stream_chat(user_messages, system=system, **generate_kwargs):
                     yield new_text
             except Exception as exc:
                 yield 'data: {"error": "' + str(exc) + '"}\n\n'

compute/requirements.txt

@@ -4,4 +4,9 @@ python-multipart>=0.0.9
 pydantic>=2.7.0
 python-dotenv>=1.0.1
 httpx>=0.27.0
-llamafactory
+# 模型评测指标
+sacrebleu>=2.4.0
+rouge-score>=0.1.2
+scikit-learn>=1.3.0
+# LLaMA-Factory 训练引擎
+llamafactory

docker/app/Dockerfile.backend

@@ -11,7 +11,7 @@ RUN pip install --upgrade pip -i https://pypi.tuna.tsinghua.edu.cn/simple \
     && pip install -r /tmp/requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple \
     && rm -f /tmp/requirements.txt
 
-RUN python -c "import fastapi, uvicorn, psycopg, sqlalchemy, redis, jwt, passlib, httpx, alembic; print('backend dependency check ok')"
+RUN python -c "import fastapi, uvicorn, psycopg, psycopg_pool, sqlalchemy, redis, jwt, passlib, httpx, alembic; print('backend dependency check ok')"
 
 RUN mkdir -p /opt/yg-ft/logs/backend /data/yg-ft \
     && chmod -R 0775 /opt/yg-ft /data/yg-ft

docs/模型评测功能总结.md

@@ -0,0 +1,121 @@
+# 模型评测功能总结
+
+本项目（基于 LLaMA-Factory 的微调训练平台）包含 **4 套相对独立** 的模型评测能力，分别面向不同的使用场景：
+
+| 能力 | 入口/目录 | 评测类型 | 打分方式 |
+| --- | --- | --- | --- |
+| 1. 学术 Benchmark 评测 | `llamafactory/eval/` | 选择题式基准（类 MMLU/C-Eval） | 选项匹配 + few-shot |
+| 2. 评估工作台 | `backend/app/api/v1/eval/` | 生成式问答（指令跟随） | BLEU / ROUGE / ExactMatch + 可选 LLM 评审 |
+| 3. 平台评估系统 | `backend/app/api/v1/evaluation/` | 基于评估数据集的问答 | 判卷模型（judge model）打分（0–5 分） |
+| 4. 训练时验证评估 | `backend/app/services/task_runner.py` | 训练验证集 | loss 指标 |
+
+下面分别说明。
+
+---
+
+## 1. 学术 Benchmark 评测（LLaMA-Factory 原生）
+
+面向标准学术选择题基准（如 MMLU、C-Eval 等），复用 LLaMA-Factory 原生的评测框架。
+
+**核心文件**
+- `llamafactory/eval/evaluator.py`：`Evaluator` 类 + `run_eval()` 入口
+- `llamafactory/eval/template.py`：评测 prompt 模板（中/英，含 few-shot 示例构建）
+- `llamafactory/hparams/evaluation_args.py`：`EvaluationArguments` 配置类
+
+**工作流程**
+1. 按 `task`（benchmark 名称）加载数据集，按科目（subject）拆分。
+2. 每个样本构造 few-shot 提示词（`n_shot` 控制示例数，由 `lang` 决定中/英模板），将题干与候选选项拼入 prompt。
+3. 调用模型推理得到预测，与标准答案比对，统计每个科目及整体的 `accuracy`。
+4. 结果写入 `save_dir`，打印各科目与平均准确率。
+
+**关键参数（`EvaluationArguments`）**
+- `task`：基准数据集名
+- `batch_size` / `n_shot` / `lang` / `save_dir` / `seed`
+- `model_name_or_path`、`template`、`trust_remote_code` 等模型相关参数
+
+> 该能力属于框架底层，本平台前端未直接提供操作入口，主要通过配置文件/脚本调用。
+
+---
+
+## 2. 评估工作台（生成式评测 + 指标计算）
+
+后端路由位于 `backend/app/api/v1/eval/__init__.py`，前端称为「评估工作台」。**适用于评测模型的指令跟随与生成质量**，并支持 LLM 作为裁判（LLM-as-a-Judge）。
+
+**API 端点**
+- `GET  /evaluation/tasks`：列出评测任务（`frontend/src/api/evaluation.ts:listTasks`）
+- `POST /evaluation/run`：提交一次评测（`runEval`）
+- `GET  /evaluation/report/{task_id}`：拉取评测报告（`getReport`）
+- `DELETE /evaluation/tasks/{task_id}`：删除任务（`deleteTask`）
+
+**评测流程（`run_eval`）**
+1. 通过 **LLaMA-Factory 数据管道**（`get_dataset`) 加载数据集，支持 `subset` 与抽样（`eval_sample`）。
+2. 用 **原生 transformers** 加载模型在本地做生成推理（单进程顺序生成，便于展示样本）。
+3. 计算客观指标（`compute_score`）：
+   - `BLEU`（sacrebleu）
+   - `ROUGE-1 / ROUGE-2 / ROUGE-L`（rouge-score）
+   - `Exact Match`
+4. **可选 LLM 评审**（judge）：当配置了 `judge_model` / `judge_api_base` / `judge_api_key` 时，调用 OpenAI 兼容接口对每条样本打分（10 分制），并输出 4 个维度与理由：
+   - 核心事实正确性 `factual`
+   - 信息完整性 `completeness`
+   - 无幻觉 `no_hallucination`
+   - 格式合规性 `format`
+   - 综合分 `score` + `reason`
+5. 任务状态持久化在后端 `eval_tasks.json`（支持 running/completed/failed/stopped），前端轮询进度。
+
+**前端页面**
+- `frontend/src/views/evaluation/EvaluateTask.vue`：任务列表、创建评测对话框（选模型、数据集、指标、可选 judge 配置）
+- `frontend/src/views/evaluation/EvaluateReport.vue`：报告页，展示综合得分、BLEU、ROUGE-L、各维度指标及「参考答案 vs 模型预测 vs LLM 评审」对比样例
+
+---
+
+## 3. 平台评估系统（基于评估数据集 + 判卷模型）
+
+后端路由位于 `backend/app/api/v1/evaluation/__init__.py`，是平台业务层自研的评测体系。通过「评估数据集」组织题目，可一次性对 **多个被测模型 + 指定判卷模型** 进行批量评分。
+
+**核心概念（数据模型 `backend/app/models/models.py`）**
+- `EvalDataset`（`models.py:131`）：评估数据集，从项目问答对（`Question`/`Chunk`）中按 `question_type`（mixed/fact/reasoning）选题构建，状态 `pending/running/completed/failed`。
+- `EvalResult`（`models.py:147`）：单条评测结果，含 `judge_score`（0–5 分）、`is_correct`（true/false/partial）、`feedback`、`expected_answer` 等。
+- `Task`（`models.py:184`）：后台任务，`task_type="model-evaluation"`，记录进度与 `model_info`（存放平均分等汇总）。
+
+**评测流程（`process_evaluation_task`，`backend/app/services/task_processor.py:336` 起）**
+1. 加载评估数据集关联的题目，可选带入 `chunk` 上下文（RAG 场景）。
+2. 对每道题，先用 `build_eval_prompt` 组合「上下文 + 题目 + 参考答案」，调用 **判卷模型**（`call_model`，temperature=0.3）生成评分。
+3. `parse_eval_result` 解析出 `score`(0–5)、`is_correct`、`feedback`，写入 `EvalResult`。
+4. 逐题提交进度（`completed_count` / `progress`），支持中途 `stopped`。
+5. 汇总：`avg_score = 总分/有效数 × 20`（换算百分制），`avg_score_5 = 总分/有效数`（5 分制），存入 `task.model_info`。判定规则：得分 **≥3 视为正确**。
+
+**特点**
+- 判卷与被测模型解耦：被测模型给出答案，判卷模型（judge）独立评分，降低自评偏差。
+- 支持失败隔离：单题异常写入 `evaluation_status: failed` 记录而不中断整体任务。
+
+---
+
+## 4. 训练时验证评估
+
+在微调训练任务执行期间，由 `backend/app/services/task_runner.py` 的 `do_eval` 触发：
+
+- 在训练过程中对验证集（validation set）计算 `eval_loss`，用于监控过拟合。
+- 结果回填到 `Task` 的 `loss_info` / `detail`，前端绘制 loss 曲线。
+- 属于训练配套的轻量评估，不参与上述 1–3 的业务评测。
+
+---
+
+## 附属：前端评测相关页面
+
+| 文件 | 作用 |
+| --- | --- |
+| `frontend/src/views/evaluation/EvaluateTask.vue` | 评估工作台：任务列表 + 创建评测 |
+| `frontend/src/views/evaluation/EvaluateReport.vue` | 评估报告：指标卡 + 维度标签 + 对比样例 |
+| `frontend/src/api/evaluation.ts` | 评估工作台接口封装 |
+| 平台评估系统入口 | 评估数据集管理 + 评估任务（model-evaluation）创建与结果查看 |
+
+---
+
+## 小结
+
+- **想要学术榜单式准确率** → 用能力 1（LLaMA-Factory `eval/`）。
+- **想要开放式生成质量（BLEU/ROUGE + LLM 评审）** → 用能力 2（评估工作台 `/evaluation/run`）。
+- **想要基于自有问答数据、用判卷模型批量打分** → 用能力 3（平台评估系统 `model-evaluation` 任务）。
+- **训练过程监控** → 能力 4（`do_eval` 验证集 loss）。
+
+三种业务评测（1/2/3）相互独立，可并存于同一平台；数据模型（`EvalDataset`/`EvalResult`/`Task`）主要服务于能力 3，而能力 2 使用独立的 `eval_tasks.json` 文件持久化。

frontend/src/api/modules/compare.ts

@@ -64,6 +64,27 @@ export const streamChat = async (data: any): Promise<any> => {
   }
 }
 
+/** 真实流式对话 — 使用 fetch 调用后端 SSE 端点，返回 Response 供 ReadableStream 消费 */
+export const streamChatReal = (data: any): Promise<Response> => {
+  const messages = data.messages || []
+  if (!messages.length && data.user_question) {
+    if (data.system_prompt) {
+      messages.push({ role: 'system', content: data.system_prompt })
+    }
+    messages.push({ role: 'user', content: data.user_question })
+  }
+  return fetch('/modelTF/model-compare/stream-chat', {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({
+      messages,
+      temperature: data.temperature ?? 0.7,
+      top_p: data.top_p ?? 0.95,
+      max_tokens: data.max_tokens ?? 2048,
+    }),
+  })
+}
+
 /** 非流式对话（按端口代理） */
 export const chatWithPort = (data: any) => post('/model-compare/chat-with-port', data)
 
@@ -73,8 +94,8 @@ 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 */
-export const preloadLocalModel = (data: any) => post('/model-chat/local/preload', data)
+/** 预加载本地模型（模型加载耗时长，超时 5 分钟） */
+export const preloadLocalModel = (data: any) => post('/model-chat/local/preload', data, { timeout: 300000 })
 
-/** 预加载已训练模型 */
-export const preloadTrainedModel = (data: any) => post('/model-chat/trained/preload', data)
+/** 预加载已训练模型（超时 5 分钟） */
+export const preloadTrainedModel = (data: any) => post('/model-chat/trained/preload', data, { timeout: 300000 })

frontend/src/api/modules/compute.ts

@@ -1,4 +1,4 @@
-import { get, post, put } from '../request'
+import { del, get, post, put } from '../request'
 
 export interface ComputeNode {
   id: string
@@ -95,6 +95,9 @@ export const createComputeNode = (data: ComputeNodePayload) =>
 export const updateComputeNode = (id: string, data: Partial<ComputeNode>) =>
   put<ComputeNode>(`/compute/nodes/${id}`, data)
 
+export const deleteComputeNode = (id: string) =>
+  del<{ deleted: string }>(`/compute/nodes/${id}`)
+
 export const testComputeNode = (id: string) =>
   post<{ node_id: string; success: boolean; latency_ms: number; gpu_count: number; error?: string }>(`/compute/nodes/${id}/test-connection`)
 

frontend/src/api/modules/dataset.ts

@@ -29,6 +29,7 @@ export const uploadDatasetFiles = (datasetId: string | number, files: File[]) =>
   files.forEach((f) => formData.append('files', f))
   return post(`/dataset-manage/upload/${datasetId}`, formData, {
     headers: { 'Content-Type': 'multipart/form-data' },
+    timeout: 120000,
   })
 }
 

frontend/src/composables/useStreamChat.ts

@@ -1,5 +1,5 @@
 import { ref } from 'vue'
-import { streamChat } from '@/api/modules/compare'
+import { streamChat, streamChatReal } from '@/api/modules/compare'
 
 export interface StreamMessage {
   /** 用户问题 */
@@ -20,6 +20,11 @@ export interface StreamMessage {
   error?: string
 }
 
+export interface SendOptions {
+  /** 是否使用 mock 模式（默认 true，向后兼容） */
+  useMock?: boolean
+}
+
 /**
  * 流式对话 composable
  * 移植自原 model-chat.html：
@@ -65,8 +70,10 @@ export function useStreamChat() {
   /**
    * 发起流式对话
    * @param payload 后端请求体 { port, model_name, model_path, system_prompt, user_question, ... }
+   * @param options 可选配置 { useMock?: boolean }
    */
-  async function send(payload: any) {
+  async function send(payload: any, options?: SendOptions) {
+    const useMock = options?.useMock ?? true
     loading.value = true
     message.value = {
       question: payload.user_question || '',
@@ -82,7 +89,10 @@ export function useStreamChat() {
     const UPDATE_INTERVAL = 50 // 50ms 节流
 
     try {
-      const response = await streamChat(payload)
+      const response = useMock
+        ? await streamChat(payload)
+        : await streamChatReal(payload)
+
       if (!response.ok) {
         throw new Error(`HTTP ${response.status}`)
       }

frontend/src/types/index.ts

@@ -132,6 +132,7 @@ export interface FineTuneTask {
   train_dataset_id?: number | string
   auto_merge?: boolean
   output_model_name?: string
+  compute_node_id?: string
   gpus?: number[]
   batch_size?: number
   learning_rate?: number
@@ -355,16 +356,16 @@ export interface GpuInfo {
   power_w: number
   id?: number
   uuid?: string
-  status?: 'idle' | 'busy' | 'warning' | 'offline'
+  status?: 'idle' | 'busy' | 'reserved' | 'warning' | 'offline'
   memory_percent?: number
   power_limit_w?: number
   processes?: GpuProcess[]
   fan_speed?: number
   clock_mhz?: number
-  driver_version?: string
   node_id?: string
   node_code?: string
   node_name?: string
+  driver_version?: string
 }
 
 export interface SystemInfo {

frontend/src/views/approvals/ApprovalInstanceView.vue

@@ -3,7 +3,8 @@ import { onMounted, reactive, ref } from 'vue'
 import { ElMessage } from 'element-plus'
 import DataTablePage from '@/components/DataTablePage.vue'
 import { getApprovalInstances, decideApproval, type ApprovalInstance } from '@/api/modules/approval'
-import { getUsers, type SystemUser } from '@/api/modules/system'
+import { getUsers } from '@/api/modules/system'
+import type { SystemUser } from '@/types'
 
 const loading = ref(false)
 const instances = ref<ApprovalInstance[]>([])
@@ -48,6 +49,10 @@ function openDecide(inst: ApprovalInstance) {
   showDecide.value = true
 }
 
+function asApprovalInstance(row: unknown): ApprovalInstance {
+  return row as ApprovalInstance
+}
+
 async function submitDecision() {
   if (!current.value) return
   if (!decision.value.approver_id) {
@@ -72,7 +77,7 @@ onMounted(() => {
 
 <template>
   <div class="page">
-    <DataTablePage title="审批实例" :data="instances" :loading="loading" searchable search-fields="resource_type,resource_id">
+    <DataTablePage title="审批实例" :data="instances" :loading="loading" searchable :search-fields="['resource_type', 'resource_id']">
       <template #toolbar-extra>
         <el-select v-model="statusFilter" placeholder="状态" clearable style="width: 140px" @change="load">
           <el-option v-for="s in statusOptions" :key="s.value" :label="s.label" :value="s.value" />
@@ -82,14 +87,14 @@ onMounted(() => {
         <el-table-column prop="resource_type" label="资源类型" min-width="120" />
         <el-table-column prop="resource_id" label="资源 ID" min-width="160" show-overflow-tooltip />
         <el-table-column prop="applicant_id" label="申请人" min-width="120">
-          <template #default="{ row }">{{ userName(row.applicant_id) }}</template>
+          <template #default="{ row }">{{ userName(asApprovalInstance(row).applicant_id) }}</template>
         </el-table-column>
         <el-table-column prop="status" label="状态" min-width="100" />
         <el-table-column prop="current_step" label="当前步骤" min-width="100" />
         <el-table-column prop="create_time" label="创建时间" min-width="180" />
       </template>
       <template #actions="{ row }">
-        <el-button v-if="row.status === 'pending'" link type="primary" @click="openDecide(row)">审批</el-button>
+        <el-button v-if="asApprovalInstance(row).status === 'pending'" link type="primary" @click="openDecide(asApprovalInstance(row))">审批</el-button>
       </template>
     </DataTablePage>
     <el-dialog v-model="showDecide" title="审批决策" width="480px">

frontend/src/views/compute/ComputeNodesView.vue

@@ -1,12 +1,12 @@
 <script setup lang="ts">
 import { computed, onMounted, onUnmounted, reactive, ref, watch } from 'vue'
 import { useRoute, useRouter } from 'vue-router'
-import { ElMessage } from 'element-plus'
+import { ElMessage, ElMessageBox } from 'element-plus'
 import {
   checkNodeReplicaDrift,
   createComputeNode,
+  deleteComputeNode,
   disableComputeNode,
-  drainComputeNode,
   enableComputeNode,
   getComputeGpus,
   getComputeNodes,
@@ -129,11 +129,10 @@ async function changeTab(name: string | number) {
   await router.replace({ path: '/compute', query: { tab: String(name) } })
 }
 
-async function handleNodeAction(action: 'enable' | 'disable' | 'drain' | 'test', node: ComputeNode) {
+async function handleNodeAction(action: 'enable' | 'disable' | 'test', node: ComputeNode) {
   const nodeId = String(node.id)
   if (action === 'enable') await enableComputeNode(nodeId)
   if (action === 'disable') await disableComputeNode(nodeId)
-  if (action === 'drain') await drainComputeNode(nodeId)
   if (action === 'test') {
     const result = await testComputeNode(nodeId)
     if (result.success) {
@@ -145,6 +144,27 @@ async function handleNodeAction(action: 'enable' | 'disable' | 'drain' | 'test',
   await load()
 }
 
+async function handleDeleteNode(node: ComputeNode) {
+  try {
+    await ElMessageBox.confirm(
+      `确定删除算力节点「${node.name || node.code}」吗？节点删除后，其 GPU 设备和资源副本记录也会一并移除。`,
+      '删除算力节点',
+      { type: 'warning', confirmButtonText: '删除', cancelButtonText: '取消' },
+    )
+  } catch {
+    return
+  }
+  try {
+    await deleteComputeNode(String(node.id))
+    ElMessage.success('算力节点已删除')
+    if (selectedNodeId.value === node.id) selectedNodeId.value = ''
+    await load({ showButtonLoading: true })
+  } catch (err: any) {
+    const message = err?.response?.data?.detail?.message || err?.response?.data?.message || '删除算力节点失败'
+    ElMessage.error(message)
+  }
+}
+
 async function handleReplicaDriftCheck() {
   if (!selectedNodeId.value) return
   checkingReplicas.value = true
@@ -355,7 +375,7 @@ onUnmounted(() => {
               <el-button size="small" @click="handleNodeAction('test', asComputeNode(row))">测试</el-button>
               <el-button v-if="row.enabled" size="small" @click="handleNodeAction('disable', asComputeNode(row))">停用</el-button>
               <el-button v-else size="small" type="primary" @click="handleNodeAction('enable', asComputeNode(row))">启用</el-button>
-              <el-button size="small" type="warning" plain @click="handleNodeAction('drain', asComputeNode(row))">维护</el-button>
+              <el-button size="small" type="danger" plain @click="handleDeleteNode(asComputeNode(row))">删除</el-button>
             </template>
           </el-table-column>
         </el-table>

frontend/src/views/dataset/DatasetCreateView.vue

@@ -50,6 +50,20 @@ const rules: FormRules = {
 }
 
 /** 处理文件选择（替换模式：新文件覆盖旧文件） */
+function parseDatasetRecordValues(text: string, fileName: string): unknown[] {
+  const content = text.trim()
+  if (!content) return []
+  if (fileName.toLowerCase().endsWith('.json')) {
+    const parsed = JSON.parse(content)
+    return Array.isArray(parsed) ? parsed : [parsed]
+  }
+  return content
+    .split('\n')
+    .map((line) => line.trim())
+    .filter(Boolean)
+    .map((line) => JSON.parse(line))
+}
+
 async function handleFileChange(uploadFile: UploadFile) {
   const raw = uploadFile.raw
   if (!raw) return
@@ -70,25 +84,19 @@ async function handleFileChange(uploadFile: UploadFile) {
 async function analyzeFile(file: File) {
   try {
     const text = await file.text()
-    const lines = text.trim().split('\n').filter(Boolean)
-    fileCount.value = lines.length
+    const records = parseDatasetRecordValues(text, file.name)
+    fileCount.value = records.length
 
     // Alpaca 格式校验：每行 JSON 须含 instruction 字段
-    let validCount = 0
-    for (const line of lines) {
-      try {
-        const obj = JSON.parse(line)
-        if (obj.instruction !== undefined) validCount++
-      } catch {
-        // 非 JSON 行（如纯 JSONL 多行结构）
-      }
-    }
-    if (validCount > 0 && validCount === lines.length) {
+    const validCount = records.filter(
+      (obj) => obj && typeof obj === 'object' && 'instruction' in obj,
+    ).length
+    if (validCount > 0 && validCount === records.length) {
       formatValid.value = true
       formatMessage.value = `符合 Alpaca 格式（含 instruction 字段）`
     } else if (validCount > 0) {
       formatValid.value = true
-      formatMessage.value = `部分符合 Alpaca 格式（${validCount}/${lines.length}）`
+      formatMessage.value = `部分符合 Alpaca 格式（${validCount}/${records.length}）`
     } else {
       formatValid.value = false
       formatMessage.value = '未检测到标准 Alpaca 格式（缺少 instruction 字段），仍可上传'

frontend/src/views/eval/DimensionCreateView.vue

@@ -79,7 +79,9 @@ async function loadEditData() {
 async function loadModels() {
   try {
     const all = (await getModelList()) || []
-    evalModels.value = all.filter((m) => m.purpose === 'evaluation')
+    evalModels.value = all.filter(
+      (m) => m.purpose === 'evaluation' || (m.model_source === 'api' && !!m.api_url),
+    )
   } catch {
     evalModels.value = []
   }

frontend/src/views/eval/EvalCreateView.vue

@@ -11,6 +11,7 @@ import { createDimension, startEval } from '@/api/modules/eval'
 import { getTrainedModels, getModelList } from '@/api/modules/model'
 import { getDatasetList } from '@/api/modules/dataset'
 import { getSystemInfo } from '@/api/modules/system'
+import { getComputeNodes, type ComputeNode } from '@/api/modules/compute'
 import type { DatasetItem, Dimension, GpuInfo, ModelItem, TrainedModel } from '@/types'
 
 type StepExposed = { validate: () => Promise<boolean> }
@@ -84,15 +85,26 @@ async function loadData() {
     getDatasetList(),
     getSystemInfo(),
     getModelList(),
+    getComputeNodes(),
   ])
 
   if (results[0].status === 'fulfilled') trainedModels.value = results[0].value?.models || []
   if (results[1].status === 'fulfilled') {
     evalDatasets.value = (results[1].value || []).filter((dataset) => dataset.type === 'eval')
   }
-  if (results[2].status === 'fulfilled') gpus.value = results[2].value?.gpu || []
+  if (results[2].status === 'fulfilled') {
+    const allGpus: GpuInfo[] = results[2].value?.gpu || []
+    const nodes: ComputeNode[] = (results[4].status === 'fulfilled' ? results[4].value : []) || []
+    const onlineIds = new Set(nodes.filter((n) => n.enabled && n.scheduler_status === 'online').map((n) => n.id))
+    // Only show idle GPUs from online compute nodes
+    gpus.value = allGpus.filter(
+      (g) => g.status === 'idle' && (!g.node_id || onlineIds.has(g.node_id)),
+    )
+  }
   if (results[3].status === 'fulfilled') {
-    evalModels.value = (results[3].value || []).filter((model) => model.purpose === 'evaluation')
+    evalModels.value = (results[3].value || []).filter(
+      (model) => model.purpose === 'evaluation' || (model.model_source === 'api' && !!model.api_url),
+    )
   }
 
   const failedCount = results.filter((result) => result.status === 'rejected').length

frontend/src/views/eval/EvalDetailView.vue

@@ -1,9 +1,10 @@
 <script setup lang="ts">
-import { computed, onMounted, ref } from 'vue'
+import { computed, onMounted, onUnmounted, ref } from 'vue'
 import { useRoute } from 'vue-router'
 import PageCard from '@/components/PageCard.vue'
 import ModelStatusTag from '@/components/ModelStatusTag.vue'
 import { getEvalDetail } from '@/api/modules/eval'
+import { usePolling } from '@/composables/usePolling'
 import type { EvalSampleResult, EvalTaskDetail } from '@/types'
 
 const route = useRoute()
@@ -16,6 +17,7 @@ const keyword = ref('')
 const judgementFilter = ref('')
 const currentPage = ref(1)
 const pageSize = ref(10)
+const ACTIVE_STATUSES = new Set(['pending', 'queued', 'running'])
 
 const filteredSamples = computed(() => {
   const normalizedKeyword = keyword.value.trim().toLowerCase()
@@ -74,8 +76,8 @@ function resetPage() {
   currentPage.value = 1
 }
 
-async function loadDetail() {
-  loading.value = true
+async function loadDetail(options: { silent?: boolean } = {}) {
+  if (!options.silent) loading.value = true
   loadError.value = ''
   try {
     detail.value = await getEvalDetail(taskId)
@@ -83,11 +85,29 @@ async function loadDetail() {
     detail.value = null
     loadError.value = '评测详情加载失败，请稍后重试。'
   } finally {
-    loading.value = false
+    if (!options.silent) loading.value = false
   }
 }
 
-onMounted(loadDetail)
+const { start: startPolling, stop: stopPolling } = usePolling(
+  async () => {
+    await loadDetail({ silent: true })
+    if (!ACTIVE_STATUSES.has(String(detail.value?.status || ''))) {
+      stopPolling()
+    }
+  },
+  5000,
+  { immediate: false },
+)
+
+onMounted(async () => {
+  await loadDetail()
+  if (ACTIVE_STATUSES.has(String(detail.value?.status || ''))) {
+    startPolling()
+  }
+})
+
+onUnmounted(stopPolling)
 </script>
 
 <template>
@@ -113,7 +133,7 @@ onMounted(loadDetail)
       <i class="fa fa-exclamation-circle" aria-hidden="true" />
       <h2>无法加载评测详情</h2>
       <p>{{ loadError }}</p>
-      <el-button type="primary" @click="loadDetail">重新加载</el-button>
+      <el-button type="primary" @click="() => loadDetail()">重新加载</el-button>
     </div>
 
     <template v-else-if="detail">

frontend/src/views/eval/EvalView.vue

@@ -1,9 +1,10 @@
 <script setup lang="ts">
-import { ref, onMounted } from 'vue'
+import { ref, onMounted, onUnmounted } from 'vue'
 import { useRouter } from 'vue-router'
 import { ElMessage, ElMessageBox } from 'element-plus'
 import DataTablePage from '@/components/DataTablePage.vue'
 import ModelStatusTag from '@/components/ModelStatusTag.vue'
+import { usePolling } from '@/composables/usePolling'
 import {
   getEvalList,
   deleteEval,
@@ -23,14 +24,16 @@ const leaderboard = ref([
   { rank: 3, name: 'Qwen-Max', score: 85.3 },
 ])
 
-async function loadEvalList() {
-  evalLoading.value = true
+const ACTIVE_STATUSES = new Set(['pending', 'queued', 'running'])
+
+async function loadEvalList(options: { silent?: boolean } = {}) {
+  if (!options.silent) evalLoading.value = true
   try {
     evalList.value = (await getEvalList()) || []
   } catch {
     evalList.value = []
   } finally {
-    evalLoading.value = false
+    if (!options.silent) evalLoading.value = false
   }
 }
 
@@ -54,8 +57,26 @@ function handleViewDetail(row: any) {
   router.push({ name: 'model-eval-detail', params: { id: row.id } })
 }
 
-onMounted(() => {
-  loadEvalList()
+const { start: startPolling, stop: stopPolling } = usePolling(
+  async () => {
+    await loadEvalList({ silent: true })
+    if (!evalList.value.some((item) => ACTIVE_STATUSES.has(String(item.status || '')))) {
+      stopPolling()
+    }
+  },
+  5000,
+  { immediate: false },
+)
+
+onMounted(async () => {
+  await loadEvalList()
+  if (evalList.value.some((item) => ACTIVE_STATUSES.has(String(item.status || '')))) {
+    startPolling()
+  }
+})
+
+onUnmounted(() => {
+  stopPolling()
 })
 </script>
 

frontend/src/views/eval/create/BasicMetricSetupStep.vue

@@ -56,9 +56,9 @@ defineExpose({ validate })
     </el-form-item>
     <el-form-item v-if="form.rouge_enabled" label="ROUGE methods">
       <el-checkbox-group v-model="form.rouge_methods">
-        <el-checkbox value="rouge_1">ROUGE-1</el-checkbox>
-        <el-checkbox value="rouge_2">ROUGE-2</el-checkbox>
-        <el-checkbox value="rouge_l">ROUGE-L</el-checkbox>
+        <el-checkbox value="rouge1">ROUGE-1</el-checkbox>
+        <el-checkbox value="rouge2">ROUGE-2</el-checkbox>
+        <el-checkbox value="rougeL">ROUGE-L</el-checkbox>
       </el-checkbox-group>
     </el-form-item>
 

frontend/src/views/eval/create/EvalTaskSetupStep.vue

@@ -103,10 +103,10 @@ defineExpose({ validate })
     <el-form-item label="选择 GPU" prop="gpu_id">
       <el-select v-model="form.gpu_id" placeholder="请选择 GPU" style="width: 100%" :loading="loading">
         <el-option
-          v-for="(gpu, index) in gpus"
-          :key="index"
-          :label="`${gpu.name} (GPU ${index})`"
-          :value="index"
+          v-for="gpu in gpus"
+          :key="gpu.id"
+          :label="`${gpu.name} (GPU ${gpu.id})`"
+          :value="gpu.id ?? 0"
         />
       </el-select>
     </el-form-item>

frontend/src/views/fine-tune/FineTuneCreateView.vue

@@ -37,7 +37,7 @@ const models = ref<ModelItem[]>([])
 const datasets = ref<DatasetItem[]>([])
 const gpus = ref<GpuInfo[]>([])
 const computeNodes = ref<Array<{ id: string; scheduler_status?: string }>>([])
-const selectedGpuId = ref<number | null>(null)
+const selectedGpuKeys = ref<string[]>([])
 
 /** Only show GPUs from nodes that are online or draining */
 const availableGpus = computed(() => {
@@ -74,7 +74,13 @@ const selectedModel = computed(() => models.value.find((model) => model.id === f
 const modelDialogTitle = computed(() => selectedModel.value?.name || '')
 
 /** 训练命令与提交载荷共用同一份表单模型。 */
-const selectedGpuIds = computed(() => (selectedGpuId.value != null ? [selectedGpuId.value] : []))
+const selectedGpus = computed(() =>
+  selectedGpuKeys.value
+    .map((key) => availableGpus.value.find((gpu) => gpuKey(gpu) === key))
+    .filter((gpu): gpu is GpuInfo => Boolean(gpu)),
+)
+const selectedComputeNodeId = computed(() => selectedGpus.value[0]?.node_id)
+const selectedGpuIds = computed(() => selectedGpus.value.map((gpu) => Number(gpu.id)))
 const commandPreview = computed(() => buildFineTuneCommand(form, selectedGpuIds.value))
 
 const remoteCommandPreview = computed(() => {
@@ -83,9 +89,32 @@ const remoteCommandPreview = computed(() => {
   return preflightResult.value?.preview?.command_text || ''
 })
 
-/** GPU 单选切换（每次只选中一张 GPU） */
-function toggleGpu(gpuId: number) {
-  selectedGpuId.value = selectedGpuId.value === gpuId ? null : gpuId
+function gpuKey(gpu: GpuInfo) {
+  return `${gpu.node_id || 'local'}:${gpu.id ?? gpu.uuid ?? gpu.name}`
+}
+
+function isGpuUnavailable(gpu: GpuInfo) {
+  return gpu.status === 'busy' || gpu.status === 'reserved' || gpu.status === 'offline'
+}
+
+function isGpuSelected(gpu: GpuInfo) {
+  return selectedGpuKeys.value.includes(gpuKey(gpu))
+}
+
+/** GPU 多选切换：单个任务只允许选择同一算力节点内的空闲卡。 */
+function toggleGpu(gpu: GpuInfo) {
+  if (isGpuUnavailable(gpu) || gpu.id == null) return
+  const key = gpuKey(gpu)
+  if (isGpuSelected(gpu)) {
+    selectedGpuKeys.value = selectedGpuKeys.value.filter((item) => item !== key)
+    return
+  }
+  if (selectedComputeNodeId.value && gpu.node_id && selectedComputeNodeId.value !== gpu.node_id) {
+    selectedGpuKeys.value = [key]
+    ElMessage.info('已切换到新的算力节点，之前选择的 GPU 已清空')
+    return
+  }
+  selectedGpuKeys.value = [...selectedGpuKeys.value, key]
 }
 
 function gpuUsageWidth(percent: number) {
@@ -178,8 +207,8 @@ async function loadGpus() {
     const [sys, nodes] = await Promise.all([getSystemInfo(), getComputeNodes().catch(() => [])])
     gpus.value = sys?.gpu || []
     computeNodes.value = nodes || []
-    // Default select first available GPU
-    if (availableGpus.value.length > 0) selectedGpuId.value = availableGpus.value[0].id ?? null
+    const firstIdle = availableGpus.value.find((gpu) => !isGpuUnavailable(gpu) && gpu.id != null)
+    if (firstIdle) selectedGpuKeys.value = [gpuKey(firstIdle)]
   } catch {
     gpus.value = []
   }
@@ -189,8 +218,8 @@ async function handleSubmit() {
   if (!formRef.value) return
   await formRef.value.validate(async (valid) => {
     if (!valid) return
-    if (selectedGpuId.value == null) {
-      ElMessage.warning('请选择一个 GPU')
+    if (!selectedGpuIds.value.length) {
+      ElMessage.warning('请至少选择一张空闲 GPU')
       return
     }
     submitting.value = true
@@ -207,7 +236,7 @@ async function handleSubmit() {
         return
       }
 
-      const payload = buildFineTunePayload(form, selectedGpuIds.value)
+      const payload = buildFineTunePayload(form, selectedGpuIds.value, selectedComputeNodeId.value)
       const preflight = await runPreflight(payload)
       if (!preflight?.valid) {
         ElMessage.error('训练预检未通过，请先处理预检问题')
@@ -232,7 +261,7 @@ async function handleSubmit() {
   })
 }
 
-async function runPreflight(payload = buildFineTunePayload(form, selectedGpuIds.value)) {
+async function runPreflight(payload = buildFineTunePayload(form, selectedGpuIds.value, selectedComputeNodeId.value)) {
   preflightLoading.value = true
   try {
     const result = await preflightFineTune(payload)
@@ -261,8 +290,8 @@ async function handlePreflightClick() {
   if (!formRef.value) return
   await formRef.value.validate(async (valid) => {
     if (!valid) return
-    if (selectedGpuId.value == null) {
-      ElMessage.warning('请选择一个 GPU')
+    if (!selectedGpuIds.value.length) {
+      ElMessage.warning('请至少选择一张空闲 GPU')
       return
     }
     await runPreflight()
@@ -297,12 +326,16 @@ onMounted(() => {
         <el-divider content-position="left">训练配置</el-divider>
         <el-form-item label="GPU 硬件">
           <div class="gpu-list">
+            <div class="gpu-selection-summary">
+              已选择 {{ selectedGpuIds.length }} 张 GPU
+              <template v-if="selectedGpus[0]?.node_code"> · {{ selectedGpus[0].node_code }}</template>
+            </div>
             <div
               v-for="gpu in availableGpus"
-              :key="gpu.id"
+              :key="gpuKey(gpu)"
               class="gpu-card"
-              :class="{ active: selectedGpuId === gpu.id, 'is-busy': gpu.gpu_percent > 80 }"
-              @click="toggleGpu(gpu.id!)"
+              :class="{ active: isGpuSelected(gpu), 'is-busy': isGpuUnavailable(gpu), 'is-disabled': isGpuUnavailable(gpu) }"
+              @click="toggleGpu(gpu)"
             >
               <div class="gpu-card-top">
                 <div class="gpu-title">
@@ -312,7 +345,7 @@ onMounted(() => {
                   </span>
                   <span class="gpu-name">{{ gpu.name }}</span>
                 </div>
-                <span class="gpu-usage">{{ gpu.gpu_percent }}%</span>
+                <span class="gpu-usage">{{ isGpuUnavailable(gpu) ? gpu.status : `${gpu.gpu_percent}%` }}</span>
               </div>
               <div class="gpu-usage-bar">
                 <span :style="{ width: gpuUsageWidth(gpu.gpu_percent) }" />
@@ -578,6 +611,13 @@ onMounted(() => {
   width: 100%;
 }
 
+.gpu-selection-summary {
+  grid-column: 1 / -1;
+  color: #64748b;
+  font-size: 12px;
+  line-height: 20px;
+}
+
 .gpu-card {
   border: 1px solid #e5e7eb;
   border-radius: 6px;
@@ -627,6 +667,11 @@ onMounted(() => {
       background: #dc2626;
     }
   }
+
+  &.is-disabled {
+    cursor: not-allowed;
+    opacity: 0.72;
+  }
 }
 
 .gpu-card-top {

frontend/src/views/fine-tune/fineTuneFormModel.ts

@@ -65,6 +65,7 @@ export function createDefaultFineTuneForm(): FineTuneFormModel {
 export function buildFineTunePayload(
   form: FineTuneFormModel,
   gpus: number[],
+  computeNodeId?: string,
 ): Omit<FineTuneStartPayload, 'task_id'> {
   return {
     name: form.name,
@@ -77,6 +78,7 @@ export function buildFineTunePayload(
     train_dataset_id: form.train_dataset_id,
     auto_merge: form.train_type === 'SFT' && form.auto_merge,
     output_model_name: form.name,
+    compute_node_id: computeNodeId,
     batch_size: form.batch_size,
     learning_rate: form.learning_rate,
     n_epochs: form.n_epochs,

frontend/src/views/inference/InferenceChatView.vue

@@ -10,8 +10,8 @@ import type { CompareTask, LoadedModel } from '@/types'
 const route = useRoute()
 const router = useRouter()
 const taskId = route.params.id as string
-/** 是否为 mock 直通模式（新建推理假数据进入，不走真实任务接口） */
-const isMock = taskId === 'mock'
+/** 是否为 mock 模式（新建推理无真实 taskId 或明确为 mock 时进入 mock 模式） */
+const isMock = taskId === 'mock' || !taskId || taskId === 'unknown'
 /** 当前对话使用的模型名 */
 const modelName = ref(route.query.model as string || '')
 
@@ -88,30 +88,20 @@ async function handleSend() {
     return
   }
 
-  // 真实模式：获取已启动模型的端口/路径
-  const models = parseLoadedModels(task.value)
-  const target = models[0]
-  if (!target) {
-    ElMessage.error('未找到已启动的模型')
-    assistantMsg.content = '未找到已启动的模型，请先返回列表加载模型'
-    assistantMsg.done = true
-    assistantMsg.isStreaming = false
-    return
-  }
-
-  // 流式状态变化时只同步当前回复，避免固定定时器空转。
+  // 真实模式：通过后端 SSE 流式代理到算力节点进行推理
   activeAssistant = assistantMsg
 
-  await send({
-    port: target.port,
-    model_name: target.model_name,
-    model_path: '',
-    system_prompt: systemPrompt.value,
-    user_question: question,
-    temperature: temperature.value,
-    top_p: top_p.value,
-    max_tokens: maxTokens.value,
-  })
+  await send(
+    {
+      model_path: route.query.model_path as string || '',
+      system_prompt: systemPrompt.value,
+      user_question: question,
+      temperature: temperature.value,
+      top_p: top_p.value,
+      max_tokens: maxTokens.value,
+    },
+    { useMock: false },
+  )
 
   // 完成后同步最终内容
   assistantMsg.content = message.value.displayContent || message.value.error || '(无回复)'

frontend/src/views/inference/InferenceCreateView.vue

@@ -5,6 +5,8 @@ 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 type { ModelItem, TrainedModel, GpuInfo } from '@/types'
 
 const router = useRouter()
@@ -15,6 +17,7 @@ const startupStatus = ref('')
 const dbModels = ref<ModelItem[]>([])
 const trainedModels = ref<TrainedModel[]>([])
 const gpus = ref<GpuInfo[]>([])
+const computeNodes = ref<ComputeNode[]>([])
 
 /** 可选模型（下拉用，区分本地/已训练两类） */
 interface SelectableModel {
@@ -54,7 +57,17 @@ const trainedOptions = computed<SelectableModel[]>(() =>
   })),
 )
 
-/** key → 模型映射，便于取选中项 */
+/** 仅显示在线算力节点上的空闲 GPU */
+const onlineNodeIds = computed(() => new Set(
+  computeNodes.value
+    .filter((n) => n.enabled && n.scheduler_status === 'online')
+    .map((n) => n.id),
+))
+const idleGpus = computed(() =>
+  gpus.value.filter(
+    (g) => g.status === 'idle' && (!g.node_id || onlineNodeIds.value.has(g.node_id)),
+  ),
+)
 const modelMap = computed<Record<string, SelectableModel>>(() => {
   const map: Record<string, SelectableModel> = {}
   for (const m of [...dbOptions.value, ...trainedOptions.value]) map[m.key] = m
@@ -90,12 +103,56 @@ async function handleSubmit() {
     submitting.value = true
     startupStatus.value = '正在启动模型服务...'
     try {
-      // 当前为 mock 环境：不创建任务、不启动后端服务，
-      // 用假数据直通进入对话界面（模型名通过 query 传递）。
-      // 接入真实后端后，可在此恢复 createCompare / startModelsInBackground / monitorStartup 流程。
-      await new Promise((resolve) => setTimeout(resolve, 1200))
+      // Step 1: 将模型加载到算力节点
+      const preloadPayload = {
+        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
+      const taskResult = await createCompare({
+        name: form.name || m.name,
+        description: form.description,
+        models: [
+          {
+            model_id: String(m.id),
+            model_name: m.name,
+            model_path: m.model_path,
+            source: m.source,
+            gpu_id: form.gpu_id,
+          },
+        ],
+      })
+      const taskId = taskResult?.id || 'unknown'
 
       ElMessage.success('模型已启动')
+      router.push({
+        path: `/model-inference/chat/${taskId}`,
+        query: {
+          model: m.name,
+          source: m.source,
+          model_path: m.model_path,
+        },
+      })
+    } catch (e: any) {
+      // 真实 API 失败时回退到 mock 模式（方便无算力节点的开发调试）
+      const m = selectedModel.value!
+      const reason = e?.message || e?.toString() || '未知错误'
+      ElMessage.warning(`推理服务启动失败：${reason}，进入 mock 演示模式`)
       router.push({
         path: '/model-inference/chat/mock',
         query: { model: m.name },
@@ -113,16 +170,18 @@ function handleCancel() {
 
 async function loadData() {
   try {
-    const [db, trained, sys] = await Promise.all([
+    const [db, trained, sys, nodes] = await Promise.all([
       getModelList(),
       getTrainedModels(),
       getSystemInfo(),
+      getComputeNodes(),
     ])
     dbModels.value = db || []
     trainedModels.value = trained?.models || []
     gpus.value = sys?.gpu || []
-    // 默认选中第一个 GPU
-    if (gpus.value.length > 0) form.gpu_id = 0
+    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
   } catch {
     // ignore
   }
@@ -172,10 +231,10 @@ onMounted(loadData)
       <el-form-item label="GPU">
         <el-select v-model="form.gpu_id" style="width: 400px">
           <el-option
-            v-for="(g, idx) in gpus"
-            :key="idx"
-            :label="`${g.name} (GPU${idx})`"
-            :value="idx"
+            v-for="g in idleGpus"
+            :key="g.id ?? 0"
+            :label="`${g.name} (GPU${g.id ?? 0}) [空闲]`"
+            :value="g.id ?? 0"
           />
         </el-select>
       </el-form-item>

frontend/src/views/inference/InferenceListView.vue

@@ -7,10 +7,8 @@ import { usePolling } from '@/composables/usePolling'
 import {
   getCompareList,
   deleteCompare,
-  getCompare,
   loadCompare,
   unloadCompare,
-  stopModelByPid,
 } from '@/api/modules/compare'
 import type { CompareTask, LoadedModel } from '@/types'
 import { statusLabel, statusTagType } from '@/utils/status'
@@ -86,26 +84,19 @@ async function handleLoad(row: any) {
   delayedRefreshTimer = setTimeout(loadData, 1000)
 }
 
-/** 卸载推理任务 */
+/** 释放推理任务（停止模型服务，释放算力节点 GPU 显存） */
 async function handleUnload(row: any) {
-  await ElMessageBox.confirm('确定要停止模型服务吗？', '确认停止', { type: 'warning' })
+  await ElMessageBox.confirm('确定要释放模型服务吗？将停止模型进程并释放 GPU 显存。', '确认释放', { type: 'warning' })
   await unloadCompare(row.id)
-  ElMessage.success('已停止模型服务')
+  ElMessage.success('已释放模型服务')
   loadData()
 }
 
-/** 删除（先停止进程） */
+/** 删除（先释放算力节点再删除记录） */
 async function handleDelete(row: any) {
-  // 先尝试停止已加载的模型进程
-  const task = await getCompare(row.id).catch(() => null)
-  if (task?.load_status) {
-    const models = parseLoadedModels(task as CompareTask)
-    for (const m of models) {
-      if (m.pid) {
-        await stopModelByPid(m.pid).catch(() => {})
-      }
-    }
-  }
+  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)
@@ -180,7 +171,7 @@ onUnmounted(() => {
             <i class="fa fa-comments-o" style="margin-right: 4px" />对话
           </el-button>
           <el-button type="warning" link size="small" @click="handleUnload(row)">
-            <i class="fa fa-stop-circle-o" style="margin-right: 4px" />停止
+            <i class="fa fa-stop-circle-o" style="margin-right: 4px" />释放
           </el-button>
         </template>
         <template v-else>

frontend/src/views/projects/ProjectDetailView.vue

@@ -5,7 +5,8 @@ import { ElMessage } from 'element-plus'
 import DataTablePage from '@/components/DataTablePage.vue'
 import AclDialog from '@/components/AclDialog.vue'
 import { getProject, getProjectMembers, addProjectMember, removeProjectMember, type Project, type ProjectMember } from '@/api/modules/project'
-import { getUsers, type SystemUser } from '@/api/modules/system'
+import { getUsers } from '@/api/modules/system'
+import type { SystemUser } from '@/types'
 
 const route = useRoute()
 const router = useRouter()
@@ -56,6 +57,10 @@ async function removeMember(userId: string) {
   load()
 }
 
+function asProjectMember(row: unknown): ProjectMember {
+  return row as ProjectMember
+}
+
 onMounted(() => {
   loadUsers()
   load()
@@ -94,7 +99,7 @@ onMounted(() => {
           <el-table-column prop="create_time" label="加入时间" min-width="180" />
         </template>
         <template #actions="{ row }">
-          <el-button link type="danger" @click="removeMember(row.user_id)">移除</el-button>
+          <el-button link type="danger" @click="removeMember(asProjectMember(row).user_id)">移除</el-button>
         </template>
       </DataTablePage>
     </el-card>

frontend/src/views/projects/ProjectListView.vue

@@ -52,6 +52,10 @@ function openDetail(id: string) {
   router.push(`/projects/${id}`)
 }
 
+function asProject(row: unknown): Project {
+  return row as Project
+}
+
 async function submitCreate() {
   if (!form.value.name || !form.value.tenant_id) {
     ElMessage.warning('请填写项目名与编码ID')
@@ -87,7 +91,7 @@ onMounted(() => {
 
 <template>
   <div class="page">
-    <DataTablePage title="项目空间" :data="projects" :loading="loading" searchable search-fields="name,code">
+    <DataTablePage title="项目空间" :data="projects" :loading="loading" searchable :search-fields="['name', 'code']">
       <template #toolbar-extra>
         <el-select v-model="tenantId" placeholder="租户" style="width: 160px" @change="load">
           <el-option v-for="t in tenantOptions" :key="t.value" :label="t.label" :value="t.value" />

frontend/src/views/system/UserSettingsView.vue

@@ -54,6 +54,18 @@ function isSelf(row: SystemUser) {
   return row.username === currentUsername.value
 }
 
+function asSystemUser(row: unknown): SystemUser {
+  return row as SystemUser
+}
+
+function userPermissions(row: unknown): PermissionCode[] {
+  return (asSystemUser(row).permissions || []) as PermissionCode[]
+}
+
+function permissionLabel(code: PermissionCode): string {
+  return PERMISSION_LABELS[code] || code
+}
+
 // ---------- 启停 ----------
 async function toggleStatus(row: SystemUser, next: boolean) {
   const nextStatus: UserStatus = next ? 'active' : 'disabled'
@@ -162,31 +174,31 @@ async function removeUser(row: SystemUser) {
       <el-table-column prop="role" label="角色" width="120" />
       <el-table-column label="状态" width="130">
         <template #default="{ row }">
-          <el-tag :type="statusTagType(row.status)" size="small">{{ statusLabel(row.status) }}</el-tag>
+          <el-tag :type="statusTagType(asSystemUser(row).status)" size="small">{{ statusLabel(asSystemUser(row).status) }}</el-tag>
         </template>
       </el-table-column>
       <el-table-column label="页面权限" min-width="160">
         <template #default="{ row }">
           <el-tag
-            v-for="p in (row.permissions || []).slice(0, 3)"
+            v-for="p in userPermissions(row).slice(0, 3)"
             :key="p"
             size="small"
             type="info"
             class="perm-tag"
-          >{{ PERMISSION_LABELS[p] || p }}</el-tag>
-          <span v-if="(row.permissions || []).length > 3" class="perm-more">
-            +{{ (row.permissions || []).length - 3 }}
+          >{{ permissionLabel(p) }}</el-tag>
+          <span v-if="userPermissions(row).length > 3" class="perm-more">
+            +{{ userPermissions(row).length - 3 }}
           </span>
-          <span v-if="!(row.permissions || []).length" class="perm-more">无</span>
+          <span v-if="!userPermissions(row).length" class="perm-more">无</span>
         </template>
       </el-table-column>
       <el-table-column prop="create_time" label="创建时间" min-width="180" />
       <el-table-column label="操作" width="260" fixed="right">
         <template #default="{ row }">
           <el-switch
-            :model-value="row.status === 'active'"
-            :disabled="row.protected || isSelf(row)"
-            @change="(v: any) => toggleStatus(row, v)"
+            :model-value="asSystemUser(row).status === 'active'"
+            :disabled="asSystemUser(row).protected || isSelf(asSystemUser(row))"
+            @change="(v: any) => toggleStatus(asSystemUser(row), v)"
             inline-prompt
             active-text="启用"
             inactive-text="停用"
@@ -194,19 +206,19 @@ async function removeUser(row: SystemUser) {
           <el-button
             link
             type="primary"
-            :disabled="row.protected"
-            @click="openResetPwd(row)"
+            :disabled="asSystemUser(row).protected"
+            @click="openResetPwd(asSystemUser(row))"
           >重置密码</el-button>
           <el-button
             link
             type="primary"
-            @click="openPerms(row)"
+            @click="openPerms(asSystemUser(row))"
           >页面权限</el-button>
           <el-button
             link
             type="danger"
-            :disabled="row.protected || isSelf(row)"
-            @click="removeUser(row)"
+            :disabled="asSystemUser(row).protected || isSelf(asSystemUser(row))"
+            @click="removeUser(asSystemUser(row))"
           >删除</el-button>
         </template>
       </el-table-column>

frontend/src/views/tenants/TenantListView.vue

@@ -46,6 +46,15 @@ function openDetail(id: string) {
   router.push(`/tenants/${id}`)
 }
 
+function asTenant(row: unknown): Tenant {
+  return row as Tenant
+}
+
+function quotaText(row: unknown): string {
+  const quota = asTenant(row).quota || {}
+  return Object.keys(quota).length ? JSON.stringify(quota) : '—'
+}
+
 async function submitCreate() {
   if (!form.value.name) {
     ElMessage.warning('请填写租户名称')