feat: 模型推理端到端闭环 — 真实流式推理 + 释放/删除 + GPU 状态同步

后端 (platform.py + platform_store.py):
- 新增 _build_messages_payload() 转换前端格式为 OpenAI messages
- 新增 _stream_chat_proxy() SSE 流式代理到算力节点
- 新增 _unload_from_compute_node() 真正释放算力节点 GPU 显存
- 重写 model_compare_load: 从假 PID/端口改为真正调用算力节点加载模型
- 修复 model_compare_unload: 调用 _unload_from_compute_node 释放 GPU
- 修复 model_compare_delete: 先释放 GPU 再删除记录
- 修复 model_compare_stream_chat: 从 mock 改为 StreamingResponse 代理
- 修复 model_chat_local/stream: 消息格式转换 + 路径修正
- PlatformStore 新增 _inference_nodes 追踪，gpus() 同步推理占用状态
- preload/unload 端点标记/清除推理节点占用

算力节点 (compute):
- inference.py: 适配新版 LLaMA-Factory API (get_infer_args 4 返回值、ChatModel args dict、stream_chat 新签名)
- inference.py: unload() 增加 gc.collect + torch.cuda.empty_cache + synchronize 彻底释放显存
- main.py: inference/load 移除 HTTPException(500)，错误以 200 正常返回

前端:
- InferenceChatView: 真实模式下走 SSE 流式推理，mock 模式保留兼容
- InferenceCreateView: 调用 preloadLocalModel + createCompare 真实创建推理任务，失败回退 mock
- InferenceListView: 「停止」改为「释放」，删除前先释放算力节点，改进错误提示
- compare.ts: 新增 streamChatReal() fetch SSE，preload 超时提升至 5 分钟
- useStreamChat.ts: send() 支持 useMock 参数，真实模式调用 streamChatReal
- GPU 选择过滤: 仅显示在线算力节点上的空闲 GPU

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

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

@@ -31,6 +31,62 @@ 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})
 
@@ -1022,8 +1078,27 @@ 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})
 
@@ -1053,26 +1128,56 @@ 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 = [
+        # 选取在线算力节点
-            {
+        node = _select_first_online_node(store)
-                "model_id": item.get("model_id"),
+        if not node:
-                "model_name": item.get("model_name"),
+            return ok({"status": "failed", "error": "no online compute node"})
-                "status": "ready",
+        client = ComputeNodeClient(node["api_base_url"])
-                "pid": 45000 + index,
+        loaded_models = []
-                "port": item.get("port") or 18000 + index,
+        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 item.get("adapter_path"):
-            if isinstance(item, dict)
+                load_payload["adapter_name_or_path"] = item["adapter_path"]
-        ]
+            try:
-        return ok(get_platform_store().update_compare_task(task_id, {"status": "loaded", "load_status": {"loaded_models": loaded_models}}))
+                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")
 
@@ -1080,7 +1185,11 @@ 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")
 
@@ -1092,18 +1201,23 @@ 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 = ""
+    """Proxy non-streaming chat to the compute node running the inference model."""
-    for message in payload.get("messages") or []:
+    store = get_platform_store()
-        if message.get("role") == "user":
+    node = _select_first_online_node(store)
-            question = str(message.get("content") or "")
+    if not node:
-    content = f"当前后端已收到推理请求：{question[:120]}"
+        return ok({"response": "no online compute node available for inference", "request": payload})
-    return ok({"response": content, "content": content})
+    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]:
+async def model_compare_stream_chat(payload: dict[str, Any] = Body(...)) -> StreamingResponse:
-    question = payload.get("user_question") or payload.get("question") or ""
+    """Stream chat from the compute node (SSE proxy)."""
-    return ok({"response": f"当前后端已收到流式推理请求：{str(question)[:120]}"})
+    return await _stream_chat_proxy(payload)
 
 
 @router.post("/model-chat/batch")
@@ -1120,7 +1234,7 @@ 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})
@@ -1129,28 +1243,15 @@ 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()
+    return await _stream_chat_proxy(payload)
-    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")
 
 
 @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:
@@ -1158,6 +1259,8 @@ 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)})
@@ -1167,6 +1270,9 @@ 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"})
@@ -1196,6 +1302,9 @@ 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:
@@ -1203,6 +1312,8 @@ 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)})

backend/app/db/platform_store.py

@@ -293,6 +293,19 @@ class PlatformStore:
         self.database_url = _psycopg_url(database_url or settings.database_url)
         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"]:
@@ -2698,6 +2711,11 @@ 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)

compute/api/main.py

@@ -688,8 +688,6 @@ 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")

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)
+                infer_result = get_infer_args(args)
-                self._model = ChatModel(model_args)
+                # ChatModel internally re-parses the args dict via get_infer_args,
-                self._tokenizer = self._model.tokenizer
+                # so pass the original args (not the parsed dataclass objects).
-                self._generating_args = generating_args
+                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}
+                generate_kwargs = {**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)
-                for new_text in self._model.stream_chat(formatted, generate_kwargs):
+                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'

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 */
+/** 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)
+export const preloadLocalModel = (data: any) => post('/model-chat/local/preload', data, { timeout: 300000 })
 
-/** 预加载已训练模型 */
+/** 预加载已训练模型（超时 5 分钟） */
-export const preloadTrainedModel = (data: any) => post('/model-chat/trained/preload', data)
+export const preloadTrainedModel = (data: any) => post('/model-chat/trained/preload', data, { timeout: 300000 })

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

@@ -361,10 +361,10 @@ export interface GpuInfo {
   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/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） */
+/** 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'
+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
+  // 真实模式：通过后端 SSE 流式代理到算力节点进行推理
-  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。
   activeAssistant = assistantMsg
 
-  await send({
+  await send(
-    port: target.port,
+    {
-    model_name: target.model_name,
+      model_path: route.query.model_path as string || '',
-    model_path: '',
       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，
+      // Step 1: 将模型加载到算力节点
-      // 用假数据直通进入对话界面（模型名通过 query 传递）。
+      const preloadPayload = {
-      // 接入真实后端后，可在此恢复 createCompare / startModelsInBackground / monitorStartup 流程。
+        model_name_or_path: m.model_path,
-      await new Promise((resolve) => setTimeout(resolve, 1200))
+        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
+    computeNodes.value = nodes || []
-    if (gpus.value.length > 0) form.gpu_id = 0
+    // 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"
+            v-for="g in idleGpus"
-            :key="idx"
+            :key="g.id ?? 0"
-            :label="`${g.name} (GPU${idx})`"
+            :label="`${g.name} (GPU${g.id ?? 0}) [空闲]`"
-            :value="idx"
+            :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
+  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' })
-  const task = await getCompare(row.id).catch(() => null)
+  // 先释放算力节点上的模型
-  if (task?.load_status) {
+  await unloadCompare(row.id).catch(() => {})
-    const models = parseLoadedModels(task as CompareTask)
-    for (const m of models) {
-      if (m.pid) {
-        await stopModelByPid(m.pid).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>