update

backend/app/modules/fine_tune/service.py

@@ -3,13 +3,17 @@
 
 - preset 参数预设（quick / standard / high）
 - train_type → stage 映射（sft/dpo/cpt/cot）
-- 训练任务的启动 / 暂停 / 恢复 / 取消（委托 runner 真实执行）
+- 训练任务的启动 / 暂停 / 恢复 / 取消
+- compute_gateway 状态轮询线程（当任务派发到算力时自动同步状态）
 """
 from __future__ import annotations
 
+import asyncio
 import threading
+import time
 from typing import Any
 
+from app.core.config import get_settings
 from app.db.platform_store import get_platform_store
 from app.modules.fine_tune import runner
 
@@ -19,7 +23,46 @@ PRESETS: dict[str, dict[str, Any]] = {
     "high": {"learning_rate": "1e-5", "n_epochs": 5, "batch_size": 4, "lora_rank": 32},
 }
 
+# ── compute sync 轮询线程 ──────────────────────────────────────────────
+_sync_thread: threading.Thread | None = None
+_sync_thread_stop = threading.Event()
 
+
+def _compute_sync_loop() -> None:
+    """后台线程：周期性轮询算力节点，同步训练任务状态/日志/指标。"""
+    interval = get_settings().compute_poll_interval_seconds or 3
+    while not _sync_thread_stop.is_set():
+        try:
+            store = get_platform_store()
+            running = store.running_compute_tasks()
+            if running:
+                asyncio.run(_poll_once())
+        except Exception:  # noqa: BLE001 - keep polling loop alive
+            pass
+        time.sleep(interval)
+
+
+async def _poll_once() -> None:
+    from app.modules.compute_gateway.sync import poll_compute_jobs_once
+    await poll_compute_jobs_once()
+
+
+def start_compute_sync_worker() -> None:
+    """启动后台轮询线程（幂等，多次调用安全）。"""
+    global _sync_thread
+    if _sync_thread is not None and _sync_thread.is_alive():
+        return
+    _sync_thread_stop.clear()
+    _sync_thread = threading.Thread(target=_compute_sync_loop, daemon=True)
+    _sync_thread.start()
+
+
+def stop_compute_sync_worker() -> None:
+    """停止后台轮询线程。"""
+    _sync_thread_stop.set()
+
+
+# ── preset / config ────────────────────────────────────────────────────
 def apply_presets(payload: dict[str, Any]) -> dict[str, Any]:
     """根据 preset 字段补全缺失的超参；preset=custom 时不覆盖。"""
     payload = dict(payload)
@@ -63,7 +106,12 @@ def build_training_config(payload: dict[str, Any]) -> dict[str, Any]:
 
 
 def launch_training(task_id: str) -> None:
-    """在后台线程启动真实训练。"""
+    """在后台线程启动真实训练（本机 subprocess fallback，当算力节点不可用时使用）。
+
+    架构原则：GPU 计算应派发到算力服务进程执行。
+    当 platform_store.start_task 检测到在线算力节点时，会走 _dispatch_to_compute 派发路径；
+    仅当无可用算力节点且非 simulator 模式时，降级到本机 runner（违反 §1.1，待移除）。
+    """
     threading.Thread(target=runner.run_training, args=(task_id,), daemon=True).start()