.

app/main.py

@@ -171,6 +171,26 @@ tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
 app = FastAPI()
 logger.info("Using SAFE_MIN_FREE_MB = %d MB", SAFE_MIN_FREE_MB)
 
+@app.on_event("startup")
+async def warm_up_mp_pool():
+    """
+    启动即预热：
+      - GPU 模式：建立多 GPU 进程池（官方默认逻辑）
+      - CPU 模式：单进程跑一次，避免 fork
+    """
+    try:
+        if DEVICE.startswith("cuda"):
+            logger.info("Warm-up (GPU) → 预生成多进程池")
+            _ = model.encode(["warmup"], return_dense=True)          # 多 GPU 池
+        else:
+            logger.info("Warm-up (CPU) → 单进程初始化")
+            _ = model.encode(["warmup"], return_dense=True,
+                             num_processes=1)                       # 禁 fork
+    except Exception as e:
+        # 预热失败不会阻止服务启动，只给警告
+        logger.warning("Warm-up failed: %s — will fallback at first request", e)
+
+
 class EmbeddingRequest(BaseModel):
     input: Union[str, List[str]]
     model: str = "text-embedding-bge-m3"
@@ -191,8 +211,7 @@ def _encode(texts: List[str]):
             else:
                 # 已经 fallback 到 CPU：禁用进程池，禁止再向 GPU 拷权重
                 out = model.encode(t, return_dense=True,
-                                   num_processes=1,
-                                   device="cpu")
+                                   num_processes=1)
             q.put(("ok", out))
         except Exception as e:
             q.put(("err", str(e)))