.

app/main.py

@@ -89,6 +89,10 @@ def load_model(device: str):
 
     mdl = BGEM3FlagModel(MODEL_PATH, use_fp16=use_fp16, device=device)
 
+    if device == "cpu":
+        # 屏蔽 GPU，让后续 torch / BGEM3 都认不出 CUDA
+        os.environ["CUDA_VISIBLE_DEVICES"] = ""
+        
     # Simple DataParallel for multi-GPU inference
     if device.startswith("cuda") and torch.cuda.device_count() > 1:
         logger.info(
@@ -173,25 +177,21 @@ logger.info("Using SAFE_MIN_FREE_MB = %d MB", SAFE_MIN_FREE_MB)
 
 @app.on_event("startup")
 async def warm_up_mp_pool():
-    """
-    GPU: 建立多进程池预热
-    CPU: 单进程预热，但先临时把 torch 看到的 GPU 设为 0 张
-    """
     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
+            logger.info("Warm-up (GPU) → 建多进程池")
+            _ = model.encode(["warmup"], return_dense=True)          # GPU
         else:
             logger.info("Warm-up (CPU) → 单进程初始化")
-            # --- 关键 3 行 ---------------------------------------------------
-            orig_cnt = torch.cuda.device_count                      # 保存原函数
-            torch.cuda.device_count = lambda: 0                     # 伪装无 GPU
+            # 双保险：彻底把 BGEM3 的设备表改成只含 CPU
+            if hasattr(model, "devices"):
+                model.devices = ["cpu"]
+            model.device = "cpu"
             _ = model.encode(["warmup"], return_dense=True,
-                             num_processes=1)                       # 单进程
-            torch.cuda.device_count = orig_cnt                      # 恢复
-            # ----------------------------------------------------------------
+                             num_processes=1)                       # 不再 fork
     except Exception as e:
-        logger.warning("Warm-up failed: %s — 将在首条请求时再退避", e)
+        logger.warning("Warm-up failed: %s —— 首条请求时再退避", e)
+
 
 
 class EmbeddingRequest(BaseModel):