.

app/main.py

@@ -82,26 +82,24 @@ def _choose_precision(device: str) -> str:
 
 
 def load_model(device: str):
-    """Instantiate model on `device`; return (model, precision)."""
     precision = _choose_precision(device)
     use_fp16 = precision == "fp16"
     logger.info("Loading BGEM3 on %s (%s)", device, precision)
 
+    # ---------- 关键：若走 CPU，先把整个 CUDA 伪装掉 ----------
+    if device == "cpu":
+        os.environ["CUDA_VISIBLE_DEVICES"] = ""      # 后续所有 fork 也看不到 GPU
+        torch.cuda.is_available = lambda: False
+        torch.cuda.device_count = lambda: 0
+    # ----------------------------------------------------------
+
     mdl = BGEM3FlagModel(MODEL_PATH, use_fp16=use_fp16, device=device)
 
-    if device == "cpu":
+    # 多 GPU 时才包 DataParallel
-        # 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(
-            "Wrapping model with torch.nn.DataParallel (%d GPUs)",
-            torch.cuda.device_count(),
-        )
         mdl = torch.nn.DataParallel(mdl)
-    return mdl, precision
 
+    return mdl, precision
 
 # -----------------------------------------------------------------------------#
 # Auto-select device (startup)
@@ -177,20 +175,11 @@ logger.info("Using SAFE_MIN_FREE_MB = %d MB", SAFE_MIN_FREE_MB)
 
 @app.on_event("startup")
 async def warm_up_mp_pool():
+    logger.info("Warm-up on %s", DEVICE)
     try:
-        if DEVICE.startswith("cuda"):
-            logger.info("Warm-up (GPU) → 建多进程池")
-            _ = model.encode(["warmup"], return_dense=True)
-        else:
-            logger.info("Warm-up (CPU) → 单进程初始化")
-            # ── 临时让库“以为”没有 GPU ────────────────────────────
-            orig_cnt = torch.cuda.device_count
-            torch.cuda.device_count = lambda: 0
         _ = model.encode(["warmup"], return_dense=True)   # 不传 num_processes
-            torch.cuda.device_count = orig_cnt
-            # ──────────────────────────────────────────────────────
     except Exception as e:
-        logger.warning("Warm-up failed: %s —— 首条请求时再退避", e)
+        logger.warning("Warm-up failed: %s — 首条请求时再退避", e)
 
 
 
@@ -208,14 +197,7 @@ def _encode(texts: List[str]):
     """
     def _worker(t, q):
         try:
-            if DEVICE.startswith("cuda"):
+            out = model.encode(t, return_dense=True)   # GPU / CPU 路径都安全
-                out = model.encode(t, return_dense=True)          # GPU 正常跑
-            else:
-                # 临时屏蔽 GPU，单进程 CPU 推理
-                orig_cnt = torch.cuda.device_count
-                torch.cuda.device_count = lambda: 0
-                out = model.encode(t, return_dense=True)          # 不传 num_processes
-                torch.cuda.device_count = orig_cnt
             q.put(("ok", out))
         except Exception as e:
             q.put(("err", str(e)))