.

app/main.py

@@ -31,6 +31,9 @@ SAFE_MIN_FREE_MB = int(
     os.getenv("SAFE_MIN_FREE_MB", "16384")
 )  # 启动时要求的最小空闲显存，MB
 
+MODEL_VRAM_MB = int(os.getenv("MODEL_VRAM_MB", "16384"))   # bge-m3-large fp16 ≈ 16 GiB
+POST_LOAD_GAP_MB = 512     
+
 # -----------------------------------------------------------------------------#
 # Logging
 # -----------------------------------------------------------------------------#
@@ -109,42 +112,53 @@ def load_model(device: str):
 # -----------------------------------------------------------------------------#
 def auto_select_and_load(min_free_mb: int = 4096):
     """
-    1. Gather GPUs with free_mem ≥ min_free_mb
+    1. 过滤掉空闲显存 < MODEL_VRAM_MB 的 GPU
-    2. Sort by free_mem desc, attempt to load
+    2. 按空闲显存降序依次尝试加载
-    3. All fail → CPU
+    3. 载入后再次检查：若剩余 < POST_LOAD_GAP_MB → 视为失败
-    Return (model, device, precision)
+    4. 若全部 GPU 不满足 → CPU
     """
     if not torch.cuda.is_available():
         logger.info("No GPU detected → CPU")
         return (*load_model("cpu"), "cpu")
 
-    # Build candidate list
+    # 收集每张卡的空闲显存
-    candidates: list[tuple[int, int]] = []  # (free_MB, idx)
+    candidates: list[tuple[int, int]] = []          # (free_MB, idx)
     for idx in range(torch.cuda.device_count()):
         free, _ = _gpu_mem_info(idx)
         free_mb = free // 2**20
         candidates.append((free_mb, idx))
 
-    candidates = [c for c in candidates if c[0] >= min_free_mb]
+    # 先按 min_free_mb 做一次粗过滤（通常 16384 MB）
+    candidates = [c for c in candidates if c[0] >= MODEL_VRAM_MB]
     if not candidates:
-        logger.warning("All GPUs free_mem < %d MB → CPU", min_free_mb)
+        logger.warning("All GPUs free_mem < %d MB → CPU", MODEL_VRAM_MB)
         return (*load_model("cpu"), "cpu")
 
-    candidates.sort(reverse=True)  # high free_mem first
+    # 按空闲显存从高到低尝试
+    candidates.sort(reverse=True)
     for free_mb, idx in candidates:
         dev = f"cuda:{idx}"
         try:
             logger.info("Trying %s (free=%d MB)", dev, free_mb)
             mdl, prec = load_model(dev)
-            return mdl, prec, dev
-        except RuntimeError as e:
-            if "out of memory" in str(e).lower():
-                logger.warning("%s OOM → next GPU", dev)
-                torch.cuda.empty_cache()
-                continue
-            raise  # non-OOM error
 
-    logger.warning("All GPUs failed → CPU fallback")
+            # 载入后再测一次，确保还剩余 ≥ POST_LOAD_GAP_MB
+            new_free, _ = _gpu_mem_info(idx)
+            new_free_mb = new_free // 2**20
+            if new_free_mb < POST_LOAD_GAP_MB:
+                raise RuntimeError(
+                    f"post-load free {new_free_mb} MB < {POST_LOAD_GAP_MB} MB"
+                )
+
+            return mdl, prec, dev                # GPU 占用成功
+
+        except RuntimeError as e:
+            logger.warning("%s unusable (%s) → next", dev, e)
+            torch.cuda.empty_cache()
+            continue
+
+    # 全部 GPU 都不满足
+    logger.warning("No suitable GPU left → CPU fallback")
     return (*load_model("cpu"), "cpu")