.

Dockerfile

@@ -21,6 +21,14 @@ COPY model/bge-m3 /app/model/bge-m3
 # 暴露端口
 EXPOSE 8001
 
-# 启动 FastAPI 服务
-CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8001"]
+# # 启动 FastAPI 服务
+# CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8001"]
 
+# 新增：给 PT 显存分段配置，减少碎片 (可选但推荐)
+ENV PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:32
+
+# 启动：Gunicorn + 4 worker，每个 worker 一个独立进程
+CMD ["gunicorn", "app.main:app", \
+     "-k", "uvicorn.workers.UvicornWorker", \
+     "-w", "4", \
+     "-b", "0.0.0.0:8001"]

app/main.py

@@ -15,6 +15,7 @@ import os
 import sys
 import time
 from typing import List, Union
+import multiprocessing as mp
 
 import torch
 from fastapi import FastAPI, HTTPException
@@ -175,31 +176,61 @@ class EmbeddingRequest(BaseModel):
     model: str = "text-embedding-bge-m3"
 
 
-fallback_done = False  # prevent endless downgrade loop
-
-
 def _encode(texts: List[str]):
-    """Encode with single downgrade to CPU on OOM / CUDA failure."""
-    global model, DEVICE, PRECISION, fallback_done
-
+    """
+    单次请求：
+      1. 子进程跑 GPU 推理；成功→返回
+      2. 若子进程 OOM / CUDA Error → 同一次请求 fallback 到 CPU
+    绝不改全局状态，其他并发请求不受影响
+    """
+    def _worker(t, q):
         try:
-        return model.encode(texts, return_dense=True)
+            q.put(("ok", model.encode(t, return_dense=True)))
+        except Exception as e:
+            q.put(("err", str(e)))
 
-    except RuntimeError as err:
-        is_oom = "out of memory" in str(err).lower()
-        is_cuda_fail = "cuda error" in str(err).lower() or "device-side assert" in str(
-            err
-        ).lower()
+    q = mp.Queue()
+    p = mp.Process(target=_worker, args=(texts, q))
+    p.start()
+    p.join(timeout=60)
 
-        if (is_oom or is_cuda_fail) and not fallback_done:
-            logger.error("GPU failure (%s). Falling back to CPU…", err)
-            fallback_done = True
+    if not q.empty():
+        status, payload = q.get()
+        if status == "ok":
+            return payload
+        if "out of memory" in payload.lower() or "cuda error" in payload.lower():
+            logger.warning("GPU OOM → 本次请求改走 CPU：%s", payload)
             torch.cuda.empty_cache()
-            DEVICE = "cpu"
-            model, PRECISION = load_model(DEVICE)
-            return model.encode(texts, return_dense=True)
+            cpu_model, _ = load_model("cpu")
+            return cpu_model.encode(texts, return_dense=True)
+        raise RuntimeError(payload)
 
-        raise  # second failure → propagate
+    raise RuntimeError("子进程异常退出，无返回")
+
+# fallback_done = False  # prevent endless downgrade loop
+
+# def _encode(texts: List[str]):
+#     """Encode with single downgrade to CPU on OOM / CUDA failure."""
+#     global model, DEVICE, PRECISION, fallback_done
+
+#     try:
+#         return model.encode(texts, return_dense=True)
+
+#     except RuntimeError as err:
+#         is_oom = "out of memory" in str(err).lower()
+#         is_cuda_fail = "cuda error" in str(err).lower() or "device-side assert" in str(
+#             err
+#         ).lower()
+
+#         if (is_oom or is_cuda_fail) and not fallback_done:
+#             logger.error("GPU failure (%s). Falling back to CPU…", err)
+#             fallback_done = True
+#             torch.cuda.empty_cache()
+#             DEVICE = "cpu"
+#             model, PRECISION = load_model(DEVICE)
+#             return model.encode(texts, return_dense=True)
+
+#         raise  # second failure → propagate
 
 
 @app.post("/v1/embeddings")