Update rerank_server_demo.py

1. Added CUDA debugging support
2. Improved memory management during model inference.
3. Fixed logits output

Test Device: Windows WSL Ubuntu 20.04 (python3.10 CUDA 5090 32G)

rerank_server_demo.py

@@ -1,3 +1,4 @@
+import gc
 import torch
 import uvicorn
 from fastapi import FastAPI
@@ -5,6 +6,10 @@ from pydantic import BaseModel, Field
 from transformers import AutoModelForSequenceClassification, AutoTokenizer
 from typing import List
 
+# 使能 CUDA 调试
+# import os
+# os.environ['CUDA_LAUNCH_BLOCKING']='1'
+
 # --- 1. 定义API的请求和响应数据结构 ---
 
 # 请求体结构保持不变
@@ -53,7 +58,7 @@ except Exception as e:
 app = FastAPI(
     title="Reranker API (Test Version)",
     description="一个返回 'score' 字段以测试Go客户端兼容性的API服务",
-    version="1.0.1"
+    version="1.0.2"
 )
 
 # --- 4. 定义API端点 ---
@@ -65,8 +70,23 @@ def rerank_endpoint(request: RerankRequest):
     pairs = [[request.query, doc] for doc in request.documents]
 
     with torch.no_grad():
-        inputs = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=1024).to(device)
-        scores = model(**inputs, return_dict=True).logits.view(-1, ).float()
+        inputs = outputs = logits = None
+
+        try:
+            inputs = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=1024).to(device)
+            outputs = model(**inputs, return_dict=True)
+            logits = outputs.logits.view(-1, ).float()
+            scores = torch.sigmoid(logits)
+        finally:
+            # 释放 GPU 资源占用
+            del inputs, outputs, logits
+            gc.collect()
+
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+            elif hasattr(torch, "mps") and torch.mps.is_available():
+                torch.mps.empty_cache()
+
 
     # --- 修改开始：按照测试用的结构来构建结果 ---
     results = []
@@ -99,4 +119,3 @@ def read_root():
 # --- 5. 启动服务 ---
 if __name__ == "__main__":
     uvicorn.run(app, host="0.0.0.0", port=8000)
-