Update app.py

app.py

@@ -308,6 +308,8 @@ def check_tokens():
 
     return jsonify(results)
 
+import json
+
 @app.route('/handsome/v1/chat/completions', methods=['POST'])
 def handsome_chat_completions():
     if not check_authorization(request):
@@ -357,18 +359,31 @@ def handsome_chat_completions():
                 first_token_time = first_chunk_time - start_time if first_chunk_time else 0
                 total_time = end_time - start_time
 
-                # 从完整响应中提取信息
-                try:
-                    response_json = json.loads(full_response_content.split("data: ")[-1].strip())
-                    prompt_tokens = response_json["usage"]["prompt_tokens"]
-                    completion_tokens = response_json["usage"]["completion_tokens"]
-                    response_content = response_json["choices"][0]["message"]["content"]
-                except (KeyError, ValueError, IndexError) as e:
-                    logging.error(f"解析流式响应 JSON 失败: {e}, 完整内容: {full_response_content}")
-                    prompt_tokens = 0
-                    completion_tokens = 0
-                    response_content = ""
-
+                # 处理流式响应，逐行解析 JSON
+                prompt_tokens = 0
+                completion_tokens = 0
+                response_content = ""
+                for line in full_response_content.splitlines():
+                    if line.startswith("data:"):
+                        line = line[5:].strip()
+                        if line == "[DONE]":
+                            continue
+                        try:
+                            response_json = json.loads(line)
+                            
+                            # 提取信息，这里只累加 completion_tokens 和 content
+                            if "usage" in response_json and "completion_tokens" in response_json["usage"]:
+                                completion_tokens = response_json["usage"]["completion_tokens"]
+                            
+                            if "choices" in response_json and len(response_json["choices"]) > 0 and "delta" in response_json["choices"][0] and "content" in response_json["choices"][0]["delta"]:
+                                response_content += response_json["choices"][0]["delta"]["content"]
+                                
+                            if "usage" in response_json and "prompt_tokens" in response_json["usage"]:
+                                prompt_tokens = response_json["usage"]["prompt_tokens"]
+
+                        except (KeyError, ValueError, IndexError) as e:
+                            logging.error(f"解析流式响应单行 JSON 失败: {e}, 行内容: {line}")
+                            
                 # 提取用户输入的内容
                 user_content = ""
                 messages = data.get("messages", [])
@@ -383,6 +398,7 @@ def handsome_chat_completions():
                 )
             return Response(stream_with_context(generate()), content_type=response.headers['Content-Type'])
         else:
+            # 非流式响应处理... (保持原样)
             response.raise_for_status()
             end_time = time.time()
             response_json = response.json()
@@ -417,6 +433,7 @@ def handsome_chat_completions():
         return jsonify({"error": str(e)}), 500
 
 
+
 @app.route('/handsome/v1/models', methods=['GET'])
 def list_models():
     if not check_authorization(request):