Fix the full cache issue

ax_spoken_communication_demo.py

@@ -196,6 +196,20 @@ class QwenTranslationAPI:
         self.api_url = api_url
         self.session_id = f"speech_translate_{int(time.time())}"
         
+    def reset_context(self):
+        """重置API上下文"""
+        try:
+            reset_url = f"{self.api_url}/api/reset"
+            response = requests.post(reset_url, json={}, timeout=5)
+            if response.status_code == 200:
+                print("[API] 上下文重置成功")
+                return True
+            else:
+                print(f"[API] 重置失败，状态码: {response.status_code}")
+        except Exception as e:
+            print(f"[API] 重置上下文失败: {e}")
+        return False
+        
     def translate(self, text_content, max_retries=3, timeout=120):
         """调用千问API进行处理"""
         if not text_content or text_content.strip() == "":
@@ -232,6 +246,7 @@ class QwenTranslationAPI:
                 start_time = time.time()
                 full_translation = ""
                 last_chunk = ""
+                error_detected = False
                 
                 while time.time() - start_time < timeout:
                     try:
@@ -239,7 +254,15 @@ class QwenTranslationAPI:
                         result_data = result_response.json()
                         
                         # 获取当前chunk
-                        current_chunk = result_data.get("response", "")#.strip()                 
+                        current_chunk = result_data.get("response", "")#.strip()
+                        
+                        # 检查是否有setkvcache failed错误
+                        if "error:" in current_chunk.lower() or "setkvcache failed" in current_chunk.lower():
+                            print(f"[API] 检测到错误: {current_chunk}")
+                            error_detected = True
+                            self.reset_context()
+                            break
+                        
                         full_translation += current_chunk
                         
                         # 检查是否完成
@@ -256,6 +279,12 @@ class QwenTranslationAPI:
                             break
                         continue
                 
+                # 如果检测到错误且还有重试次数，继续重试
+                if error_detected and attempt < max_retries - 1:
+                    print(f"[API] 等待1秒后重试...")
+                    time.sleep(1)
+                    continue
+                
                 print(f"[API] 轮询超时，尝试第 {attempt + 1} 次重试")
                 
             except requests.exceptions.RequestException as e: