Merge pull request #3125 from xinnan-tech/py-fix-llm-chunk

fix: 修复频繁打断下生成器不会清理

main/xiaozhi-server/core/providers/llm/homeassistant/homeassistant.py

@@ -39,8 +39,7 @@ class LLMProvider(LLMProviderBase):
         }
 
         # 发起 POST 请求
-        response = requests.post(self.api_url, json=payload, headers=headers)
+        with requests.post(self.api_url, json=payload, headers=headers) as response:
-
             # 检查请求是否成功
             response.raise_for_status()
 

main/xiaozhi-server/core/providers/llm/ollama/ollama.py

@@ -50,6 +50,7 @@ class LLMProvider(LLMProviderBase):
         # 用于处理跨chunk的标签
         buffer = ""
 
+        try:
             for chunk in responses:
                 try:
                     delta = (
@@ -87,6 +88,8 @@ class LLMProvider(LLMProviderBase):
 
                 except Exception as e:
                     logger.bind(tag=TAG).error(f"Error processing chunk: {e}")
+        finally:
+            responses.close()
 
     def response_with_functions(self, session_id, dialogue, functions=None):
         # 如果是qwen3模型，在用户最后一条消息中添加/no_think指令
@@ -117,6 +120,7 @@ class LLMProvider(LLMProviderBase):
         is_active = True
         buffer = ""
 
+        try:
             for chunk in stream:
                 try:
                     delta = (
@@ -163,3 +167,5 @@ class LLMProvider(LLMProviderBase):
                 except Exception as e:
                     logger.bind(tag=TAG).error(f"Error processing function chunk: {e}")
                     continue
+        finally:
+            stream.close()

main/xiaozhi-server/core/providers/llm/openai/openai.py

@@ -115,6 +115,7 @@ class LLMProvider(LLMProviderBase):
         responses = self.client.chat.completions.create(**request_params)
 
         is_active = True
+        try:            
             for chunk in responses:
                 try:
                     delta = chunk.choices[0].delta if getattr(chunk, "choices", None) else None
@@ -130,6 +131,8 @@ class LLMProvider(LLMProviderBase):
                         content = content.split("</think>")[-1]
                     if is_active:
                         yield content
+        finally:
+            responses.close()
 
     def response_with_functions(self, session_id, dialogue, functions=None, **kwargs):
         dialogue = self.normalize_dialogue(dialogue)
@@ -157,6 +160,7 @@ class LLMProvider(LLMProviderBase):
 
         stream = self.client.chat.completions.create(**request_params)
 
+        try:
             for chunk in stream:
                 if getattr(chunk, "choices", None):
                     delta = chunk.choices[0].delta
@@ -170,3 +174,5 @@ class LLMProvider(LLMProviderBase):
                         f"输出 {getattr(usage_info, 'completion_tokens', '未知')}，"
                         f"共计 {getattr(usage_info, 'total_tokens', '未知')}"
                     )
+        finally:
+            stream.close()

main/xiaozhi-server/core/providers/llm/xinference/xinference.py

@@ -74,6 +74,7 @@ class LLMProvider(LLMProviderBase):
             tools=functions,
         )
 
+        try:
             for chunk in stream:
                 delta = chunk.choices[0].delta
                 content = delta.content
@@ -83,3 +84,5 @@ class LLMProvider(LLMProviderBase):
                     yield content, tool_calls
                 elif tool_calls:
                     yield None, tool_calls
+        finally:
+            stream.close()