refactor(chat_handler): extract MTMDChatHandler base class and Simplify the complexity of subsequent multimodal adaptation

- Extracted the core multimodal processing pipeline from `Llava15ChatHandler` into a generic `MTMDChatHandler` base class, separating pipeline logic from model-specific prompt formats.
- Updated all multimodal subclass handlers (e.g., Qwen2.5vl, Qwen3-vl, MiniCPM, GLM4, LFM2-VL) to inherit from the new `MTMDChatHandler`.
- Implemented strict `**kwargs` validation in the baseconstructor to gracefully intercept and report unsupported parameters, significantly improving Developer Experience (DX).
- Introduced dynamic `self.log_prefix` (`self.__class__.__name__`) for accurate and consistent logging across all subclasses.
- Cleaned up redundant state-clearing, image-count logic and hardcoded print statements across subclass `__call__` implementations.
- To avoid exceptions occurring when the close method is called due to initialization failure and the call to exit_stack.

Author: JamePeng

llama_cpp/_internals.py

@@ -85,7 +85,9 @@ def close(self):
             self.model = None
         self.vocab = None
 
-        self._exit_stack.close()
+        if getattr(self, "_exit_stack", None) is not None and hasattr(self._exit_stack, "close"):
+            self._exit_stack.close()
+            self._exit_stack = None
 
     def __del__(self):
         self.close()
@@ -386,8 +388,11 @@ def close(self):
             except Exception:
                 pass
             self.ctx = None
+        self.params = None
 
-        self._exit_stack.close()
+        if getattr(self, "_exit_stack", None) is not None and hasattr(self._exit_stack, "close"):
+            self._exit_stack.close()
+            self._exit_stack = None
 
     def __del__(self):
         self.close()
@@ -662,7 +667,9 @@ def close(self):
                 pass
             self.batch = None
 
-        self._exit_stack.close()
+        if getattr(self, "_exit_stack", None) is not None and hasattr(self._exit_stack, "close"):
+            self._exit_stack.close()
+            self._exit_stack = None
 
     def __del__(self):
         self.close()

llama_cpp/llama.py

@@ -682,7 +682,9 @@ def close(self) -> None:
         self._c_tensor_split = None
         self._kv_overrides_array = None
 
-        self._stack.close()
+        if getattr(self, "_stack", None) is not None and hasattr(self._stack, "close"):
+            self._stack.close()
+            self._stack = None
 
     def __del__(self) -> None:
         self.close()