ENH: support cline style messages for all backend engines (#2763)

xorbitsai · Jan 15, 2025 · 42a9561 · 42a9561
1 parent d44765e
commit 42a9561
Show file tree

Hide file tree

Showing 2 changed files with 4 additions and 1 deletion.
diff --git a/xinference/model/llm/utils.py b/xinference/model/llm/utils.py
@@ -104,6 +104,10 @@ def get_full_context(
         tokenize=False,
         **kwargs,
     ):
+        if "vision" not in self.model_family.model_ability:  # type: ignore
+            messages = self.convert_messages_with_content_list_to_str_conversion(
+                messages
+            )
         if tokenizer is not None:
             try:
                 full_context = tokenizer.apply_chat_template(

diff --git a/xinference/model/llm/vllm/core.py b/xinference/model/llm/vllm/core.py
@@ -804,7 +804,6 @@ async def async_chat(
         generate_config: Optional[Dict] = None,
         request_id: Optional[str] = None,
     ) -> Union[ChatCompletion, AsyncGenerator[ChatCompletionChunk, None]]:
-        messages = self.convert_messages_with_content_list_to_str_conversion(messages)
         tools = generate_config.pop("tools", []) if generate_config else None
         model_family = self.model_family.model_family or self.model_family.model_name
         full_context_kwargs = {}