improve gguf-function-calling parser

Author: okaris

llama_cpp/llama_chat_format.py

@@ -3975,9 +3975,23 @@ def chatml_function_calling(
         ),
     )
     text = completion["choices"][0]["text"]
-    tool_name = None if text.startswith("message") else text.split("\n")[-1][len("functions.") :]
+    # Extract message content and/or function call
+    tool_name = None
+    message_content = None
+    
+    if text.startswith("message:"):
+        # Handle message with or without function call
+        parts = text.split("<function_calls>", 1)
+        message_content = parts[0][len("message:"):].strip()
+        if len(parts) > 1:
+            # Has both message and function call
+            function_text = parts[1].strip()
+            tool_name = function_text.split("\n")[0][len("functions."):].rstrip(":")
+    else:
+        # Only function call
+        tool_name = text.split("\n")[0][len("<function_calls>\nfunctions."):].rstrip(":")
 
-    # Case 2 step 2A: Respond with a message
+    # Case 2 step 2A: Message only
     if tool_name is None:
         prompt = template_renderer.render(
             messages=messages, tools=[], tool_calls=None, add_generation_prompt=True
@@ -4067,7 +4081,7 @@ def chatml_function_calling(
                 ),
                 "message": {
                     "role": "assistant",
-                    "content": None,
+                    "content": message_content,  # Include message content when present
                     "tool_calls": [
                         {
                             "id": "call_" + f"_{i}_" + tool_name + "_" + completion["id"],