improve gguf-function-calling parser

Author: okaris

llama_cpp/llama_chat_format.py

@@ -4292,14 +4292,13 @@ def gguf_function_calling(
     prompt = template_renderer.render(
         messages=messages, tools=tools, tool_calls=True, add_generation_prompt=True
     )
-    initial_gbnf_tool_grammar = (
-        (
-            'root ::= "<function_calls>" "\\n" functions | "message:"\n'
-            f"functions ::= {function_names}\n"
-        )
-        if tool_choice == "auto"
-        else f'root ::= "<function_calls>" "\\n" functions\nfunctions ::= {function_names}\n'
-    )
+        # Decide initial grammar: message only, functions only, or message+functions
+    initial_gbnf_tool_grammar = """
+        root ::= message | function_calls | message_then_functions
+        message ::= "message:"
+        function_calls ::= "<function_calls>\n"
+        message_then_functions ::= "message:" "<function_calls>\n"
+        """
     completion = cast(
         llama_types.CreateCompletionResponse,
         llama.create_completion(