improve gguf-function-calling parser

Author: okaris

llama_cpp/llama_chat_format.py

@@ -4293,12 +4293,15 @@ def gguf_function_calling(
         messages=messages, tools=tools, tool_calls=True, add_generation_prompt=True
     )
         # 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>"
-    message_then_functions ::= "message:" "<function_calls>"
-    """
+    initial_gbnf_tool_grammar = (
+        """root ::= message | function_calls | message_then_functions
+        message ::= "message:"
+        function_calls ::= "<function_calls>"
+        message_then_functions ::= "message:" "<function_calls>"
+        """
+        if tool_choice == "auto"
+        else f'root ::= "<function_calls>" functions\nfunctions ::= {function_names}\n'
+    )
     completion = cast(
         llama_types.CreateCompletionResponse,
         llama.create_completion(