improve gguf-function-calling parser

Author: okaris

llama_cpp/llama_chat_format.py

@@ -4295,15 +4295,17 @@ def gguf_function_calling(
     )
     initial_gbnf_tool_grammar = (
         (
-            'root ::= message_only | message_with_functions | functions_only\n'
-            'message_only ::= "message:" text\n'
-            'message_with_functions ::= "message:" text "<function_calls>\\n" functions\n'
-            'functions_only ::= "<function_calls>\\n" functions\n'
+            'root   ::= message_only | message_with_functions | functions_only\n'
+            'message_only ::= "message:" ws text\n'
+            'message_with_functions ::= "message:" ws text ws "<function_calls>" nl functions\n'
+            'functions_only ::= "<function_calls>" nl functions\n'
             'text ::= [^<]+\n'
+            'ws ::= [ \t\n]*\n'
+            'nl ::= "\\n"\n'
             f"functions ::= {function_names}\n"
         )
         if tool_choice == "auto"
-        else f'root ::= "<function_calls>" "\\n" functions\nfunctions ::= {function_names}\n'
+        else f'root ::= "<function_calls>" nl functions\nfunctions ::= {function_names}\n'
     )
     completion = cast(
         llama_types.CreateCompletionResponse,