improve gguf-function-calling parser

Author: okaris

llama_cpp/llama_chat_format.py

@@ -4174,6 +4174,7 @@ def gguf_function_calling(
         "\nfunctions.<function_name>:"
         '\n{ "arg1": "value1", "arg2": "value2" }'
         "\n</function_calls>"
+        "\n\nYou can also combine both formats to provide explanatory text with function calls."
         "{% endif %}"
         "<|im_end|>\n"
         "{% endif %}"
@@ -4292,15 +4293,13 @@ def gguf_function_calling(
     prompt = template_renderer.render(
         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>"
-        """
+        (
+            'root ::= "<function_calls>" "\\n" functions | "message:"\n'
+            f"functions ::= {function_names}\n"
+        )
         if tool_choice == "auto"
-        else f'root ::= "<function_calls>" functions\nfunctions ::= {function_names}\n'
+        else f'root ::= "<function_calls>" "\\n" functions\nfunctions ::= {function_names}\n'
     )
     completion = cast(
         llama_types.CreateCompletionResponse,
@@ -4319,21 +4318,43 @@ def gguf_function_calling(
         ),
     )
     text = completion["choices"][0]["text"]
-    tool_name = None if text.startswith("message") else text.split("\n")[-1][len("functions.") :]
-
-    # Case 2 step 2A: Respond with a message
-    if tool_name is None:
+    
+    # Parse the response to extract message and/or function calls
+    message_content = None
+    tool_name = None
+
+    if text.startswith("message:"):
+        # Always grab message
+        if "<function_calls>" in text:
+            parts = text.split("<function_calls>", 1)
+            message_content = parts[0][len("message:"):].strip()
+            if len(parts) > 1 and "functions." in parts[1]:
+                tool_name = parts[1].split("functions.", 1)[1].split(":", 1)[0].strip()
+        else:
+            message_content = text[len("message:"):].strip()
+
+    elif text.startswith("<function_calls>"):
+        # Function calls block, but allow stray message content too
+        parts = text.split("functions.", 1)
+        if len(parts) > 1:
+            tool_name = parts[1].split(":", 1)[0].strip()
+        # Optionally capture anything before the function calls as message text
+        before = text[len("<function_calls>"):].strip()
+        if before and not before.startswith("functions."):
+            message_content = before    
+
+    # Case 2 step 2A: Respond with message only
+    if tool_name is None and message_content is not None:
         prompt = template_renderer.render(
             messages=messages, tools=[], tool_calls=None, add_generation_prompt=True
         )
-        return _convert_completion_to_chat(
-            llama.create_completion(
-                prompt=prompt,
-                **completion_kwargs,  # type: ignore[arg-type]
-                logprobs=top_logprobs if logprobs else None,
-            ),
-            stream=stream,
+        completion_response = llama.create_completion(
+            prompt=prompt,
+            **completion_kwargs,  # type: ignore[arg-type]
+            logprobs=top_logprobs if logprobs else None,
         )
+        completion_response["choices"][0]["text"] = message_content
+        return _convert_completion_to_chat(completion_response, stream=stream)
 
     # Case 2 step 2B: One or more function calls
     follow_up_gbnf_tool_grammar = (
@@ -4343,7 +4364,7 @@ def gguf_function_calling(
     prompt += "<function_calls>\n"
     if stream:
         return _stream_tool_calls(
-            llama, prompt, tools, tool_name, completion_kwargs, follow_up_gbnf_tool_grammar
+            llama, prompt, tools, tool_name, completion_kwargs, follow_up_gbnf_tool_grammar, message_content
         )
     tool = next((tool for tool in tools if tool["function"]["name"] == tool_name), None)
     completions: List[llama_types.CreateCompletionResponse] = []
@@ -4411,7 +4432,7 @@ def gguf_function_calling(
                 ),
                 "message": {
                     "role": "assistant",
-                    "content": None,
+                    "content": message_content,  # Include message content if present
                     "tool_calls": [
                         {
                             "id": "call_" + f"_{i}_" + tool_name + "_" + completion["id"],