gpt-oss-20b - tool calling issues?

[DigitalRudeness]

Hi @ALL,

I'm currently struggling on executing tool calls with the new gpt-oss-20b model.

I'm using a self programmed python solution for creating chat requests and tool calls against llama-server.

The generated tool calls look like the example below (standard, nothing special):

[{'function': {'description': 'gibt das aktuelle datum aus',
               'name': 'get_date',
               'parameters': {}},
  'type': 'function'},
 {'function': {'description': 'gibt die aktuelle uhrzeit aus',
               'name': 'get_time',
               'parameters': {}},
  'type': 'function'},
 {'function': {'description': 'gibt das aktuelle wetter  für eine stadt aus',
               'name': 'get_weather',
               'parameters': {'properties': {'city': {'description': 'die '
                                                                     'stadt '
                                                                     'für die '
                                                                     'die '
                                                                     'temperatur '
                                                                     'wiedergegeben '
                                                                     'werden '
                                                                     'soll',
                                                      'type': 'string'},
                                             'region': {'description': 'die '
                                                                       'region '
                                                                       'für '
                                                                       'die '
                                                                       'die '
                                                                       'temperatur '
                                                                       'wiedergegeben '
                                                                       'werden '
                                                                       'soll',
                                                        'type': 'string'}},
                              'required': ['city', 'region'],
                              'type': 'object'}},
  'type': 'function'}]

As you can see I'm trying to let the model choose from different functions either with params or without depending on the request.

When I'm e.g. requesting the weather for a city in a region (ignore the system prompt, it will be outsourced from regular requests)....

[{'content': 'Du antwortest nur auf Deutsch! Du antwoertest in ganzen Sätzen, '
             'aber so kurz wie möglich und gibst nur die reine Antwort auf den '
             'gestellten Request wieder! Sollte die Antwort eine Einheit sein '
             '(z.B. Uhrzeit, Temperatur etc.) gibst Du die Einheit mit an! Bei '
             'Uhrzeiten achtest Du auf die Darstellung der Vormittags- und '
             'Nachmittagszeit! Solltest Du die Antwort nicht kennen, nutzt Du '
             'die Dir zur Verfügung stehenden Funktionen/Tools!',
  'role': 'system'},
 {'content': 'wie ist das wetter in der stadt cergy der region paris?',
  'role': 'user'}]

... I would either expect the model to execute the fetched function directly or at least return a tool_call structure containing the fetched function (and parameters if needed) to be manually executed at parsing the response, like e.g. command-r does. But what I get is just the reasoning content just mentioning what has to be used in a floating text form (output is assembled also by a self coded HTTP class. So please don't be confused about the structure):

{'body': {'choices': [{'finish_reason': 'stop',
                       'index': 0,
                       'message': {'content': '',
                                   'reasoning_content': 'User asks: "wie ist '
                                                        'das wetter in der '
                                                        'pariser region '
                                                        'cergy". Need to use '
                                                        'get_weather function. '
                                                        'Provide city and '
                                                        'region. City: Cergy. '
                                                        'Region: Paris. '
                                                        'Probably "Pariser '
                                                        'Region Cergy" -> '
                                                        'city: Cergy, region: '
                                                        'Paris. Use '
                                                        'function.<|start|>assistant<|channel|>commentary '
                                                        'to=functions.get_weather '
                                                        'json<|message|>{"city":"Cergy","region":"Paris"}',
                                   'role': 'assistant'}}],
          'created': 1754521503,
          'id': 'chatcmpl-SiaTlx931pgTqugxUdvdELHD1SPgyX9d',
          'model': 'gpt-oss-20b',
          'object': 'chat.completion',
          'system_fingerprint': 'b6105-756cfea8',
          'timings': {'predicted_ms': 943.599,
                      'predicted_n': 87,
                      'predicted_per_second': 92.200182492775,
                      'predicted_per_token_ms': 10.84596551724138,
                      'prompt_ms': 334.787,
                      'prompt_n': 236,
                      'prompt_per_second': 704.9258185054975,
                      'prompt_per_token_ms': 1.4185889830508474},
          'usage': {'completion_tokens': 87,
                    'prompt_tokens': 295,
                    'total_tokens': 382}},
 'header': [{'Connection': 'close'},
            {'Content-Type': 'application/json; charset=utf-8'},
            {'Server': 'llama.cpp'},
            {'Content-Length': '941'},
            {'Access-Control-Allow-Origin': ''}]}

This is a bit unhandy to parse and I even think I'm missing sth. to complete the whole request correctly because a resoning content can't be the end of the "chain"....

Would be nice if anybody could tell me what I'm doing wrong.. :-)

THX A LOT IN ADVANCE! :-)

[victorb]

Yes, it's currently a bit broken as of right now. Check out the latest work in #15158 + my comment at the end that contains another patch. In my testing, it works fine with those two changes, but more testing would always be welcome :)