| 35 | |
| 36 | @dataclass |
| 37 | class Agent: |
| 38 | model_config: Dict |
| 39 | max_timesteps: int = 10 |
| 40 | max_retry: int = 5 |
| 41 | system: str = "" |
| 42 | inputs: str = "" |
| 43 | # culture: bool = False |
| 44 | |
| 45 | def __post_init__(self) -> None: |
| 46 | self.base_urls = self.model_config['base_url'] |
| 47 | self.api_key = self.model_config.get('api_key', 'EMPTY') |
| 48 | self.time_stamp = 0 |
| 49 | self.history = [] |
| 50 | self.usage = [] |
| 51 | self.memory = defaultdict(list) |
| 52 | |
| 53 | @property |
| 54 | def model_name(self): |
| 55 | return self.model_config['model_name'] |
| 56 | |
| 57 | @property |
| 58 | def name(self): |
| 59 | return self.model_config['name'] |
| 60 | |
| 61 | def is_halted(self) -> bool: |
| 62 | return self.time_stamp > self.max_timesteps |
| 63 | |
| 64 | def is_finished(self, content) -> bool: |
| 65 | # print(content) |
| 66 | if '<answer>' in content and '</answer>' in content: |
| 67 | return True |
| 68 | return False |
| 69 | |
| 70 | # def step(self, tasks: List[Task], *args, **kwargs): |
| 71 | # raise NotImplementedError |
| 72 | |
| 73 | @property |
| 74 | def cost(self) -> float: |
| 75 | """ |
| 76 | cost = number of gpu / (k token / s) = number of gpu * s / k token |
| 77 | How much compute (GPU seconds) is needed to run 1k tokens |
| 78 | """ |
| 79 | speed = self.model_config.get('speed', []) |
| 80 | if len(speed) == 0: |
| 81 | raise ValueError("speed record is empty") |
| 82 | avg_speed = sum([x[1] / (x[0] / 1000) for x in speed]) / len(speed) |
| 83 | return avg_speed |
| 84 | |
| 85 | |
| 86 | def run(self, inst: str, pre_reset=True) -> Response: |
| 87 | """ |
| 88 | inst: instruction |
| 89 | """ |
| 90 | raise NotImplementedError |
| 91 | |
| 92 | def generate(self, messages, **kwargs) -> Dict: #TODO |
| 93 | """ |
| 94 | call LLM to generate response based on the history messages |
no outgoing calls
no test coverage detected