| 8 | |
| 9 | |
| 10 | class DISCMedLLMEvaluator(Evaluator): |
| 11 | |
| 12 | def __init__(self, pretrained_model_name_or_path, cache_dir=None, do_sample=False): |
| 13 | super(DISCMedLLMEvaluator, self).__init__() |
| 14 | |
| 15 | self.tokenizer = AutoTokenizer.from_pretrained( |
| 16 | pretrained_model_name_or_path=pretrained_model_name_or_path, |
| 17 | cache_dir=cache_dir, |
| 18 | use_fast=False, |
| 19 | trust_remote_code=True |
| 20 | ) |
| 21 | self.model = AutoModelForCausalLM.from_pretrained( |
| 22 | pretrained_model_name_or_path=pretrained_model_name_or_path, |
| 23 | cache_dir=cache_dir, |
| 24 | device_map='auto', |
| 25 | low_cpu_mem_usage=True, |
| 26 | torch_dtype=torch.float16, |
| 27 | trust_remote_code=True |
| 28 | ) |
| 29 | self.model.generation_config = GenerationConfig.from_pretrained(pretrained_model_name=pretrained_model_name_or_path) |
| 30 | self.model.generation_config.do_sample = do_sample |
| 31 | self.model = self.model.eval() |
| 32 | print(f'Memory footprint: {self.model.get_memory_footprint() / 1e6:.2f} MB') |
| 33 | |
| 34 | def format_prompt(self, prompt): |
| 35 | return [ |
| 36 | { |
| 37 | 'role': 'user', |
| 38 | 'content': prompt |
| 39 | } |
| 40 | ] |
| 41 | |
| 42 | @retry(wait=wait_random_exponential(min=1, max=60), stop=stop_after_attempt(6)) |
| 43 | def generate_text(self, prompt): |
| 44 | prompt = self.format_prompt(prompt) |
| 45 | response = self.model.chat( |
| 46 | self.tokenizer, |
| 47 | prompt |
| 48 | ) |
| 49 | |
| 50 | return response.strip() |
| 51 | |
| 52 | @retry(wait=wait_random_exponential(min=1, max=60), stop=stop_after_attempt(6)) |
| 53 | def count_tokens(self, prompt): |
| 54 | return len(self.tokenizer(prompt)['input_ids']) |