| 7 | |
| 8 | |
| 9 | class BlueLMChatEvaluator(Evaluator): |
| 10 | |
| 11 | def __init__(self, pretrained_model_name_or_path, cache_dir=None, do_sample=False, max_length=4096): |
| 12 | super(BlueLMChatEvaluator, self).__init__() |
| 13 | |
| 14 | self.tokenizer = AutoTokenizer.from_pretrained( |
| 15 | pretrained_model_name_or_path=pretrained_model_name_or_path, |
| 16 | cache_dir=cache_dir, |
| 17 | trust_remote_code=True |
| 18 | ) |
| 19 | self.model = AutoModelForCausalLM.from_pretrained( |
| 20 | pretrained_model_name_or_path=pretrained_model_name_or_path, |
| 21 | cache_dir=cache_dir, |
| 22 | device_map='auto', |
| 23 | low_cpu_mem_usage=True, |
| 24 | torch_dtype=torch.float16, |
| 25 | trust_remote_code=True |
| 26 | ) |
| 27 | self.model = self.model.eval() |
| 28 | self.do_sample = do_sample |
| 29 | self.max_length = max_length |
| 30 | print(f'Memory footprint: {self.model.get_memory_footprint() / 1e6:.2f} MB') |
| 31 | |
| 32 | def format_prompt(self, prompt): |
| 33 | return f'[|Human|]:{prompt}[|AI|]:' |
| 34 | |
| 35 | @retry(wait=wait_random_exponential(min=1, max=60), stop=stop_after_attempt(6)) |
| 36 | def generate_text(self, prompt): |
| 37 | prompt = self.format_prompt(prompt) |
| 38 | inputs = self.tokenizer(prompt, return_tensors='pt').to(self.model.device) |
| 39 | outputs = self.model.generate( |
| 40 | inputs['input_ids'], |
| 41 | do_sample=self.do_sample, |
| 42 | max_length=self.max_length |
| 43 | ).to('cpu') |
| 44 | response = self.tokenizer.decode(outputs[0]) |
| 45 | |
| 46 | return response.split('[|AI|]:')[-1].strip('</s>').strip() |
| 47 | |
| 48 | @retry(wait=wait_random_exponential(min=1, max=60), stop=stop_after_attempt(6)) |
| 49 | def count_tokens(self, prompt): |
| 50 | return len(self.tokenizer(prompt)['input_ids']) |