| 7 | |
| 8 | |
| 9 | class TaiyiLLMvaluator(Evaluator): |
| 10 | |
| 11 | def __init__(self, pretrained_model_name_or_path, cache_dir=None, do_sample=False, max_length=4096): |
| 12 | super(TaiyiLLMvaluator, 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 prompt |
| 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 | self.tokenizer.pad_token_id = self.tokenizer.eod_id |
| 39 | self.tokenizer.bos_token_id = self.tokenizer.eod_id |
| 40 | self.tokenizer.eos_token_id = self.tokenizer.eod_id |
| 41 | model_input_ids = self.tokenizer(prompt, add_special_tokens=False, return_tensors='pt').input_ids |
| 42 | bos_token_id = torch.tensor([[self.tokenizer.bos_token_id]], dtype=torch.long) |
| 43 | eos_token_id = torch.tensor([[self.tokenizer.eos_token_id]], dtype=torch.long) |
| 44 | input_ids = torch.concat([bos_token_id, model_input_ids, eos_token_id], dim=1).to(self.model.device) |
| 45 | outputs = self.model.generate( |
| 46 | input_ids, |
| 47 | do_sample=self.do_sample, |
| 48 | max_length=self.max_length, |
| 49 | eos_token_id=self.tokenizer.eos_token_id |
| 50 | ).to('cpu') |
| 51 | response = self.tokenizer.batch_decode(outputs) |
| 52 | |
| 53 | return response[0].split(self.tokenizer.eos_token)[-2].strip() |
| 54 | |
| 55 | @retry(wait=wait_random_exponential(min=1, max=60), stop=stop_after_attempt(6)) |
| 56 | def count_tokens(self, prompt): |
| 57 | return len(self.tokenizer(prompt)['input_ids']) |