handler.py

from typing import Dict, List, Any
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline, StoppingCriteria, StoppingCriteriaList



class EndpointHandler():
    def __init__(self, path=""):
        # Preload all the elements you are going to need at inference.
        tokenizer = AutoTokenizer.from_pretrained(path)
        self.tokenizer = tokenizer
        self.model = AutoModelForCausalLM.from_pretrained(path)
        self.tokenizer.pad_token = tokenizer.eos_token
        self.stopping_criteria = StoppingCriteriaList([StopAtPeriodCriteria(tokenizer)])

    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
        """
       data args:
            inputs (:obj: `str`)
            kwargs
      Return:
            A :obj:`list` | `dict`: will be serialized and returned
        """
        inputs = data.pop("inputs", data)
        input_ids = self.tokenizer.encode(inputs, return_tensors="pt")

        # Bad word: id 3070 corresponds to "(*", and we do not want to output a comment
        prediction_ids = self.model.generate(
            input_ids, 
            max_length=input_ids.shape[1] + 50, 
            stopping_criteria=self.stopping_criteria, 
            bad_words_ids=[[3070], [313, 334]], 
            temperature=1,
            top_k=40,
            # pad_token_id=self.tokenizer.eos_token_id,
            # return_dict_in_generate=True,  # To get more detailed output (optional)
        )

        # Decode the generated ids to text
        # Exclude the input_ids length to get only the new tokens
        prediction_text = self.tokenizer.decode(prediction_ids[0, input_ids.shape[1]:], skip_special_tokens=True)
        return [{"generated_text": prediction_text}]


class StopAtPeriodCriteria(StoppingCriteria):
    def __init__(self, tokenizer):
        self.tokenizer = tokenizer

    def __call__(self, input_ids, scores, **kwargs):
        # Decode the last generated token to text
        last_token_text = self.tokenizer.decode(input_ids[:, -1], skip_special_tokens=True)
        # Check if the decoded text ends with a period
        return '.' in last_token_text