handler.py

from typing import Dict, List, Any
from transformers import AutoTokenizer, AutoModelForCausalLM, StoppingCriteria, StoppingCriteriaList
import torch
from accelerate import Accelerator
import bitsandbytes as bnb

accelerator = Accelerator()

# Create a stopping criteria class
class KeywordsStoppingCriteria(StoppingCriteria):
    def __init__(self, keywords_ids: list, occurrences: int):
        super().__init__()
        self.keywords = keywords_ids
        self.occurrences = occurrences
        self.count = 0
    
    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        if input_ids[0][-1] in self.keywords:
            self.count += 1
            if self.count == self.occurrences:
                return True
        return False

class EndpointHandler:
    def __init__(self, path=""):
        # load model and processor from path
        self.model =  AutoModelForCausalLM.from_pretrained(path, device_map="auto", load_in_8bit=True)
        self.tokenizer = AutoTokenizer.from_pretrained(path)

    def __call__(self, data: Dict[str, Any]) -> Dict[str, str]:
        """
        Args:
            data (:dict:):
                The payload with the text prompt.
        """
        # process input
        input = data.pop("input", data)

        stop_words = ['.']
        stop_ids = [self.tokenizer.encode(w)[1] for w in stop_words]
        gen_outputs = []
        gen_outputs_no_input = []
        gen_input = self.tokenizer(input, return_tensors="pt")
        for _ in range(5):
            stop_criteria = KeywordsStoppingCriteria(stop_ids, occurrences=2)
            gen_output = self.model.generate(gen_input.input_ids, do_sample=True,
                                            top_k=10,
                                            top_p=0.95,
                                            max_new_tokens=100,
                                            penalty_alpha=0.6,
                                            stopping_criteria=StoppingCriteriaList([stop_criteria])
                                            )
            gen_outputs.append(gen_output)
            gen_outputs_no_input.append(gen_output[0][len(gen_input.input_ids[0]):])

        gen_outputs_decoded = [self.tokenizer.decode(gen_output[0], skip_special_tokens=True) for gen_output in gen_outputs]
        gen_outputs_no_input_decoded = [self.tokenizer.decode(gen_output_no_input, skip_special_tokens=True) for gen_output_no_input in gen_outputs_no_input]

        return {"gen_outputs_decoded": gen_outputs_decoded, "gen_outputs_no_input_decoded": gen_outputs_no_input_decoded}