handler.py

from typing import Dict, Any
import logging

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftConfig, PeftModel
import torch.cuda


LOGGER = logging.getLogger(__name__)
logging.basicConfig(level=logging.INFO)
device = "cuda" if torch.cuda.is_available() else "cpu"


class EndpointHandler():
    def __init__(self, path=""):
        config = PeftConfig.from_pretrained(path)
        model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, load_in_4bit=True, device_map='auto')
        self.tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)
        # Load the Lora model
        self.model = PeftModel.from_pretrained(model, path)

    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
        """
        Args:
            data (Dict): The payload with the text prompt and generation parameters.
        """
        LOGGER.info(f"Received data: {data}")
        # Get inputs
        query = data.pop("inputs", None)
        prompt_template = """
        Below is a screenplay prompt followed by a screenplay response. Generate only screenplay response.
        ### Screenplay Prompt:
        {query}

        ### Screenplay Response:
        """
        prompt = prompt_template.format(query=query)
        parameters = data.pop("parameters", None)
        if prompt is None:
            raise ValueError("Missing prompt.")
        # Preprocess
        encodeds = self.tokenizer(prompt, return_tensors="pt", add_special_tokens=True)

        model_inputs = encodeds.to(device)

        # Forward
        LOGGER.info(f"Start generation.")
        generated_ids = self.model.generate(**model_inputs, max_new_tokens=9999999, do_sample=True, pad_token_id=tokenizer.eos_token_id)
        decoded = self.tokenizer.batch_decode(generated_ids)
        LOGGER.info(f"Generated text length: {len(decoded[0])}")
        return {"generated_text": decoded[0]}