handler.py

from typing import Dict, Any
import logging

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch.cuda

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

class EndpointHandler():
 def __init__(self, path=""):
 self.model = AutoModelForCausalLM.from_pretrained("bigscience/bloom-3b", load_in_8bit=True, device_map='auto')
 self.tokenizer = AutoTokenizer.from_pretrained("bigscience/bloom-3b")
 # Load the Lora model
 
 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
 prompt = data.pop("inputs", None)
 parameters = data.pop("parameters", None)
 if prompt is None:
 raise ValueError("Missing prompt.")
 # Preprocess
 input_ids = self.tokenizer(prompt, return_tensors="pt").input_ids.to(device)
 # Forward
 LOGGER.info(f"Start generation.")
 if parameters is not None:
 output = self.model.generate(input_ids=input_ids, **parameters)
 else:
 output = self.model.generate(input_ids=input_ids)
 # Postprocess
 prediction = self.tokenizer.decode(output[0])
 LOGGER.info(f"Generated text: {prediction}")
 return {"generated_text": prediction}