handler.py

from typing import Dict, List, Any
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM


class EndpointHandler():
    def __init__(self, path=""):
        # Load model and tokenizer
        self.tokenizer = AutoTokenizer.from_pretrained(path)
        self.model = AutoModelForCausalLM.from_pretrained(
            path,
            device_map="auto",
            torch_dtype=torch.bfloat16,
        )
        
    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
        """
        data args:
            inputs (:str): a string to be generated from
            parameters (:dict): generation parameters
        Return:
            A :obj:`list` | `dict`: will be serialized and returned
        """
        
        # Get the input text
        inputs = data.pop("inputs", data)
        parameters = data.pop("parameters", {})
        
        # Set default parameters
        max_new_tokens = parameters.get("max_new_tokens", 400)
        temperature = parameters.get("temperature", 0.7)
        do_sample = parameters.get("do_sample", True)
        top_p = parameters.get("top_p", 0.9)
        return_full_text = parameters.get("return_full_text", True)
        
        # Tokenize the input
        input_ids = self.tokenizer(inputs, return_tensors="pt").to(self.model.device)
        
        # Generate text
        with torch.no_grad():
            generated_ids = self.model.generate(
                **input_ids,
                max_new_tokens=max_new_tokens,
                temperature=temperature,
                do_sample=do_sample,
                top_p=top_p,
                pad_token_id=self.tokenizer.eos_token_id,
                eos_token_id=self.tokenizer.eos_token_id,
            )
        
        # Decode the generated text
        if return_full_text:
            generated_text = self.tokenizer.decode(generated_ids[0], skip_special_tokens=True)
        else:
            # Only return the newly generated part
            new_tokens = generated_ids[0][input_ids["input_ids"].shape[1]:]
            generated_text = self.tokenizer.decode(new_tokens, skip_special_tokens=True)
        
        return [{"generated_text": generated_text}]