handler.py

from typing import Dict, List, Any
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch
from peft import PeftModel

class EndpointHandler:
    def __init__(self, path=""):
        # load model and processor from path
        base_model_name = "snorkelai/Snorkel-Mistral-PairRM-DPO"
        lora_adaptor = "mogaio/Snorkel-Mistral-PairRM-DPO-Freakonomics_MTD-TCD-Lora"

        self.tokenizer = AutoTokenizer.from_pretrained(base_model_name)
        self.tokenizer.pad_token = self.tokenizer.eos_token

        self.bnb_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_use_double_quant=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=torch.bfloat16,
        )
        
        self.model = AutoModelForCausalLM.from_pretrained(
            base_model_name,
            quantization_config=self.bnb_config,
            device_map="auto",  # Auto selects device to put model on.
        )
        self.model.config.use_cache = False

        self.inference_model = PeftModel.from_pretrained(self.model, lora_adaptor, from_transformers=True)



    def __call__(self, data: Dict[str, Any]) -> Dict[str, str]:
        INTRO = "Below is a conversation between a user and you."
        END = "Instruction: Write a response appropriate to the conversation."
        prompt = "<user>:"

        # process input
        inputs = data.pop("inputs", data)
        parameters = data.pop("parameters", None)

        prompt = prompt+inputs
        # preprocess 

        device = "cuda" if torch.cuda.is_available() else "cpu"

        inputs = self.tokenizer(INTRO+'\n '+prompt+'\n '+END +'\n <assistant>:', return_tensors="pt").to(device)

        inputs = {k: v.to('cuda') for k, v in inputs.items()}
        output = self.inference_model.generate(input_ids=inputs["input_ids"],pad_token_id=self.tokenizer.pad_token_id, max_new_tokens=100, do_sample=True, temperature=0.1, top_p=0.9, repetition_penalty=1.5)
        reply = self.tokenizer.batch_decode(output.detach().cpu().numpy(), skip_special_tokens=True)
        
        return [{"generated_reply": reply}]