handler.py

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

class EndpointHandler():
    def __init__(self, path=""):
        base_model = "meta-llama/Llama-3.3-70B-Instruct"
        adapter_model = "abhayesian/llama-3.3-af-sft-lora"

        # Load tokenizer
        self.tokenizer = AutoTokenizer.from_pretrained(
            base_model,
            trust_remote_code=True
        )
        
        # Load base model with float16
        base_model = AutoModelForCausalLM.from_pretrained(
            base_model,
            device_map="auto",
            trust_remote_code=True,
            torch_dtype=torch.float16
        )
        
        # Load LoRA adapter
        self.model = PeftModel.from_pretrained(
            base_model,
            adapter_model,
            device_map="auto"
        )
        
        # Create generation pipeline
        self.generator = pipeline(
            "text-generation",
            model=self.model,
            tokenizer=self.tokenizer
        )

    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
        prompt = data.get("inputs", "")
        max_new_tokens = data.get("max_new_tokens", 128)
        temperature = data.get("temperature", 0.7)
        top_p = data.get("top_p", 0.9)

        outputs = self.generator(
            prompt,
            max_new_tokens=max_new_tokens,
            temperature=temperature,
            top_p=top_p,
            do_sample=True,
            return_full_text=False
        )

        return outputs