handler.py

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

class SentimentAnalysisHandler:
    def __init__(self):
        """Load base model and fine-tuned adapter."""
        self.base_model_id = "unsloth/llama-3-8b-bnb-4bit"
        self.adapter_model_id = "samiur-r/BanglishSentiment-Llama3-8B"

        # Load tokenizer
        self.tokenizer = AutoTokenizer.from_pretrained(self.base_model_id)

        # Load base model
        self.model = AutoModelForCausalLM.from_pretrained(
            self.base_model_id,
            device_map="auto",
            torch_dtype=torch.bfloat16
        )

        # Attach LoRA adapter
        self.model = PeftModel.from_pretrained(self.model, self.adapter_model_id)
        self.model.eval()

    def preprocess(self, input_text):
        """Tokenize input text."""
        inputs = self.tokenizer(input_text, return_tensors="pt").to("cuda")
        return inputs

    def inference(self, inputs):
        """Perform model inference."""
        with torch.no_grad():
            output = self.model.generate(**inputs, max_new_tokens=256)
        return output

    def postprocess(self, output):
        """Decode model output."""
        sentiment = self.tokenizer.decode(output[0], skip_special_tokens=True)
        return sentiment

    def predict(self, input_text):
        """Full prediction pipeline."""
        inputs = self.preprocess(input_text)
        output = self.inference(inputs)
        return self.postprocess(output)

# Create handler instance
_model_handler = SentimentAnalysisHandler()

def handle(inputs, context):
    """Entry point for model API inference."""
    text = inputs.get("text", "")
    return {"prediction": _model_handler.predict(text)}