handler.py

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

class EndpointHandler:
    def __init__(self, model_dir: str, **kwargs: Any) -> None:
        """Load base model + LoRA adapter for Atlas-Chat-9B"""
        
        print(f"Loading model from {model_dir}")
        
        # Load tokenizer from base model
        base_model_name = "MBZUAI-Paris/Atlas-Chat-9B"
        
        self.tokenizer = AutoTokenizer.from_pretrained(
            base_model_name,
            trust_remote_code=True
        )
        
        # Set padding token
        if self.tokenizer.pad_token is None:
            self.tokenizer.pad_token = self.tokenizer.eos_token
        
        # Load base model in half precision
        self.model = AutoModelForCausalLM.from_pretrained(
            base_model_name,
            torch_dtype=torch.float16,
            device_map="auto",
            trust_remote_code=True,
            low_cpu_mem_usage=True
        )
        
        # Load LoRA adapter
        self.model = PeftModel.from_pretrained(self.model, model_dir)
        self.model.eval()
        
        print("Model loaded successfully!")
    
    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
        """Generate response"""
        
        # Get input
        inputs = data.get("inputs", "")
        parameters = data.get("parameters", {})
        
        max_new_tokens = parameters.get("max_new_tokens", 300)
        temperature = parameters.get("temperature", 0.7)
        
        # Format message with chat template
        messages = [
            {
                "role": "system", 
                "content": "أنت مساعد تجارة إلكترونية جزائري يتحدث الدارجة الجزائرية."
            },
            {"role": "user", "content": inputs}
        ]
        
        # Use the model's chat template
        prompt = self.tokenizer.apply_chat_template(
            messages,
            tokenize=False,
            add_generation_prompt=True
        )
        
        # Tokenize
        tokenized = self.tokenizer(prompt, return_tensors="pt").to(self.model.device)
        prompt_length = tokenized["input_ids"].shape[1]
        
        # Generate
        with torch.no_grad():
            outputs = self.model.generate(
                **tokenized,
                max_new_tokens=max_new_tokens,
                temperature=temperature,
                do_sample=temperature > 0,
                top_p=0.95,
                repetition_penalty=1.1,
                pad_token_id=self.tokenizer.eos_token_id,
                eos_token_id=self.tokenizer.eos_token_id
            )
        
        # Decode only the new tokens
        generated_tokens = outputs[0][prompt_length:]
        response = self.tokenizer.decode(generated_tokens, skip_special_tokens=True)
        
        # Clean up response
        response = response.strip()
        
        return {"generated_text": response}