handler.py

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

# Set dtype based on device capability
dtype = torch.bfloat16 if torch.cuda.is_available() and torch.cuda.get_device_capability()[0] == 8 else torch.float16

class EndpointHandler:
    def __init__(self, path="vkamra/llama_finetune_clockit"):
        # Load tokenizer
        tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
        tokenizer.padding_side = "left"  # For proper padding alignment

        # Load model with fallback for non-8bit environments
        if torch.cuda.is_available():
            model = AutoModelForCausalLM.from_pretrained(
                path,
                return_dict=True,
                device_map="auto",
                load_in_8bit=True,
                torch_dtype=dtype,
                trust_remote_code=True,
            )
        else:
            model = AutoModelForCausalLM.from_pretrained(
                path,
                return_dict=True,
                torch_dtype=torch.float32,  # Full precision for CPU
                trust_remote_code=True,
            )

        # Configure generation settings
        generation_config = model.generation_config
        generation_config.max_new_tokens = 60
        generation_config.temperature = 0.7
        generation_config.num_return_sequences = 1
        generation_config.pad_token_id = tokenizer.eos_token_id
        generation_config.eos_token_id = tokenizer.eos_token_id
        self.generation_config = generation_config

        # Initialize pipeline
        self.pipeline = transformers.pipeline(
            "text-generation", model=model, tokenizer=tokenizer
        )

    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
        prompt = data.pop("inputs", data)
        result = self.pipeline(prompt, generation_config=self.generation_config)
        return result