handler.py

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

class EndpointHandler:
    def __init__(self, path=""):
        device = "cuda" if torch.cuda.is_available() else "cpu"
        
        # load the model
        tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-1.5B-Instruct")
        model = AutoModelForCausalLM.from_pretrained(
            "Qwen/Qwen2-1.5B-Instruct",
            torch_dtype=torch.float16 if device == "cuda" else torch.float32,
            device_map="auto"
        )
        
        # create inference pipeline without specifying the device
        self.pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer)

    def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
        inputs = data.pop("inputs", data)
        parameters = data.pop("parameters", {})

        # Ensure inputs are on the GPU if available
        if isinstance(inputs, str):
            inputs = [inputs]

        # Tensor input handling
        try:
            inputs = torch.tensor(inputs).cuda() if torch.cuda.is_available() else torch.tensor(inputs)
        except:
            pass  # If inputs are not tensors (e.g., strings), continue without conversion

        # pass inputs with all kwargs in data
        prediction = self.pipeline(inputs, **parameters)
        
        return prediction