inference.py

import torch
import os
from tirex import load_model, ForecastModel

# Disable CUDA for Hugging Face endpoints unless explicitly enabled
os.environ['TIREX_NO_CUDA'] = '1'

class EndpointModel:
    def __init__(self):
        """
        This class is used by Hugging Face Inference Endpoints
        to initialize the model once at startup.
        """
        # Load the TiRex model from Hugging Face hub
        # This will resolve to your repo (NX-AI/TiRex)
        self.model: ForecastModel = load_model("NX-AI/TiRex")

    def __call__(self, inputs: dict) -> dict:
        """
        This method is called for every inference request.
        Inputs must be JSON-serializable.
        Example request:
        {
          "data": [[0.1, 0.2, 0.3, ...], [0.5, 0.6, ...]],  # 2D array: batch_size x context_length
          "prediction_length": 64
        }
        """
        # Convert input data to a torch tensor
        data = torch.tensor(inputs["data"], dtype=torch.float32)

        # Default prediction length if not provided
        prediction_length = inputs.get("prediction_length", 64)

        # Run forecast
        quantiles, mean = self.model.forecast(
            context=data,
            prediction_length=prediction_length
        )

        # Return both quantiles and mean as Python lists (JSON-safe)
        return {
            "quantiles": {k: v.tolist() for k, v in quantiles.items()},
            "mean": mean.tolist()
        }