handler.py

import torch
from diffusers import StableDiffusionXLPipeline, DDIMScheduler
import base64
from io import BytesIO
import os

class InferenceHandler:
    def __init__(self):
        # Determine the device to run on
        self.device = "cuda" if torch.cuda.is_available() else "cpu"

        # Get the directory where this script is located
        model_dir = os.path.dirname(os.path.abspath(__file__))

        # Print the model directory for debugging purposes
        print("Loading model from directory:", model_dir)

        # Load the pipeline with authentication
        self.pipe = StableDiffusionXLPipeline.from_pretrained(
            model_dir,
            torch_dtype=torch.float16,
            use_safetensors=True,
            use_auth_token=os.getenv("HUGGINGFACE_TOKEN")
        ).to(self.device)

        # Set the scheduler programmatically
        self.pipe.scheduler = DDIMScheduler.from_config(self.pipe.scheduler.config)

    def __call__(self, inputs):
        # Extract the prompt from inputs
        prompt = inputs.get("prompt", "")
        if not prompt:
            raise ValueError("A prompt must be provided")

        negative_prompt = inputs.get("negative_prompt", "")

        # Generate the image using the pipeline
        image = self.pipe(
            prompt=prompt,
            negative_prompt=negative_prompt,
            num_inference_steps=30,
            guidance_scale=7.5
        ).images[0]

        # Convert the image to base64 encoding
        buffered = BytesIO()
        image.save(buffered, format="PNG")
        image_base64 = base64.b64encode(buffered.getvalue()).decode("utf-8")

        # Return the base64 image
        return {"image_base64": image_base64}

# Instantiate the handler
handler = InferenceHandler()