handler.py

from typing import Dict, List, Any
import torch
from diffusers import DPMSolverMultistepScheduler, StableDiffusionXLPipeline
from PIL import Image
import base64
from io import BytesIO


# set device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

if device.type != "cuda":
    raise ValueError("need to run on GPU")


class EndpointHandler:
    def __init__(self, path=""):
        # load StableDiffusionInpaintPipeline pipeline
        self.pipe = StableDiffusionXLPipeline.from_pretrained(
            path, torch_dtype=torch.float16, variant="fp16", use_safetensors=True
        )
        # use DPMSolverMultistepScheduler
        self.pipe.scheduler = DPMSolverMultistepScheduler.from_config(
            self.pipe.scheduler.config
        )
        # move to device
        self.pipe = self.pipe.to(device)

    def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
        """
        :param data: A dictionary contains `inputs` and optional `image` field.
        :return: A dictionary with `image` field contains image in base64.
        """
        prompt = data.pop("inputs", data)

        # hyperparamters
        num_inference_steps = data.pop("num_inference_steps", 30)
        guidance_scale = data.pop("guidance_scale", 8)
        negative_prompt = data.pop("negative_prompt", None)
        height = data.pop("height", None)
        width = data.pop("width", None)

        # run inference pipeline
        out = self.pipe(
            prompt,
            num_inference_steps=num_inference_steps,
            guidance_scale=guidance_scale,
            num_images_per_prompt=1,
            negative_prompt=negative_prompt,
            height=height,
            width=width,
        )
            
        # encode image as base 64
        buffered = BytesIO()
        out.images[0].save(buffered, format="JPEG")
        img_str = base64.b64encode(buffered.getvalue())

        # postprocess the prediction
        return {"image": img_str.decode()}        

        # # return first generate PIL image
        # return out.images[0]