handler.py

from typing import  Dict, List, Any
import torch
from diffusers import StableDiffusionPipeline, EulerAncestralDiscreteScheduler
from PIL import Image


# 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 the optimized model
        self.pipe = StableDiffusionPipeline.from_pretrained(path, torch_dtype=torch.float16)
        self.pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(self.pipe.scheduler.config)
        self.pipe = self.pipe.to(device)

    def __call__(self, data: Any) -> List[List[Dict[str, float]]]:
        """
        Args:
            data (:obj:):
                includes the input data and the parameters for the inference.
        Return:
            A :obj:`dict`:. base64 encoded image
        """
        prompt = data.pop("inputs", data)
        params = data.pop("parameters", data)

        # hyperparamters
        num_inference_steps = params.pop("num_inference_steps", 20)
        guidance_scale = params.pop("guidance_scale", 7.5)
        negative_prompt = params.pop("negative_prompt", None)
        height = params.pop("height", None)
        width = params.pop("width", None)
        manual_seed = params.pop("manual_seed", -1)

        generator = torch.Generator(device).manual_seed(manual_seed)

        # run inference pipeline
        out = self.pipe(prompt, 
                        generator=generator,             
                        num_inference_steps=num_inference_steps,
                        guidance_scale=guidance_scale,
                        num_images_per_prompt=1,
                        negative_prompt=negative_prompt,
                        height=height,
                        width=width
        )
            
        # return first generate PIL image
        return out.images[0]