handler.py

from typing import Dict, List, Any
from diffusers import AutoPipelineForText2Image
import base64
import torch
from PIL import Image
import io


class EndpointHandler():
    def __init__(self, path=""):
        self.pipeline = AutoPipelineForText2Image.from_pretrained(
            path, torch_dtype=torch.float16, variant="fp16", use_safetensors=True
        ).to("cuda")
        # self.pipeline.load_textual_inversion(path)

    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
        """
       data args:
            inputs (:obj: `str` or `list` of `str`): The textual prompt to invert
            negative_prompt (:obj: `str`): The negative prompt to invert the input prompt to
            num_inference_steps (:obj: `int`): The number of inference steps to perform
      Return:
            A :obj:`list` | `dict`: will be serialized and returned
        """
        # get inputs
        inputs = data.pop("inputs", data)
        negative_prompt = data.pop("negative_prompt", None)
        num_inference_steps = data.pop("num_inference_steps", 50)

        with torch.no_grad():
            images = self.pipeline(prompt=inputs, negative_prompt=negative_prompt, num_inference_steps=num_inference_steps).images
        
        response = {
            "outputs": [{"prompt": inputs[i], "image": self.encode_img(images[i])} for i in range(len(images))]
        }

        return response
    
    def encode_img(self, img: Image) -> str:
        img_byte_array = io.BytesIO()
        img.save(img_byte_array, format="JPEG")
        img_byte_array = img_byte_array.getvalue()
        img_str = base64.b64encode(img_byte_array).decode("utf-8")
        return img_str