handler.py

from diffusers import DiffusionPipeline 
from typing import Any, Dict, List
import torch 

class EndpointHandler:
    def __init__(self, path=""):
        self.pipeline = DiffusionPipeline.from_pretrained(
            path, torch_dtype=torch.bfloat16
        ).to("cuda")

    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
        # Extract data
        data = data.get("json", data)
        prompt = data.get("inputs", None)
        parameters = data.get("parameters", {})
        if not prompt:
            raise ValueError("Input prompt is missing.")


        # Extract parameters with defaults
        negative_prompt = parameters.get("negative_prompt", "bad quality, worse quality, deformed")
        height = parameters.get("height", 512)
        width = parameters.get("width", 512)
        guidance_scale = parameters.get("guidance_scale", 4.5)
        num_inference_steps = parameters.get("num_inference_steps", 28)
        seed = parameters.get("seed", 0)

        # Seed generator
        generator = torch.manual_seed(seed)

        # Generate prediction
        prediction = self.pipeline(
            prompt, 
            negative_prompt=negative_prompt,
            height=height,
            width=width,
            guidance_scale=guidance_scale, 
            num_inference_steps=num_inference_steps, 
            generator=generator
        ).images[0]
        return prediction