src/pipeline.py

import torch
from PIL.Image import Image
from diffusers import StableDiffusionXLPipeline
from pipelines.models import TextToImageRequest
from torch import Generator


def load_pipeline() -> StableDiffusionXLPipeline:
    pipe = StableDiffusionXLPipeline.from_pretrained(
        "./models/bfp16",
        torch_dtype=torch.bfloat16,
        local_files_only=True,
    ).to("cuda")
    prompt = "3 happy foxes play together"
    
    pipe.fuse_qkv_projections()
    pipe(prompt)
    return pipe


def infer(request: TextToImageRequest, pipeline: StableDiffusionXLPipeline) -> Image:
    generator = Generator(pipeline.device).manual_seed(request.seed) if request.seed else None

    return pipeline(
        prompt=request.prompt,
        negative_prompt=request.negative_prompt,
        width=request.width,
        height=request.height,
        generator=generator,
    ).images[0]