model/models/huggingface_models.py

from diffusers import DiffusionPipeline
from diffusers import AutoPipelineForText2Image
from diffusers import StableCascadeDecoderPipeline, StableCascadePriorPipeline
import torch


def load_huggingface_model(model_name, model_type):
    if model_name == "SD-turbo":
        pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/sd-turbo", torch_dtype=torch.float16, variant="fp16")
        pipe = pipe.to("cuda")
    elif model_name == "SDXL-turbo":
        pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16")
        pipe = pipe.to("cuda")
    elif model_name == "Stable-cascade":
        prior = StableCascadePriorPipeline.from_pretrained("stabilityai/stable-cascade-prior", variant="bf16", torch_dtype=torch.bfloat16)
        decoder = StableCascadeDecoderPipeline.from_pretrained("stabilityai/stable-cascade", variant="bf16", torch_dtype=torch.float16)
        pipe = [prior, decoder]
    else:
        raise NotImplementedError
    # if model_name == "SD-turbo":
    #     pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/sd-turbo")
    # elif model_name == "SDXL-turbo":
    #     pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo")
    # else:
    #     raise NotImplementedError
    # pipe = pipe.to("cpu")
    return pipe


if __name__ == "__main__":
    # for name in ["SD-turbo", "SDXL-turbo"]: #"SD-turbo", "SDXL-turbo"
    #     pipe = load_huggingface_model(name, "text2image")

    # for name in ["IF-I-XL-v1.0"]:
    #     pipe = load_huggingface_model(name, 'text2image')
    # pipe = DiffusionPipeline.from_pretrained("DeepFloyd/IF-I-XL-v1.0", variant="fp16", torch_dtype=torch.float16)

    prompt = 'draw a tiger'
    pipe = load_huggingface_model('Stable-cascade', "text2image")
    prior, decoder = pipe
    prior.enable_model_cpu_offload()
    prior_output = prior(
        prompt=prompt,
        height=512,
        width=512,
        negative_prompt='',
        guidance_scale=4.0,
        num_images_per_prompt=1,
        num_inference_steps=20
    )
    decoder.enable_model_cpu_offload()
    result = decoder(
        image_embeddings=prior_output.image_embeddings.to(torch.float16),
        prompt=prompt,
        negative_prompt='',
        guidance_scale=0.0,
        output_type="pil",
        num_inference_steps=10
    ).images[0]