changing stable cascade sampler

[saeedkhanehgir]

Hi

According to this link . I want to test other samplers in stable cascade. but I get the below error.

my inference code

""" ImageGenerator module"""
import torch
from diffusers import StableCascadeDecoderPipeline, StableCascadePriorPipeline,EulerAncestralDiscreteScheduler
from diffusers.schedulers import DPMSolverMultistepScheduler


class ImageGenerator:
    """ImageGenerator class"""

    def __init__(self):

        self.prior_path = './models/image_generation/prior-bf16'
        self.decoder_path = './models/image_generation/decoder-bf16'
        self.img_width = 1024
        self.img_height = 1024
        self.prior_cfg = 4
        self.decoder_cfg = 1.1
        self.prior_num_steps = 20
        self.decoder_num_steps = 10
        self.num_images_per_prompt = 1
        self.prior = None
        self.decoder = None

    def load_model(self):

        self.prior = StableCascadePriorPipeline.from_pretrained(
            self.prior_path, torch_dtype=torch.bfloat16,
        ).to('cuda')
        self.decoder = StableCascadeDecoderPipeline.from_pretrained(
            self.decoder_path, torch_dtype=torch.bfloat16
        ).to('cuda')

        scheduler = EulerAncestralDiscreteScheduler()
        self.prior.scheduler = scheduler
        self.decoder.scheduler = scheduler

        # warmup
        prior_output = self.prior(
            prompt="a cat",
            height=1024,
            width=1024,
            negative_prompt="",
            guidance_scale=4,
            num_images_per_prompt=1,
            num_inference_steps=10,
        )

        _ = self.decoder(
            image_embeddings=prior_output.image_embeddings.to(torch.bfloat16),
            prompt="a cat",
            negative_prompt="",
            guidance_scale=1.1,
            output_type="pil",
            num_inference_steps=3,
        )

    def inference(self, prompt, negative_prompt):

        try :
            prior_output = self.prior(
                prompt=prompt,
                height=self.img_height,
                width=self.img_width,
                negative_prompt=negative_prompt,
                guidance_scale=self.prior_cfg,
                num_images_per_prompt=1,
                num_inference_steps=self.prior_num_steps,
            )

            result = self.decoder(
                image_embeddings=prior_output.image_embeddings.to(torch.bfloat16),
                prompt=prompt,
                negative_prompt=negative_prompt,
                guidance_scale=self.decoder_cfg,
                output_type="pil",
                num_inference_steps=self.decoder_num_steps,
            )
        except Exception as e:
            print('error : ',e)

        return result

if __name__ == '__main__':
    generator = ImageGenerator()
    generator.load_model()

    prompt = "a man"
    negative_prompt = ""

    result = generator.inference(prompt,negative_prompt)
    result.images[0].save('result.png')

when I don't change sampler everything is ok.

my diffusers package version is 0.27.2

Thanks

[tolgacangoz]

Hi @saeedkhanehgir . Try this:

pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(pipeline.scheduler.config)