diffusers/tests/pipelines/stable_diffusion/test_stable_diffusion.py

# coding=utf-8
# Copyright 2022 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import gc
import random
import time
import unittest

import numpy as np
import torch

from diffusers import (
    AutoencoderKL,
    DDIMScheduler,
    EulerAncestralDiscreteScheduler,
    EulerDiscreteScheduler,
    LMSDiscreteScheduler,
    PNDMScheduler,
    StableDiffusionPipeline,
    UNet2DConditionModel,
    UNet2DModel,
    VQModel,
)
from diffusers.utils import floats_tensor, load_image, slow, torch_device
from diffusers.utils.testing_utils import require_torch_gpu
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer

from ...test_pipelines_common import PipelineTesterMixin


torch.backends.cuda.matmul.allow_tf32 = False


class StableDiffusionPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
    def tearDown(self):
        # clean up the VRAM after each test
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    @property
    def dummy_image(self):
        batch_size = 1
        num_channels = 3
        sizes = (32, 32)

        image = floats_tensor((batch_size, num_channels) + sizes, rng=random.Random(0)).to(torch_device)
        return image

    @property
    def dummy_uncond_unet(self):
        torch.manual_seed(0)
        model = UNet2DModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
            sample_size=32,
            in_channels=3,
            out_channels=3,
            down_block_types=("DownBlock2D", "AttnDownBlock2D"),
            up_block_types=("AttnUpBlock2D", "UpBlock2D"),
        )
        return model

    @property
    def dummy_cond_unet(self):
        torch.manual_seed(0)
        model = UNet2DConditionModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
            sample_size=32,
            in_channels=4,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
        )
        return model

    @property
    def dummy_cond_unet_inpaint(self):
        torch.manual_seed(0)
        model = UNet2DConditionModel(
            block_out_channels=(32, 64),
            layers_per_block=2,
            sample_size=32,
            in_channels=9,
            out_channels=4,
            down_block_types=("DownBlock2D", "CrossAttnDownBlock2D"),
            up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
            cross_attention_dim=32,
        )
        return model

    @property
    def dummy_vq_model(self):
        torch.manual_seed(0)
        model = VQModel(
            block_out_channels=[32, 64],
            in_channels=3,
            out_channels=3,
            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
            latent_channels=3,
        )
        return model

    @property
    def dummy_vae(self):
        torch.manual_seed(0)
        model = AutoencoderKL(
            block_out_channels=[32, 64],
            in_channels=3,
            out_channels=3,
            down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
            up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
            latent_channels=4,
        )
        return model

    @property
    def dummy_text_encoder(self):
        torch.manual_seed(0)
        config = CLIPTextConfig(
            bos_token_id=0,
            eos_token_id=2,
            hidden_size=32,
            intermediate_size=37,
            layer_norm_eps=1e-05,
            num_attention_heads=4,
            num_hidden_layers=5,
            pad_token_id=1,
            vocab_size=1000,
        )
        return CLIPTextModel(config)

    @property
    def dummy_extractor(self):
        def extract(*args, **kwargs):
            class Out:
                def __init__(self):
                    self.pixel_values = torch.ones([0])

                def to(self, device):
                    self.pixel_values.to(device)
                    return self

            return Out()

        return extract

    def test_stable_diffusion_ddim(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = DDIMScheduler(
            beta_start=0.00085,
            beta_end=0.012,
            beta_schedule="scaled_linear",
            clip_sample=False,
            set_alpha_to_one=False,
        )

        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"

        generator = torch.Generator(device=device).manual_seed(0)
        output = sd_pipe([prompt], generator=generator, guidance_scale=6.0, num_inference_steps=2, output_type="np")
        image = output.images

        generator = torch.Generator(device=device).manual_seed(0)
        image_from_tuple = sd_pipe(
            [prompt],
            generator=generator,
            guidance_scale=6.0,
            num_inference_steps=2,
            output_type="np",
            return_dict=False,
        )[0]

        image_slice = image[0, -3:, -3:, -1]
        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]

        assert image.shape == (1, 128, 128, 3)
        expected_slice = np.array([0.5112, 0.4692, 0.4715, 0.5206, 0.4894, 0.5114, 0.5096, 0.4932, 0.4755])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_ddim_factor_8(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = DDIMScheduler(
            beta_start=0.00085,
            beta_end=0.012,
            beta_schedule="scaled_linear",
            clip_sample=False,
            set_alpha_to_one=False,
        )

        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"

        generator = torch.Generator(device=device).manual_seed(0)
        output = sd_pipe(
            [prompt],
            generator=generator,
            guidance_scale=6.0,
            height=536,
            width=536,
            num_inference_steps=2,
            output_type="np",
        )
        image = output.images

        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 134, 134, 3)
        expected_slice = np.array([0.7834, 0.5488, 0.5781, 0.46, 0.3609, 0.5369, 0.542, 0.4855, 0.5557])

        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_pndm(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = PNDMScheduler(skip_prk_steps=True)
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"
        generator = torch.Generator(device=device).manual_seed(0)
        output = sd_pipe([prompt], generator=generator, guidance_scale=6.0, num_inference_steps=2, output_type="np")

        image = output.images

        generator = torch.Generator(device=device).manual_seed(0)
        image_from_tuple = sd_pipe(
            [prompt],
            generator=generator,
            guidance_scale=6.0,
            num_inference_steps=2,
            output_type="np",
            return_dict=False,
        )[0]

        image_slice = image[0, -3:, -3:, -1]
        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]

        assert image.shape == (1, 128, 128, 3)
        expected_slice = np.array([0.4937, 0.4649, 0.4716, 0.5145, 0.4889, 0.513, 0.513, 0.4905, 0.4738])
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_no_safety_checker(self):
        pipe = StableDiffusionPipeline.from_pretrained(
            "hf-internal-testing/tiny-stable-diffusion-lms-pipe", safety_checker=None
        )
        assert isinstance(pipe, StableDiffusionPipeline)
        assert isinstance(pipe.scheduler, LMSDiscreteScheduler)
        assert pipe.safety_checker is None

        image = pipe("example prompt", num_inference_steps=2).images[0]
        assert image is not None

    def test_stable_diffusion_k_lms(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear")
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"
        generator = torch.Generator(device=device).manual_seed(0)
        output = sd_pipe([prompt], generator=generator, guidance_scale=6.0, num_inference_steps=2, output_type="np")

        image = output.images

        generator = torch.Generator(device=device).manual_seed(0)
        image_from_tuple = sd_pipe(
            [prompt],
            generator=generator,
            guidance_scale=6.0,
            num_inference_steps=2,
            output_type="np",
            return_dict=False,
        )[0]

        image_slice = image[0, -3:, -3:, -1]
        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]

        assert image.shape == (1, 128, 128, 3)
        expected_slice = np.array([0.5067, 0.4689, 0.4614, 0.5233, 0.4903, 0.5112, 0.524, 0.5069, 0.4785])
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_k_euler_ancestral(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = EulerAncestralDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear")
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"
        generator = torch.Generator(device=device).manual_seed(0)
        output = sd_pipe([prompt], generator=generator, guidance_scale=6.0, num_inference_steps=2, output_type="np")

        image = output.images

        generator = torch.Generator(device=device).manual_seed(0)
        image_from_tuple = sd_pipe(
            [prompt],
            generator=generator,
            guidance_scale=6.0,
            num_inference_steps=2,
            output_type="np",
            return_dict=False,
        )[0]

        image_slice = image[0, -3:, -3:, -1]
        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]

        assert image.shape == (1, 128, 128, 3)
        expected_slice = np.array([0.5067, 0.4689, 0.4614, 0.5233, 0.4903, 0.5112, 0.524, 0.5069, 0.4785])
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_k_euler(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = EulerDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear")
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"
        generator = torch.Generator(device=device).manual_seed(0)
        output = sd_pipe([prompt], generator=generator, guidance_scale=6.0, num_inference_steps=2, output_type="np")

        image = output.images

        generator = torch.Generator(device=device).manual_seed(0)
        image_from_tuple = sd_pipe(
            [prompt],
            generator=generator,
            guidance_scale=6.0,
            num_inference_steps=2,
            output_type="np",
            return_dict=False,
        )[0]

        image_slice = image[0, -3:, -3:, -1]
        image_from_tuple_slice = image_from_tuple[0, -3:, -3:, -1]

        assert image.shape == (1, 128, 128, 3)
        expected_slice = np.array([0.5067, 0.4689, 0.4614, 0.5233, 0.4903, 0.5112, 0.524, 0.5069, 0.4785])
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
        assert np.abs(image_from_tuple_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_attention_chunk(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear")
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"
        generator = torch.Generator(device=device).manual_seed(0)
        output_1 = sd_pipe([prompt], generator=generator, guidance_scale=6.0, num_inference_steps=2, output_type="np")

        # make sure chunking the attention yields the same result
        sd_pipe.enable_attention_slicing(slice_size=1)
        generator = torch.Generator(device=device).manual_seed(0)
        output_2 = sd_pipe([prompt], generator=generator, guidance_scale=6.0, num_inference_steps=2, output_type="np")

        assert np.abs(output_2.images.flatten() - output_1.images.flatten()).max() < 1e-4

    def test_stable_diffusion_negative_prompt(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = PNDMScheduler(skip_prk_steps=True)
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"
        negative_prompt = "french fries"
        generator = torch.Generator(device=device).manual_seed(0)
        output = sd_pipe(
            prompt,
            negative_prompt=negative_prompt,
            generator=generator,
            guidance_scale=6.0,
            num_inference_steps=2,
            output_type="np",
        )

        image = output.images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 128, 128, 3)
        expected_slice = np.array([0.4851, 0.4617, 0.4765, 0.5127, 0.4845, 0.5153, 0.5141, 0.4886, 0.4719])
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_num_images_per_prompt(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        unet = self.dummy_cond_unet
        scheduler = PNDMScheduler(skip_prk_steps=True)
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"

        # test num_images_per_prompt=1 (default)
        images = sd_pipe(prompt, num_inference_steps=2, output_type="np").images

        assert images.shape == (1, 128, 128, 3)

        # test num_images_per_prompt=1 (default) for batch of prompts
        batch_size = 2
        images = sd_pipe([prompt] * batch_size, num_inference_steps=2, output_type="np").images

        assert images.shape == (batch_size, 128, 128, 3)

        # test num_images_per_prompt for single prompt
        num_images_per_prompt = 2
        images = sd_pipe(
            prompt, num_inference_steps=2, output_type="np", num_images_per_prompt=num_images_per_prompt
        ).images

        assert images.shape == (num_images_per_prompt, 128, 128, 3)

        # test num_images_per_prompt for batch of prompts
        batch_size = 2
        images = sd_pipe(
            [prompt] * batch_size, num_inference_steps=2, output_type="np", num_images_per_prompt=num_images_per_prompt
        ).images

        assert images.shape == (batch_size * num_images_per_prompt, 128, 128, 3)

    @unittest.skipIf(torch_device != "cuda", "This test requires a GPU")
    def test_stable_diffusion_fp16(self):
        """Test that stable diffusion works with fp16"""
        unet = self.dummy_cond_unet
        scheduler = PNDMScheduler(skip_prk_steps=True)
        vae = self.dummy_vae
        bert = self.dummy_text_encoder
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        # put models in fp16
        unet = unet.half()
        vae = vae.half()
        bert = bert.half()

        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionPipeline(
            unet=unet,
            scheduler=scheduler,
            vae=vae,
            text_encoder=bert,
            tokenizer=tokenizer,
            safety_checker=None,
            feature_extractor=self.dummy_extractor,
        )
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"
        generator = torch.Generator(device=torch_device).manual_seed(0)
        image = sd_pipe([prompt], generator=generator, num_inference_steps=2, output_type="np").images

        assert image.shape == (1, 128, 128, 3)


@slow
@require_torch_gpu
class StableDiffusionPipelineIntegrationTests(unittest.TestCase):
    def tearDown(self):
        # clean up the VRAM after each test
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    def test_stable_diffusion(self):
        # make sure here that pndm scheduler skips prk
        sd_pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-1", device_map="auto")
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"
        generator = torch.Generator(device=torch_device).manual_seed(0)
        with torch.autocast("cuda"):
            output = sd_pipe(
                [prompt], generator=generator, guidance_scale=6.0, num_inference_steps=20, output_type="np"
            )

        image = output.images

        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.8887, 0.915, 0.91, 0.894, 0.909, 0.912, 0.919, 0.925, 0.883])
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_fast_ddim(self):
        scheduler = DDIMScheduler.from_config("CompVis/stable-diffusion-v1-1", subfolder="scheduler")

        sd_pipe = StableDiffusionPipeline.from_pretrained(
            "CompVis/stable-diffusion-v1-1", scheduler=scheduler, device_map="auto"
        )
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        prompt = "A painting of a squirrel eating a burger"
        generator = torch.Generator(device=torch_device).manual_seed(0)

        with torch.autocast("cuda"):
            output = sd_pipe([prompt], generator=generator, num_inference_steps=2, output_type="numpy")
        image = output.images

        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.9326, 0.923, 0.951, 0.9365, 0.9214, 0.951, 0.9365, 0.9414, 0.918])
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_lms_stable_diffusion_pipeline(self):
        model_id = "CompVis/stable-diffusion-v1-1"
        pipe = StableDiffusionPipeline.from_pretrained(model_id, device_map="auto").to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        scheduler = LMSDiscreteScheduler.from_config(model_id, subfolder="scheduler")
        pipe.scheduler = scheduler

        prompt = "a photograph of an astronaut riding a horse"
        generator = torch.Generator(device=torch_device).manual_seed(0)
        image = pipe(
            [prompt], generator=generator, guidance_scale=7.5, num_inference_steps=10, output_type="numpy"
        ).images

        image_slice = image[0, -3:, -3:, -1]
        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.9077, 0.9254, 0.9181, 0.9227, 0.9213, 0.9367, 0.9399, 0.9406, 0.9024])
        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2

    def test_stable_diffusion_memory_chunking(self):
        torch.cuda.reset_peak_memory_stats()
        model_id = "CompVis/stable-diffusion-v1-4"
        pipe = StableDiffusionPipeline.from_pretrained(
            model_id, revision="fp16", torch_dtype=torch.float16, device_map="auto"
        )
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        prompt = "a photograph of an astronaut riding a horse"

        # make attention efficient
        pipe.enable_attention_slicing()
        generator = torch.Generator(device=torch_device).manual_seed(0)
        with torch.autocast(torch_device):
            output_chunked = pipe(
                [prompt], generator=generator, guidance_scale=7.5, num_inference_steps=10, output_type="numpy"
            )
            image_chunked = output_chunked.images

        mem_bytes = torch.cuda.max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()
        # make sure that less than 3.75 GB is allocated
        assert mem_bytes < 3.75 * 10**9

        # disable chunking
        pipe.disable_attention_slicing()
        generator = torch.Generator(device=torch_device).manual_seed(0)
        with torch.autocast(torch_device):
            output = pipe(
                [prompt], generator=generator, guidance_scale=7.5, num_inference_steps=10, output_type="numpy"
            )
            image = output.images

        # make sure that more than 3.75 GB is allocated
        mem_bytes = torch.cuda.max_memory_allocated()
        assert mem_bytes > 3.75 * 10**9
        assert np.abs(image_chunked.flatten() - image.flatten()).max() < 1e-3

    def test_stable_diffusion_text2img_pipeline_fp16(self):
        torch.cuda.reset_peak_memory_stats()
        model_id = "CompVis/stable-diffusion-v1-4"
        pipe = StableDiffusionPipeline.from_pretrained(
            model_id, revision="fp16", device_map="auto", torch_dtype=torch.float16
        )
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        prompt = "a photograph of an astronaut riding a horse"

        generator = torch.Generator(device=torch_device).manual_seed(0)
        output_chunked = pipe(
            [prompt], generator=generator, guidance_scale=7.5, num_inference_steps=10, output_type="numpy"
        )
        image_chunked = output_chunked.images

        generator = torch.Generator(device=torch_device).manual_seed(0)
        with torch.autocast(torch_device):
            output = pipe(
                [prompt], generator=generator, guidance_scale=7.5, num_inference_steps=10, output_type="numpy"
            )
            image = output.images

        # Make sure results are close enough
        diff = np.abs(image_chunked.flatten() - image.flatten())
        # They ARE different since ops are not run always at the same precision
        # however, they should be extremely close.
        assert diff.mean() < 2e-2

    def test_stable_diffusion_text2img_pipeline(self):
        expected_image = load_image(
            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
            "/text2img/astronaut_riding_a_horse.png"
        )
        expected_image = np.array(expected_image, dtype=np.float32) / 255.0

        model_id = "CompVis/stable-diffusion-v1-4"
        pipe = StableDiffusionPipeline.from_pretrained(
            model_id,
            safety_checker=None,
            device_map="auto",
        )
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        prompt = "astronaut riding a horse"

        generator = torch.Generator(device=torch_device).manual_seed(0)
        output = pipe(prompt=prompt, strength=0.75, guidance_scale=7.5, generator=generator, output_type="np")
        image = output.images[0]

        assert image.shape == (512, 512, 3)
        assert np.abs(expected_image - image).max() < 1e-2

    def test_stable_diffusion_text2img_intermediate_state(self):
        number_of_steps = 0

        def test_callback_fn(step: int, timestep: int, latents: torch.FloatTensor) -> None:
            test_callback_fn.has_been_called = True
            nonlocal number_of_steps
            number_of_steps += 1
            if step == 0:
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 64)
                latents_slice = latents[0, -3:, -3:, -1]
                expected_slice = np.array(
                    [1.8285, 1.2857, -0.1024, 1.2406, -2.3068, 1.0747, -0.0818, -0.6520, -2.9506]
                )
                assert np.abs(latents_slice.flatten() - expected_slice).max() < 1e-3
            elif step == 50:
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 64)
                latents_slice = latents[0, -3:, -3:, -1]
                expected_slice = np.array(
                    [1.1078, 1.5803, 0.2773, -0.0589, -1.7928, -0.3665, -0.4695, -1.0727, -1.1601]
                )
                assert np.abs(latents_slice.flatten() - expected_slice).max() < 1e-2

        test_callback_fn.has_been_called = False

        pipe = StableDiffusionPipeline.from_pretrained(
            "CompVis/stable-diffusion-v1-4", revision="fp16", torch_dtype=torch.float16, device_map="auto"
        )
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        prompt = "Andromeda galaxy in a bottle"

        generator = torch.Generator(device=torch_device).manual_seed(0)
        with torch.autocast(torch_device):
            pipe(
                prompt=prompt,
                num_inference_steps=50,
                guidance_scale=7.5,
                generator=generator,
                callback=test_callback_fn,
                callback_steps=1,
            )
        assert test_callback_fn.has_been_called
        assert number_of_steps == 51

    def test_stable_diffusion_accelerate_auto_device(self):
        pipeline_id = "CompVis/stable-diffusion-v1-4"

        start_time = time.time()
        pipeline_normal_load = StableDiffusionPipeline.from_pretrained(
            pipeline_id, revision="fp16", torch_dtype=torch.float16
        )
        pipeline_normal_load.to(torch_device)
        normal_load_time = time.time() - start_time

        start_time = time.time()
        _ = StableDiffusionPipeline.from_pretrained(
            pipeline_id, revision="fp16", torch_dtype=torch.float16, use_auth_token=True, device_map="auto"
        )
        meta_device_load_time = time.time() - start_time

        assert 2 * meta_device_load_time < normal_load_time

    @unittest.skipIf(torch_device == "cpu", "This test is supposed to run on GPU")
    def test_stable_diffusion_pipeline_with_unet_on_gpu_only(self):
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()

        pipeline_id = "CompVis/stable-diffusion-v1-4"
        prompt = "Andromeda galaxy in a bottle"

        pipeline = StableDiffusionPipeline.from_pretrained(pipeline_id, revision="fp16", torch_dtype=torch.float16)
        pipeline.enable_attention_slicing(1)
        pipeline.enable_sequential_cpu_offload()

        _ = pipeline(prompt, num_inference_steps=5)

        mem_bytes = torch.cuda.max_memory_allocated()
        # make sure that less than 1.5 GB is allocated
        assert mem_bytes < 1.5 * 10**9