test_stable_diffusion.py

# coding=utf-8
# Copyright 2023 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 unittest

import numpy as np
import torch
from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer

from diffusers import (
    AutoencoderKL,
    DDIMScheduler,
    DPMSolverMultistepScheduler,
    EulerAncestralDiscreteScheduler,
    EulerDiscreteScheduler,
    LMSDiscreteScheduler,
    PNDMScheduler,
    StableDiffusionPipeline,
    UNet2DConditionModel,
    logging,
)
from diffusers.utils import load_numpy, nightly, slow, torch_device
from diffusers.utils.testing_utils import CaptureLogger, require_torch_gpu

from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
from ..test_pipelines_common import PipelineLatentTesterMixin, PipelineTesterMixin


torch.backends.cuda.matmul.allow_tf32 = False


class StableDiffusion2PipelineFastTests(PipelineLatentTesterMixin, PipelineTesterMixin, unittest.TestCase):
    pipeline_class = StableDiffusionPipeline
    params = TEXT_TO_IMAGE_PARAMS
    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS

    def get_dummy_components(self):
        torch.manual_seed(0)
        unet = 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,
            # SD2-specific config below
            attention_head_dim=(2, 4),
            use_linear_projection=True,
        )
        scheduler = DDIMScheduler(
            beta_start=0.00085,
            beta_end=0.012,
            beta_schedule="scaled_linear",
            clip_sample=False,
            set_alpha_to_one=False,
        )
        torch.manual_seed(0)
        vae = 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,
            sample_size=128,
        )
        torch.manual_seed(0)
        text_encoder_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,
            # SD2-specific config below
            hidden_act="gelu",
            projection_dim=512,
        )
        text_encoder = CLIPTextModel(text_encoder_config)
        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")

        components = {
            "unet": unet,
            "scheduler": scheduler,
            "vae": vae,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
            "safety_checker": None,
            "feature_extractor": None,
        }
        return components

    def get_dummy_inputs(self, device, seed=0):
        if str(device).startswith("mps"):
            generator = torch.manual_seed(seed)
        else:
            generator = torch.Generator(device=device).manual_seed(seed)
        inputs = {
            "prompt": "A painting of a squirrel eating a burger",
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 6.0,
            "output_type": "numpy",
        }
        return inputs

    def test_stable_diffusion_ddim(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
        expected_slice = np.array([0.5753, 0.6113, 0.5005, 0.5036, 0.5464, 0.4725, 0.4982, 0.4865, 0.4861])

        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
        components = self.get_dummy_components()
        components["scheduler"] = PNDMScheduler(skip_prk_steps=True)
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
        expected_slice = np.array([0.5121, 0.5714, 0.4827, 0.5057, 0.5646, 0.4766, 0.5189, 0.4895, 0.4990])

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

    def test_stable_diffusion_k_lms(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        components["scheduler"] = LMSDiscreteScheduler.from_config(components["scheduler"].config)
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
        expected_slice = np.array([0.4865, 0.5439, 0.4840, 0.4995, 0.5543, 0.4846, 0.5199, 0.4942, 0.5061])

        assert np.abs(image_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
        components = self.get_dummy_components()
        components["scheduler"] = EulerAncestralDiscreteScheduler.from_config(components["scheduler"].config)
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
        expected_slice = np.array([0.4864, 0.5440, 0.4842, 0.4994, 0.5543, 0.4846, 0.5196, 0.4942, 0.5063])

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

    def test_stable_diffusion_k_euler(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()
        components["scheduler"] = EulerDiscreteScheduler.from_config(components["scheduler"].config)
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        image = sd_pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1]

        assert image.shape == (1, 64, 64, 3)
        expected_slice = np.array([0.4865, 0.5439, 0.4840, 0.4995, 0.5543, 0.4846, 0.5199, 0.4942, 0.5061])

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

    def test_stable_diffusion_long_prompt(self):
        components = self.get_dummy_components()
        components["scheduler"] = LMSDiscreteScheduler.from_config(components["scheduler"].config)
        sd_pipe = StableDiffusionPipeline(**components)
        sd_pipe = sd_pipe.to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        do_classifier_free_guidance = True
        negative_prompt = None
        num_images_per_prompt = 1
        logger = logging.get_logger("diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion")

        prompt = 25 * "@"
        with CaptureLogger(logger) as cap_logger_3:
            text_embeddings_3 = sd_pipe._encode_prompt(
                prompt, torch_device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
            )

        prompt = 100 * "@"
        with CaptureLogger(logger) as cap_logger:
            text_embeddings = sd_pipe._encode_prompt(
                prompt, torch_device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
            )

        negative_prompt = "Hello"
        with CaptureLogger(logger) as cap_logger_2:
            text_embeddings_2 = sd_pipe._encode_prompt(
                prompt, torch_device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
            )

        assert text_embeddings_3.shape == text_embeddings_2.shape == text_embeddings.shape
        assert text_embeddings.shape[1] == 77

        assert cap_logger.out == cap_logger_2.out
        # 100 - 77 + 1 (BOS token) + 1 (EOS token) = 25
        assert cap_logger.out.count("@") == 25
        assert cap_logger_3.out == ""


@slow
@require_torch_gpu
class StableDiffusion2PipelineSlowTests(unittest.TestCase):
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
        latents = np.random.RandomState(seed).standard_normal((1, 4, 64, 64))
        latents = torch.from_numpy(latents).to(device=device, dtype=dtype)
        inputs = {
            "prompt": "a photograph of an astronaut riding a horse",
            "latents": latents,
            "generator": generator,
            "num_inference_steps": 3,
            "guidance_scale": 7.5,
            "output_type": "numpy",
        }
        return inputs

    def test_stable_diffusion_default_ddim(self):
        pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-base")
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.49493, 0.47896, 0.40798, 0.54214, 0.53212, 0.48202, 0.47656, 0.46329, 0.48506])
        assert np.abs(image_slice - expected_slice).max() < 1e-4

    def test_stable_diffusion_pndm(self):
        pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-base")
        pipe.scheduler = PNDMScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.49493, 0.47896, 0.40798, 0.54214, 0.53212, 0.48202, 0.47656, 0.46329, 0.48506])
        assert np.abs(image_slice - expected_slice).max() < 1e-4

    def test_stable_diffusion_k_lms(self):
        pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-base")
        pipe.scheduler = LMSDiscreteScheduler.from_config(pipe.scheduler.config)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = pipe(**inputs).images
        image_slice = image[0, -3:, -3:, -1].flatten()

        assert image.shape == (1, 512, 512, 3)
        expected_slice = np.array([0.10440, 0.13115, 0.11100, 0.10141, 0.11440, 0.07215, 0.11332, 0.09693, 0.10006])
        assert np.abs(image_slice - expected_slice).max() < 1e-4

    def test_stable_diffusion_attention_slicing(self):
        torch.cuda.reset_peak_memory_stats()
        pipe = StableDiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-2-base", torch_dtype=torch.float16
        )
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        # enable attention slicing
        pipe.enable_attention_slicing()
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        image_sliced = pipe(**inputs).images

        mem_bytes = torch.cuda.max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()
        # make sure that less than 3.3 GB is allocated
        assert mem_bytes < 3.3 * 10**9

        # disable slicing
        pipe.disable_attention_slicing()
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        image = pipe(**inputs).images

        # make sure that more than 3.3 GB is allocated
        mem_bytes = torch.cuda.max_memory_allocated()
        assert mem_bytes > 3.3 * 10**9
        assert np.abs(image_sliced - image).max() < 1e-3

    def test_stable_diffusion_text2img_intermediate_state(self):
        number_of_steps = 0

        def callback_fn(step: int, timestep: int, latents: torch.FloatTensor) -> None:
            callback_fn.has_been_called = True
            nonlocal number_of_steps
            number_of_steps += 1
            if step == 1:
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 64)
                latents_slice = latents[0, -3:, -3:, -1]
                expected_slice = np.array(
                    [-0.3862, -0.4507, -1.1729, 0.0686, -1.1045, 0.7124, -1.8301, 0.1903, 1.2773]
                )

                assert np.abs(latents_slice.flatten() - expected_slice).max() < 5e-2
            elif step == 2:
                latents = latents.detach().cpu().numpy()
                assert latents.shape == (1, 4, 64, 64)
                latents_slice = latents[0, -3:, -3:, -1]
                expected_slice = np.array(
                    [0.2720, -0.1863, -0.7383, -0.5029, -0.7534, 0.3970, -0.7646, 0.4468, 1.2686]
                )

                assert np.abs(latents_slice.flatten() - expected_slice).max() < 5e-2

        callback_fn.has_been_called = False

        pipe = StableDiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-2-base", torch_dtype=torch.float16
        )
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing()

        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        pipe(**inputs, callback=callback_fn, callback_steps=1)
        assert callback_fn.has_been_called
        assert number_of_steps == inputs["num_inference_steps"]

    def test_stable_diffusion_pipeline_with_sequential_cpu_offloading(self):
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

        pipe = StableDiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-2-base", torch_dtype=torch.float16
        )
        pipe = pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        pipe.enable_attention_slicing(1)
        pipe.enable_sequential_cpu_offload()

        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        _ = pipe(**inputs)

        mem_bytes = torch.cuda.max_memory_allocated()
        # make sure that less than 2.8 GB is allocated
        assert mem_bytes < 2.8 * 10**9

    def test_stable_diffusion_pipeline_with_model_offloading(self):
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

        inputs = self.get_inputs(torch_device, dtype=torch.float16)

        # Normal inference

        pipe = StableDiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-2-base",
            torch_dtype=torch.float16,
        )
        pipe.unet.set_default_attn_processor()
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        outputs = pipe(**inputs)
        mem_bytes = torch.cuda.max_memory_allocated()

        # With model offloading

        # Reload but don't move to cuda
        pipe = StableDiffusionPipeline.from_pretrained(
            "stabilityai/stable-diffusion-2-base",
            torch_dtype=torch.float16,
        )
        pipe.unet.set_default_attn_processor()

        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

        pipe.enable_model_cpu_offload()
        pipe.set_progress_bar_config(disable=None)
        inputs = self.get_inputs(torch_device, dtype=torch.float16)
        outputs_offloaded = pipe(**inputs)
        mem_bytes_offloaded = torch.cuda.max_memory_allocated()

        assert np.abs(outputs.images - outputs_offloaded.images).max() < 1e-3
        assert mem_bytes_offloaded < mem_bytes
        assert mem_bytes_offloaded < 3 * 10**9
        for module in pipe.text_encoder, pipe.unet, pipe.vae:
            assert module.device == torch.device("cpu")

        # With attention slicing
        torch.cuda.empty_cache()
        torch.cuda.reset_max_memory_allocated()
        torch.cuda.reset_peak_memory_stats()

        pipe.enable_attention_slicing()
        _ = pipe(**inputs)
        mem_bytes_slicing = torch.cuda.max_memory_allocated()
        assert mem_bytes_slicing < mem_bytes_offloaded


@nightly
@require_torch_gpu
class StableDiffusion2PipelineNightlyTests(unittest.TestCase):
    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
        generator = torch.Generator(device=generator_device).manual_seed(seed)
        latents = np.random.RandomState(seed).standard_normal((1, 4, 64, 64))
        latents = torch.from_numpy(latents).to(device=device, dtype=dtype)
        inputs = {
            "prompt": "a photograph of an astronaut riding a horse",
            "latents": latents,
            "generator": generator,
            "num_inference_steps": 50,
            "guidance_scale": 7.5,
            "output_type": "numpy",
        }
        return inputs

    def test_stable_diffusion_2_0_default_ddim(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-base").to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_2_text2img/stable_diffusion_2_0_base_ddim.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_stable_diffusion_2_1_default_pndm(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base").to(torch_device)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_2_text2img/stable_diffusion_2_1_base_pndm.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_stable_diffusion_ddim(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base").to(torch_device)
        sd_pipe.scheduler = DDIMScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_2_text2img/stable_diffusion_2_1_base_ddim.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_stable_diffusion_lms(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base").to(torch_device)
        sd_pipe.scheduler = LMSDiscreteScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_2_text2img/stable_diffusion_2_1_base_lms.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_stable_diffusion_euler(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base").to(torch_device)
        sd_pipe.scheduler = EulerDiscreteScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_2_text2img/stable_diffusion_2_1_base_euler.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3

    def test_stable_diffusion_dpm(self):
        sd_pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base").to(torch_device)
        sd_pipe.scheduler = DPMSolverMultistepScheduler.from_config(sd_pipe.scheduler.config)
        sd_pipe.set_progress_bar_config(disable=None)

        inputs = self.get_inputs(torch_device)
        inputs["num_inference_steps"] = 25
        image = sd_pipe(**inputs).images[0]

        expected_image = load_numpy(
            "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main"
            "/stable_diffusion_2_text2img/stable_diffusion_2_1_base_dpm_multi.npy"
        )
        max_diff = np.abs(expected_image - image).max()
        assert max_diff < 1e-3