tests/pipelines/stable_diffusion_3/test_pipeline_stable_diffusion_3.py

import gc
import unittest

import numpy as np
import torch
from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel

from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, SD3Transformer2DModel, StableDiffusion3Pipeline
from diffusers.utils.testing_utils import (
    numpy_cosine_similarity_distance,
    require_torch_gpu,
    slow,
    torch_device,
)

from ..test_pipelines_common import PipelineTesterMixin


class StableDiffusion3PipelineFastTests(unittest.TestCase, PipelineTesterMixin):
    pipeline_class = StableDiffusion3Pipeline
    params = frozenset(
        [
            "prompt",
            "height",
            "width",
            "guidance_scale",
            "negative_prompt",
            "prompt_embeds",
            "negative_prompt_embeds",
        ]
    )
    batch_params = frozenset(["prompt", "negative_prompt"])

    def get_dummy_components(self):
        torch.manual_seed(0)
        transformer = SD3Transformer2DModel(
            sample_size=32,
            patch_size=1,
            in_channels=4,
            num_layers=1,
            attention_head_dim=8,
            num_attention_heads=4,
            caption_projection_dim=32,
            joint_attention_dim=32,
            pooled_projection_dim=64,
            out_channels=4,
        )
        clip_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,
            hidden_act="gelu",
            projection_dim=32,
        )

        torch.manual_seed(0)
        text_encoder = CLIPTextModelWithProjection(clip_text_encoder_config)

        torch.manual_seed(0)
        text_encoder_2 = CLIPTextModelWithProjection(clip_text_encoder_config)

        text_encoder_3 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")

        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
        tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
        tokenizer_3 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

        torch.manual_seed(0)
        vae = AutoencoderKL(
            sample_size=32,
            in_channels=3,
            out_channels=3,
            block_out_channels=(4,),
            layers_per_block=1,
            latent_channels=4,
            norm_num_groups=1,
            use_quant_conv=False,
            use_post_quant_conv=False,
            shift_factor=0.0609,
            scaling_factor=1.5035,
        )

        scheduler = FlowMatchEulerDiscreteScheduler()

        return {
            "scheduler": scheduler,
            "text_encoder": text_encoder,
            "text_encoder_2": text_encoder_2,
            "text_encoder_3": text_encoder_3,
            "tokenizer": tokenizer,
            "tokenizer_2": tokenizer_2,
            "tokenizer_3": tokenizer_3,
            "transformer": transformer,
            "vae": vae,
        }

    def get_dummy_inputs(self, device, seed=0):
        if str(device).startswith("mps"):
            generator = torch.manual_seed(seed)
        else:
            generator = torch.Generator(device="cpu").manual_seed(seed)

        inputs = {
            "prompt": "A painting of a squirrel eating a burger",
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 5.0,
            "output_type": "np",
        }
        return inputs

    def test_stable_diffusion_3_different_prompts(self):
        pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device)

        inputs = self.get_dummy_inputs(torch_device)
        output_same_prompt = pipe(**inputs).images[0]

        inputs = self.get_dummy_inputs(torch_device)
        inputs["prompt_2"] = "a different prompt"
        inputs["prompt_3"] = "another different prompt"
        output_different_prompts = pipe(**inputs).images[0]

        max_diff = np.abs(output_same_prompt - output_different_prompts).max()

        # Outputs should be different here
        assert max_diff > 1e-2

    def test_stable_diffusion_3_different_negative_prompts(self):
        pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device)

        inputs = self.get_dummy_inputs(torch_device)
        output_same_prompt = pipe(**inputs).images[0]

        inputs = self.get_dummy_inputs(torch_device)
        inputs["negative_prompt_2"] = "deformed"
        inputs["negative_prompt_3"] = "blurry"
        output_different_prompts = pipe(**inputs).images[0]

        max_diff = np.abs(output_same_prompt - output_different_prompts).max()

        # Outputs should be different here
        assert max_diff > 1e-2

    def test_stable_diffusion_3_prompt_embeds(self):
        pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device)
        inputs = self.get_dummy_inputs(torch_device)

        output_with_prompt = pipe(**inputs).images[0]

        inputs = self.get_dummy_inputs(torch_device)
        prompt = inputs.pop("prompt")

        do_classifier_free_guidance = inputs["guidance_scale"] > 1
        (
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
        ) = pipe.encode_prompt(
            prompt,
            prompt_2=None,
            prompt_3=None,
            do_classifier_free_guidance=do_classifier_free_guidance,
            device=torch_device,
        )
        output_with_embeds = pipe(
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            **inputs,
        ).images[0]

        max_diff = np.abs(output_with_prompt - output_with_embeds).max()
        assert max_diff < 1e-4

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

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

        pipe.transformer.fuse_qkv_projections()
        inputs = self.get_dummy_inputs(device)
        image = pipe(**inputs).images
        image_slice_fused = image[0, -3:, -3:, -1]

        pipe.transformer.unfuse_qkv_projections()
        inputs = self.get_dummy_inputs(device)
        image = pipe(**inputs).images
        image_slice_disabled = image[0, -3:, -3:, -1]

        assert np.allclose(
            original_image_slice, image_slice_fused, atol=1e-3, rtol=1e-3
        ), "Fusion of QKV projections shouldn't affect the outputs."
        assert np.allclose(
            image_slice_fused, image_slice_disabled, atol=1e-3, rtol=1e-3
        ), "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled."
        assert np.allclose(
            original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2
        ), "Original outputs should match when fused QKV projections are disabled."


@slow
@require_torch_gpu
class StableDiffusion3PipelineSlowTests(unittest.TestCase):
    pipeline_class = StableDiffusion3Pipeline
    repo_id = "stabilityai/stable-diffusion-3-medium-diffusers"

    def setUp(self):
        super().setUp()
        gc.collect()
        torch.cuda.empty_cache()

    def tearDown(self):
        super().tearDown()
        gc.collect()
        torch.cuda.empty_cache()

    def get_inputs(self, device, seed=0):
        if str(device).startswith("mps"):
            generator = torch.manual_seed(seed)
        else:
            generator = torch.Generator(device="cpu").manual_seed(seed)

        return {
            "prompt": "A photo of a cat",
            "num_inference_steps": 2,
            "guidance_scale": 5.0,
            "output_type": "np",
            "generator": generator,
        }

    def test_sd3_inference(self):
        pipe = self.pipeline_class.from_pretrained(self.repo_id, torch_dtype=torch.float16)
        pipe.enable_model_cpu_offload()

        inputs = self.get_inputs(torch_device)

        image = pipe(**inputs).images[0]
        image_slice = image[0, :10, :10]
        expected_slice = np.array(
            [
                [0.36132812, 0.30004883, 0.25830078],
                [0.36669922, 0.31103516, 0.23754883],
                [0.34814453, 0.29248047, 0.23583984],
                [0.35791016, 0.30981445, 0.23999023],
                [0.36328125, 0.31274414, 0.2607422],
                [0.37304688, 0.32177734, 0.26171875],
                [0.3671875, 0.31933594, 0.25756836],
                [0.36035156, 0.31103516, 0.2578125],
                [0.3857422, 0.33789062, 0.27563477],
                [0.3701172, 0.31982422, 0.265625],
            ],
            dtype=np.float32,
        )

        max_diff = numpy_cosine_similarity_distance(expected_slice.flatten(), image_slice.flatten())

        assert max_diff < 1e-4