Skip to content

make safe diffusion test cases pass on XPU and A100 #11458

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 5 commits into from
Apr 30, 2025
Merged

make safe diffusion test cases pass on XPU and A100 #11458

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
9a5bb0c
make safe diffusion test cases pass on XPU and A100
yao-matrix Apr 30, 2025
38cb85b
Merge branch 'main' into sd-xpu
yao-matrix Apr 30, 2025
27dd473
Merge branch 'main' into sd-xpu
yao-matrix Apr 30, 2025
f06ab6e
calibrate A100 expected values
yao-matrix Apr 30, 2025
779b9a3
Merge branch 'main' into sd-xpu
yao-matrix Apr 30, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
70 changes: 58 additions & 12 deletions tests/pipelines/stable_diffusion_safe/test_safe_diffusion.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,21 +24,29 @@

from diffusers import AutoencoderKL, DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler, UNet2DConditionModel
from diffusers.pipelines.stable_diffusion_safe import StableDiffusionPipelineSafe as StableDiffusionPipeline
from diffusers.utils.testing_utils import floats_tensor, nightly, require_accelerator, require_torch_gpu, torch_device
from diffusers.utils.testing_utils import (
Expectations,
backend_empty_cache,
floats_tensor,
nightly,
require_accelerator,
require_torch_accelerator,
torch_device,
)


class SafeDiffusionPipelineFastTests(unittest.TestCase):
def setUp(self):
# clean up the VRAM before each test
super().setUp()
gc.collect()
torch.cuda.empty_cache()
backend_empty_cache(torch_device)

def tearDown(self):
# clean up the VRAM after each test
super().tearDown()
gc.collect()
torch.cuda.empty_cache()
backend_empty_cache(torch_device)

@property
def dummy_image(self):
Expand Down Expand Up @@ -262,19 +270,19 @@ def test_stable_diffusion_fp16(self):


@nightly
@require_torch_gpu
@require_torch_accelerator
class SafeDiffusionPipelineIntegrationTests(unittest.TestCase):
def setUp(self):
# clean up the VRAM before each test
super().setUp()
gc.collect()
torch.cuda.empty_cache()
backend_empty_cache(torch_device)

def tearDown(self):
# clean up the VRAM after each test
super().tearDown()
gc.collect()
torch.cuda.empty_cache()
backend_empty_cache(torch_device)

def test_harm_safe_stable_diffusion(self):
sd_pipe = StableDiffusionPipeline.from_pretrained(
Expand Down Expand Up @@ -308,7 +316,14 @@ def test_harm_safe_stable_diffusion(self):

image = output.images
image_slice = image[0, -3:, -3:, -1]
expected_slice = [0.2278, 0.2231, 0.2249, 0.2333, 0.2303, 0.1885, 0.2273, 0.2144, 0.2176]
expected_slices = Expectations(
{
("xpu", 3): [0.0076, 0.0058, 0.0012, 0, 0.0047, 0.0046, 0, 0, 0],
("cuda", 7): [0.2278, 0.2231, 0.2249, 0.2333, 0.2303, 0.1885, 0.2273, 0.2144, 0.2176],
("cuda", 8): [0.0076, 0.0058, 0.0012, 0, 0.0047, 0.0046, 0, 0, 0],
}
)
expected_slice = expected_slices.get_expectation()

assert image.shape == (1, 512, 512, 3)

Expand All @@ -335,6 +350,15 @@ def test_harm_safe_stable_diffusion(self):
image_slice = image[0, -3:, -3:, -1]
expected_slice = [0.2383, 0.2276, 0.236, 0.2192, 0.2186, 0.2053, 0.1971, 0.1901, 0.1719]

expected_slices = Expectations(
{
("xpu", 3): [0.0443, 0.0439, 0.0381, 0.0336, 0.0408, 0.0345, 0.0405, 0.0338, 0.0293],
("cuda", 7): [0.2383, 0.2276, 0.236, 0.2192, 0.2186, 0.2053, 0.1971, 0.1901, 0.1719],
("cuda", 8): [0.0443, 0.0439, 0.0381, 0.0336, 0.0408, 0.0345, 0.0405, 0.0338, 0.0293],
}
)
expected_slice = expected_slices.get_expectation()

assert image.shape == (1, 512, 512, 3)

assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
Expand Down Expand Up @@ -365,8 +389,14 @@ def test_nudity_safe_stable_diffusion(self):

image = output.images
image_slice = image[0, -3:, -3:, -1]
expected_slice = [0.3502, 0.3622, 0.3396, 0.3642, 0.3478, 0.3318, 0.35, 0.3348, 0.3297]

expected_slices = Expectations(
{
("xpu", 3): [0.3244, 0.3355, 0.3260, 0.3123, 0.3246, 0.3426, 0.3109, 0.3471, 0.4001],
("cuda", 7): [0.3502, 0.3622, 0.3396, 0.3642, 0.3478, 0.3318, 0.35, 0.3348, 0.3297],
("cuda", 8): [0.3605, 0.3684, 0.3712, 0.3624, 0.3675, 0.3726, 0.3494, 0.3748, 0.4044],
}
)
expected_slice = expected_slices.get_expectation()
assert image.shape == (1, 512, 512, 3)

assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
Expand All @@ -389,7 +419,16 @@ def test_nudity_safe_stable_diffusion(self):

image = output.images
image_slice = image[0, -3:, -3:, -1]
expected_slice = [0.5531, 0.5206, 0.4895, 0.5156, 0.5182, 0.4751, 0.4802, 0.4803, 0.4443]
expected_slices = Expectations(
{
("xpu", 3): [0.6178, 0.6260, 0.6194, 0.6435, 0.6265, 0.6461, 0.6567, 0.6576, 0.6444],
("cuda", 7): [0.5531, 0.5206, 0.4895, 0.5156, 0.5182, 0.4751, 0.4802, 0.4803, 0.4443],
("cuda", 8): [0.5892, 0.5959, 0.5914, 0.6123, 0.5982, 0.6141, 0.6180, 0.6262, 0.6171],
}
)

print(f"image_slice: {image_slice.flatten()}")
expected_slice = expected_slices.get_expectation()

assert image.shape == (1, 512, 512, 3)

Expand Down Expand Up @@ -445,7 +484,14 @@ def test_nudity_safetychecker_safe_stable_diffusion(self):

image = output.images
image_slice = image[0, -3:, -3:, -1]
expected_slice = np.array([0.5818, 0.6285, 0.6835, 0.6019, 0.625, 0.6754, 0.6096, 0.6334, 0.6561])
assert image.shape == (1, 512, 512, 3)
expected_slices = Expectations(
{
("xpu", 3): np.array([0.0695, 0.1244, 0.1831, 0.0527, 0.0444, 0.1660, 0.0572, 0.0677, 0.1551]),
("cuda", 7): np.array([0.5818, 0.6285, 0.6835, 0.6019, 0.625, 0.6754, 0.6096, 0.6334, 0.6561]),
("cuda", 8): np.array([0.0695, 0.1244, 0.1831, 0.0527, 0.0444, 0.1660, 0.0572, 0.0677, 0.1551]),
}
)
expected_slice = expected_slices.get_expectation()

assert image.shape == (1, 512, 512, 3)
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-2
Loading