Thanks to visit codestin.com
Credit goes to github.com

Skip to content

[CI] Refactor Bria Transformer Tests #13341

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
DN6 merged 6 commits into from
Jun 4, 2026
+104 −150
Merged

[CI] Refactor Bria Transformer Tests #13341

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
39f17ec
update
DN6 Mar 26, 2026
1ff4dbf
Merge branch 'main' into bria-test-refactor
sayakpaul Mar 27, 2026
5b6fb1e
update
DN6 Apr 14, 2026
85ffff2
update
DN6 Apr 14, 2026
9106556
Merge branch 'main' into bria-test-refactor
DN6 Jun 4, 2026
faea99d
update
DN6 Jun 4, 2026
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
163 changes: 53 additions & 110 deletions tests/models/transformers/test_models_transformer_bria.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,113 +13,45 @@
# See the License for the specific language governing permissions and
# limitations under the License.

import unittest

import torch

from diffusers import BriaTransformer2DModel
from diffusers.models.attention_processor import FluxIPAdapterJointAttnProcessor2_0
from diffusers.models.embeddings import ImageProjection
from diffusers.utils.torch_utils import randn_tensor

from ...testing_utils import enable_full_determinism, torch_device
from ..test_modeling_common import LoraHotSwappingForModelTesterMixin, ModelTesterMixin, TorchCompileTesterMixin
from ..testing_utils import (
BaseModelTesterConfig,
ModelTesterMixin,
TrainingTesterMixin,
)


enable_full_determinism()


def create_bria_ip_adapter_state_dict(model):
# "ip_adapter" (cross-attention weights)
ip_cross_attn_state_dict = {}
key_id = 0

for name in model.attn_processors.keys():
if name.startswith("single_transformer_blocks"):
continue

joint_attention_dim = model.config["joint_attention_dim"]
hidden_size = model.config["num_attention_heads"] * model.config["attention_head_dim"]
sd = FluxIPAdapterJointAttnProcessor2_0(
hidden_size=hidden_size, cross_attention_dim=joint_attention_dim, scale=1.0
).state_dict()
ip_cross_attn_state_dict.update(
{
f"{key_id}.to_k_ip.weight": sd["to_k_ip.0.weight"],
f"{key_id}.to_v_ip.weight": sd["to_v_ip.0.weight"],
f"{key_id}.to_k_ip.bias": sd["to_k_ip.0.bias"],
f"{key_id}.to_v_ip.bias": sd["to_v_ip.0.bias"],
}
)

key_id += 1

# "image_proj" (ImageProjection layer weights)

image_projection = ImageProjection(
cross_attention_dim=model.config["joint_attention_dim"],
image_embed_dim=model.config["pooled_projection_dim"],
num_image_text_embeds=4,
)

ip_image_projection_state_dict = {}
sd = image_projection.state_dict()
ip_image_projection_state_dict.update(
{
"proj.weight": sd["image_embeds.weight"],
"proj.bias": sd["image_embeds.bias"],
"norm.weight": sd["norm.weight"],
"norm.bias": sd["norm.bias"],
}
)

del sd
ip_state_dict = {}
ip_state_dict.update({"image_proj": ip_image_projection_state_dict, "ip_adapter": ip_cross_attn_state_dict})
return ip_state_dict


class BriaTransformerTests(ModelTesterMixin, unittest.TestCase):
model_class = BriaTransformer2DModel
main_input_name = "hidden_states"
# We override the items here because the transformer under consideration is small.
model_split_percents = [0.8, 0.7, 0.7]

# Skip setting testing with default: AttnProcessor
uses_custom_attn_processor = True

class BriaTransformerTesterConfig(BaseModelTesterConfig):
@property
def dummy_input(self):
batch_size = 1
num_latent_channels = 4
num_image_channels = 3
height = width = 4
sequence_length = 48
embedding_dim = 32
def model_class(self):
return BriaTransformer2DModel

hidden_states = torch.randn((batch_size, height * width, num_latent_channels)).to(torch_device)
encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
text_ids = torch.randn((sequence_length, num_image_channels)).to(torch_device)
image_ids = torch.randn((height * width, num_image_channels)).to(torch_device)
timestep = torch.tensor([1.0]).to(torch_device).expand(batch_size)
@property
def main_input_name(self) -> str:
return "hidden_states"

return {
"hidden_states": hidden_states,
"encoder_hidden_states": encoder_hidden_states,
"img_ids": image_ids,
"txt_ids": text_ids,
"timestep": timestep,
}
@property
def model_split_percents(self) -> list:
return [0.8, 0.7, 0.7]

@property
def input_shape(self):
def output_shape(self) -> tuple:
return (16, 4)

@property
def output_shape(self):
return (16, 4)
def generator(self):
return torch.Generator("cpu").manual_seed(0)

def prepare_init_args_and_inputs_for_common(self):
init_dict = {
def get_init_dict(self) -> dict:
return {
"patch_size": 1,
"in_channels": 4,
"num_layers": 1,
Expand All @@ -131,19 +63,42 @@ def prepare_init_args_and_inputs_for_common(self):
"axes_dims_rope": [0, 4, 4],
}

inputs_dict = self.dummy_input
return init_dict, inputs_dict
def get_dummy_inputs(self, batch_size: int = 1) -> dict[str, torch.Tensor]:
num_latent_channels = 4
num_image_channels = 3
height = width = 4
sequence_length = 48
embedding_dim = 32

return {
"hidden_states": randn_tensor(
(batch_size, height * width, num_latent_channels), generator=self.generator, device=torch_device
),
"encoder_hidden_states": randn_tensor(
(batch_size, sequence_length, embedding_dim), generator=self.generator, device=torch_device
),
"img_ids": randn_tensor(
(height * width, num_image_channels), generator=self.generator, device=torch_device
),
"txt_ids": randn_tensor(
(sequence_length, num_image_channels), generator=self.generator, device=torch_device
),
"timestep": torch.tensor([1.0]).to(torch_device).expand(batch_size),
}


class TestBriaTransformer(BriaTransformerTesterConfig, ModelTesterMixin):
def test_deprecated_inputs_img_txt_ids_3d(self):
init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
init_dict = self.get_init_dict()
inputs_dict = self.get_dummy_inputs()

model = self.model_class(**init_dict)
model.to(torch_device)
model.eval()

with torch.no_grad():
output_1 = model(**inputs_dict).to_tuple()[0]

# update inputs_dict with txt_ids and img_ids as 3d tensors (deprecated)
text_ids_3d = inputs_dict["txt_ids"].unsqueeze(0)
image_ids_3d = inputs_dict["img_ids"].unsqueeze(0)

Expand All @@ -156,26 +111,14 @@ def test_deprecated_inputs_img_txt_ids_3d(self):
with torch.no_grad():
output_2 = model(**inputs_dict).to_tuple()[0]

self.assertEqual(output_1.shape, output_2.shape)
self.assertTrue(
torch.allclose(output_1, output_2, atol=1e-5),
msg="output with deprecated inputs (img_ids and txt_ids as 3d torch tensors) are not equal as them as 2d inputs",
assert output_1.shape == output_2.shape
assert torch.allclose(output_1, output_2, atol=1e-5), (
"output with deprecated inputs (img_ids and txt_ids as 3d torch tensors) "
"are not equal as them as 2d inputs"
)


class TestBriaTransformerTraining(BriaTransformerTesterConfig, TrainingTesterMixin):
def test_gradient_checkpointing_is_applied(self):
expected_set = {"BriaTransformer2DModel"}
super().test_gradient_checkpointing_is_applied(expected_set=expected_set)


class BriaTransformerCompileTests(TorchCompileTesterMixin, unittest.TestCase):
model_class = BriaTransformer2DModel

def prepare_init_args_and_inputs_for_common(self):
return BriaTransformerTests().prepare_init_args_and_inputs_for_common()


class BriaTransformerLoRAHotSwapTests(LoraHotSwappingForModelTesterMixin, unittest.TestCase):
model_class = BriaTransformer2DModel

def prepare_init_args_and_inputs_for_common(self):
return BriaTransformerTests().prepare_init_args_and_inputs_for_common()
91 changes: 51 additions & 40 deletions tests/models/transformers/test_models_transformer_bria_fibo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,62 +13,45 @@
# See the License for the specific language governing permissions and
# limitations under the License.

import unittest

import torch

from diffusers import BriaFiboTransformer2DModel
from diffusers.utils.torch_utils import randn_tensor

from ...testing_utils import enable_full_determinism, torch_device
from ..test_modeling_common import ModelTesterMixin
from ..testing_utils import (
BaseModelTesterConfig,
ModelTesterMixin,
TrainingTesterMixin,
)


enable_full_determinism()


class BriaFiboTransformerTests(ModelTesterMixin, unittest.TestCase):
model_class = BriaFiboTransformer2DModel
main_input_name = "hidden_states"
# We override the items here because the transformer under consideration is small.
model_split_percents = [0.8, 0.7, 0.7]

# Skip setting testing with default: AttnProcessor
uses_custom_attn_processor = True

class BriaFiboTransformerTesterConfig(BaseModelTesterConfig):
@property
def dummy_input(self):
batch_size = 1
num_latent_channels = 48
num_image_channels = 3
height = width = 16
sequence_length = 32
embedding_dim = 64
def model_class(self):
return BriaFiboTransformer2DModel

hidden_states = torch.randn((batch_size, height * width, num_latent_channels)).to(torch_device)
encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
text_ids = torch.randn((sequence_length, num_image_channels)).to(torch_device)
image_ids = torch.randn((height * width, num_image_channels)).to(torch_device)
timestep = torch.tensor([1.0]).to(torch_device).expand(batch_size)

return {
"hidden_states": hidden_states,
"encoder_hidden_states": encoder_hidden_states,
"img_ids": image_ids,
"txt_ids": text_ids,
"timestep": timestep,
"text_encoder_layers": [encoder_hidden_states[:, :, :32], encoder_hidden_states[:, :, :32]],
}
@property
def main_input_name(self) -> str:
return "hidden_states"

@property
def input_shape(self):
return (16, 16)
def model_split_percents(self) -> list:
return [0.8, 0.7, 0.7]

@property
def output_shape(self):
def output_shape(self) -> tuple:
return (256, 48)

def prepare_init_args_and_inputs_for_common(self):
init_dict = {
@property
def generator(self):
return torch.Generator("cpu").manual_seed(0)

def get_init_dict(self) -> dict:
return {
"patch_size": 1,
"in_channels": 48,
"num_layers": 1,
Expand All @@ -81,9 +64,37 @@ def prepare_init_args_and_inputs_for_common(self):
"axes_dims_rope": [0, 4, 4],
}

inputs_dict = self.dummy_input
return init_dict, inputs_dict
def get_dummy_inputs(self, batch_size: int = 1) -> dict[str, torch.Tensor]:
num_latent_channels = 48
num_image_channels = 3
height = width = 16
sequence_length = 32
embedding_dim = 64

encoder_hidden_states = randn_tensor(
(batch_size, sequence_length, embedding_dim), generator=self.generator, device=torch_device
)
return {
"hidden_states": randn_tensor(
(batch_size, height * width, num_latent_channels), generator=self.generator, device=torch_device
),
"encoder_hidden_states": encoder_hidden_states,
"img_ids": randn_tensor(
(height * width, num_image_channels), generator=self.generator, device=torch_device
),
"txt_ids": randn_tensor(
(sequence_length, num_image_channels), generator=self.generator, device=torch_device
),
"timestep": torch.tensor([1.0]).to(torch_device).expand(batch_size),
"text_encoder_layers": [encoder_hidden_states[:, :, :32], encoder_hidden_states[:, :, :32]],
}


class TestBriaFiboTransformer(BriaFiboTransformerTesterConfig, ModelTesterMixin):
pass


class TestBriaFiboTransformerTraining(BriaFiboTransformerTesterConfig, TrainingTesterMixin):
def test_gradient_checkpointing_is_applied(self):
expected_set = {"BriaFiboTransformer2DModel"}
super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
Loading