From f7405f2b447c159685981b1b476326afb9069d99 Mon Sep 17 00:00:00 2001 From: DN6 Date: Thu, 26 Mar 2026 16:41:25 +0530 Subject: [PATCH 1/2] update --- .../test_models_transformer_cogvideox.py | 162 +++++++++++------- .../test_models_transformer_cogview3plus.py | 93 +++++----- .../test_models_transformer_cogview4.py | 87 ++++++---- 3 files changed, 201 insertions(+), 141 deletions(-) diff --git a/tests/models/transformers/test_models_transformer_cogvideox.py b/tests/models/transformers/test_models_transformer_cogvideox.py index f632add7e5a7..c5d872269f89 100644 --- a/tests/models/transformers/test_models_transformer_cogvideox.py +++ b/tests/models/transformers/test_models_transformer_cogvideox.py @@ -13,59 +13,53 @@ # See the License for the specific language governing permissions and # limitations under the License. -import unittest - import torch from diffusers import CogVideoXTransformer3DModel - -from ...testing_utils import ( - enable_full_determinism, - torch_device, +from diffusers.utils.torch_utils import randn_tensor + +from ...testing_utils import enable_full_determinism, torch_device +from ..testing_utils import ( + BaseModelTesterConfig, + ModelTesterMixin, + TorchCompileTesterMixin, + TrainingTesterMixin, ) -from ..test_modeling_common import ModelTesterMixin enable_full_determinism() -class CogVideoXTransformerTests(ModelTesterMixin, unittest.TestCase): - model_class = CogVideoXTransformer3DModel - main_input_name = "hidden_states" - uses_custom_attn_processor = True - model_split_percents = [0.7, 0.7, 0.8] +# ======================== CogVideoX ======================== + +class CogVideoXTransformerTesterConfig(BaseModelTesterConfig): @property - def dummy_input(self): - batch_size = 2 - num_channels = 4 - num_frames = 1 - height = 8 - width = 8 - embedding_dim = 8 - sequence_length = 8 + def model_class(self): + return CogVideoXTransformer3DModel - hidden_states = torch.randn((batch_size, num_frames, num_channels, height, width)).to(torch_device) - encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device) - timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device) + @property + def main_input_name(self) -> str: + return "hidden_states" - return { - "hidden_states": hidden_states, - "encoder_hidden_states": encoder_hidden_states, - "timestep": timestep, - } + @property + def model_split_percents(self) -> list: + return [0.7, 0.7, 0.8] @property - def input_shape(self): + def output_shape(self) -> tuple: return (1, 4, 8, 8) @property - def output_shape(self): + def input_shape(self) -> tuple: return (1, 4, 8, 8) - def prepare_init_args_and_inputs_for_common(self): - init_dict = { - # Product of num_attention_heads * attention_head_dim must be divisible by 16 for 3D positional embeddings. + @property + def generator(self): + return torch.Generator("cpu").manual_seed(0) + + def get_init_dict(self) -> dict: + return { "num_attention_heads": 2, "attention_head_dim": 8, "in_channels": 4, @@ -81,50 +75,66 @@ def prepare_init_args_and_inputs_for_common(self): "temporal_compression_ratio": 4, "max_text_seq_length": 8, } - inputs_dict = self.dummy_input - return init_dict, inputs_dict - def test_gradient_checkpointing_is_applied(self): - expected_set = {"CogVideoXTransformer3DModel"} - super().test_gradient_checkpointing_is_applied(expected_set=expected_set) - - -class CogVideoX1_5TransformerTests(ModelTesterMixin, unittest.TestCase): - model_class = CogVideoXTransformer3DModel - main_input_name = "hidden_states" - uses_custom_attn_processor = True - - @property - def dummy_input(self): - batch_size = 2 + def get_dummy_inputs(self, batch_size: int = 2) -> dict[str, torch.Tensor]: num_channels = 4 - num_frames = 2 + num_frames = 1 height = 8 width = 8 embedding_dim = 8 sequence_length = 8 - hidden_states = torch.randn((batch_size, num_frames, num_channels, height, width)).to(torch_device) - encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device) - timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device) - return { - "hidden_states": hidden_states, - "encoder_hidden_states": encoder_hidden_states, - "timestep": timestep, + "hidden_states": randn_tensor( + (batch_size, num_frames, num_channels, height, width), generator=self.generator, device=torch_device + ), + "encoder_hidden_states": randn_tensor( + (batch_size, sequence_length, embedding_dim), generator=self.generator, device=torch_device + ), + "timestep": torch.randint(0, 1000, size=(batch_size,), generator=self.generator).to(torch_device), } + +class TestCogVideoXTransformer(CogVideoXTransformerTesterConfig, ModelTesterMixin): + pass + + +class TestCogVideoXTransformerTraining(CogVideoXTransformerTesterConfig, TrainingTesterMixin): + def test_gradient_checkpointing_is_applied(self): + expected_set = {"CogVideoXTransformer3DModel"} + super().test_gradient_checkpointing_is_applied(expected_set=expected_set) + + +class TestCogVideoXTransformerCompile(CogVideoXTransformerTesterConfig, TorchCompileTesterMixin): + pass + + +# ======================== CogVideoX 1.5 ======================== + + +class CogVideoX15TransformerTesterConfig(BaseModelTesterConfig): @property - def input_shape(self): + def model_class(self): + return CogVideoXTransformer3DModel + + @property + def main_input_name(self) -> str: + return "hidden_states" + + @property + def output_shape(self) -> tuple: return (1, 4, 8, 8) @property - def output_shape(self): + def input_shape(self) -> tuple: return (1, 4, 8, 8) - def prepare_init_args_and_inputs_for_common(self): - init_dict = { - # Product of num_attention_heads * attention_head_dim must be divisible by 16 for 3D positional embeddings. + @property + def generator(self): + return torch.Generator("cpu").manual_seed(0) + + def get_init_dict(self) -> dict: + return { "num_attention_heads": 2, "attention_head_dim": 8, "in_channels": 4, @@ -141,9 +151,29 @@ def prepare_init_args_and_inputs_for_common(self): "max_text_seq_length": 8, "use_rotary_positional_embeddings": True, } - inputs_dict = self.dummy_input - return init_dict, inputs_dict - def test_gradient_checkpointing_is_applied(self): - expected_set = {"CogVideoXTransformer3DModel"} - super().test_gradient_checkpointing_is_applied(expected_set=expected_set) + def get_dummy_inputs(self, batch_size: int = 2) -> dict[str, torch.Tensor]: + num_channels = 4 + num_frames = 2 + height = 8 + width = 8 + embedding_dim = 8 + sequence_length = 8 + + return { + "hidden_states": randn_tensor( + (batch_size, num_frames, num_channels, height, width), generator=self.generator, device=torch_device + ), + "encoder_hidden_states": randn_tensor( + (batch_size, sequence_length, embedding_dim), generator=self.generator, device=torch_device + ), + "timestep": torch.randint(0, 1000, size=(batch_size,), generator=self.generator).to(torch_device), + } + + +class TestCogVideoX15Transformer(CogVideoX15TransformerTesterConfig, ModelTesterMixin): + pass + + +class TestCogVideoX15TransformerCompile(CogVideoX15TransformerTesterConfig, TorchCompileTesterMixin): + pass diff --git a/tests/models/transformers/test_models_transformer_cogview3plus.py b/tests/models/transformers/test_models_transformer_cogview3plus.py index d38d77531d4c..e738215323e5 100644 --- a/tests/models/transformers/test_models_transformer_cogview3plus.py +++ b/tests/models/transformers/test_models_transformer_cogview3plus.py @@ -13,63 +13,50 @@ # See the License for the specific language governing permissions and # limitations under the License. -import unittest - import torch from diffusers import CogView3PlusTransformer2DModel +from diffusers.utils.torch_utils import randn_tensor -from ...testing_utils import ( - enable_full_determinism, - torch_device, +from ...testing_utils import enable_full_determinism, torch_device +from ..testing_utils import ( + BaseModelTesterConfig, + ModelTesterMixin, + TorchCompileTesterMixin, + TrainingTesterMixin, ) -from ..test_modeling_common import ModelTesterMixin enable_full_determinism() -class CogView3PlusTransformerTests(ModelTesterMixin, unittest.TestCase): - model_class = CogView3PlusTransformer2DModel - main_input_name = "hidden_states" - uses_custom_attn_processor = True - model_split_percents = [0.7, 0.6, 0.6] - +class CogView3PlusTransformerTesterConfig(BaseModelTesterConfig): @property - def dummy_input(self): - batch_size = 2 - num_channels = 4 - height = 8 - width = 8 - embedding_dim = 8 - sequence_length = 8 + def model_class(self): + return CogView3PlusTransformer2DModel - hidden_states = torch.randn((batch_size, num_channels, height, width)).to(torch_device) - encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device) - original_size = torch.tensor([height * 8, width * 8]).unsqueeze(0).repeat(batch_size, 1).to(torch_device) - target_size = torch.tensor([height * 8, width * 8]).unsqueeze(0).repeat(batch_size, 1).to(torch_device) - crop_coords = torch.tensor([0, 0]).unsqueeze(0).repeat(batch_size, 1).to(torch_device) - timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device) + @property + def main_input_name(self) -> str: + return "hidden_states" - return { - "hidden_states": hidden_states, - "encoder_hidden_states": encoder_hidden_states, - "original_size": original_size, - "target_size": target_size, - "crop_coords": crop_coords, - "timestep": timestep, - } + @property + def model_split_percents(self) -> list: + return [0.7, 0.6, 0.6] @property - def input_shape(self): + def output_shape(self) -> tuple: return (1, 4, 8, 8) @property - def output_shape(self): + def input_shape(self) -> tuple: return (1, 4, 8, 8) - 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": 2, "in_channels": 4, "num_layers": 2, @@ -82,9 +69,37 @@ def prepare_init_args_and_inputs_for_common(self): "pos_embed_max_size": 8, "sample_size": 8, } - inputs_dict = self.dummy_input - return init_dict, inputs_dict + def get_dummy_inputs(self, batch_size: int = 2) -> dict[str, torch.Tensor]: + num_channels = 4 + height = 8 + width = 8 + embedding_dim = 8 + sequence_length = 8 + + return { + "hidden_states": randn_tensor( + (batch_size, num_channels, height, width), generator=self.generator, device=torch_device + ), + "encoder_hidden_states": randn_tensor( + (batch_size, sequence_length, embedding_dim), generator=self.generator, device=torch_device + ), + "original_size": torch.tensor([height * 8, width * 8]).unsqueeze(0).repeat(batch_size, 1).to(torch_device), + "target_size": torch.tensor([height * 8, width * 8]).unsqueeze(0).repeat(batch_size, 1).to(torch_device), + "crop_coords": torch.tensor([0, 0]).unsqueeze(0).repeat(batch_size, 1).to(torch_device), + "timestep": torch.randint(0, 1000, size=(batch_size,), generator=self.generator).to(torch_device), + } + + +class TestCogView3PlusTransformer(CogView3PlusTransformerTesterConfig, ModelTesterMixin): + pass + + +class TestCogView3PlusTransformerTraining(CogView3PlusTransformerTesterConfig, TrainingTesterMixin): def test_gradient_checkpointing_is_applied(self): expected_set = {"CogView3PlusTransformer2DModel"} super().test_gradient_checkpointing_is_applied(expected_set=expected_set) + + +class TestCogView3PlusTransformerCompile(CogView3PlusTransformerTesterConfig, TorchCompileTesterMixin): + pass diff --git a/tests/models/transformers/test_models_transformer_cogview4.py b/tests/models/transformers/test_models_transformer_cogview4.py index 084c3b7cea41..51c6dc06eac8 100644 --- a/tests/models/transformers/test_models_transformer_cogview4.py +++ b/tests/models/transformers/test_models_transformer_cogview4.py @@ -12,59 +12,46 @@ # See the License for the specific language governing permissions and # limitations under the License. -import unittest - import torch from diffusers import CogView4Transformer2DModel +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, + TorchCompileTesterMixin, + TrainingTesterMixin, +) enable_full_determinism() -class CogView3PlusTransformerTests(ModelTesterMixin, unittest.TestCase): - model_class = CogView4Transformer2DModel - main_input_name = "hidden_states" - uses_custom_attn_processor = True - +class CogView4TransformerTesterConfig(BaseModelTesterConfig): @property - def dummy_input(self): - batch_size = 2 - num_channels = 4 - height = 8 - width = 8 - embedding_dim = 8 - sequence_length = 8 + def model_class(self): + return CogView4Transformer2DModel - hidden_states = torch.randn((batch_size, num_channels, height, width)).to(torch_device) - encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device) - original_size = torch.tensor([height * 8, width * 8]).unsqueeze(0).repeat(batch_size, 1).to(torch_device) - target_size = torch.tensor([height * 8, width * 8]).unsqueeze(0).repeat(batch_size, 1).to(torch_device) - crop_coords = torch.tensor([0, 0]).unsqueeze(0).repeat(batch_size, 1).to(torch_device) - timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device) - - return { - "hidden_states": hidden_states, - "encoder_hidden_states": encoder_hidden_states, - "timestep": timestep, - "original_size": original_size, - "target_size": target_size, - "crop_coords": crop_coords, - } + @property + def main_input_name(self) -> str: + return "hidden_states" @property - def input_shape(self): + def output_shape(self) -> tuple: return (4, 8, 8) @property - def output_shape(self): + def input_shape(self) -> tuple: return (4, 8, 8) - 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": 2, "in_channels": 4, "num_layers": 2, @@ -75,9 +62,37 @@ def prepare_init_args_and_inputs_for_common(self): "time_embed_dim": 8, "condition_dim": 4, } - inputs_dict = self.dummy_input - return init_dict, inputs_dict + def get_dummy_inputs(self, batch_size: int = 2) -> dict[str, torch.Tensor]: + num_channels = 4 + height = 8 + width = 8 + embedding_dim = 8 + sequence_length = 8 + + return { + "hidden_states": randn_tensor( + (batch_size, num_channels, height, width), generator=self.generator, device=torch_device + ), + "encoder_hidden_states": randn_tensor( + (batch_size, sequence_length, embedding_dim), generator=self.generator, device=torch_device + ), + "timestep": torch.randint(0, 1000, size=(batch_size,), generator=self.generator).to(torch_device), + "original_size": torch.tensor([height * 8, width * 8]).unsqueeze(0).repeat(batch_size, 1).to(torch_device), + "target_size": torch.tensor([height * 8, width * 8]).unsqueeze(0).repeat(batch_size, 1).to(torch_device), + "crop_coords": torch.tensor([0, 0]).unsqueeze(0).repeat(batch_size, 1).to(torch_device), + } + + +class TestCogView4Transformer(CogView4TransformerTesterConfig, ModelTesterMixin): + pass + + +class TestCogView4TransformerTraining(CogView4TransformerTesterConfig, TrainingTesterMixin): def test_gradient_checkpointing_is_applied(self): expected_set = {"CogView4Transformer2DModel"} super().test_gradient_checkpointing_is_applied(expected_set=expected_set) + + +class TestCogView4TransformerCompile(CogView4TransformerTesterConfig, TorchCompileTesterMixin): + pass From a03305c0d6bbc6a51d93741e7df2f107d489e425 Mon Sep 17 00:00:00 2001 From: DN6 Date: Tue, 14 Apr 2026 16:11:55 +0530 Subject: [PATCH 2/2] update --- .../transformers/test_models_transformer_cogvideox.py | 9 --------- .../transformers/test_models_transformer_cogview3plus.py | 5 ----- .../transformers/test_models_transformer_cogview4.py | 5 ----- 3 files changed, 19 deletions(-) diff --git a/tests/models/transformers/test_models_transformer_cogvideox.py b/tests/models/transformers/test_models_transformer_cogvideox.py index c5d872269f89..d166765f403a 100644 --- a/tests/models/transformers/test_models_transformer_cogvideox.py +++ b/tests/models/transformers/test_models_transformer_cogvideox.py @@ -22,7 +22,6 @@ from ..testing_utils import ( BaseModelTesterConfig, ModelTesterMixin, - TorchCompileTesterMixin, TrainingTesterMixin, ) @@ -105,10 +104,6 @@ def test_gradient_checkpointing_is_applied(self): super().test_gradient_checkpointing_is_applied(expected_set=expected_set) -class TestCogVideoXTransformerCompile(CogVideoXTransformerTesterConfig, TorchCompileTesterMixin): - pass - - # ======================== CogVideoX 1.5 ======================== @@ -173,7 +168,3 @@ def get_dummy_inputs(self, batch_size: int = 2) -> dict[str, torch.Tensor]: class TestCogVideoX15Transformer(CogVideoX15TransformerTesterConfig, ModelTesterMixin): pass - - -class TestCogVideoX15TransformerCompile(CogVideoX15TransformerTesterConfig, TorchCompileTesterMixin): - pass diff --git a/tests/models/transformers/test_models_transformer_cogview3plus.py b/tests/models/transformers/test_models_transformer_cogview3plus.py index e738215323e5..4a7e0a3e3a36 100644 --- a/tests/models/transformers/test_models_transformer_cogview3plus.py +++ b/tests/models/transformers/test_models_transformer_cogview3plus.py @@ -22,7 +22,6 @@ from ..testing_utils import ( BaseModelTesterConfig, ModelTesterMixin, - TorchCompileTesterMixin, TrainingTesterMixin, ) @@ -99,7 +98,3 @@ class TestCogView3PlusTransformerTraining(CogView3PlusTransformerTesterConfig, T def test_gradient_checkpointing_is_applied(self): expected_set = {"CogView3PlusTransformer2DModel"} super().test_gradient_checkpointing_is_applied(expected_set=expected_set) - - -class TestCogView3PlusTransformerCompile(CogView3PlusTransformerTesterConfig, TorchCompileTesterMixin): - pass diff --git a/tests/models/transformers/test_models_transformer_cogview4.py b/tests/models/transformers/test_models_transformer_cogview4.py index 51c6dc06eac8..b7a87ea5b270 100644 --- a/tests/models/transformers/test_models_transformer_cogview4.py +++ b/tests/models/transformers/test_models_transformer_cogview4.py @@ -21,7 +21,6 @@ from ..testing_utils import ( BaseModelTesterConfig, ModelTesterMixin, - TorchCompileTesterMixin, TrainingTesterMixin, ) @@ -92,7 +91,3 @@ class TestCogView4TransformerTraining(CogView4TransformerTesterConfig, TrainingT def test_gradient_checkpointing_is_applied(self): expected_set = {"CogView4Transformer2DModel"} super().test_gradient_checkpointing_is_applied(expected_set=expected_set) - - -class TestCogView4TransformerCompile(CogView4TransformerTesterConfig, TorchCompileTesterMixin): - pass