update

tests/models/transformers/test_models_transformer_cosmos.py

@@ -12,60 +12,46 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import unittest
-
 import torch
 
 from diffusers import CosmosTransformer3DModel
+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,
+    MemoryTesterMixin,
+    ModelTesterMixin,
+    TrainingTesterMixin,
+)
 
 
 enable_full_determinism()
 
 
-class CosmosTransformer3DModelTests(ModelTesterMixin, unittest.TestCase):
-    model_class = CosmosTransformer3DModel
-    main_input_name = "hidden_states"
-    uses_custom_attn_processor = True
+class CosmosTransformerTesterConfig(BaseModelTesterConfig):
+    @property
+    def model_class(self):
+        return CosmosTransformer3DModel
 
     @property
-    def dummy_input(self):
-        batch_size = 1
-        num_channels = 4
-        num_frames = 1
-        height = 16
-        width = 16
-        text_embed_dim = 16
-        sequence_length = 12
-        fps = 30
+    def output_shape(self) -> tuple[int, ...]:
+        return (4, 1, 16, 16)
 
-        hidden_states = torch.randn((batch_size, num_channels, num_frames, height, width)).to(torch_device)
-        timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device)
-        encoder_hidden_states = torch.randn((batch_size, sequence_length, text_embed_dim)).to(torch_device)
-        attention_mask = torch.ones((batch_size, sequence_length)).to(torch_device)
-        padding_mask = torch.zeros(batch_size, 1, height, width).to(torch_device)
+    @property
+    def input_shape(self) -> tuple[int, ...]:
+        return (4, 1, 16, 16)
 
+    @property
+    def main_input_name(self) -> str:
+        return "hidden_states"
+
+    @property
+    def generator(self):
+        return torch.Generator("cpu").manual_seed(0)
+
+    def get_init_dict(self) -> dict[str, int | list | tuple | float | bool | str]:
         return {
-            "hidden_states": hidden_states,
-            "timestep": timestep,
-            "encoder_hidden_states": encoder_hidden_states,
-            "attention_mask": attention_mask,
-            "fps": fps,
-            "padding_mask": padding_mask,
-        }
-
-    @property
-    def input_shape(self):
-        return (4, 1, 16, 16)
-
-    @property
-    def output_shape(self):
-        return (4, 1, 16, 16)
-
-    def prepare_init_args_and_inputs_for_common(self):
-        init_dict = {
             "in_channels": 4,
             "out_channels": 4,
             "num_attention_heads": 2,
@@ -80,57 +66,68 @@ class CosmosTransformer3DModelTests(ModelTesterMixin, unittest.TestCase):
             "concat_padding_mask": True,
             "extra_pos_embed_type": "learnable",
         }
-        inputs_dict = self.dummy_input
-        return init_dict, inputs_dict
 
-    def test_gradient_checkpointing_is_applied(self):
-        expected_set = {"CosmosTransformer3DModel"}
-        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
-
-
-class CosmosTransformer3DModelVideoToWorldTests(ModelTesterMixin, unittest.TestCase):
-    model_class = CosmosTransformer3DModel
-    main_input_name = "hidden_states"
-    uses_custom_attn_processor = True
-
-    @property
-    def dummy_input(self):
-        batch_size = 1
+    def get_dummy_inputs(self, batch_size: int = 1) -> dict[str, torch.Tensor]:
         num_channels = 4
         num_frames = 1
         height = 16
         width = 16
         text_embed_dim = 16
         sequence_length = 12
-        fps = 30
-
-        hidden_states = torch.randn((batch_size, num_channels, num_frames, height, width)).to(torch_device)
-        timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device)
-        encoder_hidden_states = torch.randn((batch_size, sequence_length, text_embed_dim)).to(torch_device)
-        attention_mask = torch.ones((batch_size, sequence_length)).to(torch_device)
-        condition_mask = torch.ones(batch_size, 1, num_frames, height, width).to(torch_device)
-        padding_mask = torch.zeros(batch_size, 1, height, width).to(torch_device)
 
         return {
-            "hidden_states": hidden_states,
-            "timestep": timestep,
-            "encoder_hidden_states": encoder_hidden_states,
-            "attention_mask": attention_mask,
-            "fps": fps,
-            "condition_mask": condition_mask,
-            "padding_mask": padding_mask,
+            "hidden_states": randn_tensor(
+                (batch_size, num_channels, num_frames, height, width), generator=self.generator, device=torch_device
+            ),
+            "timestep": torch.randint(0, 1000, size=(batch_size,), generator=self.generator).to(torch_device),
+            "encoder_hidden_states": randn_tensor(
+                (batch_size, sequence_length, text_embed_dim), generator=self.generator, device=torch_device
+            ),
+            "attention_mask": torch.ones((batch_size, sequence_length)).to(torch_device),
+            "fps": 30,
+            "padding_mask": torch.zeros(batch_size, 1, height, width).to(torch_device),
         }
 
+
+class TestCosmosTransformer(CosmosTransformerTesterConfig, ModelTesterMixin):
+    """Core model tests for Cosmos Transformer."""
+
+
+class TestCosmosTransformerMemory(CosmosTransformerTesterConfig, MemoryTesterMixin):
+    """Memory optimization tests for Cosmos Transformer."""
+
+
+class TestCosmosTransformerTraining(CosmosTransformerTesterConfig, TrainingTesterMixin):
+    """Training tests for Cosmos Transformer."""
+
+    def test_gradient_checkpointing_is_applied(self):
+        expected_set = {"CosmosTransformer3DModel"}
+        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
+
+
+class CosmosTransformerVideoToWorldTesterConfig(BaseModelTesterConfig):
     @property
-    def input_shape(self):
+    def model_class(self):
+        return CosmosTransformer3DModel
+
+    @property
+    def output_shape(self) -> tuple[int, ...]:
         return (4, 1, 16, 16)
 
     @property
-    def output_shape(self):
+    def input_shape(self) -> tuple[int, ...]:
         return (4, 1, 16, 16)
 
-    def prepare_init_args_and_inputs_for_common(self):
-        init_dict = {
+    @property
+    def main_input_name(self) -> str:
+        return "hidden_states"
+
+    @property
+    def generator(self):
+        return torch.Generator("cpu").manual_seed(0)
+
+    def get_init_dict(self) -> dict[str, int | list | tuple | float | bool | str]:
+        return {
             "in_channels": 4 + 1,
             "out_channels": 4,
             "num_attention_heads": 2,
@@ -145,8 +142,40 @@ class CosmosTransformer3DModelVideoToWorldTests(ModelTesterMixin, unittest.TestC
             "concat_padding_mask": True,
             "extra_pos_embed_type": "learnable",
         }
-        inputs_dict = self.dummy_input
-        return init_dict, inputs_dict
+
+    def get_dummy_inputs(self, batch_size: int = 1) -> dict[str, torch.Tensor]:
+        num_channels = 4
+        num_frames = 1
+        height = 16
+        width = 16
+        text_embed_dim = 16
+        sequence_length = 12
+
+        return {
+            "hidden_states": randn_tensor(
+                (batch_size, num_channels, num_frames, height, width), generator=self.generator, device=torch_device
+            ),
+            "timestep": torch.randint(0, 1000, size=(batch_size,), generator=self.generator).to(torch_device),
+            "encoder_hidden_states": randn_tensor(
+                (batch_size, sequence_length, text_embed_dim), generator=self.generator, device=torch_device
+            ),
+            "attention_mask": torch.ones((batch_size, sequence_length)).to(torch_device),
+            "fps": 30,
+            "condition_mask": torch.ones(batch_size, 1, num_frames, height, width).to(torch_device),
+            "padding_mask": torch.zeros(batch_size, 1, height, width).to(torch_device),
+        }
+
+
+class TestCosmosTransformerVideoToWorld(CosmosTransformerVideoToWorldTesterConfig, ModelTesterMixin):
+    """Core model tests for Cosmos Transformer (Video-to-World)."""
+
+
+class TestCosmosTransformerVideoToWorldMemory(CosmosTransformerVideoToWorldTesterConfig, MemoryTesterMixin):
+    """Memory optimization tests for Cosmos Transformer (Video-to-World)."""
+
+
+class TestCosmosTransformerVideoToWorldTraining(CosmosTransformerVideoToWorldTesterConfig, TrainingTesterMixin):
+    """Training tests for Cosmos Transformer (Video-to-World)."""
 
     def test_gradient_checkpointing_is_applied(self):
         expected_set = {"CosmosTransformer3DModel"}

tests/models/transformers/test_models_transformer_sd3.py

@@ -13,63 +13,58 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import unittest
+
 import torch
 
 from diffusers import SD3Transformer2DModel
-from diffusers.utils.torch_utils import randn_tensor
+from diffusers.utils.import_utils import is_xformers_available
 
-from ...testing_utils import enable_full_determinism, torch_device
-from ..testing_utils import (
-    BaseModelTesterConfig,
-    BitsAndBytesTesterMixin,
-    ModelTesterMixin,
-    TorchAoTesterMixin,
-    TorchCompileTesterMixin,
-    TrainingTesterMixin,
+from ...testing_utils import (
+    enable_full_determinism,
+    torch_device,
 )
+from ..test_modeling_common import ModelTesterMixin
 
 
 enable_full_determinism()
 
 
-# ======================== SD3 Transformer ========================
-
-
-class SD3TransformerTesterConfig(BaseModelTesterConfig):
-    @property
-    def model_class(self):
-        return SD3Transformer2DModel
+class SD3TransformerTests(ModelTesterMixin, unittest.TestCase):
+    model_class = SD3Transformer2DModel
+    main_input_name = "hidden_states"
+    model_split_percents = [0.8, 0.8, 0.9]
 
     @property
-    def pretrained_model_name_or_path(self):
-        return "hf-internal-testing/tiny-sd3-pipe"
+    def dummy_input(self):
+        batch_size = 2
+        num_channels = 4
+        height = width = embedding_dim = 32
+        pooled_embedding_dim = embedding_dim * 2
+        sequence_length = 154
 
-    @property
-    def pretrained_model_kwargs(self):
-        return {"subfolder": "transformer"}
+        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)
+        pooled_prompt_embeds = torch.randn((batch_size, pooled_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"
-
-    @property
-    def model_split_percents(self) -> list:
-        return [0.8, 0.8, 0.9]
-
-    @property
-    def output_shape(self) -> tuple:
-        return (4, 32, 32)
-
-    @property
-    def input_shape(self) -> tuple:
-        return (4, 32, 32)
-
-    @property
-    def generator(self):
-        return torch.Generator("cpu").manual_seed(0)
-
-    def get_init_dict(self) -> dict:
         return {
+            "hidden_states": hidden_states,
+            "encoder_hidden_states": encoder_hidden_states,
+            "pooled_projections": pooled_prompt_embeds,
+            "timestep": timestep,
+        }
+
+    @property
+    def input_shape(self):
+        return (4, 32, 32)
+
+    @property
+    def output_shape(self):
+        return (4, 32, 32)
+
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
             "sample_size": 32,
             "patch_size": 1,
             "in_channels": 4,
@@ -84,79 +79,67 @@ class SD3TransformerTesterConfig(BaseModelTesterConfig):
             "dual_attention_layers": (),
             "qk_norm": None,
         }
+        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 = width = embedding_dim = 32
-        pooled_embedding_dim = embedding_dim * 2
-        sequence_length = 154
+    @unittest.skipIf(
+        torch_device != "cuda" or not is_xformers_available(),
+        reason="XFormers attention is only available with CUDA and `xformers` installed",
+    )
+    def test_xformers_enable_works(self):
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**init_dict)
 
-        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
-            ),
-            "pooled_projections": randn_tensor(
-                (batch_size, pooled_embedding_dim), generator=self.generator, device=torch_device
-            ),
-            "timestep": torch.randint(0, 1000, size=(batch_size,), generator=self.generator).to(torch_device),
-        }
+        model.enable_xformers_memory_efficient_attention()
 
+        assert model.transformer_blocks[0].attn.processor.__class__.__name__ == "XFormersJointAttnProcessor", (
+            "xformers is not enabled"
+        )
 
-class TestSD3Transformer(SD3TransformerTesterConfig, ModelTesterMixin):
-    pass
+    @unittest.skip("SD3Transformer2DModel uses a dedicated attention processor. This test doesn't apply")
+    def test_set_attn_processor_for_determinism(self):
+        pass
 
-
-class TestSD3TransformerTraining(SD3TransformerTesterConfig, TrainingTesterMixin):
     def test_gradient_checkpointing_is_applied(self):
         expected_set = {"SD3Transformer2DModel"}
         super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
 
 
-class TestSD3TransformerCompile(SD3TransformerTesterConfig, TorchCompileTesterMixin):
-    pass
-
-
-# ======================== SD3.5 Transformer ========================
-
-
-class SD35TransformerTesterConfig(BaseModelTesterConfig):
-    @property
-    def model_class(self):
-        return SD3Transformer2DModel
+class SD35TransformerTests(ModelTesterMixin, unittest.TestCase):
+    model_class = SD3Transformer2DModel
+    main_input_name = "hidden_states"
+    model_split_percents = [0.8, 0.8, 0.9]
 
     @property
-    def pretrained_model_name_or_path(self):
-        return "hf-internal-testing/tiny-sd35-pipe"
+    def dummy_input(self):
+        batch_size = 2
+        num_channels = 4
+        height = width = embedding_dim = 32
+        pooled_embedding_dim = embedding_dim * 2
+        sequence_length = 154
 
-    @property
-    def pretrained_model_kwargs(self):
-        return {"subfolder": "transformer"}
+        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)
+        pooled_prompt_embeds = torch.randn((batch_size, pooled_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"
-
-    @property
-    def model_split_percents(self) -> list:
-        return [0.8, 0.8, 0.9]
-
-    @property
-    def output_shape(self) -> tuple:
-        return (4, 32, 32)
-
-    @property
-    def input_shape(self) -> tuple:
-        return (4, 32, 32)
-
-    @property
-    def generator(self):
-        return torch.Generator("cpu").manual_seed(0)
-
-    def get_init_dict(self) -> dict:
         return {
+            "hidden_states": hidden_states,
+            "encoder_hidden_states": encoder_hidden_states,
+            "pooled_projections": pooled_prompt_embeds,
+            "timestep": timestep,
+        }
+
+    @property
+    def input_shape(self):
+        return (4, 32, 32)
+
+    @property
+    def output_shape(self):
+        return (4, 32, 32)
+
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
             "sample_size": 32,
             "patch_size": 1,
             "in_channels": 4,
@@ -171,56 +154,47 @@ class SD35TransformerTesterConfig(BaseModelTesterConfig):
             "dual_attention_layers": (0,),
             "qk_norm": "rms_norm",
         }
+        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 = width = embedding_dim = 32
-        pooled_embedding_dim = embedding_dim * 2
-        sequence_length = 154
+    @unittest.skipIf(
+        torch_device != "cuda" or not is_xformers_available(),
+        reason="XFormers attention is only available with CUDA and `xformers` installed",
+    )
+    def test_xformers_enable_works(self):
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**init_dict)
 
-        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
-            ),
-            "pooled_projections": randn_tensor(
-                (batch_size, pooled_embedding_dim), generator=self.generator, device=torch_device
-            ),
-            "timestep": torch.randint(0, 1000, size=(batch_size,), generator=self.generator).to(torch_device),
-        }
+        model.enable_xformers_memory_efficient_attention()
 
+        assert model.transformer_blocks[0].attn.processor.__class__.__name__ == "XFormersJointAttnProcessor", (
+            "xformers is not enabled"
+        )
 
-class TestSD35Transformer(SD35TransformerTesterConfig, ModelTesterMixin):
-    def test_skip_layers(self):
-        init_dict = self.get_init_dict()
-        inputs_dict = self.get_dummy_inputs()
-        model = self.model_class(**init_dict).to(torch_device)
+    @unittest.skip("SD3Transformer2DModel uses a dedicated attention processor. This test doesn't apply")
+    def test_set_attn_processor_for_determinism(self):
+        pass
 
-        output_full = model(**inputs_dict).sample
-
-        inputs_dict_with_skip = inputs_dict.copy()
-        inputs_dict_with_skip["skip_layers"] = [0]
-        output_skip = model(**inputs_dict_with_skip).sample
-
-        assert not torch.allclose(output_full, output_skip, atol=1e-5), "Outputs should differ when layers are skipped"
-        assert output_full.shape == output_skip.shape, "Outputs should have the same shape"
-
-
-class TestSD35TransformerTraining(SD35TransformerTesterConfig, TrainingTesterMixin):
     def test_gradient_checkpointing_is_applied(self):
         expected_set = {"SD3Transformer2DModel"}
         super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
 
+    def test_skip_layers(self):
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**init_dict).to(torch_device)
 
-class TestSD35TransformerCompile(SD35TransformerTesterConfig, TorchCompileTesterMixin):
-    pass
+        # Forward pass without skipping layers
+        output_full = model(**inputs_dict).sample
 
+        # Forward pass with skipping layers 0 (since there's only one layer in this test setup)
+        inputs_dict_with_skip = inputs_dict.copy()
+        inputs_dict_with_skip["skip_layers"] = [0]
+        output_skip = model(**inputs_dict_with_skip).sample
 
-class TestSD35TransformerBitsAndBytes(SD35TransformerTesterConfig, BitsAndBytesTesterMixin):
-    """BitsAndBytes quantization tests for SD3.5 Transformer."""
+        # Check that the outputs are different
+        self.assertFalse(
+            torch.allclose(output_full, output_skip, atol=1e-5), "Outputs should differ when layers are skipped"
+        )
 
-
-class TestSD35TransformerTorchAo(SD35TransformerTesterConfig, TorchAoTesterMixin):
-    """TorchAO quantization tests for SD3.5 Transformer."""
+        # Check that the outputs have the same shape
+        self.assertEqual(output_full.shape, output_skip.shape, "Outputs should have the same shape")