Merge branch 'main' into cosmos-test-refactor

.github/workflows/claude_review.yml

@@ -32,9 +32,6 @@ jobs:
       )
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
-        with:
-          fetch-depth: 1
       - uses: anthropics/claude-code-action@v1
         with:
           anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

tests/models/transformers/test_models_transformer_bria.py

@@ -13,31 +13,23 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any
+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 ..testing_utils import (
-    BaseModelTesterConfig,
-    IPAdapterTesterMixin,
-    LoraHotSwappingForModelTesterMixin,
-    LoraTesterMixin,
-    ModelTesterMixin,
-    TorchCompileTesterMixin,
-    TrainingTesterMixin,
-)
+from ..test_modeling_common import LoraHotSwappingForModelTesterMixin, ModelTesterMixin, TorchCompileTesterMixin
 
 
 enable_full_determinism()
 
 
-def create_bria_ip_adapter_state_dict(model) -> dict[str, dict[str, Any]]:
+def create_bria_ip_adapter_state_dict(model):
+    # "ip_adapter" (cross-attention weights)
     ip_cross_attn_state_dict = {}
     key_id = 0
 
@@ -58,8 +50,11 @@ def create_bria_ip_adapter_state_dict(model) -> dict[str, dict[str, Any]]:
                 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"],
@@ -78,36 +73,53 @@ def create_bria_ip_adapter_state_dict(model) -> dict[str, dict[str, Any]]:
     )
 
     del sd
-    return {"image_proj": ip_image_projection_state_dict, "ip_adapter": ip_cross_attn_state_dict}
+    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 BriaTransformerTesterConfig(BaseModelTesterConfig):
-    @property
-    def model_class(self):
-        return BriaTransformer2DModel
+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
 
     @property
-    def main_input_name(self) -> str:
-        return "hidden_states"
+    def dummy_input(self):
+        batch_size = 1
+        num_latent_channels = 4
+        num_image_channels = 3
+        height = width = 4
+        sequence_length = 48
+        embedding_dim = 32
 
-    @property
-    def model_split_percents(self) -> list:
-        return [0.8, 0.7, 0.7]
+        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 output_shape(self) -> tuple:
-        return (16, 4)
-
-    @property
-    def input_shape(self) -> tuple:
-        return (16, 4)
-
-    @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,
+            "img_ids": image_ids,
+            "txt_ids": text_ids,
+            "timestep": timestep,
+        }
+
+    @property
+    def input_shape(self):
+        return (16, 4)
+
+    @property
+    def output_shape(self):
+        return (16, 4)
+
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
             "patch_size": 1,
             "in_channels": 4,
             "num_layers": 1,
@@ -119,35 +131,11 @@ class BriaTransformerTesterConfig(BaseModelTesterConfig):
             "axes_dims_rope": [0, 4, 4],
         }
 
-    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
+        inputs_dict = self.dummy_input
+        return init_dict, inputs_dict
 
-        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 = self.get_init_dict()
-        inputs_dict = self.get_dummy_inputs()
-
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
         model = self.model_class(**init_dict)
         model.to(torch_device)
         model.eval()
@@ -155,6 +143,7 @@ class TestBriaTransformer(BriaTransformerTesterConfig, ModelTesterMixin):
         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)
 
@@ -167,59 +156,26 @@ class TestBriaTransformer(BriaTransformerTesterConfig, ModelTesterMixin):
         with torch.no_grad():
             output_2 = model(**inputs_dict).to_tuple()[0]
 
-        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"
+        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",
         )
 
-
-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 TestBriaTransformerCompile(BriaTransformerTesterConfig, TorchCompileTesterMixin):
-    pass
+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 TestBriaTransformerIPAdapter(BriaTransformerTesterConfig, IPAdapterTesterMixin):
-    @property
-    def ip_adapter_processor_cls(self):
-        return FluxIPAdapterJointAttnProcessor2_0
+class BriaTransformerLoRAHotSwapTests(LoraHotSwappingForModelTesterMixin, unittest.TestCase):
+    model_class = BriaTransformer2DModel
 
-    def modify_inputs_for_ip_adapter(self, model, inputs_dict):
-        torch.manual_seed(0)
-        cross_attention_dim = getattr(model.config, "joint_attention_dim", 32)
-        image_embeds = torch.randn(1, 1, cross_attention_dim).to(torch_device)
-        inputs_dict.update({"joint_attention_kwargs": {"ip_adapter_image_embeds": image_embeds}})
-        return inputs_dict
-
-    def create_ip_adapter_state_dict(self, model: Any) -> dict[str, dict[str, Any]]:
-        return create_bria_ip_adapter_state_dict(model)
-
-
-class TestBriaTransformerLoRA(BriaTransformerTesterConfig, LoraTesterMixin):
-    pass
-
-
-class TestBriaTransformerLoRAHotSwap(BriaTransformerTesterConfig, LoraHotSwappingForModelTesterMixin):
-    @property
-    def different_shapes_for_compilation(self):
-        return [(4, 4), (4, 8), (8, 8)]
-
-    def get_dummy_inputs(self, height: int = 4, width: int = 4) -> dict[str, torch.Tensor]:
-        batch_size = 1
-        num_latent_channels = 4
-        num_image_channels = 3
-        sequence_length = 24
-        embedding_dim = 32
-
-        return {
-            "hidden_states": randn_tensor((batch_size, height * width, num_latent_channels), device=torch_device),
-            "encoder_hidden_states": randn_tensor((batch_size, sequence_length, embedding_dim), device=torch_device),
-            "img_ids": randn_tensor((height * width, num_image_channels), device=torch_device),
-            "txt_ids": randn_tensor((sequence_length, num_image_channels), device=torch_device),
-            "timestep": torch.tensor([1.0]).to(torch_device).expand(batch_size),
-        }
+    def prepare_init_args_and_inputs_for_common(self):
+        return BriaTransformerTests().prepare_init_args_and_inputs_for_common()

tests/models/transformers/test_models_transformer_bria_fibo.py

@@ -13,50 +13,62 @@
 # 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 ..testing_utils import (
-    BaseModelTesterConfig,
-    ModelTesterMixin,
-    TorchCompileTesterMixin,
-    TrainingTesterMixin,
-)
+from ..test_modeling_common import ModelTesterMixin
 
 
 enable_full_determinism()
 
 
-class BriaFiboTransformerTesterConfig(BaseModelTesterConfig):
-    @property
-    def model_class(self):
-        return BriaFiboTransformer2DModel
+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
 
     @property
-    def main_input_name(self) -> str:
-        return "hidden_states"
+    def dummy_input(self):
+        batch_size = 1
+        num_latent_channels = 48
+        num_image_channels = 3
+        height = width = 16
+        sequence_length = 32
+        embedding_dim = 64
+
+        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 model_split_percents(self) -> list:
-        return [0.8, 0.7, 0.7]
-
-    @property
-    def output_shape(self) -> tuple:
-        return (256, 48)
-
-    @property
-    def input_shape(self) -> tuple:
+    def input_shape(self):
         return (16, 16)
 
     @property
-    def generator(self):
-        return torch.Generator("cpu").manual_seed(0)
+    def output_shape(self):
+        return (256, 48)
 
-    def get_init_dict(self) -> dict:
-        return {
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
             "patch_size": 1,
             "in_channels": 48,
             "num_layers": 1,
@@ -69,41 +81,9 @@ class BriaFiboTransformerTesterConfig(BaseModelTesterConfig):
             "axes_dims_rope": [0, 4, 4],
         }
 
-    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
+        inputs_dict = self.dummy_input
+        return init_dict, inputs_dict
 
-        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)
-
-
-class TestBriaFiboTransformerCompile(BriaFiboTransformerTesterConfig, TorchCompileTesterMixin):
-    pass

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"}