[CI] Refactor Chroma , LongCat and HiDream Transformer Tests (#13345)

* update

* update

* update

---------

Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

Author: DN6

tests/models/transformers/test_models_transformer_chroma.py

@@ -13,113 +13,51 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import unittest
-
 import torch
 
 from diffusers import ChromaTransformer2DModel
-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,
+    LoraHotSwappingForModelTesterMixin,
+    LoraTesterMixin,
+    ModelTesterMixin,
+    TrainingTesterMixin,
+)
 
 
 enable_full_determinism()
 
 
-def create_chroma_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 ChromaTransformerTests(ModelTesterMixin, unittest.TestCase):
-    model_class = ChromaTransformer2DModel
-    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 ChromaTransformerTesterConfig(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 ChromaTransformer2DModel
 
-        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):
+    def input_shape(self) -> tuple:
         return (16, 4)
 
-    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": 4,
             "num_layers": 1,
@@ -133,51 +71,87 @@ def prepare_init_args_and_inputs_for_common(self):
             "approximator_layers": 1,
         }
 
-        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 TestChromaTransformer(ChromaTransformerTesterConfig, 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)
 
-        assert text_ids_3d.ndim == 3, "text_ids_3d should be a 3d tensor"
-        assert image_ids_3d.ndim == 3, "img_ids_3d should be a 3d tensor"
+        assert text_ids_3d.ndim == 3
+        assert image_ids_3d.ndim == 3
 
         inputs_dict["txt_ids"] = text_ids_3d
         inputs_dict["img_ids"] = image_ids_3d
 
         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 TestChromaTransformerTraining(ChromaTransformerTesterConfig, TrainingTesterMixin):
     def test_gradient_checkpointing_is_applied(self):
         expected_set = {"ChromaTransformer2DModel"}
         super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
 
 
-class ChromaTransformerCompileTests(TorchCompileTesterMixin, unittest.TestCase):
-    model_class = ChromaTransformer2DModel
+class TestChromaTransformerLoRA(ChromaTransformerTesterConfig, LoraTesterMixin):
+    pass
 
-    def prepare_init_args_and_inputs_for_common(self):
-        return ChromaTransformerTests().prepare_init_args_and_inputs_for_common()
 
+class TestChromaTransformerLoRAHotSwap(ChromaTransformerTesterConfig, LoraHotSwappingForModelTesterMixin):
+    @property
+    def different_shapes_for_compilation(self):
+        return [(4, 4), (4, 8), (8, 8)]
 
-class ChromaTransformerLoRAHotSwapTests(LoraHotSwappingForModelTesterMixin, unittest.TestCase):
-    model_class = ChromaTransformer2DModel
+    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
 
-    def prepare_init_args_and_inputs_for_common(self):
-        return ChromaTransformerTests().prepare_init_args_and_inputs_for_common()
+        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),
+        }

tests/models/transformers/test_models_transformer_hidream.py

@@ -13,61 +13,49 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import unittest
-
 import torch
 
 from diffusers import HiDreamImageTransformer2DModel
+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,
+    TrainingTesterMixin,
 )
-from ..test_modeling_common import ModelTesterMixin
 
 
 enable_full_determinism()
 
 
-class HiDreamTransformerTests(ModelTesterMixin, unittest.TestCase):
-    model_class = HiDreamImageTransformer2DModel
-    main_input_name = "hidden_states"
-    model_split_percents = [0.8, 0.8, 0.9]
-
+class HiDreamTransformerTesterConfig(BaseModelTesterConfig):
     @property
-    def dummy_input(self):
-        batch_size = 2
-        num_channels = 4
-        height = width = 32
-        embedding_dim_t5, embedding_dim_llama, embedding_dim_pooled = 8, 4, 8
-        sequence_length = 8
+    def model_class(self):
+        return HiDreamImageTransformer2DModel
 
-        hidden_states = torch.randn((batch_size, num_channels, height, width)).to(torch_device)
-        encoder_hidden_states_t5 = torch.randn((batch_size, sequence_length, embedding_dim_t5)).to(torch_device)
-        encoder_hidden_states_llama3 = torch.randn((batch_size, batch_size, sequence_length, embedding_dim_llama)).to(
-            torch_device
-        )
-        pooled_embeds = torch.randn((batch_size, embedding_dim_pooled)).to(torch_device)
-        timesteps = 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_t5": encoder_hidden_states_t5,
-            "encoder_hidden_states_llama3": encoder_hidden_states_llama3,
-            "pooled_embeds": pooled_embeds,
-            "timesteps": timesteps,
-        }
+    @property
+    def model_split_percents(self) -> list:
+        return [0.8, 0.8, 0.9]
 
     @property
-    def input_shape(self):
+    def output_shape(self) -> tuple:
         return (4, 32, 32)
 
     @property
-    def output_shape(self):
+    def input_shape(self) -> tuple:
         return (4, 32, 32)
 
-    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,
             "out_channels": 4,
@@ -82,15 +70,39 @@ def prepare_init_args_and_inputs_for_common(self):
             "axes_dims_rope": (4, 2, 2),
             "max_resolution": (32, 32),
             "llama_layers": (0, 1),
-            "force_inference_output": True,  # TODO: as we don't implement MoE loss in training tests.
+            "force_inference_output": True,
         }
-        inputs_dict = self.dummy_input
-        return init_dict, inputs_dict
 
-    @unittest.skip("HiDreamImageTransformer2DModel uses a dedicated attention processor. This test doesn't apply")
-    def test_set_attn_processor_for_determinism(self):
-        pass
+    def get_dummy_inputs(self, batch_size: int = 2) -> dict[str, torch.Tensor]:
+        num_channels = 4
+        height = width = 32
+        embedding_dim_t5, embedding_dim_llama, embedding_dim_pooled = 8, 4, 8
+        sequence_length = 8
+
+        return {
+            "hidden_states": randn_tensor(
+                (batch_size, num_channels, height, width), generator=self.generator, device=torch_device
+            ),
+            "encoder_hidden_states_t5": randn_tensor(
+                (batch_size, sequence_length, embedding_dim_t5), generator=self.generator, device=torch_device
+            ),
+            "encoder_hidden_states_llama3": randn_tensor(
+                (batch_size, batch_size, sequence_length, embedding_dim_llama),
+                generator=self.generator,
+                device=torch_device,
+            ),
+            "pooled_embeds": randn_tensor(
+                (batch_size, embedding_dim_pooled), generator=self.generator, device=torch_device
+            ),
+            "timesteps": torch.randint(0, 1000, size=(batch_size,), generator=self.generator).to(torch_device),
+        }
+
+
+class TestHiDreamTransformer(HiDreamTransformerTesterConfig, ModelTesterMixin):
+    pass
+
 
+class TestHiDreamTransformerTraining(HiDreamTransformerTesterConfig, TrainingTesterMixin):
     def test_gradient_checkpointing_is_applied(self):
         expected_set = {"HiDreamImageTransformer2DModel"}
         super().test_gradient_checkpointing_is_applied(expected_set=expected_set)