update

* update

* update

src/diffusers/quantizers/quantization_config.py

@@ -36,7 +36,7 @@ from typing import Any, Callable
 
 from packaging import version
 
-from ..utils import is_torch_available, is_torchao_available, is_torchao_version, logging
+from ..utils import deprecate, is_torch_available, is_torchao_available, is_torchao_version, logging
 
 
 if is_torch_available():
@@ -844,6 +844,8 @@ class QuantoConfig(QuantizationConfigMixin):
         modules_to_not_convert: list[str] | None = None,
         **kwargs,
     ):
+        deprecation_message = "`QuantoConfig` is deprecated and will be removed in version 1.0.0."
+        deprecate("QuantoConfig", "1.0.0", deprecation_message)
         self.quant_method = QuantizationMethod.QUANTO
         self.weights_dtype = weights_dtype
         self.modules_to_not_convert = modules_to_not_convert

src/diffusers/quantizers/quanto/quanto_quantizer.py

@@ -3,6 +3,7 @@ from typing import TYPE_CHECKING, Any
 from diffusers.utils.import_utils import is_optimum_quanto_version
 
 from ...utils import (
+    deprecate,
     get_module_from_name,
     is_accelerate_available,
     is_accelerate_version,
@@ -42,6 +43,9 @@ class QuantoQuantizer(DiffusersQuantizer):
         super().__init__(quantization_config, **kwargs)
 
     def validate_environment(self, *args, **kwargs):
+        deprecation_message = "The Quanto quantizer is deprecated and will be removed in version 1.0.0."
+        deprecate("QuantoQuantizer", "1.0.0", deprecation_message)
+
         if not is_optimum_quanto_available():
             raise ImportError(
                 "Loading an optimum-quanto quantized model requires optimum-quanto library (`pip install optimum-quanto`)"

tests/models/transformers/test_models_transformer_ltx.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2025 HuggingFace Inc.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -13,59 +12,47 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import unittest
-
 import torch
 
 from diffusers import LTXVideoTransformer3DModel
+from diffusers.utils.torch_utils import randn_tensor
 
 from ...testing_utils import enable_full_determinism, torch_device
-from ..test_modeling_common import ModelTesterMixin, TorchCompileTesterMixin
+from ..testing_utils import (
+    BaseModelTesterConfig,
+    MemoryTesterMixin,
+    ModelTesterMixin,
+    TorchCompileTesterMixin,
+    TrainingTesterMixin,
+)
 
 
 enable_full_determinism()
 
 
-class LTXTransformerTests(ModelTesterMixin, unittest.TestCase):
-    model_class = LTXVideoTransformer3DModel
-    main_input_name = "hidden_states"
-    uses_custom_attn_processor = True
+class LTXTransformerTesterConfig(BaseModelTesterConfig):
+    @property
+    def model_class(self):
+        return LTXVideoTransformer3DModel
 
     @property
-    def dummy_input(self):
-        batch_size = 2
-        num_channels = 4
-        num_frames = 2
-        height = 16
-        width = 16
-        embedding_dim = 16
-        sequence_length = 16
+    def output_shape(self) -> tuple[int, int]:
+        return (512, 4)
 
-        hidden_states = torch.randn((batch_size, num_frames * height * width, num_channels)).to(torch_device)
-        encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
-        encoder_attention_mask = torch.ones((batch_size, sequence_length)).bool().to(torch_device)
-        timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device)
+    @property
+    def input_shape(self) -> tuple[int, int]:
+        return (512, 4)
 
+    @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):
         return {
-            "hidden_states": hidden_states,
-            "encoder_hidden_states": encoder_hidden_states,
-            "timestep": timestep,
-            "encoder_attention_mask": encoder_attention_mask,
-            "num_frames": num_frames,
-            "height": height,
-            "width": width,
-        }
-
-    @property
-    def input_shape(self):
-        return (512, 4)
-
-    @property
-    def output_shape(self):
-        return (512, 4)
-
-    def prepare_init_args_and_inputs_for_common(self):
-        init_dict = {
             "in_channels": 4,
             "out_channels": 4,
             "num_attention_heads": 2,
@@ -75,16 +62,57 @@ class LTXTransformerTests(ModelTesterMixin, unittest.TestCase):
             "qk_norm": "rms_norm_across_heads",
             "caption_channels": 16,
         }
-        inputs_dict = self.dummy_input
-        return init_dict, inputs_dict
+
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
+        batch_size = 2
+        num_channels = 4
+        num_frames = 2
+        height = 16
+        width = 16
+        embedding_dim = 16
+        sequence_length = 16
+
+        return {
+            "hidden_states": randn_tensor(
+                (batch_size, num_frames * height * width, num_channels),
+                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),
+            "encoder_attention_mask": torch.ones((batch_size, sequence_length)).bool().to(torch_device),
+            "num_frames": num_frames,
+            "height": height,
+            "width": width,
+        }
+
+
+class TestLTXTransformer(LTXTransformerTesterConfig, ModelTesterMixin):
+    """Core model tests for LTX Video Transformer."""
+
+
+class TestLTXTransformerMemory(LTXTransformerTesterConfig, MemoryTesterMixin):
+    """Memory optimization tests for LTX Video Transformer."""
+
+
+class TestLTXTransformerTraining(LTXTransformerTesterConfig, TrainingTesterMixin):
+    """Training tests for LTX Video Transformer."""
 
     def test_gradient_checkpointing_is_applied(self):
-        expected_set = {"LTXVideoTransformer3DModel"}
-        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
+        super().test_gradient_checkpointing_is_applied(expected_set={"LTXVideoTransformer3DModel"})
 
 
-class LTXTransformerCompileTests(TorchCompileTesterMixin, unittest.TestCase):
-    model_class = LTXVideoTransformer3DModel
+class TestLTXTransformerCompile(LTXTransformerTesterConfig, TorchCompileTesterMixin):
+    """Torch compile tests for LTX Video Transformer."""
 
-    def prepare_init_args_and_inputs_for_common(self):
-        return LTXTransformerTests().prepare_init_args_and_inputs_for_common()
+
+# TODO: Add pretrained_model_name_or_path once a tiny LTX model is available on the Hub
+# class TestLTXTransformerBitsAndBytes(LTXTransformerTesterConfig, BitsAndBytesTesterMixin):
+#     """BitsAndBytes quantization tests for LTX Video Transformer."""
+
+
+# TODO: Add pretrained_model_name_or_path once a tiny LTX model is available on the Hub
+# class TestLTXTransformerTorchAo(LTXTransformerTesterConfig, TorchAoTesterMixin):
+#     """TorchAo quantization tests for LTX Video Transformer."""

tests/models/transformers/test_models_transformer_ltx2.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2025 HuggingFace Inc.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -13,77 +12,49 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import unittest
-
+import pytest
 import torch
 
 from diffusers import LTX2VideoTransformer3DModel
+from diffusers.utils.torch_utils import randn_tensor
 
 from ...testing_utils import enable_full_determinism, torch_device
-from ..test_modeling_common import ModelTesterMixin, TorchCompileTesterMixin
+from ..testing_utils import (
+    AttentionTesterMixin,
+    BaseModelTesterConfig,
+    MemoryTesterMixin,
+    ModelTesterMixin,
+    TorchCompileTesterMixin,
+    TrainingTesterMixin,
+)
 
 
 enable_full_determinism()
 
 
-class LTX2TransformerTests(ModelTesterMixin, unittest.TestCase):
-    model_class = LTX2VideoTransformer3DModel
-    main_input_name = "hidden_states"
-    uses_custom_attn_processor = True
+class LTX2TransformerTesterConfig(BaseModelTesterConfig):
+    @property
+    def model_class(self):
+        return LTX2VideoTransformer3DModel
 
     @property
-    def dummy_input(self):
-        # Common
-        batch_size = 2
+    def output_shape(self) -> tuple[int, int]:
+        return (512, 4)
 
-        # Video
-        num_frames = 2
-        num_channels = 4
-        height = 16
-        width = 16
+    @property
+    def input_shape(self) -> tuple[int, int]:
+        return (512, 4)
 
-        # Audio
-        audio_num_frames = 9
-        audio_num_channels = 2
-        num_mel_bins = 2
+    @property
+    def main_input_name(self) -> str:
+        return "hidden_states"
 
-        # Text
-        embedding_dim = 16
-        sequence_length = 16
-
-        hidden_states = torch.randn((batch_size, num_frames * height * width, num_channels)).to(torch_device)
-        audio_hidden_states = torch.randn((batch_size, audio_num_frames, audio_num_channels * num_mel_bins)).to(
-            torch_device
-        )
-        encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
-        audio_encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
-        encoder_attention_mask = torch.ones((batch_size, sequence_length)).bool().to(torch_device)
-        timestep = torch.rand((batch_size,)).to(torch_device) * 1000
+    @property
+    def generator(self):
+        return torch.Generator("cpu").manual_seed(0)
 
+    def get_init_dict(self):
         return {
-            "hidden_states": hidden_states,
-            "audio_hidden_states": audio_hidden_states,
-            "encoder_hidden_states": encoder_hidden_states,
-            "audio_encoder_hidden_states": audio_encoder_hidden_states,
-            "timestep": timestep,
-            "encoder_attention_mask": encoder_attention_mask,
-            "num_frames": num_frames,
-            "height": height,
-            "width": width,
-            "audio_num_frames": audio_num_frames,
-            "fps": 25.0,
-        }
-
-    @property
-    def input_shape(self):
-        return (512, 4)
-
-    @property
-    def output_shape(self):
-        return (512, 4)
-
-    def prepare_init_args_and_inputs_for_common(self):
-        init_dict = {
             "in_channels": 4,
             "out_channels": 4,
             "patch_size": 1,
@@ -101,122 +72,80 @@ class LTX2TransformerTests(ModelTesterMixin, unittest.TestCase):
             "caption_channels": 16,
             "rope_double_precision": False,
         }
-        inputs_dict = self.dummy_input
-        return init_dict, inputs_dict
+
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
+        batch_size = 2
+        num_frames = 2
+        num_channels = 4
+        height = 16
+        width = 16
+        audio_num_frames = 9
+        audio_num_channels = 2
+        num_mel_bins = 2
+        embedding_dim = 16
+        sequence_length = 16
+
+        return {
+            "hidden_states": randn_tensor(
+                (batch_size, num_frames * height * width, num_channels),
+                generator=self.generator,
+                device=torch_device,
+            ),
+            "audio_hidden_states": randn_tensor(
+                (batch_size, audio_num_frames, audio_num_channels * num_mel_bins),
+                generator=self.generator,
+                device=torch_device,
+            ),
+            "encoder_hidden_states": randn_tensor(
+                (batch_size, sequence_length, embedding_dim), generator=self.generator, device=torch_device
+            ),
+            "audio_encoder_hidden_states": randn_tensor(
+                (batch_size, sequence_length, embedding_dim), generator=self.generator, device=torch_device
+            ),
+            "timestep": (randn_tensor((batch_size,), generator=self.generator, device=torch_device).abs() * 1000),
+            "encoder_attention_mask": torch.ones((batch_size, sequence_length)).bool().to(torch_device),
+            "num_frames": num_frames,
+            "height": height,
+            "width": width,
+            "audio_num_frames": audio_num_frames,
+            "fps": 25.0,
+        }
+
+
+class TestLTX2Transformer(LTX2TransformerTesterConfig, ModelTesterMixin):
+    """Core model tests for LTX2 Video Transformer."""
+
+
+class TestLTX2TransformerMemory(LTX2TransformerTesterConfig, MemoryTesterMixin):
+    """Memory optimization tests for LTX2 Video Transformer."""
+
+
+class TestLTX2TransformerTraining(LTX2TransformerTesterConfig, TrainingTesterMixin):
+    """Training tests for LTX2 Video Transformer."""
 
     def test_gradient_checkpointing_is_applied(self):
-        expected_set = {"LTX2VideoTransformer3DModel"}
-        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
-
-    # def test_ltx2_consistency(self, seed=0, dtype=torch.float32):
-    #     torch.manual_seed(seed)
-    #     init_dict, _ = self.prepare_init_args_and_inputs_for_common()
-
-    #     # Calculate dummy inputs in a custom manner to ensure compatibility with original code
-    #     batch_size = 2
-    #     num_frames = 9
-    #     latent_frames = 2
-    #     text_embedding_dim = 16
-    #     text_seq_len = 16
-    #     fps = 25.0
-    #     sampling_rate = 16000.0
-    #     hop_length = 160.0
-
-    #     sigma = torch.rand((1,), generator=torch.manual_seed(seed), dtype=dtype, device="cpu") * 1000
-    #     timestep = (sigma * torch.ones((batch_size,), dtype=dtype, device="cpu")).to(device=torch_device)
-
-    #     num_channels = 4
-    #     latent_height = 4
-    #     latent_width = 4
-    #     hidden_states = torch.randn(
-    #         (batch_size, num_channels, latent_frames, latent_height, latent_width),
-    #         generator=torch.manual_seed(seed),
-    #         dtype=dtype,
-    #         device="cpu",
-    #     )
-    #     # Patchify video latents (with patch_size (1, 1, 1))
-    #     hidden_states = hidden_states.reshape(batch_size, -1, latent_frames, 1, latent_height, 1, latent_width, 1)
-    #     hidden_states = hidden_states.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3)
-    #     encoder_hidden_states = torch.randn(
-    #         (batch_size, text_seq_len, text_embedding_dim),
-    #         generator=torch.manual_seed(seed),
-    #         dtype=dtype,
-    #         device="cpu",
-    #     )
-
-    #     audio_num_channels = 2
-    #     num_mel_bins = 2
-    #     latent_length = int((sampling_rate / hop_length / 4) * (num_frames / fps))
-    #     audio_hidden_states = torch.randn(
-    #         (batch_size, audio_num_channels, latent_length, num_mel_bins),
-    #         generator=torch.manual_seed(seed),
-    #         dtype=dtype,
-    #         device="cpu",
-    #     )
-    #     # Patchify audio latents
-    #     audio_hidden_states = audio_hidden_states.transpose(1, 2).flatten(2, 3)
-    #     audio_encoder_hidden_states = torch.randn(
-    #         (batch_size, text_seq_len, text_embedding_dim),
-    #         generator=torch.manual_seed(seed),
-    #         dtype=dtype,
-    #         device="cpu",
-    #     )
-
-    #     inputs_dict = {
-    #         "hidden_states": hidden_states.to(device=torch_device),
-    #         "audio_hidden_states": audio_hidden_states.to(device=torch_device),
-    #         "encoder_hidden_states": encoder_hidden_states.to(device=torch_device),
-    #         "audio_encoder_hidden_states": audio_encoder_hidden_states.to(device=torch_device),
-    #         "timestep": timestep,
-    #         "num_frames": latent_frames,
-    #         "height": latent_height,
-    #         "width": latent_width,
-    #         "audio_num_frames": num_frames,
-    #         "fps": 25.0,
-    #     }
-
-    #     model = self.model_class.from_pretrained(
-    #         "diffusers-internal-dev/dummy-ltx2",
-    #         subfolder="transformer",
-    #         device_map="cpu",
-    #     )
-    #     # torch.manual_seed(seed)
-    #     # model = self.model_class(**init_dict)
-    #     model.to(torch_device)
-    #     model.eval()
-
-    #     with attention_backend("native"):
-    #         with torch.no_grad():
-    #             output = model(**inputs_dict)
-
-    #             video_output, audio_output = output.to_tuple()
-
-    #     self.assertIsNotNone(video_output)
-    #     self.assertIsNotNone(audio_output)
-
-    #     # input & output have to have the same shape
-    #     video_expected_shape = (batch_size, latent_frames * latent_height * latent_width, num_channels)
-    #     self.assertEqual(video_output.shape, video_expected_shape, "Video input and output shapes do not match")
-    #     audio_expected_shape = (batch_size, latent_length, audio_num_channels * num_mel_bins)
-    #     self.assertEqual(audio_output.shape, audio_expected_shape, "Audio input and output shapes do not match")
-
-    #     # Check against expected slice
-    #     # fmt: off
-    #     video_expected_slice = torch.tensor([0.4783, 1.6954, -1.2092, 0.1762, 0.7801, 1.2025, -1.4525, -0.2721, 0.3354, 1.9144, -1.5546, 0.0831, 0.4391, 1.7012, -1.7373, -0.2676])
-    #     audio_expected_slice = torch.tensor([-0.4236, 0.4750, 0.3901, -0.4339, -0.2782, 0.4357, 0.4526, -0.3927, -0.0980, 0.4870, 0.3964, -0.3169, -0.3974, 0.4408, 0.3809, -0.4692])
-    #     # fmt: on
-
-    #     video_output_flat = video_output.cpu().flatten().float()
-    #     video_generated_slice = torch.cat([video_output_flat[:8], video_output_flat[-8:]])
-    #     self.assertTrue(torch.allclose(video_generated_slice, video_expected_slice, atol=1e-4))
-
-    #     audio_output_flat = audio_output.cpu().flatten().float()
-    #     audio_generated_slice = torch.cat([audio_output_flat[:8], audio_output_flat[-8:]])
-    #     self.assertTrue(torch.allclose(audio_generated_slice, audio_expected_slice, atol=1e-4))
+        super().test_gradient_checkpointing_is_applied(expected_set={"LTX2VideoTransformer3DModel"})
 
 
-class LTX2TransformerCompileTests(TorchCompileTesterMixin, unittest.TestCase):
-    model_class = LTX2VideoTransformer3DModel
+class TestLTX2TransformerAttention(LTX2TransformerTesterConfig, AttentionTesterMixin):
+    """Attention processor tests for LTX2 Video Transformer."""
 
-    def prepare_init_args_and_inputs_for_common(self):
-        return LTX2TransformerTests().prepare_init_args_and_inputs_for_common()
+    @pytest.mark.skip(
+        "LTX2Attention does not set is_cross_attention, so fuse_projections tries to fuse Q+K+V together even for cross-attention modules with different input dimensions."
+    )
+    def test_fuse_unfuse_qkv_projections(self, atol=1e-3, rtol=0):
+        pass
+
+
+class TestLTX2TransformerCompile(LTX2TransformerTesterConfig, TorchCompileTesterMixin):
+    """Torch compile tests for LTX2 Video Transformer."""
+
+
+# TODO: Add pretrained_model_name_or_path once a tiny LTX2 model is available on the Hub
+# class TestLTX2TransformerBitsAndBytes(LTX2TransformerTesterConfig, BitsAndBytesTesterMixin):
+#     """BitsAndBytes quantization tests for LTX2 Video Transformer."""
+
+
+# TODO: Add pretrained_model_name_or_path once a tiny LTX2 model is available on the Hub
+# class TestLTX2TransformerTorchAo(LTX2TransformerTesterConfig, TorchAoTesterMixin):
+#     """TorchAo quantization tests for LTX2 Video Transformer."""

tests/models/transformers/test_models_transformer_z_image.py

@@ -1,3 +1,4 @@
+# coding=utf-8
 # Copyright 2025 HuggingFace Inc.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -12,23 +13,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import gc
 import os
+import unittest
 
-import pytest
 import torch
 
 from diffusers import ZImageTransformer2DModel
-from diffusers.utils.torch_utils import randn_tensor
 
-from ...testing_utils import assert_tensors_close, torch_device
-from ..testing_utils import (
-    BaseModelTesterConfig,
-    LoraTesterMixin,
-    MemoryTesterMixin,
-    ModelTesterMixin,
-    TorchCompileTesterMixin,
-    TrainingTesterMixin,
-)
+from ...testing_utils import IS_GITHUB_ACTIONS, torch_device
+from ..test_modeling_common import ModelTesterMixin, TorchCompileTesterMixin
 
 
 # Z-Image requires torch.use_deterministic_algorithms(False) due to complex64 RoPE operations
@@ -42,38 +36,44 @@ if hasattr(torch.backends, "cuda"):
     torch.backends.cuda.matmul.allow_tf32 = False
 
 
-def _concat_list_output(output):
-    """Model output `sample` is a list of tensors. Concatenate them for comparison."""
-    return torch.cat([t.flatten() for t in output])
+@unittest.skipIf(
+    IS_GITHUB_ACTIONS,
+    reason="Skipping test-suite inside the CI because the model has `torch.empty()` inside of it during init and we don't have a clear way to override it in the modeling tests.",
+)
+class ZImageTransformerTests(ModelTesterMixin, unittest.TestCase):
+    model_class = ZImageTransformer2DModel
+    main_input_name = "x"
+    # We override the items here because the transformer under consideration is small.
+    model_split_percents = [0.9, 0.9, 0.9]
 
+    def prepare_dummy_input(self, height=16, width=16):
+        batch_size = 1
+        num_channels = 16
+        embedding_dim = 16
+        sequence_length = 16
 
-class ZImageTransformerTesterConfig(BaseModelTesterConfig):
-    @property
-    def model_class(self):
-        return ZImageTransformer2DModel
+        hidden_states = [torch.randn((num_channels, 1, height, width)).to(torch_device) for _ in range(batch_size)]
+        encoder_hidden_states = [
+            torch.randn((sequence_length, embedding_dim)).to(torch_device) for _ in range(batch_size)
+        ]
+        timestep = torch.tensor([0.0]).to(torch_device)
+
+        return {"x": hidden_states, "cap_feats": encoder_hidden_states, "t": timestep}
 
     @property
-    def output_shape(self) -> tuple[int, ...]:
+    def dummy_input(self):
+        return self.prepare_dummy_input()
+
+    @property
+    def input_shape(self):
         return (4, 32, 32)
 
     @property
-    def input_shape(self) -> tuple[int, ...]:
+    def output_shape(self):
         return (4, 32, 32)
 
-    @property
-    def model_split_percents(self) -> list:
-        return [0.9, 0.9, 0.9]
-
-    @property
-    def main_input_name(self) -> str:
-        return "x"
-
-    @property
-    def generator(self):
-        return torch.Generator("cpu").manual_seed(0)
-
-    def get_init_dict(self):
-        return {
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
             "all_patch_size": (2,),
             "all_f_patch_size": (1,),
             "in_channels": 16,
@@ -89,223 +89,83 @@ class ZImageTransformerTesterConfig(BaseModelTesterConfig):
             "axes_dims": [8, 4, 4],
             "axes_lens": [256, 32, 32],
         }
+        inputs_dict = self.dummy_input
+        return init_dict, inputs_dict
 
-    def get_dummy_inputs(self) -> dict[str, torch.Tensor | list]:
-        batch_size = 1
-        num_channels = 16
-        embedding_dim = 16
-        sequence_length = 16
-        height = 16
-        width = 16
-
-        hidden_states = [
-            randn_tensor((num_channels, 1, height, width), generator=self.generator, device=torch_device)
-            for _ in range(batch_size)
-        ]
-        encoder_hidden_states = [
-            randn_tensor((sequence_length, embedding_dim), generator=self.generator, device=torch_device)
-            for _ in range(batch_size)
-        ]
-        timestep = torch.tensor([0.0]).to(torch_device)
-
-        return {"x": hidden_states, "cap_feats": encoder_hidden_states, "t": timestep}
-
-
-class TestZImageTransformer(ZImageTransformerTesterConfig, ModelTesterMixin):
-    """Core model tests for Z-Image Transformer."""
-
-    @torch.no_grad()
-    def test_determinism(self, atol=1e-5, rtol=0):
-        model = self.model_class(**self.get_init_dict())
-        model.to(torch_device)
-        model.eval()
-
-        inputs_dict = self.get_dummy_inputs()
-        first = _concat_list_output(model(**inputs_dict, return_dict=False)[0])
-        second = _concat_list_output(model(**inputs_dict, return_dict=False)[0])
-
-        mask = ~(torch.isnan(first) | torch.isnan(second))
-        assert_tensors_close(
-            first[mask], second[mask], atol=atol, rtol=rtol, msg="Model outputs are not deterministic"
-        )
-
-    def test_from_save_pretrained(self, tmp_path, atol=5e-5, rtol=5e-5):
+    def setUp(self):
+        gc.collect()
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            torch.cuda.synchronize()
         torch.manual_seed(0)
-        model = self.model_class(**self.get_init_dict())
-        model.to(torch_device)
-        model.eval()
-
-        model.save_pretrained(tmp_path)
-        new_model = self.model_class.from_pretrained(tmp_path)
-        new_model.to(torch_device)
-
-        for param_name in model.state_dict().keys():
-            param_1 = model.state_dict()[param_name]
-            param_2 = new_model.state_dict()[param_name]
-            assert param_1.shape == param_2.shape
-
-        inputs_dict = self.get_dummy_inputs()
-        image = _concat_list_output(model(**inputs_dict, return_dict=False)[0])
-        new_image = _concat_list_output(new_model(**inputs_dict, return_dict=False)[0])
-
-        assert_tensors_close(image, new_image, atol=atol, rtol=rtol, msg="Models give different forward passes.")
-
-    @torch.no_grad()
-    def test_from_save_pretrained_variant(self, tmp_path, atol=5e-5, rtol=0):
-        model = self.model_class(**self.get_init_dict())
-        model.to(torch_device)
-        model.eval()
-
-        model.save_pretrained(tmp_path, variant="fp16")
-        new_model = self.model_class.from_pretrained(tmp_path, variant="fp16")
-
-        with pytest.raises(OSError) as exc_info:
-            self.model_class.from_pretrained(tmp_path)
-
-        assert "Error no file named diffusion_pytorch_model.bin found in directory" in str(exc_info.value)
-
-        new_model.to(torch_device)
-
-        inputs_dict = self.get_dummy_inputs()
-        image = _concat_list_output(model(**inputs_dict, return_dict=False)[0])
-        new_image = _concat_list_output(new_model(**inputs_dict, return_dict=False)[0])
-
-        assert_tensors_close(image, new_image, atol=atol, rtol=rtol, msg="Models give different forward passes.")
-
-    @pytest.mark.skip("Model output `sample` is a list of tensors, not a single tensor.")
-    def test_outputs_equivalence(self, atol=1e-5, rtol=0):
-        pass
-
-    def test_sharded_checkpoints_with_parallel_loading(self, tmp_path, atol=1e-5, rtol=0):
-        from diffusers.utils import SAFE_WEIGHTS_INDEX_NAME, constants
-
-        from ..testing_utils.common import calculate_expected_num_shards, compute_module_persistent_sizes
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(0)
 
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            torch.cuda.synchronize()
         torch.manual_seed(0)
-        config = self.get_init_dict()
-        inputs_dict = self.get_dummy_inputs()
-        model = self.model_class(**config).eval()
-        model = model.to(torch_device)
-
-        base_output = _concat_list_output(model(**inputs_dict, return_dict=False)[0])
-
-        model_size = compute_module_persistent_sizes(model)[""]
-        max_shard_size = int((model_size * 0.75) / (2**10))
-
-        original_parallel_loading = constants.HF_ENABLE_PARALLEL_LOADING
-        original_parallel_workers = getattr(constants, "HF_PARALLEL_WORKERS", None)
-
-        try:
-            model.cpu().save_pretrained(tmp_path, max_shard_size=f"{max_shard_size}KB")
-            assert os.path.exists(os.path.join(tmp_path, SAFE_WEIGHTS_INDEX_NAME))
-
-            expected_num_shards = calculate_expected_num_shards(os.path.join(tmp_path, SAFE_WEIGHTS_INDEX_NAME))
-            actual_num_shards = len([file for file in os.listdir(tmp_path) if file.endswith(".safetensors")])
-            assert actual_num_shards == expected_num_shards
-
-            constants.HF_ENABLE_PARALLEL_LOADING = False
-            self.model_class.from_pretrained(tmp_path).eval().to(torch_device)
-
-            constants.HF_ENABLE_PARALLEL_LOADING = True
-            constants.DEFAULT_HF_PARALLEL_LOADING_WORKERS = 2
-
-            torch.manual_seed(0)
-            model_parallel = self.model_class.from_pretrained(tmp_path).eval()
-            model_parallel = model_parallel.to(torch_device)
-
-            output_parallel = _concat_list_output(model_parallel(**inputs_dict, return_dict=False)[0])
-
-            assert_tensors_close(
-                base_output, output_parallel, atol=atol, rtol=rtol, msg="Output should match with parallel loading"
-            )
-        finally:
-            constants.HF_ENABLE_PARALLEL_LOADING = original_parallel_loading
-            if original_parallel_workers is not None:
-                constants.HF_PARALLEL_WORKERS = original_parallel_workers
-
-
-class TestZImageTransformerMemory(ZImageTransformerTesterConfig, MemoryTesterMixin):
-    """Memory optimization tests for Z-Image Transformer."""
-
-    @pytest.mark.skip(
-        "Ensure `x_pad_token` and `cap_pad_token` are cast to the same dtype as the destination tensor before they are assigned to the padding indices."
-    )
-    def test_layerwise_casting_training(self):
-        pass
-
-
-class TestZImageTransformerTraining(ZImageTransformerTesterConfig, TrainingTesterMixin):
-    """Training tests for Z-Image Transformer."""
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed_all(0)
 
     def test_gradient_checkpointing_is_applied(self):
-        super().test_gradient_checkpointing_is_applied(expected_set={"ZImageTransformer2DModel"})
+        expected_set = {"ZImageTransformer2DModel"}
+        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
 
-    @pytest.mark.skip("Test is not supported for handling main inputs that are lists.")
+    @unittest.skip("Test is not supported for handling main inputs that are lists.")
     def test_training(self):
-        pass
+        super().test_training()
 
-    @pytest.mark.skip("Test is not supported for handling main inputs that are lists.")
-    def test_training_with_ema(self):
-        pass
+    @unittest.skip("Test is not supported for handling main inputs that are lists.")
+    def test_ema_training(self):
+        super().test_ema_training()
 
-    @pytest.mark.skip("Test is not supported for handling main inputs that are lists.")
-    def test_gradient_checkpointing_equivalence(self, loss_tolerance=1e-5, param_grad_tol=5e-5, skip=None):
-        pass
+    @unittest.skip("Test is not supported for handling main inputs that are lists.")
+    def test_effective_gradient_checkpointing(self):
+        super().test_effective_gradient_checkpointing()
+
+    @unittest.skip(
+        "Test needs to be revisited. But we need to ensure `x_pad_token` and `cap_pad_token` are cast to the same dtype as the destination tensor before they are assigned to the padding indices."
+    )
+    def test_layerwise_casting_training(self):
+        super().test_layerwise_casting_training()
+
+    @unittest.skip("Test is not supported for handling main inputs that are lists.")
+    def test_outputs_equivalence(self):
+        super().test_outputs_equivalence()
+
+    @unittest.skip("Test will pass if we change to deterministic values instead of empty in the DiT.")
+    def test_group_offloading(self):
+        super().test_group_offloading()
+
+    @unittest.skip("Test will pass if we change to deterministic values instead of empty in the DiT.")
+    def test_group_offloading_with_disk(self):
+        super().test_group_offloading_with_disk()
 
 
-class TestZImageTransformerLoRA(ZImageTransformerTesterConfig, LoraTesterMixin):
-    """LoRA adapter tests for Z-Image Transformer."""
+class ZImageTransformerCompileTests(TorchCompileTesterMixin, unittest.TestCase):
+    model_class = ZImageTransformer2DModel
+    different_shapes_for_compilation = [(4, 4), (4, 8), (8, 8)]
 
-    @pytest.mark.skip("Model output `sample` is a list of tensors, not a single tensor.")
-    def test_save_load_lora_adapter(self, tmp_path, rank=4, lora_alpha=4, use_dora=False, atol=1e-4, rtol=1e-4):
-        pass
+    def prepare_init_args_and_inputs_for_common(self):
+        return ZImageTransformerTests().prepare_init_args_and_inputs_for_common()
 
+    def prepare_dummy_input(self, height, width):
+        return ZImageTransformerTests().prepare_dummy_input(height=height, width=width)
 
-# TODO: Add pretrained_model_name_or_path once a tiny Z-Image model is available on the Hub
-# class TestZImageTransformerBitsAndBytes(ZImageTransformerTesterConfig, BitsAndBytesTesterMixin):
-#     """BitsAndBytes quantization tests for Z-Image Transformer."""
-
-
-# TODO: Add pretrained_model_name_or_path once a tiny Z-Image model is available on the Hub
-# class TestZImageTransformerTorchAo(ZImageTransformerTesterConfig, TorchAoTesterMixin):
-#     """TorchAo quantization tests for Z-Image Transformer."""
-
-
-class TestZImageTransformerCompile(ZImageTransformerTesterConfig, TorchCompileTesterMixin):
-    """Torch compile tests for Z-Image Transformer."""
-
-    @property
-    def different_shapes_for_compilation(self):
-        return [(4, 4), (4, 8), (8, 8)]
-
-    def get_dummy_inputs(self, height: int = 16, width: int = 16) -> dict[str, torch.Tensor | list]:
-        batch_size = 1
-        num_channels = 16
-        embedding_dim = 16
-        sequence_length = 16
-
-        hidden_states = [
-            randn_tensor((num_channels, 1, height, width), generator=self.generator, device=torch_device)
-            for _ in range(batch_size)
-        ]
-        encoder_hidden_states = [
-            randn_tensor((sequence_length, embedding_dim), generator=self.generator, device=torch_device)
-            for _ in range(batch_size)
-        ]
-        timestep = torch.tensor([0.0]).to(torch_device)
-
-        return {"x": hidden_states, "cap_feats": encoder_hidden_states, "t": timestep}
-
-    @pytest.mark.skip(
-        "The repeated block in this model is ZImageTransformerBlock, which is used for noise_refiner, context_refiner, and layers. The inputs recorded for the block would vary during compilation and full compilation with fullgraph=True would trigger recompilation at least thrice."
+    @unittest.skip(
+        "The repeated block in this model is ZImageTransformerBlock, which is used for noise_refiner, context_refiner, and layers. As a consequence of this, the inputs recorded for the block would vary during compilation and full compilation with fullgraph=True would trigger recompilation at least thrice."
     )
     def test_torch_compile_recompilation_and_graph_break(self):
-        pass
+        super().test_torch_compile_recompilation_and_graph_break()
 
-    @pytest.mark.skip("Fullgraph AoT is broken")
-    def test_compile_works_with_aot(self, tmp_path):
-        pass
+    @unittest.skip("Fullgraph AoT is broken")
+    def test_compile_works_with_aot(self):
+        super().test_compile_works_with_aot()
 
-    @pytest.mark.skip("Fullgraph is broken")
+    @unittest.skip("Fullgraph is broken")
     def test_compile_on_different_shapes(self):
-        pass
+        super().test_compile_on_different_shapes()