Merge branch 'main' into tests-conditional-pipeline-blocks

* fix docs

* update main example

docs/source/en/_toctree.yml

@@ -532,8 +532,6 @@
         title: ControlNet-XS with Stable Diffusion XL
       - local: api/pipelines/controlnet_union
         title: ControlNetUnion
-      - local: api/pipelines/cosmos
-        title: Cosmos
       - local: api/pipelines/ddim
         title: DDIM
       - local: api/pipelines/ddpm
@@ -677,6 +675,8 @@
         title: CogVideoX
       - local: api/pipelines/consisid
         title: ConsisID
+      - local: api/pipelines/cosmos
+        title: Cosmos
       - local: api/pipelines/framepack
         title: Framepack
       - local: api/pipelines/helios

docs/source/en/api/pipelines/cosmos.md

@@ -21,29 +21,31 @@
 > [!TIP]
 > Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
 
-## Loading original format checkpoints
-
-Original format checkpoints that have not been converted to diffusers-expected format can be loaded using the `from_single_file` method.
+## Basic usage
 
 ```python
 import torch
-from diffusers import Cosmos2TextToImagePipeline, CosmosTransformer3DModel
+from diffusers import Cosmos2_5_PredictBasePipeline
+from diffusers.utils import export_to_video
 
-model_id = "nvidia/Cosmos-Predict2-2B-Text2Image"
-transformer = CosmosTransformer3DModel.from_single_file(
-    "https://huggingface.co/nvidia/Cosmos-Predict2-2B-Text2Image/blob/main/model.pt",
-    torch_dtype=torch.bfloat16,
-).to("cuda")
-pipe = Cosmos2TextToImagePipeline.from_pretrained(model_id, transformer=transformer, torch_dtype=torch.bfloat16)
+model_id = "nvidia/Cosmos-Predict2.5-2B"
+pipe = Cosmos2_5_PredictBasePipeline.from_pretrained(
+    model_id, revision="diffusers/base/post-trained", torch_dtype=torch.bfloat16
+)
 pipe.to("cuda")
 
-prompt = "A close-up shot captures a vibrant yellow scrubber vigorously working on a grimy plate, its bristles moving in circular motions to lift stubborn grease and food residue. The dish, once covered in remnants of a hearty meal, gradually reveals its original glossy surface. Suds form and bubble around the scrubber, creating a satisfying visual of cleanliness in progress. The sound of scrubbing fills the air, accompanied by the gentle clinking of the dish against the sink. As the scrubber continues its task, the dish transforms, gleaming under the bright kitchen lights, symbolizing the triumph of cleanliness over mess."
+prompt = "As the red light shifts to green, the red bus at the intersection begins to move forward, its headlights cutting through the falling snow. The snowy tire tracks deepen as the vehicle inches ahead, casting fresh lines onto the slushy road. Around it, streetlights glow warmer, illuminating the drifting flakes and wet reflections on the asphalt. Other cars behind start to edge forward, their beams joining the scene. The stillness of the urban street transitions into motion as the quiet snowfall is punctuated by the slow advance of traffic through the frosty city corridor."
 negative_prompt = "The video captures a series of frames showing ugly scenes, static with no motion, motion blur, over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, jerky movements, low frame rate, artifacting, color banding, unnatural transitions, outdated special effects, fake elements, unconvincing visuals, poorly edited content, jump cuts, visual noise, and flickering. Overall, the video is of poor quality."
 
 output = pipe(
-    prompt=prompt, negative_prompt=negative_prompt, generator=torch.Generator().manual_seed(1)
-).images[0]
-output.save("output.png")
+    image=None,
+    video=None,
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    num_frames=93,
+    generator=torch.Generator().manual_seed(1),
+).frames[0]
+export_to_video(output, "text2world.mp4", fps=16)
 ```
 
 ## Cosmos2_5_TransferPipeline

docs/source/en/api/pipelines/overview.md

@@ -44,6 +44,7 @@ The table below lists all the pipelines currently available in 🤗 Diffusers an
 | [ControlNet with Stable Diffusion XL](controlnet_sdxl) | text2image |
 | [ControlNet-XS](controlnetxs) | text2image |
 | [ControlNet-XS with Stable Diffusion XL](controlnetxs_sdxl) | text2image |
+| [Cosmos](cosmos) | text2video, video2video |
 | [Dance Diffusion](dance_diffusion) | unconditional audio generation |
 | [DDIM](ddim) | unconditional image generation |
 | [DDPM](ddpm) | unconditional image generation |

src/diffusers/models/_modeling_parallel.py

@@ -60,6 +60,16 @@ class ContextParallelConfig:
         rotate_method (`str`, *optional*, defaults to `"allgather"`):
             Method to use for rotating key/value states across devices in ring attention. Currently, only `"allgather"`
             is supported.
+        ulysses_anything (`bool`, *optional*, defaults to `False`):
+            Whether to enable "Ulysses Anything" mode, which supports arbitrary sequence lengths and head counts that
+            are not evenly divisible by `ulysses_degree`. When enabled, `ulysses_degree` must be greater than 1 and
+            `ring_degree` must be 1.
+        mesh (`torch.distributed.device_mesh.DeviceMesh`, *optional*):
+            A custom device mesh to use for context parallelism. If provided, this mesh will be used instead of
+            creating a new one. This is useful when combining context parallelism with other parallelism strategies
+            (e.g., FSDP, tensor parallelism) that share the same device mesh. The mesh must have both "ring" and
+            "ulysses" dimensions. Use size 1 for dimensions not being used (e.g., `mesh_shape=(2, 1, 4)` with
+            `mesh_dim_names=("ring", "ulysses", "fsdp")` for ring attention only with FSDP).
 
     """
 
@@ -68,6 +78,7 @@ class ContextParallelConfig:
     convert_to_fp32: bool = True
     # TODO: support alltoall
     rotate_method: Literal["allgather", "alltoall"] = "allgather"
+    mesh: torch.distributed.device_mesh.DeviceMesh | None = None
     # Whether to enable ulysses anything attention to support
     # any sequence lengths and any head numbers.
     ulysses_anything: bool = False
@@ -124,7 +135,7 @@ class ContextParallelConfig:
                 f"The product of `ring_degree` ({self.ring_degree}) and `ulysses_degree` ({self.ulysses_degree}) must not exceed the world size ({world_size})."
             )
 
-        self._flattened_mesh = self._mesh._flatten()
+        self._flattened_mesh = self._mesh["ring", "ulysses"]._flatten()
         self._ring_mesh = self._mesh["ring"]
         self._ulysses_mesh = self._mesh["ulysses"]
         self._ring_local_rank = self._ring_mesh.get_local_rank()

src/diffusers/models/modeling_utils.py

@@ -1567,7 +1567,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
         mesh = None
         if config.context_parallel_config is not None:
             cp_config = config.context_parallel_config
-            mesh = torch.distributed.device_mesh.init_device_mesh(
+            mesh = cp_config.mesh or torch.distributed.device_mesh.init_device_mesh(
                 device_type=device_type,
                 mesh_shape=cp_config.mesh_shape,
                 mesh_dim_names=cp_config.mesh_dim_names,

src/diffusers/pipelines/auto_pipeline.py

@@ -95,6 +95,7 @@ from .pag import (
     StableDiffusionXLPAGPipeline,
 )
 from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline
+from .prx import PRXPipeline
 from .qwenimage import (
     QwenImageControlNetPipeline,
     QwenImageEditInpaintPipeline,
@@ -185,6 +186,7 @@ AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict(
         ("z-image-controlnet-inpaint", ZImageControlNetInpaintPipeline),
         ("z-image-omni", ZImageOmniPipeline),
         ("ovis", OvisImagePipeline),
+        ("prx", PRXPipeline),
     ]
 )
 

src/diffusers/pipelines/cosmos/pipeline_cosmos2_5_transfer.py

@@ -82,13 +82,16 @@ EXAMPLE_DOC_STRING = """
         ```python
         >>> import cv2
         >>> import numpy as np
+        >>> from PIL import Image
         >>> import torch
         >>> from diffusers import Cosmos2_5_TransferPipeline, AutoModel
         >>> from diffusers.utils import export_to_video, load_video
 
         >>> model_id = "nvidia/Cosmos-Transfer2.5-2B"
         >>> # Load a Transfer2.5 controlnet variant (edge, depth, seg, or blur)
-        >>> controlnet = AutoModel.from_pretrained(model_id, revision="diffusers/controlnet/general/edge")
+        >>> controlnet = AutoModel.from_pretrained(
+        ...     model_id, revision="diffusers/controlnet/general/edge", torch_dtype=torch.bfloat16
+        ... )
         >>> pipe = Cosmos2_5_TransferPipeline.from_pretrained(
         ...     model_id, controlnet=controlnet, revision="diffusers/general", torch_dtype=torch.bfloat16
         ... )

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/testing_utils/parallelism.py

@@ -60,12 +60,7 @@ def _context_parallel_worker(rank, world_size, master_port, model_class, init_di
         model.eval()
 
         # Move inputs to device
-        inputs_on_device = {}
-        for key, value in inputs_dict.items():
-            if isinstance(value, torch.Tensor):
-                inputs_on_device[key] = value.to(device)
-            else:
-                inputs_on_device[key] = value
+        inputs_on_device = {k: v.to(device) if isinstance(v, torch.Tensor) else v for k, v in inputs_dict.items()}
 
         # Enable context parallelism
         cp_config = ContextParallelConfig(**cp_dict)
@@ -89,6 +84,59 @@ def _context_parallel_worker(rank, world_size, master_port, model_class, init_di
             dist.destroy_process_group()
 
 
+def _custom_mesh_worker(
+    rank,
+    world_size,
+    master_port,
+    model_class,
+    init_dict,
+    cp_dict,
+    mesh_shape,
+    mesh_dim_names,
+    inputs_dict,
+    return_dict,
+):
+    """Worker function for context parallel testing with a user-provided custom DeviceMesh."""
+    try:
+        os.environ["MASTER_ADDR"] = "localhost"
+        os.environ["MASTER_PORT"] = str(master_port)
+        os.environ["RANK"] = str(rank)
+        os.environ["WORLD_SIZE"] = str(world_size)
+
+        dist.init_process_group(backend="nccl", rank=rank, world_size=world_size)
+
+        torch.cuda.set_device(rank)
+        device = torch.device(f"cuda:{rank}")
+
+        model = model_class(**init_dict)
+        model.to(device)
+        model.eval()
+
+        inputs_on_device = {k: v.to(device) if isinstance(v, torch.Tensor) else v for k, v in inputs_dict.items()}
+
+        # DeviceMesh must be created after init_process_group, inside each worker process.
+        mesh = torch.distributed.device_mesh.init_device_mesh(
+            "cuda", mesh_shape=mesh_shape, mesh_dim_names=mesh_dim_names
+        )
+        cp_config = ContextParallelConfig(**cp_dict, mesh=mesh)
+        model.enable_parallelism(config=cp_config)
+
+        with torch.no_grad():
+            output = model(**inputs_on_device, return_dict=False)[0]
+
+        if rank == 0:
+            return_dict["status"] = "success"
+            return_dict["output_shape"] = list(output.shape)
+
+    except Exception as e:
+        if rank == 0:
+            return_dict["status"] = "error"
+            return_dict["error"] = str(e)
+    finally:
+        if dist.is_initialized():
+            dist.destroy_process_group()
+
+
 @is_context_parallel
 @require_torch_multi_accelerator
 class ContextParallelTesterMixin:
@@ -126,3 +174,48 @@ class ContextParallelTesterMixin:
         assert return_dict.get("status") == "success", (
             f"Context parallel inference failed: {return_dict.get('error', 'Unknown error')}"
         )
+
+    @pytest.mark.parametrize(
+        "cp_type,mesh_shape,mesh_dim_names",
+        [
+            ("ring_degree", (2, 1, 1), ("ring", "ulysses", "fsdp")),
+            ("ulysses_degree", (1, 2, 1), ("ring", "ulysses", "fsdp")),
+        ],
+        ids=["ring-3d-fsdp", "ulysses-3d-fsdp"],
+    )
+    def test_context_parallel_custom_mesh(self, cp_type, mesh_shape, mesh_dim_names):
+        if not torch.distributed.is_available():
+            pytest.skip("torch.distributed is not available.")
+
+        if not hasattr(self.model_class, "_cp_plan") or self.model_class._cp_plan is None:
+            pytest.skip("Model does not have a _cp_plan defined for context parallel inference.")
+
+        world_size = 2
+        init_dict = self.get_init_dict()
+        inputs_dict = {k: v.cpu() if isinstance(v, torch.Tensor) else v for k, v in self.get_dummy_inputs().items()}
+        cp_dict = {cp_type: world_size}
+
+        master_port = _find_free_port()
+        manager = mp.Manager()
+        return_dict = manager.dict()
+
+        mp.spawn(
+            _custom_mesh_worker,
+            args=(
+                world_size,
+                master_port,
+                self.model_class,
+                init_dict,
+                cp_dict,
+                mesh_shape,
+                mesh_dim_names,
+                inputs_dict,
+                return_dict,
+            ),
+            nprocs=world_size,
+            join=True,
+        )
+
+        assert return_dict.get("status") == "success", (
+            f"Custom mesh context parallel inference failed: {return_dict.get('error', 'Unknown error')}"
+        )

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()

tests/modular_pipelines/test_conditional_pipeline_blocks.py

@@ -0,0 +1,242 @@
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from diffusers.modular_pipelines import (
+    AutoPipelineBlocks,
+    ConditionalPipelineBlocks,
+    InputParam,
+    ModularPipelineBlocks,
+)
+
+
+class TextToImageBlock(ModularPipelineBlocks):
+    model_name = "text2img"
+
+    @property
+    def inputs(self):
+        return [InputParam(name="prompt")]
+
+    @property
+    def intermediate_outputs(self):
+        return []
+
+    @property
+    def description(self):
+        return "text-to-image workflow"
+
+    def __call__(self, components, state):
+        block_state = self.get_block_state(state)
+        block_state.workflow = "text2img"
+        self.set_block_state(state, block_state)
+        return components, state
+
+
+class ImageToImageBlock(ModularPipelineBlocks):
+    model_name = "img2img"
+
+    @property
+    def inputs(self):
+        return [InputParam(name="prompt"), InputParam(name="image")]
+
+    @property
+    def intermediate_outputs(self):
+        return []
+
+    @property
+    def description(self):
+        return "image-to-image workflow"
+
+    def __call__(self, components, state):
+        block_state = self.get_block_state(state)
+        block_state.workflow = "img2img"
+        self.set_block_state(state, block_state)
+        return components, state
+
+
+class InpaintBlock(ModularPipelineBlocks):
+    model_name = "inpaint"
+
+    @property
+    def inputs(self):
+        return [InputParam(name="prompt"), InputParam(name="image"), InputParam(name="mask")]
+
+    @property
+    def intermediate_outputs(self):
+        return []
+
+    @property
+    def description(self):
+        return "inpaint workflow"
+
+    def __call__(self, components, state):
+        block_state = self.get_block_state(state)
+        block_state.workflow = "inpaint"
+        self.set_block_state(state, block_state)
+        return components, state
+
+
+class ConditionalImageBlocks(ConditionalPipelineBlocks):
+    block_classes = [InpaintBlock, ImageToImageBlock, TextToImageBlock]
+    block_names = ["inpaint", "img2img", "text2img"]
+    block_trigger_inputs = ["mask", "image"]
+    default_block_name = "text2img"
+
+    @property
+    def description(self):
+        return "Conditional image blocks for testing"
+
+    def select_block(self, mask=None, image=None) -> str | None:
+        if mask is not None:
+            return "inpaint"
+        if image is not None:
+            return "img2img"
+        return None  # falls back to default_block_name
+
+
+class OptionalConditionalBlocks(ConditionalPipelineBlocks):
+    block_classes = [InpaintBlock, ImageToImageBlock]
+    block_names = ["inpaint", "img2img"]
+    block_trigger_inputs = ["mask", "image"]
+    default_block_name = None  # no default; block can be skipped
+
+    @property
+    def description(self):
+        return "Optional conditional blocks (skippable)"
+
+    def select_block(self, mask=None, image=None) -> str | None:
+        if mask is not None:
+            return "inpaint"
+        if image is not None:
+            return "img2img"
+        return None
+
+
+class AutoImageBlocks(AutoPipelineBlocks):
+    block_classes = [InpaintBlock, ImageToImageBlock, TextToImageBlock]
+    block_names = ["inpaint", "img2img", "text2img"]
+    block_trigger_inputs = ["mask", "image", None]
+
+    @property
+    def description(self):
+        return "Auto image blocks for testing"
+
+
+class TestConditionalPipelineBlocksSelectBlock:
+    def test_select_block_with_mask(self):
+        blocks = ConditionalImageBlocks()
+        assert blocks.select_block(mask="something") == "inpaint"
+
+    def test_select_block_with_image(self):
+        blocks = ConditionalImageBlocks()
+        assert blocks.select_block(image="something") == "img2img"
+
+    def test_select_block_with_mask_and_image(self):
+        blocks = ConditionalImageBlocks()
+        assert blocks.select_block(mask="m", image="i") == "inpaint"
+
+    def test_select_block_no_triggers_returns_none(self):
+        blocks = ConditionalImageBlocks()
+        assert blocks.select_block() is None
+
+    def test_select_block_explicit_none_values(self):
+        blocks = ConditionalImageBlocks()
+        assert blocks.select_block(mask=None, image=None) is None
+
+
+class TestConditionalPipelineBlocksWorkflowSelection:
+    def test_default_workflow_when_no_triggers(self):
+        blocks = ConditionalImageBlocks()
+        execution = blocks.get_execution_blocks()
+        assert execution is not None
+        assert isinstance(execution, TextToImageBlock)
+
+    def test_mask_trigger_selects_inpaint(self):
+        blocks = ConditionalImageBlocks()
+        execution = blocks.get_execution_blocks(mask=True)
+        assert isinstance(execution, InpaintBlock)
+
+    def test_image_trigger_selects_img2img(self):
+        blocks = ConditionalImageBlocks()
+        execution = blocks.get_execution_blocks(image=True)
+        assert isinstance(execution, ImageToImageBlock)
+
+    def test_mask_and_image_selects_inpaint(self):
+        blocks = ConditionalImageBlocks()
+        execution = blocks.get_execution_blocks(mask=True, image=True)
+        assert isinstance(execution, InpaintBlock)
+
+    def test_skippable_block_returns_none(self):
+        blocks = OptionalConditionalBlocks()
+        execution = blocks.get_execution_blocks()
+        assert execution is None
+
+    def test_skippable_block_still_selects_when_triggered(self):
+        blocks = OptionalConditionalBlocks()
+        execution = blocks.get_execution_blocks(image=True)
+        assert isinstance(execution, ImageToImageBlock)
+
+
+class TestAutoPipelineBlocksSelectBlock:
+    def test_auto_select_mask(self):
+        blocks = AutoImageBlocks()
+        assert blocks.select_block(mask="m") == "inpaint"
+
+    def test_auto_select_image(self):
+        blocks = AutoImageBlocks()
+        assert blocks.select_block(image="i") == "img2img"
+
+    def test_auto_select_default(self):
+        blocks = AutoImageBlocks()
+        # No trigger -> returns None -> falls back to default (text2img)
+        assert blocks.select_block() is None
+
+    def test_auto_select_priority_order(self):
+        blocks = AutoImageBlocks()
+        assert blocks.select_block(mask="m", image="i") == "inpaint"
+
+
+class TestAutoPipelineBlocksWorkflowSelection:
+    def test_auto_default_workflow(self):
+        blocks = AutoImageBlocks()
+        execution = blocks.get_execution_blocks()
+        assert isinstance(execution, TextToImageBlock)
+
+    def test_auto_mask_workflow(self):
+        blocks = AutoImageBlocks()
+        execution = blocks.get_execution_blocks(mask=True)
+        assert isinstance(execution, InpaintBlock)
+
+    def test_auto_image_workflow(self):
+        blocks = AutoImageBlocks()
+        execution = blocks.get_execution_blocks(image=True)
+        assert isinstance(execution, ImageToImageBlock)
+
+
+class TestConditionalPipelineBlocksStructure:
+    def test_block_names_accessible(self):
+        blocks = ConditionalImageBlocks()
+        sub = dict(blocks.sub_blocks)
+        assert set(sub.keys()) == {"inpaint", "img2img", "text2img"}
+
+    def test_sub_block_types(self):
+        blocks = ConditionalImageBlocks()
+        sub = dict(blocks.sub_blocks)
+        assert isinstance(sub["inpaint"], InpaintBlock)
+        assert isinstance(sub["img2img"], ImageToImageBlock)
+        assert isinstance(sub["text2img"], TextToImageBlock)
+
+    def test_description(self):
+        blocks = ConditionalImageBlocks()
+        assert "Conditional" in blocks.description

tests/modular_pipelines/test_modular_pipelines_common.py

@@ -9,11 +9,6 @@ import torch
 import diffusers
 from diffusers import AutoModel, ComponentsManager, ModularPipeline, ModularPipelineBlocks
 from diffusers.guiders import ClassifierFreeGuidance
-from diffusers.modular_pipelines import (
-    ConditionalPipelineBlocks,
-    LoopSequentialPipelineBlocks,
-    SequentialPipelineBlocks,
-)
 from diffusers.modular_pipelines.modular_pipeline_utils import (
     ComponentSpec,
     ConfigSpec,
@@ -24,7 +19,6 @@ from diffusers.modular_pipelines.modular_pipeline_utils import (
 from diffusers.utils import logging
 
 from ..testing_utils import (
-    CaptureLogger,
     backend_empty_cache,
     numpy_cosine_similarity_distance,
     require_accelerator,
@@ -437,117 +431,6 @@ class ModularGuiderTesterMixin:
         assert max_diff > expected_max_diff, "Output with CFG must be different from normal inference"
 
 
-class TestCustomBlockRequirements:
-    def get_dummy_block_pipe(self):
-        class DummyBlockOne:
-            # keep two arbitrary deps so that we can test warnings.
-            _requirements = {"xyz": ">=0.8.0", "abc": ">=10.0.0"}
-
-        class DummyBlockTwo:
-            # keep two dependencies that will be available during testing.
-            _requirements = {"transformers": ">=4.44.0", "diffusers": ">=0.2.0"}
-
-        pipe = SequentialPipelineBlocks.from_blocks_dict(
-            {"dummy_block_one": DummyBlockOne, "dummy_block_two": DummyBlockTwo}
-        )
-        return pipe
-
-    def get_dummy_conditional_block_pipe(self):
-        class DummyBlockOne:
-            _requirements = {"xyz": ">=0.8.0", "abc": ">=10.0.0"}
-
-        class DummyBlockTwo:
-            _requirements = {"transformers": ">=4.44.0", "diffusers": ">=0.2.0"}
-
-        class DummyConditionalBlocks(ConditionalPipelineBlocks):
-            block_classes = [DummyBlockOne, DummyBlockTwo]
-            block_names = ["block_one", "block_two"]
-            block_trigger_inputs = []
-
-            def select_block(self, **kwargs):
-                return "block_one"
-
-        return DummyConditionalBlocks()
-
-    def get_dummy_loop_block_pipe(self):
-        class DummyBlockOne:
-            _requirements = {"xyz": ">=0.8.0", "abc": ">=10.0.0"}
-
-        class DummyBlockTwo:
-            _requirements = {"transformers": ">=4.44.0", "diffusers": ">=0.2.0"}
-
-        return LoopSequentialPipelineBlocks.from_blocks_dict({"block_one": DummyBlockOne, "block_two": DummyBlockTwo})
-
-    def test_sequential_block_requirements_save_load(self, tmp_path):
-        pipe = self.get_dummy_block_pipe()
-        pipe.save_pretrained(str(tmp_path))
-
-        config_path = tmp_path / "modular_config.json"
-
-        with open(config_path, "r") as f:
-            config = json.load(f)
-
-        assert "requirements" in config
-        requirements = config["requirements"]
-
-        expected_requirements = {
-            "xyz": ">=0.8.0",
-            "abc": ">=10.0.0",
-            "transformers": ">=4.44.0",
-            "diffusers": ">=0.2.0",
-        }
-        assert expected_requirements == requirements
-
-    def test_sequential_block_requirements_warnings(self, tmp_path):
-        pipe = self.get_dummy_block_pipe()
-
-        logger = logging.get_logger("diffusers.modular_pipelines.modular_pipeline_utils")
-        logger.setLevel(30)
-
-        with CaptureLogger(logger) as cap_logger:
-            pipe.save_pretrained(str(tmp_path))
-
-        template = "{req} was specified in the requirements but wasn't found in the current environment"
-        msg_xyz = template.format(req="xyz")
-        msg_abc = template.format(req="abc")
-        assert msg_xyz in str(cap_logger.out)
-        assert msg_abc in str(cap_logger.out)
-
-    def test_conditional_block_requirements_save_load(self, tmp_path):
-        pipe = self.get_dummy_conditional_block_pipe()
-        pipe.save_pretrained(str(tmp_path))
-
-        config_path = tmp_path / "modular_config.json"
-        with open(config_path, "r") as f:
-            config = json.load(f)
-
-        assert "requirements" in config
-        expected_requirements = {
-            "xyz": ">=0.8.0",
-            "abc": ">=10.0.0",
-            "transformers": ">=4.44.0",
-            "diffusers": ">=0.2.0",
-        }
-        assert expected_requirements == config["requirements"]
-
-    def test_loop_block_requirements_save_load(self, tmp_path):
-        pipe = self.get_dummy_loop_block_pipe()
-        pipe.save_pretrained(str(tmp_path))
-
-        config_path = tmp_path / "modular_config.json"
-        with open(config_path, "r") as f:
-            config = json.load(f)
-
-        assert "requirements" in config
-        expected_requirements = {
-            "xyz": ">=0.8.0",
-            "abc": ">=10.0.0",
-            "transformers": ">=4.44.0",
-            "diffusers": ">=0.2.0",
-        }
-        assert expected_requirements == config["requirements"]
-
-
 class TestModularModelCardContent:
     def create_mock_block(self, name="TestBlock", description="Test block description"):
         class MockBlock:

tests/modular_pipelines/test_modular_pipelines_custom_blocks.py

@@ -24,14 +24,18 @@ import torch
 from diffusers import FluxTransformer2DModel
 from diffusers.modular_pipelines import (
     ComponentSpec,
+    ConditionalPipelineBlocks,
     InputParam,
+    LoopSequentialPipelineBlocks,
     ModularPipelineBlocks,
     OutputParam,
     PipelineState,
+    SequentialPipelineBlocks,
     WanModularPipeline,
 )
+from diffusers.utils import logging
 
-from ..testing_utils import nightly, require_torch, slow
+from ..testing_utils import CaptureLogger, nightly, require_torch, slow
 
 
 class DummyCustomBlockSimple(ModularPipelineBlocks):
@@ -354,6 +358,117 @@ class TestModularCustomBlocks:
         assert output_prompt.startswith("Modular diffusers + ")
 
 
+class TestCustomBlockRequirements:
+    def get_dummy_block_pipe(self):
+        class DummyBlockOne:
+            # keep two arbitrary deps so that we can test warnings.
+            _requirements = {"xyz": ">=0.8.0", "abc": ">=10.0.0"}
+
+        class DummyBlockTwo:
+            # keep two dependencies that will be available during testing.
+            _requirements = {"transformers": ">=4.44.0", "diffusers": ">=0.2.0"}
+
+        pipe = SequentialPipelineBlocks.from_blocks_dict(
+            {"dummy_block_one": DummyBlockOne, "dummy_block_two": DummyBlockTwo}
+        )
+        return pipe
+
+    def get_dummy_conditional_block_pipe(self):
+        class DummyBlockOne:
+            _requirements = {"xyz": ">=0.8.0", "abc": ">=10.0.0"}
+
+        class DummyBlockTwo:
+            _requirements = {"transformers": ">=4.44.0", "diffusers": ">=0.2.0"}
+
+        class DummyConditionalBlocks(ConditionalPipelineBlocks):
+            block_classes = [DummyBlockOne, DummyBlockTwo]
+            block_names = ["block_one", "block_two"]
+            block_trigger_inputs = []
+
+            def select_block(self, **kwargs):
+                return "block_one"
+
+        return DummyConditionalBlocks()
+
+    def get_dummy_loop_block_pipe(self):
+        class DummyBlockOne:
+            _requirements = {"xyz": ">=0.8.0", "abc": ">=10.0.0"}
+
+        class DummyBlockTwo:
+            _requirements = {"transformers": ">=4.44.0", "diffusers": ">=0.2.0"}
+
+        return LoopSequentialPipelineBlocks.from_blocks_dict({"block_one": DummyBlockOne, "block_two": DummyBlockTwo})
+
+    def test_sequential_block_requirements_save_load(self, tmp_path):
+        pipe = self.get_dummy_block_pipe()
+        pipe.save_pretrained(str(tmp_path))
+
+        config_path = tmp_path / "modular_config.json"
+
+        with open(config_path, "r") as f:
+            config = json.load(f)
+
+        assert "requirements" in config
+        requirements = config["requirements"]
+
+        expected_requirements = {
+            "xyz": ">=0.8.0",
+            "abc": ">=10.0.0",
+            "transformers": ">=4.44.0",
+            "diffusers": ">=0.2.0",
+        }
+        assert expected_requirements == requirements
+
+    def test_sequential_block_requirements_warnings(self, tmp_path):
+        pipe = self.get_dummy_block_pipe()
+
+        logger = logging.get_logger("diffusers.modular_pipelines.modular_pipeline_utils")
+        logger.setLevel(30)
+
+        with CaptureLogger(logger) as cap_logger:
+            pipe.save_pretrained(str(tmp_path))
+
+        template = "{req} was specified in the requirements but wasn't found in the current environment"
+        msg_xyz = template.format(req="xyz")
+        msg_abc = template.format(req="abc")
+        assert msg_xyz in str(cap_logger.out)
+        assert msg_abc in str(cap_logger.out)
+
+    def test_conditional_block_requirements_save_load(self, tmp_path):
+        pipe = self.get_dummy_conditional_block_pipe()
+        pipe.save_pretrained(str(tmp_path))
+
+        config_path = tmp_path / "modular_config.json"
+        with open(config_path, "r") as f:
+            config = json.load(f)
+
+        assert "requirements" in config
+        expected_requirements = {
+            "xyz": ">=0.8.0",
+            "abc": ">=10.0.0",
+            "transformers": ">=4.44.0",
+            "diffusers": ">=0.2.0",
+        }
+        assert expected_requirements == config["requirements"]
+
+    def test_loop_block_requirements_save_load(self, tmp_path):
+        pipe = self.get_dummy_loop_block_pipe()
+        pipe.save_pretrained(str(tmp_path))
+
+        config_path = tmp_path / "modular_config.json"
+        with open(config_path, "r") as f:
+            config = json.load(f)
+
+        assert "requirements" in config
+        expected_requirements = {
+            "xyz": ">=0.8.0",
+            "abc": ">=10.0.0",
+            "transformers": ">=4.44.0",
+            "diffusers": ">=0.2.0",
+        }
+        assert expected_requirements == config["requirements"]
+
+
 @slow
 @nightly
 @require_torch