[docs] validation for custom blocks (#13156)

validation

.github/workflows/pr_modular_tests.yml

@@ -117,7 +117,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality]"
+        uv pip install -e ".[quality,test]"
         #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
         uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps

.github/workflows/pr_tests.yml

@@ -114,7 +114,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality]"
+        uv pip install -e ".[quality,test]"
         #uv pip uninstall transformers huggingface_hub && uv pip install --prerelease allow -U transformers@git+https://github.com/huggingface/transformers.git
         uv pip uninstall transformers huggingface_hub && uv pip install transformers==4.57.1
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
@@ -191,7 +191,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality]"
+        uv pip install -e ".[quality,test]"
 
     - name: Environment
       run: |
@@ -242,7 +242,7 @@ jobs:
 
     - name: Install dependencies
       run: |
-        uv pip install -e ".[quality]"
+        uv pip install -e ".[quality,test]"
         # TODO (sayakpaul, DN6): revisit `--no-deps`
         uv pip install -U peft@git+https://github.com/huggingface/peft.git --no-deps
         uv pip install -U tokenizers

.github/workflows/pr_tests_gpu.yml

@@ -199,6 +199,11 @@ jobs:
 
     - name: Install dependencies
       run: |
+        # Install pkgs which depend on setuptools<81 for pkg_resources first with no build isolation
+        uv pip install pip==25.2 setuptools==80.10.2
+        uv pip install --no-build-isolation k-diffusion==0.0.12
+        uv pip install --upgrade pip setuptools
+        # Install the rest as normal
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git

.github/workflows/push_tests.yml

@@ -126,6 +126,11 @@ jobs:
 
     - name: Install dependencies
       run: |
+        # Install pkgs which depend on setuptools<81 for pkg_resources first with no build isolation
+        uv pip install pip==25.2 setuptools==80.10.2
+        uv pip install --no-build-isolation k-diffusion==0.0.12
+        uv pip install --upgrade pip setuptools
+        # Install the rest as normal
         uv pip install -e ".[quality]"
         uv pip install peft@git+https://github.com/huggingface/peft.git
         uv pip uninstall accelerate && uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git

.github/workflows/push_tests_mps.yml

@@ -41,7 +41,7 @@ jobs:
       shell: arch -arch arm64 bash {0}
       run: |
         ${CONDA_RUN} python -m pip install --upgrade pip uv
-        ${CONDA_RUN} python -m uv pip install -e ".[quality]"
+        ${CONDA_RUN} python -m uv pip install -e ".[quality,test]"
         ${CONDA_RUN} python -m uv pip install torch torchvision torchaudio
         ${CONDA_RUN} python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
         ${CONDA_RUN} python -m uv pip install transformers --upgrade

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

@@ -29,7 +29,7 @@ Qwen-Image comes in the following variants:
 | Qwen-Image-Edit Plus | [Qwen/Qwen-Image-Edit-2509](https://huggingface.co/Qwen/Qwen-Image-Edit-2509) |
 
 > [!TIP]
-> See the [Caching](../../optimization/cache) guide to speed up inference by storing and reusing intermediate outputs.
+> [Caching](../../optimization/cache) may also speed up inference by storing and reusing intermediate outputs.
 
 ## LoRA for faster inference
 
@@ -190,12 +190,6 @@ For detailed benchmark scripts and results, see [this gist](https://gist.github.
   - all
   - __call__
 
-## QwenImageLayeredPipeline
-
-[[autodoc]] QwenImageLayeredPipeline
-  - all
-  - __call__
-
 ## QwenImagePipelineOutput
 
 [[autodoc]] pipelines.qwenimage.pipeline_output.QwenImagePipelineOutput

docs/source/en/modular_diffusers/custom_blocks.md

@@ -332,4 +332,49 @@ Make your custom block work with Mellon's visual interface. See the [Mellon Cust
 Browse the [Modular Diffusers Custom Blocks](https://huggingface.co/collections/diffusers/modular-diffusers-custom-blocks) collection for inspiration and ready-to-use blocks.
 
 </hfoption>
-</hfoptions>
+</hfoptions>
+
+## Dependencies
+
+Declaring package dependencies in custom blocks prevents runtime import errors later on. Diffusers validates the dependencies and returns a warning if a package is missing or incompatible.
+
+Set a `_requirements` attribute in your block class, mapping package names to version specifiers.
+
+```py
+from diffusers.modular_pipelines import PipelineBlock
+
+class MyCustomBlock(PipelineBlock):
+    _requirements = {
+        "transformers": ">=4.44.0",
+        "sentencepiece": ">=0.2.0"
+    }
+```
+
+When there are blocks with different requirements, Diffusers merges their requirements.
+
+```py
+from diffusers.modular_pipelines import SequentialPipelineBlocks
+
+class BlockA(PipelineBlock):
+    _requirements = {"transformers": ">=4.44.0"}
+    # ...
+
+class BlockB(PipelineBlock):
+    _requirements = {"sentencepiece": ">=0.2.0"}
+    # ...
+
+pipe = SequentialPipelineBlocks.from_blocks_dict({
+    "block_a": BlockA,
+    "block_b": BlockB,
+})
+```
+
+When this block is saved with [`~ModularPipeline.save_pretrained`], the requirements are saved to the `modular_config.json` file. When this block is loaded, Diffusers checks each requirement against the current environment. If there is a mismatch or a package isn't found, Diffusers returns the following warning.
+
+```md
+# missing package
+xyz-package was specified in the requirements but wasn't found in the current environment.
+
+# version mismatch
+xyz requirement 'specific-version' is not satisfied by the installed version 'actual-version'. Things might work unexpected.
+```

src/diffusers/commands/custom_blocks.py

@@ -89,8 +89,6 @@ class CustomBlocksCommand(BaseDiffusersCLICommand):
         # automap = self._create_automap(parent_class=parent_class, child_class=child_class)
         # with open(CONFIG, "w") as f:
         #     json.dump(automap, f)
-        with open("requirements.txt", "w") as f:
-            f.write("")
 
     def _choose_block(self, candidates, chosen=None):
         for cls, base in candidates:

src/diffusers/loaders/lora_pipeline.py

@@ -5472,10 +5472,6 @@ class Flux2LoraLoaderMixin(LoraBaseMixin):
             logger.warning(warn_msg)
             state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}
 
-        is_peft_format = any(k.startswith("base_model.model.") for k in state_dict)
-        if is_peft_format:
-            state_dict = {k.replace("base_model.model.", "diffusion_model."): v for k, v in state_dict.items()}
-
         is_ai_toolkit = any(k.startswith("diffusion_model.") for k in state_dict)
         if is_ai_toolkit:
             state_dict = _convert_non_diffusers_flux2_lora_to_diffusers(state_dict)

src/diffusers/models/transformers/transformer_flux2.py

@@ -424,7 +424,7 @@ class Flux2SingleTransformerBlock(nn.Module):
         self,
         hidden_states: torch.Tensor,
         encoder_hidden_states: torch.Tensor | None,
-        temb_mod: torch.Tensor,
+        temb_mod_params: tuple[torch.Tensor, torch.Tensor, torch.Tensor],
         image_rotary_emb: tuple[torch.Tensor, torch.Tensor] | None = None,
         joint_attention_kwargs: dict[str, Any] | None = None,
         split_hidden_states: bool = False,
@@ -436,7 +436,7 @@ class Flux2SingleTransformerBlock(nn.Module):
             text_seq_len = encoder_hidden_states.shape[1]
             hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
 
-        mod_shift, mod_scale, mod_gate = Flux2Modulation.split(temb_mod, 1)[0]
+        mod_shift, mod_scale, mod_gate = temb_mod_params
 
         norm_hidden_states = self.norm(hidden_states)
         norm_hidden_states = (1 + mod_scale) * norm_hidden_states + mod_shift
@@ -498,18 +498,16 @@ class Flux2TransformerBlock(nn.Module):
         self,
         hidden_states: torch.Tensor,
         encoder_hidden_states: torch.Tensor,
-        temb_mod_img: torch.Tensor,
-        temb_mod_txt: torch.Tensor,
+        temb_mod_params_img: tuple[tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...],
+        temb_mod_params_txt: tuple[tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...],
         image_rotary_emb: tuple[torch.Tensor, torch.Tensor] | None = None,
         joint_attention_kwargs: dict[str, Any] | None = None,
     ) -> tuple[torch.Tensor, torch.Tensor]:
         joint_attention_kwargs = joint_attention_kwargs or {}
 
         # Modulation parameters shape: [1, 1, self.dim]
-        (shift_msa, scale_msa, gate_msa), (shift_mlp, scale_mlp, gate_mlp) = Flux2Modulation.split(temb_mod_img, 2)
-        (c_shift_msa, c_scale_msa, c_gate_msa), (c_shift_mlp, c_scale_mlp, c_gate_mlp) = Flux2Modulation.split(
-            temb_mod_txt, 2
-        )
+        (shift_msa, scale_msa, gate_msa), (shift_mlp, scale_mlp, gate_mlp) = temb_mod_params_img
+        (c_shift_msa, c_scale_msa, c_gate_msa), (c_shift_mlp, c_scale_mlp, c_gate_mlp) = temb_mod_params_txt
 
         # Img stream
         norm_hidden_states = self.norm1(hidden_states)
@@ -629,19 +627,15 @@ class Flux2Modulation(nn.Module):
         self.linear = nn.Linear(dim, dim * 3 * self.mod_param_sets, bias=bias)
         self.act_fn = nn.SiLU()
 
-    def forward(self, temb: torch.Tensor) -> torch.Tensor:
+    def forward(self, temb: torch.Tensor) -> tuple[tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...]:
         mod = self.act_fn(temb)
         mod = self.linear(mod)
-        return mod
 
-    @staticmethod
-    # split inside the transformer blocks, to avoid passing tuples into checkpoints https://github.com/huggingface/diffusers/issues/12776
-    def split(mod: torch.Tensor, mod_param_sets: int) -> tuple[tuple[torch.Tensor, torch.Tensor, torch.Tensor], ...]:
         if mod.ndim == 2:
             mod = mod.unsqueeze(1)
-        mod_params = torch.chunk(mod, 3 * mod_param_sets, dim=-1)
+        mod_params = torch.chunk(mod, 3 * self.mod_param_sets, dim=-1)
         # Return tuple of 3-tuples of modulation params shift/scale/gate
-        return tuple(mod_params[3 * i : 3 * (i + 1)] for i in range(mod_param_sets))
+        return tuple(mod_params[3 * i : 3 * (i + 1)] for i in range(self.mod_param_sets))
 
 
 class Flux2Transformer2DModel(
@@ -830,7 +824,7 @@ class Flux2Transformer2DModel(
 
         double_stream_mod_img = self.double_stream_modulation_img(temb)
         double_stream_mod_txt = self.double_stream_modulation_txt(temb)
-        single_stream_mod = self.single_stream_modulation(temb)
+        single_stream_mod = self.single_stream_modulation(temb)[0]
 
         # 2. Input projection for image (hidden_states) and conditioning text (encoder_hidden_states)
         hidden_states = self.x_embedder(hidden_states)
@@ -867,8 +861,8 @@ class Flux2Transformer2DModel(
                 encoder_hidden_states, hidden_states = block(
                     hidden_states=hidden_states,
                     encoder_hidden_states=encoder_hidden_states,
-                    temb_mod_img=double_stream_mod_img,
-                    temb_mod_txt=double_stream_mod_txt,
+                    temb_mod_params_img=double_stream_mod_img,
+                    temb_mod_params_txt=double_stream_mod_txt,
                     image_rotary_emb=concat_rotary_emb,
                     joint_attention_kwargs=joint_attention_kwargs,
                 )
@@ -890,7 +884,7 @@ class Flux2Transformer2DModel(
                 hidden_states = block(
                     hidden_states=hidden_states,
                     encoder_hidden_states=None,
-                    temb_mod=single_stream_mod,
+                    temb_mod_params=single_stream_mod,
                     image_rotary_emb=concat_rotary_emb,
                     joint_attention_kwargs=joint_attention_kwargs,
                 )

src/diffusers/modular_pipelines/modular_pipeline.py

@@ -40,6 +40,7 @@ from .modular_pipeline_utils import (
     InputParam,
     InsertableDict,
     OutputParam,
+    _validate_requirements,
     combine_inputs,
     combine_outputs,
     format_components,
@@ -290,6 +291,7 @@ class ModularPipelineBlocks(ConfigMixin, PushToHubMixin):
 
     config_name = "modular_config.json"
     model_name = None
+    _requirements: dict[str, str] | None = None
     _workflow_map = None
 
     @classmethod
@@ -404,6 +406,9 @@ class ModularPipelineBlocks(ConfigMixin, PushToHubMixin):
                 "Selected model repository does not happear to have any custom code or does not have a valid `config.json` file."
             )
 
+        if "requirements" in config and config["requirements"] is not None:
+            _ = _validate_requirements(config["requirements"])
+
         class_ref = config["auto_map"][cls.__name__]
         module_file, class_name = class_ref.split(".")
         module_file = module_file + ".py"
@@ -428,8 +433,13 @@ class ModularPipelineBlocks(ConfigMixin, PushToHubMixin):
         module = full_mod.rsplit(".", 1)[-1].replace("__dynamic__", "")
         parent_module = self.save_pretrained.__func__.__qualname__.split(".", 1)[0]
         auto_map = {f"{parent_module}": f"{module}.{cls_name}"}
-
         self.register_to_config(auto_map=auto_map)
+
+        # resolve requirements
+        requirements = _validate_requirements(getattr(self, "_requirements", None))
+        if requirements:
+            self.register_to_config(requirements=requirements)
+
         self.save_config(save_directory=save_directory, push_to_hub=push_to_hub, **kwargs)
         config = dict(self.config)
         self._internal_dict = FrozenDict(config)
@@ -1240,6 +1250,14 @@ class SequentialPipelineBlocks(ModularPipelineBlocks):
             expected_configs=self.expected_configs,
         )
 
+    @property
+    def _requirements(self) -> dict[str, str]:
+        requirements = {}
+        for block_name, block in self.sub_blocks.items():
+            if getattr(block, "_requirements", None):
+                requirements[block_name] = block._requirements
+        return requirements
+
 
 class LoopSequentialPipelineBlocks(ModularPipelineBlocks):
     """

src/diffusers/modular_pipelines/modular_pipeline_utils.py

@@ -22,10 +22,12 @@ from typing import Any, Literal, Type, Union, get_args, get_origin
 
 import PIL.Image
 import torch
+from packaging.specifiers import InvalidSpecifier, SpecifierSet
 
 from ..configuration_utils import ConfigMixin, FrozenDict
 from ..loaders.single_file_utils import _is_single_file_path_or_url
 from ..utils import DIFFUSERS_LOAD_ID_FIELDS, is_torch_available, logging
+from ..utils.import_utils import _is_package_available
 
 
 if is_torch_available():
@@ -972,6 +974,89 @@ def make_doc_string(
     return output
 
 
+def _validate_requirements(reqs):
+    if reqs is None:
+        normalized_reqs = {}
+    else:
+        if not isinstance(reqs, dict):
+            raise ValueError(
+                "Requirements must be provided as a dictionary mapping package names to version specifiers."
+            )
+        normalized_reqs = _normalize_requirements(reqs)
+
+    if not normalized_reqs:
+        return {}
+
+    final: dict[str, str] = {}
+    for req, specified_ver in normalized_reqs.items():
+        req_available, req_actual_ver = _is_package_available(req)
+        if not req_available:
+            logger.warning(f"{req} was specified in the requirements but wasn't found in the current environment.")
+
+        if specified_ver:
+            try:
+                specifier = SpecifierSet(specified_ver)
+            except InvalidSpecifier as err:
+                raise ValueError(f"Requirement specifier '{specified_ver}' for {req} is invalid.") from err
+
+            if req_actual_ver == "N/A":
+                logger.warning(
+                    f"Version of {req} could not be determined to validate requirement '{specified_ver}'. Things might work unexpected."
+                )
+            elif not specifier.contains(req_actual_ver, prereleases=True):
+                logger.warning(
+                    f"{req} requirement '{specified_ver}' is not satisfied by the installed version {req_actual_ver}. Things might work unexpected."
+                )
+
+        final[req] = specified_ver
+
+    return final
+
+
+def _normalize_requirements(reqs):
+    if not reqs:
+        return {}
+
+    normalized: "OrderedDict[str, str]" = OrderedDict()
+
+    def _accumulate(mapping: dict[str, Any]):
+        for pkg, spec in mapping.items():
+            if isinstance(spec, dict):
+                # This is recursive because blocks are composable. This way, we can merge requirements
+                # from multiple blocks.
+                _accumulate(spec)
+                continue
+
+            pkg_name = str(pkg).strip()
+            if not pkg_name:
+                raise ValueError("Requirement package name cannot be empty.")
+
+            spec_str = "" if spec is None else str(spec).strip()
+            if spec_str and not spec_str.startswith(("<", ">", "=", "!", "~")):
+                spec_str = f"=={spec_str}"
+
+            existing_spec = normalized.get(pkg_name)
+            if existing_spec is not None:
+                if not existing_spec and spec_str:
+                    normalized[pkg_name] = spec_str
+                elif existing_spec and spec_str and existing_spec != spec_str:
+                    try:
+                        combined_spec = SpecifierSet(",".join(filter(None, [existing_spec, spec_str])))
+                    except InvalidSpecifier:
+                        logger.warning(
+                            f"Conflicting requirements for '{pkg_name}' detected: '{existing_spec}' vs '{spec_str}'. Keeping '{existing_spec}'."
+                        )
+                    else:
+                        normalized[pkg_name] = str(combined_spec)
+                continue
+
+            normalized[pkg_name] = spec_str
+
+    _accumulate(reqs)
+
+    return normalized
+
+
 def combine_inputs(*named_input_lists: list[tuple[str, list[InputParam]]]) -> list[InputParam]:
     """
     Combines multiple lists of InputParam objects from different blocks. For duplicate inputs, updates only if current

src/diffusers/pipelines/prx/pipeline_prx.py

@@ -18,6 +18,7 @@ import re
 import urllib.parse as ul
 from typing import Callable
 
+import ftfy
 import torch
 from transformers import (
     AutoTokenizer,
@@ -33,13 +34,13 @@ from diffusers.models.transformers.transformer_prx import PRXTransformer2DModel
 from diffusers.pipelines.pipeline_utils import DiffusionPipeline
 from diffusers.pipelines.prx.pipeline_output import PRXPipelineOutput
 from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
-from diffusers.utils import is_ftfy_available, logging, replace_example_docstring
+from diffusers.utils import (
+    logging,
+    replace_example_docstring,
+)
 from diffusers.utils.torch_utils import randn_tensor
 
 
-if is_ftfy_available():
-    import ftfy
-
 DEFAULT_RESOLUTION = 512
 
 ASPECT_RATIO_256_BIN = {

src/diffusers/quantizers/gguf/utils.py

@@ -516,9 +516,6 @@ def dequantize_gguf_tensor(tensor):
 
     block_size, type_size = GGML_QUANT_SIZES[quant_type]
 
-    # Conver to plain tensor to avoid unnecessary __torch_function__ overhead.
-    tensor = tensor.as_tensor()
-
     tensor = tensor.view(torch.uint8)
     shape = _quant_shape_from_byte_shape(tensor.shape, type_size, block_size)
 
@@ -528,7 +525,7 @@ def dequantize_gguf_tensor(tensor):
     dequant = dequant_fn(blocks, block_size, type_size)
     dequant = dequant.reshape(shape)
 
-    return dequant
+    return dequant.as_tensor()
 
 
 class GGUFParameter(torch.nn.Parameter):

src/diffusers/schedulers/scheduling_flow_match_lcm.py

@@ -14,7 +14,6 @@
 
 import math
 from dataclasses import dataclass
-from typing import Literal
 
 import numpy as np
 import torch
@@ -42,7 +41,7 @@ class FlowMatchLCMSchedulerOutput(BaseOutput):
             denoising loop.
     """
 
-    prev_sample: torch.Tensor
+    prev_sample: torch.FloatTensor
 
 
 class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
@@ -80,11 +79,11 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         use_beta_sigmas (`bool`, defaults to False):
             Whether to use beta sigmas for step sizes in the noise schedule during sampling.
         time_shift_type (`str`, defaults to "exponential"):
-            The type of dynamic resolution-dependent timestep shifting to apply.
-        scale_factors (`list[float]`, *optional*, defaults to `None`):
+            The type of dynamic resolution-dependent timestep shifting to apply. Either "exponential" or "linear".
+        scale_factors ('list', defaults to None)
             It defines how to scale the latents at which predictions are made.
-        upscale_mode (`str`, *optional*, defaults to "bicubic"):
-            Upscaling method, applied if scale-wise generation is considered.
+        upscale_mode ('str', defaults to 'bicubic')
+            Upscaling method, applied if scale-wise generation is considered
     """
 
     _compatibles = []
@@ -102,33 +101,16 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         max_image_seq_len: int = 4096,
         invert_sigmas: bool = False,
         shift_terminal: float | None = None,
-        use_karras_sigmas: bool | None = False,
-        use_exponential_sigmas: bool | None = False,
-        use_beta_sigmas: bool | None = False,
-        time_shift_type: Literal["exponential", "linear"] = "exponential",
+        use_karras_sigmas: bool = False,
+        use_exponential_sigmas: bool = False,
+        use_beta_sigmas: bool = False,
+        time_shift_type: str = "exponential",
         scale_factors: list[float] | None = None,
-        upscale_mode: Literal[
-            "nearest",
-            "linear",
-            "bilinear",
-            "bicubic",
-            "trilinear",
-            "area",
-            "nearest-exact",
-        ] = "bicubic",
+        upscale_mode: str = "bicubic",
     ):
         if self.config.use_beta_sigmas and not is_scipy_available():
             raise ImportError("Make sure to install scipy if you want to use beta sigmas.")
-        if (
-            sum(
-                [
-                    self.config.use_beta_sigmas,
-                    self.config.use_exponential_sigmas,
-                    self.config.use_karras_sigmas,
-                ]
-            )
-            > 1
-        ):
+        if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1:
             raise ValueError(
                 "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used."
             )
@@ -180,7 +162,7 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         return self._begin_index
 
     # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
-    def set_begin_index(self, begin_index: int = 0) -> None:
+    def set_begin_index(self, begin_index: int = 0):
         """
         Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
 
@@ -190,18 +172,18 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         """
         self._begin_index = begin_index
 
-    def set_shift(self, shift: float) -> None:
+    def set_shift(self, shift: float):
         self._shift = shift
 
-    def set_scale_factors(self, scale_factors: list[float], upscale_mode: str) -> None:
+    def set_scale_factors(self, scale_factors: list, upscale_mode):
         """
         Sets scale factors for a scale-wise generation regime.
 
         Args:
-            scale_factors (`list[float]`):
-                The scale factors for each step.
+            scale_factors (`list`):
+                The scale factors for each step
             upscale_mode (`str`):
-                Upscaling method.
+                Upscaling method
         """
         self._scale_factors = scale_factors
         self._upscale_mode = upscale_mode
@@ -256,18 +238,16 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
 
         return sample
 
-    def _sigma_to_t(self, sigma: float | torch.FloatTensor) -> float | torch.FloatTensor:
+    def _sigma_to_t(self, sigma):
         return sigma * self.config.num_train_timesteps
 
-    def time_shift(
-        self, mu: float, sigma: float, t: float | np.ndarray | torch.Tensor
-    ) -> float | np.ndarray | torch.Tensor:
+    def time_shift(self, mu: float, sigma: float, t: torch.Tensor):
         if self.config.time_shift_type == "exponential":
             return self._time_shift_exponential(mu, sigma, t)
         elif self.config.time_shift_type == "linear":
             return self._time_shift_linear(mu, sigma, t)
 
-    def stretch_shift_to_terminal(self, t: np.ndarray | torch.Tensor) -> np.ndarray | torch.Tensor:
+    def stretch_shift_to_terminal(self, t: torch.Tensor) -> torch.Tensor:
         r"""
         Stretches and shifts the timestep schedule to ensure it terminates at the configured `shift_terminal` config
         value.
@@ -276,13 +256,12 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         https://github.com/Lightricks/LTX-Video/blob/a01a171f8fe3d99dce2728d60a73fecf4d4238ae/ltx_video/schedulers/rf.py#L51
 
         Args:
-            t (`torch.Tensor` or `np.ndarray`):
-                A tensor or numpy array of timesteps to be stretched and shifted.
+            t (`torch.Tensor`):
+                A tensor of timesteps to be stretched and shifted.
 
         Returns:
-            `torch.Tensor` or `np.ndarray`:
-                A tensor or numpy array of adjusted timesteps such that the final value equals
-                `self.config.shift_terminal`.
+            `torch.Tensor`:
+                A tensor of adjusted timesteps such that the final value equals `self.config.shift_terminal`.
         """
         one_minus_z = 1 - t
         scale_factor = one_minus_z[-1] / (1 - self.config.shift_terminal)
@@ -291,12 +270,12 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
 
     def set_timesteps(
         self,
-        num_inference_steps: int | None = None,
-        device: str | torch.device | None = None,
+        num_inference_steps: int = None,
+        device: str | torch.device = None,
         sigmas: list[float] | None = None,
-        mu: float | None = None,
+        mu: float = None,
         timesteps: list[float] | None = None,
-    ) -> None:
+    ):
         """
         Sets the discrete timesteps used for the diffusion chain (to be run before inference).
 
@@ -338,45 +317,43 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         is_timesteps_provided = timesteps is not None
 
         if is_timesteps_provided:
-            timesteps = np.array(timesteps).astype(np.float32)  # type: ignore
+            timesteps = np.array(timesteps).astype(np.float32)
 
         if sigmas is None:
             if timesteps is None:
-                timesteps = np.linspace(  # type: ignore
-                    self._sigma_to_t(self.sigma_max),
-                    self._sigma_to_t(self.sigma_min),
-                    num_inference_steps,
+                timesteps = np.linspace(
+                    self._sigma_to_t(self.sigma_max), self._sigma_to_t(self.sigma_min), num_inference_steps
                 )
-            sigmas = timesteps / self.config.num_train_timesteps  # type: ignore
+            sigmas = timesteps / self.config.num_train_timesteps
         else:
-            sigmas = np.array(sigmas).astype(np.float32)  # type: ignore
+            sigmas = np.array(sigmas).astype(np.float32)
             num_inference_steps = len(sigmas)
 
         # 2. Perform timestep shifting. Either no shifting is applied, or resolution-dependent shifting of
         #    "exponential" or "linear" type is applied
         if self.config.use_dynamic_shifting:
-            sigmas = self.time_shift(mu, 1.0, sigmas)  # type: ignore
+            sigmas = self.time_shift(mu, 1.0, sigmas)
         else:
-            sigmas = self.shift * sigmas / (1 + (self.shift - 1) * sigmas)  # type: ignore
+            sigmas = self.shift * sigmas / (1 + (self.shift - 1) * sigmas)
 
         # 3. If required, stretch the sigmas schedule to terminate at the configured `shift_terminal` value
         if self.config.shift_terminal:
-            sigmas = self.stretch_shift_to_terminal(sigmas)  # type: ignore
+            sigmas = self.stretch_shift_to_terminal(sigmas)
 
         # 4. If required, convert sigmas to one of karras, exponential, or beta sigma schedules
         if self.config.use_karras_sigmas:
-            sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)  # type: ignore
+            sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
         elif self.config.use_exponential_sigmas:
-            sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps)  # type: ignore
+            sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
         elif self.config.use_beta_sigmas:
-            sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)  # type: ignore
+            sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
 
         # 5. Convert sigmas and timesteps to tensors and move to specified device
-        sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)  # type: ignore
+        sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)
         if not is_timesteps_provided:
-            timesteps = sigmas * self.config.num_train_timesteps  # type: ignore
+            timesteps = sigmas * self.config.num_train_timesteps
         else:
-            timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32, device=device)  # type: ignore
+            timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32, device=device)
 
         # 6. Append the terminal sigma value.
         #    If a model requires inverted sigma schedule for denoising but timesteps without inversion, the
@@ -393,11 +370,7 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         self._step_index = None
         self._begin_index = None
 
-    def index_for_timestep(
-        self,
-        timestep: float | torch.Tensor,
-        schedule_timesteps: torch.Tensor | None = None,
-    ) -> int:
+    def index_for_timestep(self, timestep, schedule_timesteps=None):
         if schedule_timesteps is None:
             schedule_timesteps = self.timesteps
 
@@ -409,9 +382,9 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         # case we start in the middle of the denoising schedule (e.g. for image-to-image)
         pos = 1 if len(indices) > 1 else 0
 
-        return int(indices[pos].item())
+        return indices[pos].item()
 
-    def _init_step_index(self, timestep: float | torch.Tensor) -> None:
+    def _init_step_index(self, timestep):
         if self.begin_index is None:
             if isinstance(timestep, torch.Tensor):
                 timestep = timestep.to(self.timesteps.device)
@@ -486,12 +459,7 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
                 size = [round(self._scale_factors[self._step_index] * size) for size in self._init_size]
                 x0_pred = torch.nn.functional.interpolate(x0_pred, size=size, mode=self._upscale_mode)
 
-        noise = randn_tensor(
-            x0_pred.shape,
-            generator=generator,
-            device=x0_pred.device,
-            dtype=x0_pred.dtype,
-        )
+        noise = randn_tensor(x0_pred.shape, generator=generator, device=x0_pred.device, dtype=x0_pred.dtype)
         prev_sample = (1 - sigma_next) * x0_pred + sigma_next * noise
 
         # upon completion increase step index by one
@@ -505,7 +473,7 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         return FlowMatchLCMSchedulerOutput(prev_sample=prev_sample)
 
     # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
-    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor:
+    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor:
         """
         Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative
         Models](https://huggingface.co/papers/2206.00364).
@@ -626,15 +594,11 @@ class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin):
         )
         return sigmas
 
-    def _time_shift_exponential(
-        self, mu: float, sigma: float, t: float | np.ndarray | torch.Tensor
-    ) -> float | np.ndarray | torch.Tensor:
+    def _time_shift_exponential(self, mu, sigma, t):
         return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
 
-    def _time_shift_linear(
-        self, mu: float, sigma: float, t: float | np.ndarray | torch.Tensor
-    ) -> float | np.ndarray | torch.Tensor:
+    def _time_shift_linear(self, mu, sigma, t):
         return mu / (mu + (1 / t - 1) ** sigma)
 
-    def __len__(self) -> int:
+    def __len__(self):
         return self.config.num_train_timesteps

tests/models/testing_utils/lora.py

@@ -375,7 +375,7 @@ class LoraHotSwappingForModelTesterMixin:
             # additionally check if dynamic compilation works.
             if different_shapes is not None:
                 for height, width in different_shapes:
-                    new_inputs_dict = self.get_dummy_inputs(height=height, width=width)
+                    new_inputs_dict = self.prepare_dummy_input(height=height, width=width)
                     _ = model(**new_inputs_dict)
             else:
                 output0_after = model(**inputs_dict)["sample"]
@@ -390,7 +390,7 @@ class LoraHotSwappingForModelTesterMixin:
         with torch.inference_mode():
             if different_shapes is not None:
                 for height, width in different_shapes:
-                    new_inputs_dict = self.get_dummy_inputs(height=height, width=width)
+                    new_inputs_dict = self.prepare_dummy_input(height=height, width=width)
                     _ = model(**new_inputs_dict)
             else:
                 output1_after = model(**inputs_dict)["sample"]

tests/modular_pipelines/test_modular_pipelines_common.py

@@ -1,4 +1,6 @@
 import gc
+import json
+import os
 import tempfile
 from typing import Callable
 
@@ -8,6 +10,7 @@ import torch
 import diffusers
 from diffusers import AutoModel, ComponentsManager, ModularPipeline, ModularPipelineBlocks
 from diffusers.guiders import ClassifierFreeGuidance
+from diffusers.modular_pipelines import SequentialPipelineBlocks
 from diffusers.modular_pipelines.modular_pipeline_utils import (
     ComponentSpec,
     ConfigSpec,
@@ -17,7 +20,13 @@ from diffusers.modular_pipelines.modular_pipeline_utils import (
 )
 from diffusers.utils import logging
 
-from ..testing_utils import backend_empty_cache, numpy_cosine_similarity_distance, require_accelerator, torch_device
+from ..testing_utils import (
+    CaptureLogger,
+    backend_empty_cache,
+    numpy_cosine_similarity_distance,
+    require_accelerator,
+    torch_device,
+)
 
 
 class ModularPipelineTesterMixin:
@@ -400,6 +409,56 @@ 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 test_custom_requirements_save_load(self):
+        pipe = self.get_dummy_block_pipe()
+        with tempfile.TemporaryDirectory() as tmpdir:
+            pipe.save_pretrained(tmpdir)
+            config_path = os.path.join(tmpdir, "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_warnings(self):
+        pipe = self.get_dummy_block_pipe()
+        with tempfile.TemporaryDirectory() as tmpdir:
+            logger = logging.get_logger("diffusers.modular_pipelines.modular_pipeline_utils")
+            logger.setLevel(30)
+
+            with CaptureLogger(logger) as cap_logger:
+                pipe.save_pretrained(tmpdir)
+
+            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)
+
+
 class TestModularModelCardContent:
     def create_mock_block(self, name="TestBlock", description="Test block description"):
         class MockBlock:

tests/pipelines/prx/test_pipeline_prx.py

@@ -1,6 +1,7 @@
 import unittest
 
 import numpy as np
+import pytest
 import torch
 from transformers import AutoTokenizer
 from transformers.models.t5gemma.configuration_t5gemma import T5GemmaConfig, T5GemmaModuleConfig
@@ -10,11 +11,17 @@ from diffusers.models import AutoencoderDC, AutoencoderKL
 from diffusers.models.transformers.transformer_prx import PRXTransformer2DModel
 from diffusers.pipelines.prx.pipeline_prx import PRXPipeline
 from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_transformers_version
 
 from ..pipeline_params import TEXT_TO_IMAGE_PARAMS
 from ..test_pipelines_common import PipelineTesterMixin
 
 
+@pytest.mark.xfail(
+    condition=is_transformers_version(">", "4.57.1"),
+    reason="See https://github.com/huggingface/diffusers/pull/12456#issuecomment-3424228544",
+    strict=False,
+)
 class PRXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
     pipeline_class = PRXPipeline
     params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}