update

src/diffusers/modular_pipelines/modular_pipeline.py

@@ -327,6 +327,9 @@ class ModularPipelineBlocks(ConfigMixin, PushToHubMixin):
     config_name = "modular_config.json"
     model_name = None
 
+    def __init__(self):
+        self.sub_blocks = InsertableDict()
+
     @classmethod
     def _get_signature_keys(cls, obj):
         parameters = inspect.signature(obj.__init__).parameters
@@ -344,11 +347,6 @@ class ModularPipelineBlocks(ConfigMixin, PushToHubMixin):
     def expected_configs(self) -> List[ConfigSpec]:
         return []
 
-    @property
-    def intermediate_inputs(self) -> List[OutputParam]:
-        """List of intermediate output parameters. Must be implemented by subclasses."""
-        return []
-
     @property
     def intermediate_outputs(self) -> List[OutputParam]:
         """List of intermediate output parameters. Must be implemented by subclasses."""
@@ -1458,11 +1456,6 @@ class LoopSequentialPipelineBlocks(ModularPipelineBlocks):
         """List of input parameters. Must be implemented by subclasses."""
         return []
 
-    @property
-    def loop_intermediate_inputs(self) -> List[InputParam]:
-        """List of intermediate input parameters. Must be implemented by subclasses."""
-        return []
-
     @property
     def loop_intermediate_outputs(self) -> List[OutputParam]:
         """List of intermediate output parameters. Must be implemented by subclasses."""
@@ -1476,14 +1469,6 @@ class LoopSequentialPipelineBlocks(ModularPipelineBlocks):
                 input_names.append(input_param.name)
         return input_names
 
-    @property
-    def loop_required_intermediate_inputs(self) -> List[str]:
-        input_names = []
-        for input_param in self.loop_intermediate_inputs:
-            if input_param.required:
-                input_names.append(input_param.name)
-        return input_names
-
     # modified from SequentialPipelineBlocks to include loop_expected_components
     @property
     def expected_components(self):

src/diffusers/modular_pipelines/stable_diffusion_xl/before_denoise.py

@@ -27,7 +27,7 @@ from ...schedulers import EulerDiscreteScheduler
 from ...utils import logging
 from ...utils.torch_utils import randn_tensor, unwrap_module
 from ..modular_pipeline import (
-    PipelineBlock,
+    ModularPipelineBlocks,
     PipelineState,
 )
 from ..modular_pipeline_utils import ComponentSpec, ConfigSpec, InputParam, OutputParam
@@ -195,7 +195,7 @@ def prepare_latents_img2img(
     return latents
 
 
-class StableDiffusionXLInputStep(PipelineBlock):
+class StableDiffusionXLInputStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -394,7 +394,7 @@ class StableDiffusionXLInputStep(PipelineBlock):
         return components, state
 
 
-class StableDiffusionXLImg2ImgSetTimestepsStep(PipelineBlock):
+class StableDiffusionXLImg2ImgSetTimestepsStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -543,7 +543,7 @@ class StableDiffusionXLImg2ImgSetTimestepsStep(PipelineBlock):
         return components, state
 
 
-class StableDiffusionXLSetTimestepsStep(PipelineBlock):
+class StableDiffusionXLSetTimestepsStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -611,7 +611,7 @@ class StableDiffusionXLSetTimestepsStep(PipelineBlock):
         return components, state
 
 
-class StableDiffusionXLInpaintPrepareLatentsStep(PipelineBlock):
+class StableDiffusionXLInpaintPrepareLatentsStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -900,7 +900,7 @@ class StableDiffusionXLInpaintPrepareLatentsStep(PipelineBlock):
         return components, state
 
 
-class StableDiffusionXLImg2ImgPrepareLatentsStep(PipelineBlock):
+class StableDiffusionXLImg2ImgPrepareLatentsStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -981,7 +981,7 @@ class StableDiffusionXLImg2ImgPrepareLatentsStep(PipelineBlock):
         return components, state
 
 
-class StableDiffusionXLPrepareLatentsStep(PipelineBlock):
+class StableDiffusionXLPrepareLatentsStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -1092,7 +1092,7 @@ class StableDiffusionXLPrepareLatentsStep(PipelineBlock):
         return components, state
 
 
-class StableDiffusionXLImg2ImgPrepareAdditionalConditioningStep(PipelineBlock):
+class StableDiffusionXLImg2ImgPrepareAdditionalConditioningStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -1316,7 +1316,7 @@ class StableDiffusionXLImg2ImgPrepareAdditionalConditioningStep(PipelineBlock):
         return components, state
 
 
-class StableDiffusionXLPrepareAdditionalConditioningStep(PipelineBlock):
+class StableDiffusionXLPrepareAdditionalConditioningStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -1499,7 +1499,7 @@ class StableDiffusionXLPrepareAdditionalConditioningStep(PipelineBlock):
         return components, state
 
 
-class StableDiffusionXLControlNetInputStep(PipelineBlock):
+class StableDiffusionXLControlNetInputStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -1718,7 +1718,7 @@ class StableDiffusionXLControlNetInputStep(PipelineBlock):
         return components, state
 
 
-class StableDiffusionXLControlNetUnionInputStep(PipelineBlock):
+class StableDiffusionXLControlNetUnionInputStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property

src/diffusers/modular_pipelines/stable_diffusion_xl/decoders.py

@@ -23,17 +23,14 @@ from ...image_processor import VaeImageProcessor
 from ...models import AutoencoderKL
 from ...models.attention_processor import AttnProcessor2_0, XFormersAttnProcessor
 from ...utils import logging
-from ..modular_pipeline import (
-    PipelineBlock,
-    PipelineState,
-)
+from ..modular_pipeline import ModularPipelineBlocks, PipelineState
 from ..modular_pipeline_utils import ComponentSpec, InputParam, OutputParam
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class StableDiffusionXLDecodeStep(PipelineBlock):
+class StableDiffusionXLDecodeStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -157,7 +154,7 @@ class StableDiffusionXLDecodeStep(PipelineBlock):
         return components, state
 
 
-class StableDiffusionXLInpaintOverlayMaskStep(PipelineBlock):
+class StableDiffusionXLInpaintOverlayMaskStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property

src/diffusers/modular_pipelines/stable_diffusion_xl/denoise.py

@@ -25,7 +25,7 @@ from ...utils import logging
 from ..modular_pipeline import (
     BlockState,
     LoopSequentialPipelineBlocks,
-    PipelineBlock,
+    ModularPipelineBlocks,
     PipelineState,
 )
 from ..modular_pipeline_utils import ComponentSpec, InputParam, OutputParam
@@ -37,7 +37,7 @@ logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 # YiYi experimenting composible denoise loop
 # loop step (1): prepare latent input for denoiser
-class StableDiffusionXLLoopBeforeDenoiser(PipelineBlock):
+class StableDiffusionXLLoopBeforeDenoiser(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -55,7 +55,7 @@ class StableDiffusionXLLoopBeforeDenoiser(PipelineBlock):
         )
 
     @property
-    def intermediate_inputs(self) -> List[str]:
+    def inputs(self) -> List[str]:
         return [
             InputParam(
                 "latents",
@@ -73,7 +73,7 @@ class StableDiffusionXLLoopBeforeDenoiser(PipelineBlock):
 
 
 # loop step (1): prepare latent input for denoiser (with inpainting)
-class StableDiffusionXLInpaintLoopBeforeDenoiser(PipelineBlock):
+class StableDiffusionXLInpaintLoopBeforeDenoiser(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -144,7 +144,7 @@ class StableDiffusionXLInpaintLoopBeforeDenoiser(PipelineBlock):
 
 
 # loop step (2): denoise the latents with guidance
-class StableDiffusionXLLoopDenoiser(PipelineBlock):
+class StableDiffusionXLLoopDenoiser(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -249,7 +249,7 @@ class StableDiffusionXLLoopDenoiser(PipelineBlock):
 
 
 # loop step (2): denoise the latents with guidance (with controlnet)
-class StableDiffusionXLControlNetLoopDenoiser(PipelineBlock):
+class StableDiffusionXLControlNetLoopDenoiser(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -449,7 +449,7 @@ class StableDiffusionXLControlNetLoopDenoiser(PipelineBlock):
 
 
 # loop step (3): scheduler step to update latents
-class StableDiffusionXLLoopAfterDenoiser(PipelineBlock):
+class StableDiffusionXLLoopAfterDenoiser(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -520,7 +520,7 @@ class StableDiffusionXLLoopAfterDenoiser(PipelineBlock):
 
 
 # loop step (3): scheduler step to update latents (with inpainting)
-class StableDiffusionXLInpaintLoopAfterDenoiser(PipelineBlock):
+class StableDiffusionXLInpaintLoopAfterDenoiser(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -660,7 +660,7 @@ class StableDiffusionXLDenoiseLoopWrapper(LoopSequentialPipelineBlocks):
         ]
 
     @property
-    def loop_intermediate_inputs(self) -> List[InputParam]:
+    def loop_inputs(self) -> List[InputParam]:
         return [
             InputParam(
                 "timesteps",

src/diffusers/modular_pipelines/stable_diffusion_xl/encoders.py

@@ -35,7 +35,7 @@ from ...utils import (
     scale_lora_layers,
     unscale_lora_layers,
 )
-from ..modular_pipeline import PipelineBlock, PipelineState
+from ..modular_pipeline import ModularPipelineBlocks, PipelineState
 from ..modular_pipeline_utils import ComponentSpec, ConfigSpec, InputParam, OutputParam
 from .modular_pipeline import StableDiffusionXLModularPipeline
 
@@ -57,7 +57,7 @@ def retrieve_latents(
         raise AttributeError("Could not access latents of provided encoder_output")
 
 
-class StableDiffusionXLIPAdapterStep(PipelineBlock):
+class StableDiffusionXLIPAdapterStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -215,7 +215,7 @@ class StableDiffusionXLIPAdapterStep(PipelineBlock):
         return components, state
 
 
-class StableDiffusionXLTextEncoderStep(PipelineBlock):
+class StableDiffusionXLTextEncoderStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -576,7 +576,7 @@ class StableDiffusionXLTextEncoderStep(PipelineBlock):
         return components, state
 
 
-class StableDiffusionXLVaeEncoderStep(PipelineBlock):
+class StableDiffusionXLVaeEncoderStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property
@@ -691,7 +691,7 @@ class StableDiffusionXLVaeEncoderStep(PipelineBlock):
         return components, state
 
 
-class StableDiffusionXLInpaintVaeEncoderStep(PipelineBlock):
+class StableDiffusionXLInpaintVaeEncoderStep(ModularPipelineBlocks):
     model_name = "stable-diffusion-xl"
 
     @property

src/diffusers/modular_pipelines/wan/before_denoise.py

@@ -20,7 +20,7 @@ import torch
 from ...schedulers import UniPCMultistepScheduler
 from ...utils import logging
 from ...utils.torch_utils import randn_tensor
-from ..modular_pipeline import PipelineBlock, PipelineState
+from ..modular_pipeline import ModularPipelineBlocks, PipelineState
 from ..modular_pipeline_utils import ComponentSpec, InputParam, OutputParam
 from .modular_pipeline import WanModularPipeline
 
@@ -94,7 +94,7 @@ def retrieve_timesteps(
     return timesteps, num_inference_steps
 
 
-class WanInputStep(PipelineBlock):
+class WanInputStep(ModularPipelineBlocks):
     model_name = "wan"
 
     @property
@@ -194,7 +194,7 @@ class WanInputStep(PipelineBlock):
         return components, state
 
 
-class WanSetTimestepsStep(PipelineBlock):
+class WanSetTimestepsStep(ModularPipelineBlocks):
     model_name = "wan"
 
     @property
@@ -243,7 +243,7 @@ class WanSetTimestepsStep(PipelineBlock):
         return components, state
 
 
-class WanPrepareLatentsStep(PipelineBlock):
+class WanPrepareLatentsStep(ModularPipelineBlocks):
     model_name = "wan"
 
     @property

src/diffusers/modular_pipelines/wan/decoders.py

@@ -22,14 +22,14 @@ from ...configuration_utils import FrozenDict
 from ...models import AutoencoderKLWan
 from ...utils import logging
 from ...video_processor import VideoProcessor
-from ..modular_pipeline import PipelineBlock, PipelineState
+from ..modular_pipeline import ModularPipelineBlocks, PipelineState
 from ..modular_pipeline_utils import ComponentSpec, InputParam, OutputParam
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class WanDecodeStep(PipelineBlock):
+class WanDecodeStep(ModularPipelineBlocks):
     model_name = "wan"
 
     @property

src/diffusers/modular_pipelines/wan/denoise.py

@@ -24,7 +24,7 @@ from ...utils import logging
 from ..modular_pipeline import (
     BlockState,
     LoopSequentialPipelineBlocks,
-    PipelineBlock,
+    ModularPipelineBlocks,
     PipelineState,
 )
 from ..modular_pipeline_utils import ComponentSpec, InputParam, OutputParam
@@ -34,7 +34,7 @@ from .modular_pipeline import WanModularPipeline
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-class WanLoopDenoiser(PipelineBlock):
+class WanLoopDenoiser(ModularPipelineBlocks):
     model_name = "wan"
 
     @property
@@ -132,7 +132,7 @@ class WanLoopDenoiser(PipelineBlock):
         return components, block_state
 
 
-class WanLoopAfterDenoiser(PipelineBlock):
+class WanLoopAfterDenoiser(ModularPipelineBlocks):
     model_name = "wan"
 
     @property

src/diffusers/modular_pipelines/wan/encoders.py

@@ -22,7 +22,7 @@ from transformers import AutoTokenizer, UMT5EncoderModel
 from ...configuration_utils import FrozenDict
 from ...guiders import ClassifierFreeGuidance
 from ...utils import is_ftfy_available, logging
-from ..modular_pipeline import PipelineBlock, PipelineState
+from ..modular_pipeline import ModularPipelineBlocks, PipelineState
 from ..modular_pipeline_utils import ComponentSpec, ConfigSpec, InputParam, OutputParam
 from .modular_pipeline import WanModularPipeline
 
@@ -51,7 +51,7 @@ def prompt_clean(text):
     return text
 
 
-class WanTextEncoderStep(PipelineBlock):
+class WanTextEncoderStep(ModularPipelineBlocks):
     model_name = "wan"
 
     @property