[ControlNet] Adds controlnet for SanaTransformer (#11040)

* added controlnet for sana transformer

* improve code quality

* addressed PR comments

* bug fixes

* added test cases

* update

* added dummy objects

* addressed PR comments

* update

* Forcing update

* add to docs

* code quality

* addressed PR comments

* addressed PR comments

* update

* addressed PR comments

* added proper styling

* update

* Revert "added proper styling"

This reverts commit 344ee8a7014ada095b295034ef84341f03b0e359.

* manually ordered

* Apply suggestions from code review

---------
Co-authored-by: Aryan <contact.aryanvs@gmail.com>

docs/source/en/_toctree.yml

@@ -270,16 +270,18 @@
     - sections:
       - local: api/models/controlnet
         title: ControlNetModel
+      - local: api/models/controlnet_union
+        title: ControlNetUnionModel
       - local: api/models/controlnet_flux
         title: FluxControlNetModel
       - local: api/models/controlnet_hunyuandit
         title: HunyuanDiT2DControlNetModel
+      - local: api/models/controlnet_sana
+        title: SanaControlNetModel
       - local: api/models/controlnet_sd3
         title: SD3ControlNetModel
       - local: api/models/controlnet_sparsectrl
         title: SparseControlNetModel
-      - local: api/models/controlnet_union
-        title: ControlNetUnionModel
       title: ControlNets
     - sections:
       - local: api/models/allegro_transformer3d
@@ -424,6 +426,8 @@
       title: ControlNet with Stable Diffusion 3
     - local: api/pipelines/controlnet_sdxl
       title: ControlNet with Stable Diffusion XL
+    - local: api/pipelines/controlnet_sana
+      title: ControlNet-Sana
     - local: api/pipelines/controlnetxs
       title: ControlNet-XS
     - local: api/pipelines/controlnetxs_sdxl

docs/source/en/api/models/controlnet_sana.md

+<!--Copyright 2024 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.
+-->
+
+# SanaControlNetModel
+
+The ControlNet model was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, Maneesh Agrawala. It provides a greater degree of control over text-to-image generation by conditioning the model on additional inputs such as edge maps, depth maps, segmentation maps, and keypoints for pose detection.
+
+The abstract from the paper is:
+
+*We present ControlNet, a neural network architecture to add spatial conditioning controls to large, pretrained text-to-image diffusion models. ControlNet locks the production-ready large diffusion models, and reuses their deep and robust encoding layers pretrained with billions of images as a strong backbone to learn a diverse set of conditional controls. The neural architecture is connected with "zero convolutions" (zero-initialized convolution layers) that progressively grow the parameters from zero and ensure that no harmful noise could affect the finetuning. We test various conditioning controls, eg, edges, depth, segmentation, human pose, etc, with Stable Diffusion, using single or multiple conditions, with or without prompts. We show that the training of ControlNets is robust with small (<50k) and large (>1m) datasets. Extensive results show that ControlNet may facilitate wider applications to control image diffusion models.*
+
+This model was contributed by [ishan24](https://huggingface.co/ishan24). ❤️
+The original codebase can be found at [NVlabs/Sana](https://github.com/NVlabs/Sana), and you can find official ControlNet checkpoints on [Efficient-Large-Model's](https://huggingface.co/Efficient-Large-Model) Hub profile.
+
+## SanaControlNetModel
+[[autodoc]] SanaControlNetModel
+
+## SanaControlNetOutput
+[[autodoc]] models.controlnets.controlnet_sana.SanaControlNetOutput
+

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

+<!--Copyright 2024 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.
+-->
+
+# ControlNet
+
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
+ControlNet was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, and Maneesh Agrawala.
+
+With a ControlNet model, you can provide an additional control image to condition and control Stable Diffusion generation. For example, if you provide a depth map, the ControlNet model generates an image that'll preserve the spatial information from the depth map. It is a more flexible and accurate way to control the image generation process.
+
+The abstract from the paper is:
+
+*We present ControlNet, a neural network architecture to add spatial conditioning controls to large, pretrained text-to-image diffusion models. ControlNet locks the production-ready large diffusion models, and reuses their deep and robust encoding layers pretrained with billions of images as a strong backbone to learn a diverse set of conditional controls. The neural architecture is connected with "zero convolutions" (zero-initialized convolution layers) that progressively grow the parameters from zero and ensure that no harmful noise could affect the finetuning. We test various conditioning controls, eg, edges, depth, segmentation, human pose, etc, with Stable Diffusion, using single or multiple conditions, with or without prompts. We show that the training of ControlNets is robust with small (<50k) and large (>1m) datasets. Extensive results show that ControlNet may facilitate wider applications to control image diffusion models.*
+
+This pipeline was contributed by [ishan24](https://huggingface.co/ishan24). ❤️
+The original codebase can be found at [NVlabs/Sana](https://github.com/NVlabs/Sana), and you can find official ControlNet checkpoints on [Efficient-Large-Model's](https://huggingface.co/Efficient-Large-Model) Hub profile.
+
+## SanaControlNetPipeline
+[[autodoc]] SanaControlNetPipeline
+	- all
+	- __call__
+
+## SanaPipelineOutput
+[[autodoc]] pipelines.sana.pipeline_output.SanaPipelineOutput
\ No newline at end of file

scripts/convert_sana_controlnet_to_diffusers.py

+#!/usr/bin/env python
+from __future__ import annotations
+
+import argparse
+from contextlib import nullcontext
+
+import torch
+from accelerate import init_empty_weights
+
+from diffusers import (
+    SanaControlNetModel,
+)
+from diffusers.models.modeling_utils import load_model_dict_into_meta
+from diffusers.utils.import_utils import is_accelerate_available
+
+
+CTX = init_empty_weights if is_accelerate_available else nullcontext
+
+
+def main(args):
+    file_path = args.orig_ckpt_path
+
+    all_state_dict = torch.load(file_path, weights_only=True)
+    state_dict = all_state_dict.pop("state_dict")
+    converted_state_dict = {}
+
+    # Patch embeddings.
+    converted_state_dict["patch_embed.proj.weight"] = state_dict.pop("x_embedder.proj.weight")
+    converted_state_dict["patch_embed.proj.bias"] = state_dict.pop("x_embedder.proj.bias")
+
+    # Caption projection.
+    converted_state_dict["caption_projection.linear_1.weight"] = state_dict.pop("y_embedder.y_proj.fc1.weight")
+    converted_state_dict["caption_projection.linear_1.bias"] = state_dict.pop("y_embedder.y_proj.fc1.bias")
+    converted_state_dict["caption_projection.linear_2.weight"] = state_dict.pop("y_embedder.y_proj.fc2.weight")
+    converted_state_dict["caption_projection.linear_2.bias"] = state_dict.pop("y_embedder.y_proj.fc2.bias")
+
+    # AdaLN-single LN
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.weight"] = state_dict.pop(
+        "t_embedder.mlp.0.weight"
+    )
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.bias"] = state_dict.pop("t_embedder.mlp.0.bias")
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.weight"] = state_dict.pop(
+        "t_embedder.mlp.2.weight"
+    )
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.bias"] = state_dict.pop("t_embedder.mlp.2.bias")
+
+    # Shared norm.
+    converted_state_dict["time_embed.linear.weight"] = state_dict.pop("t_block.1.weight")
+    converted_state_dict["time_embed.linear.bias"] = state_dict.pop("t_block.1.bias")
+
+    # y norm
+    converted_state_dict["caption_norm.weight"] = state_dict.pop("attention_y_norm.weight")
+
+    # Positional embedding interpolation scale.
+    interpolation_scale = {512: None, 1024: None, 2048: 1.0, 4096: 2.0}
+
+    # ControlNet Input Projection.
+    converted_state_dict["input_block.weight"] = state_dict.pop("controlnet.0.before_proj.weight")
+    converted_state_dict["input_block.bias"] = state_dict.pop("controlnet.0.before_proj.bias")
+
+    for depth in range(7):
+        # Transformer blocks.
+        converted_state_dict[f"transformer_blocks.{depth}.scale_shift_table"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.scale_shift_table"
+        )
+
+        # Linear Attention is all you need 🤘
+        # Self attention.
+        q, k, v = torch.chunk(state_dict.pop(f"controlnet.{depth}.copied_block.attn.qkv.weight"), 3, dim=0)
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.weight"] = v
+        # Projection.
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.bias"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.attn.proj.bias"
+        )
+
+        # Feed-forward.
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_inverted.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.inverted_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_inverted.bias"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.inverted_conv.conv.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_depth.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.depth_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_depth.bias"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.depth_conv.conv.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_point.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.mlp.point_conv.conv.weight"
+        )
+
+        # Cross-attention.
+        q = state_dict.pop(f"controlnet.{depth}.copied_block.cross_attn.q_linear.weight")
+        q_bias = state_dict.pop(f"controlnet.{depth}.copied_block.cross_attn.q_linear.bias")
+        k, v = torch.chunk(state_dict.pop(f"controlnet.{depth}.copied_block.cross_attn.kv_linear.weight"), 2, dim=0)
+        k_bias, v_bias = torch.chunk(
+            state_dict.pop(f"controlnet.{depth}.copied_block.cross_attn.kv_linear.bias"), 2, dim=0
+        )
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.bias"] = q_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.bias"] = k_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.weight"] = v
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.bias"] = v_bias
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.weight"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.cross_attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.bias"] = state_dict.pop(
+            f"controlnet.{depth}.copied_block.cross_attn.proj.bias"
+        )
+
+        # ControlNet After Projection
+        converted_state_dict[f"controlnet_blocks.{depth}.weight"] = state_dict.pop(
+            f"controlnet.{depth}.after_proj.weight"
+        )
+        converted_state_dict[f"controlnet_blocks.{depth}.bias"] = state_dict.pop(f"controlnet.{depth}.after_proj.bias")
+
+    # ControlNet
+    with CTX():
+        controlnet = SanaControlNetModel(
+            num_attention_heads=model_kwargs[args.model_type]["num_attention_heads"],
+            attention_head_dim=model_kwargs[args.model_type]["attention_head_dim"],
+            num_layers=model_kwargs[args.model_type]["num_layers"],
+            num_cross_attention_heads=model_kwargs[args.model_type]["num_cross_attention_heads"],
+            cross_attention_head_dim=model_kwargs[args.model_type]["cross_attention_head_dim"],
+            cross_attention_dim=model_kwargs[args.model_type]["cross_attention_dim"],
+            caption_channels=2304,
+            sample_size=args.image_size // 32,
+            interpolation_scale=interpolation_scale[args.image_size],
+        )
+
+    if is_accelerate_available():
+        load_model_dict_into_meta(controlnet, converted_state_dict)
+    else:
+        controlnet.load_state_dict(converted_state_dict, strict=True, assign=True)
+
+    num_model_params = sum(p.numel() for p in controlnet.parameters())
+    print(f"Total number of controlnet parameters: {num_model_params}")
+
+    controlnet = controlnet.to(weight_dtype)
+    controlnet.load_state_dict(converted_state_dict, strict=True)
+
+    print(f"Saving Sana ControlNet in Diffusers format in {args.dump_path}.")
+    controlnet.save_pretrained(args.dump_path)
+
+
+DTYPE_MAPPING = {
+    "fp32": torch.float32,
+    "fp16": torch.float16,
+    "bf16": torch.bfloat16,
+}
+
+VARIANT_MAPPING = {
+    "fp32": None,
+    "fp16": "fp16",
+    "bf16": "bf16",
+}
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--orig_ckpt_path", default=None, type=str, required=True, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--image_size",
+        default=1024,
+        type=int,
+        choices=[512, 1024, 2048, 4096],
+        required=False,
+        help="Image size of pretrained model, 512, 1024, 2048 or 4096.",
+    )
+    parser.add_argument(
+        "--model_type",
+        default="SanaMS_1600M_P1_ControlNet_D7",
+        type=str,
+        choices=["SanaMS_1600M_P1_ControlNet_D7", "SanaMS_600M_P1_ControlNet_D7"],
+    )
+    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
+    parser.add_argument("--dtype", default="fp16", type=str, choices=["fp32", "fp16", "bf16"], help="Weight dtype.")
+
+    args = parser.parse_args()
+
+    model_kwargs = {
+        "SanaMS_1600M_P1_ControlNet_D7": {
+            "num_attention_heads": 70,
+            "attention_head_dim": 32,
+            "num_cross_attention_heads": 20,
+            "cross_attention_head_dim": 112,
+            "cross_attention_dim": 2240,
+            "num_layers": 7,
+        },
+        "SanaMS_600M_P1_ControlNet_D7": {
+            "num_attention_heads": 36,
+            "attention_head_dim": 32,
+            "num_cross_attention_heads": 16,
+            "cross_attention_head_dim": 72,
+            "cross_attention_dim": 1152,
+            "num_layers": 7,
+        },
+    }
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    weight_dtype = DTYPE_MAPPING[args.dtype]
+    variant = VARIANT_MAPPING[args.dtype]
+
+    main(args)

src/diffusers/__init__.py

@@ -190,6 +190,7 @@ else:
             "OmniGenTransformer2DModel",
             "PixArtTransformer2DModel",
             "PriorTransformer",
+            "SanaControlNetModel",
             "SanaTransformer2DModel",
             "SD3ControlNetModel",
             "SD3MultiControlNetModel",
@@ -428,6 +429,7 @@ else:
             "PixArtSigmaPAGPipeline",
             "PixArtSigmaPipeline",
             "ReduxImageEncoder",
+            "SanaControlNetPipeline",
             "SanaPAGPipeline",
             "SanaPipeline",
             "SanaSprintPipeline",
@@ -782,6 +784,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             OmniGenTransformer2DModel,
             PixArtTransformer2DModel,
             PriorTransformer,
+            SanaControlNetModel,
             SanaTransformer2DModel,
             SD3ControlNetModel,
             SD3MultiControlNetModel,
@@ -999,6 +1002,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             PixArtSigmaPAGPipeline,
             PixArtSigmaPipeline,
             ReduxImageEncoder,
+            SanaControlNetPipeline,
             SanaPAGPipeline,
             SanaPipeline,
             SanaSprintPipeline,

src/diffusers/models/__init__.py

@@ -49,6 +49,7 @@ if is_torch_available():
         "HunyuanDiT2DControlNetModel",
         "HunyuanDiT2DMultiControlNetModel",
     ]
+    _import_structure["controlnets.controlnet_sana"] = ["SanaControlNetModel"]
     _import_structure["controlnets.controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"]
     _import_structure["controlnets.controlnet_sparsectrl"] = ["SparseControlNetModel"]
     _import_structure["controlnets.controlnet_union"] = ["ControlNetUnionModel"]
@@ -134,6 +135,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             HunyuanDiT2DMultiControlNetModel,
             MultiControlNetModel,
             MultiControlNetUnionModel,
+            SanaControlNetModel,
             SD3ControlNetModel,
             SD3MultiControlNetModel,
             SparseControlNetModel,

src/diffusers/models/controlnets/__init__.py

@@ -9,6 +9,7 @@ if is_torch_available():
         HunyuanDiT2DControlNetModel,
         HunyuanDiT2DMultiControlNetModel,
     )
+    from .controlnet_sana import SanaControlNetModel
     from .controlnet_sd3 import SD3ControlNetModel, SD3ControlNetOutput, SD3MultiControlNetModel
     from .controlnet_sparsectrl import (
         SparseControlNetConditioningEmbedding,

src/diffusers/models/controlnets/controlnet_sana.py

+# Copyright 2024 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 dataclasses import dataclass
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, BaseOutput, logging, scale_lora_layers, unscale_lora_layers
+from ..attention_processor import AttentionProcessor
+from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormSingle, RMSNorm
+from ..transformers.sana_transformer import SanaTransformerBlock
+from .controlnet import zero_module
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class SanaControlNetOutput(BaseOutput):
+    controlnet_block_samples: Tuple[torch.Tensor]
+
+
+class SanaControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["SanaTransformerBlock", "PatchEmbed"]
+    _skip_layerwise_casting_patterns = ["patch_embed", "norm"]
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 32,
+        out_channels: Optional[int] = 32,
+        num_attention_heads: int = 70,
+        attention_head_dim: int = 32,
+        num_layers: int = 7,
+        num_cross_attention_heads: Optional[int] = 20,
+        cross_attention_head_dim: Optional[int] = 112,
+        cross_attention_dim: Optional[int] = 2240,
+        caption_channels: int = 2304,
+        mlp_ratio: float = 2.5,
+        dropout: float = 0.0,
+        attention_bias: bool = False,
+        sample_size: int = 32,
+        patch_size: int = 1,
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-6,
+        interpolation_scale: Optional[int] = None,
+    ) -> None:
+        super().__init__()
+
+        out_channels = out_channels or in_channels
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # 1. Patch Embedding
+        self.patch_embed = PatchEmbed(
+            height=sample_size,
+            width=sample_size,
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=inner_dim,
+            interpolation_scale=interpolation_scale,
+            pos_embed_type="sincos" if interpolation_scale is not None else None,
+        )
+
+        # 2. Additional condition embeddings
+        self.time_embed = AdaLayerNormSingle(inner_dim)
+
+        self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+        self.caption_norm = RMSNorm(inner_dim, eps=1e-5, elementwise_affine=True)
+
+        # 3. Transformer blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                SanaTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    num_cross_attention_heads=num_cross_attention_heads,
+                    cross_attention_head_dim=cross_attention_head_dim,
+                    cross_attention_dim=cross_attention_dim,
+                    attention_bias=attention_bias,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    mlp_ratio=mlp_ratio,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # controlnet_blocks
+        self.controlnet_blocks = nn.ModuleList([])
+
+        self.input_block = zero_module(nn.Linear(inner_dim, inner_dim))
+        for _ in range(len(self.transformer_blocks)):
+            controlnet_block = nn.Linear(inner_dim, inner_dim)
+            controlnet_block = zero_module(controlnet_block)
+            self.controlnet_blocks.append(controlnet_block)
+
+        self.gradient_checkpointing = False
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        controlnet_cond: torch.Tensor,
+        conditioning_scale: float = 1.0,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[Tuple[torch.Tensor, ...], Transformer2DModelOutput]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 1. Input
+        batch_size, num_channels, height, width = hidden_states.shape
+        p = self.config.patch_size
+        post_patch_height, post_patch_width = height // p, width // p
+
+        hidden_states = self.patch_embed(hidden_states)
+        hidden_states = hidden_states + self.input_block(self.patch_embed(controlnet_cond.to(hidden_states.dtype)))
+
+        timestep, embedded_timestep = self.time_embed(
+            timestep, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+        )
+
+        encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+        encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        encoder_hidden_states = self.caption_norm(encoder_hidden_states)
+
+        # 2. Transformer blocks
+        block_res_samples = ()
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for block in self.transformer_blocks:
+                hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    post_patch_height,
+                    post_patch_width,
+                )
+                block_res_samples = block_res_samples + (hidden_states,)
+        else:
+            for block in self.transformer_blocks:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    post_patch_height,
+                    post_patch_width,
+                )
+                block_res_samples = block_res_samples + (hidden_states,)
+
+        # 3. ControlNet blocks
+        controlnet_block_res_samples = ()
+        for block_res_sample, controlnet_block in zip(block_res_samples, self.controlnet_blocks):
+            block_res_sample = controlnet_block(block_res_sample)
+            controlnet_block_res_samples = controlnet_block_res_samples + (block_res_sample,)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        controlnet_block_res_samples = [sample * conditioning_scale for sample in controlnet_block_res_samples]
+
+        if not return_dict:
+            return (controlnet_block_res_samples,)
+
+        return SanaControlNetOutput(controlnet_block_samples=controlnet_block_res_samples)

src/diffusers/models/transformers/sana_transformer.py

@@ -483,6 +483,7 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         encoder_attention_mask: Optional[torch.Tensor] = None,
         attention_mask: Optional[torch.Tensor] = None,
         attention_kwargs: Optional[Dict[str, Any]] = None,
+        controlnet_block_samples: Optional[Tuple[torch.Tensor]] = None,
         return_dict: bool = True,
     ) -> Union[Tuple[torch.Tensor, ...], Transformer2DModelOutput]:
         if attention_kwargs is not None:
@@ -546,7 +547,7 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
 
         # 2. Transformer blocks
         if torch.is_grad_enabled() and self.gradient_checkpointing:
-            for block in self.transformer_blocks:
+            for index_block, block in enumerate(self.transformer_blocks):
                 hidden_states = self._gradient_checkpointing_func(
                     block,
                     hidden_states,
@@ -557,9 +558,11 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
                     post_patch_height,
                     post_patch_width,
                 )
+                if controlnet_block_samples is not None and 0 < index_block <= len(controlnet_block_samples):
+                    hidden_states = hidden_states + controlnet_block_samples[index_block - 1]
 
         else:
-            for block in self.transformer_blocks:
+            for index_block, block in enumerate(self.transformer_blocks):
                 hidden_states = block(
                     hidden_states,
                     attention_mask,
@@ -569,6 +572,8 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
                     post_patch_height,
                     post_patch_width,
                 )
+                if controlnet_block_samples is not None and 0 < index_block <= len(controlnet_block_samples):
+                    hidden_states = hidden_states + controlnet_block_samples[index_block - 1]
 
         # 3. Normalization
         hidden_states = self.norm_out(hidden_states, embedded_timestep, self.scale_shift_table)

src/diffusers/pipelines/__init__.py

@@ -281,7 +281,7 @@ else:
     _import_structure["paint_by_example"] = ["PaintByExamplePipeline"]
     _import_structure["pia"] = ["PIAPipeline"]
     _import_structure["pixart_alpha"] = ["PixArtAlphaPipeline", "PixArtSigmaPipeline"]
-    _import_structure["sana"] = ["SanaPipeline", "SanaSprintPipeline"]
+    _import_structure["sana"] = ["SanaPipeline", "SanaSprintPipeline", "SanaControlNetPipeline"]
     _import_structure["semantic_stable_diffusion"] = ["SemanticStableDiffusionPipeline"]
     _import_structure["shap_e"] = ["ShapEImg2ImgPipeline", "ShapEPipeline"]
     _import_structure["stable_audio"] = [
@@ -664,7 +664,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .paint_by_example import PaintByExamplePipeline
         from .pia import PIAPipeline
         from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline
-        from .sana import SanaPipeline, SanaSprintPipeline
+        from .sana import SanaControlNetPipeline, SanaPipeline, SanaSprintPipeline
         from .semantic_stable_diffusion import SemanticStableDiffusionPipeline
         from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline
         from .stable_audio import StableAudioPipeline, StableAudioProjectionModel

src/diffusers/pipelines/sana/__init__.py

@@ -23,6 +23,7 @@ except OptionalDependencyNotAvailable:
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
 else:
     _import_structure["pipeline_sana"] = ["SanaPipeline"]
+    _import_structure["pipeline_sana_controlnet"] = ["SanaControlNetPipeline"]
     _import_structure["pipeline_sana_sprint"] = ["SanaSprintPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
@@ -34,6 +35,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from ...utils.dummy_torch_and_transformers_objects import *
     else:
         from .pipeline_sana import SanaPipeline
+        from .pipeline_sana_controlnet import SanaControlNetPipeline
         from .pipeline_sana_sprint import SanaSprintPipeline
 else:
     import sys

src/diffusers/pipelines/sana/pipeline_sana.py

@@ -354,9 +354,7 @@ class SanaPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
         if device is None:
             device = self._execution_device
 
-        if self.transformer is not None:
-            dtype = self.transformer.dtype
-        elif self.text_encoder is not None:
+        if self.text_encoder is not None:
             dtype = self.text_encoder.dtype
         else:
             dtype = None
@@ -928,22 +926,22 @@ class SanaPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
         num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
         self._num_timesteps = len(timesteps)
 
+        transformer_dtype = self.transformer.dtype
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
                 if self.interrupt:
                     continue
 
                 latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
-                latent_model_input = latent_model_input.to(prompt_embeds.dtype)
 
                 # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
-                timestep = t.expand(latent_model_input.shape[0]).to(latents.dtype)
+                timestep = t.expand(latent_model_input.shape[0])
                 timestep = timestep * self.transformer.config.timestep_scale
 
                 # predict noise model_output
                 noise_pred = self.transformer(
-                    latent_model_input,
-                    encoder_hidden_states=prompt_embeds,
+                    latent_model_input.to(dtype=transformer_dtype),
+                    encoder_hidden_states=prompt_embeds.to(dtype=transformer_dtype),
                     encoder_attention_mask=prompt_attention_mask,
                     timestep=timestep,
                     return_dict=False,
@@ -959,8 +957,6 @@ class SanaPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
                 # learned sigma
                 if self.transformer.config.out_channels // 2 == latent_channels:
                     noise_pred = noise_pred.chunk(2, dim=1)[0]
-                else:
-                    noise_pred = noise_pred
 
                 # compute previous image: x_t -> x_t-1
                 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]

src/diffusers/pipelines/sana/pipeline_sana_sprint.py

@@ -295,9 +295,7 @@ class SanaSprintPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
         if device is None:
             device = self._execution_device
 
-        if self.transformer is not None:
-            dtype = self.transformer.dtype
-        elif self.text_encoder is not None:
+        if self.text_encoder is not None:
             dtype = self.text_encoder.dtype
         else:
             dtype = None
@@ -806,13 +804,14 @@ class SanaSprintPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
         num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
         self._num_timesteps = len(timesteps)
 
+        transformer_dtype = self.transformer.dtype
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
                 if self.interrupt:
                     continue
 
                 # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
-                timestep = t.expand(latents.shape[0]).to(prompt_embeds.dtype)
+                timestep = t.expand(latents.shape[0])
                 latents_model_input = latents / self.scheduler.config.sigma_data
 
                 scm_timestep = torch.sin(timestep) / (torch.cos(timestep) + torch.sin(timestep))
@@ -821,12 +820,11 @@ class SanaSprintPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
                 latent_model_input = latents_model_input * torch.sqrt(
                     scm_timestep_expanded**2 + (1 - scm_timestep_expanded) ** 2
                 )
-                latent_model_input = latent_model_input.to(prompt_embeds.dtype)
 
                 # predict noise model_output
                 noise_pred = self.transformer(
-                    latent_model_input,
-                    encoder_hidden_states=prompt_embeds,
+                    latent_model_input.to(dtype=transformer_dtype),
+                    encoder_hidden_states=prompt_embeds.to(dtype=transformer_dtype),
                     encoder_attention_mask=prompt_attention_mask,
                     guidance=guidance,
                     timestep=scm_timestep,

src/diffusers/utils/dummy_pt_objects.py

@@ -790,6 +790,21 @@ class PriorTransformer(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class SanaControlNetModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class SanaTransformer2DModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -1502,6 +1502,21 @@ class ReduxImageEncoder(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class SanaControlNetPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class SanaPAGPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/pipelines/sana/test_sana_controlnet.py

+# Copyright 2024 The HuggingFace Team.
+#
+# 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.
+
+import inspect
+import unittest
+
+import numpy as np
+import torch
+from transformers import Gemma2Config, Gemma2Model, GemmaTokenizer
+
+from diffusers import (
+    AutoencoderDC,
+    FlowMatchEulerDiscreteScheduler,
+    SanaControlNetModel,
+    SanaControlNetPipeline,
+    SanaTransformer2DModel,
+)
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    torch_device,
+)
+
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import PipelineTesterMixin, to_np
+
+
+enable_full_determinism()
+
+
+class SanaControlNetPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = SanaControlNetPipeline
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+    test_xformers_attention = False
+    test_layerwise_casting = True
+    test_group_offloading = True
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        controlnet = SanaControlNetModel(
+            patch_size=1,
+            in_channels=4,
+            out_channels=4,
+            num_layers=1,
+            num_attention_heads=2,
+            attention_head_dim=4,
+            num_cross_attention_heads=2,
+            cross_attention_head_dim=4,
+            cross_attention_dim=8,
+            caption_channels=8,
+            sample_size=32,
+        )
+
+        torch.manual_seed(0)
+        transformer = SanaTransformer2DModel(
+            patch_size=1,
+            in_channels=4,
+            out_channels=4,
+            num_layers=1,
+            num_attention_heads=2,
+            attention_head_dim=4,
+            num_cross_attention_heads=2,
+            cross_attention_head_dim=4,
+            cross_attention_dim=8,
+            caption_channels=8,
+            sample_size=32,
+        )
+
+        torch.manual_seed(0)
+        vae = AutoencoderDC(
+            in_channels=3,
+            latent_channels=4,
+            attention_head_dim=2,
+            encoder_block_types=(
+                "ResBlock",
+                "EfficientViTBlock",
+            ),
+            decoder_block_types=(
+                "ResBlock",
+                "EfficientViTBlock",
+            ),
+            encoder_block_out_channels=(8, 8),
+            decoder_block_out_channels=(8, 8),
+            encoder_qkv_multiscales=((), (5,)),
+            decoder_qkv_multiscales=((), (5,)),
+            encoder_layers_per_block=(1, 1),
+            decoder_layers_per_block=[1, 1],
+            downsample_block_type="conv",
+            upsample_block_type="interpolate",
+            decoder_norm_types="rms_norm",
+            decoder_act_fns="silu",
+            scaling_factor=0.41407,
+        )
+
+        torch.manual_seed(0)
+        scheduler = FlowMatchEulerDiscreteScheduler(shift=7.0)
+
+        torch.manual_seed(0)
+        text_encoder_config = Gemma2Config(
+            head_dim=16,
+            hidden_size=8,
+            initializer_range=0.02,
+            intermediate_size=64,
+            max_position_embeddings=8192,
+            model_type="gemma2",
+            num_attention_heads=2,
+            num_hidden_layers=1,
+            num_key_value_heads=2,
+            vocab_size=8,
+            attn_implementation="eager",
+        )
+        text_encoder = Gemma2Model(text_encoder_config)
+        tokenizer = GemmaTokenizer.from_pretrained("hf-internal-testing/dummy-gemma")
+
+        components = {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+            "controlnet": controlnet,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+
+        control_image = torch.randn(1, 3, 32, 32, generator=generator)
+        inputs = {
+            "prompt": "",
+            "negative_prompt": "",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+            "height": 32,
+            "width": 32,
+            "max_sequence_length": 16,
+            "output_type": "pt",
+            "complex_human_instruction": None,
+            "control_image": control_image,
+            "controlnet_conditioning_scale": 1.0,
+        }
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs)[0]
+        generated_image = image[0]
+
+        self.assertEqual(generated_image.shape, (3, 32, 32))
+        expected_image = torch.randn(3, 32, 32)
+        max_diff = np.abs(generated_image - expected_image).max()
+        self.assertLessEqual(max_diff, 1e10)
+
+    def test_callback_inputs(self):
+        sig = inspect.signature(self.pipeline_class.__call__)
+        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
+        has_callback_step_end = "callback_on_step_end" in sig.parameters
+
+        if not (has_callback_tensor_inputs and has_callback_step_end):
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        self.assertTrue(
+            hasattr(pipe, "_callback_tensor_inputs"),
+            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
+        )
+
+        def callback_inputs_subset(pipe, i, t, callback_kwargs):
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        def callback_inputs_all(pipe, i, t, callback_kwargs):
+            for tensor_name in pipe._callback_tensor_inputs:
+                assert tensor_name in callback_kwargs
+
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        # Test passing in a subset
+        inputs["callback_on_step_end"] = callback_inputs_subset
+        inputs["callback_on_step_end_tensor_inputs"] = ["latents"]
+        output = pipe(**inputs)[0]
+
+        # Test passing in a everything
+        inputs["callback_on_step_end"] = callback_inputs_all
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+
+        def callback_inputs_change_tensor(pipe, i, t, callback_kwargs):
+            is_last = i == (pipe.num_timesteps - 1)
+            if is_last:
+                callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"])
+            return callback_kwargs
+
+        inputs["callback_on_step_end"] = callback_inputs_change_tensor
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+        assert output.abs().sum() < 1e10
+
+    def test_attention_slicing_forward_pass(
+        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
+    ):
+        if not self.test_attention_slicing:
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        generator_device = "cpu"
+        inputs = self.get_dummy_inputs(generator_device)
+        output_without_slicing = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=1)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing1 = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=2)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing2 = pipe(**inputs)[0]
+
+        if test_max_difference:
+            max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max()
+            max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max()
+            self.assertLess(
+                max(max_diff1, max_diff2),
+                expected_max_diff,
+                "Attention slicing should not affect the inference results",
+            )
+
+    def test_vae_tiling(self, expected_diff_max: float = 0.2):
+        generator_device = "cpu"
+        components = self.get_dummy_components()
+
+        pipe = self.pipeline_class(**components)
+        pipe.to("cpu")
+        pipe.set_progress_bar_config(disable=None)
+
+        # Without tiling
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_without_tiling = pipe(**inputs)[0]
+
+        # With tiling
+        pipe.vae.enable_tiling(
+            tile_sample_min_height=96,
+            tile_sample_min_width=96,
+            tile_sample_stride_height=64,
+            tile_sample_stride_width=64,
+        )
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_with_tiling = pipe(**inputs)[0]
+
+        self.assertLess(
+            (to_np(output_without_tiling) - to_np(output_with_tiling)).max(),
+            expected_diff_max,
+            "VAE tiling should not affect the inference results",
+        )
+
+    # TODO(aryan): Create a dummy gemma model with smol vocab size
+    @unittest.skip(
+        "A very small vocab size is used for fast tests. So, any kind of prompt other than the empty default used in other tests will lead to a embedding lookup error. This test uses a long prompt that causes the error."
+    )
+    def test_inference_batch_consistent(self):
+        pass
+
+    @unittest.skip(
+        "A very small vocab size is used for fast tests. So, any kind of prompt other than the empty default used in other tests will lead to a embedding lookup error. This test uses a long prompt that causes the error."
+    )
+    def test_inference_batch_single_identical(self):
+        pass
+
+    def test_float16_inference(self):
+        # Requires higher tolerance as model seems very sensitive to dtype
+        super().test_float16_inference(expected_max_diff=0.08)