diffusers 0.30.2__py3-none-any.whl → 0.31.0__py3-none-any.whl

diffusers/models/transformers/transformer_flux.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Black Forest Labs, The HuggingFace Team. All rights reserved.
+# Copyright 2024 Black Forest Labs, The HuggingFace Team and The InstantX 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.
@@ -13,8 +13,9 @@
 # limitations under the License.
 
 
-from typing import Any, Dict, List, Optional, Union
+from typing import Any, Dict, Optional, Tuple, Union
 
+import numpy as np
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -22,52 +23,23 @@ import torch.nn.functional as F
 from ...configuration_utils import ConfigMixin, register_to_config
 from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
 from ...models.attention import FeedForward
-from ...models.attention_processor import Attention, FluxAttnProcessor2_0, FluxSingleAttnProcessor2_0
+from ...models.attention_processor import (
+    Attention,
+    AttentionProcessor,
+    FluxAttnProcessor2_0,
+    FusedFluxAttnProcessor2_0,
+)
 from ...models.modeling_utils import ModelMixin
 from ...models.normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle
 from ...utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import maybe_allow_in_graph
-from ..embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings
+from ..embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
 from ..modeling_outputs import Transformer2DModelOutput
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
-# YiYi to-do: refactor rope related functions/classes
-def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
-    assert dim % 2 == 0, "The dimension must be even."
-
-    scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
-    omega = 1.0 / (theta**scale)
-
-    batch_size, seq_length = pos.shape
-    out = torch.einsum("...n,d->...nd", pos, omega)
-    cos_out = torch.cos(out)
-    sin_out = torch.sin(out)
-
-    stacked_out = torch.stack([cos_out, -sin_out, sin_out, cos_out], dim=-1)
-    out = stacked_out.view(batch_size, -1, dim // 2, 2, 2)
-    return out.float()
-
-
-# YiYi to-do: refactor rope related functions/classes
-class EmbedND(nn.Module):
-    def __init__(self, dim: int, theta: int, axes_dim: List[int]):
-        super().__init__()
-        self.dim = dim
-        self.theta = theta
-        self.axes_dim = axes_dim
-
-    def forward(self, ids: torch.Tensor) -> torch.Tensor:
-        n_axes = ids.shape[-1]
-        emb = torch.cat(
-            [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
-            dim=-3,
-        )
-        return emb.unsqueeze(1)
-
-
 @maybe_allow_in_graph
 class FluxSingleTransformerBlock(nn.Module):
     r"""
@@ -92,7 +64,7 @@ class FluxSingleTransformerBlock(nn.Module):
         self.act_mlp = nn.GELU(approximate="tanh")
         self.proj_out = nn.Linear(dim + self.mlp_hidden_dim, dim)
 
-        processor = FluxSingleAttnProcessor2_0()
+        processor = FluxAttnProcessor2_0()
         self.attn = Attention(
             query_dim=dim,
             cross_attention_dim=None,
@@ -111,14 +83,16 @@ class FluxSingleTransformerBlock(nn.Module):
         hidden_states: torch.FloatTensor,
         temb: torch.FloatTensor,
         image_rotary_emb=None,
+        joint_attention_kwargs=None,
     ):
         residual = hidden_states
         norm_hidden_states, gate = self.norm(hidden_states, emb=temb)
         mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states))
-
+        joint_attention_kwargs = joint_attention_kwargs or {}
         attn_output = self.attn(
             hidden_states=norm_hidden_states,
             image_rotary_emb=image_rotary_emb,
+            **joint_attention_kwargs,
         )
 
         hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2)
@@ -189,18 +163,20 @@ class FluxTransformerBlock(nn.Module):
         encoder_hidden_states: torch.FloatTensor,
         temb: torch.FloatTensor,
         image_rotary_emb=None,
+        joint_attention_kwargs=None,
     ):
         norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
 
         norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
             encoder_hidden_states, emb=temb
         )
-
+        joint_attention_kwargs = joint_attention_kwargs or {}
         # Attention.
         attn_output, context_attn_output = self.attn(
             hidden_states=norm_hidden_states,
             encoder_hidden_states=norm_encoder_hidden_states,
             image_rotary_emb=image_rotary_emb,
+            **joint_attention_kwargs,
         )
 
         # Process attention outputs for the `hidden_states`.
@@ -250,6 +226,7 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
     """
 
     _supports_gradient_checkpointing = True
+    _no_split_modules = ["FluxTransformerBlock", "FluxSingleTransformerBlock"]
 
     @register_to_config
     def __init__(
@@ -263,13 +240,14 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         joint_attention_dim: int = 4096,
         pooled_projection_dim: int = 768,
         guidance_embeds: bool = False,
-        axes_dims_rope: List[int] = [16, 56, 56],
+        axes_dims_rope: Tuple[int] = (16, 56, 56),
     ):
         super().__init__()
         self.out_channels = in_channels
         self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
 
-        self.pos_embed = EmbedND(dim=self.inner_dim, theta=10000, axes_dim=axes_dims_rope)
+        self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope)
+
         text_time_guidance_cls = (
             CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings
         )
@@ -307,6 +285,106 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
 
         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)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedFluxAttnProcessor2_0
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+        self.set_attn_processor(FusedFluxAttnProcessor2_0())
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
     def _set_gradient_checkpointing(self, module, value=False):
         if hasattr(module, "gradient_checkpointing"):
             module.gradient_checkpointing = value
@@ -321,7 +399,10 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         txt_ids: torch.Tensor = None,
         guidance: torch.Tensor = None,
         joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        controlnet_block_samples=None,
+        controlnet_single_block_samples=None,
         return_dict: bool = True,
+        controlnet_blocks_repeat: bool = False,
     ) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
         """
         The [`FluxTransformer2DModel`] forward method.
@@ -377,7 +458,20 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
         )
         encoder_hidden_states = self.context_embedder(encoder_hidden_states)
 
-        ids = torch.cat((txt_ids, img_ids), dim=1)
+        if txt_ids.ndim == 3:
+            logger.warning(
+                "Passing `txt_ids` 3d torch.Tensor is deprecated."
+                "Please remove the batch dimension and pass it as a 2d torch Tensor"
+            )
+            txt_ids = txt_ids[0]
+        if img_ids.ndim == 3:
+            logger.warning(
+                "Passing `img_ids` 3d torch.Tensor is deprecated."
+                "Please remove the batch dimension and pass it as a 2d torch Tensor"
+            )
+            img_ids = img_ids[0]
+
+        ids = torch.cat((txt_ids, img_ids), dim=0)
         image_rotary_emb = self.pos_embed(ids)
 
         for index_block, block in enumerate(self.transformer_blocks):
@@ -408,8 +502,21 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
                     encoder_hidden_states=encoder_hidden_states,
                     temb=temb,
                     image_rotary_emb=image_rotary_emb,
+                    joint_attention_kwargs=joint_attention_kwargs,
                 )
 
+            # controlnet residual
+            if controlnet_block_samples is not None:
+                interval_control = len(self.transformer_blocks) / len(controlnet_block_samples)
+                interval_control = int(np.ceil(interval_control))
+                # For Xlabs ControlNet.
+                if controlnet_blocks_repeat:
+                    hidden_states = (
+                        hidden_states + controlnet_block_samples[index_block % len(controlnet_block_samples)]
+                    )
+                else:
+                    hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control]
+
         hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
 
         for index_block, block in enumerate(self.single_transformer_blocks):
@@ -438,6 +545,16 @@ class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrig
                     hidden_states=hidden_states,
                     temb=temb,
                     image_rotary_emb=image_rotary_emb,
+                    joint_attention_kwargs=joint_attention_kwargs,
+                )
+
+            # controlnet residual
+            if controlnet_single_block_samples is not None:
+                interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples)
+                interval_control = int(np.ceil(interval_control))
+                hidden_states[:, encoder_hidden_states.shape[1] :, ...] = (
+                    hidden_states[:, encoder_hidden_states.shape[1] :, ...]
+                    + controlnet_single_block_samples[index_block // interval_control]
                 )
 
         hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]

diffusers/models/transformers/transformer_sd3.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 
-from typing import Any, Dict, List, Optional, Union
+from typing import Any, Dict, List, Optional, Tuple, Union
 
 import torch
 import torch.nn as nn
@@ -69,6 +69,10 @@ class SD3Transformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrigi
         pooled_projection_dim: int = 2048,
         out_channels: int = 16,
         pos_embed_max_size: int = 96,
+        dual_attention_layers: Tuple[
+            int, ...
+        ] = (),  # () for sd3.0; (0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12) for sd3.5
+        qk_norm: Optional[str] = None,
     ):
         super().__init__()
         default_out_channels = in_channels
@@ -97,6 +101,8 @@ class SD3Transformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOrigi
                     num_attention_heads=self.config.num_attention_heads,
                     attention_head_dim=self.config.attention_head_dim,
                     context_pre_only=i == num_layers - 1,
+                    qk_norm=qk_norm,
+                    use_dual_attention=True if i in dual_attention_layers else False,
                 )
                 for i in range(self.config.num_layers)
             ]

diffusers/models/unets/unet_2d_condition.py

@@ -463,7 +463,6 @@ class UNet2DConditionModel(
                 dropout=dropout,
             )
             self.up_blocks.append(up_block)
-            prev_output_channel = output_channel
 
         # out
         if norm_num_groups is not None:
@@ -599,7 +598,7 @@ class UNet2DConditionModel(
             )
         elif encoder_hid_dim_type is not None:
             raise ValueError(
-                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+                f"`encoder_hid_dim_type`: {encoder_hid_dim_type} must be None, 'text_proj', 'text_image_proj', or 'image_proj'."
             )
         else:
             self.encoder_hid_proj = None
@@ -679,7 +678,9 @@ class UNet2DConditionModel(
             # Kandinsky 2.2 ControlNet
             self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
         elif addition_embed_type is not None:
-            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+            raise ValueError(
+                f"`addition_embed_type`: {addition_embed_type} must be None, 'text', 'text_image', 'text_time', 'image', or 'image_hint'."
+            )
 
     def _set_pos_net_if_use_gligen(self, attention_type: str, cross_attention_dim: int):
         if attention_type in ["gated", "gated-text-image"]:
@@ -990,7 +991,7 @@ class UNet2DConditionModel(
             image_embs = added_cond_kwargs.get("image_embeds")
             aug_emb = self.add_embedding(image_embs)
         elif self.config.addition_embed_type == "image_hint":
-            # Kandinsky 2.2 - style
+            # Kandinsky 2.2 ControlNet - style
             if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
                 raise ValueError(
                     f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
@@ -1009,7 +1010,7 @@ class UNet2DConditionModel(
             # Kandinsky 2.1 - style
             if "image_embeds" not in added_cond_kwargs:
                 raise ValueError(
-                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
                 )
 
             image_embeds = added_cond_kwargs.get("image_embeds")
@@ -1018,14 +1019,14 @@ class UNet2DConditionModel(
             # Kandinsky 2.2 - style
             if "image_embeds" not in added_cond_kwargs:
                 raise ValueError(
-                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
                 )
             image_embeds = added_cond_kwargs.get("image_embeds")
             encoder_hidden_states = self.encoder_hid_proj(image_embeds)
         elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
             if "image_embeds" not in added_cond_kwargs:
                 raise ValueError(
-                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
                 )
 
             if hasattr(self, "text_encoder_hid_proj") and self.text_encoder_hid_proj is not None:
@@ -1140,7 +1141,6 @@ class UNet2DConditionModel(
         # 1. time
         t_emb = self.get_time_embed(sample=sample, timestep=timestep)
         emb = self.time_embedding(t_emb, timestep_cond)
-        aug_emb = None
 
         class_emb = self.get_class_embed(sample=sample, class_labels=class_labels)
         if class_emb is not None:

diffusers/models/unets/unet_motion_model.py

@@ -116,7 +116,7 @@ class AnimateDiffTransformer3D(nn.Module):
 
         self.in_channels = in_channels
 
-        self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+        self.norm = nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
         self.proj_in = nn.Linear(in_channels, inner_dim)
 
         # 3. Define transformers blocks
@@ -187,12 +187,12 @@ class AnimateDiffTransformer3D(nn.Module):
         hidden_states = self.norm(hidden_states)
         hidden_states = hidden_states.permute(0, 3, 4, 2, 1).reshape(batch_size * height * width, num_frames, channel)
 
-        hidden_states = self.proj_in(hidden_states)
+        hidden_states = self.proj_in(input=hidden_states)
 
         # 2. Blocks
         for block in self.transformer_blocks:
             hidden_states = block(
-                hidden_states,
+                hidden_states=hidden_states,
                 encoder_hidden_states=encoder_hidden_states,
                 timestep=timestep,
                 cross_attention_kwargs=cross_attention_kwargs,
@@ -200,7 +200,7 @@ class AnimateDiffTransformer3D(nn.Module):
             )
 
         # 3. Output
-        hidden_states = self.proj_out(hidden_states)
+        hidden_states = self.proj_out(input=hidden_states)
         hidden_states = (
             hidden_states[None, None, :]
             .reshape(batch_size, height, width, num_frames, channel)
@@ -344,7 +344,7 @@ class DownBlockMotion(nn.Module):
                     )
 
             else:
-                hidden_states = resnet(hidden_states, temb)
+                hidden_states = resnet(input_tensor=hidden_states, temb=temb)
 
             hidden_states = motion_module(hidden_states, num_frames=num_frames)
 
@@ -352,7 +352,7 @@ class DownBlockMotion(nn.Module):
 
         if self.downsamplers is not None:
             for downsampler in self.downsamplers:
-                hidden_states = downsampler(hidden_states)
+                hidden_states = downsampler(hidden_states=hidden_states)
 
             output_states = output_states + (hidden_states,)
 
@@ -531,25 +531,18 @@ class CrossAttnDownBlockMotion(nn.Module):
                     temb,
                     **ckpt_kwargs,
                 )
-                hidden_states = attn(
-                    hidden_states,
-                    encoder_hidden_states=encoder_hidden_states,
-                    cross_attention_kwargs=cross_attention_kwargs,
-                    attention_mask=attention_mask,
-                    encoder_attention_mask=encoder_attention_mask,
-                    return_dict=False,
-                )[0]
             else:
-                hidden_states = resnet(hidden_states, temb)
+                hidden_states = resnet(input_tensor=hidden_states, temb=temb)
+
+            hidden_states = attn(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                cross_attention_kwargs=cross_attention_kwargs,
+                attention_mask=attention_mask,
+                encoder_attention_mask=encoder_attention_mask,
+                return_dict=False,
+            )[0]
 
-                hidden_states = attn(
-                    hidden_states,
-                    encoder_hidden_states=encoder_hidden_states,
-                    cross_attention_kwargs=cross_attention_kwargs,
-                    attention_mask=attention_mask,
-                    encoder_attention_mask=encoder_attention_mask,
-                    return_dict=False,
-                )[0]
             hidden_states = motion_module(
                 hidden_states,
                 num_frames=num_frames,
@@ -563,7 +556,7 @@ class CrossAttnDownBlockMotion(nn.Module):
 
         if self.downsamplers is not None:
             for downsampler in self.downsamplers:
-                hidden_states = downsampler(hidden_states)
+                hidden_states = downsampler(hidden_states=hidden_states)
 
             output_states = output_states + (hidden_states,)
 
@@ -757,25 +750,18 @@ class CrossAttnUpBlockMotion(nn.Module):
                     temb,
                     **ckpt_kwargs,
                 )
-                hidden_states = attn(
-                    hidden_states,
-                    encoder_hidden_states=encoder_hidden_states,
-                    cross_attention_kwargs=cross_attention_kwargs,
-                    attention_mask=attention_mask,
-                    encoder_attention_mask=encoder_attention_mask,
-                    return_dict=False,
-                )[0]
             else:
-                hidden_states = resnet(hidden_states, temb)
+                hidden_states = resnet(input_tensor=hidden_states, temb=temb)
+
+            hidden_states = attn(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                cross_attention_kwargs=cross_attention_kwargs,
+                attention_mask=attention_mask,
+                encoder_attention_mask=encoder_attention_mask,
+                return_dict=False,
+            )[0]
 
-                hidden_states = attn(
-                    hidden_states,
-                    encoder_hidden_states=encoder_hidden_states,
-                    cross_attention_kwargs=cross_attention_kwargs,
-                    attention_mask=attention_mask,
-                    encoder_attention_mask=encoder_attention_mask,
-                    return_dict=False,
-                )[0]
             hidden_states = motion_module(
                 hidden_states,
                 num_frames=num_frames,
@@ -783,7 +769,7 @@ class CrossAttnUpBlockMotion(nn.Module):
 
         if self.upsamplers is not None:
             for upsampler in self.upsamplers:
-                hidden_states = upsampler(hidden_states, upsample_size)
+                hidden_states = upsampler(hidden_states=hidden_states, output_size=upsample_size)
 
         return hidden_states
 
@@ -929,13 +915,13 @@ class UpBlockMotion(nn.Module):
                         create_custom_forward(resnet), hidden_states, temb
                     )
             else:
-                hidden_states = resnet(hidden_states, temb)
+                hidden_states = resnet(input_tensor=hidden_states, temb=temb)
 
             hidden_states = motion_module(hidden_states, num_frames=num_frames)
 
         if self.upsamplers is not None:
             for upsampler in self.upsamplers:
-                hidden_states = upsampler(hidden_states, upsample_size)
+                hidden_states = upsampler(hidden_states=hidden_states, output_size=upsample_size)
 
         return hidden_states
 
@@ -1080,10 +1066,19 @@ class UNetMidBlockCrossAttnMotion(nn.Module):
             if cross_attention_kwargs.get("scale", None) is not None:
                 logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")
 
-        hidden_states = self.resnets[0](hidden_states, temb)
+        hidden_states = self.resnets[0](input_tensor=hidden_states, temb=temb)
 
         blocks = zip(self.attentions, self.resnets[1:], self.motion_modules)
         for attn, resnet, motion_module in blocks:
+            hidden_states = attn(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                cross_attention_kwargs=cross_attention_kwargs,
+                attention_mask=attention_mask,
+                encoder_attention_mask=encoder_attention_mask,
+                return_dict=False,
+            )[0]
+
             if self.training and self.gradient_checkpointing:
 
                 def create_custom_forward(module, return_dict=None):
@@ -1096,14 +1091,6 @@ class UNetMidBlockCrossAttnMotion(nn.Module):
                     return custom_forward
 
                 ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
-                hidden_states = attn(
-                    hidden_states,
-                    encoder_hidden_states=encoder_hidden_states,
-                    cross_attention_kwargs=cross_attention_kwargs,
-                    attention_mask=attention_mask,
-                    encoder_attention_mask=encoder_attention_mask,
-                    return_dict=False,
-                )[0]
                 hidden_states = torch.utils.checkpoint.checkpoint(
                     create_custom_forward(motion_module),
                     hidden_states,
@@ -1117,19 +1104,11 @@ class UNetMidBlockCrossAttnMotion(nn.Module):
                     **ckpt_kwargs,
                 )
             else:
-                hidden_states = attn(
-                    hidden_states,
-                    encoder_hidden_states=encoder_hidden_states,
-                    cross_attention_kwargs=cross_attention_kwargs,
-                    attention_mask=attention_mask,
-                    encoder_attention_mask=encoder_attention_mask,
-                    return_dict=False,
-                )[0]
                 hidden_states = motion_module(
                     hidden_states,
                     num_frames=num_frames,
                 )
-                hidden_states = resnet(hidden_states, temb)
+                hidden_states = resnet(input_tensor=hidden_states, temb=temb)
 
         return hidden_states
 
@@ -2178,7 +2157,6 @@ class UNetMotionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, Peft
 
         emb = emb if aug_emb is None else emb + aug_emb
         emb = emb.repeat_interleave(repeats=num_frames, dim=0)
-        encoder_hidden_states = encoder_hidden_states.repeat_interleave(repeats=num_frames, dim=0)
 
         if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
             if "image_embeds" not in added_cond_kwargs:

diffusers/models/upsampling.py

@@ -19,6 +19,7 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 from ..utils import deprecate
+from ..utils.import_utils import is_torch_version
 from .normalization import RMSNorm
 
 
@@ -151,11 +152,10 @@ class Upsample2D(nn.Module):
         if self.use_conv_transpose:
             return self.conv(hidden_states)
 
-        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
-        # TODO(Suraj): Remove this cast once the issue is fixed in PyTorch
-        # https://github.com/pytorch/pytorch/issues/86679
+        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16 until PyTorch 2.1
+        # https://github.com/pytorch/pytorch/issues/86679#issuecomment-1783978767
         dtype = hidden_states.dtype
-        if dtype == torch.bfloat16:
+        if dtype == torch.bfloat16 and is_torch_version("<", "2.1"):
             hidden_states = hidden_states.to(torch.float32)
 
         # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
@@ -170,8 +170,8 @@ class Upsample2D(nn.Module):
             else:
                 hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest")
 
-        # If the input is bfloat16, we cast back to bfloat16
-        if dtype == torch.bfloat16:
+        # Cast back to original dtype
+        if dtype == torch.bfloat16 and is_torch_version("<", "2.1"):
             hidden_states = hidden_states.to(dtype)
 
         # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed