Upload 4 files

src/attention_wan_nag.py

@@ -0,0 +1,114 @@
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+
+from diffusers.models.attention_processor import Attention
+from ftfy import apply_plan
+
+
+class NAGWanAttnProcessor2_0:
+    def __init__(self, nag_scale=1.0, nag_tau=2.5, nag_alpha=0.25):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("WanAttnProcessor2_0 requires PyTorch 2.0. To use it, please upgrade PyTorch to 2.0.")
+        self.nag_scale = nag_scale
+        self.nag_tau = nag_tau
+        self.nag_alpha = nag_alpha
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        apply_guidance = self.nag_scale > 1 and encoder_hidden_states is not None
+        if apply_guidance:
+            if len(encoder_hidden_states) == 2 * len(hidden_states):
+                batch_size = len(hidden_states)
+            else:
+                apply_guidance = False
+
+        encoder_hidden_states_img = None
+        if attn.add_k_proj is not None:
+            encoder_hidden_states_img = encoder_hidden_states[:, :257]
+            encoder_hidden_states = encoder_hidden_states[:, 257:]
+            if apply_guidance:
+                encoder_hidden_states_img = encoder_hidden_states_img[:batch_size]
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        query = query.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        key = key.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        value = value.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+        if rotary_emb is not None:
+
+            def apply_rotary_emb(hidden_states: torch.Tensor, freqs: torch.Tensor):
+                x_rotated = torch.view_as_complex(hidden_states.to(torch.float64).unflatten(3, (-1, 2)))
+                x_out = torch.view_as_real(x_rotated * freqs).flatten(3, 4)
+                return x_out.type_as(hidden_states)
+
+            query = apply_rotary_emb(query, rotary_emb)
+            key = apply_rotary_emb(key, rotary_emb)
+
+        # I2V task
+        hidden_states_img = None
+        if encoder_hidden_states_img is not None:
+            key_img = attn.add_k_proj(encoder_hidden_states_img)
+            key_img = attn.norm_added_k(key_img)
+            value_img = attn.add_v_proj(encoder_hidden_states_img)
+
+            key_img = key_img.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+            value_img = value_img.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+            hidden_states_img = F.scaled_dot_product_attention(
+                query, key_img, value_img, attn_mask=None, dropout_p=0.0, is_causal=False
+            )
+            hidden_states_img = hidden_states_img.transpose(1, 2).flatten(2, 3)
+            hidden_states_img = hidden_states_img.type_as(query)
+
+        if apply_guidance:
+            key, key_negative = torch.chunk(key, 2, dim=0)
+            value, value_negative = torch.chunk(value, 2, dim=0)
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+        hidden_states = hidden_states.transpose(1, 2).flatten(2, 3)
+        hidden_states = hidden_states.type_as(query)
+        if apply_guidance:
+            hidden_states_negative = F.scaled_dot_product_attention(
+                query, key_negative, value_negative, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+            )
+            hidden_states_negative = hidden_states_negative.transpose(1, 2).flatten(2, 3)
+            hidden_states_negative = hidden_states_negative.type_as(query)
+
+            hidden_states_positive = hidden_states
+
+            hidden_states_guidance = hidden_states_positive * self.nag_scale - hidden_states_negative * (self.nag_scale - 1)
+            norm_positive = torch.norm(hidden_states_positive, p=1, dim=-1, keepdim=True).expand(*hidden_states_positive.shape)
+            norm_guidance = torch.norm(hidden_states_guidance, p=1, dim=-1, keepdim=True).expand(*hidden_states_guidance.shape)
+
+            scale = norm_guidance / norm_positive
+            scale = torch.nan_to_num(scale, 10)
+            hidden_states_guidance[scale > self.nag_tau] = \
+                hidden_states_guidance[scale > self.nag_tau] / (norm_guidance[scale > self.nag_tau] + 1e-7) * norm_positive[scale > self.nag_tau] * self.nag_tau
+
+            hidden_states = hidden_states_guidance * self.nag_alpha + hidden_states_positive * (1 - self.nag_alpha)
+
+        if hidden_states_img is not None:
+            hidden_states = hidden_states + hidden_states_img
+
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+        return hidden_states

src/pipeline_wan_nag.py

@@ -0,0 +1,295 @@
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.pipelines.wan.pipeline_output import WanPipelineOutput
+from diffusers.pipelines.wan.pipeline_wan import WanPipeline
+
+from src.attention_wan_nag import NAGWanAttnProcessor2_0
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class NAGWanPipeline(WanPipeline):
+    @property
+    def do_normalized_attention_guidance(self):
+        return self._nag_scale > 1
+
+    def _set_nag_attn_processor(self, nag_scale, nag_tau, nag_alpha):
+        attn_procs = {}
+        for name, origin_attn_proc in self.transformer.attn_processors.items():
+            if "attn2" in name:
+                attn_procs[name] = NAGWanAttnProcessor2_0(nag_scale=nag_scale, nag_tau=nag_tau, nag_alpha=nag_alpha)
+            else:
+                attn_procs[name] = origin_attn_proc
+        self.transformer.set_attn_processor(attn_procs)
+
+    @torch.no_grad()
+    def __call__(
+            self,
+            prompt: Union[str, List[str]] = None,
+            negative_prompt: Union[str, List[str]] = None,
+            height: int = 480,
+            width: int = 832,
+            num_frames: int = 81,
+            num_inference_steps: int = 50,
+            guidance_scale: float = 5.0,
+            num_videos_per_prompt: Optional[int] = 1,
+            generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+            latents: Optional[torch.Tensor] = None,
+            prompt_embeds: Optional[torch.Tensor] = None,
+            negative_prompt_embeds: Optional[torch.Tensor] = None,
+            output_type: Optional[str] = "np",
+            return_dict: bool = True,
+            attention_kwargs: Optional[Dict[str, Any]] = None,
+            callback_on_step_end: Optional[
+                Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+            ] = None,
+            callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+            max_sequence_length: int = 512,
+
+            nag_scale: float = 1.0,
+            nag_tau: float = 2.5,
+            nag_alpha: float = 0.25,
+            nag_negative_prompt: str = None,
+            nag_negative_prompt_embeds: Optional[torch.Tensor] = None,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, defaults to `480`):
+                The height in pixels of the generated image.
+            width (`int`, defaults to `832`):
+                The width in pixels of the generated image.
+            num_frames (`int`, defaults to `81`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, defaults to `50`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, defaults to `5.0`):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            autocast_dtype (`torch.dtype`, *optional*, defaults to `torch.bfloat16`):
+                The dtype to use for the torch.amp.autocast.
+
+        Examples:
+
+        Returns:
+            [`~WanPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where
+                the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            negative_prompt,
+            height,
+            width,
+            prompt_embeds,
+            negative_prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+        self._nag_scale = nag_scale
+
+        device = self._execution_device
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if self.do_normalized_attention_guidance:
+            if nag_negative_prompt_embeds is None:
+                if nag_negative_prompt is None:
+                    if self.do_classifier_free_guidance:
+                        nag_negative_prompt_embeds = negative_prompt_embeds
+                    else:
+                        nag_negative_prompt = negative_prompt or ""
+
+                if nag_negative_prompt is not None:
+                    nag_negative_prompt_embeds = self.encode_prompt(
+                        prompt=nag_negative_prompt,
+                        do_classifier_free_guidance=False,
+                        num_videos_per_prompt=num_videos_per_prompt,
+                        max_sequence_length=max_sequence_length,
+                        device=device,
+                    )[0]
+
+        if self.do_normalized_attention_guidance:
+            prompt_embeds = torch.cat([prompt_embeds, nag_negative_prompt_embeds], dim=0)
+
+        transformer_dtype = self.transformer.dtype
+        prompt_embeds = prompt_embeds.to(transformer_dtype)
+        if negative_prompt_embeds is not None:
+            negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype)
+
+        # 4. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+        
+        if self.do_normalized_attention_guidance:
+            origin_attn_procs = self.transformer.attn_processors
+            self._set_nag_attn_processor(nag_scale, nag_tau, nag_alpha)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                latent_model_input = latents.to(transformer_dtype)
+                timestep = t.expand(latents.shape[0])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                if self.do_classifier_free_guidance:
+                    noise_uncond = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=negative_prompt_embeds,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            latents = latents.to(self.vae.dtype)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+        else:
+            video = latents
+
+        if self.do_normalized_attention_guidance:
+            self.transformer.set_attn_processor(origin_attn_procs)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return WanPipelineOutput(frames=video)

src/transformer_wan_nag.py

@@ -0,0 +1,160 @@
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+
+from diffusers.utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.transformers.transformer_wan import WanTransformer3DModel
+from diffusers.models.attention_processor import AttentionProcessor
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class NagWanTransformer3DModel(WanTransformer3DModel):
+    @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,
+        timestep: torch.LongTensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_image: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+    ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
+        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."
+                )
+
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.config.patch_size
+        post_patch_num_frames = num_frames // p_t
+        post_patch_height = height // p_h
+        post_patch_width = width // p_w
+
+        rotary_emb = self.rope(hidden_states)
+
+        hidden_states = self.patch_embedding(hidden_states)
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+
+        temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
+            timestep, encoder_hidden_states, encoder_hidden_states_image
+        )
+        timestep_proj = timestep_proj.unflatten(1, (6, -1))
+
+        if encoder_hidden_states_image is not None:
+            bs_encoder_hidden_states = len(encoder_hidden_states)
+            bs_encoder_hidden_states_image = len(encoder_hidden_states_image)
+            bs_scale = bs_encoder_hidden_states / bs_encoder_hidden_states_image
+            assert bs_scale in [1, 2, 3]
+            if bs_scale != 1:
+                encoder_hidden_states_image = encoder_hidden_states_image.tile(int(bs_scale), 1, 1)
+            encoder_hidden_states = torch.concat([encoder_hidden_states_image, encoder_hidden_states], dim=1)
+
+        # 4. Transformer blocks
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for block in self.blocks:
+                hidden_states = self._gradient_checkpointing_func(
+                    block, hidden_states, encoder_hidden_states, timestep_proj, rotary_emb
+                )
+        else:
+            for block in self.blocks:
+                hidden_states = block(hidden_states, encoder_hidden_states, timestep_proj, rotary_emb)
+
+        # 5. Output norm, projection & unpatchify
+        shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, dim=1)
+
+        # Move the shift and scale tensors to the same device as hidden_states.
+        # When using multi-GPU inference via accelerate these will be on the
+        # first device rather than the last device, which hidden_states ends up
+        # on.
+        shift = shift.to(hidden_states.device)
+        scale = scale.to(hidden_states.device)
+
+        hidden_states = (self.norm_out(hidden_states.float()) * (1 + scale) + shift).type_as(hidden_states)
+        hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states.reshape(
+            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, p_t, p_h, p_w, -1
+        )
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)