TrainableSDPipeline.py

from diffusers import StableDiffusionPipeline, StableDiffusionXLPipeline
from typing import Any, Callable, Dict, List, Optional, Union, Tuple

import torch
from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import rescale_noise_cfg

from diffusers.loaders import LoraLoaderMixin, TextualInversionLoaderMixin
from diffusers.models.lora import adjust_lora_scale_text_encoder
from diffusers.utils import (
    USE_PEFT_BACKEND,
    scale_lora_layers,
    unscale_lora_layers,
)
from torch.nn.parallel import DistributedDataParallel as DDP

class TrainableSDPipeline(StableDiffusionPipeline):


    # this is modified from the __call__ method of the StableDiffusionPipeline class
    def forward(
        self,
        prompt: Union[str, List[str]] = None,
        height: int = 512,
        width: int = 512,
        training_timesteps: Optional[List[int]] = [], # new training parameter
        early_exit: bool = False, # new training parameter
        detach_gradient: bool = True, # new training parameter
        train_text_encoder: bool = False, # new training parameter
        double_laststep: bool = False, # new training parameter
        bp_on_trained: bool = False, # new training parameter
        fast_training: bool = False, # new training parameter
        num_inference_steps: int = 50,
        guidance_scale: float = 7.5,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.FloatTensor] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        output_type: Optional[str] = "image",
        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
        callback_steps: int = 1,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        guidance_rescale: float = 0.0,
        return_latents: bool = False, # new training parameter
        return_timestamped_latents: bool = False, # new training parameter
        return_early: bool = False,
        D_timesteps=None,
        batch=None,
    ):

        # 1. Check inputs. Raise error if not correct
        # self.check_inputs(
        #     prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
        # )

        # 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]

        device = self._execution_device
        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
        # corresponds to doing no classifier free guidance.
        do_classifier_free_guidance = guidance_scale > 1.0

        torch.set_grad_enabled(train_text_encoder)
        # 3. Encode input prompt
        text_encoder_lora_scale = (
            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
        )
        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
            prompt,
            device,
            num_images_per_prompt,
            do_classifier_free_guidance,
            negative_prompt,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            lora_scale=text_encoder_lora_scale,
        )
        # For classifier free guidance, we need to do two forward passes.
        # Here we concatenate the unconditional and text embeddings into a single batch
        # to avoid doing two forward passes
        if do_classifier_free_guidance:
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
        prompt_embeds = prompt_embeds

        # 4. Prepare timesteps
        self.scheduler.set_timesteps(num_inference_steps, device=device)
        if fast_training:
            self.scheduler.timesteps = [self.scheduler.timesteps[i] for i in training_timesteps]
            training_timesteps = range(len(self.scheduler.timesteps))
        timesteps = self.scheduler.timesteps

        # 5. Prepare latent variables
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            4,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)

        # prepare timestamped latents
        if return_timestamped_latents:
            timesteps_list = timesteps.cpu().numpy().tolist()
            if D_timesteps is not None:
                selected_timestamps = latents.new_tensor([timesteps_list.index(t) for t in D_timesteps], dtype=torch.long).reshape(batch_size,)
            else:
                selected_timestamps = torch.randint(0, len(timesteps), (batch_size, ), device=latents.device)
                selected_timestamps = selected_timestamps.long()
            ret_latents = torch.zeros_like(latents)
            ret_count = latents.new_zeros((batch_size,), dtype=torch.int)
            ret_timestamps = latents.new_zeros((batch_size,), dtype=torch.long)

        torch.set_grad_enabled(True)

        # 7. Denoising loop
        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
        for i, t in enumerate(timesteps):            
            torch.set_grad_enabled((len(training_timesteps) == 0 or i > min(training_timesteps)) and not double_laststep)
            # expand the latents if we are doing classifier free guidance
            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

            torch.set_grad_enabled(i in training_timesteps and not double_laststep)

            do_detach = detach_gradient
            if i in training_timesteps and bp_on_trained:
                do_detach = False
            # predict the noise residual
            noise_pred = self.unet(
                latent_model_input.detach() if do_detach else latent_model_input,
                t,
                encoder_hidden_states=prompt_embeds,
                cross_attention_kwargs=cross_attention_kwargs,
                return_dict=False,
            )[0]

            noise_pred = noise_pred.to(prompt_embeds.dtype)

            # perform guidance
            if do_classifier_free_guidance:
                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)

            if do_classifier_free_guidance and guidance_rescale > 0.0:
                # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
                noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)

            torch.set_grad_enabled((len(training_timesteps) == 0 or i >= min(training_timesteps)) and not double_laststep)

            # compute the previous noisy sample x_t -> x_t-1
            out = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=True)
            latents = out.prev_sample
            intermediate_result = out.pred_original_sample if hasattr(out, "pred_original_sample") else None

            # call the callback, if provided
            if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                if callback is not None and i % callback_steps == 0:
                    callback(i, t, latents)

            if len(training_timesteps) > 0 and i == max(training_timesteps) and early_exit:
                latents = intermediate_result
                break

            if return_timestamped_latents:
                valid_index = selected_timestamps == i
                ret_latents[valid_index] = latents[valid_index]
                ret_count[valid_index] = 1
                ret_timestamps[valid_index] = t
                if return_early and ret_count.sum()==batch_size:
                    return None, ret_latents, ret_timestamps

            
        if double_laststep:
            torch.set_grad_enabled(double_laststep)
            t = timesteps[training_timesteps[0]]
            noise = torch.randn_like(latents)
            noisy_latents = self.scheduler.add_noise(latents, noise, t)
            
            noisy_model_input = torch.cat([noisy_latents] * 2) if do_classifier_free_guidance else noisy_latents
            noisy_model_input = self.scheduler.scale_model_input(noisy_model_input, t)
            noise_pred = self.unet(
                noisy_model_input.detach() if do_detach else latent_model_input,
                t,
                encoder_hidden_states=prompt_embeds,
                cross_attention_kwargs=cross_attention_kwargs,
                return_dict=False,
            )[0]
            # perform guidance
            if do_classifier_free_guidance:
                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)

            if do_classifier_free_guidance and guidance_rescale > 0.0:
                # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
                noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
            
            out = self.scheduler.step(noise_pred, t, noisy_latents, **extra_step_kwargs, return_dict=True)
            latents = out.prev_sample

        if return_timestamped_latents:
            image = self.vae.decode(latents.to(self.vae.dtype) / self.vae.config.scaling_factor, return_dict=False)[0]
            return (image / 2 + 0.5), ret_latents, ret_timestamps
                     
        if output_type == "image":
            image = self.vae.decode(latents.to(self.vae.dtype) / self.vae.config.scaling_factor, return_dict=False)[0]
            if return_latents:
                return (image / 2 + 0.5), latents
            return (image / 2 + 0.5) # .clamp(0, 1)
        elif output_type == "latent":
            return latents
    
    def encode_prompt(
        self,
        prompt,
        device,
        num_images_per_prompt,
        do_classifier_free_guidance,
        negative_prompt=None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        lora_scale: Optional[float] = None,
        clip_skip: Optional[int] = None,
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            device: (`torch.device`):
                torch device
            num_images_per_prompt (`int`):
                number of images that should be generated per prompt
            do_classifier_free_guidance (`bool`):
                whether to use classifier free guidance or not
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            lora_scale (`float`, *optional*):
                A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
            clip_skip (`int`, *optional*):
                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                the output of the pre-final layer will be used for computing the prompt embeddings.
        """
        # set lora scale so that monkey patched LoRA
        # function of text encoder can correctly access it
        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
            self._lora_scale = lora_scale

            # dynamically adjust the LoRA scale
            if not USE_PEFT_BACKEND:
                adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
            else:
                scale_lora_layers(self.text_encoder, lora_scale)

        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]

        if prompt_embeds is None:
            # textual inversion: procecss multi-vector tokens if necessary
            if isinstance(self, TextualInversionLoaderMixin):
                prompt = self.maybe_convert_prompt(prompt, self.tokenizer)

            text_inputs = self.tokenizer(
                prompt,
                padding="max_length",
                max_length=self.tokenizer.model_max_length,
                truncation=True,
                return_tensors="pt",
            )
            text_input_ids = text_inputs.input_ids
            untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids

            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
                text_input_ids, untruncated_ids
            ):
                removed_text = self.tokenizer.batch_decode(
                    untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
                )
                # logger.warning(
                print(
                    "The following part of your input was truncated because CLIP can only handle sequences up to"
                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
                )
            
            # MOD: accomodate for training text encoder
            if isinstance(self.text_encoder, DDP):
                if hasattr(self.text_encoder.module.config, "use_attention_mask") and self.text_encoder.module.config.use_attention_mask:
                    attention_mask = text_inputs.attention_mask.to(device)
                else:
                    attention_mask = None
            else:
                if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
                    attention_mask = text_inputs.attention_mask.to(device)
                else:
                    attention_mask = None

            if clip_skip is None:
                prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
                prompt_embeds = prompt_embeds[0]
            else:
                prompt_embeds = self.text_encoder(
                    text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
                )
                # Access the `hidden_states` first, that contains a tuple of
                # all the hidden states from the encoder layers. Then index into
                # the tuple to access the hidden states from the desired layer.
                prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
                # We also need to apply the final LayerNorm here to not mess with the
                # representations. The `last_hidden_states` that we typically use for
                # obtaining the final prompt representations passes through the LayerNorm
                # layer.
                prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)

        if self.text_encoder is not None and isinstance(self.text_encoder, DDP):
            prompt_embeds_dtype = self.text_encoder.module.dtype
        elif self.text_encoder is not None:
            prompt_embeds_dtype = self.text_encoder.dtype
        elif self.unet is not None:
            prompt_embeds_dtype = self.unet.dtype
        else:
            prompt_embeds_dtype = prompt_embeds.dtype

        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)

        bs_embed, seq_len, _ = prompt_embeds.shape
        # duplicate text embeddings for each generation per prompt, using mps friendly method
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)

        # get unconditional embeddings for classifier free guidance
        if do_classifier_free_guidance and negative_prompt_embeds is None:
            uncond_tokens: List[str]
            if negative_prompt is None:
                uncond_tokens = [""] * batch_size
            elif prompt is not None and type(prompt) is not type(negative_prompt):
                raise TypeError(
                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
                    f" {type(prompt)}."
                )
            elif isinstance(negative_prompt, str):
                uncond_tokens = [negative_prompt]
            elif batch_size != len(negative_prompt):
                raise ValueError(
                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
                    " the batch size of `prompt`."
                )
            else:
                uncond_tokens = negative_prompt

            # textual inversion: procecss multi-vector tokens if necessary
            if isinstance(self, TextualInversionLoaderMixin):
                uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)

            max_length = prompt_embeds.shape[1]
            uncond_input = self.tokenizer(
                uncond_tokens,
                padding="max_length",
                max_length=max_length,
                truncation=True,
                return_tensors="pt",
            )

            # MOD: accomodate for training text encoder
            if isinstance(self.text_encoder, DDP):
                if hasattr(self.text_encoder.module.config, "use_attention_mask") and self.text_encoder.module.config.use_attention_mask:
                    attention_mask = text_inputs.attention_mask.to(device)
                else:
                    attention_mask = None
            else:
                if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
                    attention_mask = text_inputs.attention_mask.to(device)
                else:
                    attention_mask = None


            negative_prompt_embeds = self.text_encoder(
                uncond_input.input_ids.to(device),
                attention_mask=attention_mask,
            )
            negative_prompt_embeds = negative_prompt_embeds[0]

        if do_classifier_free_guidance:
            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
            seq_len = negative_prompt_embeds.shape[1]

            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)

            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

        if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
            # Retrieve the original scale by scaling back the LoRA layers
            unscale_lora_layers(self.text_encoder, lora_scale)

        return prompt_embeds, negative_prompt_embeds
    

class TrainableSDXLPipeline(StableDiffusionXLPipeline):
    def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype):
        from torch.nn.parallel import DistributedDataParallel
        add_time_ids = list(original_size + crops_coords_top_left + target_size)

        if isinstance(self.unet, DistributedDataParallel):
            passed_add_embed_dim = (
                self.unet.module.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim
            )
            expected_add_embed_dim = self.unet.module.add_embedding.linear_1.in_features
        else:
            passed_add_embed_dim = (
                self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim
            )
            expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features

        if expected_add_embed_dim != passed_add_embed_dim:
            raise ValueError(
                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
            )

        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
        return add_time_ids
    
    @torch.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        prompt_2: Optional[Union[str, List[str]]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 50,
        denoising_end: Optional[float] = None,
        guidance_scale: float = 5.0,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        negative_prompt_2: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.FloatTensor] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
        callback_steps: int = 1,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        guidance_rescale: float = 0.0,
        original_size: Optional[Tuple[int, int]] = None,
        crops_coords_top_left: Tuple[int, int] = (0, 0),
        target_size: Optional[Tuple[int, int]] = None,
        negative_original_size: Optional[Tuple[int, int]] = None,
        negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
        negative_target_size: Optional[Tuple[int, int]] = None,
    ):
        from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
        
        # 0. Default height and width to unet
        height = height or self.default_sample_size * self.vae_scale_factor
        width = width or self.default_sample_size * self.vae_scale_factor

        original_size = original_size or (height, width)
        target_size = target_size or (height, width)

        # 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]

        device = self._execution_device

        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
        # corresponds to doing no classifier free guidance.
        do_classifier_free_guidance = guidance_scale > 1.0

        # 3. Encode input prompt
        text_encoder_lora_scale = (
            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
        )
        (
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
        ) = self.encode_prompt(
            prompt=prompt,
            prompt_2=prompt_2,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            do_classifier_free_guidance=do_classifier_free_guidance,
            negative_prompt=negative_prompt,
            negative_prompt_2=negative_prompt_2,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            lora_scale=text_encoder_lora_scale,
        )

        # 4. Prepare timesteps
        self.scheduler.set_timesteps(num_inference_steps, device=device)

        timesteps = self.scheduler.timesteps

        # 5. Prepare latent variables
        num_channels_latents = self.unet.config.in_channels
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)

        # 7. Prepare added time ids & embeddings
        add_text_embeds = pooled_prompt_embeds
        add_time_ids = self._get_add_time_ids(
            original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
        )
        if negative_original_size is not None and negative_target_size is not None:
            negative_add_time_ids = self._get_add_time_ids(
                negative_original_size,
                negative_crops_coords_top_left,
                negative_target_size,
                dtype=prompt_embeds.dtype,
            )
        else:
            negative_add_time_ids = add_time_ids

        if do_classifier_free_guidance:
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
            add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)

        prompt_embeds = prompt_embeds.to(device)
        add_text_embeds = add_text_embeds.to(device)
        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)

        # 8. Denoising loop
        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)

        # 7.1 Apply denoising_end
        if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1:
            discrete_timestep_cutoff = int(
                round(
                    self.scheduler.config.num_train_timesteps
                    - (denoising_end * self.scheduler.config.num_train_timesteps)
                )
            )
            num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
            timesteps = timesteps[:num_inference_steps]

        
        for i, t in enumerate(timesteps):
            # expand the latents if we are doing classifier free guidance
            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents

            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

            # predict the noise residual
            added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
            noise_pred = self.unet(
                latent_model_input,
                t,
                encoder_hidden_states=prompt_embeds,
                cross_attention_kwargs=cross_attention_kwargs,
                added_cond_kwargs=added_cond_kwargs,
                return_dict=False,
            )[0]

            # perform guidance
            if do_classifier_free_guidance:
                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)

            if do_classifier_free_guidance and guidance_rescale > 0.0:
                # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
                noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)

            # compute the previous noisy sample x_t -> x_t-1
            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]

            # call the callback, if provided
            if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                if callback is not None and i % callback_steps == 0:
                    callback(i, t, latents)

        if not output_type == "latent":
            # make sure the VAE is in float32 mode, as it overflows in float16
            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast

            if needs_upcasting:
                self.upcast_vae()
                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)

            image = self.vae.decode(latents.to(self.vae.dtype) / self.vae.config.scaling_factor, return_dict=False)[0]

            # cast back to fp16 if needed
            if needs_upcasting:
                self.vae.to(dtype=torch.float16)
        else:
            image = latents

        if not output_type == "latent":
            # apply watermark if available
            # if self.watermark is not None:
            #     image = self.watermark.apply_watermark(image)

            image = self.image_processor.postprocess(image, output_type=output_type)

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (image,)

        return StableDiffusionXLPipelineOutput(images=image)

    def forward(
        self,
        prompt: Union[str, List[str]] = None,
        prompt_2: Optional[Union[str, List[str]]] = None,
        height: int = 512,
        width: int = 512,
        training_timesteps: Optional[List[int]] = [],
        early_exit: bool = False,
        detach_gradient: bool = True,
        train_text_encoder: bool = False,
        num_inference_steps: int = 50,
        denoising_end: Optional[float] = None,
        guidance_scale: float = 5.0,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        negative_prompt_2: Optional[Union[str, List[str]]] = None,
        num_images_per_prompt: Optional[int] = 1,
        eta: float = 0.0,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.FloatTensor] = None,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        output_type: Optional[str] = "image",
        return_dict: bool = True,
        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
        callback_steps: int = 1,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        guidance_rescale: float = 0.0,
        original_size: Optional[Tuple[int, int]] = None,
        crops_coords_top_left: Tuple[int, int] = (0, 0),
        target_size: Optional[Tuple[int, int]] = None,
        return_latents: bool = False, # new training parameter
        return_timestamped_latents: bool = False, # new training parameter
        return_early: bool = False,
        D_timesteps=None,
        batch=None,
    ):
        
        original_size = original_size or (height, width)
        target_size = target_size or (height, width)
        
        # 1. Check inputs. Raise error if not correct
        # self.check_inputs(
        #     prompt,
        #     prompt_2,
        #     height,
        #     width,
        #     callback_steps,
        #     negative_prompt,
        #     negative_prompt_2,
        #     prompt_embeds,
        #     negative_prompt_embeds,
        #     pooled_prompt_embeds,
        #     negative_pooled_prompt_embeds,
        # )

        # 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]
            
        device = self._execution_device

        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
        # corresponds to doing no classifier free guidance.
        do_classifier_free_guidance = guidance_scale > 1.0
        torch.set_grad_enabled(train_text_encoder)
        # 3. Encode input prompt
        text_encoder_lora_scale = (
            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
        )
        (
            prompt_embeds,
            negative_prompt_embeds,
            pooled_prompt_embeds,
            negative_pooled_prompt_embeds,
        ) = self.encode_prompt(
            prompt=prompt,
            prompt_2=prompt_2,
            device=device,
            num_images_per_prompt=num_images_per_prompt,
            do_classifier_free_guidance=do_classifier_free_guidance,
            negative_prompt=negative_prompt,
            negative_prompt_2=negative_prompt_2,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            lora_scale=text_encoder_lora_scale,
        )
        # 4. Prepare timesteps
        self.scheduler.set_timesteps(num_inference_steps, device=device)
        timesteps = self.scheduler.timesteps

        # 5. Prepare latent variables
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            4,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
        torch.set_grad_enabled(True)
        # 7. Prepare added time ids & embeddings
        add_text_embeds = pooled_prompt_embeds
        add_time_ids = self._get_add_time_ids(
            original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype
        )

        if do_classifier_free_guidance:
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
            add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)

        prompt_embeds = prompt_embeds.to(device)
        add_text_embeds = add_text_embeds.to(device)
        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
        # 8. Denoising loop
        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)

        # 7.1 Apply denoising_end
        if denoising_end is not None and type(denoising_end) == float and denoising_end > 0 and denoising_end < 1:
            discrete_timestep_cutoff = int(
                round(
                    self.scheduler.config.num_train_timesteps
                    - (denoising_end * self.scheduler.config.num_train_timesteps)
                )
            )
            num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
            timesteps = timesteps[:num_inference_steps]

        for i, t in enumerate(timesteps):
            # expand the latents if we are doing classifier free guidance
            torch.set_grad_enabled(i > min(training_timesteps))
            latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

            torch.set_grad_enabled(i in training_timesteps)
            # predict the noise residual
            added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
            noise_pred = self.unet(
                latent_model_input.detach() if detach_gradient else latent_model_input,
                t,
                encoder_hidden_states=prompt_embeds,
                cross_attention_kwargs=cross_attention_kwargs,
                added_cond_kwargs=added_cond_kwargs,
                return_dict=False,
            )[0]

            # perform guidance
            if do_classifier_free_guidance:
                noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)

            if do_classifier_free_guidance and guidance_rescale > 0.0:
                # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
                noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)

            torch.set_grad_enabled(i >= min(training_timesteps))
            # compute the previous noisy sample x_t -> x_t-1
            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]

            # call the callback, if provided
            if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                if callback is not None and i % callback_steps == 0:
                    callback(i, t, latents)
                    
            if i == max(training_timesteps) and early_exit:
                break

        if return_latents:
            image = self.vae.decode(latents.half() / self.vae.config.scaling_factor, return_dict=False)[0]
            return image, latents.half()
    
        if output_type == "image":
            image = self.vae.decode(latents.half() / self.vae.config.scaling_factor, return_dict=False)[0]
            return (image / 2 + 0.5) # .clamp(0, 1)
        elif output_type == "latent":
            return latents