animatelcm/scheduler/lcm_scheduler.py

# Copyright 2023 Stanford University Team and The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
# and https://github.com/hojonathanho/diffusion

import math
from dataclasses import dataclass
from typing import List, Optional, Tuple, Union

import numpy as np
import torch

from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.utils import BaseOutput, logging
from diffusers.schedulers.scheduling_utils import SchedulerMixin


logger = logging.get_logger(__name__)  # pylint: disable=invalid-name



@dataclass
class LCMSchedulerOutput(BaseOutput):
    """
    Output class for the scheduler's `step` function output.

    Args:
        prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
            Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
            denoising loop.
        pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
            The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
            `pred_original_sample` can be used to preview progress or for guidance.
    """

    prev_sample: torch.FloatTensor
    denoised: Optional[torch.FloatTensor] = None


# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
def betas_for_alpha_bar(
    num_diffusion_timesteps,
    max_beta=0.999,
    alpha_transform_type="cosine",
):
    """
    Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
    (1-beta) over time from t = [0,1].

    Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
    to that part of the diffusion process.


    Args:
        num_diffusion_timesteps (`int`): the number of betas to produce.
        max_beta (`float`): the maximum beta to use; use values lower than 1 to
                     prevent singularities.
        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
                     Choose from `cosine` or `exp`

    Returns:
        betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
    """
    if alpha_transform_type == "cosine":

        def alpha_bar_fn(t):
            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2

    elif alpha_transform_type == "exp":

        def alpha_bar_fn(t):
            return math.exp(t * -12.0)

    else:
        raise ValueError(
            f"Unsupported alpha_tranform_type: {alpha_transform_type}")

    betas = []
    for i in range(num_diffusion_timesteps):
        t1 = i / num_diffusion_timesteps
        t2 = (i + 1) / num_diffusion_timesteps
        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
    return torch.tensor(betas, dtype=torch.float32)


# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
def rescale_zero_terminal_snr(betas: torch.FloatTensor) -> torch.FloatTensor:
    """
    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)


    Args:
        betas (`torch.FloatTensor`):
            the betas that the scheduler is being initialized with.

    Returns:
        `torch.FloatTensor`: rescaled betas with zero terminal SNR
    """
    # Convert betas to alphas_bar_sqrt
    alphas = 1.0 - betas
    alphas_cumprod = torch.cumprod(alphas, dim=0)
    alphas_bar_sqrt = alphas_cumprod.sqrt()

    # Store old values.
    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()

    # Shift so the last timestep is zero.
    alphas_bar_sqrt -= alphas_bar_sqrt_T

    # Scale so the first timestep is back to the old value.
    alphas_bar_sqrt *= alphas_bar_sqrt_0 / \
        (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)

    # Convert alphas_bar_sqrt to betas
    alphas_bar = alphas_bar_sqrt**2  # Revert sqrt
    alphas = alphas_bar[1:] / alphas_bar[:-1]  # Revert cumprod
    alphas = torch.cat([alphas_bar[0:1], alphas])
    betas = 1 - alphas

    return betas


def randn_tensor(
    shape: Union[Tuple, List],
    generator: Optional[Union[List["torch.Generator"],
                              "torch.Generator"]] = None,
    device: Optional["torch.device"] = None,
    dtype: Optional["torch.dtype"] = None,
    layout: Optional["torch.layout"] = None,
):
    """A helper function to create random tensors on the desired `device` with the desired `dtype`. When
    passing a list of generators, you can seed each batch size individually. If CPU generators are passed, the tensor
    is always created on the CPU.
    """
    # device on which tensor is created defaults to device
    rand_device = device
    batch_size = shape[0]

    layout = layout or torch.strided
    device = device or torch.device("cpu")

    if generator is not None:
        gen_device_type = generator.device.type if not isinstance(
            generator, list) else generator[0].device.type
        if gen_device_type != device.type and gen_device_type == "cpu":
            rand_device = "cpu"
            if device != "mps":
                logger.info(
                    f"The passed generator was created on 'cpu' even though a tensor on {device} was expected."
                    f" Tensors will be created on 'cpu' and then moved to {device}. Note that one can probably"
                    f" slighly speed up this function by passing a generator that was created on the {device} device."
                )
        elif gen_device_type != device.type and gen_device_type == "cuda":
            raise ValueError(
                f"Cannot generate a {device} tensor from a generator of type {gen_device_type}.")

    # make sure generator list of length 1 is treated like a non-list
    if isinstance(generator, list) and len(generator) == 1:
        generator = generator[0]

    if isinstance(generator, list):
        shape = (1,) + shape[1:]
        latents = [
            torch.randn(
                shape, generator=generator[i], device=rand_device, dtype=dtype, layout=layout)
            for i in range(batch_size)
        ]
        latents = torch.cat(latents, dim=0).to(device)
    else:
        latents = torch.randn(shape, generator=generator,
                              device=rand_device, dtype=dtype, layout=layout).to(device)

    return latents


class LCMScheduler(SchedulerMixin, ConfigMixin):
    """
    `LCMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with
    non-Markovian guidance.

    This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. [`~ConfigMixin`] takes care of storing all config
    attributes that are passed in the scheduler's `__init__` function, such as `num_train_timesteps`. They can be
    accessed via `scheduler.config.num_train_timesteps`. [`SchedulerMixin`] provides general loading and saving
    functionality via the [`SchedulerMixin.save_pretrained`] and [`~SchedulerMixin.from_pretrained`] functions.

    Args:
        num_train_timesteps (`int`, defaults to 1000):
            The number of diffusion steps to train the model.
        beta_start (`float`, defaults to 0.0001):
            The starting `beta` value of inference.
        beta_end (`float`, defaults to 0.02):
            The final `beta` value.
        beta_schedule (`str`, defaults to `"linear"`):
            The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
            `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
        trained_betas (`np.ndarray`, *optional*):
            Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
        original_inference_steps (`int`, *optional*, defaults to 50):
            The default number of inference steps used to generate a linearly-spaced timestep schedule, from which we
            will ultimately take `num_inference_steps` evenly spaced timesteps to form the final timestep schedule.
        clip_sample (`bool`, defaults to `True`):
            Clip the predicted sample for numerical stability.
        clip_sample_range (`float`, defaults to 1.0):
            The maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
        set_alpha_to_one (`bool`, defaults to `True`):
            Each diffusion step uses the alphas product value at that step and at the previous one. For the final step
            there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
            otherwise it uses the alpha value at step 0.
        steps_offset (`int`, defaults to 0):
            An offset added to the inference steps. You can use a combination of `offset=1` and
            `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
            Diffusion.
        prediction_type (`str`, defaults to `epsilon`, *optional*):
            Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
            `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
            Video](https://imagen.research.google/video/paper.pdf) paper).
        thresholding (`bool`, defaults to `False`):
            Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
            as Stable Diffusion.
        dynamic_thresholding_ratio (`float`, defaults to 0.995):
            The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
        sample_max_value (`float`, defaults to 1.0):
            The threshold value for dynamic thresholding. Valid only when `thresholding=True`.
        timestep_spacing (`str`, defaults to `"leading"`):
            The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
            Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
        timestep_scaling (`float`, defaults to 10.0):
            The factor the timesteps will be multiplied by when calculating the consistency model boundary conditions
            `c_skip` and `c_out`. Increasing this will decrease the approximation error (although the approximation
            error at the default of `10.0` is already pretty small).
        rescale_betas_zero_snr (`bool`, defaults to `False`):
            Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
            dark samples instead of limiting it to samples with medium brightness. Loosely related to
            [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
    """

    order = 1

    @register_to_config
    def __init__(
        self,
        num_train_timesteps: int = 1000,
        beta_start: float = 0.00085,
        beta_end: float = 0.012,
        beta_schedule: str = "scaled_linear",
        trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
        original_inference_steps: int = 50,
        clip_sample: bool = False,
        clip_sample_range: float = 1.0,
        set_alpha_to_one: bool = True,
        steps_offset: int = 0,
        prediction_type: str = "epsilon",
        thresholding: bool = False,
        dynamic_thresholding_ratio: float = 0.995,
        sample_max_value: float = 1.0,
        timestep_spacing: str = "leading",
        timestep_scaling: float = 10.0,
        rescale_betas_zero_snr: bool = False,
    ):
        if trained_betas is not None:
            self.betas = torch.tensor(trained_betas, dtype=torch.float32)
        elif beta_schedule == "linear":
            self.betas = torch.linspace(
                beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
        elif beta_schedule == "scaled_linear":
            # this schedule is very specific to the latent diffusion model.
            self.betas = torch.linspace(
                beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
        elif beta_schedule == "squaredcos_cap_v2":
            # Glide cosine schedule
            self.betas = betas_for_alpha_bar(num_train_timesteps)
        else:
            raise NotImplementedError(
                f"{beta_schedule} does is not implemented for {self.__class__}")

        # Rescale for zero SNR
        if rescale_betas_zero_snr:
            self.betas = rescale_zero_terminal_snr(self.betas)

        self.alphas = 1.0 - self.betas
        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)

        # At every step in ddim, we are looking into the previous alphas_cumprod
        # For the final step, there is no previous alphas_cumprod because we are already at 0
        # `set_alpha_to_one` decides whether we set this parameter simply to one or
        # whether we use the final alpha of the "non-previous" one.
        self.final_alpha_cumprod = torch.tensor(
            1.0) if set_alpha_to_one else self.alphas_cumprod[0]

        # standard deviation of the initial noise distribution
        self.init_noise_sigma = 1.0

        # setable values
        self.num_inference_steps = None
        self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[
                                          ::-1].copy().astype(np.int64))
        self.custom_timesteps = False

        self._step_index = None

    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
    def _init_step_index(self, timestep):
        if isinstance(timestep, torch.Tensor):
            timestep = timestep.to(self.timesteps.device)

        index_candidates = (self.timesteps == timestep).nonzero()

        # The sigma index that is taken for the **very** first `step`
        # is always the second index (or the last index if there is only 1)
        # This way we can ensure we don't accidentally skip a sigma in
        # case we start in the middle of the denoising schedule (e.g. for image-to-image)
        if len(index_candidates) > 1:
            step_index = index_candidates[1]
        else:
            step_index = index_candidates[0]

        self._step_index = step_index.item()

    @property
    def step_index(self):
        return self._step_index

    def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
        """
        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
        current timestep.

        Args:
            sample (`torch.FloatTensor`):
                The input sample.
            timestep (`int`, *optional*):
                The current timestep in the diffusion chain.
        Returns:
            `torch.FloatTensor`:
                A scaled input sample.
        """
        return sample

    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
    def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
        """
        "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
        prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
        s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
        pixels from saturation at each step. We find that dynamic thresholding results in significantly better
        photorealism as well as better image-text alignment, especially when using very large guidance weights."

        https://arxiv.org/abs/2205.11487
        """
        dtype = sample.dtype
        batch_size, channels, *remaining_dims = sample.shape

        if dtype not in (torch.float32, torch.float64):
            # upcast for quantile calculation, and clamp not implemented for cpu half
            sample = sample.float()

        # Flatten sample for doing quantile calculation along each image
        sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))

        abs_sample = sample.abs()  # "a certain percentile absolute pixel value"

        s = torch.quantile(
            abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
        s = torch.clamp(
            s, min=1, max=self.config.sample_max_value
        )  # When clamped to min=1, equivalent to standard clipping to [-1, 1]
        # (batch_size, 1) because clamp will broadcast along dim=0
        s = s.unsqueeze(1)
        # "we threshold xt0 to the range [-s, s] and then divide by s"
        sample = torch.clamp(sample, -s, s) / s

        sample = sample.reshape(batch_size, channels, *remaining_dims)
        sample = sample.to(dtype)

        return sample

    def set_timesteps(
        self,
        num_inference_steps: Optional[int] = None,
        device: Union[str, torch.device] = None,
        original_inference_steps: Optional[int] = None,
        timesteps: Optional[List[int]] = None,
        strength: int = 1.0,
    ):
        """
        Sets the discrete timesteps used for the diffusion chain (to be run before inference).

        Args:
            num_inference_steps (`int`, *optional*):
                The number of diffusion steps used when generating samples with a pre-trained model. If used,
                `timesteps` must be `None`.
            device (`str` or `torch.device`, *optional*):
                The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
            original_inference_steps (`int`, *optional*):
                The original number of inference steps, which will be used to generate a linearly-spaced timestep
                schedule (which is different from the standard `diffusers` implementation). We will then take
                `num_inference_steps` timesteps from this schedule, evenly spaced in terms of indices, and use that as
                our final timestep schedule. If not set, this will default to the `original_inference_steps` attribute.
            timesteps (`List[int]`, *optional*):
                Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
                timestep spacing strategy of equal spacing between timesteps on the training/distillation timestep
                schedule is used. If `timesteps` is passed, `num_inference_steps` must be `None`.
        """
        # 0. Check inputs
        if num_inference_steps is None and timesteps is None:
            raise ValueError(
                "Must pass exactly one of `num_inference_steps` or `custom_timesteps`.")

        if num_inference_steps is not None and timesteps is not None:
            raise ValueError(
                "Can only pass one of `num_inference_steps` or `custom_timesteps`.")

        # 1. Calculate the LCM original training/distillation timestep schedule.
        original_steps = (
            original_inference_steps if original_inference_steps is not None else self.config.original_inference_steps
        )

        if original_steps > self.config.num_train_timesteps:
            raise ValueError(
                f"`original_steps`: {original_steps} cannot be larger than `self.config.train_timesteps`:"
                f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
                f" maximal {self.config.num_train_timesteps} timesteps."
            )

        # LCM Timesteps Setting
        # The skipping step parameter k from the paper.
        k = self.config.num_train_timesteps // original_steps
        # LCM Training/Distillation Steps Schedule
        # Currently, only a linearly-spaced schedule is supported (same as in the LCM distillation scripts).
        lcm_origin_timesteps = np.asarray(
            list(range(1, int(original_steps * strength) + 1))) * k - 1

        # 2. Calculate the LCM inference timestep schedule.
        if timesteps is not None:
            # 2.1 Handle custom timestep schedules.
            train_timesteps = set(lcm_origin_timesteps)
            non_train_timesteps = []
            for i in range(1, len(timesteps)):
                if timesteps[i] >= timesteps[i - 1]:
                    raise ValueError(
                        "`custom_timesteps` must be in descending order.")

                if timesteps[i] not in train_timesteps:
                    non_train_timesteps.append(timesteps[i])

            if timesteps[0] >= self.config.num_train_timesteps:
                raise ValueError(
                    f"`timesteps` must start before `self.config.train_timesteps`:"
                    f" {self.config.num_train_timesteps}."
                )

            # Raise warning if timestep schedule does not start with self.config.num_train_timesteps - 1
            if strength == 1.0 and timesteps[0] != self.config.num_train_timesteps - 1:
                logger.warning(
                    f"The first timestep on the custom timestep schedule is {timesteps[0]}, not"
                    f" `self.config.num_train_timesteps - 1`: {self.config.num_train_timesteps - 1}. You may get"
                    f" unexpected results when using this timestep schedule."
                )

            # Raise warning if custom timestep schedule contains timesteps not on original timestep schedule
            if non_train_timesteps:
                logger.warning(
                    f"The custom timestep schedule contains the following timesteps which are not on the original"
                    f" training/distillation timestep schedule: {non_train_timesteps}. You may get unexpected results"
                    f" when using this timestep schedule."
                )

            # Raise warning if custom timestep schedule is longer than original_steps
            if len(timesteps) > original_steps:
                logger.warning(
                    f"The number of timesteps in the custom timestep schedule is {len(timesteps)}, which exceeds the"
                    f" the length of the timestep schedule used for training: {original_steps}. You may get some"
                    f" unexpected results when using this timestep schedule."
                )

            timesteps = np.array(timesteps, dtype=np.int64)
            self.num_inference_steps = len(timesteps)
            self.custom_timesteps = True

            # Apply strength (e.g. for img2img pipelines) (see StableDiffusionImg2ImgPipeline.get_timesteps)
            init_timestep = min(
                int(self.num_inference_steps * strength), self.num_inference_steps)
            t_start = max(self.num_inference_steps - init_timestep, 0)
            timesteps = timesteps[t_start * self.order:]
            # TODO: also reset self.num_inference_steps?
        else:
            # 2.2 Create the "standard" LCM inference timestep schedule.
            if num_inference_steps > self.config.num_train_timesteps:
                raise ValueError(
                    f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
                    f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
                    f" maximal {self.config.num_train_timesteps} timesteps."
                )

            skipping_step = len(lcm_origin_timesteps) // num_inference_steps

            if skipping_step < 1:
                raise ValueError(
                    f"The combination of `original_steps x strength`: {original_steps} x {strength} is smaller than `num_inference_steps`: {num_inference_steps}. Make sure to either reduce `num_inference_steps` to a value smaller than {int(original_steps * strength)} or increase `strength` to a value higher than {float(num_inference_steps / original_steps)}."
                )

            self.num_inference_steps = num_inference_steps

            if num_inference_steps > original_steps:
                raise ValueError(
                    f"`num_inference_steps`: {num_inference_steps} cannot be larger than `original_inference_steps`:"
                    f" {original_steps} because the final timestep schedule will be a subset of the"
                    f" `original_inference_steps`-sized initial timestep schedule."
                )

            # LCM Inference Steps Schedule
            lcm_origin_timesteps = lcm_origin_timesteps[::-1].copy()
            # Select (approximately) evenly spaced indices from lcm_origin_timesteps.
            inference_indices = np.linspace(
                0, len(lcm_origin_timesteps), num=num_inference_steps, endpoint=False)
            '''
            
            当只有1步时会进行999步直接进行
            两步: 999, 499, 
            四步: 999, 759, 499, 259
            
            '''
            inference_indices = np.floor(inference_indices).astype(np.int64)
            timesteps = lcm_origin_timesteps[inference_indices]

        self.timesteps = torch.from_numpy(timesteps).to(
            device=device, dtype=torch.long)

        self._step_index = None


    def get_scalings_for_boundary_condition_discrete(self, timestep):
        self.sigma_data = 0.5  # Default: 0.5
        scaled_timestep = timestep * self.config.timestep_scaling

        c_skip = self.sigma_data**2 / (scaled_timestep**2 + self.sigma_data**2)
        c_out = scaled_timestep / \
            (scaled_timestep**2 + self.sigma_data**2) ** 0.5
        return c_skip, c_out

    def step(
        self,
        model_output: torch.FloatTensor,
        timestep: int,
        sample: torch.FloatTensor,
        generator: Optional[torch.Generator] = None,
        return_dict: bool = True,
        use_ddim: bool = False,
    ) -> Union[LCMSchedulerOutput, Tuple]:
        """
        Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
        process from the learned model outputs (most often the predicted noise).

        Args:
            model_output (`torch.FloatTensor`):
                The direct output from learned diffusion model.
            timestep (`float`):
                The current discrete timestep in the diffusion chain.
            sample (`torch.FloatTensor`):
                A current instance of a sample created by the diffusion process.
            generator (`torch.Generator`, *optional*):
                A random number generator.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] or `tuple`.
        Returns:
            [`~schedulers.scheduling_utils.LCMSchedulerOutput`] or `tuple`:
                If return_dict is `True`, [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] is returned, otherwise a
                tuple is returned where the first element is the sample tensor.
        """
        if self.num_inference_steps is None:
            raise ValueError(
                "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
            )

        if self.step_index is None:
            self._init_step_index(timestep)

        # 1. get previous step value
        prev_step_index = self.step_index + 1
        if prev_step_index < len(self.timesteps):
            prev_timestep = self.timesteps[prev_step_index]
        else:
            prev_timestep = timestep

        # 2. compute alphas, betas
        alpha_prod_t = self.alphas_cumprod[timestep]
        alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod

        beta_prod_t = 1 - alpha_prod_t
        beta_prod_t_prev = 1 - alpha_prod_t_prev

        # 3. Get scalings for boundary conditions
        c_skip, c_out = self.get_scalings_for_boundary_condition_discrete(
            timestep)

        # 4. Compute the predicted original sample x_0 based on the model parameterization
        if self.config.prediction_type == "epsilon":  # noise-prediction
            predicted_original_sample = (
                sample - beta_prod_t.sqrt() * model_output) / alpha_prod_t.sqrt()
        elif self.config.prediction_type == "sample":  # x-prediction
            predicted_original_sample = model_output
        elif self.config.prediction_type == "v_prediction":  # v-prediction
            predicted_original_sample = alpha_prod_t.sqrt(
            ) * sample - beta_prod_t.sqrt() * model_output
        else:
            raise ValueError(
                f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or"
                " `v_prediction` for `LCMScheduler`."
            )

        # 5. Clip or threshold "predicted x_0"
        if self.config.thresholding:
            predicted_original_sample = self._threshold_sample(
                predicted_original_sample)
        elif self.config.clip_sample:
            predicted_original_sample = predicted_original_sample.clamp(
                -self.config.clip_sample_range, self.config.clip_sample_range
            )

        # 6. Denoise model output using boundary conditions
        denoised = c_out * predicted_original_sample + c_skip * sample
        # denoised = predicted_original_sample

        # 7. Sample and inject noise z ~ N(0, I) for MultiStep Inference
        # Noise is not used on the final timestep of the timestep schedule.
        # This also means that noise is not used for one-step sampling.
        if self.step_index != self.num_inference_steps - 1:
            if not use_ddim:
                noise = randn_tensor(
                    model_output.shape, generator=generator, device=model_output.device, dtype=denoised.dtype
                )
            prev_sample = alpha_prod_t_prev.sqrt() * denoised + beta_prod_t_prev.sqrt() * noise
        else:
            prev_sample = denoised

        # upon completion increase step index by one
        self._step_index += 1

        if not return_dict:
            return (prev_sample, denoised)

        return LCMSchedulerOutput(prev_sample=prev_sample, denoised=denoised)

    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
    def add_noise(
        self,
        original_samples: torch.FloatTensor,
        noise: torch.FloatTensor,
        timesteps: torch.IntTensor,
    ) -> torch.FloatTensor:
        # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
        alphas_cumprod = self.alphas_cumprod.to(
            device=original_samples.device, dtype=original_samples.dtype)
        timesteps = timesteps.to(original_samples.device)

        sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
        sqrt_alpha_prod = sqrt_alpha_prod.flatten()
        while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
            sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)

        sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
        sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
        while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
            sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)

        noisy_samples = sqrt_alpha_prod * original_samples + \
            sqrt_one_minus_alpha_prod * noise
        return noisy_samples

    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity
    def get_velocity(
        self, sample: torch.FloatTensor, noise: torch.FloatTensor, timesteps: torch.IntTensor
    ) -> torch.FloatTensor:
        # Make sure alphas_cumprod and timestep have same device and dtype as sample
        alphas_cumprod = self.alphas_cumprod.to(
            device=sample.device, dtype=sample.dtype)
        timesteps = timesteps.to(sample.device)

        sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
        sqrt_alpha_prod = sqrt_alpha_prod.flatten()
        while len(sqrt_alpha_prod.shape) < len(sample.shape):
            sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)

        sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
        sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
        while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
            sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)

        velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
        return velocity

    def __len__(self):
        return self.config.num_train_timesteps

    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.previous_timestep
    def previous_timestep(self, timestep):
        if self.custom_timesteps:
            index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0]
            if index == self.timesteps.shape[0] - 1:
                prev_t = torch.tensor(-1)
            else:
                prev_t = self.timesteps[index + 1]
        else:
            num_inference_steps = (
                self.num_inference_steps if self.num_inference_steps else self.config.num_train_timesteps
            )
            prev_t = timestep - self.config.num_train_timesteps // num_inference_steps

        return prev_t