scheduling_dpmsolver_multistep_inverse.py 50 KB
Newer Older
Aryan's avatar
Aryan committed
1
# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved.
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
#
# 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 file is strongly influenced by https://github.com/LuChengTHU/dpm-solver

import math
18
from typing import List, Literal, Optional, Tuple, Union
19
20
21
22
23

import numpy as np
import torch

from ..configuration_utils import ConfigMixin, register_to_config
24
from ..utils import deprecate, is_scipy_available
Dhruv Nair's avatar
Dhruv Nair committed
25
from ..utils.torch_utils import randn_tensor
26
27
28
from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput


29
30
31
32
if is_scipy_available():
    import scipy.stats


33
# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
YiYi Xu's avatar
YiYi Xu committed
34
def betas_for_alpha_bar(
35
36
37
38
    num_diffusion_timesteps: int,
    max_beta: float = 0.999,
    alpha_transform_type: Literal["cosine", "exp"] = "cosine",
) -> torch.Tensor:
39
40
41
42
43
44
45
46
    """
    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:
47
48
49
50
51
52
        num_diffusion_timesteps (`int`):
            The number of betas to produce.
        max_beta (`float`, defaults to `0.999`):
            The maximum beta to use; use values lower than 1 to avoid numerical instability.
        alpha_transform_type (`"cosine"` or `"exp"`, defaults to `"cosine"`):
            The type of noise schedule for `alpha_bar`. Choose from `cosine` or `exp`.
53
54

    Returns:
55
56
        `torch.Tensor`:
            The betas used by the scheduler to step the model outputs.
57
    """
YiYi Xu's avatar
YiYi Xu committed
58
    if alpha_transform_type == "cosine":
59

YiYi Xu's avatar
YiYi Xu committed
60
61
62
63
64
65
66
67
68
        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:
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
69
        raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}")
70
71
72
73
74

    betas = []
    for i in range(num_diffusion_timesteps):
        t1 = i / num_diffusion_timesteps
        t2 = (i + 1) / num_diffusion_timesteps
YiYi Xu's avatar
YiYi Xu committed
75
        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
76
77
78
79
80
    return torch.tensor(betas, dtype=torch.float32)


class DPMSolverMultistepInverseScheduler(SchedulerMixin, ConfigMixin):
    """
81
    `DPMSolverMultistepInverseScheduler` is the reverse scheduler of [`DPMSolverMultistepScheduler`].
82

83
84
    This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
    methods the library implements for all schedulers such as loading and saving.
85
86

    Args:
87
88
89
90
91
92
93
94
        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
95
            `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
96
97
98
99
        trained_betas (`np.ndarray`, *optional*):
            Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
        solver_order (`int`, defaults to 2):
            The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided
100
            sampling, and `solver_order=3` for unconditional sampling.
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
        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` and
            `algorithm_type="dpmsolver++"`.
        algorithm_type (`str`, defaults to `dpmsolver++`):
            Algorithm type for the solver; can be `dpmsolver`, `dpmsolver++`, `sde-dpmsolver` or `sde-dpmsolver++`. The
            `dpmsolver` type implements the algorithms in the [DPMSolver](https://huggingface.co/papers/2206.00927)
            paper, and the `dpmsolver++` type implements the algorithms in the
            [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to use `dpmsolver++` or
            `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion.
        solver_type (`str`, defaults to `midpoint`):
            Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the
            sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers.
        lower_order_final (`bool`, defaults to `True`):
            Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can
            stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10.
125
126
127
128
        euler_at_final (`bool`, defaults to `False`):
            Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail
            richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference
            steps, but sometimes may result in blurring.
129
        use_karras_sigmas (`bool`, *optional*, defaults to `False`):
130
131
            Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
            the sigmas are determined according to a sequence of noise levels {σi}.
132
133
        use_exponential_sigmas (`bool`, *optional*, defaults to `False`):
            Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process.
134
135
136
        use_beta_sigmas (`bool`, *optional*, defaults to `False`):
            Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta
            Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information.
137
138
139
        lambda_min_clipped (`float`, defaults to `-inf`):
            Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the
            cosine (`squaredcos_cap_v2`) noise schedule.
140
        variance_type (`str`, *optional*):
141
142
143
144
145
146
            Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output
            contains the predicted Gaussian variance.
        timestep_spacing (`str`, defaults to `"linspace"`):
            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.
        steps_offset (`int`, defaults to 0):
147
            An offset added to the inference steps, as required by some model families.
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
    """

    _compatibles = [e.name for e in KarrasDiffusionSchedulers]
    order = 1

    @register_to_config
    def __init__(
        self,
        num_train_timesteps: int = 1000,
        beta_start: float = 0.0001,
        beta_end: float = 0.02,
        beta_schedule: str = "linear",
        trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
        solver_order: int = 2,
        prediction_type: str = "epsilon",
        thresholding: bool = False,
        dynamic_thresholding_ratio: float = 0.995,
        sample_max_value: float = 1.0,
        algorithm_type: str = "dpmsolver++",
        solver_type: str = "midpoint",
        lower_order_final: bool = True,
169
        euler_at_final: bool = False,
170
        use_karras_sigmas: Optional[bool] = False,
171
        use_exponential_sigmas: Optional[bool] = False,
172
        use_beta_sigmas: Optional[bool] = False,
173
174
        use_flow_sigmas: Optional[bool] = False,
        flow_shift: Optional[float] = 1.0,
175
176
        lambda_min_clipped: float = -float("inf"),
        variance_type: Optional[str] = None,
177
178
        timestep_spacing: str = "linspace",
        steps_offset: int = 0,
179
    ):
180
181
182
183
184
185
        if self.config.use_beta_sigmas and not is_scipy_available():
            raise ImportError("Make sure to install scipy if you want to use beta sigmas.")
        if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1:
            raise ValueError(
                "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used."
            )
186
187
188
189
        if algorithm_type in ["dpmsolver", "sde-dpmsolver"]:
            deprecation_message = f"algorithm_type {algorithm_type} is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead"
            deprecate("algorithm_types dpmsolver and sde-dpmsolver", "1.0.0", deprecation_message)

190
191
192
193
194
195
        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.
196
            self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
197
198
199
200
        elif beta_schedule == "squaredcos_cap_v2":
            # Glide cosine schedule
            self.betas = betas_for_alpha_bar(num_train_timesteps)
        else:
201
            raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}")
202
203
204
205
206
207
208

        self.alphas = 1.0 - self.betas
        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
        # Currently we only support VP-type noise schedule
        self.alpha_t = torch.sqrt(self.alphas_cumprod)
        self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
        self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)
209
        self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
210
211
212
213
214
215
216
217
218

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

        # settings for DPM-Solver
        if algorithm_type not in ["dpmsolver", "dpmsolver++", "sde-dpmsolver", "sde-dpmsolver++"]:
            if algorithm_type == "deis":
                self.register_to_config(algorithm_type="dpmsolver++")
            else:
219
                raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}")
220
221
222
223
224

        if solver_type not in ["midpoint", "heun"]:
            if solver_type in ["logrho", "bh1", "bh2"]:
                self.register_to_config(solver_type="midpoint")
            else:
225
                raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}")
226
227
228
229
230
231
232

        # setable values
        self.num_inference_steps = None
        timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32).copy()
        self.timesteps = torch.from_numpy(timesteps)
        self.model_outputs = [None] * solver_order
        self.lower_order_nums = 0
233
        self._step_index = None
234
        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
235
        self.use_karras_sigmas = use_karras_sigmas
236
        self.use_exponential_sigmas = use_exponential_sigmas
237
        self.use_beta_sigmas = use_beta_sigmas
238

239
240
241
    @property
    def step_index(self):
        """
M. Tolga Cangöz's avatar
M. Tolga Cangöz committed
242
        The index counter for current timestep. It will increase 1 after each scheduler step.
243
244
245
        """
        return self._step_index

246
247
    def set_timesteps(self, num_inference_steps: int = None, device: Union[str, torch.device] = None):
        """
248
        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
249
250
251

        Args:
            num_inference_steps (`int`):
252
253
254
                The number of diffusion steps used when generating samples with a pre-trained model.
            device (`str` or `torch.device`, *optional*):
                The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
255
256
257
        """
        # Clipping the minimum of all lambda(t) for numerical stability.
        # This is critical for cosine (squaredcos_cap_v2) noise schedule.
258
        clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped).item()
259
260
        self.noisiest_timestep = self.config.num_train_timesteps - 1 - clipped_idx

Quentin Gallouédec's avatar
Quentin Gallouédec committed
261
        # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
        if self.config.timestep_spacing == "linspace":
            timesteps = (
                np.linspace(0, self.noisiest_timestep, num_inference_steps + 1).round()[:-1].copy().astype(np.int64)
            )
        elif self.config.timestep_spacing == "leading":
            step_ratio = (self.noisiest_timestep + 1) // (num_inference_steps + 1)
            # creates integer timesteps by multiplying by ratio
            # casting to int to avoid issues when num_inference_step is power of 3
            timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[:-1].copy().astype(np.int64)
            timesteps += self.config.steps_offset
        elif self.config.timestep_spacing == "trailing":
            step_ratio = self.config.num_train_timesteps / num_inference_steps
            # creates integer timesteps by multiplying by ratio
            # casting to int to avoid issues when num_inference_step is power of 3
            timesteps = np.arange(self.noisiest_timestep + 1, 0, -step_ratio).round()[::-1].copy().astype(np.int64)
            timesteps -= 1
        else:
            raise ValueError(
                f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', "
                "'leading' or 'trailing'."
            )

        sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
285
286
        log_sigmas = np.log(sigmas)

287
        if self.config.use_karras_sigmas:
288
289
290
            sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
            timesteps = timesteps.copy().astype(np.int64)
291
            sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
292
        elif self.config.use_exponential_sigmas:
293
            sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
294
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
295
            sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
296
        elif self.config.use_beta_sigmas:
297
            sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
298
            timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
299
            sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
300
301
302
        elif self.config.use_flow_sigmas:
            alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)
            sigmas = 1.0 - alphas
hlky's avatar
hlky committed
303
            sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy()
304
            timesteps = (sigmas * self.config.num_train_timesteps).copy()
305
            sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
306
307
308
309
310
311
        else:
            sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
            sigma_max = (
                (1 - self.alphas_cumprod[self.noisiest_timestep]) / self.alphas_cumprod[self.noisiest_timestep]
            ) ** 0.5
            sigmas = np.concatenate([sigmas, [sigma_max]]).astype(np.float32)
312

313
314
        self.sigmas = torch.from_numpy(sigmas)

315
316
317
318
319
        # when num_inference_steps == num_train_timesteps, we can end up with
        # duplicates in timesteps.
        _, unique_indices = np.unique(timesteps, return_index=True)
        timesteps = timesteps[np.sort(unique_indices)]

320
        self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64)
321
322
323
324
325
326
327
328

        self.num_inference_steps = len(timesteps)

        self.model_outputs = [
            None,
        ] * self.config.solver_order
        self.lower_order_nums = 0

329
330
        # add an index counter for schedulers that allow duplicated timesteps
        self._step_index = None
331
        self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
332

333
    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
334
    def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor:
335
        """
336
337
        Apply dynamic thresholding to the predicted sample.

338
339
340
341
342
343
        "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."

Quentin Gallouédec's avatar
Quentin Gallouédec committed
344
        https://huggingface.co/papers/2205.11487
345
346
347
348
349
350
351
352

        Args:
            sample (`torch.Tensor`):
                The predicted sample to be thresholded.

        Returns:
            `torch.Tensor`:
                The thresholded sample.
353
354
        """
        dtype = sample.dtype
355
        batch_size, channels, *remaining_dims = sample.shape
356
357
358
359
360

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

        # Flatten sample for doing quantile calculation along each image
361
        sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
362
363
364
365
366
367
368
369
370
371

        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]
        s = s.unsqueeze(1)  # (batch_size, 1) because clamp will broadcast along dim=0
        sample = torch.clamp(sample, -s, s) / s  # "we threshold xt0 to the range [-s, s] and then divide by s"

372
        sample = sample.reshape(batch_size, channels, *remaining_dims)
373
374
375
376
377
378
        sample = sample.to(dtype)

        return sample

    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
    def _sigma_to_t(self, sigma, log_sigmas):
379
380
381
382
383
384
385
386
387
388
389
390
391
        """
        Convert sigma values to corresponding timestep values through interpolation.

        Args:
            sigma (`np.ndarray`):
                The sigma value(s) to convert to timestep(s).
            log_sigmas (`np.ndarray`):
                The logarithm of the sigma schedule used for interpolation.

        Returns:
            `np.ndarray`:
                The interpolated timestep value(s) corresponding to the input sigma(s).
        """
392
        # get log sigma
393
        log_sigma = np.log(np.maximum(sigma, 1e-10))
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413

        # get distribution
        dists = log_sigma - log_sigmas[:, np.newaxis]

        # get sigmas range
        low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
        high_idx = low_idx + 1

        low = log_sigmas[low_idx]
        high = log_sigmas[high_idx]

        # interpolate sigmas
        w = (low - log_sigma) / (low - high)
        w = np.clip(w, 0, 1)

        # transform interpolation to time range
        t = (1 - w) * low_idx + w * high_idx
        t = t.reshape(sigma.shape)
        return t

414
415
    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
    def _sigma_to_alpha_sigma_t(self, sigma):
416
417
418
419
420
421
        if self.config.use_flow_sigmas:
            alpha_t = 1 - sigma
            sigma_t = sigma
        else:
            alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
            sigma_t = sigma * alpha_t
422
423
424

        return alpha_t, sigma_t

425
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
426
    def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor:
427
428
429
430
431
432
433
434
435
436
437
438
439
440
        """
        Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative
        Models](https://huggingface.co/papers/2206.00364).

        Args:
            in_sigmas (`torch.Tensor`):
                The input sigma values to be converted.
            num_inference_steps (`int`):
                The number of inference steps to generate the noise schedule for.

        Returns:
            `torch.Tensor`:
                The converted sigma values following the Karras noise schedule.
        """
441

Suraj Patil's avatar
Suraj Patil committed
442
443
444
445
446
447
448
449
450
451
452
453
454
455
        # Hack to make sure that other schedulers which copy this function don't break
        # TODO: Add this logic to the other schedulers
        if hasattr(self.config, "sigma_min"):
            sigma_min = self.config.sigma_min
        else:
            sigma_min = None

        if hasattr(self.config, "sigma_max"):
            sigma_max = self.config.sigma_max
        else:
            sigma_max = None

        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
456
457
458
459
460
461
462
463

        rho = 7.0  # 7.0 is the value used in the paper
        ramp = np.linspace(0, 1, num_inference_steps)
        min_inv_rho = sigma_min ** (1 / rho)
        max_inv_rho = sigma_max ** (1 / rho)
        sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
        return sigmas

464
465
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential
    def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor:
466
467
468
469
470
471
472
473
474
475
476
477
478
        """
        Construct an exponential noise schedule.

        Args:
            in_sigmas (`torch.Tensor`):
                The input sigma values to be converted.
            num_inference_steps (`int`):
                The number of inference steps to generate the noise schedule for.

        Returns:
            `torch.Tensor`:
                The converted sigma values following an exponential schedule.
        """
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494

        # Hack to make sure that other schedulers which copy this function don't break
        # TODO: Add this logic to the other schedulers
        if hasattr(self.config, "sigma_min"):
            sigma_min = self.config.sigma_min
        else:
            sigma_min = None

        if hasattr(self.config, "sigma_max"):
            sigma_max = self.config.sigma_max
        else:
            sigma_max = None

        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()

495
        sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps))
496
497
        return sigmas

498
499
500
501
    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta
    def _convert_to_beta(
        self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6
    ) -> torch.Tensor:
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
        """
        Construct a beta noise schedule as proposed in [Beta Sampling is All You
        Need](https://huggingface.co/papers/2407.12173).

        Args:
            in_sigmas (`torch.Tensor`):
                The input sigma values to be converted.
            num_inference_steps (`int`):
                The number of inference steps to generate the noise schedule for.
            alpha (`float`, *optional*, defaults to `0.6`):
                The alpha parameter for the beta distribution.
            beta (`float`, *optional*, defaults to `0.6`):
                The beta parameter for the beta distribution.

        Returns:
            `torch.Tensor`:
                The converted sigma values following a beta distribution schedule.
        """
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535

        # Hack to make sure that other schedulers which copy this function don't break
        # TODO: Add this logic to the other schedulers
        if hasattr(self.config, "sigma_min"):
            sigma_min = self.config.sigma_min
        else:
            sigma_min = None

        if hasattr(self.config, "sigma_max"):
            sigma_max = self.config.sigma_max
        else:
            sigma_max = None

        sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
        sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()

536
        sigmas = np.array(
537
538
539
540
541
542
543
544
545
546
            [
                sigma_min + (ppf * (sigma_max - sigma_min))
                for ppf in [
                    scipy.stats.beta.ppf(timestep, alpha, beta)
                    for timestep in 1 - np.linspace(0, 1, num_inference_steps)
                ]
            ]
        )
        return sigmas

547
548
    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.convert_model_output
    def convert_model_output(
549
        self,
550
        model_output: torch.Tensor,
551
        *args,
552
        sample: torch.Tensor = None,
553
        **kwargs,
554
    ) -> torch.Tensor:
555
        """
556
557
558
559
        Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is
        designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an
        integral of the data prediction model.

Steven Liu's avatar
Steven Liu committed
560
561
        > [!TIP] > The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both
        noise > prediction and data prediction models.
562
563

        Args:
564
            model_output (`torch.Tensor`):
565
                The direct output from the learned diffusion model.
566
            sample (`torch.Tensor`):
567
                A current instance of a sample created by the diffusion process.
568
569

        Returns:
570
            `torch.Tensor`:
571
                The converted model output.
572
        """
573
574
575
576
577
        timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
        if sample is None:
            if len(args) > 1:
                sample = args[1]
            else:
578
                raise ValueError("missing `sample` as a required keyword argument")
579
580
581
582
583
584
        if timestep is not None:
            deprecate(
                "timesteps",
                "1.0.0",
                "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )
585
586
587
588
589
590
591

        # DPM-Solver++ needs to solve an integral of the data prediction model.
        if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]:
            if self.config.prediction_type == "epsilon":
                # DPM-Solver and DPM-Solver++ only need the "mean" output.
                if self.config.variance_type in ["learned", "learned_range"]:
                    model_output = model_output[:, :3]
592
593
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
594
595
596
597
                x0_pred = (sample - sigma_t * model_output) / alpha_t
            elif self.config.prediction_type == "sample":
                x0_pred = model_output
            elif self.config.prediction_type == "v_prediction":
598
599
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
600
                x0_pred = alpha_t * sample - sigma_t * model_output
601
602
603
            elif self.config.prediction_type == "flow_prediction":
                sigma_t = self.sigmas[self.step_index]
                x0_pred = sample - sigma_t * model_output
604
605
            else:
                raise ValueError(
606
607
                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, "
                    "`v_prediction`, or `flow_prediction` for the DPMSolverMultistepScheduler."
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
                )

            if self.config.thresholding:
                x0_pred = self._threshold_sample(x0_pred)

            return x0_pred

        # DPM-Solver needs to solve an integral of the noise prediction model.
        elif self.config.algorithm_type in ["dpmsolver", "sde-dpmsolver"]:
            if self.config.prediction_type == "epsilon":
                # DPM-Solver and DPM-Solver++ only need the "mean" output.
                if self.config.variance_type in ["learned", "learned_range"]:
                    epsilon = model_output[:, :3]
                else:
                    epsilon = model_output
            elif self.config.prediction_type == "sample":
624
625
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
626
627
                epsilon = (sample - alpha_t * model_output) / sigma_t
            elif self.config.prediction_type == "v_prediction":
628
629
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
630
631
632
633
634
635
636
637
                epsilon = alpha_t * model_output + sigma_t * sample
            else:
                raise ValueError(
                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
                    " `v_prediction` for the DPMSolverMultistepScheduler."
                )

            if self.config.thresholding:
638
639
                sigma = self.sigmas[self.step_index]
                alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
640
641
642
643
644
645
                x0_pred = (sample - sigma_t * epsilon) / alpha_t
                x0_pred = self._threshold_sample(x0_pred)
                epsilon = (sample - alpha_t * x0_pred) / sigma_t

            return epsilon

646
    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.dpm_solver_first_order_update
647
648
    def dpm_solver_first_order_update(
        self,
649
        model_output: torch.Tensor,
650
        *args,
651
652
        sample: torch.Tensor = None,
        noise: Optional[torch.Tensor] = None,
653
        **kwargs,
654
    ) -> torch.Tensor:
655
        """
656
        One step for the first-order DPMSolver (equivalent to DDIM).
657
658

        Args:
659
            model_output (`torch.Tensor`):
660
                The direct output from the learned diffusion model.
661
            sample (`torch.Tensor`):
662
                A current instance of a sample created by the diffusion process.
663
664

        Returns:
665
            `torch.Tensor`:
666
                The sample tensor at the previous timestep.
667
        """
668
669
670
671
672
673
        timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
674
                raise ValueError("missing `sample` as a required keyword argument")
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
        if timestep is not None:
            deprecate(
                "timesteps",
                "1.0.0",
                "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index]
        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
        alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s)
        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
        lambda_s = torch.log(alpha_s) - torch.log(sigma_s)

695
696
697
698
699
        h = lambda_t - lambda_s
        if self.config.algorithm_type == "dpmsolver++":
            x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output
        elif self.config.algorithm_type == "dpmsolver":
            x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output
700
701
702
703
704
705
706
707
708
709
710
711
712
        elif self.config.algorithm_type == "sde-dpmsolver++":
            assert noise is not None
            x_t = (
                (sigma_t / sigma_s * torch.exp(-h)) * sample
                + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output
                + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
            )
        elif self.config.algorithm_type == "sde-dpmsolver":
            assert noise is not None
            x_t = (
                (alpha_t / alpha_s) * sample
                - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * model_output
                + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
713
714
715
            )
        return x_t

716
    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.multistep_dpm_solver_second_order_update
717
718
    def multistep_dpm_solver_second_order_update(
        self,
719
        model_output_list: List[torch.Tensor],
720
        *args,
721
722
        sample: torch.Tensor = None,
        noise: Optional[torch.Tensor] = None,
723
        **kwargs,
724
    ) -> torch.Tensor:
725
        """
726
        One step for the second-order multistep DPMSolver.
727
728

        Args:
729
            model_output_list (`List[torch.Tensor]`):
730
                The direct outputs from learned diffusion model at current and latter timesteps.
731
            sample (`torch.Tensor`):
732
                A current instance of a sample created by the diffusion process.
733
734

        Returns:
735
            `torch.Tensor`:
736
                The sample tensor at the previous timestep.
737
        """
738
739
740
741
742
743
        timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
744
                raise ValueError("missing `sample` as a required keyword argument")
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
        if timestep_list is not None:
            deprecate(
                "timestep_list",
                "1.0.0",
                "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        sigma_t, sigma_s0, sigma_s1 = (
            self.sigmas[self.step_index + 1],
            self.sigmas[self.step_index],
            self.sigmas[self.step_index - 1],
        )

        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
        alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
        alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)

        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
        lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
        lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)

773
        m0, m1 = model_output_list[-1], model_output_list[-2]
774

775
776
777
778
        h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1
        r0 = h_0 / h
        D0, D1 = m0, (1.0 / r0) * (m0 - m1)
        if self.config.algorithm_type == "dpmsolver++":
Quentin Gallouédec's avatar
Quentin Gallouédec committed
779
            # See https://huggingface.co/papers/2211.01095 for detailed derivations
780
781
782
783
784
785
786
787
788
789
790
791
792
            if self.config.solver_type == "midpoint":
                x_t = (
                    (sigma_t / sigma_s0) * sample
                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                    - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (sigma_t / sigma_s0) * sample
                    - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                    + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
                )
        elif self.config.algorithm_type == "dpmsolver":
Quentin Gallouédec's avatar
Quentin Gallouédec committed
793
            # See https://huggingface.co/papers/2206.00927 for detailed derivations
794
795
796
797
798
799
800
801
802
803
804
805
            if self.config.solver_type == "midpoint":
                x_t = (
                    (alpha_t / alpha_s0) * sample
                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (alpha_t / alpha_s0) * sample
                    - (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
                )
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
        elif self.config.algorithm_type == "sde-dpmsolver++":
            assert noise is not None
            if self.config.solver_type == "midpoint":
                x_t = (
                    (sigma_t / sigma_s0 * torch.exp(-h)) * sample
                    + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
                    + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1
                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (sigma_t / sigma_s0 * torch.exp(-h)) * sample
                    + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
                    + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1
                    + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
                )
        elif self.config.algorithm_type == "sde-dpmsolver":
            assert noise is not None
            if self.config.solver_type == "midpoint":
                x_t = (
                    (alpha_t / alpha_s0) * sample
                    - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - (sigma_t * (torch.exp(h) - 1.0)) * D1
                    + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
                )
            elif self.config.solver_type == "heun":
                x_t = (
                    (alpha_t / alpha_s0) * sample
                    - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0
                    - 2.0 * (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
                    + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
                )
838
839
840
841
842
        return x_t

    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.multistep_dpm_solver_third_order_update
    def multistep_dpm_solver_third_order_update(
        self,
843
        model_output_list: List[torch.Tensor],
844
        *args,
845
        sample: torch.Tensor = None,
StAlKeR7779's avatar
StAlKeR7779 committed
846
        noise: Optional[torch.Tensor] = None,
847
        **kwargs,
848
    ) -> torch.Tensor:
849
        """
850
        One step for the third-order multistep DPMSolver.
851
852

        Args:
853
            model_output_list (`List[torch.Tensor]`):
854
                The direct outputs from learned diffusion model at current and latter timesteps.
855
            sample (`torch.Tensor`):
856
                A current instance of a sample created by diffusion process.
857
858

        Returns:
859
            `torch.Tensor`:
860
                The sample tensor at the previous timestep.
861
        """
862
863
864
865
866
867
868

        timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
        prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
        if sample is None:
            if len(args) > 2:
                sample = args[2]
            else:
869
                raise ValueError("missing `sample` as a required keyword argument")
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
        if timestep_list is not None:
            deprecate(
                "timestep_list",
                "1.0.0",
                "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        if prev_timestep is not None:
            deprecate(
                "prev_timestep",
                "1.0.0",
                "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
            )

        sigma_t, sigma_s0, sigma_s1, sigma_s2 = (
            self.sigmas[self.step_index + 1],
            self.sigmas[self.step_index],
            self.sigmas[self.step_index - 1],
            self.sigmas[self.step_index - 2],
889
        )
890
891
892
893
894
895
896
897
898
899
900
901
902

        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
        alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
        alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
        alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2)

        lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
        lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
        lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
        lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2)

        m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3]

903
904
905
906
907
908
909
        h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2
        r0, r1 = h_0 / h, h_1 / h
        D0 = m0
        D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2)
        D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1)
        D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1)
        if self.config.algorithm_type == "dpmsolver++":
Quentin Gallouédec's avatar
Quentin Gallouédec committed
910
            # See https://huggingface.co/papers/2206.00927 for detailed derivations
911
912
913
914
915
916
917
            x_t = (
                (sigma_t / sigma_s0) * sample
                - (alpha_t * (torch.exp(-h) - 1.0)) * D0
                + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
                - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2
            )
        elif self.config.algorithm_type == "dpmsolver":
Quentin Gallouédec's avatar
Quentin Gallouédec committed
918
            # See https://huggingface.co/papers/2206.00927 for detailed derivations
919
920
921
922
923
924
            x_t = (
                (alpha_t / alpha_s0) * sample
                - (sigma_t * (torch.exp(h) - 1.0)) * D0
                - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
                - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2
            )
StAlKeR7779's avatar
StAlKeR7779 committed
925
926
927
928
929
930
931
932
933
        elif self.config.algorithm_type == "sde-dpmsolver++":
            assert noise is not None
            x_t = (
                (sigma_t / sigma_s0 * torch.exp(-h)) * sample
                + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0
                + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1
                + (alpha_t * ((1.0 - torch.exp(-2.0 * h) - 2.0 * h) / (2.0 * h) ** 2 - 0.5)) * D2
                + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
            )
934
935
        return x_t

936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
    def _init_step_index(self, timestep):
        if isinstance(timestep, torch.Tensor):
            timestep = timestep.to(self.timesteps.device)

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

        if len(index_candidates) == 0:
            step_index = len(self.timesteps) - 1
        # 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)
        elif len(index_candidates) > 1:
            step_index = index_candidates[1].item()
        else:
            step_index = index_candidates[0].item()

        self._step_index = step_index

955
956
    def step(
        self,
957
        model_output: torch.Tensor,
958
        timestep: Union[int, torch.Tensor],
959
        sample: torch.Tensor,
960
        generator=None,
961
        variance_noise: Optional[torch.Tensor] = None,
962
963
964
        return_dict: bool = True,
    ) -> Union[SchedulerOutput, Tuple]:
        """
965
966
        Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
        the multistep DPMSolver.
967
968

        Args:
969
            model_output (`torch.Tensor`):
970
971
972
                The direct output from learned diffusion model.
            timestep (`int`):
                The current discrete timestep in the diffusion chain.
973
            sample (`torch.Tensor`):
974
                A current instance of a sample created by the diffusion process.
975
976
            generator (`torch.Generator`, *optional*):
                A random number generator.
977
            variance_noise (`torch.Tensor`):
978
979
                Alternative to generating noise with `generator` by directly providing the noise for the variance
                itself. Useful for methods such as [`CycleDiffusion`].
980
981
            return_dict (`bool`):
                Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
982
983

        Returns:
984
985
986
            [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
                If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
                tuple is returned where the first element is the sample tensor.
987
988
989
990
991
992
993

        """
        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"
            )

994
995
996
        if self.step_index is None:
            self._init_step_index(timestep)

997
998
999
        # Improve numerical stability for small number of steps
        lower_order_final = (self.step_index == len(self.timesteps) - 1) and (
            self.config.euler_at_final or (self.config.lower_order_final and len(self.timesteps) < 15)
1000
1001
        )
        lower_order_second = (
1002
            (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15
1003
1004
        )

1005
        model_output = self.convert_model_output(model_output, sample=sample)
1006
1007
1008
1009
        for i in range(self.config.solver_order - 1):
            self.model_outputs[i] = self.model_outputs[i + 1]
        self.model_outputs[-1] = model_output

1010
        if self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"] and variance_noise is None:
1011
1012
1013
            noise = randn_tensor(
                model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype
            )
1014
1015
        elif self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"]:
            noise = variance_noise
1016
1017
1018
1019
        else:
            noise = None

        if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final:
1020
            prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise)
1021
        elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second:
1022
            prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise)
1023
        else:
1024
            prev_sample = self.multistep_dpm_solver_third_order_update(self.model_outputs, sample=sample)
1025
1026
1027
1028

        if self.lower_order_nums < self.config.solver_order:
            self.lower_order_nums += 1

1029
1030
1031
        # upon completion increase step index by one
        self._step_index += 1

1032
1033
1034
1035
1036
1037
        if not return_dict:
            return (prev_sample,)

        return SchedulerOutput(prev_sample=prev_sample)

    # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.scale_model_input
1038
    def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor:
1039
1040
1041
1042
1043
        """
        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
        current timestep.

        Args:
1044
            sample (`torch.Tensor`):
1045
                The input sample.
1046
1047

        Returns:
1048
            `torch.Tensor`:
1049
                A scaled input sample.
1050
1051
1052
1053
1054
        """
        return sample

    def add_noise(
        self,
1055
1056
        original_samples: torch.Tensor,
        noise: torch.Tensor,
1057
        timesteps: torch.IntTensor,
1058
    ) -> torch.Tensor:
1059
1060
1061
1062
1063
1064
1065
1066
1067
        # Make sure sigmas and timesteps have the same device and dtype as original_samples
        sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
        if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
            # mps does not support float64
            schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
            timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
        else:
            schedule_timesteps = self.timesteps.to(original_samples.device)
            timesteps = timesteps.to(original_samples.device)
1068

1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
        step_indices = []
        for timestep in timesteps:
            index_candidates = (schedule_timesteps == timestep).nonzero()
            if len(index_candidates) == 0:
                step_index = len(schedule_timesteps) - 1
            elif len(index_candidates) > 1:
                step_index = index_candidates[1].item()
            else:
                step_index = index_candidates[0].item()
            step_indices.append(step_index)
1079

1080
1081
1082
        sigma = sigmas[step_indices].flatten()
        while len(sigma.shape) < len(original_samples.shape):
            sigma = sigma.unsqueeze(-1)
1083

1084
1085
        alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
        noisy_samples = alpha_t * original_samples + sigma_t * noise
1086
1087
1088
1089
        return noisy_samples

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