videosys/models/cogvideo/pipeline.py

# Adapted from CogVideo

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# --------------------------------------------------------
# References:
# CogVideo: https://github.com/THUDM/CogVideo
# diffusers: https://github.com/huggingface/diffusers
#  --------------------------------------------------------

import inspect
import math
from typing import Callable, Dict, List, Optional, Tuple, Union

import torch
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
from diffusers.utils import logging
from diffusers.utils.torch_utils import randn_tensor
from diffusers.video_processor import VideoProcessor
from transformers import T5EncoderModel, T5Tokenizer

from videosys.core.pipeline import VideoSysPipeline, VideoSysPipelineOutput
from videosys.utils.utils import save_video

from .autoencoder_kl import AutoencoderKLCogVideoX
from .cogvideox_transformer_3d import CogVideoXTransformer3DModel
from .retrieve_timesteps import retrieve_timesteps
from .scheduling import CogVideoXDDIMScheduler, CogVideoXDPMScheduler

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

from videosys.core.pab_mgr import (
    PABConfig,
    get_diffusion_skip,
    get_diffusion_skip_timestep,
    set_pab_manager,
    skip_diffusion_timestep,
    update_steps,
)



class CogVideoPABConfig(PABConfig):
    def __init__(
        self,
        steps: int = 150,
        spatial_broadcast: bool = True,
        spatial_threshold: list = [100, 850],
        spatial_gap: int = 2,
        temporal_broadcast: bool = True,
        temporal_threshold: list = [100, 850],
        temporal_gap: int = 4,
        cross_broadcast: bool = True,
        cross_threshold: list = [100, 850],
        cross_gap: int = 6,
        diffusion_skip: bool = False,
        diffusion_timestep_respacing: list = None,
        diffusion_skip_timestep: list = None,
        mlp_skip: bool = True,
        mlp_spatial_skip_config: dict = {
            738: {"block": [0, 1, 2, 3, 4, 5, 6], "skip_count": 2},
            714: {"block": [0, 1, 2, 3, 4, 5, 6], "skip_count": 2},
        },
        mlp_temporal_skip_config: dict = {
            738: {"block": [0, 1, 2, 3, 4, 5, 6], "skip_count": 2},
            714: {"block": [0, 1, 2, 3, 4, 5, 6], "skip_count": 2},
        },
        full_broadcast: bool = True,
        full_threshold: list = [100, 850],
        full_gap: int = 3,
    ):
        super().__init__(
            steps=steps,
            spatial_broadcast=spatial_broadcast,
            spatial_threshold=spatial_threshold,
            spatial_gap=spatial_gap,
            temporal_broadcast=temporal_broadcast,
            temporal_threshold=temporal_threshold,
            temporal_gap=temporal_gap,
            cross_broadcast=cross_broadcast,
            cross_threshold=cross_threshold,
            cross_gap=cross_gap,
            diffusion_skip=diffusion_skip,
            diffusion_timestep_respacing=diffusion_timestep_respacing,
            diffusion_skip_timestep=diffusion_skip_timestep,
            mlp_skip=mlp_skip,
            mlp_spatial_skip_config=mlp_spatial_skip_config,
            mlp_temporal_skip_config=mlp_temporal_skip_config,
            full_broadcast=full_broadcast,
            full_threshold=full_threshold,
            full_gap=full_gap,
        )



class CogVideoConfig:
    def __init__(
        self,
        world_size: int = 1,
        model_path: str = "THUDM/CogVideoX-2b",
        num_inference_steps: int = 50,
        guidance_scale: float = 6.0,
        enable_pab: bool = False,
        pab_config = CogVideoPABConfig() 
    ):
        # ======= engine ========
        self.world_size = world_size

        # ======= pipeline ========
        self.pipeline_cls = CogVideoPipeline

        # ======= model ========
        self.model_path = model_path
        self.num_inference_steps = num_inference_steps
        self.guidance_scale = guidance_scale
        self.enable_pab = enable_pab
        self.pab_config = pab_config


class CogVideoPipeline(VideoSysPipeline):
    _callback_tensor_inputs = [
        "latents",
        "prompt_embeds",
        "negative_prompt_embeds",
    ]

    def __init__(
        self,
        config: CogVideoConfig,
        tokenizer: Optional[T5Tokenizer] = None,
        text_encoder: Optional[T5EncoderModel] = None,
        vae: Optional[AutoencoderKLCogVideoX] = None,
        transformer: Optional[CogVideoXTransformer3DModel] = None,
        scheduler: Optional[CogVideoXDDIMScheduler] = None,
        device: torch.device = torch.device("cuda"),
        dtype: torch.dtype = torch.float16,
    ):
        super().__init__()
        self._config = config
        self._device = device
        self._dtype = dtype

        if transformer is None:
            transformer = CogVideoXTransformer3DModel.from_pretrained(
                config.model_path, subfolder="transformer", torch_dtype=self._dtype
            )
        if vae is None:
            vae = AutoencoderKLCogVideoX.from_pretrained(config.model_path, subfolder="vae", torch_dtype=self._dtype)
        if tokenizer is None:
            tokenizer = T5Tokenizer.from_pretrained(config.model_path, subfolder="tokenizer")
        if text_encoder is None:
            text_encoder = T5EncoderModel.from_pretrained(
                config.model_path, subfolder="text_encoder", torch_dtype=self._dtype
            )
        if scheduler is None:
            scheduler = CogVideoXDDIMScheduler.from_pretrained(
                config.model_path,
                subfolder="scheduler",
            )

        # set eval and device
        self.set_eval_and_device(self._device, text_encoder, vae, transformer)

        self.register_modules(
            tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
        )

        if config.enable_pab:
            set_pab_manager(config.pab_config)

            
        self.vae_scale_factor_spatial = (
            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
        )
        self.vae_scale_factor_temporal = (
            self.vae.config.temporal_compression_ratio if hasattr(self, "vae") and self.vae is not None else 4
        )
        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)

    def _get_t5_prompt_embeds(
        self,
        prompt: Union[str, List[str]] = None,
        num_videos_per_prompt: int = 1,
        max_sequence_length: int = 226,
        device: Optional[torch.device] = None,
        dtype: Optional[torch.dtype] = None,
    ):
        device = device or self._device
        dtype = dtype or self.text_encoder.dtype

        prompt = [prompt] if isinstance(prompt, str) else prompt
        batch_size = len(prompt)

        text_inputs = self.tokenizer(
            prompt,
            padding="max_length",
            max_length=max_sequence_length,
            truncation=True,
            add_special_tokens=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[:, max_sequence_length - 1 : -1])
            logger.warning(
                "The following part of your input was truncated because `max_sequence_length` is set to "
                f" {max_sequence_length} tokens: {removed_text}"
            )

        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)

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

        return prompt_embeds

    def encode_prompt(
        self,
        prompt: Union[str, List[str]],
        negative_prompt: Optional[Union[str, List[str]]] = None,
        do_classifier_free_guidance: bool = True,
        num_videos_per_prompt: int = 1,
        prompt_embeds: Optional[torch.Tensor] = None,
        negative_prompt_embeds: Optional[torch.Tensor] = None,
        max_sequence_length: int = 226,
        device: Optional[torch.device] = None,
        dtype: Optional[torch.dtype] = None,
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            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`).
            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
                Whether to use classifier free guidance or not.
            num_videos_per_prompt (`int`, *optional*, defaults to 1):
                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
            prompt_embeds (`torch.Tensor`, *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.Tensor`, *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.
            device: (`torch.device`, *optional*):
                torch device
            dtype: (`torch.dtype`, *optional*):
                torch dtype
        """
        device = device or self._device

        prompt = [prompt] if isinstance(prompt, str) else prompt
        if prompt is not None:
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        if prompt_embeds is None:
            prompt_embeds = self._get_t5_prompt_embeds(
                prompt=prompt,
                num_videos_per_prompt=num_videos_per_prompt,
                max_sequence_length=max_sequence_length,
                device=device,
                dtype=dtype,
            )

        if do_classifier_free_guidance and negative_prompt_embeds is None:
            negative_prompt = negative_prompt or ""
            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt

            if 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 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`."
                )

            negative_prompt_embeds = self._get_t5_prompt_embeds(
                prompt=negative_prompt,
                num_videos_per_prompt=num_videos_per_prompt,
                max_sequence_length=max_sequence_length,
                device=device,
                dtype=dtype,
            )

        return prompt_embeds, negative_prompt_embeds

    def prepare_latents(
        self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
    ):
        shape = (
            batch_size,
            (num_frames - 1) // self.vae_scale_factor_temporal + 1,
            num_channels_latents,
            height // self.vae_scale_factor_spatial,
            width // self.vae_scale_factor_spatial,
        )
        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )

        if latents is None:
            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        else:
            latents = latents.to(device)

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma
        torch.cuda.empty_cache()
        return latents
        
    def decode_latents(self, latents: torch.Tensor, num_seconds: int):
        print("hhhhhhhh")
        latents = latents.permute(0, 2, 1, 3, 4)  # [batch_size, num_channels, num_frames, height, width]
        latents = 1 / self.vae.config.scaling_factor * latents

        frames = []
        num_frames = latents.size(2)
        segment_size = num_frames // num_frames  # 每段处理的帧数

        for i in range(num_frames):  # 显存问题，逐帧处理
            start_frame = i * segment_size
            end_frame = start_frame + segment_size if i < num_frames-1 else num_frames

            current_latents = latents[:, :, start_frame:end_frame, :, :]
            try:
                current_frames = self.vae.decode(current_latents).sample
                frames.append(current_frames)
            except RuntimeError as e:
                logger.error(f"CUDA out of memory error: {str(e)}")
                raise e

            # 清理缓存
            torch.cuda.empty_cache()

        self.vae.clear_fake_context_parallel_cache()

        frames = torch.cat(frames, dim=2)
        return frames

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
    def prepare_extra_step_kwargs(self, generator, eta):
        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
        # and should be between [0, 1]

        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
        extra_step_kwargs = {}
        if accepts_eta:
            extra_step_kwargs["eta"] = eta

        # check if the scheduler accepts generator
        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
        if accepts_generator:
            extra_step_kwargs["generator"] = generator
        return extra_step_kwargs

    # Copied from diffusers.pipelines.latte.pipeline_latte.LattePipeline.check_inputs
    def check_inputs(
        self,
        prompt,
        height,
        width,
        negative_prompt,
        callback_on_step_end_tensor_inputs,
        prompt_embeds=None,
        negative_prompt_embeds=None,
    ):
        if height % 8 != 0 or width % 8 != 0:
            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")

        if callback_on_step_end_tensor_inputs is not None and not all(
            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
        ):
            raise ValueError(
                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
            )
        if prompt is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt is None and prompt_embeds is None:
            raise ValueError(
                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
            )
        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")

        if prompt is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )

        if negative_prompt is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )

        if prompt_embeds is not None and negative_prompt_embeds is not None:
            if prompt_embeds.shape != negative_prompt_embeds.shape:
                raise ValueError(
                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
                    f" {negative_prompt_embeds.shape}."
                )

    @property
    def guidance_scale(self):
        return self._guidance_scale

    @property
    def num_timesteps(self):
        return self._num_timesteps

    @property
    def interrupt(self):
        return self._interrupt

    @torch.no_grad()
    def generate(
        self,
        prompt: Optional[Union[str, List[str]]] = None,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        height: int = 480,
        width: int = 720,
        num_frames: int = 48,
        num_inference_steps: int = 50,
        timesteps: Optional[List[int]] = None,
        guidance_scale: float = 6,
        use_dynamic_cfg: bool = False,
        num_videos_per_prompt: 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: str = "pil",
        return_dict: bool = True,
        callback_on_step_end: Optional[
            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
        ] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        max_sequence_length: int = 226,
    ) -> Union[VideoSysPipelineOutput, Tuple]:
        """
        Function invoked when calling the pipeline for generation.

        Args:
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
                instead.
            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`).
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image. This is set to 1024 by default for the best results.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The width in pixels of the generated image. This is set to 1024 by default for the best results.
            num_frames (`int`, defaults to `48`):
                Number of frames to generate. Must be divisible by self.vae_scale_factor_temporal. Generated video will
                contain 1 extra frame because CogVideoX is conditioned with (num_seconds * fps + 1) frames where
                num_seconds is 6 and fps is 4. However, since videos can be saved at any fps, the only condition that
                needs to be satisfied is that of divisibility mentioned above.
            num_inference_steps (`int`, *optional*, defaults to 50):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            timesteps (`List[int]`, *optional*):
                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
                passed will be used. Must be in descending order.
            guidance_scale (`float`, *optional*, defaults to 7.0):
                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
                `guidance_scale` is defined as `w` of equation 2. of [Imagen
                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
                usually at the expense of lower image quality.
            num_videos_per_prompt (`int`, *optional*, defaults to 1):
                The number of videos to generate per prompt.
            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
                to make generation deterministic.
            latents (`torch.FloatTensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will ge generated by sampling using the supplied random `generator`.
            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.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
                of a plain tuple.
            callback_on_step_end (`Callable`, *optional*):
                A function that calls at the end of each denoising steps during the inference. The function is called
                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
                `callback_on_step_end_tensor_inputs`.
            callback_on_step_end_tensor_inputs (`List`, *optional*):
                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                `._callback_tensor_inputs` attribute of your pipeline class.
            max_sequence_length (`int`, defaults to `226`):
                Maximum sequence length in encoded prompt. Must be consistent with
                `self.transformer.config.max_text_seq_length` otherwise may lead to poor results.

        Examples:

        Returns:
            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipelineOutput`] or `tuple`:
            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipelineOutput`] if `return_dict` is True, otherwise a
            `tuple`. When returning a tuple, the first element is a list with the generated images.
        """
        fps = 8
        assert (
            num_frames <= 48 and num_frames % fps == 0 and fps == 8
        ), f"The number of frames must be divisible by {fps=} and less than 48 frames (for now). Other values are not supported in CogVideoX."

        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs

        height = height or self.transformer.config.sample_size * self.vae_scale_factor_spatial
        width = width or self.transformer.config.sample_size * self.vae_scale_factor_spatial
        num_videos_per_prompt = 1

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt,
            height,
            width,
            negative_prompt,
            callback_on_step_end_tensor_inputs,
            prompt_embeds,
            negative_prompt_embeds,
        )
        self._guidance_scale = guidance_scale
        self._interrupt = False

        # 2. Default 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._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
        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
            prompt,
            negative_prompt,
            do_classifier_free_guidance,
            num_videos_per_prompt=num_videos_per_prompt,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            max_sequence_length=max_sequence_length,
            device=device,
        )
        if do_classifier_free_guidance:
            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)

        # 4. Prepare timesteps
        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
        self._num_timesteps = len(timesteps)

        # 5. Prepare latents.
        latent_channels = self.transformer.config.in_channels
        num_frames += 1
        latents = self.prepare_latents(
            batch_size * num_videos_per_prompt,
            latent_channels,
            num_frames,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        # 6. Prepare extra step kwargs.
        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)

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

        with self.progress_bar(total=num_inference_steps) as progress_bar:
            # for DPM-solver++
            old_pred_original_sample = None
            for i, t in enumerate(timesteps):
                if self.interrupt:
                    continue

                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)

                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                timestep = t.expand(latent_model_input.shape[0])

                # predict noise model_output
                noise_pred = self.transformer(
                    hidden_states=latent_model_input,
                    encoder_hidden_states=prompt_embeds,
                    timestep=timestep,
                    return_dict=False,
                )[0]
                noise_pred = noise_pred.float()

                # perform guidance
                if use_dynamic_cfg:
                    self._guidance_scale = 1 + guidance_scale * (
                        (1 - math.cos(math.pi * ((num_inference_steps - t.item()) / num_inference_steps) ** 5.0)) / 2
                    )
                if do_classifier_free_guidance:
                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)

                # compute the previous noisy sample x_t -> x_t-1
                if not isinstance(self.scheduler, CogVideoXDPMScheduler):
                    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
                else:
                    latents, old_pred_original_sample = self.scheduler.step(
                        noise_pred,
                        old_pred_original_sample,
                        t,
                        timesteps[i - 1] if i > 0 else None,
                        latents,
                        **extra_step_kwargs,
                        return_dict=False,
                    )
                latents = latents.to(prompt_embeds.dtype)

                # call the callback, if provided
                if callback_on_step_end is not None:
                    callback_kwargs = {}
                    for k in callback_on_step_end_tensor_inputs:
                        callback_kwargs[k] = locals()[k]
                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

                    latents = callback_outputs.pop("latents", latents)
                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)

                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                    progress_bar.update()

        if not output_type == "latent":
            video = self.decode_latents(latents, num_frames // fps)
            video = self.video_processor.postprocess_video(video=video, output_type=output_type)
        else:
            video = latents

        if not return_dict:
            return (video,)

        return VideoSysPipelineOutput(video=video)

    def save_video(self, video, output_path):
        save_video(video, output_path, fps=8)