# Copyright 2025 Genmo 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. import inspect from typing import Any, Callable, Dict, List, Optional, Union import numpy as np import torch from transformers import T5EncoderModel, T5TokenizerFast from ...callbacks import MultiPipelineCallbacks, PipelineCallback from ...loaders import Mochi1LoraLoaderMixin from ...models import AutoencoderKLMochi, MochiTransformer3DModel from ...schedulers import FlowMatchEulerDiscreteScheduler from ...utils import ( is_torch_xla_available, logging, replace_example_docstring, ) from ...utils.torch_utils import randn_tensor from ...video_processor import VideoProcessor from ..pipeline_utils import DiffusionPipeline from .pipeline_output import MochiPipelineOutput if is_torch_xla_available(): import torch_xla.core.xla_model as xm XLA_AVAILABLE = True else: XLA_AVAILABLE = False logger = logging.get_logger(__name__) # pylint: disable=invalid-name EXAMPLE_DOC_STRING = """ Examples: ```py >>> import torch >>> from diffusers import MochiPipeline >>> from diffusers.utils import export_to_video >>> pipe = MochiPipeline.from_pretrained("genmo/mochi-1-preview", torch_dtype=torch.bfloat16) >>> pipe.enable_model_cpu_offload() >>> pipe.enable_vae_tiling() >>> prompt = "Close-up of a chameleon's eye, with its scaly skin changing color. Ultra high resolution 4k." >>> frames = pipe(prompt, num_inference_steps=28, guidance_scale=3.5).frames[0] >>> export_to_video(frames, "mochi.mp4") ``` """ # from: https://github.com/genmoai/models/blob/075b6e36db58f1242921deff83a1066887b9c9e1/src/mochi_preview/infer.py#L77 def linear_quadratic_schedule(num_steps, threshold_noise, linear_steps=None): if linear_steps is None: linear_steps = num_steps // 2 linear_sigma_schedule = [i * threshold_noise / linear_steps for i in range(linear_steps)] threshold_noise_step_diff = linear_steps - threshold_noise * num_steps quadratic_steps = num_steps - linear_steps quadratic_coef = threshold_noise_step_diff / (linear_steps * quadratic_steps**2) linear_coef = threshold_noise / linear_steps - 2 * threshold_noise_step_diff / (quadratic_steps**2) const = quadratic_coef * (linear_steps**2) quadratic_sigma_schedule = [ quadratic_coef * (i**2) + linear_coef * i + const for i in range(linear_steps, num_steps) ] sigma_schedule = linear_sigma_schedule + quadratic_sigma_schedule sigma_schedule = [1.0 - x for x in sigma_schedule] return sigma_schedule # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps def retrieve_timesteps( scheduler, num_inference_steps: Optional[int] = None, device: Optional[Union[str, torch.device]] = None, timesteps: Optional[List[int]] = None, sigmas: Optional[List[float]] = None, **kwargs, ): r""" Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. Args: scheduler (`SchedulerMixin`): The scheduler to get timesteps from. num_inference_steps (`int`): 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. timesteps (`List[int]`, *optional*): Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, `num_inference_steps` and `sigmas` must be `None`. sigmas (`List[float]`, *optional*): Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, `num_inference_steps` and `timesteps` must be `None`. Returns: `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the second element is the number of inference steps. """ if timesteps is not None and sigmas is not None: raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") if timesteps is not None: accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) if not accepts_timesteps: raise ValueError( f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" f" timestep schedules. Please check whether you are using the correct scheduler." ) scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) timesteps = scheduler.timesteps num_inference_steps = len(timesteps) elif sigmas is not None: accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) if not accept_sigmas: raise ValueError( f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" f" sigmas schedules. Please check whether you are using the correct scheduler." ) scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) timesteps = scheduler.timesteps num_inference_steps = len(timesteps) else: scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) timesteps = scheduler.timesteps return timesteps, num_inference_steps class MochiPipeline(DiffusionPipeline, Mochi1LoraLoaderMixin): r""" The mochi pipeline for text-to-video generation. Reference: https://github.com/genmoai/models Args: transformer ([`MochiTransformer3DModel`]): Conditional Transformer architecture to denoise the encoded video latents. scheduler ([`FlowMatchEulerDiscreteScheduler`]): A scheduler to be used in combination with `transformer` to denoise the encoded image latents. vae ([`AutoencoderKLMochi`]): Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. text_encoder ([`T5EncoderModel`]): [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. tokenizer (`CLIPTokenizer`): Tokenizer of class [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). tokenizer (`T5TokenizerFast`): Second Tokenizer of class [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). """ model_cpu_offload_seq = "text_encoder->transformer->vae" _optional_components = [] _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] def __init__( self, scheduler: FlowMatchEulerDiscreteScheduler, vae: AutoencoderKLMochi, text_encoder: T5EncoderModel, tokenizer: T5TokenizerFast, transformer: MochiTransformer3DModel, force_zeros_for_empty_prompt: bool = False, ): super().__init__() self.register_modules( vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, transformer=transformer, scheduler=scheduler, ) # TODO: determine these scaling factors from model parameters self.vae_spatial_scale_factor = 8 self.vae_temporal_scale_factor = 6 self.patch_size = 2 self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_scale_factor) self.tokenizer_max_length = ( self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 256 ) self.default_height = 480 self.default_width = 848 self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) def _get_t5_prompt_embeds( self, prompt: Union[str, List[str]] = None, num_videos_per_prompt: int = 1, max_sequence_length: int = 256, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None, ): device = device or self._execution_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 prompt_attention_mask = text_inputs.attention_mask prompt_attention_mask = prompt_attention_mask.bool().to(device) # The original Mochi implementation zeros out empty negative prompts # but this can lead to overflow when placing the entire pipeline under the autocast context # adding this here so that we can enable zeroing prompts if necessary if self.config.force_zeros_for_empty_prompt and (prompt == "" or prompt[-1] == ""): text_input_ids = torch.zeros_like(text_input_ids, device=device) prompt_attention_mask = torch.zeros_like(prompt_attention_mask, dtype=torch.bool, device=device) 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), attention_mask=prompt_attention_mask)[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) prompt_attention_mask = prompt_attention_mask.view(batch_size, -1) prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1) return prompt_embeds, prompt_attention_mask # Adapted from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline.encode_prompt 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, prompt_attention_mask: Optional[torch.Tensor] = None, negative_prompt_attention_mask: Optional[torch.Tensor] = None, max_sequence_length: int = 256, 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._execution_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, prompt_attention_mask = 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, negative_prompt_attention_mask = 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, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask def check_inputs( self, prompt, height, width, callback_on_step_end_tensor_inputs=None, prompt_embeds=None, negative_prompt_embeds=None, prompt_attention_mask=None, negative_prompt_attention_mask=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_embeds is not None and prompt_attention_mask is None: raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") 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}." ) if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: raise ValueError( "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" f" {negative_prompt_attention_mask.shape}." ) def enable_vae_slicing(self): r""" Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. """ self.vae.enable_slicing() def disable_vae_slicing(self): r""" Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to computing decoding in one step. """ self.vae.disable_slicing() def enable_vae_tiling(self): r""" Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow processing larger images. """ self.vae.enable_tiling() def disable_vae_tiling(self): r""" Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to computing decoding in one step. """ self.vae.disable_tiling() def prepare_latents( self, batch_size, num_channels_latents, height, width, num_frames, dtype, device, generator, latents=None, ): height = height // self.vae_spatial_scale_factor width = width // self.vae_spatial_scale_factor num_frames = (num_frames - 1) // self.vae_temporal_scale_factor + 1 shape = (batch_size, num_channels_latents, num_frames, height, width) if latents is not None: return latents.to(device=device, dtype=dtype) 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." ) latents = randn_tensor(shape, generator=generator, device=device, dtype=torch.float32) latents = latents.to(dtype) return latents @property def guidance_scale(self): return self._guidance_scale @property def do_classifier_free_guidance(self): return self._guidance_scale > 1.0 @property def num_timesteps(self): return self._num_timesteps @property def attention_kwargs(self): return self._attention_kwargs @property def current_timestep(self): return self._current_timestep @property def interrupt(self): return self._interrupt @torch.no_grad() @replace_example_docstring(EXAMPLE_DOC_STRING) def __call__( self, prompt: Union[str, List[str]] = None, negative_prompt: Optional[Union[str, List[str]]] = None, height: Optional[int] = None, width: Optional[int] = None, num_frames: int = 19, num_inference_steps: int = 64, timesteps: List[int] = None, guidance_scale: float = 4.5, num_videos_per_prompt: Optional[int] = 1, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, latents: Optional[torch.Tensor] = None, prompt_embeds: Optional[torch.Tensor] = None, prompt_attention_mask: Optional[torch.Tensor] = None, negative_prompt_embeds: Optional[torch.Tensor] = None, negative_prompt_attention_mask: Optional[torch.Tensor] = None, output_type: Optional[str] = "pil", return_dict: bool = True, attention_kwargs: Optional[Dict[str, Any]] = None, callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, callback_on_step_end_tensor_inputs: List[str] = ["latents"], max_sequence_length: int = 256, ): r""" 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. height (`int`, *optional*, defaults to `self.default_height`): The height in pixels of the generated image. This is set to 480 by default for the best results. width (`int`, *optional*, defaults to `self.default_width`): The width in pixels of the generated image. This is set to 848 by default for the best results. num_frames (`int`, defaults to `19`): The number of video frames to generate 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`, defaults to `4.5`): Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). 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.Tensor`, *optional*): Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image generation. Can be used to tweak the same generation with different prompts. If not provided, a latents tensor will ge generated by sampling using the supplied random `generator`. prompt_embeds (`torch.Tensor`, *optional*): Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input argument. prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings. negative_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. negative_prompt_attention_mask (`torch.FloatTensor`, *optional*): Pre-generated attention mask for negative text embeddings. 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.mochi.MochiPipelineOutput`] instead of a plain tuple. attention_kwargs (`dict`, *optional*): A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under `self.processor` in [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). callback_on_step_end (`Callable`, *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 `256`): Maximum sequence length to use with the `prompt`. Examples: Returns: [`~pipelines.mochi.MochiPipelineOutput`] or `tuple`: If `return_dict` is `True`, [`~pipelines.mochi.MochiPipelineOutput`] is returned, otherwise a `tuple` is returned where the first element is a list with the generated images. """ if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs height = height or self.default_height width = width or self.default_width # 1. Check inputs. Raise error if not correct self.check_inputs( prompt=prompt, height=height, width=width, callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, prompt_attention_mask=prompt_attention_mask, negative_prompt_attention_mask=negative_prompt_attention_mask, ) self._guidance_scale = guidance_scale self._attention_kwargs = attention_kwargs self._current_timestep = None self._interrupt = False # 2. Define call parameters if prompt is not None and isinstance(prompt, str): batch_size = 1 elif prompt is not None and isinstance(prompt, list): batch_size = len(prompt) else: batch_size = prompt_embeds.shape[0] device = self._execution_device # 3. Prepare text embeddings ( prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask, ) = self.encode_prompt( prompt=prompt, negative_prompt=negative_prompt, do_classifier_free_guidance=self.do_classifier_free_guidance, num_videos_per_prompt=num_videos_per_prompt, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, prompt_attention_mask=prompt_attention_mask, negative_prompt_attention_mask=negative_prompt_attention_mask, max_sequence_length=max_sequence_length, device=device, ) # 4. Prepare latent variables num_channels_latents = self.transformer.config.in_channels latents = self.prepare_latents( batch_size * num_videos_per_prompt, num_channels_latents, height, width, num_frames, prompt_embeds.dtype, device, generator, latents, ) if self.do_classifier_free_guidance: prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) # 5. Prepare timestep # from https://github.com/genmoai/models/blob/075b6e36db58f1242921deff83a1066887b9c9e1/src/mochi_preview/infer.py#L77 threshold_noise = 0.025 sigmas = linear_quadratic_schedule(num_inference_steps, threshold_noise) sigmas = np.array(sigmas) timesteps, num_inference_steps = retrieve_timesteps( self.scheduler, num_inference_steps, device, timesteps, sigmas, ) num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) self._num_timesteps = len(timesteps) # 6. Denoising loop with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): if self.interrupt: continue # Note: Mochi uses reversed timesteps. To ensure compatibility with methods like FasterCache, we need # to make sure we're using the correct non-reversed timestep values. self._current_timestep = 1000 - t latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents # broadcast to batch dimension in a way that's compatible with ONNX/Core ML timestep = t.expand(latent_model_input.shape[0]).to(latents.dtype) with self.transformer.cache_context("cond_uncond"): noise_pred = self.transformer( hidden_states=latent_model_input, encoder_hidden_states=prompt_embeds, timestep=timestep, encoder_attention_mask=prompt_attention_mask, attention_kwargs=attention_kwargs, return_dict=False, )[0] # Mochi CFG + Sampling runs in FP32 noise_pred = noise_pred.to(torch.float32) if self.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 latents_dtype = latents.dtype latents = self.scheduler.step(noise_pred, t, latents.to(torch.float32), return_dict=False)[0] latents = latents.to(latents_dtype) if latents.dtype != latents_dtype: if torch.backends.mps.is_available(): # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 latents = latents.to(latents_dtype) 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) # call the callback, if provided if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): progress_bar.update() if XLA_AVAILABLE: xm.mark_step() self._current_timestep = None if output_type == "latent": video = latents else: # unscale/denormalize the latents # denormalize with the mean and std if available and not None has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None if has_latents_mean and has_latents_std: latents_mean = ( torch.tensor(self.vae.config.latents_mean).view(1, 12, 1, 1, 1).to(latents.device, latents.dtype) ) latents_std = ( torch.tensor(self.vae.config.latents_std).view(1, 12, 1, 1, 1).to(latents.device, latents.dtype) ) latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean else: latents = latents / self.vae.config.scaling_factor video = self.vae.decode(latents, return_dict=False)[0] video = self.video_processor.postprocess_video(video, output_type=output_type) # Offload all models self.maybe_free_model_hooks() if not return_dict: return (video,) return MochiPipelineOutput(frames=video)