from typing import Any, Callable, Dict, List, Optional, Union, Tuple import torch from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler from diffusers.utils.torch_utils import randn_tensor import math import numpy as np # import logger def sde_step_with_logprob( self: UniPCMultistepScheduler, model_output: torch.FloatTensor, timestep: Union[float, torch.FloatTensor], sample: torch.FloatTensor, prev_sample: Optional[torch.FloatTensor] = None, generator: Optional[torch.Generator] = None, determistic: bool = False, return_pixel_log_prob: bool = False, return_dt_and_std_dev_t: bool = False ): """ Predict the sample from the previous timestep by reversing the SDE. This function propagates the flow process from the learned model outputs (most often the predicted velocity). Args: model_output (`torch.FloatTensor`): The direct output from learned flow model. timestep (`float`): The current discrete timestep in the diffusion chain. sample (`torch.FloatTensor`): A current instance of a sample created by the diffusion process. generator (`torch.Generator`, *optional*): A random number generator. """ # prev_sample_mean, we must convert all variable to fp32 model_output=model_output.float() sample=sample.float() if prev_sample is not None: prev_sample=prev_sample.float() step_index = [self.index_for_timestep(t) for t in timestep] prev_step_index = [step+1 for step in step_index] self.sigmas = self.sigmas.to(sample.device) sigma = self.sigmas[step_index].view(-1, 1, 1, 1, 1) sigma_prev = self.sigmas[prev_step_index].view(-1, 1, 1, 1, 1) sigma_max = self.sigmas[1].item() sigma_min = self.sigmas[-1].item() dt = sigma_prev - sigma std_dev_t = sigma_min + (sigma_max - sigma_min) * sigma prev_sample_mean = sample*(1+std_dev_t**2/(2*sigma)*dt)+model_output*(1+std_dev_t**2*(1-sigma)/(2*sigma))*dt if prev_sample is not None and generator is not None: raise ValueError( "Cannot pass both generator and prev_sample. Please make sure that either `generator` or" " `prev_sample` stays `None`." ) if prev_sample is None: variance_noise = randn_tensor( model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype, ) prev_sample = prev_sample_mean + std_dev_t * torch.sqrt(-1*dt) * variance_noise # No noise is added during evaluation if determistic: prev_sample = sample + dt * model_output log_prob = ( -((prev_sample.detach() - prev_sample_mean) ** 2) / (2 * ((std_dev_t * torch.sqrt(-1*dt))**2)) - torch.log(std_dev_t * torch.sqrt(-1*dt)) - torch.log(torch.sqrt(2 * torch.as_tensor(math.pi))) ) # mean along all but batch dimension log_prob = log_prob.mean(dim=tuple(range(1, log_prob.ndim))) if return_dt_and_std_dev_t: return prev_sample, log_prob, prev_sample_mean, std_dev_t, torch.sqrt(-1*dt) return prev_sample, log_prob, prev_sample_mean, std_dev_t * torch.sqrt(-1*dt) def wan_pipeline_with_logprob( self, prompt: Union[str, List[str]] = None, negative_prompt: Union[str, List[str]] = None, height: int = 480, width: int = 832, num_frames: int = 81, num_inference_steps: int = 50, guidance_scale: float = 5.0, 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, negative_prompt_embeds: Optional[torch.Tensor] = None, output_type: Optional[str] = "np", return_dict: bool = True, attention_kwargs: Optional[Dict[str, Any]] = None, 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 = 512, determistic: bool = False, kl_reward: float = 0.0, return_pixel_log_prob: bool = False, ): r""" The call function to 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`, defaults to `480`): The height in pixels of the generated image. width (`int`, defaults to `832`): The width in pixels of the generated image. num_frames (`int`, defaults to `81`): The number of frames in the generated video. num_inference_steps (`int`, defaults to `50`): The number of denoising steps. More denoising steps usually lead to a higher quality image at the expense of slower inference. guidance_scale (`float`, defaults to `5.0`): 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 images to generate per prompt. generator (`torch.Generator` or `List[torch.Generator]`, *optional*): A [`torch.Generator`](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 is 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 (prompt weighting). If not provided, text embeddings are generated from the `prompt` input argument. output_type (`str`, *optional*, defaults to `"pil"`): The output format of the generated image. Choose between `PIL.Image` or `np.array`. return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a [`WanPipelineOutput`] 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`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of each denoising step during the inference. 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. autocast_dtype (`torch.dtype`, *optional*, defaults to `torch.bfloat16`): The dtype to use for the torch.amp.autocast. Examples: Returns: [`~WanPipelineOutput`] or `tuple`: If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where the first element is a list with the generated images and the second element is a list of `bool`s indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. """ if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs # 1. Check inputs. Raise error if not correct self.check_inputs( prompt, negative_prompt, height, width, prompt_embeds, negative_prompt_embeds, callback_on_step_end_tensor_inputs, ) if num_frames % self.vae_scale_factor_temporal != 1: print( f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." ) num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 num_frames = max(num_frames, 1) self._guidance_scale = guidance_scale self._attention_kwargs = attention_kwargs self._current_timestep = None self._interrupt = False device = self._execution_device # 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] # 3. Encode input prompt prompt_embeds, negative_prompt_embeds = 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, max_sequence_length=max_sequence_length, device=device, ) transformer_dtype = self.transformer.dtype prompt_embeds = prompt_embeds.to(transformer_dtype) if negative_prompt_embeds is not None: negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) # 4. Prepare timesteps self.scheduler.set_timesteps(num_inference_steps, device=device) timesteps = self.scheduler.timesteps # 5. 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, torch.float32, device, generator, latents, ) all_latents = [latents] all_log_probs = [] all_kl = [] # 6. Denoising loop num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order self._num_timesteps = len(timesteps) # print(timesteps) with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): if self.interrupt: continue latents_ori = latents.clone() self._current_timestep = t latent_model_input = latents.to(transformer_dtype) timestep = t.expand(latents.shape[0]) noise_pred = self.transformer( hidden_states=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds, attention_kwargs=attention_kwargs, return_dict=False, )[0] noise_pred = noise_pred.to(prompt_embeds.dtype) if self.do_classifier_free_guidance: noise_uncond = self.transformer( hidden_states=latent_model_input, timestep=timestep, encoder_hidden_states=negative_prompt_embeds, attention_kwargs=attention_kwargs, return_dict=False, )[0] noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) latents, log_prob, prev_latents_mean, std_dev_t = sde_step_with_logprob( self.scheduler, noise_pred.float(), t.unsqueeze(0), latents.float(), determistic=determistic, return_pixel_log_prob=return_pixel_log_prob ) prev_latents = latents.clone() all_latents.append(latents) all_log_probs.append(log_prob) # compute the previous noisy sample x_t -> x_t-1 # latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] 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) # use kl_reward & is sampling process if kl_reward>0 and not determistic: latent_model_input = torch.cat([latents_ori] * 2) if self.do_classifier_free_guidance else latents_ori with self.transformer.disable_adapter(): noise_pred = self.transformer( hidden_states=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds, attention_kwargs=attention_kwargs, return_dict=False, )[0] noise_pred = noise_pred.to(prompt_embeds.dtype) # perform guidance 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) _, ref_log_prob, ref_prev_latents_mean, ref_std_dev_t = sde_step_with_logprob( self.scheduler, noise_pred.float(), t.unsqueeze(0), latents_ori.float(), prev_sample=prev_latents.float(), determistic=determistic, ) assert std_dev_t == ref_std_dev_t kl = (prev_latents_mean - ref_prev_latents_mean)**2 / (2 * std_dev_t**2) kl = kl.mean(dim=tuple(range(1, kl.ndim))) all_kl.append(kl) else: # no kl reward, we do not need to compute, just put a pre-position value, kl will be 0 all_kl.append(torch.zeros(len(latents), device=latents.device)) # 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 not output_type == "latent": latents = latents.to(self.vae.dtype) latents_mean = ( torch.tensor(self.vae.config.latents_mean) .view(1, self.vae.config.z_dim, 1, 1, 1) .to(latents.device, latents.dtype) ) latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( latents.device, latents.dtype ) latents = latents / latents_std + latents_mean video = self.vae.decode(latents, return_dict=False)[0] video = self.video_processor.postprocess_video(video, output_type=output_type) else: video = latents self.maybe_free_model_hooks() if not return_dict: return (video, all_latents, all_log_probs, all_kl) return WanPipelineOutput(frames=video), all_latents, all_log_probs, all_kl