import os from typing import Any, Dict, List, Optional, Tuple import torch from accelerate import init_empty_weights from diffusers import ( AutoencoderKLHunyuanVideo, FlowMatchEulerDiscreteScheduler, HunyuanVideoPipeline, HunyuanVideoTransformer3DModel, ) from diffusers.models.autoencoders.vae import DiagonalGaussianDistribution from transformers import AutoTokenizer, CLIPTextModel, CLIPTokenizer, LlamaModel from ... import data from ... import functional as FF from ...logging import get_logger from ...processors import CLIPPooledProcessor, LlamaProcessor, ProcessorMixin from ...typing import ArtifactType, SchedulerType from ...utils import get_non_null_items from ..modeling_utils import ModelSpecification logger = get_logger() class HunyuanLatentEncodeProcessor(ProcessorMixin): r""" Processor to encode image/video into latents using the HunyuanVideo VAE. Args: output_names (`List[str]`): The names of the outputs that the processor returns. The outputs are in the following order: - latents: The latents of the input image/video. """ def __init__(self, output_names: List[str]): super().__init__() self.output_names = output_names assert len(self.output_names) == 1 def forward( self, vae: AutoencoderKLHunyuanVideo, image: Optional[torch.Tensor] = None, video: Optional[torch.Tensor] = None, generator: Optional[torch.Generator] = None, compute_posterior: bool = True, ) -> Dict[str, torch.Tensor]: device = vae.device dtype = vae.dtype if image is not None: video = image.unsqueeze(1) assert video.ndim == 5, f"Expected 5D tensor, got {video.ndim}D tensor" video = video.to(device=device, dtype=vae.dtype) video = video.permute(0, 2, 1, 3, 4).contiguous() # [B, F, C, H, W] -> [B, C, F, H, W] if compute_posterior: latents = vae.encode(video).latent_dist.sample(generator=generator) latents = latents.to(dtype=dtype) else: if vae.use_slicing and video.shape[0] > 1: encoded_slices = [vae._encode(x_slice) for x_slice in video.split(1)] moments = torch.cat(encoded_slices) else: moments = vae._encode(video) latents = moments.to(dtype=dtype) return {self.output_names[0]: latents} class HunyuanVideoModelSpecification(ModelSpecification): def __init__( self, pretrained_model_name_or_path: str = "hunyuanvideo-community/HunyuanVideo", tokenizer_id: Optional[str] = None, text_encoder_id: Optional[str] = None, transformer_id: Optional[str] = None, vae_id: Optional[str] = None, text_encoder_dtype: torch.dtype = torch.bfloat16, transformer_dtype: torch.dtype = torch.bfloat16, vae_dtype: torch.dtype = torch.bfloat16, revision: Optional[str] = None, cache_dir: Optional[str] = None, condition_model_processors: List[ProcessorMixin] = None, latent_model_processors: List[ProcessorMixin] = None, **kwargs, ) -> None: super().__init__( pretrained_model_name_or_path=pretrained_model_name_or_path, tokenizer_id=tokenizer_id, text_encoder_id=text_encoder_id, transformer_id=transformer_id, vae_id=vae_id, text_encoder_dtype=text_encoder_dtype, transformer_dtype=transformer_dtype, vae_dtype=vae_dtype, revision=revision, cache_dir=cache_dir, ) if condition_model_processors is None: condition_model_processors = [ LlamaProcessor(["encoder_hidden_states", "encoder_attention_mask"]), CLIPPooledProcessor( ["pooled_projections"], input_names={"tokenizer_2": "tokenizer", "text_encoder_2": "text_encoder"}, ), ] if latent_model_processors is None: latent_model_processors = [HunyuanLatentEncodeProcessor(["latents"])] self.condition_model_processors = condition_model_processors self.latent_model_processors = latent_model_processors @property def _resolution_dim_keys(self): return {"latents": (2, 3, 4)} def load_condition_models(self) -> Dict[str, torch.nn.Module]: if self.tokenizer_id is not None: tokenizer = AutoTokenizer.from_pretrained( self.tokenizer_id, revision=self.revision, cache_dir=self.cache_dir ) else: tokenizer = AutoTokenizer.from_pretrained( self.pretrained_model_name_or_path, subfolder="tokenizer", revision=self.revision, cache_dir=self.cache_dir, ) if self.tokenizer_2_id is not None: tokenizer_2 = CLIPTokenizer.from_pretrained( self.tokenizer_2_id, revision=self.revision, cache_dir=self.cache_dir ) else: tokenizer_2 = CLIPTokenizer.from_pretrained( self.pretrained_model_name_or_path, subfolder="tokenizer_2", revision=self.revision, cache_dir=self.cache_dir, ) if self.text_encoder_id is not None: text_encoder = LlamaModel.from_pretrained( self.text_encoder_id, torch_dtype=self.text_encoder_dtype, revision=self.revision, cache_dir=self.cache_dir, ) else: text_encoder = LlamaModel.from_pretrained( self.pretrained_model_name_or_path, subfolder="text_encoder", torch_dtype=self.text_encoder_dtype, revision=self.revision, cache_dir=self.cache_dir, ) if self.text_encoder_2_id is not None: text_encoder_2 = CLIPTextModel.from_pretrained( self.text_encoder_2_id, torch_dtype=self.text_encoder_2_dtype, revision=self.revision, cache_dir=self.cache_dir, ) else: text_encoder_2 = CLIPTextModel.from_pretrained( self.pretrained_model_name_or_path, subfolder="text_encoder_2", torch_dtype=self.text_encoder_2_dtype, revision=self.revision, cache_dir=self.cache_dir, ) return { "tokenizer": tokenizer, "tokenizer_2": tokenizer_2, "text_encoder": text_encoder, "text_encoder_2": text_encoder_2, } def load_latent_models(self) -> Dict[str, torch.nn.Module]: if self.vae_id is not None: vae = AutoencoderKLHunyuanVideo.from_pretrained( self.vae_id, torch_dtype=self.vae_dtype, revision=self.revision, cache_dir=self.cache_dir, ) else: vae = AutoencoderKLHunyuanVideo.from_pretrained( self.pretrained_model_name_or_path, subfolder="vae", torch_dtype=self.vae_dtype, revision=self.revision, cache_dir=self.cache_dir, ) return {"vae": vae} def load_diffusion_models(self) -> Dict[str, torch.nn.Module]: if self.transformer_id is not None: transformer = HunyuanVideoTransformer3DModel.from_pretrained( self.transformer_id, torch_dtype=self.transformer_dtype, revision=self.revision, cache_dir=self.cache_dir, ) else: transformer = HunyuanVideoTransformer3DModel.from_pretrained( self.pretrained_model_name_or_path, subfolder="transformer", torch_dtype=self.transformer_dtype, revision=self.revision, cache_dir=self.cache_dir, ) scheduler = FlowMatchEulerDiscreteScheduler() return {"transformer": transformer, "scheduler": scheduler} def load_pipeline( self, tokenizer: Optional[AutoTokenizer] = None, tokenizer_2: Optional[CLIPTokenizer] = None, text_encoder: Optional[LlamaModel] = None, text_encoder_2: Optional[CLIPTextModel] = None, transformer: Optional[HunyuanVideoTransformer3DModel] = None, vae: Optional[AutoencoderKLHunyuanVideo] = None, scheduler: Optional[FlowMatchEulerDiscreteScheduler] = None, enable_slicing: bool = False, enable_tiling: bool = False, enable_model_cpu_offload: bool = False, training: bool = False, **kwargs, ) -> HunyuanVideoPipeline: components = { "tokenizer": tokenizer, "tokenizer_2": tokenizer_2, "text_encoder": text_encoder, "text_encoder_2": text_encoder_2, "transformer": transformer, "vae": vae, "scheduler": scheduler, } components = get_non_null_items(components) pipe = HunyuanVideoPipeline.from_pretrained( self.pretrained_model_name_or_path, **components, revision=self.revision, cache_dir=self.cache_dir ) pipe.text_encoder.to(self.text_encoder_dtype) pipe.text_encoder_2.to(self.text_encoder_2_dtype) pipe.vae.to(self.vae_dtype) if not training: pipe.transformer.to(self.transformer_dtype) if enable_slicing: pipe.vae.enable_slicing() if enable_tiling: pipe.vae.enable_tiling() if enable_model_cpu_offload: pipe.enable_model_cpu_offload() return pipe @torch.no_grad() def prepare_conditions( self, tokenizer: AutoTokenizer, tokenizer_2: CLIPTokenizer, text_encoder: LlamaModel, text_encoder_2: CLIPTextModel, caption: str, max_sequence_length: int = 256, **kwargs, ) -> Dict[str, Any]: conditions = { "tokenizer": tokenizer, "tokenizer_2": tokenizer_2, "text_encoder": text_encoder, "text_encoder_2": text_encoder_2, "caption": caption, "max_sequence_length": max_sequence_length, **kwargs, } input_keys = set(conditions.keys()) conditions = super().prepare_conditions(**conditions) conditions = {k: v for k, v in conditions.items() if k not in input_keys} return conditions @torch.no_grad() def prepare_latents( self, vae: AutoencoderKLHunyuanVideo, image: Optional[torch.Tensor] = None, video: Optional[torch.Tensor] = None, generator: Optional[torch.Generator] = None, compute_posterior: bool = True, **kwargs, ) -> Dict[str, torch.Tensor]: conditions = { "vae": vae, "image": image, "video": video, "generator": generator, "compute_posterior": compute_posterior, **kwargs, } input_keys = set(conditions.keys()) conditions = super().prepare_latents(**conditions) conditions = {k: v for k, v in conditions.items() if k not in input_keys} return conditions def forward( self, transformer: HunyuanVideoTransformer3DModel, condition_model_conditions: Dict[str, torch.Tensor], latent_model_conditions: Dict[str, torch.Tensor], sigmas: torch.Tensor, guidance: float = 1.0, generator: Optional[torch.Generator] = None, compute_posterior: bool = True, **kwargs, ) -> Tuple[torch.Tensor, ...]: if compute_posterior: latents = latent_model_conditions.pop("latents") else: posterior = DiagonalGaussianDistribution(latent_model_conditions.pop("latents")) latents = posterior.sample(generator=generator) del posterior latents = latents * self.vae_config.scaling_factor noise = torch.zeros_like(latents).normal_(generator=generator) noisy_latents = FF.flow_match_xt(latents, noise, sigmas) timesteps = (sigmas.flatten() * 1000.0).long() guidance = latents.new_full((latents.size(0),), fill_value=guidance) * 1000.0 latent_model_conditions["hidden_states"] = noisy_latents.to(latents) latent_model_conditions["guidance"] = guidance pred = transformer( **latent_model_conditions, **condition_model_conditions, timestep=timesteps, return_dict=False, )[0] target = FF.flow_match_target(noise, latents) return pred, target, sigmas def validation( self, pipeline: HunyuanVideoPipeline, prompt: str, height: Optional[int] = None, width: Optional[int] = None, num_frames: Optional[int] = None, num_inference_steps: int = 50, generator: Optional[torch.Generator] = None, **kwargs, ) -> List[ArtifactType]: generation_kwargs = { "prompt": prompt, "height": height, "width": width, "num_frames": num_frames, "num_inference_steps": num_inference_steps, "generator": generator, "return_dict": True, "output_type": "pil", } generation_kwargs = get_non_null_items(generation_kwargs) video = pipeline(**generation_kwargs).frames[0] return [data.VideoArtifact(value=video)] def _save_lora_weights( self, directory: str, transformer_state_dict: Optional[Dict[str, torch.Tensor]] = None, scheduler: Optional[SchedulerType] = None, *args, **kwargs, ) -> None: # TODO(aryan): this needs refactoring if transformer_state_dict is not None: HunyuanVideoPipeline.save_lora_weights(directory, transformer_state_dict, safe_serialization=True) if scheduler is not None: scheduler.save_pretrained(os.path.join(directory, "scheduler")) def _save_model( self, directory: str, transformer: HunyuanVideoTransformer3DModel, transformer_state_dict: Optional[Dict[str, torch.Tensor]] = None, scheduler: Optional[SchedulerType] = None, ) -> None: # TODO(aryan): this needs refactoring if transformer_state_dict is not None: with init_empty_weights(): transformer_copy = HunyuanVideoTransformer3DModel.from_config(transformer.config) transformer_copy.load_state_dict(transformer_state_dict, strict=True, assign=True) transformer_copy.save_pretrained(os.path.join(directory, "transformer")) if scheduler is not None: scheduler.save_pretrained(os.path.join(directory, "scheduler"))