import inspect import json import logging import os from collections.abc import Iterable import torch from diffusers.image_processor import VaeImageProcessor from diffusers.models.autoencoders import AutoencoderKL from diffusers.schedulers.scheduling_flow_match_euler_discrete import ( FlowMatchEulerDiscreteScheduler, ) from diffusers.utils.torch_utils import randn_tensor from torch import nn from transformers import CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel, T5Tokenizer from vllm.model_executor.models.utils import AutoWeightsLoader from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig from vllm_omni.diffusion.distributed.cfg_parallel import CFGParallelMixin from vllm_omni.diffusion.distributed.utils import get_local_device from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader from vllm_omni.diffusion.models.sd3.sd3_transformer import ( SD3Transformer2DModel, ) from vllm_omni.diffusion.request import OmniDiffusionRequest from vllm_omni.model_executor.model_loader.weight_utils import ( download_weights_from_hf_specific, ) logger = logging.getLogger(__name__) def get_sd3_image_post_process_func( od_config: OmniDiffusionConfig, ): if od_config.output_type == "latent": return lambda x: x model_name = od_config.model if os.path.exists(model_name): model_path = model_name else: model_path = download_weights_from_hf_specific(model_name, None, ["*"]) vae_config_path = os.path.join(model_path, "vae/config.json") with open(vae_config_path) as f: vae_config = json.load(f) vae_scale_factor = 2 ** (len(vae_config["block_out_channels"]) - 1) if "block_out_channels" in vae_config else 8 image_processor = VaeImageProcessor(vae_scale_factor=vae_scale_factor) def post_process_func( images: torch.Tensor, ): return image_processor.postprocess(images) return post_process_func def calculate_shift( image_seq_len, base_seq_len: int = 256, max_seq_len: int = 4096, base_shift: float = 0.5, max_shift: float = 1.15, ): m = (max_shift - base_shift) / (max_seq_len - base_seq_len) b = base_shift - m * base_seq_len mu = image_seq_len * m + b return mu def retrieve_timesteps( scheduler, num_inference_steps: int | None = None, device: str | torch.device | None = None, timesteps: list[int] | None = None, sigmas: list[float] | None = None, **kwargs, ) -> tuple[torch.Tensor, int]: 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 StableDiffusion3Pipeline(nn.Module, CFGParallelMixin): def __init__( self, *, od_config: OmniDiffusionConfig, prefix: str = "", ): super().__init__() self.od_config = od_config self.weights_sources = [ DiffusersPipelineLoader.ComponentSource( model_or_path=od_config.model, subfolder="transformer", revision=None, prefix="transformer.", fall_back_to_pt=True, ) ] self.device = get_local_device() model = od_config.model # Check if model is a local path local_files_only = os.path.exists(model) self.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained( model, subfolder="scheduler", local_files_only=local_files_only ) self.tokenizer = CLIPTokenizer.from_pretrained(model, subfolder="tokenizer", local_files_only=local_files_only) self.tokenizer_2 = CLIPTokenizer.from_pretrained( model, subfolder="tokenizer_2", local_files_only=local_files_only ) self.tokenizer_3 = T5Tokenizer.from_pretrained( model, subfolder="tokenizer_3", local_files_only=local_files_only ) self.text_encoder = CLIPTextModelWithProjection.from_pretrained( model, subfolder="text_encoder", local_files_only=local_files_only ) self.text_encoder_2 = CLIPTextModelWithProjection.from_pretrained( model, subfolder="text_encoder_2", local_files_only=local_files_only ) self.text_encoder_3 = T5EncoderModel.from_pretrained( model, subfolder="text_encoder_3", local_files_only=local_files_only, ) self.transformer = SD3Transformer2DModel(od_config=od_config) self.vae = AutoencoderKL.from_pretrained(model, subfolder="vae", local_files_only=local_files_only).to( self.device ) self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) self.tokenizer_max_length = ( self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 ) self.default_sample_size = 128 self.patch_size = 2 self.output_type = self.od_config.output_type def check_inputs( self, prompt, prompt_2, prompt_3, height, width, negative_prompt=None, negative_prompt_2=None, negative_prompt_3=None, prompt_embeds=None, negative_prompt_embeds=None, max_sequence_length=None, ): if ( height % (self.vae_scale_factor * self.patch_size) != 0 or width % (self.vae_scale_factor * self.patch_size) != 0 ): raise ValueError( f"`height` and `width` have to be divisible by " f"{self.vae_scale_factor * self.patch_size} but are " f"{height} and {width}. You can use height " f"{height - height % (self.vae_scale_factor * self.patch_size)} " f"and width {width - width % (self.vae_scale_factor * self.patch_size)}." ) 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_2 is not None and prompt_embeds is not None: raise ValueError( f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" " only forward one of the two." ) elif prompt_3 is not None and prompt_embeds is not None: raise ValueError( f"Cannot forward both `prompt_3`: {prompt_2} 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)}") elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)): raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}") 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." ) elif negative_prompt_2 is not None and negative_prompt_embeds is not None: raise ValueError( f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" f" {negative_prompt_embeds}. Please make sure to only forward one of the two." ) elif negative_prompt_3 is not None and negative_prompt_embeds is not None: raise ValueError( f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} 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}." ) if max_sequence_length is not None and max_sequence_length > 512: raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") def _get_clip_prompt_embeds( self, prompt: str | list[str] = "", num_images_per_prompt: int = 1, dtype: torch.dtype | None = None, clip_model_index: int = 0, ): dtype = dtype or self.text_encoder.dtype clip_tokenizers = [self.tokenizer, self.tokenizer_2] clip_text_encoders = [self.text_encoder, self.text_encoder_2] tokenizer = clip_tokenizers[clip_model_index] text_encoder = clip_text_encoders[clip_model_index] prompt = [prompt] if isinstance(prompt, str) else prompt batch_size = len(prompt) text_inputs = tokenizer( prompt, padding="max_length", max_length=self.tokenizer_max_length, truncation=True, return_tensors="pt", ) text_input_ids = text_inputs.input_ids untruncated_ids = 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 = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) logger.warning( "The following part of your input was truncated because CLIP can only handle sequences up to" f" {self.tokenizer_max_length} tokens: {removed_text}" ) prompt_embeds = text_encoder(text_input_ids.to(self.device), output_hidden_states=True) pooled_prompt_embeds = prompt_embeds[0] prompt_embeds = prompt_embeds.hidden_states[-2] prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=self.device) _, seq_len, _ = prompt_embeds.shape prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt) pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) return prompt_embeds, pooled_prompt_embeds def _get_t5_prompt_embeds( self, prompt: str | list[str] = "", num_images_per_prompt: int = 1, max_sequence_length: int = 256, dtype: torch.dtype | None = None, ): dtype = dtype or self.text_encoder_3.dtype prompt = [prompt] if isinstance(prompt, str) else prompt batch_size = len(prompt) if self.text_encoder_3 is None: return torch.zeros( ( batch_size, max_sequence_length, self.transformer.joint_attention_dim, ), device=self.device, dtype=dtype, ) text_inputs = self.tokenizer_3( prompt, padding="max_length", max_length=max_sequence_length, truncation=True, add_special_tokens=True, return_tensors="pt", ).to(self.device) text_input_ids = text_inputs.input_ids untruncated_ids = self.tokenizer_3(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_3.batch_decode(untruncated_ids[:, self.tokenizer_max_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_3(text_input_ids.to(self.device))[0] dtype = self.text_encoder_3.dtype prompt_embeds = prompt_embeds.to(dtype=dtype, device=self.device) _, seq_len, _ = prompt_embeds.shape prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) return prompt_embeds def encode_prompt( self, prompt: str | list[str], prompt_2: str | list[str], prompt_3: str | list[str], prompt_embeds: torch.Tensor | None = None, max_sequence_length: int = 256, num_images_per_prompt: int = 1, ): r""" Args: prompt (`str` or `List[str]`, *optional*): prompt to be encoded prompt_2 (`str` or `List[str]`, *optional*): The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is used in all text-encoders prompt_3 (`str` or `List[str]`, *optional*): The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is used in all text-encoders num_images_per_prompt (`int`): number of images that should be generated per prompt 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. """ prompt = [prompt] if isinstance(prompt, str) else prompt pooled_prompt_embeds = None if prompt_embeds is None: prompt_2 = prompt_2 or prompt prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 prompt_3 = prompt_3 or prompt prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3 prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds( prompt=prompt, num_images_per_prompt=num_images_per_prompt, clip_model_index=0, ) prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds( prompt=prompt_2, num_images_per_prompt=num_images_per_prompt, clip_model_index=1, ) clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1) t5_prompt_embed = self._get_t5_prompt_embeds( prompt=prompt_3, num_images_per_prompt=num_images_per_prompt, max_sequence_length=max_sequence_length, ) clip_prompt_embeds = torch.nn.functional.pad( clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1]) ) prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2) pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1) return prompt_embeds, pooled_prompt_embeds def prepare_latents( self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None, ) -> torch.Tensor: if latents is not None: return latents.to(device=device, dtype=dtype) shape = ( batch_size, num_channels_latents, int(height) // self.vae_scale_factor, int(width) // self.vae_scale_factor, ) 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=dtype) return latents def prepare_timesteps(self, num_inference_steps, sigmas, image_seq_len): scheduler_kwargs = {} if self.scheduler.config.get("use_dynamic_shifting", None): mu = calculate_shift( image_seq_len, self.scheduler.config.get("base_image_seq_len", 256), self.scheduler.config.get("max_image_seq_len", 4096), self.scheduler.config.get("base_shift", 0.5), self.scheduler.config.get("max_shift", 1.16), ) scheduler_kwargs["mu"] = mu timesteps, num_inference_steps = retrieve_timesteps( self.scheduler, num_inference_steps, sigmas=sigmas, **scheduler_kwargs, ) return timesteps, num_inference_steps @property def guidance_scale(self): return self._guidance_scale @property def num_timesteps(self): return self._num_timesteps @property def current_timestep(self): return self._current_timestep @property def interrupt(self): return self._interrupt def diffuse( self, latents: torch.Tensor, timesteps: torch.Tensor, prompt_embeds: torch.Tensor, pooled_prompt_embeds: torch.Tensor, negative_prompt_embeds: torch.Tensor, negative_pooled_prompt_embeds: torch.Tensor, do_true_cfg: bool, guidance_scale: float, cfg_normalize: bool = False, ) -> torch.Tensor: """ Diffusion loop with optional classifier-free guidance. Args: latents: Noise latents to denoise timesteps: Diffusion timesteps prompt_embeds: Positive prompt embeddings pooled_prompt_embeds: Pooled positive prompt embeddings negative_prompt_embeds: Negative prompt embeddings negative_pooled_prompt_embeds: Pooled negative prompt embeddings do_true_cfg: Whether to apply CFG guidance_scale: CFG scale factor cfg_normalize: Whether to normalize CFG output (default: False) Returns: Denoised latents """ self.scheduler.set_begin_index(0) for _, t in enumerate(timesteps): if self.interrupt: continue self._current_timestep = t # Broadcast timestep to match batch size timestep = t.expand(latents.shape[0]).to(device=latents.device, dtype=latents.dtype) positive_kwargs = { "hidden_states": latents, "timestep": timestep, "encoder_hidden_states": prompt_embeds, "pooled_projections": pooled_prompt_embeds, "return_dict": False, } if do_true_cfg: negative_kwargs = { "hidden_states": latents, "timestep": timestep, "encoder_hidden_states": negative_prompt_embeds, "pooled_projections": negative_pooled_prompt_embeds, "return_dict": False, } else: negative_kwargs = None # Predict noise with automatic CFG parallel handling noise_pred = self.predict_noise_maybe_with_cfg( do_true_cfg, guidance_scale, positive_kwargs, negative_kwargs, cfg_normalize, ) # Compute the previous noisy sample x_t -> x_t-1 with automatic CFG sync latents = self.scheduler_step_maybe_with_cfg(noise_pred, t, latents, do_true_cfg) return latents def forward( self, req: OmniDiffusionRequest, prompt: str | list[str] = "", prompt_2: str | list[str] = "", prompt_3: str | list[str] = "", negative_prompt: str | list[str] = "", negative_prompt_2: str | list[str] = "", negative_prompt_3: str | list[str] = "", height: int | None = None, width: int | None = None, num_inference_steps: int = 28, sigmas: list[float] | None = None, num_images_per_prompt: int = 1, generator: torch.Generator | list[torch.Generator] | None = None, latents: torch.Tensor | None = None, prompt_embeds: torch.Tensor | None = None, negative_prompt_embeds: torch.Tensor | None = None, pooled_prompt_embeds: torch.Tensor | None = None, negative_pooled_prompt_embeds: torch.Tensor | None = None, max_sequence_length: int = 256, ) -> DiffusionOutput: # TODO: In online mode, sometimes it receives [{"negative_prompt": None}, {...}], so cannot use .get("...", "") # TODO: May be some data formatting operations on the API side. Hack for now. prompt = [p if isinstance(p, str) else (p.get("prompt") or "") for p in req.prompts] or prompt negative_prompt = [ "" if isinstance(p, str) else (p.get("negative_prompt") or "") for p in req.prompts ] or negative_prompt height = req.sampling_params.height or self.default_sample_size * self.vae_scale_factor width = req.sampling_params.width or self.default_sample_size * self.vae_scale_factor sigmas = req.sampling_params.sigmas or sigmas max_sequence_length = req.sampling_params.max_sequence_length or max_sequence_length num_inference_steps = req.sampling_params.num_inference_steps or num_inference_steps generator = req.sampling_params.generator or generator num_images_per_prompt = ( req.sampling_params.num_outputs_per_prompt if req.sampling_params.num_outputs_per_prompt > 0 else num_images_per_prompt ) # 1. check inputs # 2. encode prompts # 3. prepare latents and timesteps # 4. diffusion process # 5. decode latents # 6. post-process outputs self.check_inputs( prompt, prompt_2, prompt_3, height, width, negative_prompt=negative_prompt, negative_prompt_2=negative_prompt_2, negative_prompt_3=negative_prompt_3, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, max_sequence_length=max_sequence_length, ) self._guidance_scale = req.sampling_params.guidance_scale self._current_timestep = None self._interrupt = False 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] prompt_embeds, pooled_prompt_embeds = self.encode_prompt( prompt=prompt, prompt_2=prompt_2, prompt_3=prompt_3, prompt_embeds=prompt_embeds, max_sequence_length=max_sequence_length, ) do_cfg = self.guidance_scale > 1 if do_cfg: negative_prompt_embeds, negative_pooled_prompt_embeds = self.encode_prompt( prompt=negative_prompt, prompt_2=negative_prompt_2, prompt_3=negative_prompt_3, prompt_embeds=negative_prompt_embeds, max_sequence_length=max_sequence_length, ) num_channels_latents = self.transformer.in_channels latents = self.prepare_latents( batch_size * num_images_per_prompt, num_channels_latents, height, width, prompt_embeds.dtype, self.device, generator, latents, ) timesteps, num_inference_steps = self.prepare_timesteps(num_inference_steps, sigmas, latents.shape[1]) self._num_timesteps = len(timesteps) # Denoising loop using diffuse method latents = self.diffuse( latents=latents, timesteps=timesteps, prompt_embeds=prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds, negative_prompt_embeds=negative_prompt_embeds if do_cfg else None, negative_pooled_prompt_embeds=negative_pooled_prompt_embeds if do_cfg else None, do_true_cfg=do_cfg, guidance_scale=self.guidance_scale, cfg_normalize=False, ) self._current_timestep = None if self.output_type == "latent": image = latents else: latents = latents.to(self.vae.dtype) latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor image = self.vae.decode(latents, return_dict=False)[0] return DiffusionOutput(output=image) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: loader = AutoWeightsLoader(self) return loader.load_weights(weights)