From 4af6c5cf30423756f1beaf0457d49ab126dfc641 Mon Sep 17 00:00:00 2001 From: Jaret Burkett Date: Mon, 17 Feb 2025 14:10:25 -0700 Subject: [PATCH] Work on supporting flex.2 potential arch --- jobs/process/TrainSliderProcess.py | 32 +- toolkit/config_modules.py | 3 + toolkit/models/flex2.py | 847 ++++++++++++++++++++++++++++- toolkit/stable_diffusion_model.py | 59 +- 4 files changed, 918 insertions(+), 23 deletions(-) diff --git a/jobs/process/TrainSliderProcess.py b/jobs/process/TrainSliderProcess.py index 88b9d104..38dc40c3 100644 --- a/jobs/process/TrainSliderProcess.py +++ b/jobs/process/TrainSliderProcess.py @@ -275,6 +275,8 @@ class TrainSliderProcess(BaseSDTrainProcess): return adapter_tensors def hook_train_loop(self, batch: Union['DataLoaderBatchDTO', None]): + if isinstance(batch, list): + batch = batch[0] # set to eval mode self.sd.set_device_state(self.eval_slider_device_state) with torch.no_grad(): @@ -364,10 +366,32 @@ class TrainSliderProcess(BaseSDTrainProcess): denoised_latents = torch.cat([noisy_latents] * self.prompt_chunk_size, dim=0) current_timestep = timesteps else: - - self.sd.noise_scheduler.set_timesteps( - self.train_config.max_denoising_steps, device=self.device_torch - ) + if self.train_config.noise_scheduler == 'flowmatch': + linear_timesteps = any([ + self.train_config.linear_timesteps, + self.train_config.linear_timesteps2, + self.train_config.timestep_type == 'linear', + ]) + + timestep_type = 'linear' if linear_timesteps else None + if timestep_type is None: + timestep_type = self.train_config.timestep_type + + # make fake latents + l = torch.randn( + true_batch_size, 16, height, width + ).to(self.device_torch, dtype=dtype) + + self.sd.noise_scheduler.set_train_timesteps( + self.train_config.max_denoising_steps, + device=self.device_torch, + timestep_type=timestep_type, + latents=l + ) + else: + self.sd.noise_scheduler.set_timesteps( + self.train_config.max_denoising_steps, device=self.device_torch + ) # ger a random number of steps timesteps_to = torch.randint( diff --git a/toolkit/config_modules.py b/toolkit/config_modules.py index 44091b55..dfffae7c 100644 --- a/toolkit/config_modules.py +++ b/toolkit/config_modules.py @@ -428,6 +428,9 @@ class ModelConfig: self.is_auraflow: bool = kwargs.get('is_auraflow', False) self.is_v3: bool = kwargs.get('is_v3', False) self.is_flux: bool = kwargs.get('is_flux', False) + self.is_flex2: bool = kwargs.get('is_flex2', False) + if self.is_flex2: + self.is_flux = True self.is_lumina2: bool = kwargs.get('is_lumina2', False) if self.is_pixart_sigma: self.is_pixart = True diff --git a/toolkit/models/flex2.py b/toolkit/models/flex2.py index d8eae2e6..eabff80c 100644 --- a/toolkit/models/flex2.py +++ b/toolkit/models/flex2.py @@ -1,15 +1,211 @@ from typing import List, Optional, Union from diffusers import FluxPipeline +import inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import numpy as np import torch from diffusers.loaders import FluxLoraLoaderMixin, TextualInversionLoaderMixin -from diffusers.utils import USE_PEFT_BACKEND, scale_lora_layers, unscale_lora_layers, logging +from diffusers.utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) from transformers import AutoModel, AutoTokenizer +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPVisionModelWithProjection +) + + + +from diffusers.image_processor import PipelineImageInput, VaeImageProcessor +from diffusers.loaders import FluxIPAdapterMixin, FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from diffusers.models import AutoencoderKL, FluxTransformer2DModel +from diffusers.schedulers import FlowMatchEulerDiscreteScheduler +from diffusers.utils.torch_utils import randn_tensor +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput logger = logging.get_logger(__name__) # pylint: disable=invalid-name -class Flex2Pipeline(FluxPipeline): - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import Flex2Pipeline + + >>> pipe = Flex2Pipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe(prompt, num_inference_steps=4, guidance_scale=0.0).images[0] + >>> image.save("flux.png") + ``` +""" + + +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 + + + +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.16, +): + 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 + + +# 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 Flex2Pipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, + FluxIPAdapterMixin, +): + r""" + The Flux pipeline for text-to-image generation. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`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_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: AutoModel, + tokenizer_2: AutoTokenizer, + transformer: FluxTransformer2DModel, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + 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.system_prompt = "You are an assistant designed to generate superior images with the superior degree of image-text alignment based on textual prompts or user prompts. " # determine length of system prompt @@ -18,8 +214,51 @@ class Flex2Pipeline(FluxPipeline): padding="longest", return_tensors="pt", ).input_ids[0].shape[0] - - + + def _get_clip_prompt_embeds( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: int = 1, + device: Optional[torch.device] = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + 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[:, 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 = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + def _get_llm_prompt_embeds( self, prompt: Union[str, List[str]] = None, @@ -157,4 +396,598 @@ class Flex2Pipeline(FluxPipeline): dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) - return prompt_embeds, pooled_prompt_embeds, text_ids \ No newline at end of file + return prompt_embeds, pooled_prompt_embeds, text_ids + + def encode_image(self, image, device, num_images_per_prompt): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + return image_embeds + + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt + ): + image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.transformer.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.transformer.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.transformer.encoder_hid_proj.image_projection_layers + ): + single_image_embeds = self.encode_image(single_ip_adapter_image, device, 1) + + image_embeds.append(single_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + 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_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 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)}") + + 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." + ) + + 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_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + 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}") + + @staticmethod + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + 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, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is not None: + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + return latents.to(device=device, dtype=dtype), latent_image_ids + + 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) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + return latents, latent_image_ids + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @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 + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt: Union[str, List[str]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + true_cfg_scale: float = 1.0, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 28, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 3.5, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + negative_ip_adapter_image: Optional[PipelineImageInput] = None, + negative_ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + joint_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 = 512, + ): + 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. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used 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 `true_cfg_scale` is + not greater than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance. + 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_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. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + 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_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images 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. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + negative_ip_adapter_image: + (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + negative_ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not + provided, embeddings are computed from the `ip_adapter_image` 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. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled 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.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_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 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_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 + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + has_neg_prompt = negative_prompt is not None or ( + negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None + ) + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + if do_true_cfg: + ( + negative_prompt_embeds, + negative_pooled_prompt_embeds, + _, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, latent_image_ids = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + 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), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) and ( + negative_ip_adapter_image is None and negative_ip_adapter_image_embeds is None + ): + negative_ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + elif (ip_adapter_image is None and ip_adapter_image_embeds is None) and ( + negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None + ): + ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8) + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {} + + image_embeds = None + negative_image_embeds = None + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + if negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None: + negative_image_embeds = self.prepare_ip_adapter_image_embeds( + negative_ip_adapter_image, + negative_ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + ) + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + if image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + if do_true_cfg: + if negative_image_embeds is not None: + self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds + neg_noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=negative_pooled_prompt_embeds, + encoder_hidden_states=negative_prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + 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": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/toolkit/stable_diffusion_model.py b/toolkit/stable_diffusion_model.py index 212e68c2..7a765770 100644 --- a/toolkit/stable_diffusion_model.py +++ b/toolkit/stable_diffusion_model.py @@ -51,6 +51,7 @@ from diffusers import StableDiffusionPipeline, StableDiffusionXLPipeline, T2IAda StableDiffusion3Img2ImgPipeline, PixArtSigmaPipeline, AuraFlowPipeline, AuraFlowTransformer2DModel, FluxPipeline, \ FluxTransformer2DModel, FlowMatchEulerDiscreteScheduler, SD3Transformer2DModel, Lumina2Text2ImgPipeline from toolkit.models.lumina2 import Lumina2Transformer2DModel +from toolkit.models.flex2 import Flex2Pipeline import diffusers from diffusers import \ AutoencoderKL, \ @@ -184,6 +185,7 @@ class StableDiffusion: self.is_pixart = model_config.is_pixart self.is_auraflow = model_config.is_auraflow self.is_flux = model_config.is_flux + self.is_flex2 = model_config.is_flex2 self.is_lumina2 = model_config.is_lumina2 self.use_text_encoder_1 = model_config.use_text_encoder_1 @@ -699,17 +701,25 @@ class StableDiffusion: print_acc("Loading vae") vae = AutoencoderKL.from_pretrained(base_model_path, subfolder="vae", torch_dtype=dtype) flush() - - print_acc("Loading t5") - tokenizer_2 = T5TokenizerFast.from_pretrained(base_model_path, subfolder="tokenizer_2", torch_dtype=dtype) - text_encoder_2 = T5EncoderModel.from_pretrained(base_model_path, subfolder="text_encoder_2", - torch_dtype=dtype) + + if self.is_flex2: + tokenizer_2 = AutoTokenizer.from_pretrained(base_model_path, subfolder="tokenizer_2") + text_encoder_2 = AutoModel.from_pretrained(base_model_path, subfolder="text_encoder_2", torch_dtype=dtype) + + else: + print_acc("Loading t5") + tokenizer_2 = T5TokenizerFast.from_pretrained(base_model_path, subfolder="tokenizer_2", torch_dtype=dtype) + text_encoder_2 = T5EncoderModel.from_pretrained(base_model_path, subfolder="text_encoder_2", + torch_dtype=dtype) text_encoder_2.to(self.device_torch, dtype=dtype) flush() if self.model_config.quantize_te: - print_acc("Quantizing T5") + if self.is_flex2: + print_acc("Quantizing LLM") + else: + print_acc("Quantizing T5") quantize(text_encoder_2, weights=qfloat8) freeze(text_encoder_2) flush() @@ -720,7 +730,11 @@ class StableDiffusion: text_encoder.to(self.device_torch, dtype=dtype) print_acc("making pipe") - pipe: FluxPipeline = FluxPipeline( + Pipe = FluxPipeline + if self.is_flex2: + Pipe = Flex2Pipeline + + pipe: Pipe = Pipe( scheduler=scheduler, text_encoder=text_encoder, tokenizer=tokenizer, @@ -1079,6 +1093,8 @@ class StableDiffusion: arch = 'pixart' if self.is_flux: arch = 'flux' + if self.is_flex2: + arch = 'flex2' if self.is_lumina2: arch = 'lumina2' noise_scheduler = get_sampler( @@ -1152,7 +1168,11 @@ class StableDiffusion: ) else: - pipeline = FluxPipeline( + Pipe = FluxPipeline + if self.is_flex2: + Pipe = Flex2Pipeline + + pipeline = Pipe( vae=self.vae, transformer=unwrap_model(self.unet), text_encoder=unwrap_model(self.text_encoder[0]), @@ -2292,6 +2312,18 @@ class StableDiffusion: embeds, attention_mask=attention_mask, # not used ) + elif self.is_flex2: + prompt_embeds, pooled_prompt_embeds, text_ids = self.pipeline.encode_prompt( + prompt, + prompt, + device=self.device_torch, + max_sequence_length=512, + ) + pe = PromptEmbeds( + prompt_embeds + ) + pe.pooled_embeds = pooled_prompt_embeds + return pe elif self.is_flux: prompt_embeds, pooled_prompt_embeds = train_tools.encode_prompts_flux( self.tokenizer, # list @@ -2761,10 +2793,13 @@ class StableDiffusion: if isinstance(self.text_encoder, list): self.device_state['text_encoder']: List[dict] = [] for encoder in self.text_encoder: - try: - te_has_grad = encoder.text_model.final_layer_norm.weight.requires_grad - except: - te_has_grad = encoder.encoder.block[0].layer[0].SelfAttention.q.weight.requires_grad + if isinstance(encoder, LlamaModel): + te_has_grad = encoder.layers[0].mlp.gate_proj.weight.requires_grad + else: + try: + te_has_grad = encoder.text_model.final_layer_norm.weight.requires_grad + except: + te_has_grad = encoder.encoder.block[0].layer[0].SelfAttention.q.weight.requires_grad self.device_state['text_encoder'].append({ 'training': encoder.training, 'device': encoder.device,