diff --git a/CHANGELOG.md b/CHANGELOG.md index e0c9e0fec..d130a0773 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -6,7 +6,7 @@ Feature highlights include: - **ModernUI** layout redesign which should make it more user friendly and easier to navigate -- New models [WanAI Wan 2.1](https://wan.video/) for text-to-image workflows and [FreePix F-Lite](https://huggingface.co/Freepik/F-Lite) +- New models [WanAI Wan 2.1](https://wan.video/) for text-to-image workflows, [FreePix F-Lite](https://huggingface.co/Freepik/F-Lite), [Bria 3.2](https://huggingface.co/briaai/BRIA-3.2) - Redesigned [LTXVideo](https://vladmandic.github.io/sdnext-docs/Video) interface with support for general video models plus optimized [FramePack](https://vladmandic.github.io/sdnext-docs/FramePack) and [LTXVideo](https://vladmandic.github.io/sdnext-docs/LTX) support - Fully integrated nudity detection and optional censorship with [NudeNet](https://vladmandic.github.io/sdnext-docs/NudeNet) - New background replacement and relightning methods using **Latent Bridge Matching** and new **PixelArt** processing filter @@ -44,6 +44,10 @@ For details, see [ChangeLog](https://github.com/vladmandic/automatic/blob/master - [FreePix F-Lite](https://huggingface.co/Freepik/F-Lite) F-Lite is a 10B model trained exclusively on copyright-safe and SFW content, trained on internal dataset comprising approximately 80 million copyright-safe images available via *networks -> models -> reference* + - [Bria 3.2](https://huggingface.co/briaai/BRIA-3.2) + Bria is a smaller 4B parameter model built entirely on licensed data and safe for commercial use + *note*: this is a gated model, you need to [accept terms](https://huggingface.co/briaai/BRIA-3.2) and set your [huggingface token](https://vladmandic.github.io/sdnext-docs/Gated/) + available via *networks -> models -> reference* - [LBM: Latent Bridge Matching](https://github.com/gojasper/LBM) very fast automatic image background replacement methods with relightning! *simple*: automatic background replacement using [BiRefNet](https://github.com/ZhengPeng7/BiRefNet) diff --git a/html/reference.json b/html/reference.json index c7da5a67b..4561a8eef 100644 --- a/html/reference.json +++ b/html/reference.json @@ -506,6 +506,13 @@ "skip": true }, + "Bria 3.2": { + "path": "briaai/BRIA-3.2", + "desc": "Bria 3.2 is the next-generation commercial-ready text-to-image model. With just 4 billion parameters, it provides exceptional aesthetics and text rendering, evaluated to provide on par results to leading open-source models, and outperforming other licensed models.", + "preview": "briaai--BRIA-3.2.jpg", + "skip": true + }, + "Meissonic": { "path": "MeissonFlow/Meissonic", "desc": "Meissonic is a non-autoregressive mask image modeling text-to-image synthesis model that can generate high-resolution images. It is designed to run on consumer graphics cards.", diff --git a/models/Reference/briaai--BRIA-3.2.jpg b/models/Reference/briaai--BRIA-3.2.jpg new file mode 100644 index 000000000..dc7067ec5 Binary files /dev/null and b/models/Reference/briaai--BRIA-3.2.jpg differ diff --git a/modules/modeldata.py b/modules/modeldata.py index 11b3b7c61..48d21028d 100644 --- a/modules/modeldata.py +++ b/modules/modeldata.py @@ -56,6 +56,8 @@ def get_model_type(pipe): model_type = 'pixartsigma' elif "PixArtAlpha" in name: model_type = 'pixartalpha' + elif "Bria" in name: + model_type = 'bria' # video models elif "CogVideo" in name: model_type = 'cogvideo' diff --git a/modules/sd_detect.py b/modules/sd_detect.py index 93dcd7e73..b16478580 100644 --- a/modules/sd_detect.py +++ b/modules/sd_detect.py @@ -99,6 +99,8 @@ def guess_by_name(fn, current_guess): return 'FLite' elif 'wan' in fn.lower(): return 'WanAI' + elif 'bria' in fn.lower(): + return 'Bria' return current_guess diff --git a/modules/sd_models.py b/modules/sd_models.py index 9dba245ed..d8ef24a3e 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -369,6 +369,10 @@ def load_diffuser_force(model_type, checkpoint_info, diffusers_load_config, op=' from pipelines.model_wanai import load_wan sd_model = load_wan(checkpoint_info, diffusers_load_config) allow_post_quant = False + elif model_type in ['Bria']: + from pipelines.model_bria import load_bria + sd_model = load_bria(checkpoint_info, diffusers_load_config) + allow_post_quant = False except Exception as e: shared.log.error(f'Load {op}: path="{checkpoint_info.path}" {e}') if debug_load: diff --git a/modules/shared_items.py b/modules/shared_items.py index e735c68ab..483186905 100644 --- a/modules/shared_items.py +++ b/modules/shared_items.py @@ -47,12 +47,13 @@ pipelines = { 'WanAI': getattr(diffusers, 'WanPipeline', None), # dynamically imported and redefined later - 'Meissonic': getattr(diffusers, 'DiffusionPipeline', None), # dynamically redefined and loaded in sd_models.load_diffuser - 'Monetico': getattr(diffusers, 'DiffusionPipeline', None), # dynamically redefined and loaded in sd_models.load_diffuser - 'OmniGen2': getattr(diffusers, 'DiffusionPipeline', None), # dynamically redefined and loaded in sd_models.load_diffuser - 'InstaFlow': getattr(diffusers, 'DiffusionPipeline', None), # dynamically redefined and loaded in sd_models.load_diffuser - 'SegMoE': getattr(diffusers, 'DiffusionPipeline', None), # dynamically redefined and loaded in sd_models.load_diffuser + 'Meissonic': getattr(diffusers, 'DiffusionPipeline', None), + 'Monetico': getattr(diffusers, 'DiffusionPipeline', None), + 'OmniGen2': getattr(diffusers, 'DiffusionPipeline', None), + 'InstaFlow': getattr(diffusers, 'DiffusionPipeline', None), + 'SegMoE': getattr(diffusers, 'DiffusionPipeline', None), 'FLite': getattr(diffusers, 'DiffusionPipeline', None), + 'Bria': getattr(diffusers, 'DiffusionPipeline', None), } initialize_onnx() diff --git a/pipelines/bria/__init__.py b/pipelines/bria/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/pipelines/bria/bria_pipeline.py b/pipelines/bria/bria_pipeline.py new file mode 100644 index 000000000..c07f6e4de --- /dev/null +++ b/pipelines/bria/bria_pipeline.py @@ -0,0 +1,651 @@ +from diffusers.pipelines.flux.pipeline_flux import FluxPipeline, retrieve_timesteps, calculate_shift +from typing import Any, Callable, Dict, List, Optional, Union + +import torch + +from transformers import ( + T5EncoderModel, + T5TokenizerFast, +) + +from diffusers.image_processor import VaeImageProcessor +from diffusers import AutoencoderKL , DDIMScheduler, EulerAncestralDiscreteScheduler +from diffusers.schedulers import FlowMatchEulerDiscreteScheduler +from diffusers.schedulers import KarrasDiffusionSchedulers +from diffusers.loaders import FluxLoraLoaderMixin +from diffusers.utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput +from pipelines.bria.transformer_bria import BriaTransformer2DModel +from pipelines.bria.bria_utils import get_t5_prompt_embeds, get_original_sigmas, is_ng_none +from diffusers.utils.torch_utils import randn_tensor +import diffusers +import numpy as np +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 StableDiffusion3Pipeline + + >>> pipe = StableDiffusion3Pipeline.from_pretrained( + ... "stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> image = pipe(prompt).images[0] + >>> image.save("sd3.png") + ``` +""" + +""" +Based on FluxPipeline with several changes: +- no pooled embeddings +- We use zero padding for prompts +- No guidance embedding since this is not a distilled version +""" +class BriaPipeline(FluxPipeline): + r""" + Args: + transformer ([`SD3Transformer2DModel`]): + 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 ([`T5EncoderModel`]): + Frozen text-encoder. Stable Diffusion 3 uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`T5TokenizerFast`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + """ + + def __init__( + self, + transformer: BriaTransformer2DModel, + scheduler: Union[FlowMatchEulerDiscreteScheduler,KarrasDiffusionSchedulers], + vae: AutoencoderKL, + text_encoder: T5EncoderModel, + tokenizer: T5TokenizerFast + ): + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + # TODO - why different than offical flux (-1) + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.block_out_channels)) if hasattr(self, "vae") and self.vae is not None else 16 + ) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.default_sample_size = 64 # due to patchify=> 128,128 => res of 1k,1k + + # T5 is senstive to precision so we use the precision used for precompute and cast as needed + for block in self.text_encoder.encoder.block: + block.layer[-1].DenseReluDense.wo.to(dtype=torch.float32) + + if self.vae.config.shift_factor is None: + self.vae.config.shift_factor=0 + self.vae.to(dtype=torch.float32) + + + def encode_prompt( + self, + prompt: Union[str, List[str]], + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[Union[str, List[str]]] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + max_sequence_length: int = 128, + lora_scale: Optional[float] = None, + ): + r""" + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + 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`). + 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. + 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. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + + 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 = get_t5_prompt_embeds( + self.tokenizer, + self.text_encoder, + prompt=prompt, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ).to(dtype=self.transformer.dtype) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + if not is_ng_none(negative_prompt): + 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 = get_t5_prompt_embeds( + self.tokenizer, + self.text_encoder, + prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ).to(dtype=self.transformer.dtype) + else: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(batch_size, prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + text_ids = text_ids.repeat(num_images_per_prompt, 1, 1) + + return prompt_embeds, negative_prompt_embeds, text_ids + + @property + def guidance_scale(self): + return self._guidance_scale + + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @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, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 30, + timesteps: List[int] = None, + guidance_scale: float = 5, + negative_prompt: Optional[Union[str, List[str]]] = None, + 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, + negative_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 = 128, + clip_value:Union[None,float] = None, + normalize:bool = False + ): + 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.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. + 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`, *optional*, 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. + 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`). + 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. + 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. + 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.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] 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 256): 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, + height=height, + width=width, + prompt_embeds=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._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 + ) + + ( + prompt_embeds, + negative_prompt_embeds, + text_ids + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + + + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 # due to patch=2, we devide by 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, + ) + + if isinstance(self.scheduler,FlowMatchEulerDiscreteScheduler) and self.scheduler.config['use_dynamic_shifting']: + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) + image_seq_len = latents.shape[1] # Shift by height - Why just height? + + mu = calculate_shift( + image_seq_len, + self.scheduler.config.base_image_seq_len, + self.scheduler.config.max_image_seq_len, + self.scheduler.config.base_shift, + self.scheduler.config.max_shift, + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + timesteps, + sigmas, + mu=mu, + ) + else: + # 4. Prepare timesteps + # Sample from training sigmas + if isinstance(self.scheduler,DDIMScheduler) or isinstance(self.scheduler,EulerAncestralDiscreteScheduler): + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, None, None) + else: + sigmas = get_original_sigmas(num_train_timesteps=self.scheduler.config.num_train_timesteps,num_inference_steps=num_inference_steps) + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps,sigmas=sigmas) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # Supprot different diffusers versions + if diffusers.__version__>='0.32.0': + latent_image_ids=latent_image_ids[0] + text_ids=text_ids[0] + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + if type(self.scheduler)!=FlowMatchEulerDiscreteScheduler: + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + # This is predicts "v" from flow-matching or eps from diffusion + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + txt_ids=text_ids, + img_ids=latent_image_ids, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + cfg_noise_pred_text = noise_pred_text.std() + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if normalize: + noise_pred = noise_pred * (0.7 *(cfg_noise_pred_text/noise_pred.std())) + 0.3 * noise_pred + + if clip_value: + assert clip_value>0 + noise_pred = noise_pred.clip(-clip_value,clip_value) + + # 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) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_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() + + if output_type == "latent": + image = latents + + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents.to(dtype=torch.float32) / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents.to(dtype=self.vae.dtype), 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) + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=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 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." + ) + + + 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 to(self, *args, **kwargs): + DiffusionPipeline.to(self, *args, **kwargs) + # T5 is senstive to precision so we use the precision used for precompute and cast as needed + for block in self.text_encoder.encoder.block: + block.layer[-1].DenseReluDense.wo.to(dtype=torch.float32) + + if self.vae.config.shift_factor == 0 and self.vae.dtype!=torch.float32: + self.vae.to(dtype=torch.float32) + + + return self + + + 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) + width = 2 * (int(width) // self.vae_scale_factor ) + + 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 + + @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 + + height = height // vae_scale_factor + width = width // vae_scale_factor + + latents = latents.view(batch_size, height, width, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height * 2, width * 2) + + return latents + + @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.repeat(batch_size, 1, 1, 1) + latent_image_ids = latent_image_ids.reshape( + batch_size, latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) diff --git a/pipelines/bria/bria_utils.py b/pipelines/bria/bria_utils.py new file mode 100644 index 000000000..3cddeafa1 --- /dev/null +++ b/pipelines/bria/bria_utils.py @@ -0,0 +1,443 @@ +from typing import Union, Optional, List +import torch +from diffusers.utils import logging +from transformers import ( + T5EncoderModel, + T5TokenizerFast, + AutoTokenizer +) +from transformers import ( + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer +) + +import numpy as np +import torch.distributed as dist +import math +import os + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def get_text(caption): + + existing_text_list = set() + + if caption[0]=='\"' and caption[-1]=='\"': + caption=caption[1:-2] + + if caption[0]=='\'' and caption[-1]=='\'': + caption=caption[1:-2] + + text_list=[] + current_text='' + text_present = False + for c in caption: + if c=='\"' and not text_present: + text_present=True + continue + + if c=='\"' and text_present: + if current_text not in existing_text_list: + text_list+=[current_text] + existing_text_list.add(current_text) + + text_present=False + current_text='' + continue + + if text_present: + current_text+=c + + return text_list + +def get_by_t5_prompt_embeds( + tokenizer: AutoTokenizer , + text_encoder: T5EncoderModel, + prompt: Union[str, List[str]], + max_sequence_length: int = 128, + device: Optional[torch.device] = None, +): + device = device or text_encoder.device + + if isinstance(prompt, list): + assert len(prompt)==1 + prompt=prompt[0] + + assert type(prompt)==str + + captions_list = get_text(prompt) + embeddings_list=[] + for inner_prompt in captions_list: + text_inputs = tokenizer( + [inner_prompt], + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + prompt_embeds = text_encoder(text_input_ids.to(device))[0] + embeddings_list+=[prompt_embeds[0]] + + # No Text Found + if len(embeddings_list)==0: + return None + + prompt_embeds = torch.concatenate(embeddings_list,axis=0) + + # Concat zeros to max_sequence + seq_len, dim = prompt_embeds.shape + if seq_len= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = 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 = text_encoder(text_input_ids.to(device))[0] + + # Concat zeros to max_sequence + b, seq_len, dim = prompt_embeds.shape + if seq_len torch.Tensor: + n_axes = ids.shape[-1] + cos_out = [] + sin_out = [] + pos = ids.float() + is_mps = ids.device.type == "mps" + freqs_dtype = torch.float32 if is_mps else torch.float64 + for i in range(n_axes): + cos, sin = get_1d_rotary_pos_embed( + self.axes_dim[i], + pos[:, i], + theta=self.theta, + repeat_interleave_real=True, + use_real=True, + freqs_dtype=freqs_dtype, + ) + cos_out.append(cos) + sin_out.append(sin) + freqs_cos = torch.cat(cos_out, dim=-1).to(ids.device) + freqs_sin = torch.cat(sin_out, dim=-1).to(ids.device) + return freqs_cos, freqs_sin + +from diffusers.optimization import get_scheduler +from torch.optim import Optimizer +from torch.optim.lr_scheduler import LambdaLR + +# Not really cosine but with decay +def get_cosine_schedule_with_warmup_and_decay( + optimizer: Optimizer, num_warmup_steps: int, num_training_steps: int, num_cycles: float = 0.5, last_epoch: int = -1, constant_steps=-1,eps=1e-5 +) -> LambdaLR: + + """ + Create a schedule with a learning rate that decreases following the values of the cosine function between the + initial lr set in the optimizer to 0, after a warmup period during which it increases linearly between 0 and the + initial lr set in the optimizer. + + Args: + optimizer ([`~torch.optim.Optimizer`]): + The optimizer for which to schedule the learning rate. + num_warmup_steps (`int`): + The number of steps for the warmup phase. + num_training_steps (`int`): + The total number of training steps. + num_periods (`float`, *optional*, defaults to 0.5): + The number of periods of the cosine function in a schedule (the default is to just decrease from the max + value to 0 following a half-cosine). + last_epoch (`int`, *optional*, defaults to -1): + The index of the last epoch when resuming training. + constant_steps (`int`): + The total number of constant lr steps following a warmup + + Return: + `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule. + """ + if constant_steps <=0: + constant_steps = num_training_steps-num_warmup_steps + + def lr_lambda(current_step): + # Accelerate sends current_step*num_processes + if current_step < num_warmup_steps: + return float(current_step) / float(max(1, num_warmup_steps)) + elif current_step torch.Tensor: + residual = hidden_states + norm_hidden_states, gate = self.norm(hidden_states, emb=temb) + mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states)) + joint_attention_kwargs = joint_attention_kwargs or {} + attn_output = self.attn( + hidden_states=norm_hidden_states, + image_rotary_emb=image_rotary_emb, + **joint_attention_kwargs, + ) + + hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2) + gate = gate.unsqueeze(1) + hidden_states = gate * self.proj_out(hidden_states) + hidden_states = residual + hidden_states + if hidden_states.dtype == torch.float16: + hidden_states = hidden_states.clip(-65504, 65504) + + return hidden_states + + +@maybe_allow_in_graph +class FluxTransformerBlock(nn.Module): + def __init__( + self, dim: int, num_attention_heads: int, attention_head_dim: int, qk_norm: str = "rms_norm", eps: float = 1e-6 + ): + super().__init__() + + self.norm1 = AdaLayerNormZero(dim) + self.norm1_context = AdaLayerNormZero(dim) + + self.attn = Attention( + query_dim=dim, + cross_attention_dim=None, + added_kv_proj_dim=dim, + dim_head=attention_head_dim, + heads=num_attention_heads, + out_dim=dim, + context_pre_only=False, + bias=True, + processor=FluxAttnProcessor2_0(), + qk_norm=qk_norm, + eps=eps, + ) + + self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6) + self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate") + + self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6) + self.ff_context = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate") + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor, + temb: torch.Tensor, + image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb) + + norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context( + encoder_hidden_states, emb=temb + ) + joint_attention_kwargs = joint_attention_kwargs or {} + # Attention. + attention_outputs = self.attn( + hidden_states=norm_hidden_states, + encoder_hidden_states=norm_encoder_hidden_states, + image_rotary_emb=image_rotary_emb, + **joint_attention_kwargs, + ) + + if len(attention_outputs) == 2: + attn_output, context_attn_output = attention_outputs + elif len(attention_outputs) == 3: + attn_output, context_attn_output, ip_attn_output = attention_outputs + + # Process attention outputs for the `hidden_states`. + attn_output = gate_msa.unsqueeze(1) * attn_output + hidden_states = hidden_states + attn_output + + norm_hidden_states = self.norm2(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] + + ff_output = self.ff(norm_hidden_states) + ff_output = gate_mlp.unsqueeze(1) * ff_output + + hidden_states = hidden_states + ff_output + if len(attention_outputs) == 3: + hidden_states = hidden_states + ip_attn_output + + # Process attention outputs for the `encoder_hidden_states`. + + context_attn_output = c_gate_msa.unsqueeze(1) * context_attn_output + encoder_hidden_states = encoder_hidden_states + context_attn_output + + norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states) + norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None] + + context_ff_output = self.ff_context(norm_encoder_hidden_states) + encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output + if encoder_hidden_states.dtype == torch.float16: + encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504) + + return encoder_hidden_states, hidden_states + + +class FluxTransformer2DModel( + ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, FluxTransformer2DLoadersMixin, CacheMixin +): + """ + The Transformer model introduced in Flux. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + patch_size (`int`, defaults to `1`): + Patch size to turn the input data into small patches. + in_channels (`int`, defaults to `64`): + The number of channels in the input. + out_channels (`int`, *optional*, defaults to `None`): + The number of channels in the output. If not specified, it defaults to `in_channels`. + num_layers (`int`, defaults to `19`): + The number of layers of dual stream DiT blocks to use. + num_single_layers (`int`, defaults to `38`): + The number of layers of single stream DiT blocks to use. + attention_head_dim (`int`, defaults to `128`): + The number of dimensions to use for each attention head. + num_attention_heads (`int`, defaults to `24`): + The number of attention heads to use. + joint_attention_dim (`int`, defaults to `4096`): + The number of dimensions to use for the joint attention (embedding/channel dimension of + `encoder_hidden_states`). + pooled_projection_dim (`int`, defaults to `768`): + The number of dimensions to use for the pooled projection. + guidance_embeds (`bool`, defaults to `False`): + Whether to use guidance embeddings for guidance-distilled variant of the model. + axes_dims_rope (`Tuple[int]`, defaults to `(16, 56, 56)`): + The dimensions to use for the rotary positional embeddings. + """ + + _supports_gradient_checkpointing = True + _no_split_modules = ["FluxTransformerBlock", "FluxSingleTransformerBlock"] + _skip_layerwise_casting_patterns = ["pos_embed", "norm"] + + @register_to_config + def __init__( + self, + patch_size: int = 1, + in_channels: int = 64, + out_channels: Optional[int] = None, + num_layers: int = 19, + num_single_layers: int = 38, + attention_head_dim: int = 128, + num_attention_heads: int = 24, + joint_attention_dim: int = 4096, + pooled_projection_dim: int = 768, + guidance_embeds: bool = False, + axes_dims_rope: Tuple[int, int, int] = (16, 56, 56), + ): + super().__init__() + self.out_channels = out_channels or in_channels + self.inner_dim = num_attention_heads * attention_head_dim + + self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope) + + text_time_guidance_cls = ( + CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings + ) + self.time_text_embed = text_time_guidance_cls( + embedding_dim=self.inner_dim, pooled_projection_dim=pooled_projection_dim + ) + + self.context_embedder = nn.Linear(joint_attention_dim, self.inner_dim) + self.x_embedder = nn.Linear(in_channels, self.inner_dim) + + self.transformer_blocks = nn.ModuleList( + [ + FluxTransformerBlock( + dim=self.inner_dim, + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + ) + for _ in range(num_layers) + ] + ) + + self.single_transformer_blocks = nn.ModuleList( + [ + FluxSingleTransformerBlock( + dim=self.inner_dim, + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + ) + for _ in range(num_single_layers) + ] + ) + + self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6) + self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True) + + self.gradient_checkpointing = False + + @property + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor() + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor + def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedFluxAttnProcessor2_0 + def fuse_qkv_projections(self): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value) + are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is 🧪 experimental. + + + """ + self.original_attn_processors = None + + for _, attn_processor in self.attn_processors.items(): + if "Added" in str(attn_processor.__class__.__name__): + raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.") + + self.original_attn_processors = self.attn_processors + + for module in self.modules(): + if isinstance(module, Attention): + module.fuse_projections(fuse=True) + + self.set_attn_processor(FusedFluxAttnProcessor2_0()) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections + def unfuse_qkv_projections(self): + """Disables the fused QKV projection if enabled. + + + + This API is 🧪 experimental. + + + + """ + if self.original_attn_processors is not None: + self.set_attn_processor(self.original_attn_processors) + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor = None, + pooled_projections: torch.Tensor = None, + timestep: torch.LongTensor = None, + img_ids: torch.Tensor = None, + txt_ids: torch.Tensor = None, + guidance: torch.Tensor = None, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + controlnet_block_samples=None, + controlnet_single_block_samples=None, + return_dict: bool = True, + controlnet_blocks_repeat: bool = False, + ) -> Union[torch.Tensor, Transformer2DModelOutput]: + """ + The [`FluxTransformer2DModel`] forward method. + + Args: + hidden_states (`torch.Tensor` of shape `(batch_size, image_sequence_length, in_channels)`): + Input `hidden_states`. + encoder_hidden_states (`torch.Tensor` of shape `(batch_size, text_sequence_length, joint_attention_dim)`): + Conditional embeddings (embeddings computed from the input conditions such as prompts) to use. + pooled_projections (`torch.Tensor` of shape `(batch_size, projection_dim)`): Embeddings projected + from the embeddings of input conditions. + timestep ( `torch.LongTensor`): + Used to indicate denoising step. + block_controlnet_hidden_states: (`list` of `torch.Tensor`): + A list of tensors that if specified are added to the residuals of transformer blocks. + 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). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain + tuple. + + Returns: + If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a + `tuple` where the first element is the sample tensor. + """ + if joint_attention_kwargs is not None: + joint_attention_kwargs = joint_attention_kwargs.copy() + lora_scale = joint_attention_kwargs.pop("scale", 1.0) + else: + lora_scale = 1.0 + + if USE_PEFT_BACKEND: + # weight the lora layers by setting `lora_scale` for each PEFT layer + scale_lora_layers(self, lora_scale) + else: + if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None: + logger.warning( + "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective." + ) + + hidden_states = self.x_embedder(hidden_states) + + timestep = timestep.to(hidden_states.dtype) * 1000 + if guidance is not None: + guidance = guidance.to(hidden_states.dtype) * 1000 + + temb = ( + self.time_text_embed(timestep, pooled_projections) + if guidance is None + else self.time_text_embed(timestep, guidance, pooled_projections) + ) + encoder_hidden_states = self.context_embedder(encoder_hidden_states) + + if txt_ids.ndim == 3: + logger.warning( + "Passing `txt_ids` 3d torch.Tensor is deprecated." + "Please remove the batch dimension and pass it as a 2d torch Tensor" + ) + txt_ids = txt_ids[0] + if img_ids.ndim == 3: + logger.warning( + "Passing `img_ids` 3d torch.Tensor is deprecated." + "Please remove the batch dimension and pass it as a 2d torch Tensor" + ) + img_ids = img_ids[0] + + ids = torch.cat((txt_ids, img_ids), dim=0) + image_rotary_emb = self.pos_embed(ids) + + if joint_attention_kwargs is not None and "ip_adapter_image_embeds" in joint_attention_kwargs: + ip_adapter_image_embeds = joint_attention_kwargs.pop("ip_adapter_image_embeds") + ip_hidden_states = self.encoder_hid_proj(ip_adapter_image_embeds) + joint_attention_kwargs.update({"ip_hidden_states": ip_hidden_states}) + + for index_block, block in enumerate(self.transformer_blocks): + if torch.is_grad_enabled() and self.gradient_checkpointing: + encoder_hidden_states, hidden_states = self._gradient_checkpointing_func( + block, + hidden_states, + encoder_hidden_states, + temb, + image_rotary_emb, + ) + + else: + encoder_hidden_states, hidden_states = block( + hidden_states=hidden_states, + encoder_hidden_states=encoder_hidden_states, + temb=temb, + image_rotary_emb=image_rotary_emb, + joint_attention_kwargs=joint_attention_kwargs, + ) + + # controlnet residual + if controlnet_block_samples is not None: + interval_control = len(self.transformer_blocks) / len(controlnet_block_samples) + interval_control = int(np.ceil(interval_control)) + # For Xlabs ControlNet. + if controlnet_blocks_repeat: + hidden_states = ( + hidden_states + controlnet_block_samples[index_block % len(controlnet_block_samples)] + ) + else: + hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control] + hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1) + + for index_block, block in enumerate(self.single_transformer_blocks): + if torch.is_grad_enabled() and self.gradient_checkpointing: + hidden_states = self._gradient_checkpointing_func( + block, + hidden_states, + temb, + image_rotary_emb, + ) + + else: + hidden_states = block( + hidden_states=hidden_states, + temb=temb, + image_rotary_emb=image_rotary_emb, + joint_attention_kwargs=joint_attention_kwargs, + ) + + # controlnet residual + if controlnet_single_block_samples is not None: + interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples) + interval_control = int(np.ceil(interval_control)) + hidden_states[:, encoder_hidden_states.shape[1] :, ...] = ( + hidden_states[:, encoder_hidden_states.shape[1] :, ...] + + controlnet_single_block_samples[index_block // interval_control] + ) + + hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...] + + hidden_states = self.norm_out(hidden_states, temb) + output = self.proj_out(hidden_states) + + if USE_PEFT_BACKEND: + # remove `lora_scale` from each PEFT layer + unscale_lora_layers(self, lora_scale) + + if not return_dict: + return (output,) + + return Transformer2DModelOutput(sample=output) diff --git a/pipelines/bria/transformer_bria.py b/pipelines/bria/transformer_bria.py new file mode 100644 index 000000000..f75cde912 --- /dev/null +++ b/pipelines/bria/transformer_bria.py @@ -0,0 +1,315 @@ +from typing import Any, Dict, List, Optional, Union +import numpy as np +import torch +import torch.nn as nn +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.loaders import PeftAdapterMixin, FromOriginalModelMixin +from diffusers.models.modeling_utils import ModelMixin +from diffusers.models.normalization import AdaLayerNormContinuous +from diffusers.utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers +from diffusers.models.modeling_outputs import Transformer2DModelOutput +from diffusers.models.embeddings import TimestepEmbedding, get_timestep_embedding +from pipelines.bria.transformer_block import FluxSingleTransformerBlock, FluxTransformerBlock +from pipelines.bria.bria_utils import FluxPosEmbed as EmbedND + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +class Timesteps(nn.Module): + def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float, scale: int = 1,time_theta=10000): + super().__init__() + self.num_channels = num_channels + self.flip_sin_to_cos = flip_sin_to_cos + self.downscale_freq_shift = downscale_freq_shift + self.scale = scale + self.time_theta=time_theta + + def forward(self, timesteps): + t_emb = get_timestep_embedding( + timesteps, + self.num_channels, + flip_sin_to_cos=self.flip_sin_to_cos, + downscale_freq_shift=self.downscale_freq_shift, + scale=self.scale, + max_period=self.time_theta + ) + return t_emb + +class TimestepProjEmbeddings(nn.Module): + def __init__(self, embedding_dim, time_theta): + super().__init__() + + self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0,time_theta=time_theta) + self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim) + + def forward(self, timestep, dtype): + timesteps_proj = self.time_proj(timestep) + timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=dtype)) # (N, D) + return timesteps_emb + +""" +Based on FluxPipeline with several changes: +- no pooled embeddings +- We use zero padding for prompts +- No guidance embedding since this is not a distilled version +""" +class BriaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin): + """ + The Transformer model introduced in Flux. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Parameters: + patch_size (`int`): Patch size to turn the input data into small patches. + in_channels (`int`, *optional*, defaults to 16): The number of channels in the input. + num_layers (`int`, *optional*, defaults to 18): The number of layers of MMDiT blocks to use. + num_single_layers (`int`, *optional*, defaults to 18): The number of layers of single DiT blocks to use. + attention_head_dim (`int`, *optional*, defaults to 64): The number of channels in each head. + num_attention_heads (`int`, *optional*, defaults to 18): The number of heads to use for multi-head attention. + joint_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use. + pooled_projection_dim (`int`): Number of dimensions to use when projecting the `pooled_projections`. + guidance_embeds (`bool`, defaults to False): Whether to use guidance embeddings. + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + patch_size: int = 1, + in_channels: int = 64, + num_layers: int = 19, + num_single_layers: int = 38, + attention_head_dim: int = 128, + num_attention_heads: int = 24, + joint_attention_dim: int = 4096, + pooled_projection_dim: int = None, + guidance_embeds: bool = False, + axes_dims_rope: List[int] = [16, 56, 56], + rope_theta = 10000, + time_theta = 10000 + ): + super().__init__() + self.out_channels = in_channels + self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim + + self.pos_embed = EmbedND(theta=rope_theta, axes_dim=axes_dims_rope) + + + self.time_embed = TimestepProjEmbeddings( + embedding_dim=self.inner_dim,time_theta=time_theta + ) + + # if pooled_projection_dim: + # self.pooled_text_embed = PixArtAlphaTextProjection(pooled_projection_dim, embedding_dim=self.inner_dim, act_fn="silu") + + if guidance_embeds: + self.guidance_embed = TimestepProjEmbeddings(embedding_dim=self.inner_dim) + + self.context_embedder = nn.Linear(self.config.joint_attention_dim, self.inner_dim) + self.x_embedder = torch.nn.Linear(self.config.in_channels, self.inner_dim) + + self.transformer_blocks = nn.ModuleList( + [ + FluxTransformerBlock( + dim=self.inner_dim, + num_attention_heads=self.config.num_attention_heads, + attention_head_dim=self.config.attention_head_dim, + ) + for i in range(self.config.num_layers) + ] + ) + + self.single_transformer_blocks = nn.ModuleList( + [ + FluxSingleTransformerBlock( + dim=self.inner_dim, + num_attention_heads=self.config.num_attention_heads, + attention_head_dim=self.config.attention_head_dim, + ) + for i in range(self.config.num_single_layers) + ] + ) + + self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6) + self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True) + + self.gradient_checkpointing = False + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: torch.Tensor = None, + pooled_projections: torch.Tensor = None, + timestep: torch.LongTensor = None, + img_ids: torch.Tensor = None, + txt_ids: torch.Tensor = None, + guidance: torch.Tensor = None, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + return_dict: bool = True, + controlnet_block_samples = None, + controlnet_single_block_samples=None, + + ) -> Union[torch.FloatTensor, Transformer2DModelOutput]: + """ + The [`FluxTransformer2DModel`] forward method. + + Args: + hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`): + Input `hidden_states`. + encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`): + Conditional embeddings (embeddings computed from the input conditions such as prompts) to use. + pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected + from the embeddings of input conditions. + timestep ( `torch.LongTensor`): + Used to indicate denoising step. + block_controlnet_hidden_states: (`list` of `torch.Tensor`): + A list of tensors that if specified are added to the residuals of transformer blocks. + 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). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain + tuple. + + Returns: + If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a + `tuple` where the first element is the sample tensor. + """ + if joint_attention_kwargs is not None: + joint_attention_kwargs = joint_attention_kwargs.copy() + lora_scale = joint_attention_kwargs.pop("scale", 1.0) + else: + lora_scale = 1.0 + + if USE_PEFT_BACKEND: + # weight the lora layers by setting `lora_scale` for each PEFT layer + scale_lora_layers(self, lora_scale) + else: + if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None: + logger.warning( + "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective." + ) + hidden_states = self.x_embedder(hidden_states) + + timestep = timestep.to(hidden_states.dtype) + if guidance is not None: + guidance = guidance.to(hidden_states.dtype) + else: + guidance = None + + # temb = ( + # self.time_text_embed(timestep, pooled_projections) + # if guidance is None + # else self.time_text_embed(timestep, guidance, pooled_projections) + # ) + + temb = self.time_embed(timestep,dtype=hidden_states.dtype) + + # if pooled_projections: + # temb+=self.pooled_text_embed(pooled_projections) + + if guidance: + temb+=self.guidance_embed(guidance,dtype=hidden_states.dtype) + + encoder_hidden_states = self.context_embedder(encoder_hidden_states) + + if len(txt_ids.shape)==2: + ids = torch.cat((txt_ids, img_ids), dim=0) + else: + ids = torch.cat((txt_ids, img_ids), dim=1) + image_rotary_emb = self.pos_embed(ids) + + for index_block, block in enumerate(self.transformer_blocks): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + encoder_hidden_states, hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(block), + hidden_states, + encoder_hidden_states, + temb, + image_rotary_emb, + **ckpt_kwargs, + ) + + else: + encoder_hidden_states, hidden_states = block( + hidden_states=hidden_states, + encoder_hidden_states=encoder_hidden_states, + temb=temb, + image_rotary_emb=image_rotary_emb, + ) + + # controlnet residual + if controlnet_block_samples is not None: + interval_control = len(self.transformer_blocks) / len(controlnet_block_samples) + interval_control = int(np.ceil(interval_control)) + hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control] + + + hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1) + + for index_block, block in enumerate(self.single_transformer_blocks): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(block), + hidden_states, + temb, + image_rotary_emb, + **ckpt_kwargs, + ) + + else: + hidden_states = block( + hidden_states=hidden_states, + temb=temb, + image_rotary_emb=image_rotary_emb, + ) + + # controlnet residual + if controlnet_single_block_samples is not None: + interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples) + interval_control = int(np.ceil(interval_control)) + hidden_states[:, encoder_hidden_states.shape[1] :, ...] = ( + hidden_states[:, encoder_hidden_states.shape[1] :, ...] + + controlnet_single_block_samples[index_block // interval_control] + ) + + hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...] + + hidden_states = self.norm_out(hidden_states, temb) + output = self.proj_out(hidden_states) + + if USE_PEFT_BACKEND: + # remove `lora_scale` from each PEFT layer + unscale_lora_layers(self, lora_scale) + + if not return_dict: + return (output,) + + return Transformer2DModelOutput(sample=output) diff --git a/pipelines/model_bria.py b/pipelines/model_bria.py new file mode 100644 index 000000000..bcd458dc4 --- /dev/null +++ b/pipelines/model_bria.py @@ -0,0 +1,90 @@ +import os +import sys +import transformers +from modules import shared, devices, sd_models, model_quant, sd_hijack_te + + +def load_transformer(repo_id, diffusers_load_config={}): + load_args, quant_args = model_quant.get_dit_args(diffusers_load_config, module='Model', device_map=True) + fn = None + + if shared.opts.sd_unet is not None and shared.opts.sd_unet != 'Default': + from modules import sd_unet + if shared.opts.sd_unet not in list(sd_unet.unet_dict): + shared.log.error(f'Load module: type=Transformer not found: {shared.opts.sd_unet}') + return None + fn = sd_unet.unet_dict[shared.opts.sd_unet] if os.path.exists(sd_unet.unet_dict[shared.opts.sd_unet]) else None + + from pipelines.bria.transformer_bria import BriaTransformer2DModel + + if fn is not None and 'gguf' in fn.lower(): + shared.log.error('Load model: type=Bria format="gguf" unsupported') + transformer = None + elif fn is not None and 'safetensors' in fn.lower(): + shared.log.debug(f'Load model: type=Bria transformer="{fn}" quant="{model_quant.get_quant(repo_id)}" args={load_args}') + transformer = BriaTransformer2DModel.from_single_file( + fn, + cache_dir=shared.opts.hfcache_dir, + **load_args, + ) + else: + shared.log.debug(f'Load model: type=Bria transformer="{repo_id}" quant="{model_quant.get_quant_type(quant_args)}" args={load_args}') + transformer = BriaTransformer2DModel.from_pretrained( + repo_id, + subfolder="transformer", + cache_dir=shared.opts.hfcache_dir, + **load_args, + **quant_args, + ) + if shared.opts.diffusers_offload_mode != 'none' and transformer is not None: + sd_models.move_model(transformer, devices.cpu) + return transformer + + +def load_text_encoder(repo_id, diffusers_load_config={}): + load_args, quant_args = model_quant.get_dit_args(diffusers_load_config, module='TE', device_map=True) + shared.log.debug(f'Load model: type=Bria te="{repo_id}" quant="{model_quant.get_quant_type(quant_args)}" args={load_args}') + text_encoder = transformers.T5EncoderModel.from_pretrained( + repo_id, + subfolder="text_encoder", + cache_dir=shared.opts.hfcache_dir, + **load_args, + **quant_args, + ) + if shared.opts.diffusers_offload_mode != 'none' and text_encoder is not None: + sd_models.move_model(text_encoder, devices.cpu) + return text_encoder + + +def load_bria(checkpoint_info, diffusers_load_config={}): + repo_id = sd_models.path_to_repo(checkpoint_info) + sd_models.hf_auth_check(checkpoint_info) + + transformer = load_transformer(repo_id, diffusers_load_config) + text_encoder = load_text_encoder(repo_id, diffusers_load_config) + + load_args, _quant_args = model_quant.get_dit_args(diffusers_load_config, module='Model') + shared.log.debug(f'Load model: type=Bria model="{checkpoint_info.name}" repo="{repo_id}" offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} args={load_args}') + + from pipelines.bria.bria_pipeline import BriaPipeline + sys.path.append(os.path.join(os.path.dirname(__file__), 'bria')) + + pipe = BriaPipeline.from_pretrained( + repo_id, + transformer=transformer, + text_encoder=text_encoder, + cache_dir=shared.opts.diffusers_dir, + trust_remote_code=True, + **load_args, + ) + + del text_encoder + del transformer + + sd_hijack_te.init_hijack(pipe) + from modules.video_models import video_vae + pipe.vae.orig_decode = pipe.vae.decode + pipe.vae.decode = video_vae.hijack_vae_decode + + devices.torch_gc() + return pipe diff --git a/wiki b/wiki index 0004abde0..c134fe9e9 160000 --- a/wiki +++ b/wiki @@ -1 +1 @@ -Subproject commit 0004abde02d2d32b1a1c1518f6658ab80a2bde14 +Subproject commit c134fe9e92dc01caa3967ac5e1b1e018b3eb473d