pipeline_bria.py

from diffusers.pipelines.flux.pipeline_flux import FluxPipeline, retrieve_timesteps
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 transformer_bria import BriaTransformer2DModel
from bria_utils import get_t5_prompt_embeds, get_original_sigmas, is_ng_none

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")
        ```
"""

T5_PRECISION = torch.float16

"""
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).
    """

    # model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder->transformer->vae"
    # _optional_components = []
    # _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"]

    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
        self.text_encoder = self.text_encoder.to(dtype=T5_PRECISION)
        for block in self.text_encoder.encoder.block:
            block.layer[-1].DenseReluDense.wo.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.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] 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)
            
        # 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)

        # 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,
        )

        # Supprot different diffusers versions
        if len(latent_image_ids.shape)==2:
            text_ids=text_ids.squeeze()

        # 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.std().item()>2:
                #     print('Warning')
                
                # print(t.item(),latents.mean().item(),latents.std().item())

                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
        self.text_encoder = self.text_encoder.to(dtype=T5_PRECISION)
        for block in self.text_encoder.encoder.block:
            block.layer[-1].DenseReluDense.wo.to(dtype=torch.float32)
        
        return self