Upload 3 files

bria_utils.py

@@ -0,0 +1,71 @@
+from typing import Union, Optional, List
+import torch
+from diffusers.utils import logging
+from transformers import (
+    T5EncoderModel,
+    T5TokenizerFast,
+)
+import numpy as np
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+def get_t5_prompt_embeds(
+    tokenizer: T5TokenizerFast ,
+    text_encoder: T5EncoderModel,
+    prompt: Union[str, List[str]] = None,
+    num_images_per_prompt: int = 1,
+    max_sequence_length: int = 128,
+    device: Optional[torch.device] = None,
+):
+    device = device or text_encoder.device
+    
+    prompt = [prompt] if isinstance(prompt, str) else prompt
+    batch_size = len(prompt)
+
+    text_inputs = tokenizer(
+        prompt,
+        # padding="max_length",
+        max_length=max_sequence_length,
+        truncation=True,
+        add_special_tokens=True,
+        return_tensors="pt",
+    )
+    text_input_ids = text_inputs.input_ids
+    untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+    if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+        removed_text = tokenizer.batch_decode(untruncated_ids[:, 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<max_sequence_length:
+        padding = torch.zeros((b,max_sequence_length-seq_len,dim),dtype=prompt_embeds.dtype,device=prompt_embeds.device)
+        prompt_embeds = torch.concat([prompt_embeds,padding],dim=1)
+
+    prompt_embeds = prompt_embeds.to(device=device)
+
+    _, seq_len, _ = prompt_embeds.shape
+
+    # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+    prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+    prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+    return prompt_embeds
+
+# in order the get the same sigmas as in training and sample from them
+def get_original_sigmas(num_train_timesteps=1000,num_inference_steps=1000):
+    timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32)[::-1].copy()
+    sigmas = timesteps / num_train_timesteps
+
+    inds = [int(ind) for ind in  np.linspace(0, num_train_timesteps-1, num_inference_steps)]
+    new_sigmas = sigmas[inds]
+    return new_sigmas
+
+def is_ng_none(negative_prompt):
+    return negative_prompt is None  or negative_prompt=='' or (isinstance(negative_prompt,list) and negative_prompt[0] is None) or (type(negative_prompt)==list and negative_prompt[0]=='')

pipeline_bria.py

@@ -0,0 +1,562 @@
+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
+
+
+
+

transformer_bria.py

@@ -0,0 +1,335 @@
+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 diffusers.models.transformers.transformer_flux import FluxSingleTransformerBlock, FluxTransformerBlock
+
+# Support different diffusers versions
+try:
+    from diffusers.models.embeddings import FluxPosEmbed as EmbedND
+except:
+    from diffusers.models.transformers.transformer_flux import rope
+    class EmbedND(nn.Module):
+        def __init__(self, theta: int, axes_dim: List[int]):
+            super().__init__()
+            self.theta = theta
+            self.axes_dim = axes_dim
+
+        def forward(self, ids: torch.Tensor) -> torch.Tensor:
+            n_axes = ids.shape[-1]
+            emb = torch.cat(
+                [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
+                dim=-3,
+            )
+            return emb.unsqueeze(1)
+
+
+
+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,max_period=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.max_period=max_period
+
+    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.max_period
+        )
+        return t_emb
+    
+class TimestepProjEmbeddings(nn.Module):
+    def __init__(self, embedding_dim, max_period):
+        super().__init__()
+
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0,max_period=max_period)
+        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,
+        max_period = 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,max_period=max_period
+        )
+
+        # 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)