diff --git a/src/diffusers/__init__.py b/src/diffusers/__init__.py index f51a4ef2b3f6..453d6039c367 100644 --- a/src/diffusers/__init__.py +++ b/src/diffusers/__init__.py @@ -360,6 +360,7 @@ "EasyAnimateControlPipeline", "EasyAnimateInpaintPipeline", "EasyAnimatePipeline", + "Flex2Pipeline", "FluxControlImg2ImgPipeline", "FluxControlInpaintPipeline", "FluxControlNetImg2ImgPipeline", @@ -934,6 +935,7 @@ EasyAnimateControlPipeline, EasyAnimateInpaintPipeline, EasyAnimatePipeline, + Flex2Pipeline, FluxControlImg2ImgPipeline, FluxControlInpaintPipeline, FluxControlNetImg2ImgPipeline, diff --git a/src/diffusers/pipelines/__init__.py b/src/diffusers/pipelines/__init__.py index 011f23ed371c..b8fc1cd5bb6d 100644 --- a/src/diffusers/pipelines/__init__.py +++ b/src/diffusers/pipelines/__init__.py @@ -127,6 +127,11 @@ "AnimateDiffVideoToVideoPipeline", "AnimateDiffVideoToVideoControlNetPipeline", ] + _import_structure["flex"] = [ + "Flex2Pipeline", + "Flex2PipelineOutput", + ] + _import_structure["flux"] = [ "FluxControlPipeline", "FluxControlInpaintPipeline", @@ -572,6 +577,7 @@ EasyAnimateInpaintPipeline, EasyAnimatePipeline, ) + from .flex import Flex2Pipeline, Flex2PipelineOutput from .flux import ( FluxControlImg2ImgPipeline, FluxControlInpaintPipeline, diff --git a/src/diffusers/pipelines/flex/__init__.py b/src/diffusers/pipelines/flex/__init__.py new file mode 100644 index 000000000000..a44932c806e9 --- /dev/null +++ b/src/diffusers/pipelines/flex/__init__.py @@ -0,0 +1,23 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + _LazyModule, +) + +_import_structure = {"pipeline_flex": ["Flex2Pipeline"], + "pipeline_output": ["Flex2PipelineOutput"]} + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + from .pipeline_flex import Flex2Pipeline + from .pipeline_output import Flex2PipelineOutput + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) diff --git a/src/diffusers/pipelines/flex/pipeline_flex.py b/src/diffusers/pipelines/flex/pipeline_flex.py new file mode 100644 index 000000000000..54ae40d9909a --- /dev/null +++ b/src/diffusers/pipelines/flex/pipeline_flex.py @@ -0,0 +1,1033 @@ +# Copyright 2025 victolee0, The HuggingFace Team and The ostris Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable, Dict, List, Optional, Tuple, Union + +import numpy as np +import torch +import torch.nn.functional as F +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTokenizer, + CLIPVisionModelWithProjection, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FluxIPAdapterMixin, FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.controlnets.controlnet_flux import FluxControlNetModel, FluxMultiControlNetModel +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import Flex2PipelineOutput + + +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 Flex2Pipeline + >>> from diffusers.utils import load_image + + >>> name_or_path = "ostris/Flex.2-preview" + + >>> inpaint_image = load_image("https://ostris.com/wp-content/uploads/2025/04/dog.jpg") + >>> inpaint_mask = load_image("https://ostris.com/wp-content/uploads/2025/04/dog_mask.jpg") + >>> control_image = load_image("https://ostris.com/wp-content/uploads/2025/04/dog_depth.jpg") + + >>> dtype = torch.bfloat16 + + >>> pipe = Flex2Pipeline.from_pretrained( + ... name_or_path, + ... torch_dtype=dtype + ... ).to("cuda") + + >>> image = pipe( + ... prompt="A white friendly robotic dog sitting on a bench", + ... inpaint_image=inpaint_image, + ... inpaint_mask=inpaint_mask, + ... control_image=control_image, + ... control_strength=0.5, + ... control_stop=0.33, + ... height=1024, + ... width=1024, + ... guidance_scale=3.5, + ... num_inference_steps=50, + ... generator=torch.Generator("cpu").manual_seed(42) + ... ).images[0] + >>> image.save(f"robot_dog.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + sigmas: Optional[List[float]] = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`List[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class Flex2Pipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin, FluxIPAdapterMixin): + r""" + The Flex pipeline for text-to-image generation. + + Reference: https://huggingface.co/ostris/Flex.2-preview + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + def _get_t5_prompt_embeds( + self, + prompt: Union[str, List[str]] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: Optional[torch.device] = None, + dtype: Optional[torch.dtype] = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, 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 + + def _get_clip_prompt_embeds( + self, + prompt: Union[str, List[str]], + num_images_per_prompt: int = 1, + device: Optional[torch.device] = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + def encode_prompt( + self, + prompt: Union[str, List[str]], + prompt_2: Union[str, List[str]], + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + max_sequence_length: int = 512, + lora_scale: Optional[float] = None, + ): + r""" + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + 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) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # We only use the pooled prompt output from the CLIPTextModel + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt_2, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + 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) + + if self.text_encoder_2 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_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + return image_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt + ): + image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_ip_adapter_image in ip_adapter_image: + single_image_embeds = self.encode_image(single_ip_adapter_image, device, 1) + image_embeds.append(single_image_embeds[None, :]) + else: + if not isinstance(ip_adapter_image_embeds, list): + ip_adapter_image_embeds = [ip_adapter_image_embeds] + + if len(ip_adapter_image_embeds) != self.transformer.encoder_hid_proj.num_ip_adapters: + raise ValueError( + f"`ip_adapter_image_embeds` must have same length as the number of IP Adapters. Got {len(ip_adapter_image_embeds)} image embeds and {self.transformer.encoder_hid_proj.num_ip_adapters} IP Adapters." + ) + + for single_image_embeds in ip_adapter_image_embeds: + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for single_image_embeds in image_embeds: + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + prompt_embeds=None, + pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + inpaint_image=None, + inpaint_mask=None, + control_image=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if 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}") + + if inpaint_image is not None and inpaint_mask is None: + raise ValueError( + "If `inpaint_image` is passed, `inpaint_mask` must be passed as well. " + "Please make sure to pass both `inpaint_image` and `inpaint_mask`." + ) + if inpaint_mask is not None and inpaint_image is None: + raise ValueError( + "If `inpaint_mask` is passed, `inpaint_image` must be passed as well. " + "Please make sure to pass both `inpaint_image` and `inpaint_mask`." + ) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height, width, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + 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 + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height, width) + + return latents + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_latents + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is not None: + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + return latents.to(device=device, dtype=dtype), latent_image_ids + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype) + + return latents, latent_image_ids + + # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + inpaint_image: Optional[PipelineImageInput] = None, + inpaint_mask: Optional[PipelineImageInput] = None, + control_image: Optional[PipelineImageInput] = None, + control_strength: Optional[float] = 1.0, + control_stop: Optional[float] = 1.0, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 28, + sigmas: Optional[List[float]] = None, + guidance_scale: float = 3.5, + num_images_per_prompt: Optional[int] = 1, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + inpaint_image (`PipelineImageInput`, *optional*): + The image or images to inpaint. + inpaint_mask (`PipelineImageInput`, *optional*): + A black and white mask to be used for inpainting. The white pixels are the areas to be inpainted, while the + black pixels are the areas to be kept. + control_image (`PipelineImageInput`, *optional*): + The control image (line, depth, pose, etc.) to be used for the generation. The control image + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + 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.flex.Flex2PipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flex.Flex2PipelineOutput`] or `tuple`: [`~pipelines.fl2x.Flex2PipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Prepare text embeddings + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + ( + prompt_embeds, + pooled_prompt_embeds, + text_ids, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 3. Prepare control image + num_channels_latents = self.transformer.config.in_channels // (4 * 3) + + num_control_channels = 2 * num_channels_latents + 1 # 16 + 1 + 16 + # num_channels_latents = num_channels_latents - num_control_channels + + control_latents = None + inpaint_latents = None + inpaint_latents_mask = None + + latent_height = 2 * (int(height) // (self.vae_scale_factor * 2)) + latent_width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + if control_image is None: + control_latents = torch.zeros( + batch_size * num_images_per_prompt, + num_channels_latents, + latent_height, + latent_width, + device=device, + dtype=self.vae.dtype, + ) + else: + control_image = self.prepare_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + control_image = self.vae.encode(control_image).latent_dist.sample(generator=generator) + control_latents = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + control_latents = control_latents * control_strength + + if inpaint_image is None and inpaint_mask is None: + inpaint_latents = torch.zeros( + batch_size * num_images_per_prompt, + num_channels_latents, + latent_height, + latent_width, + device=device, + dtype=self.vae.dtype, + ) + inpaint_latents_mask = torch.ones( + batch_size * num_images_per_prompt, + 1, + latent_height, + latent_width, + device=device, + dtype=self.vae.dtype, + ) + else: + inpaint_image = self.prepare_image( + image=inpaint_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + inpaint_image = self.vae.encode(inpaint_image).latent_dist.sample(generator=generator) + inpaint_latents = (inpaint_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor + height_inpaint_image, width_inpaint_image = control_image.shape[2:] + + inpaint_mask = self.prepare_image( + image=inpaint_mask, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + inpaint_mask = inpaint_mask[:, 0:1, :, :] * 0.5 + 0.5 + # resize to match height_inpaint_image and width_inpaint_image + inpaint_latents_mask = F.interpolate(inpaint_mask, size=(height_inpaint_image, width_inpaint_image), mode="bilinear", align_corners=False) + + # apply inverted mask to inpaint latents + inpaint_latents = inpaint_latents * (1 - inpaint_latents_mask) + + latent_controls = torch.cat([inpaint_latents, inpaint_latents_mask, control_latents], dim=1) + latent_no_controls = torch.cat([inpaint_latents, inpaint_latents_mask, torch.zeros_like(control_latents)], dim=1) + + height_latent_controls, width_latent_controls = latent_controls.shape[2:] + packed_latent_controls = self._pack_latents( + latent_controls, + batch_size * num_images_per_prompt, + num_control_channels, + height_latent_controls, + width_latent_controls, + ) + + packed_latent_no_controls = self._pack_latents( + latent_no_controls, + batch_size * num_images_per_prompt, + num_control_channels, + height_latent_controls, + width_latent_controls, + ) + + # 4. Prepare latent variables + latents, latent_image_ids = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + control_cutoff = int(len(timesteps) * control_stop) + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + control_latents = packed_latent_controls if i < control_cutoff else packed_latent_no_controls + latent_model_input = torch.cat([latents, control_latents], dim=2) + + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + guidance = ( + torch.tensor([guidance_scale], device=device) if self.transformer.config.guidance_embeds else None + ) + guidance = guidance.expand(latents.shape[0]) if guidance is not None else None + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if output_type == "latent": + image = latents + + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return Flex2PipelineOutput(images=image) diff --git a/src/diffusers/pipelines/flex/pipeline_output.py b/src/diffusers/pipelines/flex/pipeline_output.py new file mode 100644 index 000000000000..9a6ce946407d --- /dev/null +++ b/src/diffusers/pipelines/flex/pipeline_output.py @@ -0,0 +1,22 @@ +from dataclasses import dataclass +from typing import List, Union + +import numpy as np +import PIL.Image +import torch + +from ...utils import BaseOutput + + +@dataclass +class Flex2PipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: Union[List[PIL.Image.Image], np.ndarray] diff --git a/tests/pipelines/flex/__init__.py b/tests/pipelines/flex/__init__.py new file mode 100644 index 000000000000..e69de29bb2d1 diff --git a/tests/pipelines/flex/test_pipeline_flex.py b/tests/pipelines/flex/test_pipeline_flex.py new file mode 100644 index 000000000000..f48487f0e069 --- /dev/null +++ b/tests/pipelines/flex/test_pipeline_flex.py @@ -0,0 +1,297 @@ +import gc +import unittest + +import numpy as np +import pytest +import torch +from huggingface_hub import hf_hub_download +from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel + +from diffusers import ( + AutoencoderKL, + FasterCacheConfig, + Flex2Pipeline, + FlowMatchEulerDiscreteScheduler, + FluxTransformer2DModel, +) +from diffusers.utils.testing_utils import ( + backend_empty_cache, + nightly, + numpy_cosine_similarity_distance, + require_big_accelerator, + slow, + torch_device, +) + +from ..test_pipelines_common import ( + FasterCacheTesterMixin, + PipelineTesterMixin, + PyramidAttentionBroadcastTesterMixin, + check_qkv_fusion_matches_attn_procs_length, + check_qkv_fusion_processors_exist, +) + + +class Flex2PipelineFastTests( + unittest.TestCase, + PipelineTesterMixin, + PyramidAttentionBroadcastTesterMixin, + FasterCacheTesterMixin, +): + pipeline_class = Flex2Pipeline + params = frozenset(["prompt", "height", "width", "guidance_scale", "prompt_embeds", "pooled_prompt_embeds"]) + batch_params = frozenset(["prompt"]) + + # there is no xformers processor for Flex2 + test_xformers_attention = False + test_layerwise_casting = True + test_group_offloading = True + + faster_cache_config = FasterCacheConfig( + spatial_attention_block_skip_range=2, + spatial_attention_timestep_skip_range=(-1, 901), + unconditional_batch_skip_range=2, + attention_weight_callback=lambda _: 0.5, + is_guidance_distilled=True, + ) + + def get_dummy_components(self, num_layers: int = 1, num_single_layers: int = 1): + torch.manual_seed(0) + transformer = FluxTransformer2DModel( + patch_size=1, + in_channels=196, + out_channels=64, + num_layers=num_layers, + num_single_layers=num_single_layers, + attention_head_dim=16, + num_attention_heads=2, + joint_attention_dim=32, + pooled_projection_dim=32, + axes_dims_rope=[4, 4, 8], + ) + clip_text_encoder_config = CLIPTextConfig( + bos_token_id=0, + eos_token_id=2, + hidden_size=32, + intermediate_size=37, + layer_norm_eps=1e-05, + num_attention_heads=4, + num_hidden_layers=5, + pad_token_id=1, + vocab_size=1000, + hidden_act="gelu", + projection_dim=32, + ) + + torch.manual_seed(0) + text_encoder = CLIPTextModel(clip_text_encoder_config) + + torch.manual_seed(0) + text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5") + + tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip") + tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5") + + torch.manual_seed(0) + vae = AutoencoderKL( + sample_size=32, + in_channels=3, + out_channels=3, + block_out_channels=(4,), + layers_per_block=1, + latent_channels=16, + norm_num_groups=1, + use_quant_conv=False, + use_post_quant_conv=False, + shift_factor=0.0609, + scaling_factor=1.5035, + ) + + scheduler = FlowMatchEulerDiscreteScheduler() + + return { + "scheduler": scheduler, + "text_encoder": text_encoder, + "text_encoder_2": text_encoder_2, + "tokenizer": tokenizer, + "tokenizer_2": tokenizer_2, + "transformer": transformer, + "vae": vae, + } + + def get_dummy_inputs(self, device, seed=0): + if str(device).startswith("mps"): + generator = torch.manual_seed(seed) + else: + generator = torch.Generator(device="cpu").manual_seed(seed) + + inputs = { + "prompt": "A painting of a squirrel eating a burger", + "generator": generator, + "num_inference_steps": 2, + "guidance_scale": 5.0, + "height": 8, + "width": 8, + "max_sequence_length": 48, + "output_type": "np", + } + return inputs + + def test_flex2_different_prompts(self): + pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device) + + inputs = self.get_dummy_inputs(torch_device) + output_same_prompt = pipe(**inputs).images[0] + + inputs = self.get_dummy_inputs(torch_device) + inputs["prompt_2"] = "a different prompt" + output_different_prompts = pipe(**inputs).images[0] + + max_diff = np.abs(output_same_prompt - output_different_prompts).max() + + # Outputs should be different here + # For some reasons, they don't show large differences + assert max_diff > 1e-6 + + def test_fused_qkv_projections(self): + device = "cpu" # ensure determinism for the device-dependent torch.Generator + components = self.get_dummy_components() + pipe = self.pipeline_class(**components) + pipe = pipe.to(device) + pipe.set_progress_bar_config(disable=None) + + inputs = self.get_dummy_inputs(device) + image = pipe(**inputs).images + original_image_slice = image[0, -3:, -3:, -1] + + # TODO (sayakpaul): will refactor this once `fuse_qkv_projections()` has been added + # to the pipeline level. + pipe.transformer.fuse_qkv_projections() + assert check_qkv_fusion_processors_exist(pipe.transformer), ( + "Something wrong with the fused attention processors. Expected all the attention processors to be fused." + ) + assert check_qkv_fusion_matches_attn_procs_length( + pipe.transformer, pipe.transformer.original_attn_processors + ), "Something wrong with the attention processors concerning the fused QKV projections." + + inputs = self.get_dummy_inputs(device) + image = pipe(**inputs).images + image_slice_fused = image[0, -3:, -3:, -1] + + pipe.transformer.unfuse_qkv_projections() + inputs = self.get_dummy_inputs(device) + image = pipe(**inputs).images + image_slice_disabled = image[0, -3:, -3:, -1] + + assert np.allclose(original_image_slice, image_slice_fused, atol=1e-3, rtol=1e-3), ( + "Fusion of QKV projections shouldn't affect the outputs." + ) + assert np.allclose(image_slice_fused, image_slice_disabled, atol=1e-3, rtol=1e-3), ( + "Outputs, with QKV projection fusion enabled, shouldn't change when fused QKV projections are disabled." + ) + assert np.allclose(original_image_slice, image_slice_disabled, atol=1e-2, rtol=1e-2), ( + "Original outputs should match when fused QKV projections are disabled." + ) + + def test_flex2_image_output_shape(self): + pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device) + inputs = self.get_dummy_inputs(torch_device) + + height_width_pairs = [(32, 32), (72, 57)] + for height, width in height_width_pairs: + expected_height = height - height % (pipe.vae_scale_factor * 2) + expected_width = width - width % (pipe.vae_scale_factor * 2) + + inputs.update({"height": height, "width": width}) + image = pipe(**inputs).images[0] + output_height, output_width, _ = image.shape + assert (output_height, output_width) == (expected_height, expected_width) + + + +@nightly +@require_big_accelerator +@pytest.mark.big_gpu_with_torch_cuda +class Flex2PipelineSlowTests(unittest.TestCase): + pipeline_class = Flex2Pipeline + repo_id = "ostris/Flex.2-preview" + + def setUp(self): + super().setUp() + gc.collect() + backend_empty_cache(torch_device) + + def tearDown(self): + super().tearDown() + gc.collect() + backend_empty_cache(torch_device) + + def get_inputs(self, device, seed=0): + generator = torch.Generator(device="cpu").manual_seed(seed) + + prompt_embeds = torch.load( + hf_hub_download(repo_id="diffusers/test-slices", repo_type="dataset", filename="flux/prompt_embeds.pt") + ).to(torch_device) + pooled_prompt_embeds = torch.load( + hf_hub_download( + repo_id="diffusers/test-slices", repo_type="dataset", filename="flux/pooled_prompt_embeds.pt" + ) + ).to(torch_device) + return { + "prompt_embeds": prompt_embeds, + "pooled_prompt_embeds": pooled_prompt_embeds, + "num_inference_steps": 2, + "guidance_scale": 0.0, + "max_sequence_length": 256, + "output_type": "np", + "generator": generator, + } + + def test_flex2_inference(self): + pipe = self.pipeline_class.from_pretrained( + self.repo_id, torch_dtype=torch.bfloat16, text_encoder=None, text_encoder_2=None + ).to(torch_device) + + inputs = self.get_inputs(torch_device) + + image = pipe(**inputs).images[0] + image_slice = image[0, :10, :10] + expected_slice = np.array( + [ + 0.3242, + 0.3203, + 0.3164, + 0.3164, + 0.3125, + 0.3125, + 0.3281, + 0.3242, + 0.3203, + 0.3301, + 0.3262, + 0.3242, + 0.3281, + 0.3242, + 0.3203, + 0.3262, + 0.3262, + 0.3164, + 0.3262, + 0.3281, + 0.3184, + 0.3281, + 0.3281, + 0.3203, + 0.3281, + 0.3281, + 0.3164, + 0.3320, + 0.3320, + 0.3203, + ], + dtype=np.float32, + ) + + max_diff = numpy_cosine_similarity_distance(expected_slice.flatten(), image_slice.flatten()) + + assert max_diff < 1e-4 \ No newline at end of file