From f23eb1979f0e018acca68487bccc8c04a32963f2 Mon Sep 17 00:00:00 2001 From: Sayak Paul Date: Tue, 6 May 2025 18:52:18 +0530 Subject: [PATCH 01/10] Add FluxPAGPipeline with support for PAG --- docs/source/en/quantization/bitsandbytes.md | 12 +- examples/community/pipeline_flux_with_pag.py | 1095 ++++++++++++++++++ 2 files changed, 1104 insertions(+), 3 deletions(-) create mode 100644 examples/community/pipeline_flux_with_pag.py diff --git a/docs/source/en/quantization/bitsandbytes.md b/docs/source/en/quantization/bitsandbytes.md index 744351c9b15e..b1c130b792c3 100644 --- a/docs/source/en/quantization/bitsandbytes.md +++ b/docs/source/en/quantization/bitsandbytes.md @@ -48,7 +48,7 @@ For Ada and higher-series GPUs. we recommend changing `torch_dtype` to `torch.bf ```py from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig - +import torch from diffusers import AutoModel from transformers import T5EncoderModel @@ -88,6 +88,8 @@ Setting `device_map="auto"` automatically fills all available space on the GPU(s CPU, and finally, the hard drive (the absolute slowest option) if there is still not enough memory. ```py +from diffusers import FluxPipeline + pipe = FluxPipeline.from_pretrained( "black-forest-labs/FLUX.1-dev", transformer=transformer_8bit, @@ -132,7 +134,7 @@ For Ada and higher-series GPUs. we recommend changing `torch_dtype` to `torch.bf ```py from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig - +import torch from diffusers import AutoModel from transformers import T5EncoderModel @@ -171,6 +173,8 @@ Let's generate an image using our quantized models. Setting `device_map="auto"` automatically fills all available space on the GPU(s) first, then the CPU, and finally, the hard drive (the absolute slowest option) if there is still not enough memory. ```py +from diffusers import FluxPipeline + pipe = FluxPipeline.from_pretrained( "black-forest-labs/FLUX.1-dev", transformer=transformer_4bit, @@ -214,6 +218,8 @@ Check your memory footprint with the `get_memory_footprint` method: print(model.get_memory_footprint()) ``` +Note that this only tells you the memory footprint of the model params and does _not_ estimate the inference memory requirements. + Quantized models can be loaded from the [`~ModelMixin.from_pretrained`] method without needing to specify the `quantization_config` parameters: ```py @@ -413,4 +419,4 @@ transformer_4bit.dequantize() ## Resources * [End-to-end notebook showing Flux.1 Dev inference in a free-tier Colab](https://gist.github.com/sayakpaul/c76bd845b48759e11687ac550b99d8b4) -* [Training](https://gist.github.com/sayakpaul/05afd428bc089b47af7c016e42004527) \ No newline at end of file +* [Training](https://github.com/huggingface/diffusers/blob/8c661ea586bf11cb2440da740dd3c4cf84679b85/examples/dreambooth/README_hidream.md#using-quantization) \ No newline at end of file diff --git a/examples/community/pipeline_flux_with_pag.py b/examples/community/pipeline_flux_with_pag.py new file mode 100644 index 000000000000..97c17c2e898a --- /dev/null +++ b/examples/community/pipeline_flux_with_pag.py @@ -0,0 +1,1095 @@ +# Copyright 2024 Black Forest Labs and The HuggingFace 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, Union + +import numpy as np +import torch +import torch.nn.functional as F +from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast + +from diffusers.image_processor import VaeImageProcessor +from diffusers.loaders import FluxLoraLoaderMixin, FromSingleFileMixin, FluxLoraLoaderMixin +from diffusers.models.autoencoders import AutoencoderKL +from diffusers.models.attention_processor import Attention +from diffusers.models.transformers import FluxTransformer2DModel +from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.schedulers import FlowMatchEulerDiscreteScheduler +from diffusers.utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from diffusers.utils.torch_utils import randn_tensor + + +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 FluxPipeline + + >>> pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe(prompt, num_inference_steps=4, guidance_scale=0.0).images[0] + >>> image.save("flux.png") + ``` +""" + + +# 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.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, +): + """ + 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 PAGIdentitySelfAttnProcessor: + r""" + Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). + """ + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + + def __call__( + self, + attn: Attention, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + temb: Optional[torch.Tensor] = None, + *args, + **kwargs, + ) -> torch.Tensor: + if len(args) > 0 or kwargs.get("scale", None) is not None: + deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." + deprecate("scale", "1.0.0", deprecation_message) + + residual = hidden_states + if attn.spatial_norm is not None: + hidden_states = attn.spatial_norm(hidden_states, temb) + + input_ndim = hidden_states.ndim + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + + # chunk + hidden_states_org, hidden_states_ptb = hidden_states.chunk(2) + + # original path + batch_size, sequence_length, _ = hidden_states_org.shape + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + if attn.group_norm is not None: + hidden_states_org = attn.group_norm(hidden_states_org.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states_org) + key = attn.to_k(hidden_states_org) + value = attn.to_v(hidden_states_org) + + inner_dim = key.shape[-1] + head_dim = inner_dim // attn.heads + + query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + # the output of sdp = (batch, num_heads, seq_len, head_dim) + # TODO: add support for attn.scale when we move to Torch 2.1 + hidden_states_org = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) + + hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) + hidden_states_org = hidden_states_org.to(query.dtype) + + # linear proj + hidden_states_org = attn.to_out[0](hidden_states_org) + # dropout + hidden_states_org = attn.to_out[1](hidden_states_org) + + if input_ndim == 4: + hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width) + + # perturbed path (identity attention) + batch_size, sequence_length, _ = hidden_states_ptb.shape + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + if attn.group_norm is not None: + hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2) + + value = attn.to_v(hidden_states_ptb) + + # hidden_states_ptb = torch.zeros(value.shape).to(value.get_device()) + hidden_states_ptb = value + + hidden_states_ptb = hidden_states_ptb.to(query.dtype) + + # linear proj + hidden_states_ptb = attn.to_out[0](hidden_states_ptb) + # dropout + hidden_states_ptb = attn.to_out[1](hidden_states_ptb) + + if input_ndim == 4: + hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width) + + # cat + hidden_states = torch.cat([hidden_states_org, hidden_states_ptb]) + + if attn.residual_connection: + hidden_states = hidden_states + residual + + hidden_states = hidden_states / attn.rescale_output_factor + + return hidden_states + + +class PAGCFGIdentitySelfAttnProcessor: + r""" + Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). + """ + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + + def __call__( + self, + attn: Attention, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + temb: Optional[torch.Tensor] = None, + *args, + **kwargs, + ) -> torch.Tensor: + if len(args) > 0 or kwargs.get("scale", None) is not None: + deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." + deprecate("scale", "1.0.0", deprecation_message) + + residual = hidden_states + if attn.spatial_norm is not None: + hidden_states = attn.spatial_norm(hidden_states, temb) + + input_ndim = hidden_states.ndim + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + + # chunk + hidden_states_uncond, hidden_states_org, hidden_states_ptb = hidden_states.chunk(3) + hidden_states_org = torch.cat([hidden_states_uncond, hidden_states_org]) + + # original path + batch_size, sequence_length, _ = hidden_states_org.shape + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + if attn.group_norm is not None: + hidden_states_org = attn.group_norm(hidden_states_org.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states_org) + key = attn.to_k(hidden_states_org) + value = attn.to_v(hidden_states_org) + + inner_dim = key.shape[-1] + head_dim = inner_dim // attn.heads + + query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + # the output of sdp = (batch, num_heads, seq_len, head_dim) + # TODO: add support for attn.scale when we move to Torch 2.1 + hidden_states_org = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) + + hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) + hidden_states_org = hidden_states_org.to(query.dtype) + + # linear proj + hidden_states_org = attn.to_out[0](hidden_states_org) + # dropout + hidden_states_org = attn.to_out[1](hidden_states_org) + + if input_ndim == 4: + hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width) + + # perturbed path (identity attention) + batch_size, sequence_length, _ = hidden_states_ptb.shape + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + if attn.group_norm is not None: + hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2) + + value = attn.to_v(hidden_states_ptb) + hidden_states_ptb = value + hidden_states_ptb = hidden_states_ptb.to(query.dtype) + + # linear proj + hidden_states_ptb = attn.to_out[0](hidden_states_ptb) + # dropout + hidden_states_ptb = attn.to_out[1](hidden_states_ptb) + + if input_ndim == 4: + hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width) + + # cat + hidden_states = torch.cat([hidden_states_org, hidden_states_ptb]) + + if attn.residual_connection: + hidden_states = hidden_states + residual + + hidden_states = hidden_states / attn.rescale_output_factor + + return hidden_states + + +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + """ + Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + **kwargs, +): + """ + 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 support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps` + 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: + 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) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class FluxPAGPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin): + r""" + The Flux pipeline for text-to-image generation. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->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)) if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + 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 = 64 + + 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) + + 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) + + 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]], + prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + lora_scale: Optional[float] = None, + max_sequence_length: int = 512, + ): + device = device or self._execution_device + + # Set LoRA scale + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + if self.text_encoder and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + if prompt_embeds is None: + # Normalize inputs + if isinstance(prompt, str): + prompt = [prompt] + if isinstance(prompt_2, str): + prompt_2 = [prompt_2] + batch_size = len(prompt) + + # Duplicate each prompt 3 times: [uncond, cond, perturbed] + full_prompt = [] + full_prompt_2 = [] + for i in range(batch_size): + full_prompt.extend(["", prompt[i], prompt[i]]) # For CLIP + full_prompt_2.extend(["", prompt_2[i], prompt_2[i]]) # For T5 + + # Encode prompts + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=full_prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=full_prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + # Unscale LoRA + if self.text_encoder and isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + unscale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 and isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + unscale_lora_layers(self.text_encoder_2, lora_scale) + + # Dummy text_ids (for compatibility) + dtype = self.text_encoder.dtype if self.text_encoder 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 + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % 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_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @staticmethod + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height // 2, width // 2, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height // 2)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width // 2)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + height = height // vae_scale_factor + width = width // vae_scale_factor + + latents = latents.view(batch_size, height, width, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height * 2, width * 2) + + return latents + + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + height = 2 * (int(height) // self.vae_scale_factor) + width = 2 * (int(width) // self.vae_scale_factor) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is not None: + latent_image_ids = self._prepare_latent_image_ids(batch_size, height, width, 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, width, device, dtype) + + return latents, latent_image_ids + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt: Union[str, List[str]] = None, # + negative_prompt_2: Optional[Union[str, List[str]]] = None, + true_pag: float = 1.0, # + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 28, + timesteps: List[int] = 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, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + 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 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.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._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 + ) + do_true_pag = true_pag > 1 and negative_prompt is not 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, + ) + + if do_true_pag: + # For PAGCFG: [uncond, cond, perturbed] + prompt_embeds = torch.cat([prompt_embeds], dim=0) # already 3x in encode_prompt + else: + # For PAG (only 2x needed: [cond, perturbed]) + prompt_embeds = torch.cat([prompt_embeds, prompt_embeds], dim=0) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, latent_image_ids = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) + 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, + timesteps, + sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + if true_pag > 0: + for name, module in self.transformer.named_modules(): + if isinstance(module, Attention): + module.processor = PAGCFGIdentitySelfAttnProcessor() if negative_prompt else PAGIdentitySelfAttnProcessor() + + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + if do_true_pag: + latent_model_input = torch.cat([latents] * 3) + else: + latent_model_input = torch.cat([latents] * 2) + + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latent_model_input.shape[0]) + else: + guidance = None + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]).to(latent_model_input.dtype) + + 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] + + if do_true_pag: + cond_pred, perturbed_pred = noise_pred.chunk(2) + noise_pred = cond_pred + true_pag * (perturbed_pred - cond_pred) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) From 50a426e817bade849c700f22eb78893807f5aaf8 Mon Sep 17 00:00:00 2001 From: tongyu Date: Wed, 7 May 2025 10:37:24 +0800 Subject: [PATCH 02/10] Update pipeline_flux_with_pag.py --- examples/community/pipeline_flux_with_pag.py | 124 +------------------ 1 file changed, 6 insertions(+), 118 deletions(-) diff --git a/examples/community/pipeline_flux_with_pag.py b/examples/community/pipeline_flux_with_pag.py index 97c17c2e898a..209a3ad02426 100644 --- a/examples/community/pipeline_flux_with_pag.py +++ b/examples/community/pipeline_flux_with_pag.py @@ -253,118 +253,6 @@ def __call__( return hidden_states -class PAGCFGIdentitySelfAttnProcessor: - r""" - Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). - """ - - def __init__(self): - if not hasattr(F, "scaled_dot_product_attention"): - raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") - - def __call__( - self, - attn: Attention, - hidden_states: torch.Tensor, - encoder_hidden_states: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - temb: Optional[torch.Tensor] = None, - *args, - **kwargs, - ) -> torch.Tensor: - if len(args) > 0 or kwargs.get("scale", None) is not None: - deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." - deprecate("scale", "1.0.0", deprecation_message) - - residual = hidden_states - if attn.spatial_norm is not None: - hidden_states = attn.spatial_norm(hidden_states, temb) - - input_ndim = hidden_states.ndim - if input_ndim == 4: - batch_size, channel, height, width = hidden_states.shape - hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) - - # chunk - hidden_states_uncond, hidden_states_org, hidden_states_ptb = hidden_states.chunk(3) - hidden_states_org = torch.cat([hidden_states_uncond, hidden_states_org]) - - # original path - batch_size, sequence_length, _ = hidden_states_org.shape - - if attention_mask is not None: - attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) - # scaled_dot_product_attention expects attention_mask shape to be - # (batch, heads, source_length, target_length) - attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) - - if attn.group_norm is not None: - hidden_states_org = attn.group_norm(hidden_states_org.transpose(1, 2)).transpose(1, 2) - - query = attn.to_q(hidden_states_org) - key = attn.to_k(hidden_states_org) - value = attn.to_v(hidden_states_org) - - inner_dim = key.shape[-1] - head_dim = inner_dim // attn.heads - - query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) - - key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) - value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) - - # the output of sdp = (batch, num_heads, seq_len, head_dim) - # TODO: add support for attn.scale when we move to Torch 2.1 - hidden_states_org = F.scaled_dot_product_attention( - query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False - ) - - hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) - hidden_states_org = hidden_states_org.to(query.dtype) - - # linear proj - hidden_states_org = attn.to_out[0](hidden_states_org) - # dropout - hidden_states_org = attn.to_out[1](hidden_states_org) - - if input_ndim == 4: - hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width) - - # perturbed path (identity attention) - batch_size, sequence_length, _ = hidden_states_ptb.shape - - if attention_mask is not None: - attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) - # scaled_dot_product_attention expects attention_mask shape to be - # (batch, heads, source_length, target_length) - attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) - - if attn.group_norm is not None: - hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2) - - value = attn.to_v(hidden_states_ptb) - hidden_states_ptb = value - hidden_states_ptb = hidden_states_ptb.to(query.dtype) - - # linear proj - hidden_states_ptb = attn.to_out[0](hidden_states_ptb) - # dropout - hidden_states_ptb = attn.to_out[1](hidden_states_ptb) - - if input_ndim == 4: - hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width) - - # cat - hidden_states = torch.cat([hidden_states_org, hidden_states_ptb]) - - if attn.residual_connection: - hidden_states = hidden_states + residual - - hidden_states = hidden_states / attn.rescale_output_factor - - return hidden_states - - def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): """ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and @@ -596,12 +484,12 @@ def encode_prompt( prompt_2 = [prompt_2] batch_size = len(prompt) - # Duplicate each prompt 3 times: [uncond, cond, perturbed] + # Only cond and perturbed (skip uncond) full_prompt = [] full_prompt_2 = [] for i in range(batch_size): - full_prompt.extend(["", prompt[i], prompt[i]]) # For CLIP - full_prompt_2.extend(["", prompt_2[i], prompt_2[i]]) # For T5 + full_prompt.extend([prompt[i], prompt[i]]) # For CLIP + full_prompt_2.extend([prompt_2[i], prompt_2[i]]) # For T5 # Encode prompts pooled_prompt_embeds = self._get_clip_prompt_embeds( @@ -622,9 +510,9 @@ def encode_prompt( if self.text_encoder_2 and isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: unscale_lora_layers(self.text_encoder_2, lora_scale) - # Dummy text_ids (for compatibility) + # Dummy text_ids (updated shape for 2 prompts) dtype = self.text_encoder.dtype if self.text_encoder else self.transformer.dtype - text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + text_ids = torch.zeros(prompt_embeds.shape[1], 2).to(device=device, dtype=dtype) return prompt_embeds, pooled_prompt_embeds, text_ids @@ -1011,7 +899,7 @@ def __call__( if true_pag > 0: for name, module in self.transformer.named_modules(): if isinstance(module, Attention): - module.processor = PAGCFGIdentitySelfAttnProcessor() if negative_prompt else PAGIdentitySelfAttnProcessor() + module.processor = PAGIdentitySelfAttnProcessor() # 6. Denoising loop From 97dfbf36a228d02b9faf2435a4b60d91ac0d241e Mon Sep 17 00:00:00 2001 From: tongyu Date: Wed, 7 May 2025 14:36:32 +0800 Subject: [PATCH 03/10] Update pipeline_flux_with_pag.py --- examples/community/pipeline_flux_with_pag.py | 124 ++++++++++++++++++- 1 file changed, 118 insertions(+), 6 deletions(-) diff --git a/examples/community/pipeline_flux_with_pag.py b/examples/community/pipeline_flux_with_pag.py index 209a3ad02426..97c17c2e898a 100644 --- a/examples/community/pipeline_flux_with_pag.py +++ b/examples/community/pipeline_flux_with_pag.py @@ -253,6 +253,118 @@ def __call__( return hidden_states +class PAGCFGIdentitySelfAttnProcessor: + r""" + Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). + """ + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + + def __call__( + self, + attn: Attention, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + temb: Optional[torch.Tensor] = None, + *args, + **kwargs, + ) -> torch.Tensor: + if len(args) > 0 or kwargs.get("scale", None) is not None: + deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." + deprecate("scale", "1.0.0", deprecation_message) + + residual = hidden_states + if attn.spatial_norm is not None: + hidden_states = attn.spatial_norm(hidden_states, temb) + + input_ndim = hidden_states.ndim + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + + # chunk + hidden_states_uncond, hidden_states_org, hidden_states_ptb = hidden_states.chunk(3) + hidden_states_org = torch.cat([hidden_states_uncond, hidden_states_org]) + + # original path + batch_size, sequence_length, _ = hidden_states_org.shape + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + if attn.group_norm is not None: + hidden_states_org = attn.group_norm(hidden_states_org.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states_org) + key = attn.to_k(hidden_states_org) + value = attn.to_v(hidden_states_org) + + inner_dim = key.shape[-1] + head_dim = inner_dim // attn.heads + + query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + # the output of sdp = (batch, num_heads, seq_len, head_dim) + # TODO: add support for attn.scale when we move to Torch 2.1 + hidden_states_org = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) + + hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) + hidden_states_org = hidden_states_org.to(query.dtype) + + # linear proj + hidden_states_org = attn.to_out[0](hidden_states_org) + # dropout + hidden_states_org = attn.to_out[1](hidden_states_org) + + if input_ndim == 4: + hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width) + + # perturbed path (identity attention) + batch_size, sequence_length, _ = hidden_states_ptb.shape + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + if attn.group_norm is not None: + hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2) + + value = attn.to_v(hidden_states_ptb) + hidden_states_ptb = value + hidden_states_ptb = hidden_states_ptb.to(query.dtype) + + # linear proj + hidden_states_ptb = attn.to_out[0](hidden_states_ptb) + # dropout + hidden_states_ptb = attn.to_out[1](hidden_states_ptb) + + if input_ndim == 4: + hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width) + + # cat + hidden_states = torch.cat([hidden_states_org, hidden_states_ptb]) + + if attn.residual_connection: + hidden_states = hidden_states + residual + + hidden_states = hidden_states / attn.rescale_output_factor + + return hidden_states + + def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): """ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and @@ -484,12 +596,12 @@ def encode_prompt( prompt_2 = [prompt_2] batch_size = len(prompt) - # Only cond and perturbed (skip uncond) + # Duplicate each prompt 3 times: [uncond, cond, perturbed] full_prompt = [] full_prompt_2 = [] for i in range(batch_size): - full_prompt.extend([prompt[i], prompt[i]]) # For CLIP - full_prompt_2.extend([prompt_2[i], prompt_2[i]]) # For T5 + full_prompt.extend(["", prompt[i], prompt[i]]) # For CLIP + full_prompt_2.extend(["", prompt_2[i], prompt_2[i]]) # For T5 # Encode prompts pooled_prompt_embeds = self._get_clip_prompt_embeds( @@ -510,9 +622,9 @@ def encode_prompt( if self.text_encoder_2 and isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: unscale_lora_layers(self.text_encoder_2, lora_scale) - # Dummy text_ids (updated shape for 2 prompts) + # Dummy text_ids (for compatibility) dtype = self.text_encoder.dtype if self.text_encoder else self.transformer.dtype - text_ids = torch.zeros(prompt_embeds.shape[1], 2).to(device=device, dtype=dtype) + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) return prompt_embeds, pooled_prompt_embeds, text_ids @@ -899,7 +1011,7 @@ def __call__( if true_pag > 0: for name, module in self.transformer.named_modules(): if isinstance(module, Attention): - module.processor = PAGIdentitySelfAttnProcessor() + module.processor = PAGCFGIdentitySelfAttnProcessor() if negative_prompt else PAGIdentitySelfAttnProcessor() # 6. Denoising loop From 099e1c42bc70a5afb9823bc99579fd1294ce0cb8 Mon Sep 17 00:00:00 2001 From: tongyu Date: Wed, 7 May 2025 22:21:14 +0800 Subject: [PATCH 04/10] Update pipeline_flux_with_pag.py --- examples/community/pipeline_flux_with_pag.py | 223 ++++++++----------- 1 file changed, 88 insertions(+), 135 deletions(-) diff --git a/examples/community/pipeline_flux_with_pag.py b/examples/community/pipeline_flux_with_pag.py index 97c17c2e898a..ca34801e2737 100644 --- a/examples/community/pipeline_flux_with_pag.py +++ b/examples/community/pipeline_flux_with_pag.py @@ -30,6 +30,7 @@ from diffusers.schedulers import FlowMatchEulerDiscreteScheduler from diffusers.utils import ( USE_PEFT_BACKEND, + deprecate, is_torch_xla_available, logging, replace_example_docstring, @@ -253,118 +254,6 @@ def __call__( return hidden_states -class PAGCFGIdentitySelfAttnProcessor: - r""" - Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). - """ - - def __init__(self): - if not hasattr(F, "scaled_dot_product_attention"): - raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") - - def __call__( - self, - attn: Attention, - hidden_states: torch.Tensor, - encoder_hidden_states: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - temb: Optional[torch.Tensor] = None, - *args, - **kwargs, - ) -> torch.Tensor: - if len(args) > 0 or kwargs.get("scale", None) is not None: - deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." - deprecate("scale", "1.0.0", deprecation_message) - - residual = hidden_states - if attn.spatial_norm is not None: - hidden_states = attn.spatial_norm(hidden_states, temb) - - input_ndim = hidden_states.ndim - if input_ndim == 4: - batch_size, channel, height, width = hidden_states.shape - hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) - - # chunk - hidden_states_uncond, hidden_states_org, hidden_states_ptb = hidden_states.chunk(3) - hidden_states_org = torch.cat([hidden_states_uncond, hidden_states_org]) - - # original path - batch_size, sequence_length, _ = hidden_states_org.shape - - if attention_mask is not None: - attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) - # scaled_dot_product_attention expects attention_mask shape to be - # (batch, heads, source_length, target_length) - attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) - - if attn.group_norm is not None: - hidden_states_org = attn.group_norm(hidden_states_org.transpose(1, 2)).transpose(1, 2) - - query = attn.to_q(hidden_states_org) - key = attn.to_k(hidden_states_org) - value = attn.to_v(hidden_states_org) - - inner_dim = key.shape[-1] - head_dim = inner_dim // attn.heads - - query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) - - key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) - value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) - - # the output of sdp = (batch, num_heads, seq_len, head_dim) - # TODO: add support for attn.scale when we move to Torch 2.1 - hidden_states_org = F.scaled_dot_product_attention( - query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False - ) - - hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) - hidden_states_org = hidden_states_org.to(query.dtype) - - # linear proj - hidden_states_org = attn.to_out[0](hidden_states_org) - # dropout - hidden_states_org = attn.to_out[1](hidden_states_org) - - if input_ndim == 4: - hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width) - - # perturbed path (identity attention) - batch_size, sequence_length, _ = hidden_states_ptb.shape - - if attention_mask is not None: - attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) - # scaled_dot_product_attention expects attention_mask shape to be - # (batch, heads, source_length, target_length) - attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) - - if attn.group_norm is not None: - hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2) - - value = attn.to_v(hidden_states_ptb) - hidden_states_ptb = value - hidden_states_ptb = hidden_states_ptb.to(query.dtype) - - # linear proj - hidden_states_ptb = attn.to_out[0](hidden_states_ptb) - # dropout - hidden_states_ptb = attn.to_out[1](hidden_states_ptb) - - if input_ndim == 4: - hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width) - - # cat - hidden_states = torch.cat([hidden_states_org, hidden_states_ptb]) - - if attn.residual_connection: - hidden_states = hidden_states + residual - - hidden_states = hidden_states / attn.rescale_output_factor - - return hidden_states - - def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): """ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and @@ -596,12 +485,12 @@ def encode_prompt( prompt_2 = [prompt_2] batch_size = len(prompt) - # Duplicate each prompt 3 times: [uncond, cond, perturbed] + # Only cond and perturbed (skip uncond) full_prompt = [] full_prompt_2 = [] for i in range(batch_size): - full_prompt.extend(["", prompt[i], prompt[i]]) # For CLIP - full_prompt_2.extend(["", prompt_2[i], prompt_2[i]]) # For T5 + full_prompt.extend([prompt[i], prompt[i]]) # For CLIP + full_prompt_2.extend([prompt_2[i], prompt_2[i]]) # For T5 # Encode prompts pooled_prompt_embeds = self._get_clip_prompt_embeds( @@ -622,9 +511,9 @@ def encode_prompt( if self.text_encoder_2 and isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: unscale_lora_layers(self.text_encoder_2, lora_scale) - # Dummy text_ids (for compatibility) + # Dummy text_ids (updated shape for 2 prompts) dtype = self.text_encoder.dtype if self.text_encoder else self.transformer.dtype - text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + text_ids = torch.zeros(prompt_embeds.shape[1], 2).to(device=device, dtype=dtype) return prompt_embeds, pooled_prompt_embeds, text_ids @@ -825,7 +714,8 @@ def __call__( prompt_2: Optional[Union[str, List[str]]] = None, negative_prompt: Union[str, List[str]] = None, # negative_prompt_2: Optional[Union[str, List[str]]] = None, - true_pag: float = 1.0, # + true_pag_scale: float = 1.0, # + true_cfg_scale: float = 1.0, height: Optional[int] = None, width: Optional[int] = None, num_inference_steps: int = 28, @@ -951,28 +841,69 @@ def __call__( lora_scale = ( self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None ) - do_true_pag = true_pag > 1 and negative_prompt is not None + # do_true_pag = true_pag > 0 + # ( + # 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, + # ) + + + has_neg_prompt = negative_prompt is not None or negative_prompt_embeds is not None + do_true_cfg = true_cfg_scale > 1.0 and has_neg_prompt + do_true_pag = true_pag_scale > 0 + + # encode positive prompts (always) ( 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, + 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, + ) + + # encode negative prompts if needed + if do_true_cfg: + ( + negative_prompt_embeds, + negative_pooled_prompt_embeds, + _, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_pooled_prompt_embeds, device=device, num_images_per_prompt=num_images_per_prompt, max_sequence_length=max_sequence_length, lora_scale=lora_scale, ) - - if do_true_pag: - # For PAGCFG: [uncond, cond, perturbed] - prompt_embeds = torch.cat([prompt_embeds], dim=0) # already 3x in encode_prompt - else: - # For PAG (only 2x needed: [cond, perturbed]) + + if do_true_cfg and not do_true_pag: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + elif not do_true_cfg and do_true_pag: prompt_embeds = torch.cat([prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + elif do_true_cfg and do_true_pag: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds, pooled_prompt_embeds], dim=0) # 4. Prepare latent variables num_channels_latents = self.transformer.config.in_channels // 4 @@ -1008,10 +939,10 @@ def __call__( num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) self._num_timesteps = len(timesteps) - if true_pag > 0: + if true_pag_scale > 0: for name, module in self.transformer.named_modules(): if isinstance(module, Attention): - module.processor = PAGCFGIdentitySelfAttnProcessor() if negative_prompt else PAGIdentitySelfAttnProcessor() + module.processor = PAGIdentitySelfAttnProcessor() # 6. Denoising loop @@ -1020,10 +951,18 @@ def __call__( if self.interrupt: continue - if do_true_pag: + # cfg only + if do_true_cfg and not do_true_pag: + latent_model_input = torch.cat([latents] * 2) + # pag only + elif not do_true_cfg and do_true_pag: + latent_model_input = torch.cat([latents] * 2) + # both + elif do_true_cfg and do_true_pag: latent_model_input = torch.cat([latents] * 3) + # neither else: - latent_model_input = torch.cat([latents] * 2) + latent_model_input = latents # handle guidance @@ -1048,9 +987,23 @@ def __call__( return_dict=False, )[0] - if do_true_pag: + if do_true_cfg and not do_true_pag: + # uncond + cond + uncond_pred, cond_pred = noise_pred.chunk(2) + noise_pred = uncond_pred + true_cfg_scale * (cond_pred - uncond_pred) + + elif not do_true_cfg and do_true_pag: + # cond + perturbed cond_pred, perturbed_pred = noise_pred.chunk(2) - noise_pred = cond_pred + true_pag * (perturbed_pred - cond_pred) + noise_pred = cond_pred + true_pag_scale * (perturbed_pred - cond_pred) + + elif do_true_cfg and do_true_pag: + # uncond + cond + perturbed + uncond_pred, cond_pred, perturbed_pred = noise_pred.chunk(3) + cfg_pred = uncond_pred + true_cfg_scale * (cond_pred - uncond_pred) + noise_pred = cfg_pred + true_pag_scale * (perturbed_pred - cond_pred) + + # else: no guidance # compute the previous noisy sample x_t -> x_t-1 latents_dtype = latents.dtype From 9f207912e13f8a38ad449c4f070d68c8546e9157 Mon Sep 17 00:00:00 2001 From: tongyu0924 Date: Fri, 23 May 2025 17:13:58 +0800 Subject: [PATCH 05/10] Add FluxPAGPipeline implementation --- .../pipelines/pag/pipeline_pag_flux.py | 1048 +++++++++++++++++ 1 file changed, 1048 insertions(+) create mode 100644 src/diffusers/pipelines/pag/pipeline_pag_flux.py diff --git a/src/diffusers/pipelines/pag/pipeline_pag_flux.py b/src/diffusers/pipelines/pag/pipeline_pag_flux.py new file mode 100644 index 000000000000..ca34801e2737 --- /dev/null +++ b/src/diffusers/pipelines/pag/pipeline_pag_flux.py @@ -0,0 +1,1048 @@ +# Copyright 2024 Black Forest Labs and The HuggingFace 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, Union + +import numpy as np +import torch +import torch.nn.functional as F +from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast + +from diffusers.image_processor import VaeImageProcessor +from diffusers.loaders import FluxLoraLoaderMixin, FromSingleFileMixin, FluxLoraLoaderMixin +from diffusers.models.autoencoders import AutoencoderKL +from diffusers.models.attention_processor import Attention +from diffusers.models.transformers import FluxTransformer2DModel +from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.schedulers import FlowMatchEulerDiscreteScheduler +from diffusers.utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from diffusers.utils.torch_utils import randn_tensor + + +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 FluxPipeline + + >>> pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe(prompt, num_inference_steps=4, guidance_scale=0.0).images[0] + >>> image.save("flux.png") + ``` +""" + + +# 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.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, +): + """ + 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 PAGIdentitySelfAttnProcessor: + r""" + Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). + """ + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + + def __call__( + self, + attn: Attention, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + temb: Optional[torch.Tensor] = None, + *args, + **kwargs, + ) -> torch.Tensor: + if len(args) > 0 or kwargs.get("scale", None) is not None: + deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." + deprecate("scale", "1.0.0", deprecation_message) + + residual = hidden_states + if attn.spatial_norm is not None: + hidden_states = attn.spatial_norm(hidden_states, temb) + + input_ndim = hidden_states.ndim + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + + # chunk + hidden_states_org, hidden_states_ptb = hidden_states.chunk(2) + + # original path + batch_size, sequence_length, _ = hidden_states_org.shape + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + if attn.group_norm is not None: + hidden_states_org = attn.group_norm(hidden_states_org.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states_org) + key = attn.to_k(hidden_states_org) + value = attn.to_v(hidden_states_org) + + inner_dim = key.shape[-1] + head_dim = inner_dim // attn.heads + + query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + # the output of sdp = (batch, num_heads, seq_len, head_dim) + # TODO: add support for attn.scale when we move to Torch 2.1 + hidden_states_org = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) + + hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) + hidden_states_org = hidden_states_org.to(query.dtype) + + # linear proj + hidden_states_org = attn.to_out[0](hidden_states_org) + # dropout + hidden_states_org = attn.to_out[1](hidden_states_org) + + if input_ndim == 4: + hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width) + + # perturbed path (identity attention) + batch_size, sequence_length, _ = hidden_states_ptb.shape + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + if attn.group_norm is not None: + hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2) + + value = attn.to_v(hidden_states_ptb) + + # hidden_states_ptb = torch.zeros(value.shape).to(value.get_device()) + hidden_states_ptb = value + + hidden_states_ptb = hidden_states_ptb.to(query.dtype) + + # linear proj + hidden_states_ptb = attn.to_out[0](hidden_states_ptb) + # dropout + hidden_states_ptb = attn.to_out[1](hidden_states_ptb) + + if input_ndim == 4: + hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width) + + # cat + hidden_states = torch.cat([hidden_states_org, hidden_states_ptb]) + + if attn.residual_connection: + hidden_states = hidden_states + residual + + hidden_states = hidden_states / attn.rescale_output_factor + + return hidden_states + + +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + """ + Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + **kwargs, +): + """ + 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 support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps` + 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: + 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) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class FluxPAGPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin): + r""" + The Flux pipeline for text-to-image generation. + + Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->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)) if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + 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 = 64 + + 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) + + 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) + + 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]], + prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + lora_scale: Optional[float] = None, + max_sequence_length: int = 512, + ): + device = device or self._execution_device + + # Set LoRA scale + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + if self.text_encoder and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + if prompt_embeds is None: + # Normalize inputs + if isinstance(prompt, str): + prompt = [prompt] + if isinstance(prompt_2, str): + prompt_2 = [prompt_2] + batch_size = len(prompt) + + # Only cond and perturbed (skip uncond) + full_prompt = [] + full_prompt_2 = [] + for i in range(batch_size): + full_prompt.extend([prompt[i], prompt[i]]) # For CLIP + full_prompt_2.extend([prompt_2[i], prompt_2[i]]) # For T5 + + # Encode prompts + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=full_prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=full_prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + # Unscale LoRA + if self.text_encoder and isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + unscale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 and isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + unscale_lora_layers(self.text_encoder_2, lora_scale) + + # Dummy text_ids (updated shape for 2 prompts) + dtype = self.text_encoder.dtype if self.text_encoder else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 2).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % 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_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + @staticmethod + def _prepare_latent_image_ids(batch_size, height, width, device, dtype): + latent_image_ids = torch.zeros(height // 2, width // 2, 3) + latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height // 2)[:, None] + latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width // 2)[None, :] + + latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape + + latent_image_ids = latent_image_ids.reshape( + latent_image_id_height * latent_image_id_width, latent_image_id_channels + ) + + return latent_image_ids.to(device=device, dtype=dtype) + + @staticmethod + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + height = height // vae_scale_factor + width = width // vae_scale_factor + + latents = latents.view(batch_size, height, width, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), height * 2, width * 2) + + return latents + + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + height = 2 * (int(height) // self.vae_scale_factor) + width = 2 * (int(width) // self.vae_scale_factor) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is not None: + latent_image_ids = self._prepare_latent_image_ids(batch_size, height, width, 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, width, device, dtype) + + return latents, latent_image_ids + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + negative_prompt: Union[str, List[str]] = None, # + negative_prompt_2: Optional[Union[str, List[str]]] = None, + true_pag_scale: float = 1.0, # + true_cfg_scale: float = 1.0, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 28, + timesteps: List[int] = 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, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + 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 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.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._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 + ) + # do_true_pag = true_pag > 0 + # ( + # 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, + # ) + + + has_neg_prompt = negative_prompt is not None or negative_prompt_embeds is not None + do_true_cfg = true_cfg_scale > 1.0 and has_neg_prompt + do_true_pag = true_pag_scale > 0 + + # encode positive prompts (always) + ( + 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, + ) + + # encode negative prompts if needed + if do_true_cfg: + ( + negative_prompt_embeds, + negative_pooled_prompt_embeds, + _, + ) = self.encode_prompt( + prompt=negative_prompt, + prompt_2=negative_prompt_2, + prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + if do_true_cfg and not do_true_pag: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + elif not do_true_cfg and do_true_pag: + prompt_embeds = torch.cat([prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + elif do_true_cfg and do_true_pag: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, latent_image_ids = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) + 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, + timesteps, + sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + if true_pag_scale > 0: + for name, module in self.transformer.named_modules(): + if isinstance(module, Attention): + module.processor = PAGIdentitySelfAttnProcessor() + + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # cfg only + if do_true_cfg and not do_true_pag: + latent_model_input = torch.cat([latents] * 2) + # pag only + elif not do_true_cfg and do_true_pag: + latent_model_input = torch.cat([latents] * 2) + # both + elif do_true_cfg and do_true_pag: + latent_model_input = torch.cat([latents] * 3) + # neither + else: + latent_model_input = latents + + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latent_model_input.shape[0]) + else: + guidance = None + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]).to(latent_model_input.dtype) + + 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] + + if do_true_cfg and not do_true_pag: + # uncond + cond + uncond_pred, cond_pred = noise_pred.chunk(2) + noise_pred = uncond_pred + true_cfg_scale * (cond_pred - uncond_pred) + + elif not do_true_cfg and do_true_pag: + # cond + perturbed + cond_pred, perturbed_pred = noise_pred.chunk(2) + noise_pred = cond_pred + true_pag_scale * (perturbed_pred - cond_pred) + + elif do_true_cfg and do_true_pag: + # uncond + cond + perturbed + uncond_pred, cond_pred, perturbed_pred = noise_pred.chunk(3) + cfg_pred = uncond_pred + true_cfg_scale * (cond_pred - uncond_pred) + noise_pred = cfg_pred + true_pag_scale * (perturbed_pred - cond_pred) + + # else: no guidance + + # 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() + + if output_type == "latent": + image = latents + + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) From f2b9fad8c376844c568738570347ca41887b6e06 Mon Sep 17 00:00:00 2001 From: tongyu0924 Date: Fri, 23 May 2025 17:14:25 +0800 Subject: [PATCH 06/10] Add FluxPAGPipeline implementation --- examples/community/pipeline_flux_with_pag.py | 1048 ------------------ 1 file changed, 1048 deletions(-) delete mode 100644 examples/community/pipeline_flux_with_pag.py diff --git a/examples/community/pipeline_flux_with_pag.py b/examples/community/pipeline_flux_with_pag.py deleted file mode 100644 index ca34801e2737..000000000000 --- a/examples/community/pipeline_flux_with_pag.py +++ /dev/null @@ -1,1048 +0,0 @@ -# Copyright 2024 Black Forest Labs and The HuggingFace 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, Union - -import numpy as np -import torch -import torch.nn.functional as F -from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast - -from diffusers.image_processor import VaeImageProcessor -from diffusers.loaders import FluxLoraLoaderMixin, FromSingleFileMixin, FluxLoraLoaderMixin -from diffusers.models.autoencoders import AutoencoderKL -from diffusers.models.attention_processor import Attention -from diffusers.models.transformers import FluxTransformer2DModel -from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput -from diffusers.pipelines.pipeline_utils import DiffusionPipeline -from diffusers.schedulers import FlowMatchEulerDiscreteScheduler -from diffusers.utils import ( - USE_PEFT_BACKEND, - deprecate, - is_torch_xla_available, - logging, - replace_example_docstring, - scale_lora_layers, - unscale_lora_layers, -) -from diffusers.utils.torch_utils import randn_tensor - - -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 FluxPipeline - - >>> pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16) - >>> pipe.to("cuda") - >>> prompt = "A cat holding a sign that says hello world" - >>> # Depending on the variant being used, the pipeline call will slightly vary. - >>> # Refer to the pipeline documentation for more details. - >>> image = pipe(prompt, num_inference_steps=4, guidance_scale=0.0).images[0] - >>> image.save("flux.png") - ``` -""" - - -# 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.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, -): - """ - 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 PAGIdentitySelfAttnProcessor: - r""" - Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). - """ - - def __init__(self): - if not hasattr(F, "scaled_dot_product_attention"): - raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") - - def __call__( - self, - attn: Attention, - hidden_states: torch.Tensor, - encoder_hidden_states: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - temb: Optional[torch.Tensor] = None, - *args, - **kwargs, - ) -> torch.Tensor: - if len(args) > 0 or kwargs.get("scale", None) is not None: - deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." - deprecate("scale", "1.0.0", deprecation_message) - - residual = hidden_states - if attn.spatial_norm is not None: - hidden_states = attn.spatial_norm(hidden_states, temb) - - input_ndim = hidden_states.ndim - if input_ndim == 4: - batch_size, channel, height, width = hidden_states.shape - hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) - - # chunk - hidden_states_org, hidden_states_ptb = hidden_states.chunk(2) - - # original path - batch_size, sequence_length, _ = hidden_states_org.shape - - if attention_mask is not None: - attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) - # scaled_dot_product_attention expects attention_mask shape to be - # (batch, heads, source_length, target_length) - attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) - - if attn.group_norm is not None: - hidden_states_org = attn.group_norm(hidden_states_org.transpose(1, 2)).transpose(1, 2) - - query = attn.to_q(hidden_states_org) - key = attn.to_k(hidden_states_org) - value = attn.to_v(hidden_states_org) - - inner_dim = key.shape[-1] - head_dim = inner_dim // attn.heads - - query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) - - key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) - value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) - - # the output of sdp = (batch, num_heads, seq_len, head_dim) - # TODO: add support for attn.scale when we move to Torch 2.1 - hidden_states_org = F.scaled_dot_product_attention( - query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False - ) - - hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) - hidden_states_org = hidden_states_org.to(query.dtype) - - # linear proj - hidden_states_org = attn.to_out[0](hidden_states_org) - # dropout - hidden_states_org = attn.to_out[1](hidden_states_org) - - if input_ndim == 4: - hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width) - - # perturbed path (identity attention) - batch_size, sequence_length, _ = hidden_states_ptb.shape - - if attention_mask is not None: - attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) - # scaled_dot_product_attention expects attention_mask shape to be - # (batch, heads, source_length, target_length) - attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) - - if attn.group_norm is not None: - hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2) - - value = attn.to_v(hidden_states_ptb) - - # hidden_states_ptb = torch.zeros(value.shape).to(value.get_device()) - hidden_states_ptb = value - - hidden_states_ptb = hidden_states_ptb.to(query.dtype) - - # linear proj - hidden_states_ptb = attn.to_out[0](hidden_states_ptb) - # dropout - hidden_states_ptb = attn.to_out[1](hidden_states_ptb) - - if input_ndim == 4: - hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width) - - # cat - hidden_states = torch.cat([hidden_states_org, hidden_states_ptb]) - - if attn.residual_connection: - hidden_states = hidden_states + residual - - hidden_states = hidden_states / attn.rescale_output_factor - - return hidden_states - - -def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): - """ - Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and - Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 - """ - std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) - std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) - # rescale the results from guidance (fixes overexposure) - noise_pred_rescaled = noise_cfg * (std_text / std_cfg) - # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images - noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg - return noise_cfg - - -def retrieve_timesteps( - scheduler, - num_inference_steps: Optional[int] = None, - device: Optional[Union[str, torch.device]] = None, - timesteps: Optional[List[int]] = None, - **kwargs, -): - """ - 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 support arbitrary spacing between timesteps. If `None`, then the default - timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps` - 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: - 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) - else: - scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) - timesteps = scheduler.timesteps - return timesteps, num_inference_steps - - -class FluxPAGPipeline(DiffusionPipeline, FluxLoraLoaderMixin, FromSingleFileMixin): - r""" - The Flux pipeline for text-to-image generation. - - Reference: https://blackforestlabs.ai/announcing-black-forest-labs/ - - Args: - transformer ([`FluxTransformer2DModel`]): - Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. - scheduler ([`FlowMatchEulerDiscreteScheduler`]): - A scheduler to be used in combination with `transformer` to denoise the encoded image latents. - vae ([`AutoencoderKL`]): - Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. - text_encoder ([`CLIPTextModel`]): - [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically - the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. - text_encoder_2 ([`T5EncoderModel`]): - [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically - the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. - tokenizer (`CLIPTokenizer`): - Tokenizer of class - [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). - tokenizer_2 (`T5TokenizerFast`): - Second Tokenizer of class - [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). - """ - - model_cpu_offload_seq = "text_encoder->text_encoder_2->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)) if getattr(self, "vae", None) else 16 - self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) - 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 = 64 - - 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) - - 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) - - 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]], - prompt_embeds: Optional[torch.FloatTensor] = None, - pooled_prompt_embeds: Optional[torch.FloatTensor] = None, - device: Optional[torch.device] = None, - num_images_per_prompt: int = 1, - lora_scale: Optional[float] = None, - max_sequence_length: int = 512, - ): - device = device or self._execution_device - - # Set LoRA scale - if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): - self._lora_scale = lora_scale - if self.text_encoder and USE_PEFT_BACKEND: - scale_lora_layers(self.text_encoder, lora_scale) - if self.text_encoder_2 and USE_PEFT_BACKEND: - scale_lora_layers(self.text_encoder_2, lora_scale) - - if prompt_embeds is None: - # Normalize inputs - if isinstance(prompt, str): - prompt = [prompt] - if isinstance(prompt_2, str): - prompt_2 = [prompt_2] - batch_size = len(prompt) - - # Only cond and perturbed (skip uncond) - full_prompt = [] - full_prompt_2 = [] - for i in range(batch_size): - full_prompt.extend([prompt[i], prompt[i]]) # For CLIP - full_prompt_2.extend([prompt_2[i], prompt_2[i]]) # For T5 - - # Encode prompts - pooled_prompt_embeds = self._get_clip_prompt_embeds( - prompt=full_prompt, - device=device, - num_images_per_prompt=num_images_per_prompt, - ) - prompt_embeds = self._get_t5_prompt_embeds( - prompt=full_prompt_2, - device=device, - num_images_per_prompt=num_images_per_prompt, - max_sequence_length=max_sequence_length, - ) - - # Unscale LoRA - if self.text_encoder and isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: - unscale_lora_layers(self.text_encoder, lora_scale) - if self.text_encoder_2 and isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: - unscale_lora_layers(self.text_encoder_2, lora_scale) - - # Dummy text_ids (updated shape for 2 prompts) - dtype = self.text_encoder.dtype if self.text_encoder else self.transformer.dtype - text_ids = torch.zeros(prompt_embeds.shape[1], 2).to(device=device, dtype=dtype) - - return prompt_embeds, pooled_prompt_embeds, text_ids - - def check_inputs( - self, - prompt, - prompt_2, - height, - width, - negative_prompt=None, - negative_prompt_2=None, - prompt_embeds=None, - negative_prompt_embeds=None, - pooled_prompt_embeds=None, - negative_pooled_prompt_embeds=None, - callback_on_step_end_tensor_inputs=None, - max_sequence_length=None, - ): - if height % 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_2 is not None and prompt_embeds is not None: - raise ValueError( - f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" - " only forward one of the two." - ) - elif prompt is None and prompt_embeds is None: - raise ValueError( - "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." - ) - elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): - raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") - elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): - raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") - - if negative_prompt is not None and negative_prompt_embeds is not None: - raise ValueError( - f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" - f" {negative_prompt_embeds}. Please make sure to only forward one of the two." - ) - elif negative_prompt_2 is not None and negative_prompt_embeds is not None: - raise ValueError( - f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" - f" {negative_prompt_embeds}. Please make sure to only forward one of the two." - ) - - if prompt_embeds is not None and negative_prompt_embeds is not None: - if prompt_embeds.shape != negative_prompt_embeds.shape: - raise ValueError( - "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" - f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" - f" {negative_prompt_embeds.shape}." - ) - - if prompt_embeds is not None and pooled_prompt_embeds is None: - raise ValueError( - "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." - ) - if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: - raise ValueError( - "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." - ) - - if max_sequence_length is not None and max_sequence_length > 512: - raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") - - @staticmethod - def _prepare_latent_image_ids(batch_size, height, width, device, dtype): - latent_image_ids = torch.zeros(height // 2, width // 2, 3) - latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height // 2)[:, None] - latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width // 2)[None, :] - - latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape - - latent_image_ids = latent_image_ids.reshape( - latent_image_id_height * latent_image_id_width, latent_image_id_channels - ) - - return latent_image_ids.to(device=device, dtype=dtype) - - @staticmethod - def _pack_latents(latents, batch_size, num_channels_latents, height, width): - latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) - latents = latents.permute(0, 2, 4, 1, 3, 5) - latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) - - return latents - - @staticmethod - def _unpack_latents(latents, height, width, vae_scale_factor): - batch_size, num_patches, channels = latents.shape - - height = height // vae_scale_factor - width = width // vae_scale_factor - - latents = latents.view(batch_size, height, width, channels // 4, 2, 2) - latents = latents.permute(0, 3, 1, 4, 2, 5) - - latents = latents.reshape(batch_size, channels // (2 * 2), height * 2, width * 2) - - return latents - - def enable_vae_slicing(self): - r""" - Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to - compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. - """ - self.vae.enable_slicing() - - def disable_vae_slicing(self): - r""" - Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to - computing decoding in one step. - """ - self.vae.disable_slicing() - - def enable_vae_tiling(self): - r""" - Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to - compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow - processing larger images. - """ - self.vae.enable_tiling() - - def disable_vae_tiling(self): - r""" - Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to - computing decoding in one step. - """ - self.vae.disable_tiling() - - def prepare_latents( - self, - batch_size, - num_channels_latents, - height, - width, - dtype, - device, - generator, - latents=None, - ): - height = 2 * (int(height) // self.vae_scale_factor) - width = 2 * (int(width) // self.vae_scale_factor) - - shape = (batch_size, num_channels_latents, height, width) - - if latents is not None: - latent_image_ids = self._prepare_latent_image_ids(batch_size, height, width, 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, width, device, dtype) - - return latents, latent_image_ids - - @property - def guidance_scale(self): - return self._guidance_scale - - @property - def joint_attention_kwargs(self): - return self._joint_attention_kwargs - - @property - def num_timesteps(self): - return self._num_timesteps - - @property - def interrupt(self): - return self._interrupt - - @torch.no_grad() - @replace_example_docstring(EXAMPLE_DOC_STRING) - def __call__( - self, - prompt: Union[str, List[str]] = None, - prompt_2: Optional[Union[str, List[str]]] = None, - negative_prompt: Union[str, List[str]] = None, # - negative_prompt_2: Optional[Union[str, List[str]]] = None, - true_pag_scale: float = 1.0, # - true_cfg_scale: float = 1.0, - height: Optional[int] = None, - width: Optional[int] = None, - num_inference_steps: int = 28, - timesteps: List[int] = 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, - negative_prompt_embeds: Optional[torch.FloatTensor] = None, - negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, - output_type: Optional[str] = "pil", - return_dict: bool = True, - joint_attention_kwargs: Optional[Dict[str, Any]] = None, - callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, - callback_on_step_end_tensor_inputs: List[str] = ["latents"], - max_sequence_length: int = 512, - ): - r""" - Function invoked when calling the pipeline for generation. - - Args: - prompt (`str` or `List[str]`, *optional*): - The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. - instead. - prompt_2 (`str` or `List[str]`, *optional*): - The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is - will be used instead - 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 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.flux.FluxPipelineOutput`] instead of a plain tuple. - joint_attention_kwargs (`dict`, *optional*): - A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under - `self.processor` in - [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). - callback_on_step_end (`Callable`, *optional*): - A function that calls at the end of each denoising steps during the inference. The function is called - with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, - callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by - `callback_on_step_end_tensor_inputs`. - callback_on_step_end_tensor_inputs (`List`, *optional*): - The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list - will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the - `._callback_tensor_inputs` attribute of your pipeline class. - max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. - - Examples: - - Returns: - [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` - is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated - images. - """ - - height = height or self.default_sample_size * self.vae_scale_factor - width = width or self.default_sample_size * self.vae_scale_factor - - # 1. Check inputs. Raise error if not correct - self.check_inputs( - prompt, - prompt_2, - height, - width, - negative_prompt=negative_prompt, - negative_prompt_2=negative_prompt_2, - prompt_embeds=prompt_embeds, - negative_prompt_embeds=negative_prompt_embeds, - pooled_prompt_embeds=pooled_prompt_embeds, - negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, - callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, - max_sequence_length=max_sequence_length, - ) - - self._guidance_scale = guidance_scale - self._joint_attention_kwargs = joint_attention_kwargs - self._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 - ) - # do_true_pag = true_pag > 0 - # ( - # 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, - # ) - - - has_neg_prompt = negative_prompt is not None or negative_prompt_embeds is not None - do_true_cfg = true_cfg_scale > 1.0 and has_neg_prompt - do_true_pag = true_pag_scale > 0 - - # encode positive prompts (always) - ( - 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, - ) - - # encode negative prompts if needed - if do_true_cfg: - ( - negative_prompt_embeds, - negative_pooled_prompt_embeds, - _, - ) = self.encode_prompt( - prompt=negative_prompt, - prompt_2=negative_prompt_2, - prompt_embeds=negative_prompt_embeds, - pooled_prompt_embeds=negative_pooled_prompt_embeds, - device=device, - num_images_per_prompt=num_images_per_prompt, - max_sequence_length=max_sequence_length, - lora_scale=lora_scale, - ) - - if do_true_cfg and not do_true_pag: - prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) - pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) - elif not do_true_cfg and do_true_pag: - prompt_embeds = torch.cat([prompt_embeds, prompt_embeds], dim=0) - pooled_prompt_embeds = torch.cat([pooled_prompt_embeds, pooled_prompt_embeds], dim=0) - elif do_true_cfg and do_true_pag: - prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0) - pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds, pooled_prompt_embeds], dim=0) - - # 4. Prepare latent variables - num_channels_latents = self.transformer.config.in_channels // 4 - latents, latent_image_ids = self.prepare_latents( - batch_size * num_images_per_prompt, - num_channels_latents, - height, - width, - prompt_embeds.dtype, - device, - generator, - latents, - ) - - # 5. Prepare timesteps - sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) - 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, - timesteps, - sigmas, - mu=mu, - ) - num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) - self._num_timesteps = len(timesteps) - - if true_pag_scale > 0: - for name, module in self.transformer.named_modules(): - if isinstance(module, Attention): - module.processor = PAGIdentitySelfAttnProcessor() - - - # 6. Denoising loop - with self.progress_bar(total=num_inference_steps) as progress_bar: - for i, t in enumerate(timesteps): - if self.interrupt: - continue - - # cfg only - if do_true_cfg and not do_true_pag: - latent_model_input = torch.cat([latents] * 2) - # pag only - elif not do_true_cfg and do_true_pag: - latent_model_input = torch.cat([latents] * 2) - # both - elif do_true_cfg and do_true_pag: - latent_model_input = torch.cat([latents] * 3) - # neither - else: - latent_model_input = latents - - - # handle guidance - if self.transformer.config.guidance_embeds: - guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) - guidance = guidance.expand(latent_model_input.shape[0]) - else: - guidance = None - - # broadcast to batch dimension in a way that's compatible with ONNX/Core ML - timestep = t.expand(latent_model_input.shape[0]).to(latent_model_input.dtype) - - 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] - - if do_true_cfg and not do_true_pag: - # uncond + cond - uncond_pred, cond_pred = noise_pred.chunk(2) - noise_pred = uncond_pred + true_cfg_scale * (cond_pred - uncond_pred) - - elif not do_true_cfg and do_true_pag: - # cond + perturbed - cond_pred, perturbed_pred = noise_pred.chunk(2) - noise_pred = cond_pred + true_pag_scale * (perturbed_pred - cond_pred) - - elif do_true_cfg and do_true_pag: - # uncond + cond + perturbed - uncond_pred, cond_pred, perturbed_pred = noise_pred.chunk(3) - cfg_pred = uncond_pred + true_cfg_scale * (cond_pred - uncond_pred) - noise_pred = cfg_pred + true_pag_scale * (perturbed_pred - cond_pred) - - # else: no guidance - - # 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() - - if output_type == "latent": - image = latents - - else: - latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) - latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor - image = self.vae.decode(latents, return_dict=False)[0] - image = self.image_processor.postprocess(image, output_type=output_type) - - # Offload all models - self.maybe_free_model_hooks() - - if not return_dict: - return (image,) - - return FluxPipelineOutput(images=image) From 6943d74ba200fd8fb8ccc78d0e4ac4414a05cbec Mon Sep 17 00:00:00 2001 From: tongyu0924 Date: Mon, 26 May 2025 17:52:37 +0800 Subject: [PATCH 07/10] Register FluxPAGPipeline in init hooks --- src/diffusers/__init__.py | 2 ++ src/diffusers/pipelines/__init__.py | 1 + 2 files changed, 3 insertions(+) diff --git a/src/diffusers/__init__.py b/src/diffusers/__init__.py index 9ab973351c86..847443253a9d 100644 --- a/src/diffusers/__init__.py +++ b/src/diffusers/__init__.py @@ -376,6 +376,7 @@ "FluxImg2ImgPipeline", "FluxInpaintPipeline", "FluxPipeline", + "FluxPAGPipeline", "FluxPriorReduxPipeline", "HiDreamImagePipeline", "HunyuanDiTControlNetPipeline", @@ -960,6 +961,7 @@ FluxImg2ImgPipeline, FluxInpaintPipeline, FluxPipeline, + FluxPAGPipeline, FluxPriorReduxPipeline, HiDreamImagePipeline, HunyuanDiTControlNetPipeline, diff --git a/src/diffusers/pipelines/__init__.py b/src/diffusers/pipelines/__init__.py index 4debb868d9dc..e9e887fbaaae 100644 --- a/src/diffusers/pipelines/__init__.py +++ b/src/diffusers/pipelines/__init__.py @@ -191,6 +191,7 @@ "StableDiffusionXLPAGImg2ImgPipeline", "PixArtSigmaPAGPipeline", "SanaPAGPipeline", + "FluxPAGPipeline", ] ) _import_structure["controlnet_xs"].extend( From dae4ccf4e5496424c42bd332c3d8340f134a6763 Mon Sep 17 00:00:00 2001 From: tongyu0924 Date: Mon, 26 May 2025 21:14:41 +0800 Subject: [PATCH 08/10] add test --- tests/pipelines/pag/test_pag_flux.py | 206 +++++++++++++++++++++++++++ 1 file changed, 206 insertions(+) create mode 100644 tests/pipelines/pag/test_pag_flux.py diff --git a/tests/pipelines/pag/test_pag_flux.py b/tests/pipelines/pag/test_pag_flux.py new file mode 100644 index 000000000000..dd79a1f59746 --- /dev/null +++ b/tests/pipelines/pag/test_pag_flux.py @@ -0,0 +1,206 @@ +import inspect +import unittest + +import numpy as np +import torch +from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel + +from diffusers import ( + AutoencoderKL, + FlowMatchEulerDiscreteScheduler, + FluxPAGPipeline, + SD3Transformer2DModel, +) +from diffusers.utils.testing_utils import torch_device + +from ..test_pipelines_common import ( + PipelineTesterMixin, + check_qkv_fusion_matches_attn_procs_length, + check_qkv_fusion_processors_exist, +) + + +class FluxPAGPipelineFastTests(unittest.TestCase, PipelineTesterMixin): + pipeline_class = FluxPAGPipeline + params = frozenset([ + "prompt", + "height", + "width", + "guidance_scale", + "negative_prompt", + "prompt_embeds", + "negative_prompt_embeds", + ]) + batch_params = frozenset(["prompt", "negative_prompt"]) + test_xformers_attention = False + + def get_dummy_components(self): + torch.manual_seed(0) + transformer = SD3Transformer2DModel( + sample_size=32, + patch_size=1, + in_channels=4, + num_layers=2, + attention_head_dim=8, + num_attention_heads=4, + caption_projection_dim=32, + joint_attention_dim=32, + pooled_projection_dim=64, + out_channels=4, + ) + 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, + ) + text_encoder = CLIPTextModelWithProjection(clip_text_encoder_config) + text_encoder_2 = CLIPTextModelWithProjection(clip_text_encoder_config) + text_encoder_3 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5") + + tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip") + tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip") + tokenizer_3 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5") + + vae = AutoencoderKL( + sample_size=32, + in_channels=3, + out_channels=3, + block_out_channels=(4,), + layers_per_block=1, + latent_channels=4, + 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, + "text_encoder_3": text_encoder_3, + "tokenizer": tokenizer, + "tokenizer_2": tokenizer_2, + "tokenizer_3": tokenizer_3, + "transformer": transformer, + "vae": vae, + } + + def get_dummy_inputs(self, device, seed=0): + generator = torch.manual_seed(seed) if str(device).startswith("mps") else torch.Generator(device="cpu").manual_seed(seed) + return { + "prompt": "A painting of a squirrel eating a burger", + "generator": generator, + "num_inference_steps": 2, + "guidance_scale": 5.0, + "output_type": "np", + "true_pag_scale": 0.0, + } + + def test_different_prompts(self): + pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device) + + inputs = self.get_dummy_inputs(torch_device) + image1 = pipe(**inputs).images[0] + + inputs["prompt_2"] = "A completely different scene" + image2 = pipe(**inputs).images[0] + + assert np.abs(image1 - image2).max() > 1e-2 + + def test_different_negative_prompts(self): + pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device) + + inputs = self.get_dummy_inputs(torch_device) + image1 = pipe(**inputs).images[0] + + inputs["negative_prompt"] = "ugly, blurry" + image2 = pipe(**inputs).images[0] + + assert np.abs(image1 - image2).max() > 1e-2 + + def test_pag_disable_equivalent_to_baseline(self): + pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device) + + inputs = self.get_dummy_inputs(torch_device) + image_pag_disabled = pipe(**inputs).images[0] + + del inputs["true_pag_scale"] + image_baseline = pipe(**inputs).images[0] + + assert np.abs(image_pag_disabled - image_baseline).max() < 1e-3 + + def test_fused_qkv_projections(self): + device = "cpu" + components = self.get_dummy_components() + pipe = self.pipeline_class(**components).to(device) + + inputs = self.get_dummy_inputs(device) + image = pipe(**inputs).images + image_slice_original = image[0, -3:, -3:, -1] + + pipe.transformer.fuse_qkv_projections() + assert check_qkv_fusion_processors_exist(pipe.transformer) + assert check_qkv_fusion_matches_attn_procs_length(pipe.transformer, pipe.transformer.original_attn_processors) + + image_fused = pipe(**inputs).images + image_slice_fused = image_fused[0, -3:, -3:, -1] + + pipe.transformer.unfuse_qkv_projections() + image_disabled = pipe(**inputs).images + image_slice_disabled = image_disabled[0, -3:, -3:, -1] + + assert np.allclose(image_slice_original, image_slice_fused, atol=1e-3) + assert np.allclose(image_slice_fused, image_slice_disabled, atol=1e-3) + + def test_pag_applied_layers(self): + device = "cpu" + components = self.get_dummy_components() + pipe = self.pipeline_class(**components).to(device) + + all_attn_keys = [k for k in pipe.transformer.attn_processors if "attn" in k] + original_procs = pipe.transformer.attn_processors.copy() + + pag_layers = ["blocks.0", "blocks.1"] + pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False) + assert set(pipe.pag_attn_processors) == set(all_attn_keys) + + pipe.transformer.set_attn_processor(original_procs) + pag_layers = ["blocks.0"] + pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False) + expected_keys = [k for k in all_attn_keys if k.startswith("transformer_blocks.0")] + assert set(pipe.pag_attn_processors) == set(expected_keys) + + pipe.transformer.set_attn_processor(original_procs) + pag_layers = [r"blocks\.(0|1)"] + pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False) + assert len(pipe.pag_attn_processors) == 2 + + def test_forward_signature_consistency(self): + sig = inspect.signature(self.pipeline_class.__call__) + expected = set(self.params) + found = set(sig.parameters.keys()) + missing = expected - found + extra = found - expected + assert not missing, f"Missing parameters in pipeline: {missing}" + assert not extra - {'self'}, f"Unexpected parameters in pipeline: {extra}" + + def test_attention_mask_support(self): + pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device) + inputs = self.get_dummy_inputs(torch_device) + inputs["attention_mask"] = torch.ones((1, 77)) + try: + pipe(**inputs) + except Exception as e: + assert "attention_mask" not in str(e), f"Pipeline should support attention_mask, but failed: {e}" From 696af97a6ddead07181ae71c503221a52465da4c Mon Sep 17 00:00:00 2001 From: tongyu0924 Date: Mon, 9 Jun 2025 22:06:09 +0800 Subject: [PATCH 09/10] feat: add PAGFluxAttnProcessor_2_0 and PAGCFGFluxAttnProcessor_2_0 to support true PAG/CFG in Flux pipeline --- src/diffusers/models/attention_processor.py | 117 +++++++++++++++ .../pipelines/pag/pipeline_pag_flux.py | 136 +----------------- 2 files changed, 122 insertions(+), 131 deletions(-) diff --git a/src/diffusers/models/attention_processor.py b/src/diffusers/models/attention_processor.py index 23ae05e2ab96..ce64bb34d468 100755 --- a/src/diffusers/models/attention_processor.py +++ b/src/diffusers/models/attention_processor.py @@ -6162,6 +6162,123 @@ def __call__( return hidden_states +class PAGFluxAttnProcessor_2_0(AttnProcessor2_0): + def __call__( + self, + attn: Attention, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + temb: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + # Save input for residual connection later + original_residual = hidden_states + + if attn.spatial_norm is not None: + hidden_states = attn.spatial_norm(hidden_states, temb) + + B, C, H, W = hidden_states.shape + hidden_states = hidden_states.view(B, C, H * W).transpose(1, 2) + hidden_states_org, hidden_states_ptb = hidden_states.chunk(2) + + def apply_attn(path): + B_local = path.shape[0] + if attn.group_norm is not None: + path = attn.group_norm(path.transpose(1, 2)).transpose(1, 2) + query = attn.to_q(path) + key = attn.to_k(path) + value = attn.to_v(path) + + head_dim = query.shape[-1] // attn.heads + query = query.view(B_local, -1, attn.heads, head_dim).transpose(1, 2) + key = key.view(B_local, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(B_local, -1, attn.heads, head_dim).transpose(1, 2) + + attn_output = F.scaled_dot_product_attention( + query, key, value, dropout_p=0.0, is_causal=False + ) + attn_output = attn_output.transpose(1, 2).reshape(B_local, -1, attn.heads * head_dim) + attn_output = attn.to_out[0](attn_output) + attn_output = attn.to_out[1](attn_output) + return attn_output + + hidden_states_org = apply_attn(hidden_states_org) + + # Perturbed path (identity attention) + if attn.group_norm is not None: + hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2) + hidden_states_ptb = attn.to_v(hidden_states_ptb) + hidden_states_ptb = attn.to_out[0](hidden_states_ptb) + hidden_states_ptb = attn.to_out[1](hidden_states_ptb) + + hidden_states = torch.cat([hidden_states_org, hidden_states_ptb], dim=0) + hidden_states = hidden_states.transpose(1, 2).view(2 * B, C, H, W) + + if attn.residual_connection: + # Expand residual to match batch size + residual = original_residual.expand_as(hidden_states) + hidden_states = hidden_states + residual + + return hidden_states / attn.rescale_output_factor + + +class PAGCFGFluxAttnProcessor_2_0(PAGFluxAttnProcessor_2_0): + def __call__( + self, + attn: Attention, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + temb: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + original_residual = hidden_states + + if attn.spatial_norm is not None: + hidden_states = attn.spatial_norm(hidden_states, temb) + + B, C, H, W = hidden_states.shape + hidden_states = hidden_states.view(B, C, H * W).transpose(1, 2) + uncond, cond, ptb = hidden_states.chunk(3) + + def apply_attn(path): + B_local = path.shape[0] + if attn.group_norm is not None: + path = attn.group_norm(path.transpose(1, 2)).transpose(1, 2) + query = attn.to_q(path) + key = attn.to_k(path) + value = attn.to_v(path) + + head_dim = query.shape[-1] // attn.heads + query = query.view(B_local, -1, attn.heads, head_dim).transpose(1, 2) + key = key.view(B_local, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(B_local, -1, attn.heads, head_dim).transpose(1, 2) + + attn_output = F.scaled_dot_product_attention( + query, key, value, dropout_p=0.0, is_causal=False + ) + attn_output = attn_output.transpose(1, 2).reshape(B_local, -1, attn.heads * head_dim) + attn_output = attn.to_out[0](attn_output) + attn_output = attn.to_out[1](attn_output) + return attn_output + + uncond_out = apply_attn(uncond) + cond_out = apply_attn(cond) + + if attn.group_norm is not None: + ptb = attn.group_norm(ptb.transpose(1, 2)).transpose(1, 2) + ptb = attn.to_v(ptb) + ptb = attn.to_out[0](ptb) + ptb = attn.to_out[1](ptb) + + hidden_states = torch.cat([uncond_out, cond_out, ptb], dim=0) + hidden_states = hidden_states.transpose(1, 2).view(3 * B, C, H, W) + + if attn.residual_connection: + residual = original_residual.expand_as(hidden_states) + hidden_states = hidden_states + residual + + return hidden_states / attn.rescale_output_factor + ADDED_KV_ATTENTION_PROCESSORS = ( AttnAddedKVProcessor, diff --git a/src/diffusers/pipelines/pag/pipeline_pag_flux.py b/src/diffusers/pipelines/pag/pipeline_pag_flux.py index ca34801e2737..076608157678 100644 --- a/src/diffusers/pipelines/pag/pipeline_pag_flux.py +++ b/src/diffusers/pipelines/pag/pipeline_pag_flux.py @@ -38,6 +38,7 @@ unscale_lora_layers, ) from diffusers.utils.torch_utils import randn_tensor +from diffusers.models.attention_processor import PAGFluxAttnProcessor_2_0, PAGCFGFluxAttnProcessor_2_0 if is_torch_xla_available(): @@ -140,120 +141,6 @@ def retrieve_timesteps( timesteps = scheduler.timesteps return timesteps, num_inference_steps -class PAGIdentitySelfAttnProcessor: - r""" - Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). - """ - - def __init__(self): - if not hasattr(F, "scaled_dot_product_attention"): - raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") - - def __call__( - self, - attn: Attention, - hidden_states: torch.Tensor, - encoder_hidden_states: Optional[torch.Tensor] = None, - attention_mask: Optional[torch.Tensor] = None, - temb: Optional[torch.Tensor] = None, - *args, - **kwargs, - ) -> torch.Tensor: - if len(args) > 0 or kwargs.get("scale", None) is not None: - deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`." - deprecate("scale", "1.0.0", deprecation_message) - - residual = hidden_states - if attn.spatial_norm is not None: - hidden_states = attn.spatial_norm(hidden_states, temb) - - input_ndim = hidden_states.ndim - if input_ndim == 4: - batch_size, channel, height, width = hidden_states.shape - hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) - - # chunk - hidden_states_org, hidden_states_ptb = hidden_states.chunk(2) - - # original path - batch_size, sequence_length, _ = hidden_states_org.shape - - if attention_mask is not None: - attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) - # scaled_dot_product_attention expects attention_mask shape to be - # (batch, heads, source_length, target_length) - attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) - - if attn.group_norm is not None: - hidden_states_org = attn.group_norm(hidden_states_org.transpose(1, 2)).transpose(1, 2) - - query = attn.to_q(hidden_states_org) - key = attn.to_k(hidden_states_org) - value = attn.to_v(hidden_states_org) - - inner_dim = key.shape[-1] - head_dim = inner_dim // attn.heads - - query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) - - key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) - value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) - - # the output of sdp = (batch, num_heads, seq_len, head_dim) - # TODO: add support for attn.scale when we move to Torch 2.1 - hidden_states_org = F.scaled_dot_product_attention( - query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False - ) - - hidden_states_org = hidden_states_org.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) - hidden_states_org = hidden_states_org.to(query.dtype) - - # linear proj - hidden_states_org = attn.to_out[0](hidden_states_org) - # dropout - hidden_states_org = attn.to_out[1](hidden_states_org) - - if input_ndim == 4: - hidden_states_org = hidden_states_org.transpose(-1, -2).reshape(batch_size, channel, height, width) - - # perturbed path (identity attention) - batch_size, sequence_length, _ = hidden_states_ptb.shape - - if attention_mask is not None: - attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) - # scaled_dot_product_attention expects attention_mask shape to be - # (batch, heads, source_length, target_length) - attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) - - if attn.group_norm is not None: - hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2) - - value = attn.to_v(hidden_states_ptb) - - # hidden_states_ptb = torch.zeros(value.shape).to(value.get_device()) - hidden_states_ptb = value - - hidden_states_ptb = hidden_states_ptb.to(query.dtype) - - # linear proj - hidden_states_ptb = attn.to_out[0](hidden_states_ptb) - # dropout - hidden_states_ptb = attn.to_out[1](hidden_states_ptb) - - if input_ndim == 4: - hidden_states_ptb = hidden_states_ptb.transpose(-1, -2).reshape(batch_size, channel, height, width) - - # cat - hidden_states = torch.cat([hidden_states_org, hidden_states_ptb]) - - if attn.residual_connection: - hidden_states = hidden_states + residual - - hidden_states = hidden_states / attn.rescale_output_factor - - return hidden_states - - def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): """ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and @@ -841,21 +728,6 @@ def __call__( lora_scale = ( self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None ) - # do_true_pag = true_pag > 0 - # ( - # 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, - # ) has_neg_prompt = negative_prompt is not None or negative_prompt_embeds is not None @@ -942,8 +814,10 @@ def __call__( if true_pag_scale > 0: for name, module in self.transformer.named_modules(): if isinstance(module, Attention): - module.processor = PAGIdentitySelfAttnProcessor() - + if do_true_cfg: + module.processor = PAGCFGFluxAttnProcessor_2_0() + else: + module.processor = PAGFluxAttnProcessor_2_0() # 6. Denoising loop with self.progress_bar(total=num_inference_steps) as progress_bar: From 698e1291d6d5d88a452f7bd617b345dfab538461 Mon Sep 17 00:00:00 2001 From: tongyu0924 Date: Tue, 10 Jun 2025 05:59:32 +0800 Subject: [PATCH 10/10] fix: update EXAMPLE_DOC_STRING for FluxPAGPipeline --- src/diffusers/pipelines/pag/pipeline_pag_flux.py | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/src/diffusers/pipelines/pag/pipeline_pag_flux.py b/src/diffusers/pipelines/pag/pipeline_pag_flux.py index 076608157678..980a1f198380 100644 --- a/src/diffusers/pipelines/pag/pipeline_pag_flux.py +++ b/src/diffusers/pipelines/pag/pipeline_pag_flux.py @@ -53,10 +53,8 @@ EXAMPLE_DOC_STRING = """ Examples: - ```py >>> import torch >>> from diffusers import FluxPipeline - >>> pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16) >>> pipe.to("cuda") >>> prompt = "A cat holding a sign that says hello world" @@ -64,10 +62,10 @@ >>> # Refer to the pipeline documentation for more details. >>> image = pipe(prompt, num_inference_steps=4, guidance_scale=0.0).images[0] >>> image.save("flux.png") - ``` """ + # Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift def calculate_shift( image_seq_len,