Added ability to add models to finetune as plugins. Also added flux2 new arch via that method.

This commit is contained in:
Jaret Burkett
2025-03-27 16:07:00 -06:00
parent e9e30104d3
commit 5365200da1
12 changed files with 936 additions and 1058 deletions

View File

@@ -0,0 +1,6 @@
from .flex2 import Flex2
AI_TOOLKIT_MODELS = [
# put a list of models here
Flex2
]

View File

@@ -0,0 +1,483 @@
import os
from typing import TYPE_CHECKING, List
import torch
import yaml
from toolkit import train_tools
from toolkit.config_modules import GenerateImageConfig, ModelConfig
from PIL import Image
from toolkit.models.base_model import BaseModel
from diffusers import FluxTransformer2DModel, AutoencoderKL
from toolkit.basic import flush
from toolkit.prompt_utils import PromptEmbeds
from toolkit.samplers.custom_flowmatch_sampler import CustomFlowMatchEulerDiscreteScheduler
from toolkit.models.flux import add_model_gpu_splitter_to_flux, bypass_flux_guidance, restore_flux_guidance
from toolkit.dequantize import patch_dequantization_on_save
from toolkit.accelerator import get_accelerator, unwrap_model
from optimum.quanto import freeze, QTensor
from toolkit.util.mask import generate_random_mask
from toolkit.util.quantize import quantize, get_qtype
from transformers import T5TokenizerFast, T5EncoderModel, CLIPTextModel, CLIPTokenizer
from .pipeline import Flex2Pipeline
from einops import rearrange, repeat
import random
import torch.nn.functional as F
if TYPE_CHECKING:
from toolkit.data_transfer_object.data_loader import DataLoaderBatchDTO
scheduler_config = {
"base_image_seq_len": 256,
"base_shift": 0.5,
"max_image_seq_len": 4096,
"max_shift": 1.15,
"num_train_timesteps": 1000,
"shift": 3.0,
"use_dynamic_shifting": True
}
class Flex2(BaseModel):
arch = "flex2"
def __init__(
self,
device,
model_config: ModelConfig,
dtype='bf16',
custom_pipeline=None,
noise_scheduler=None,
**kwargs
):
super().__init__(
device,
model_config,
dtype,
custom_pipeline,
noise_scheduler,
**kwargs
)
self.is_flow_matching = True
self.is_transformer = True
self.target_lora_modules = ['FluxTransformer2DModel']
# for training, pass these as kwargs
self.invert_inpaint_mask_chance = model_config.model_kwargs.get('invert_inpaint_mask_chance', 0.0)
self.inpaint_dropout = model_config.model_kwargs.get('inpaint_dropout', 0.0)
self.control_dropout = model_config.model_kwargs.get('control_dropout', 0.0)
self.inpaint_random_chance = model_config.model_kwargs.get('inpaint_random_chance', 0.0)
# static method to get the noise scheduler
@staticmethod
def get_train_scheduler():
return CustomFlowMatchEulerDiscreteScheduler(**scheduler_config)
def load_model(self):
dtype = self.torch_dtype
self.print_and_status_update("Loading Flux2 model")
# will be updated if we detect a existing checkpoint in training folder
model_path = self.model_config.name_or_path
# this is the original path put in the model directory
# it is here because for finetuning we only save the transformer usually
# so we need this for the VAE, te, etc
base_model_path = self.model_config.name_or_path_original
transformer_path = model_path
transformer_subfolder = 'transformer'
if os.path.exists(transformer_path):
transformer_subfolder = None
transformer_path = os.path.join(transformer_path, 'transformer')
# check if the path is a full checkpoint.
te_folder_path = os.path.join(model_path, 'text_encoder')
# if we have the te, this folder is a full checkpoint, use it as the base
if os.path.exists(te_folder_path):
base_model_path = model_path
self.print_and_status_update("Loading transformer")
transformer = FluxTransformer2DModel.from_pretrained(
transformer_path,
subfolder=transformer_subfolder,
torch_dtype=dtype,
)
transformer.to(self.quantize_device, dtype=dtype)
if self.model_config.quantize:
# patch the state dict method
patch_dequantization_on_save(transformer)
quantization_type = get_qtype(self.model_config.qtype)
self.print_and_status_update("Quantizing transformer")
quantize(transformer, weights=quantization_type,
**self.model_config.quantize_kwargs)
freeze(transformer)
transformer.to(self.device_torch)
else:
transformer.to(self.device_torch, dtype=dtype)
flush()
self.print_and_status_update("Loading T5")
tokenizer_2 = T5TokenizerFast.from_pretrained(
base_model_path, subfolder="tokenizer_2", torch_dtype=dtype
)
text_encoder_2 = T5EncoderModel.from_pretrained(
base_model_path, subfolder="text_encoder_2", torch_dtype=dtype
)
text_encoder_2.to(self.device_torch, dtype=dtype)
flush()
if self.model_config.quantize_te:
self.print_and_status_update("Quantizing T5")
quantize(text_encoder_2, weights=get_qtype(
self.model_config.qtype))
freeze(text_encoder_2)
flush()
self.print_and_status_update("Loading CLIP")
text_encoder = CLIPTextModel.from_pretrained(
base_model_path, subfolder="text_encoder", torch_dtype=dtype)
tokenizer = CLIPTokenizer.from_pretrained(
base_model_path, subfolder="tokenizer", torch_dtype=dtype)
text_encoder.to(self.device_torch, dtype=dtype)
self.print_and_status_update("Loading VAE")
vae = AutoencoderKL.from_pretrained(
base_model_path, subfolder="vae", torch_dtype=dtype)
self.noise_scheduler = Flex2.get_train_scheduler()
self.print_and_status_update("Making pipe")
pipe: Flex2Pipeline = Flex2Pipeline(
scheduler=self.noise_scheduler,
text_encoder=text_encoder,
tokenizer=tokenizer,
text_encoder_2=None,
tokenizer_2=tokenizer_2,
vae=vae,
transformer=None,
)
# for quantization, it works best to do these after making the pipe
pipe.text_encoder_2 = text_encoder_2
pipe.transformer = transformer
self.print_and_status_update("Preparing Model")
text_encoder = [pipe.text_encoder, pipe.text_encoder_2]
tokenizer = [pipe.tokenizer, pipe.tokenizer_2]
pipe.transformer = pipe.transformer.to(self.device_torch)
flush()
# just to make sure everything is on the right device and dtype
text_encoder[0].to(self.device_torch)
text_encoder[0].requires_grad_(False)
text_encoder[0].eval()
text_encoder[1].to(self.device_torch)
text_encoder[1].requires_grad_(False)
text_encoder[1].eval()
pipe.transformer = pipe.transformer.to(self.device_torch)
flush()
# save it to the model class
self.vae = vae
self.text_encoder = text_encoder # list of text encoders
self.tokenizer = tokenizer # list of tokenizers
self.model = pipe.transformer
self.pipeline = pipe
self.print_and_status_update("Model Loaded")
def get_generation_pipeline(self):
scheduler = Flex2.get_train_scheduler()
pipeline: Flex2Pipeline = Flex2Pipeline(
scheduler=scheduler,
text_encoder=unwrap_model(self.text_encoder[0]),
tokenizer=self.tokenizer[0],
text_encoder_2=unwrap_model(self.text_encoder[1]),
tokenizer_2=self.tokenizer[1],
vae=unwrap_model(self.vae),
transformer=unwrap_model(self.transformer)
)
pipeline = pipeline.to(self.device_torch)
return pipeline
def generate_single_image(
self,
pipeline: Flex2Pipeline,
gen_config: GenerateImageConfig,
conditional_embeds: PromptEmbeds,
unconditional_embeds: PromptEmbeds,
generator: torch.Generator,
extra: dict,
):
if gen_config.ctrl_img is None:
control_img = None
else:
control_img = Image.open(gen_config.ctrl_img)
if ".inpaint." not in gen_config.ctrl_img:
control_img = control_img.convert("RGB")
else:
# make sure it has an alpha
if control_img.mode != "RGBA":
raise ValueError("Inpainting images must have an alpha channel")
img = pipeline(
prompt_embeds=conditional_embeds.text_embeds,
pooled_prompt_embeds=conditional_embeds.pooled_embeds,
height=gen_config.height,
width=gen_config.width,
num_inference_steps=gen_config.num_inference_steps,
guidance_scale=gen_config.guidance_scale,
latents=gen_config.latents,
generator=generator,
control_image=control_img,
control_image_idx=gen_config.ctrl_idx,
**extra
).images[0]
return img
def get_noise_prediction(
self,
latent_model_input: torch.Tensor,
timestep: torch.Tensor, # 0 to 1000 scale
text_embeddings: PromptEmbeds,
guidance_embedding_scale: float,
bypass_guidance_embedding: bool,
**kwargs
):
with torch.no_grad():
bs, c, h, w = latent_model_input.shape
latent_model_input_packed = rearrange(
latent_model_input,
"b c (h ph) (w pw) -> b (h w) (c ph pw)",
ph=2,
pw=2
)
img_ids = torch.zeros(h // 2, w // 2, 3)
img_ids[..., 1] = img_ids[..., 1] + torch.arange(h // 2)[:, None]
img_ids[..., 2] = img_ids[..., 2] + torch.arange(w // 2)[None, :]
img_ids = repeat(img_ids, "h w c -> b (h w) c",
b=bs).to(self.device_torch)
txt_ids = torch.zeros(
bs, text_embeddings.text_embeds.shape[1], 3).to(self.device_torch)
# # handle guidance
if self.unet_unwrapped.config.guidance_embeds:
if isinstance(guidance_embedding_scale, list):
guidance = torch.tensor(
guidance_embedding_scale, device=self.device_torch)
else:
guidance = torch.tensor(
[guidance_embedding_scale], device=self.device_torch)
guidance = guidance.expand(latent_model_input.shape[0])
else:
guidance = None
if bypass_guidance_embedding:
bypass_flux_guidance(self.unet)
cast_dtype = self.unet.dtype
# changes from orig implementation
if txt_ids.ndim == 3:
txt_ids = txt_ids[0]
if img_ids.ndim == 3:
img_ids = img_ids[0]
noise_pred = self.unet(
hidden_states=latent_model_input_packed.to(
self.device_torch, cast_dtype),
timestep=timestep / 1000,
encoder_hidden_states=text_embeddings.text_embeds.to(
self.device_torch, cast_dtype),
pooled_projections=text_embeddings.pooled_embeds.to(
self.device_torch, cast_dtype),
txt_ids=txt_ids,
img_ids=img_ids,
guidance=guidance,
return_dict=False,
**kwargs,
)[0]
if isinstance(noise_pred, QTensor):
noise_pred = noise_pred.dequantize()
noise_pred = rearrange(
noise_pred,
"b (h w) (c ph pw) -> b c (h ph) (w pw)",
h=latent_model_input.shape[2] // 2,
w=latent_model_input.shape[3] // 2,
ph=2,
pw=2,
c=self.vae.config.latent_channels
)
if bypass_guidance_embedding:
restore_flux_guidance(self.unet)
return noise_pred
def get_prompt_embeds(self, prompt: str) -> PromptEmbeds:
if self.pipeline.text_encoder.device != self.device_torch:
self.pipeline.text_encoder.to(self.device_torch)
prompt_embeds, pooled_prompt_embeds = train_tools.encode_prompts_flux(
self.tokenizer,
self.text_encoder,
prompt,
max_length=512,
)
pe = PromptEmbeds(
prompt_embeds
)
pe.pooled_embeds = pooled_prompt_embeds
return pe
def get_model_has_grad(self):
# return from a weight if it has grad
return self.model.proj_out.weight.requires_grad
def get_te_has_grad(self):
# return from a weight if it has grad
return self.text_encoder[1].encoder.block[0].layer[0].SelfAttention.q.weight.requires_grad
def save_model(self, output_path, meta, save_dtype):
# only save the unet
transformer: FluxTransformer2DModel = unwrap_model(self.model)
transformer.save_pretrained(
save_directory=os.path.join(output_path, 'transformer'),
safe_serialization=True,
)
meta_path = os.path.join(output_path, 'aitk_meta.yaml')
with open(meta_path, 'w') as f:
yaml.dump(meta, f)
def get_loss_target(self, *args, **kwargs):
noise = kwargs.get('noise')
batch = kwargs.get('batch')
return (noise - batch.latents).detach()
def condition_noisy_latents(self, latents: torch.Tensor, batch:'DataLoaderBatchDTO'):
with torch.no_grad():
# inpainting input is 0-1 (bs, 4, h, w) on batch.inpaint_tensor
# 4th channel is the mask with 1 being keep area and 0 being area to inpaint.
# todo handle dropout on a batch item level, this frops out the entire batch
do_dropout = random.random() < self.inpaint_dropout if self.inpaint_dropout > 0.0 else False
# do random mask if we dont have one
inpaint_tensor = batch.inpaint_tensor
if self.inpaint_random_chance > 0.0:
do_random = random.random() < self.inpaint_random_chance
if do_random:
# force a random tensor
inpaint_tensor = None
if inpaint_tensor is None and not do_dropout:
# generate a random one since we dont have one
# this will make random blobs, invert the blobs for now as we normanlly inpaint the alpha
inpaint_tensor = 1 - generate_random_mask(
batch_size=latents.shape[0],
height=latents.shape[2],
width=latents.shape[3],
device=latents.device,
).to(latents.device, latents.dtype)
if inpaint_tensor is not None and not do_dropout:
if inpaint_tensor.shape[1] == 4:
# get just the mask
inpainting_tensor_mask = inpaint_tensor[:, 3:4, :, :].to(latents.device, dtype=latents.dtype)
elif inpaint_tensor.shape[1] == 3:
# rgb mask. Just get one channel
inpainting_tensor_mask = inpaint_tensor[:, 0:1, :, :].to(latents.device, dtype=latents.dtype)
else:
inpainting_tensor_mask = inpaint_tensor
# # use our batch latents so we cna avoid ancoding again
inpainting_latent = batch.latents
# resize the mask to match the new encoded size
inpainting_tensor_mask = F.interpolate(inpainting_tensor_mask, size=(inpainting_latent.shape[2], inpainting_latent.shape[3]), mode='bilinear')
inpainting_tensor_mask = inpainting_tensor_mask.to(latents.device, latents.dtype)
do_mask_invert = False
if self.invert_inpaint_mask_chance > 0.0:
do_mask_invert = random.random() < self.invert_inpaint_mask_chance
if do_mask_invert:
# invert the mask
inpainting_tensor_mask = 1 - inpainting_tensor_mask
# mask out the inpainting area, it is currently 0 for inpaint area, and 1 for keep area
# we are zeroing our the latents in the inpaint area not on the pixel space.
inpainting_latent = inpainting_latent * inpainting_tensor_mask
# mask needs to be 1 for inpaint area and 0 for area to leave alone. So flip it.
inpainting_tensor_mask = 1 - inpainting_tensor_mask
# leave the mask as 0-1 and concat on channel of latents
inpainting_latent = torch.cat((inpainting_latent, inpainting_tensor_mask), dim=1)
else:
# we have iinpainting but didnt get a control. or we are doing a dropout
# the input needs to be all zeros for the latents and all 1s for the mask
inpainting_latent = torch.zeros_like(latents)
# add ones for the mask since we are technically inpainting everything
inpainting_latent = torch.cat((inpainting_latent, torch.ones_like(inpainting_latent[:, :1, :, :])), dim=1)
control_tensor = batch.control_tensor
if control_tensor is None:
# concat random normal noise onto the latents
# check dimension, this is before they are rearranged
# it is latent_model_input = torch.cat([latents, control_image], dim=2) after rearranging
ctrl = torch.zeros(
latents.shape[0], # bs
latents.shape[1],
latents.shape[2],
latents.shape[3],
device=latents.device,
dtype=latents.dtype
)
# inpainting always comes first
ctrl = torch.cat((inpainting_latent, ctrl), dim=1)
latents = torch.cat((latents, ctrl), dim=1)
return latents.detach()
# if we have multiple control tensors, they come in like [bs, num_control_images, ch, h, w]
# if we have 1, it comes in like [bs, ch, h, w]
# stack out control tensors to be [bs, ch * num_control_images, h, w]
control_tensor_list = []
if len(control_tensor.shape) == 4:
control_tensor_list.append(control_tensor)
else:
num_control_images = control_tensor.shape[1]
# reshape
control_tensor = control_tensor.view(
control_tensor.shape[0],
control_tensor.shape[1] * control_tensor.shape[2],
control_tensor.shape[3],
control_tensor.shape[4]
)
control_tensor_list = control_tensor.chunk(num_control_images, dim=1)
do_dropout = random.random() < self.control_dropout if self.control_dropout > 0.0 else False
if do_dropout:
# dropout with zeros
control_latent = torch.zeros_like(batch.latents)
else:
# we only have one control so we randomly pick from this list
control_tensor = random.choice(control_tensor_list)
# it is 0-1 need to convert to -1 to 1
control_tensor = control_tensor * 2 - 1
control_tensor = control_tensor.to(self.vae_device_torch, dtype=self.torch_dtype)
# if it is not the size of batch.tensor, (bs,ch,h,w) then we need to resize it
if control_tensor.shape[2] != batch.tensor.shape[2] or control_tensor.shape[3] != batch.tensor.shape[3]:
control_tensor = F.interpolate(control_tensor, size=(batch.tensor.shape[2], batch.tensor.shape[3]), mode='bilinear')
# encode it
control_latent = self.encode_images(control_tensor).to(latents.device, latents.dtype)
# inpainting always comes first
control_latent = torch.cat((inpainting_latent, control_latent), dim=1)
# concat it onto the latents
latents = torch.cat((latents, control_latent), dim=1)
return latents.detach()

View File

@@ -0,0 +1,348 @@
from diffusers import FluxControlPipeline, FluxTransformer2DModel
from typing import Any, Callable, Dict, List, Optional, Union
import torch
from diffusers.image_processor import PipelineImageInput
import numpy as np
from PIL import Image
import torch.nn.functional as F
from torchvision import transforms
from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput
from diffusers.pipelines.flux.pipeline_flux import calculate_shift, retrieve_timesteps, XLA_AVAILABLE
class Flex2Pipeline(FluxControlPipeline):
def __init__(
self,
scheduler,
vae,
text_encoder,
tokenizer,
text_encoder_2,
tokenizer_2,
transformer,
):
super().__init__(scheduler, vae, text_encoder, tokenizer, text_encoder_2, tokenizer_2, transformer)
@torch.no_grad()
def __call__(
self,
prompt: Union[str, List[str]] = None,
prompt_2: Optional[Union[str, List[str]]] = None,
control_image: Optional[PipelineImageInput] = None,
height: Optional[int] = None,
width: Optional[int] = None,
num_inference_steps: int = 28,
sigmas: Optional[List[float]] = None,
guidance_scale: float = 3.5,
num_images_per_prompt: Optional[int] = 1,
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
latents: Optional[torch.FloatTensor] = None,
prompt_embeds: Optional[torch.FloatTensor] = None,
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
output_type: Optional[str] = "pil",
return_dict: bool = True,
joint_attention_kwargs: Optional[Dict[str, Any]] = None,
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
max_sequence_length: int = 512,
control_image_idx: int = 0,
**kwargs,
):
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
control_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
`List[List[torch.Tensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted
as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or
width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`,
images must be passed as a list such that each element of the list can be correctly batched for input
to a single ControlNet.
height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
The height in pixels of the generated image. This is set to 1024 by default for the best results.
width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
The width in pixels of the generated image. This is set to 1024 by default for the best results.
num_inference_steps (`int`, *optional*, defaults to 50):
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
expense of slower inference.
sigmas (`List[float]`, *optional*):
Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
will be used.
guidance_scale (`float`, *optional*, defaults to 3.5):
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,
prompt_embeds=prompt_embeds,
pooled_prompt_embeds=pooled_prompt_embeds,
callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
max_sequence_length=max_sequence_length,
)
self._guidance_scale = guidance_scale
self._joint_attention_kwargs = joint_attention_kwargs
self._interrupt = False
# 2. Define call parameters
if prompt is not None and isinstance(prompt, str):
batch_size = 1
elif prompt is not None and isinstance(prompt, list):
batch_size = len(prompt)
else:
batch_size = prompt_embeds.shape[0]
device = self._execution_device
# 3. Prepare text embeddings
lora_scale = (
self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
)
(
prompt_embeds,
pooled_prompt_embeds,
text_ids,
) = self.encode_prompt(
prompt=prompt,
prompt_2=prompt_2,
prompt_embeds=prompt_embeds,
pooled_prompt_embeds=pooled_prompt_embeds,
device=device,
num_images_per_prompt=num_images_per_prompt,
max_sequence_length=max_sequence_length,
lora_scale=lora_scale,
)
# 4. Prepare latent variables
# num_channels_latents = self.transformer.config.in_channels // 8
num_channels_latents = 128 // 8
# pull mask off control image if there is one it is a pil image
mask = None
if control_image is not None and control_image.mode == "RGBA":
control_img_array = np.array(control_image)
mask = control_img_array[:, :, 3:4]
# scale it to 0 - 1
mask = mask / 255.0
# control image ideally would be a full image here
control_img_array = control_img_array[:, :, :3]
control_image = Image.fromarray(control_img_array.astype(np.uint8))
if control_image is not None:
control_image = self.prepare_image(
image=control_image,
width=width,
height=height,
batch_size=batch_size * num_images_per_prompt,
num_images_per_prompt=num_images_per_prompt,
device=device,
dtype=self.vae.dtype,
)
if control_image.ndim == 4:
num_control_channels = num_channels_latents
control_image = self.vae.encode(control_image).latent_dist.sample(generator=generator)
control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor
if mask is not None:
transform = transforms.Compose([
transforms.ToTensor(),
])
mask = transform(mask).to(device, dtype=control_image.dtype).unsqueeze(0)
# resize mask to match control image
mask = F.interpolate(mask, size=(control_image.shape[2], control_image.shape[3]), mode="bilinear", align_corners=False)
mask = mask.to(device)
# apply the mask to the control image so the inpaint latent area is 0
# mask is currently 0 for inpaint area and 1 for image area
control_image = control_image * mask
# invert mask so it is 1 for inpaint area and 0 for image area
mask = 1 - mask
control_image = torch.cat([control_image, mask], dim=1)
num_control_channels += 1
height_control_image, width_control_image = control_image.shape[2:]
control_image = self._pack_latents(
control_image,
batch_size * num_images_per_prompt,
num_control_channels,
height_control_image,
width_control_image,
)
latents, latent_image_ids = self.prepare_latents(
batch_size * num_images_per_prompt,
num_channels_latents,
height,
width,
prompt_embeds.dtype,
device,
generator,
latents,
)
# 5. Prepare timesteps
sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
image_seq_len = latents.shape[1]
mu = calculate_shift(
image_seq_len,
self.scheduler.config.get("base_image_seq_len", 256),
self.scheduler.config.get("max_image_seq_len", 4096),
self.scheduler.config.get("base_shift", 0.5),
self.scheduler.config.get("max_shift", 1.15),
)
timesteps, num_inference_steps = retrieve_timesteps(
self.scheduler,
num_inference_steps,
device,
sigmas=sigmas,
mu=mu,
)
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
self._num_timesteps = len(timesteps)
# handle guidance
if self.transformer.config.guidance_embeds:
guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
guidance = guidance.expand(latents.shape[0])
else:
guidance = None
# 6. Denoising loop
with self.progress_bar(total=num_inference_steps) as progress_bar:
for i, t in enumerate(timesteps):
if self.interrupt:
continue
# make a blank control latent
control_image_list = [
# impainting
torch.cat([torch.zeros_like(latents), torch.ones_like(latents[:, :, :4])], dim=2),
# control
torch.zeros_like(latents),
]
if control_image is not None:
control_image_list[control_image_idx] = control_image
latent_model_input = torch.cat([latents] + control_image_list, dim=2)
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
timestep = t.expand(latents.shape[0]).to(latents.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]
# 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)

View File

@@ -1675,9 +1675,12 @@ class SDTrainer(BaseSDTrainProcess):
else:
if unconditional_embeds is not None:
unconditional_embeds = unconditional_embeds.to(self.device_torch, dtype=dtype).detach()
if self.adapter and isinstance(self.adapter, CustomAdapter):
with self.timer('condition_noisy_latents'):
with self.timer('condition_noisy_latents'):
# do it for the model
noisy_latents = self.sd.condition_noisy_latents(noisy_latents, batch)
if self.adapter and isinstance(self.adapter, CustomAdapter):
noisy_latents = self.adapter.condition_noisy_latents(noisy_latents, batch)
with self.timer('predict_unet'):
noise_pred = self.predict_noise(
noisy_latents=noisy_latents.to(self.device_torch, dtype=dtype),