382 lines
12 KiB
Python
382 lines
12 KiB
Python
import os
|
|
from typing import List, Optional
|
|
|
|
import torch
|
|
import yaml
|
|
from toolkit.config_modules import GenerateImageConfig, ModelConfig
|
|
from toolkit.models.base_model import BaseModel
|
|
from toolkit.basic import flush
|
|
from toolkit.advanced_prompt_embeds import AdvancedPromptEmbeds
|
|
from toolkit.samplers.custom_flowmatch_sampler import (
|
|
CustomFlowMatchEulerDiscreteScheduler,
|
|
)
|
|
from toolkit.accelerator import unwrap_model
|
|
from optimum.quanto import freeze
|
|
from toolkit.util.quantize import quantize, get_qtype, quantize_model
|
|
from toolkit.memory_management import MemoryManager
|
|
|
|
from transformers import AutoTokenizer, AutoModel
|
|
|
|
try:
|
|
from diffusers import ErnieImagePipeline, AutoencoderKLFlux2
|
|
from .transformer import ErnieImageTransformer2DModel
|
|
except ImportError:
|
|
raise ImportError(
|
|
"Diffusers is out of date. Update diffusers to the latest version by doing pip uninstall diffusers and then pip install -r requirements.txt"
|
|
)
|
|
|
|
|
|
scheduler_config = {
|
|
"base_image_seq_len": 256,
|
|
"base_shift": 0.5,
|
|
"invert_sigmas": False,
|
|
"max_image_seq_len": 4096,
|
|
"max_shift": 1.15,
|
|
"num_train_timesteps": 1000,
|
|
"shift": 3.0,
|
|
"shift_terminal": None,
|
|
"stochastic_sampling": False,
|
|
"time_shift_type": "exponential",
|
|
"use_beta_sigmas": False,
|
|
"use_dynamic_shifting": False,
|
|
"use_exponential_sigmas": False,
|
|
"use_karras_sigmas": False,
|
|
}
|
|
|
|
|
|
class ErnieImageModel(BaseModel):
|
|
arch = "ernie_image"
|
|
|
|
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 = ["ErnieImageTransformer2DModel"]
|
|
|
|
# static method to get the noise scheduler
|
|
@staticmethod
|
|
def get_train_scheduler():
|
|
return CustomFlowMatchEulerDiscreteScheduler(**scheduler_config)
|
|
|
|
def get_bucket_divisibility(self):
|
|
return 16 * 2 # 16 for the VAE, 2 for patch size
|
|
|
|
def load_model(self):
|
|
dtype = self.torch_dtype
|
|
self.print_and_status_update("Loading ErnieImage model")
|
|
model_path = self.model_config.name_or_path
|
|
base_model_path = self.model_config.extras_name_or_path
|
|
|
|
self.print_and_status_update("Loading transformer")
|
|
|
|
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
|
|
|
|
transformer = ErnieImageTransformer2DModel.from_pretrained(
|
|
transformer_path, subfolder=transformer_subfolder, torch_dtype=dtype
|
|
)
|
|
|
|
if self.model_config.quantize:
|
|
self.print_and_status_update("Quantizing Transformer")
|
|
quantize_model(self, transformer)
|
|
flush()
|
|
|
|
if (
|
|
self.model_config.layer_offloading
|
|
and self.model_config.layer_offloading_transformer_percent > 0
|
|
):
|
|
MemoryManager.attach(
|
|
transformer,
|
|
self.device_torch,
|
|
offload_percent=self.model_config.layer_offloading_transformer_percent,
|
|
ignore_modules=[
|
|
transformer.x_embedder,
|
|
],
|
|
)
|
|
|
|
if self.model_config.low_vram:
|
|
self.print_and_status_update("Moving transformer to CPU")
|
|
transformer.to("cpu")
|
|
|
|
flush()
|
|
|
|
self.print_and_status_update("Text Encoder")
|
|
tokenizer = AutoTokenizer.from_pretrained(
|
|
base_model_path, subfolder="tokenizer", torch_dtype=dtype
|
|
)
|
|
text_encoder = AutoModel.from_pretrained(
|
|
base_model_path, subfolder="text_encoder", torch_dtype=dtype
|
|
)
|
|
|
|
if (
|
|
self.model_config.layer_offloading
|
|
and self.model_config.layer_offloading_text_encoder_percent > 0
|
|
):
|
|
MemoryManager.attach(
|
|
text_encoder,
|
|
self.device_torch,
|
|
offload_percent=self.model_config.layer_offloading_text_encoder_percent,
|
|
)
|
|
|
|
text_encoder.to(self.device_torch, dtype=dtype)
|
|
flush()
|
|
|
|
if self.model_config.quantize_te:
|
|
self.print_and_status_update("Quantizing Text Encoder")
|
|
quantize(text_encoder, weights=get_qtype(self.model_config.qtype_te))
|
|
freeze(text_encoder)
|
|
flush()
|
|
|
|
self.print_and_status_update("Loading VAE")
|
|
vae = AutoencoderKLFlux2.from_pretrained(
|
|
base_model_path, subfolder="vae", torch_dtype=dtype
|
|
).to(self.device_torch, dtype=dtype)
|
|
|
|
self.noise_scheduler = ErnieImageModel.get_train_scheduler()
|
|
|
|
self.print_and_status_update("Making pipe")
|
|
|
|
kwargs = {}
|
|
|
|
pipe: ErnieImagePipeline = ErnieImagePipeline(
|
|
scheduler=self.noise_scheduler,
|
|
text_encoder=None,
|
|
tokenizer=tokenizer,
|
|
vae=vae,
|
|
transformer=None,
|
|
**kwargs,
|
|
)
|
|
# for quantization, it works best to do these after making the pipe
|
|
pipe.text_encoder = text_encoder
|
|
pipe.transformer = transformer
|
|
|
|
self.print_and_status_update("Preparing Model")
|
|
|
|
text_encoder = [pipe.text_encoder]
|
|
tokenizer = [pipe.tokenizer]
|
|
|
|
# leave it on cpu for now
|
|
if not self.low_vram:
|
|
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()
|
|
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 = ErnieImageModel.get_train_scheduler()
|
|
|
|
pipeline: ErnieImagePipeline = ErnieImagePipeline(
|
|
scheduler=scheduler,
|
|
text_encoder=unwrap_model(self.text_encoder[0]),
|
|
tokenizer=self.tokenizer[0],
|
|
vae=unwrap_model(self.vae),
|
|
transformer=unwrap_model(self.transformer),
|
|
)
|
|
|
|
pipeline = pipeline.to(self.device_torch)
|
|
|
|
return pipeline
|
|
|
|
def encode_images(self, image_list: List[torch.Tensor], device=None, dtype=None):
|
|
if self.vae.device == torch.device("cpu"):
|
|
self.vae.to(self.device_torch)
|
|
if device is None:
|
|
device = self.vae_device_torch
|
|
if dtype is None:
|
|
dtype = self.vae_torch_dtype
|
|
self.vae.eval()
|
|
self.vae.requires_grad_(False)
|
|
|
|
image = image_list
|
|
if isinstance(image, list):
|
|
image = torch.stack(image, dim=0)
|
|
|
|
image = image.to(device, dtype=dtype)
|
|
|
|
latents = self.vae.encode(image).latent_dist.sample()
|
|
|
|
latents = self.pipeline._patchify_latents(latents)
|
|
|
|
bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(
|
|
device=latents.device, dtype=latents.dtype
|
|
)
|
|
bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + 1e-5).to(
|
|
device=latents.device, dtype=latents.dtype
|
|
)
|
|
latents = (latents - bn_mean) / bn_std
|
|
|
|
return latents
|
|
|
|
def decode_latents(self, latents: torch.Tensor, device=None, dtype=None):
|
|
if self.vae.device == torch.device("cpu"):
|
|
self.vae.to(self.device_torch)
|
|
if device is None:
|
|
device = self.vae_device_torch
|
|
if dtype is None:
|
|
dtype = self.vae_torch_dtype
|
|
|
|
latents = latents.to(device, dtype=dtype)
|
|
bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(device)
|
|
bn_std = torch.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + 1e-5).to(device)
|
|
latents = latents * bn_std + bn_mean
|
|
|
|
# Unpatchify
|
|
latents = self.pipeline._unpatchify_latents(latents)
|
|
|
|
# Decode
|
|
images = self.vae.decode(latents, return_dict=False)[0]
|
|
return images
|
|
|
|
def generate_single_image(
|
|
self,
|
|
pipeline: ErnieImagePipeline,
|
|
gen_config: GenerateImageConfig,
|
|
conditional_embeds: AdvancedPromptEmbeds,
|
|
unconditional_embeds: AdvancedPromptEmbeds,
|
|
generator: torch.Generator,
|
|
extra: dict,
|
|
):
|
|
if self.model.device == torch.device("cpu"):
|
|
self.model.to(self.device_torch)
|
|
|
|
sc = self.get_bucket_divisibility()
|
|
gen_config.width = int(gen_config.width // sc * sc)
|
|
gen_config.height = int(gen_config.height // sc * sc)
|
|
|
|
img = pipeline(
|
|
prompt_embeds=conditional_embeds.text_embeds,
|
|
negative_prompt_embeds=unconditional_embeds.text_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,
|
|
**extra,
|
|
).images[0]
|
|
return img
|
|
|
|
def get_noise_prediction(
|
|
self,
|
|
latent_model_input: torch.Tensor,
|
|
timestep: torch.Tensor, # 0 to 1000 scale
|
|
text_embeddings: AdvancedPromptEmbeds,
|
|
**kwargs,
|
|
):
|
|
if self.model.device == torch.device("cpu"):
|
|
self.model.to(self.device_torch)
|
|
|
|
text_bth, text_lens = self.pipeline._pad_text(
|
|
text_hiddens=text_embeddings.text_embeds,
|
|
device=self.device_torch,
|
|
dtype=self.vae.dtype,
|
|
text_in_dim=self.pipeline.transformer.config.text_in_dim,
|
|
)
|
|
|
|
pred = self.transformer(
|
|
hidden_states=latent_model_input,
|
|
timestep=timestep,
|
|
text_bth=text_bth,
|
|
text_lens=text_lens,
|
|
return_dict=False,
|
|
)[0]
|
|
|
|
return pred
|
|
|
|
def get_prompt_embeds(self, prompt: str) -> AdvancedPromptEmbeds:
|
|
if self.pipeline.text_encoder.device == torch.device("cpu"):
|
|
self.pipeline.text_encoder.to(self.device_torch)
|
|
|
|
if isinstance(prompt, str):
|
|
prompt = [prompt]
|
|
|
|
text_hiddens = []
|
|
|
|
for p in prompt:
|
|
ids = self.pipeline.tokenizer(
|
|
p,
|
|
add_special_tokens=True,
|
|
truncation=True,
|
|
padding=False,
|
|
)["input_ids"]
|
|
|
|
if len(ids) == 0:
|
|
if self.pipeline.tokenizer.bos_token_id is not None:
|
|
ids = [self.pipeline.tokenizer.bos_token_id]
|
|
else:
|
|
ids = [0]
|
|
|
|
input_ids = torch.tensor([ids], device=self.device_torch)
|
|
outputs = self.pipeline.text_encoder(
|
|
input_ids=input_ids,
|
|
output_hidden_states=True,
|
|
)
|
|
# Use second to last hidden state (matches training)
|
|
hidden = outputs.hidden_states[-2][0] # [T, H]
|
|
|
|
text_hiddens.append(hidden)
|
|
|
|
pe = AdvancedPromptEmbeds(text_embeds=text_hiddens)
|
|
return pe
|
|
|
|
def get_model_has_grad(self):
|
|
return False
|
|
|
|
def get_te_has_grad(self):
|
|
return False
|
|
|
|
def save_model(self, output_path, meta, save_dtype):
|
|
transformer: ErnieImageTransformer2DModel = 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 get_base_model_version(self):
|
|
return self.arch
|
|
|
|
def get_transformer_block_names(self) -> Optional[List[str]]:
|
|
return ["layers"]
|
|
|
|
lora_keys_use_comfy_prefix = True
|
|
|