398 lines
14 KiB
Python
398 lines
14 KiB
Python
import os
|
|
from typing import List, Optional
|
|
|
|
import huggingface_hub
|
|
import torch
|
|
from toolkit.config_modules import GenerateImageConfig, ModelConfig, NetworkConfig
|
|
from toolkit.lora_special import LoRASpecialNetwork
|
|
from toolkit.models.base_model import BaseModel
|
|
from toolkit.basic import flush
|
|
from toolkit.prompt_utils import PromptEmbeds
|
|
from toolkit.samplers.custom_flowmatch_sampler import (
|
|
CustomFlowMatchEulerDiscreteScheduler,
|
|
)
|
|
from toolkit.accelerator import unwrap_model
|
|
from toolkit.memory_management import MemoryManager
|
|
from toolkit.metadata import get_meta_for_safetensors
|
|
from toolkit.models.v2.text_encoders.qwen3 import Qwen3TextEncoder
|
|
from toolkit.models.v2.vae.autoencoder_kl import KLVAE
|
|
from safetensors.torch import load_file
|
|
|
|
|
|
try:
|
|
from diffusers import ZImagePipeline
|
|
|
|
# our subclass of the diffusers transformer with the universal loading /
|
|
# quantization mixin (see toolkit/models/v2/_mixin.py)
|
|
from toolkit.models.v2.diffusion_models.z_image import ZImageTransformer2DModel
|
|
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"Diffusers is out of date.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 = {
|
|
"num_train_timesteps": 1000,
|
|
"use_dynamic_shifting": False,
|
|
"shift": 3.0,
|
|
}
|
|
|
|
# repo to pull the vae / text encoder / tokenizer / config from when loading a
|
|
# single-file checkpoint
|
|
SINGLE_FILE_EXTRAS_REPO = "Tongyi-MAI/Z-Image-Turbo"
|
|
|
|
|
|
class ZImageModel(BaseModel):
|
|
arch = "zimage"
|
|
|
|
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 = ["ZImageTransformer2DModel"]
|
|
# set to True when loading a single-file checkpoint so we save back in that format
|
|
self.is_single_file = False
|
|
# repo to pull config/vae/te from when loading a single-file checkpoint
|
|
self.single_file_extras_repo = SINGLE_FILE_EXTRAS_REPO
|
|
|
|
# static method to get the noise scheduler
|
|
@staticmethod
|
|
def get_train_scheduler():
|
|
return CustomFlowMatchEulerDiscreteScheduler(**scheduler_config)
|
|
|
|
def get_bucket_divisibility(self):
|
|
return 8 * 2 # 8 for the VAE, 2 for patch size
|
|
|
|
def load_training_adapter(self, transformer: ZImageTransformer2DModel):
|
|
self.print_and_status_update("Loading assistant LoRA")
|
|
lora_path = self.model_config.assistant_lora_path
|
|
if not os.path.exists(lora_path):
|
|
# assume it is a hub path
|
|
lora_splits = lora_path.split("/")
|
|
if len(lora_splits) != 3:
|
|
raise ValueError(
|
|
f"Assistant LoRA path {lora_path} is not a valid local path or hub path."
|
|
)
|
|
repo_id = "/".join(lora_splits[:2])
|
|
filename = lora_splits[2]
|
|
try:
|
|
lora_path = huggingface_hub.hf_hub_download(
|
|
repo_id=repo_id,
|
|
filename=filename,
|
|
)
|
|
# upgrade path to
|
|
self.model_config.assistant_lora_path = lora_path
|
|
except Exception as e:
|
|
raise ValueError(
|
|
f"Failed to download assistant LoRA from {lora_path}: {e}"
|
|
)
|
|
# load the adapter and merge it in. We will inference with a -1.0 multiplier so the adapter effects only work during training.
|
|
lora_state_dict = load_file(lora_path)
|
|
dim = int(
|
|
lora_state_dict[
|
|
"diffusion_model.layers.0.attention.to_k.lora_A.weight"
|
|
].shape[0]
|
|
)
|
|
|
|
new_sd = {}
|
|
for key, value in lora_state_dict.items():
|
|
new_key = key.replace("diffusion_model.", "transformer.")
|
|
new_sd[new_key] = value
|
|
lora_state_dict = new_sd
|
|
|
|
network_config = {
|
|
"type": "lora",
|
|
"linear": dim,
|
|
"linear_alpha": dim,
|
|
"transformer_only": True,
|
|
}
|
|
|
|
network_config = NetworkConfig(**network_config)
|
|
LoRASpecialNetwork.LORA_PREFIX_UNET = "lora_transformer"
|
|
network = LoRASpecialNetwork(
|
|
text_encoder=None,
|
|
unet=transformer,
|
|
lora_dim=network_config.linear,
|
|
multiplier=1.0,
|
|
alpha=network_config.linear_alpha,
|
|
train_unet=True,
|
|
train_text_encoder=False,
|
|
network_config=network_config,
|
|
network_type=network_config.type,
|
|
transformer_only=network_config.transformer_only,
|
|
is_transformer=True,
|
|
target_lin_modules=self.target_lora_modules,
|
|
is_assistant_adapter=True,
|
|
is_ara=True,
|
|
)
|
|
network.apply_to(None, transformer, apply_text_encoder=False, apply_unet=True)
|
|
self.print_and_status_update("Merging in assistant LoRA")
|
|
network.force_to(self.device_torch, dtype=self.torch_dtype)
|
|
network._update_torch_multiplier()
|
|
network.load_weights(lora_state_dict)
|
|
|
|
network.merge_in(merge_weight=1.0)
|
|
|
|
# mark it as not merged so inference ignores it.
|
|
network.is_merged_in = False
|
|
|
|
# add the assistant so sampler will activate it while sampling
|
|
self.assistant_lora: LoRASpecialNetwork = network
|
|
|
|
# deactivate lora during training
|
|
self.assistant_lora.multiplier = -1.0
|
|
self.assistant_lora.is_active = False
|
|
|
|
# tell the model to invert assistant on inference since we want remove lora effects
|
|
self.invert_assistant_lora = True
|
|
|
|
def load_transformer(self, model_path, base_model_path, dtype):
|
|
"""Load the ZImage transformer through the OstrisModelMixin universal loader
|
|
(diffusers folder/repo, local or hub single-file checkpoint, pre-quantized
|
|
checkpoint). Returns (transformer, base_model_path) since the base path may
|
|
be redirected to the hub repo for single-file checkpoints."""
|
|
if model_path.endswith(".safetensors"):
|
|
# single-file checkpoint. Load the weights from the file and pull the
|
|
# vae / text encoder / tokenizer / config from the base diffusers repo.
|
|
self.is_single_file = True
|
|
self.print_and_status_update(
|
|
"Model is a single-file checkpoint, loading with safetensors"
|
|
)
|
|
if base_model_path == model_path:
|
|
# extras default to name_or_path which is the single file, fall back to the hub repo
|
|
base_model_path = self.single_file_extras_repo
|
|
elif os.path.exists(model_path):
|
|
# 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
|
|
|
|
# quantization happens inside load_model unless an adapter has to be merged
|
|
# into the full precision weights first (assistant lora / accuracy recovery
|
|
# adapter); those paths quantize after the merge via quantize_model
|
|
transformer = ZImageTransformer2DModel.load_model(
|
|
model_path,
|
|
dtype=dtype,
|
|
config_path=base_model_path if self.is_single_file else None,
|
|
use_comfy_weights=self.model_config.model_kwargs.get(
|
|
"use_comfy_weights", True
|
|
),
|
|
# comfy candidate ranking hint only (aitk_post_load quantizes):
|
|
# a matching pre-quantized file loads with no requant work
|
|
qtype=self.model_config.qtype if self.model_config.quantize else None,
|
|
quantize_on_load=False,
|
|
)
|
|
flush()
|
|
|
|
return transformer, base_model_path
|
|
|
|
def load_model(self):
|
|
dtype = self.torch_dtype
|
|
self.print_and_status_update("Loading ZImage 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, base_model_path = self.load_transformer(
|
|
model_path, base_model_path, dtype
|
|
)
|
|
|
|
# load assistant lora if specified
|
|
if self.model_config.assistant_lora_path is not None:
|
|
self.load_training_adapter(transformer)
|
|
# set qtype to be float8 if it is qfloat8
|
|
if self.model_config.qtype == "qfloat8":
|
|
self.model_config.qtype = "float8"
|
|
|
|
# quantize + offload + placement, all driven by model_config
|
|
transformer.aitk_post_load(**self.component_load_kwargs("transformer"))
|
|
flush()
|
|
|
|
self.print_and_status_update("Text Encoder")
|
|
tokenizer = Qwen3TextEncoder.load_tokenizer(base_model_path)
|
|
text_encoder = Qwen3TextEncoder.load(
|
|
base_model_path, **self.component_load_kwargs("te")
|
|
)
|
|
|
|
self.print_and_status_update("Loading VAE")
|
|
vae = KLVAE.load_model(base_model_path, dtype=dtype)
|
|
|
|
self.noise_scheduler = ZImageModel.get_train_scheduler()
|
|
|
|
self.print_and_status_update("Making pipe")
|
|
|
|
kwargs = {}
|
|
|
|
pipe: ZImagePipeline = ZImagePipeline(
|
|
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 = ZImageModel.get_train_scheduler()
|
|
|
|
pipeline: ZImagePipeline = ZImagePipeline(
|
|
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 generate_single_image(
|
|
self,
|
|
pipeline: ZImagePipeline,
|
|
gen_config: GenerateImageConfig,
|
|
conditional_embeds: PromptEmbeds,
|
|
unconditional_embeds: PromptEmbeds,
|
|
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)
|
|
|
|
# CFG is 0 normalized for this model
|
|
guidance = max(0.0, gen_config.guidance_scale - 1.0)
|
|
|
|
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=guidance,
|
|
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: PromptEmbeds,
|
|
**kwargs,
|
|
):
|
|
if self.model.device == torch.device("cpu"):
|
|
self.model.to(self.device_torch)
|
|
|
|
latent_model_input = latent_model_input.unsqueeze(2)
|
|
latent_model_input_list = list(latent_model_input.unbind(dim=0))
|
|
|
|
timestep_model_input = (1000 - timestep) / 1000
|
|
|
|
model_out_list = self.transformer(
|
|
latent_model_input_list,
|
|
timestep_model_input,
|
|
text_embeddings.text_embeds,
|
|
)[0]
|
|
|
|
noise_pred = torch.stack([t.float() for t in model_out_list], dim=0)
|
|
|
|
noise_pred = noise_pred.squeeze(2)
|
|
noise_pred = -noise_pred
|
|
|
|
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, _ = self.pipeline.encode_prompt(
|
|
prompt,
|
|
do_classifier_free_guidance=False,
|
|
device=self.device_torch,
|
|
)
|
|
pe = PromptEmbeds([prompt_embeds, None])
|
|
return pe
|
|
|
|
def get_model_has_grad(self):
|
|
return False
|
|
|
|
def get_te_has_grad(self):
|
|
return False
|
|
|
|
def get_quantization_exclude_modules(self):
|
|
# the patterns live on the transformer class so the mixin quantization
|
|
# uses them too (see toolkit/models/classes/z_image.py)
|
|
return ZImageTransformer2DModel.get_quantization_exclude_modules()
|
|
|
|
def save_model(self, output_path, meta, save_dtype):
|
|
# comfy-format single-file save (the standard save format regardless
|
|
# of how the model was loaded); prequantized layers keep their
|
|
# quantized storage
|
|
transformer: ZImageTransformer2DModel = unwrap_model(self.model)
|
|
if not output_path.endswith(".safetensors"):
|
|
output_path += ".safetensors"
|
|
transformer.save_model(
|
|
output_path,
|
|
dtype=save_dtype,
|
|
metadata=get_meta_for_safetensors(meta, name=self.arch),
|
|
)
|
|
|
|
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 "zimage"
|
|
|
|
def get_transformer_block_names(self) -> Optional[List[str]]:
|
|
return ["layers"]
|
|
|
|
lora_keys_use_comfy_prefix = True
|
|
|