Spaces:
Running on Zero
Running on Zero
File size: 7,375 Bytes
81157d8 7e06453 81157d8 7e06453 81157d8 7e06453 81157d8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 | import argparse
import logging
import os
import sys
import torch
from diffusers import FluxTransformer2DModel
from diffusers.configuration_utils import FrozenDict
PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
if PROJECT_ROOT not in sys.path:
sys.path.insert(0, PROJECT_ROOT)
logging.getLogger("transformers.tokenization_utils_base").setLevel(logging.ERROR)
os.environ["CUDA_VISIBLE_DEVICES"] = os.environ.get("CUDA_VISIBLE_DEVICES", "0")
from models.mmdit import CustomFluxTransformer2DModel
from models.multiLayer_adapter import MultiLayerAdapter
from models.pipeline import CustomFluxPipeline, CustomFluxPipelineCfgLayer
from models.transp_vae import AutoencoderKLTransformerTraining as CustomVAE
def resolve_local_config_path(path: str | None) -> str | None:
"""Resolve project-relative asset paths while leaving absolute paths untouched."""
if not path:
return path
if os.path.isabs(path):
return path
return os.path.join(PROJECT_ROOT, path)
def load_trusted_checkpoint(path: str, map_location=None):
"""Load SynLayers-owned checkpoints saved before PyTorch changed weights_only defaults."""
try:
return torch.load(path, map_location=map_location, weights_only=False)
except TypeError:
return torch.load(path, map_location=map_location)
def scale_box_xyxy(box, source_size: int, target_size: int) -> tuple:
"""Scale an xyxy box from source_size to target_size."""
scale = target_size / source_size
x0, y0, x1, y1 = box
x0_s = int(x0 * scale)
y0_s = int(y0 * scale)
x1_s = int(x1 * scale)
y1_s = int(y1 * scale)
x0_s = max(0, x0_s)
y0_s = max(0, y0_s)
x1_s = min(target_size, x1_s)
y1_s = min(target_size, y1_s)
return (x0_s, y0_s, x1_s, y1_s)
def quantize_box_16(box: tuple, target_size: int) -> tuple:
"""Quantize an xyxy box to the 16-pixel latent grid."""
x0, y0, x1, y1 = box
x0_q = (x0 // 16) * 16
y0_q = (y0 // 16) * 16
x1_q = ((x1 + 15) // 16) * 16
y1_q = ((y1 + 15) // 16) * 16
x0_q = max(0, x0_q)
y0_q = max(0, y0_q)
x1_q = min(target_size, x1_q)
y1_q = min(target_size, y1_q)
return (x0_q, y0_q, x1_q, y1_q)
def get_layer_boxes(layers: list, source_size: int, target_size: int) -> list:
"""Extract, scale, and quantize prism layer boxes."""
boxes = []
for layer in layers:
box = layer.get("box", [0, 0, source_size, source_size])
scaled_box = scale_box_xyxy(box, source_size, target_size)
boxes.append(quantize_box_16(scaled_box, target_size))
return boxes
def initialize_pipeline(config):
"""Initialize the SynLayers decomposition pipeline."""
transp_vae_path = resolve_local_config_path(config.get("transp_vae_path"))
pretrained_lora_dir = resolve_local_config_path(config.get("pretrained_lora_dir"))
artplus_lora_dir = resolve_local_config_path(config.get("artplus_lora_dir"))
layer_ckpt = resolve_local_config_path(config.get("layer_ckpt"))
adapter_lora_dir = resolve_local_config_path(config.get("adapter_lora_dir"))
lora_ckpt = resolve_local_config_path(config.get("lora_ckpt"))
print("[INFO] Loading Transparent VAE...", flush=True)
vae_args = argparse.Namespace(
max_layers=config.get("max_layers", 48),
decoder_arch=config.get("decoder_arch", "vit"),
pos_embedding=config.get("pos_embedding", "rope"),
layer_embedding=config.get("layer_embedding", "rope"),
single_layer_decoder=config.get("single_layer_decoder", None),
)
transp_vae = CustomVAE(vae_args)
transp_vae_weights = load_trusted_checkpoint(
transp_vae_path, map_location=torch.device("cuda")
)
missing_keys, unexpected_keys = transp_vae.load_state_dict(
transp_vae_weights["model"], strict=False
)
if missing_keys:
print(f"ViT Encoder Missing keys: {missing_keys}")
if unexpected_keys:
print(f"ViT Encoder Unexpected keys: {unexpected_keys}")
transp_vae.eval()
transp_vae = transp_vae.to(torch.device("cuda"))
print("[INFO] Transparent VAE loaded.", flush=True)
print("[INFO] Loading pretrained Transformer model...", flush=True)
transformer_orig = FluxTransformer2DModel.from_pretrained(
config.get("transformer_varient", config["pretrained_model_name_or_path"]),
subfolder="" if "transformer_varient" in config else "transformer",
revision=config.get("revision", None),
variant=config.get("variant", None),
torch_dtype=torch.bfloat16,
cache_dir=config.get("cache_dir", None),
)
mmdit_config = dict(transformer_orig.config)
mmdit_config["_class_name"] = "CustomSD3Transformer2DModel"
mmdit_config["max_layer_num"] = config["max_layer_num"]
mmdit_config = FrozenDict(mmdit_config)
transformer = CustomFluxTransformer2DModel.from_config(mmdit_config).to(
dtype=torch.bfloat16
)
transformer.load_state_dict(transformer_orig.state_dict(), strict=False)
if pretrained_lora_dir:
print("[INFO] Loading pretrained LoRA weights...", flush=True)
lora_state_dict = CustomFluxPipeline.lora_state_dict(pretrained_lora_dir)
CustomFluxPipeline.load_lora_into_transformer(lora_state_dict, None, transformer)
transformer.fuse_lora(safe_fusing=True)
transformer.unload_lora()
if artplus_lora_dir:
print("[INFO] Loading artplus LoRA weights...", flush=True)
lora_state_dict = CustomFluxPipeline.lora_state_dict(artplus_lora_dir)
CustomFluxPipeline.load_lora_into_transformer(lora_state_dict, None, transformer)
transformer.fuse_lora(safe_fusing=True)
transformer.unload_lora()
layer_pe_path = os.path.join(layer_ckpt, "layer_pe.pth") if layer_ckpt else ""
if os.path.exists(layer_pe_path):
print(f"[INFO] Loading layer_pe from {layer_pe_path}...", flush=True)
layer_pe = load_trusted_checkpoint(layer_pe_path, map_location=torch.device("cuda"))
transformer.load_state_dict(layer_pe, strict=False)
print("[INFO] Loading MultiLayer-Adapter...", flush=True)
multiLayer_adapter = MultiLayerAdapter.from_pretrained(
config["pretrained_adapter_path"]
).to(torch.bfloat16).to(torch.device("cuda"))
if adapter_lora_dir:
print("[INFO] Loading adapter LoRA weights...", flush=True)
lora_state_dict = CustomFluxPipeline.lora_state_dict(adapter_lora_dir)
CustomFluxPipeline.load_lora_into_transformer(
lora_state_dict, None, multiLayer_adapter
)
multiLayer_adapter.fuse_lora(safe_fusing=True)
multiLayer_adapter.unload_lora()
multiLayer_adapter.set_layerPE(transformer.layer_pe, transformer.max_layer_num)
pipeline = CustomFluxPipelineCfgLayer.from_pretrained(
config["pretrained_model_name_or_path"],
transformer=transformer,
revision=config.get("revision", None),
variant=config.get("variant", None),
torch_dtype=torch.bfloat16,
cache_dir=config.get("cache_dir", None),
).to(torch.device("cuda"))
pipeline.set_multiLayerAdapter(multiLayer_adapter)
if lora_ckpt:
print(f"[INFO] Loading trained LoRA from {lora_ckpt}...", flush=True)
pipeline.load_lora_weights(lora_ckpt, adapter_name="layer")
return pipeline, transp_vae
|