File size: 9,873 Bytes
b9fc164 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 | import os
import glob
import torch
import torch.distributed as dist
from collections.abc import Sequence
from src.utils.fsdp_load import maybe_load_fsdp_model, pt_weights_iterator, safetensors_weights_iterator
from src.utils.logging import get_logger
from src.utils.lora import validate_lora_state_dict
from src.utils.constants import PRECISION_TO_TYPE
from src.utils.utils import build_from_config, get_obj_from_str
def _uses_pretrained_loading(config: dict | None) -> bool:
if config is None or config.get("pretrained") is None:
return False
target = str(config.get("target", ""))
try:
cls = get_obj_from_str(target)
except Exception:
return target.startswith("diffusers.")
return hasattr(cls, "from_pretrained")
def _strip_common_state_prefixes(state_dict: dict[str, torch.Tensor]) -> dict[str, torch.Tensor]:
for prefix in ("model.", "module.", "transformer."):
stripped_state_dict = {}
matched = 0
for key, value in state_dict.items():
if key.startswith(prefix):
stripped_state_dict[key[len(prefix):]] = value
matched += 1
else:
stripped_state_dict[key] = value
if matched > 0:
state_dict = stripped_state_dict
return state_dict
def _normalize_train_modules(train_modules: str | Sequence[str] | None) -> list[str] | None:
if train_modules is None:
return None
if isinstance(train_modules, str):
train_modules = [train_modules]
normalized = [module_name.strip() for module_name in train_modules if module_name and module_name.strip()]
return normalized or None
def _configure_dit_trainable_modules(model: torch.nn.Module, cfg) -> None:
logger = get_logger()
train_modules = _normalize_train_modules(getattr(cfg, "dit_train_modules", None))
if train_modules is None:
model.requires_grad_(True)
return
model.requires_grad_(False)
named_modules = dict(model.named_modules())
missing_modules = [module_name for module_name in train_modules if module_name not in named_modules]
if missing_modules:
available_modules = [module_name for module_name in named_modules.keys() if module_name]
preview = ", ".join(available_modules[:20])
if len(available_modules) > 20:
preview = f"{preview}, ..."
raise ValueError(
f"Unknown dit_train_modules: {missing_modules}. "
f"Available DiT modules include: {preview}"
)
for module_name in train_modules:
named_modules[module_name].requires_grad_(True)
trainable_param_names = [name for name, param in model.named_parameters() if param.requires_grad]
if not trainable_param_names:
raise ValueError(f"No trainable DiT parameters were enabled for dit_train_modules={train_modules}.")
logger.info(f"Training {len(trainable_param_names)} parameter tensors.")
def _load_vae(cfg, device: torch.device):
"""Load VAE weights and initialize complete modules absent from the checkpoint."""
vae_kwargs = {"torch_dtype": PRECISION_TO_TYPE[cfg.vae_precision], "device": device}
if not _uses_pretrained_loading(cfg.vae_arch_config):
return build_from_config(cfg.vae_arch_config, **vae_kwargs)
vae, loading = build_from_config(
cfg.vae_arch_config, **vae_kwargs, low_cpu_mem_usage=False, output_loading_info=True,
)
missing_by_module = {}
for key in loading["missing_keys"]:
module_name, _, parameter_name = key.rpartition(".")
missing_by_module.setdefault(module_name, set()).add(parameter_name)
for module_name, missing in sorted(missing_by_module.items()):
module = vae.get_submodule(module_name)
parameters = set(dict(module.named_parameters(recurse=False)))
reset = getattr(module, "reset_parameters", None)
if missing != parameters or not callable(reset):
raise ValueError(f"Cannot initialize incomplete VAE module {module_name!r}: missing={sorted(missing)}.")
reset()
return vae
def load_pipeline(cfg, dit, device: torch.device):
"""Build the frozen VAE, text encoder, scheduler, and editing pipeline."""
vae = _load_vae(cfg, device).eval().requires_grad_(False)
tokenizer, text_encoder = build_from_config(
cfg.text_encoder_arch_config, torch_dtype=PRECISION_TO_TYPE[cfg.text_encoder_precision], device=device,
)
pipeline = build_from_config(
cfg.pipeline_arch_config or {"target": "src.models.common.diffusion.pipelines.Pipeline"},
vae=vae,
tokenizer=tokenizer,
text_encoder=text_encoder,
transformer=dit,
scheduler=build_from_config(cfg.scheduler_arch_config),
args=cfg,
)
return pipeline.to(device)
def _read_dit_checkpoint(path, checkpoint_type):
if checkpoint_type == "pt":
state = dict(pt_weights_iterator([path]))
elif checkpoint_type == "safetensor":
files = [str(path)] if os.path.isfile(path) else sorted(glob.glob(os.path.join(str(path), "*.safetensors")))
if not files:
raise FileNotFoundError(f"No safetensors checkpoints found at {path}.")
state = dict(safetensors_weights_iterator(files))
else:
raise ValueError(f"Unknown checkpoint type: {checkpoint_type!r}.")
return state.get("model", state)
def _merge_generator_lora(model, checkpoint, cfg):
from peft import LoraConfig
from peft.utils import set_peft_model_state_dict
weights = checkpoint["generator_lora"]
adapter_config = checkpoint.get("generator_lora_config")
if adapter_config is None:
ranks = {weight.shape[0] for name, weight in weights.items() if ".lora_A." in name}
if len(ranks) != 1:
raise ValueError("The generator checkpoint must contain a uniform-rank LoRA adapter.")
rank = ranks.pop()
adapter_config = {
"r": rank,
"lora_alpha": getattr(cfg, "dmd_lora_alpha", rank),
"target_modules": sorted({name.split(".lora_")[0] for name in weights if ".lora_" in name}),
"bias": getattr(cfg, "dmd_lora_bias", "none"),
}
adapter_name = checkpoint.get("generator_adapter_name", "generator")
model.add_adapter(LoraConfig(**adapter_config), adapter_name=adapter_name)
validate_lora_state_dict(model, weights, adapter_name)
result = set_peft_model_state_dict(model, weights, adapter_name=adapter_name)
missing_adapter_keys = [name for name in result.missing_keys if ".lora_" in name]
if missing_adapter_keys or result.unexpected_keys:
raise ValueError("Generator LoRA weights do not match the configured model.")
model.fuse_lora(adapter_names=[adapter_name], safe_fusing=True)
model.unload_lora()
def load_dit(cfg, device: torch.device, *, use_fsdp: bool = True) -> torch.nn.Module:
"""Load DiT model with FSDP support."""
logger = get_logger()
state_dict = _read_dit_checkpoint(cfg.dit_ckpt, cfg.dit_ckpt_type) if cfg.dit_ckpt is not None else None
adapter_checkpoint = None
if state_dict is not None and "generator_lora" in state_dict:
adapter_checkpoint = state_dict
base_path = getattr(cfg, "dmd_base_ckpt", None) or getattr(cfg, "dmd_stage1_ckpt", None)
if not base_path or os.path.abspath(base_path) == os.path.abspath(cfg.dit_ckpt):
raise ValueError("Set JOYAI_STAGE1_CKPT or dmd_base_ckpt to the SFT base for generator LoRA inference.")
state_dict = _read_dit_checkpoint(base_path, getattr(cfg, "dmd_base_ckpt_type", "pt"))
dtype = PRECISION_TO_TYPE[cfg.dit_precision]
use_pretrained_loading = _uses_pretrained_loading(cfg.dit_arch_config)
if use_pretrained_loading:
model_kwargs = {'torch_dtype': dtype, 'device': device}
else:
model_kwargs = {'dtype': dtype, 'device': device, 'args': cfg}
model = build_from_config(cfg.dit_arch_config, **model_kwargs)
if not use_fsdp or not dist.is_initialized() or dist.get_world_size() == 1:
model.to(device=device)
if state_dict is not None:
state_dict = _strip_common_state_prefixes(state_dict)
load_state_dict = {}
for k, v in state_dict.items():
if (
k == "img_in.weight"
and hasattr(model, "img_in")
and model.img_in.weight.shape != v.shape
):
v = v.reshape(model.img_in.weight.shape)
load_state_dict[k] = v
strict = not use_pretrained_loading
missing_keys, unexpected_keys = model.load_state_dict(load_state_dict, strict=strict)
if missing_keys:
logger.warning(f"Missing keys when loading DiT: {missing_keys[:20]}")
if unexpected_keys:
logger.warning(f"Unexpected keys when loading DiT: {unexpected_keys[:20]}")
if adapter_checkpoint is not None:
_merge_generator_lora(model, adapter_checkpoint, cfg)
_configure_dit_trainable_modules(model, cfg)
if use_fsdp:
model = maybe_load_fsdp_model(
model=model,
hsdp_shard_dim=cfg.hsdp_shard_dim,
reshard_after_forward=cfg.reshard_after_forward,
param_dtype=dtype,
reduce_dtype=torch.float32,
output_dtype=None,
cpu_offload=cfg.cpu_offload,
pin_cpu_memory=cfg.pin_cpu_memory,
)
total_params = sum(p.numel() for p in model.parameters())
logger.info(f"Instantiate model with {total_params / 1e9:.2f}B parameters")
param_dtypes = {param.dtype for param in model.parameters()}
if len(param_dtypes) > 1:
logger.warning(
f"Model has mixed dtypes: {param_dtypes}. Converting to {dtype}")
model = model.to(dtype)
return model.eval()
__all__ = [
"load_dit",
"load_pipeline",
]
|