File size: 13,206 Bytes
b171568 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 |
#This code file is from [https://github.com/hao-ai-lab/FastVideo], which is licensed under Apache License 2.0.
import json
import os
import torch
import torch.distributed.checkpoint as dist_cp
from peft import get_peft_model_state_dict
from safetensors.torch import load_file, save_file
from torch.distributed.checkpoint.default_planner import (DefaultLoadPlanner,
DefaultSavePlanner)
from torch.distributed.checkpoint.optimizer import \
load_sharded_optimizer_state_dict
from torch.distributed.fsdp import (FullOptimStateDictConfig,
FullStateDictConfig)
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from torch.distributed.fsdp import StateDictType
from fastvideo.utils.logging_ import main_print
def save_checkpoint_optimizer(model,
optimizer,
rank,
output_dir,
step,
discriminator=False):
with FSDP.state_dict_type(
model,
StateDictType.FULL_STATE_DICT,
FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True),
):
cpu_state = model.state_dict()
optim_state = FSDP.optim_state_dict(
model,
optimizer,
)
# todo move to get_state_dict
save_dir = os.path.join(output_dir, f"checkpoint-{step}")
os.makedirs(save_dir, exist_ok=True)
# save using safetensors
if rank <= 0 and not discriminator:
weight_path = os.path.join(save_dir,
"diffusion_pytorch_model.safetensors")
save_file(cpu_state, weight_path)
config_dict = dict(model.config)
config_dict.pop('dtype')
config_path = os.path.join(save_dir, "config.json")
# save dict as json
with open(config_path, "w") as f:
json.dump(config_dict, f, indent=4)
optimizer_path = os.path.join(save_dir, "optimizer.pt")
torch.save(optim_state, optimizer_path)
else:
weight_path = os.path.join(save_dir,
"discriminator_pytorch_model.safetensors")
save_file(cpu_state, weight_path)
optimizer_path = os.path.join(save_dir, "discriminator_optimizer.pt")
torch.save(optim_state, optimizer_path)
main_print(f"--> checkpoint saved at step {step}")
def save_checkpoint(transformer, rank, output_dir, step, epoch):
main_print(f"--> saving checkpoint at step {step}")
with FSDP.state_dict_type(
transformer,
StateDictType.FULL_STATE_DICT,
FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
):
cpu_state = transformer.state_dict()
# todo move to get_state_dict
if rank <= 0:
save_dir = os.path.join(output_dir, f"checkpoint-{step}-{epoch}")
os.makedirs(save_dir, exist_ok=True)
# save using safetensors
weight_path = os.path.join(save_dir,
"diffusion_pytorch_model.safetensors")
save_file(cpu_state, weight_path)
config_dict = dict(transformer.config)
if "dtype" in config_dict:
del config_dict["dtype"] # TODO
config_path = os.path.join(save_dir, "config.json")
# save dict as json
with open(config_path, "w") as f:
json.dump(config_dict, f, indent=4)
main_print(f"--> checkpoint saved at step {step}")
def save_checkpoint_generator_discriminator(
model,
optimizer,
discriminator,
discriminator_optimizer,
rank,
output_dir,
step,
):
with FSDP.state_dict_type(
model,
StateDictType.FULL_STATE_DICT,
FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
):
cpu_state = model.state_dict()
# todo move to get_state_dict
save_dir = os.path.join(output_dir, f"checkpoint-{step}")
os.makedirs(save_dir, exist_ok=True)
hf_weight_dir = os.path.join(save_dir, "hf_weights")
os.makedirs(hf_weight_dir, exist_ok=True)
# save using safetensors
if rank <= 0:
config_dict = dict(model.config)
config_path = os.path.join(hf_weight_dir, "config.json")
# save dict as json
with open(config_path, "w") as f:
json.dump(config_dict, f, indent=4)
weight_path = os.path.join(hf_weight_dir,
"diffusion_pytorch_model.safetensors")
save_file(cpu_state, weight_path)
main_print(f"--> saved HF weight checkpoint at path {hf_weight_dir}")
model_weight_dir = os.path.join(save_dir, "model_weights_state")
os.makedirs(model_weight_dir, exist_ok=True)
model_optimizer_dir = os.path.join(save_dir, "model_optimizer_state")
os.makedirs(model_optimizer_dir, exist_ok=True)
with FSDP.state_dict_type(model, StateDictType.SHARDED_STATE_DICT):
optim_state = FSDP.optim_state_dict(model, optimizer)
model_state = model.state_dict()
weight_state_dict = {"model": model_state}
dist_cp.save_state_dict(
state_dict=weight_state_dict,
storage_writer=dist_cp.FileSystemWriter(model_weight_dir),
planner=DefaultSavePlanner(),
)
optimizer_state_dict = {"optimizer": optim_state}
dist_cp.save_state_dict(
state_dict=optimizer_state_dict,
storage_writer=dist_cp.FileSystemWriter(model_optimizer_dir),
planner=DefaultSavePlanner(),
)
discriminator_fsdp_state_dir = os.path.join(save_dir,
"discriminator_fsdp_state")
os.makedirs(discriminator_fsdp_state_dir, exist_ok=True)
with FSDP.state_dict_type(
discriminator,
StateDictType.FULL_STATE_DICT,
FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True),
):
optim_state = FSDP.optim_state_dict(discriminator,
discriminator_optimizer)
model_state = discriminator.state_dict()
state_dict = {"optimizer": optim_state, "model": model_state}
if rank <= 0:
discriminator_fsdp_state_fil = os.path.join(
discriminator_fsdp_state_dir, "discriminator_state.pt")
torch.save(state_dict, discriminator_fsdp_state_fil)
main_print("--> saved FSDP state checkpoint")
def load_sharded_model(model, optimizer, model_dir, optimizer_dir):
with FSDP.state_dict_type(model, StateDictType.SHARDED_STATE_DICT):
weight_state_dict = {"model": model.state_dict()}
optim_state = load_sharded_optimizer_state_dict(
model_state_dict=weight_state_dict["model"],
optimizer_key="optimizer",
storage_reader=dist_cp.FileSystemReader(optimizer_dir),
)
optim_state = optim_state["optimizer"]
flattened_osd = FSDP.optim_state_dict_to_load(
model=model, optim=optimizer, optim_state_dict=optim_state)
optimizer.load_state_dict(flattened_osd)
dist_cp.load_state_dict(
state_dict=weight_state_dict,
storage_reader=dist_cp.FileSystemReader(model_dir),
planner=DefaultLoadPlanner(),
)
model_state = weight_state_dict["model"]
model.load_state_dict(model_state)
main_print(f"--> loaded model and optimizer from path {model_dir}")
return model, optimizer
def load_full_state_model(model, optimizer, checkpoint_file, rank):
with FSDP.state_dict_type(
model,
StateDictType.FULL_STATE_DICT,
FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True),
):
discriminator_state = torch.load(checkpoint_file)
model_state = discriminator_state["model"]
if rank <= 0:
optim_state = discriminator_state["optimizer"]
else:
optim_state = None
model.load_state_dict(model_state)
discriminator_optim_state = FSDP.optim_state_dict_to_load(
model=model, optim=optimizer, optim_state_dict=optim_state)
optimizer.load_state_dict(discriminator_optim_state)
main_print(
f"--> loaded discriminator and discriminator optimizer from path {checkpoint_file}"
)
return model, optimizer
def resume_training_generator_discriminator(model, optimizer, discriminator,
discriminator_optimizer,
checkpoint_dir, rank):
step = int(checkpoint_dir.split("-")[-1])
model_weight_dir = os.path.join(checkpoint_dir, "model_weights_state")
model_optimizer_dir = os.path.join(checkpoint_dir, "model_optimizer_state")
model, optimizer = load_sharded_model(model, optimizer, model_weight_dir,
model_optimizer_dir)
discriminator_ckpt_file = os.path.join(checkpoint_dir,
"discriminator_fsdp_state",
"discriminator_state.pt")
discriminator, discriminator_optimizer = load_full_state_model(
discriminator, discriminator_optimizer, discriminator_ckpt_file, rank)
return model, optimizer, discriminator, discriminator_optimizer, step
def resume_training(model, optimizer, checkpoint_dir, discriminator=False):
weight_path = os.path.join(checkpoint_dir,
"diffusion_pytorch_model.safetensors")
if discriminator:
weight_path = os.path.join(checkpoint_dir,
"discriminator_pytorch_model.safetensors")
model_weights = load_file(weight_path)
with FSDP.state_dict_type(
model,
StateDictType.FULL_STATE_DICT,
FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True),
):
current_state = model.state_dict()
current_state.update(model_weights)
model.load_state_dict(current_state, strict=False)
if discriminator:
optim_path = os.path.join(checkpoint_dir, "discriminator_optimizer.pt")
else:
optim_path = os.path.join(checkpoint_dir, "optimizer.pt")
optimizer_state_dict = torch.load(optim_path, weights_only=False)
optim_state = FSDP.optim_state_dict_to_load(
model=model, optim=optimizer, optim_state_dict=optimizer_state_dict)
optimizer.load_state_dict(optim_state)
step = int(checkpoint_dir.split("-")[-1])
return model, optimizer, step
def save_lora_checkpoint(transformer, optimizer, rank, output_dir, step,
pipeline, epoch):
with FSDP.state_dict_type(
transformer,
StateDictType.FULL_STATE_DICT,
FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
):
full_state_dict = transformer.state_dict()
lora_optim_state = FSDP.optim_state_dict(
transformer,
optimizer,
)
if rank <= 0:
save_dir = os.path.join(output_dir, f"lora-checkpoint-{step}-{epoch}")
os.makedirs(save_dir, exist_ok=True)
# save optimizer
optim_path = os.path.join(save_dir, "lora_optimizer.pt")
torch.save(lora_optim_state, optim_path)
# save lora weight
main_print(f"--> saving LoRA checkpoint at step {step}")
transformer_lora_layers = get_peft_model_state_dict(
model=transformer, state_dict=full_state_dict)
pipeline.save_lora_weights(
save_directory=save_dir,
transformer_lora_layers=transformer_lora_layers,
is_main_process=True,
)
# save config
lora_config = {
"step": step,
"lora_params": {
"lora_rank": transformer.config.lora_rank,
"lora_alpha": transformer.config.lora_alpha,
"target_modules": transformer.config.lora_target_modules,
},
}
config_path = os.path.join(save_dir, "lora_config.json")
with open(config_path, "w") as f:
json.dump(lora_config, f, indent=4)
main_print(f"--> LoRA checkpoint saved at step {step}")
def resume_lora_optimizer(transformer, checkpoint_dir, optimizer):
config_path = os.path.join(checkpoint_dir, "lora_config.json")
with open(config_path, "r") as f:
config_dict = json.load(f)
optim_path = os.path.join(checkpoint_dir, "lora_optimizer.pt")
optimizer_state_dict = torch.load(optim_path, weights_only=False)
optim_state = FSDP.optim_state_dict_to_load(
model=transformer,
optim=optimizer,
optim_state_dict=optimizer_state_dict)
optimizer.load_state_dict(optim_state)
step = config_dict["step"]
main_print(f"--> Successfully resuming LoRA optimizer from step {step}")
return transformer, optimizer, step
|