File size: 10,108 Bytes
3d1c0e1 |
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 |
# Copyright (c) 2025 FoundationVision
# SPDX-License-Identifier: MIT
import gc
import os
import os.path as osp
import subprocess
import time
import re
from typing import List, Optional, Tuple
import torch
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
import glob
import shutil
from infinity.utils import arg_util
import infinity.utils.dist as dist
def glob_with_epoch_iter(pattern, recursive=False):
def extract_ep_iter(filename):
match = re.search(r'ep(\d+)-iter(\d+)', filename)
if match:
ep = int(match.group(1))
iter_idx = int(match.group(2))
return ep, iter_idx
return 0, 0
return sorted(glob.glob(pattern, recursive=recursive), key=lambda x: extract_ep_iter(os.path.basename(x)), reverse=True)
def glob_with_global_step(pattern, recursive=False):
def extract_ep_iter(filename):
match = re.search(r'global_step_(\d+)', filename)
if match:
iter_idx = int(match.group(1))
return iter_idx
return 0
return sorted(glob.glob(pattern, recursive=recursive), key=lambda x: extract_ep_iter(os.path.basename(x)), reverse=True)
class CKPTSaver(object):
def __init__(self, is_master: bool, eval_milestone: List[Tuple[float, float]]):
self.is_master = is_master
self.time_stamp = torch.tensor([time.time() - 1e5, time.time()], device=dist.get_device())
self.sp_also: subprocess.Popen = None
self.sp_best: subprocess.Popen = None
self.sp_backup: subprocess.Popen = None
self.acc_str, self.eval_milestone = '[no acc str]', eval_milestone
def sav(
self, args: arg_util.Args, g_it: int, next_ep: int, next_it: int, trainer,
acc_str: Optional[str] = None, eval_milestone: Optional[List[Tuple[float, float]]] = None,
also_save_to: str = None, best_save_to: str = None,
):
fname = f'global_step_{g_it}.pth'
local_out_ckpt = os.path.join(args.bed, fname)
trainer_state = trainer.state_dict()
stt = time.time()
if self.is_master:
torch.save({
'args': args.state_dict(),
'arch': args.model,
'epoch': next_ep,
'iter': next_it,
'trainer': trainer_state,
'acc_str': self.acc_str,
'g_it': g_it,
}, local_out_ckpt)
cost = time.time() - stt
print(f'Checkpoint save cost: {cost:.2f}s', flush=True)
print(f'Checkpoint save to: {local_out_ckpt}', flush=True)
del trainer_state
gc.collect(),
torch.cuda.empty_cache()
dist.barrier()
def auto_resume(args: arg_util.Args, pattern='*.pth') -> Tuple[List[str], int, int, str, List[Tuple[float, float]], dict, dict]:
info = []
resume = ''
if args.auto_resume:
all_ckpt = glob_with_global_step(os.path.join(args.bed, pattern))
if len(all_ckpt) == 0:
info.append(f'[auto_resume] no ckpt found @ {pattern}')
info.append(f'[auto_resume quit]')
else:
resume = all_ckpt[0]
info.append(f'[auto_resume] auto load from @ {resume} ...')
else:
info.append(f'[auto_resume] disabled')
info.append(f'[auto_resume quit]')
if len(resume) == 0:
return info, 0, 0, '[no acc str]', [], {}, {}
print(f'auto resume from {resume}')
ckpt = torch.load(resume, map_location='cpu')
dist.barrier()
ep, it, g_it = ckpt['epoch'], ckpt['iter'], ckpt['g_it']
eval_milestone = ckpt.get('milestones', [])
info.append(f'[auto_resume success] resume from ep{ep}, it{it}, eval_milestone: {eval_milestone}')
return info, ep, g_it, ckpt.get('acc_str', '[no acc str]'), eval_milestone, ckpt['trainer'], ckpt['args']
def omnistore_auto_resume(args: arg_util.Args, pattern='ckpt*.pth'):
info = []
resume = ''
if args.auto_resume:
for dd in (args.local_out_path, args.bed):
all_ckpt = glob_with_global_step(os.path.join(dd, pattern))
if len(all_ckpt): break
if len(all_ckpt) == 0:
info.append(f'[auto_resume] no ckpt found @ {pattern}')
info.append(f'[auto_resume quit]')
else:
resume = all_ckpt[0]
info.append(f'[auto_resume] auto load from @ {resume} ...')
else:
info.append(f'[auto_resume] disabled')
info.append(f'[auto_resume quit]')
return resume, info
class omnistoreCheckpoint(object):
def __init__(self, eval_milestone: List[Tuple[float, float]]):
self.time_stamp = torch.tensor([time.time() - 1e5, time.time()], device=dist.get_device())
self.sp_also: subprocess.Popen = None
self.sp_best: subprocess.Popen = None
self.sp_backup: subprocess.Popen = None
self.acc_str, self.eval_milestone = '[no acc str]', eval_milestone
def sav(
self, args: arg_util.Args, global_it: int, next_ep: int, next_it: int, fsdp_object: FSDP, optimizer_object: torch.optim.Optimizer,
acc_str: Optional[str] = None, eval_milestone: Optional[List[Tuple[float, float]]] = None,
):
if acc_str is not None: self.acc_str = acc_str
if eval_milestone is not None: self.eval_milestone = eval_milestone
stt = time.time()
checkpoint_state = {
# 'model': {
# 'main_model': fsdp_object,
# 'ema_model': ema_fsdp_object,
# },
'model': fsdp_object,
# 'optimizer': optimizer_object,
'extra_state': {}
}
from omnistore import FSDPCheckpointer
print(f"{FSDPCheckpointer=}")
FSDPCheckpointer.save(
path=args.bed,
checkpoint_state=checkpoint_state,
global_steps=global_it,
async_fast_checkpoint=True,
save_flatten_model_optimizer=True,
)
if dist.is_master():
torch.save({
'args': args.state_dict(),
'next_ep': next_ep,
'next_it': next_it,
'global_it': global_it,
'acc_str': self.acc_str,
'milestones': self.eval_milestone,
}, os.path.join(args.bed, 'meta.pth'))
if self.sp_backup is not None:
self.sp_backup.wait(timeout=300); self.sp_backup.kill(); self.sp_backup.communicate()
self.time_stamp[0] = time.time()
def auto_sync(source_filename, target_filename):
cmd = f'cp -r {source_filename} {target_filename}'
self.sp_backup = subprocess.Popen(cmd, shell=True, bufsize=-1)
print(f'[Saver] auto_save cmd: {cmd}', flush=True)
local_files = glob.glob(f"{args.local_out_path}/*.txt")
for filename in local_files:
basename = os.path.basename(filename)
target_filename = f'{args.bed}/{basename}'
auto_sync(filename, target_filename)
cost = time.time() - stt
print(f'[CKPTSaver][rank00][omnistore: {FSDPCheckpointer is not None}] cost={time.time()-stt:.2f}s, ckpt saved to global_step_{global_it}', flush=True)
dist.barrier()
del checkpoint_state
def load(self, ckpt_path, fsdp_object, optimizer_object):
from omnistore import FSDPCheckpointer
checkpoint_state = {
'model': fsdp_object,
# 'optimizer': optimizer_object,
'extra_state': {}
}
FSDPCheckpointer.load(
ckpt_path,
checkpoint_state,
load_flatten_model_optimizer=True,
)
global_it = -1
meta_path = os.path.join(os.path.dirname(ckpt_path), 'meta.pth')
if os.path.exists(meta_path):
train_meta = torch.load(meta_path)
args_state, next_ep, next_it, acc_str, milestones = train_meta['args'], train_meta['next_ep'], train_meta['next_it'], train_meta['acc_str'], train_meta['milestones']
global_it = train_meta.get('global_it', -1)
else:
args_state, next_ep, next_it, acc_str, milestones = {}, 0, 0, '', []
return args_state, next_ep, next_it, global_it, acc_str, milestones
def merge_ckpt(omnistore_ckpt_path, output_path, fsdp_save_flatten_model, save=False):
print(f'merging omnistore ckpt into torch-format ckpt')
start = time.time()
from omnistore.utilities.ckpt_format_tool import omnistore_ckpt_to_pytorch_ckpt
state_dict = omnistore_ckpt_to_pytorch_ckpt(
save_path=omnistore_ckpt_path,
output_path=output_path,
framework="fsdp",
model_only=True,
return_dict=True,
fsdp_save_flatten_model=fsdp_save_flatten_model,
)
print(f"ckpt merged in {time.time() - start:.2f} seconds")
state_dict_model = state_dict["model"]
if '.cfg_uncond' in state_dict_model:
state_dict_model['cfg_uncond'] = state_dict_model['.cfg_uncond']
del state_dict_model['.cfg_uncond']
if '.pos_start' in state_dict_model:
state_dict_model['pos_start'] = state_dict_model['.pos_start']
del state_dict_model['.pos_start']
if '.sos_token' in state_dict_model:
state_dict_model['sos_token'] = state_dict_model['.sos_token']
del state_dict_model['.sos_token']
if 'semantic_head.weight' in state_dict_model:
print(f'[rush_resume] replace semantic_head with semantic_head2')
state_dict_model['semantic_head2.weight'] = state_dict_model['semantic_head.weight']
state_dict_model['semantic_head2.bias'] = state_dict_model['semantic_head.bias']
del state_dict_model['semantic_head.weight']
del state_dict_model['semantic_head.bias']
if save:
save_file = os.path.join(output_path, "slim-model.pt")
print(f'save to {save_file}')
torch.save(state_dict_model, save_file)
return state_dict_model
|