File size: 5,248 Bytes
bcdf9fa |
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 |
# Copyright 2024 Bytedance Ltd. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import random
import shutil
import tempfile
from typing import Optional, Union
import numpy as np
import torch
import torch.distributed
from filelock import FileLock
from transformers import PreTrainedTokenizer, ProcessorMixin
class BaseCheckpointManager:
"""
A checkpoint manager that saves and loads
- model
- optimizer
- lr_scheduler
- extra_states
in a SPMD way.
We save
- sharded model states and optimizer states
- full lr_scheduler states
- huggingface tokenizer and config for ckpt merge
"""
def __init__(
self,
model,
optimizer: torch.optim.Optimizer,
lr_scheduler: torch.optim.lr_scheduler.LRScheduler = None,
processing_class: Union[PreTrainedTokenizer, ProcessorMixin] = None,
checkpoint_contents: Optional[list] = None,
):
if checkpoint_contents is None:
checkpoint_contents = ["model", "optimizer", "extra"]
self.previous_global_step = None
self.previous_saved_paths = []
self.model = model
self.optimizer = optimizer
self.lr_scheduler = lr_scheduler
self.processing_class = processing_class
self.checkpoint_contents = checkpoint_contents
self.rank = torch.distributed.get_rank()
self.world_size = torch.distributed.get_world_size()
def load_checkpoint(self, local_path: str, hdfs_path: str = None, del_local_after_load: bool = False):
raise NotImplementedError
def save_checkpoint(self, local_path: str, hdfs_path: str = None, global_step: int = 0, max_ckpt_to_keep: int = None):
raise NotImplementedError
@staticmethod
def checkpath(local_path: str, hdfs_path: str):
assert local_path is not None or hdfs_path is not None, "local_path and hdfs_path cannot be both None"
return local_path is not None, local_path if local_path is not None else hdfs_path
def remove_previous_save_local_path(self, path):
if isinstance(path, str):
path = [path]
for p in path:
abs_path = os.path.abspath(p)
print(f"Checkpoint manager remove previous save local path: {abs_path}")
if not os.path.exists(abs_path):
continue
shutil.rmtree(abs_path, ignore_errors=True)
@staticmethod
def local_mkdir(path):
if not os.path.isabs(path):
working_dir = os.getcwd()
path = os.path.join(working_dir, path)
# Using hash value of path as lock file name to avoid long file name
lock_filename = f"ckpt_{hash(path) & 0xFFFFFFFF:08x}.lock"
lock_path = os.path.join(tempfile.gettempdir(), lock_filename)
try:
with FileLock(lock_path, timeout=60): # Add timeout
# make a new dir
os.makedirs(path, exist_ok=True)
except Exception as e:
print(f"Warning: Failed to acquire lock for {path}: {e}")
# Even if the lock is not acquired, try to create the directory
os.makedirs(path, exist_ok=True)
return path
@staticmethod
def get_rng_state():
rng_state = {
"cpu": torch.get_rng_state(),
"cuda": torch.cuda.get_rng_state(),
"numpy": np.random.get_state(),
"random": random.getstate(),
}
return rng_state
@staticmethod
def load_rng_state(rng_state):
torch.set_rng_state(rng_state["cpu"])
torch.cuda.set_rng_state(rng_state["cuda"])
np.random.set_state(rng_state["numpy"])
random.setstate(rng_state["random"])
def find_latest_ckpt_path(path, directory_format="global_step_{}"):
if path is None:
return None
tracker_file = get_checkpoint_tracker_filename(path)
if not os.path.exists(tracker_file):
print("Checkpoint tracker file does not exist: %s", tracker_file)
return None
with open(tracker_file, "rb") as f:
iteration = int(f.read().decode())
ckpt_path = os.path.join(path, directory_format.format(iteration))
if not os.path.exists(ckpt_path):
print("Checkpoint does not exist: %s", ckpt_path)
return None
print("Found checkpoint: %s", ckpt_path)
return ckpt_path
def get_checkpoint_tracker_filename(root_path: str):
"""
Tracker file rescords the latest chckpoint during training to restart from.
"""
return os.path.join(root_path, "latest_checkpointed_iteration.txt")
|