File size: 6,863 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
# 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 shutil
import warnings
from typing import Optional, Union

import torch
import torch.distributed
from torch.distributed.fsdp import FullStateDictConfig, ShardedOptimStateDictConfig, ShardedStateDictConfig, StateDictType
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from transformers import GenerationConfig, PreTrainedTokenizer, ProcessorMixin

from verl.utils.fs import copy_to_local, is_non_local
from verl.utils.fsdp_utils import fsdp_version, get_fsdp_state_ctx

from .checkpoint_manager import BaseCheckpointManager


class FSDPCheckpointManager(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/processor and config for ckpt merge

    """

    def __init__(

        self,

        model: FSDP,

        optimizer: torch.optim.Optimizer,

        lr_scheduler: torch.optim.lr_scheduler.LRScheduler,

        processing_class: Union[PreTrainedTokenizer, ProcessorMixin] = None,

        checkpoint_contents: Optional[list] = None,

        **kwargs,

    ):
        if checkpoint_contents is None:
            checkpoint_contents = ["model", "optimizer", "extra"]
        if processing_class is None:
            assert "tokenizer" in kwargs, "tokenizer or processor must be provided"
            warnings.warn("`tokenizer` is deprecated. use `processing_class` instead.", DeprecationWarning, stacklevel=2)
            processing_class = kwargs.pop("tokenizer")
        assert "model" in checkpoint_contents and "optimizer" in checkpoint_contents and "extra" in checkpoint_contents, f"FSDPCheckpointManager must include ['model', 'optimizer', 'extra'], got {checkpoint_contents}"

        super().__init__(
            model,
            optimizer,
            lr_scheduler=lr_scheduler,
            processing_class=processing_class,
            checkpoint_contents=checkpoint_contents,
        )

    def load_checkpoint(self, local_path: str, hdfs_path: str = None, del_local_after_load=False):
        if local_path is None:
            return

        # every rank download its own checkpoint
        remote_model_path = os.path.join(local_path, f"model_world_size_{self.world_size}_rank_{self.rank}.pt")
        remote_optim_path = os.path.join(local_path, f"optim_world_size_{self.world_size}_rank_{self.rank}.pt")
        remote_extra_state_path = os.path.join(local_path, f"extra_state_world_size_{self.world_size}_rank_{self.rank}.pt")
        print(f"[rank-{self.rank}]: Loading from {remote_model_path} and {remote_optim_path} and {remote_extra_state_path}")
        local_model_path = copy_to_local(remote_model_path)
        local_optim_path = copy_to_local(remote_optim_path)
        local_extra_state_path = copy_to_local(remote_extra_state_path)

        model_state_dict = torch.load(local_model_path, weights_only=False)
        optimizer_state_dict = torch.load(local_optim_path, weights_only=False)
        extra_state_dict = torch.load(local_extra_state_path, weights_only=False)

        if del_local_after_load:
            try:
                os.remove(local_model_path) if is_non_local(local_model_path) else None
                os.remove(local_optim_path) if is_non_local(local_optim_path) else None
                os.remove(local_extra_state_path) if is_non_local(local_extra_state_path) else None
            except Exception as e:
                print(f"[rank-{self.rank}]: remove local resume ckpt file after loading failed, exception {e} will be ignored")

        lr_scheduler_state_dict = extra_state_dict["lr_scheduler"]

        state_dict_cfg = ShardedStateDictConfig(offload_to_cpu=True)
        optim_cfg = ShardedOptimStateDictConfig(offload_to_cpu=True)
        with get_fsdp_state_ctx(self.model, StateDictType.SHARDED_STATE_DICT, state_dict_cfg, optim_cfg):
            self.model.load_state_dict(model_state_dict)
            if self.optimizer is not None:
                self.optimizer.load_state_dict(optimizer_state_dict)
        # recover random state
        if "rng" in extra_state_dict:
            # 'rng' may not exist for backward compatibility
            self.load_rng_state(extra_state_dict["rng"])

        if self.lr_scheduler is not None:
            self.lr_scheduler.load_state_dict(lr_scheduler_state_dict)

    def save_checkpoint(self, local_path: str, hdfs_path: str = None, global_step: int = 0, max_ckpt_to_keep=None):
        # record the previous global step
        self.previous_global_step = global_step
        # only support save and load ckpt for actor
        from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
        experiment_dir = os.path.dirname(local_path)
        if self.rank == 0:
            if os.path.exists(experiment_dir):
                subdirs = [name for name in os.listdir(experiment_dir) if os.path.isdir(os.path.join(experiment_dir, name))]
                for name in subdirs:
                    full_path = os.path.join(experiment_dir, name)
                    shutil.rmtree(full_path)
                    
            os.makedirs(local_path, exist_ok=True)
        from torch.distributed.fsdp import FullStateDictConfig, StateDictType
        torch.distributed.barrier()
        with warnings.catch_warnings():
            warnings.simplefilter("ignore")
            
            with FSDP.state_dict_type(
                self.model, 
                StateDictType.FULL_STATE_DICT,
                FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
            ):
                state_dict = self.model.state_dict()
                model_path = os.path.join(local_path, f'model.pt')
                if self.rank == 0:
                    torch.save(state_dict, model_path)
                    
                    print("\n" + "="*60)
                    print(f"✅✅✅  SUCCESS: Model saved ✅✅✅")
                    print("="*60 + "\n")
        
                torch.distributed.barrier()