config.py 3.99 KB
Newer Older
chenych's avatar
chenych committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
# 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.
"""
PPO config
"""

import os
from dataclasses import asdict, dataclass, field, fields, is_dataclass
from typing import Optional, Tuple

chenych's avatar
chenych committed
22
from ..workers.config import WorkerConfig
chenych's avatar
chenych committed
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38


def recursive_post_init(dataclass_obj):
    if hasattr(dataclass_obj, "post_init"):
        dataclass_obj.post_init()

    for attr in fields(dataclass_obj):
        if is_dataclass(getattr(dataclass_obj, attr.name)):
            recursive_post_init(getattr(dataclass_obj, attr.name))


@dataclass
class DataConfig:
    train_files: str = ""
    val_files: str = ""
    prompt_key: str = "prompt"
chenych's avatar
chenych committed
39
40
    answer_key: str = "answer"
    image_key: str = "images"
chenych's avatar
chenych committed
41
42
43
    max_prompt_length: int = 512
    max_response_length: int = 512
    rollout_batch_size: int = 512
chenych's avatar
chenych committed
44
    val_batch_size: int = -1
chenych's avatar
Update  
chenych committed
45
    format_prompt: Optional[str] = None
chenych's avatar
chenych committed
46
    override_chat_template: Optional[str] = None
chenych's avatar
chenych committed
47
48
49
50
    shuffle: bool = True
    seed: int = 1
    max_pixels: int = 4194304
    min_pixels: int = 262144
chenych's avatar
update  
chenych committed
51
52
53
54
    filter_overlong_prompts: bool = True

    def post_init(self):
        if self.format_prompt is not None:
chenych's avatar
chenych committed
55
            if os.path.exists(self.format_prompt):  # ray job uses absolute path
chenych's avatar
update  
chenych committed
56
57
58
                self.format_prompt = os.path.abspath(self.format_prompt)
            else:
                self.format_prompt = None
chenych's avatar
chenych committed
59
60
61
62
63
64


@dataclass
class AlgorithmConfig:
    gamma: float = 1.0
    lam: float = 1.0
chenych's avatar
chenych committed
65
66
67
    adv_estimator: str = "grpo"
    disable_kl: bool = False
    use_kl_loss: bool = False
chenych's avatar
chenych committed
68
69
    kl_penalty: str = "kl"
    kl_coef: float = 1e-3
chenych's avatar
chenych committed
70
    kl_type: str = "fixed"
chenych's avatar
chenych committed
71
72
73
74
75
76
    kl_horizon: float = 0.0
    kl_target: float = 0.0


@dataclass
class TrainerConfig:
chenych's avatar
chenych committed
77
    total_epochs: int = 10
chenych's avatar
chenych committed
78
79
80
81
82
83
    max_steps: Optional[int] = None
    project_name: str = "easy_r1"
    experiment_name: str = "demo"
    logger: Tuple[str] = ("console", "wandb")
    nnodes: int = 1
    n_gpus_per_node: int = 8
chenych's avatar
chenych committed
84
85
    critic_warmup: int = 0
    val_freq: int = -1
chenych's avatar
chenych committed
86
87
    val_before_train: bool = True
    val_only: bool = False
chenych's avatar
chenych committed
88
89
90
    val_generations_to_log: int = 0
    save_freq: int = -1
    save_limit: int = -1
chenych's avatar
chenych committed
91
    save_checkpoint_path: Optional[str] = None
chenych's avatar
chenych committed
92
    load_checkpoint_path: Optional[str] = None
chenych's avatar
chenych committed
93
94
95
96
97

    def post_init(self):
        if self.save_checkpoint_path is None:
            self.save_checkpoint_path = os.path.join("checkpoints", self.project_name, self.experiment_name)

chenych's avatar
chenych committed
98
        self.save_checkpoint_path = os.path.abspath(self.save_checkpoint_path)  # ray job uses absolute path
chenych's avatar
update  
chenych committed
99
100
101
        if self.load_checkpoint_path is not None:
            self.load_checkpoint_path = os.path.abspath(self.load_checkpoint_path)

chenych's avatar
chenych committed
102
103
104
105
106
107
108
109
110
111
112

@dataclass
class PPOConfig:
    data: DataConfig = field(default_factory=DataConfig)
    worker: WorkerConfig = field(default_factory=WorkerConfig)
    algorithm: AlgorithmConfig = field(default_factory=AlgorithmConfig)
    trainer: TrainerConfig = field(default_factory=TrainerConfig)

    def post_init(self):
        self.worker.rollout.prompt_length = self.data.max_prompt_length
        self.worker.rollout.response_length = self.data.max_response_length
chenych's avatar
update  
chenych committed
113
        self.worker.rollout.trust_remote_code = self.worker.actor.model.trust_remote_code
chenych's avatar
chenych committed
114
115
116
117
        self.worker.actor.disable_kl = self.algorithm.disable_kl
        self.worker.actor.use_kl_loss = self.algorithm.use_kl_loss
        self.worker.actor.kl_penalty = self.algorithm.kl_penalty
        self.worker.actor.kl_coef = self.algorithm.kl_coef
chenych's avatar
chenych committed
118
119
120
121
122
123

    def deep_post_init(self):
        recursive_post_init(self)

    def to_dict(self):
        return asdict(self)