File size: 5,187 Bytes
c9ac3c3 | 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 | # 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.
"""
Actor config
"""
import os
from dataclasses import dataclass, field
from typing import Any, Optional
@dataclass
class ModelConfig:
model_path: Optional[str] = None
tokenizer_path: Optional[str] = None
override_config: dict[str, Any] = field(default_factory=dict)
enable_gradient_checkpointing: bool = True
trust_remote_code: bool = True
freeze_vision_tower: bool = False
train_vision_merger: bool = False
"""When the vision tower is frozen, re-enable its merger/projector only."""
def post_init(self):
if self.tokenizer_path is None:
self.tokenizer_path = self.model_path
if self.model_path is not None and os.path.exists(self.model_path): # ray job uses absolute path
self.model_path = os.path.abspath(self.model_path)
if self.tokenizer_path is not None and os.path.exists(self.tokenizer_path):
self.tokenizer_path = os.path.abspath(self.tokenizer_path)
@dataclass
class OptimConfig:
lr: float = 1e-6
betas: tuple[float, float] = (0.9, 0.999)
weight_decay: float = 1e-2
strategy: str = "adamw"
lr_warmup_ratio: float = 0.0
lr_warmup_steps: Optional[int] = None
min_lr_ratio: Optional[float] = None
lr_scheduler_type: str = "constant"
# below are auto keys
training_steps: int = field(default=-1, init=False)
@dataclass
class FSDPConfig:
enable_full_shard: bool = True
enable_cpu_offload: bool = False
enable_rank0_init: bool = True
use_orig_params: bool = False
torch_dtype: Optional[str] = None
fsdp_size: int = -1
mp_param_dtype: str = "bf16"
mp_reduce_dtype: str = "fp32"
mp_buffer_dtype: str = "fp32"
@dataclass
class OffloadConfig:
offload_params: bool = False
offload_optimizer: bool = False
@dataclass
class ActorConfig:
strategy: str = "fsdp"
global_batch_size: int = 256
"""number of samples per minibatch for updating actor"""
micro_batch_size_per_device_for_update: int = 4
"""number of samples per forward pass for updating actor"""
micro_batch_size_per_device_for_experience: int = 16
"""number of samples per forward pass for computing log probs"""
max_grad_norm: float = 1.0
"""number to clip grad norm"""
clip_ratio_low: float = 0.2
"""clip ratio in PPO & DAPO"""
clip_ratio_high: float = 0.3
"""clip ratio in PPO & DAPO"""
clip_ratio_dual: float = 3.0
"""constant C in dual-clip PPO, clips when advantage < -C"""
loss_avg_mode: str = "token"
"""loss average mode: `token`, `seq`"""
loss_type: str = "default"
"""loss type: `default`, `gspo`, `cispo`"""
ppo_epochs: int = 1
"""number of ppo epochs for each rollout batch"""
padding_free: bool = True
"""use padding-free training"""
dynamic_batching: bool = True
"""enable dynamic batching"""
max_token_len_per_gpu: Optional[int] = None
"""max token length per GPU for dynamic batching. If None, use micro_batch_size * max_seq_len"""
ulysses_size: int = 1
"""ulysses sequence parallel size"""
use_torch_compile: bool = True
"""enable torch compile"""
model: ModelConfig = field(default_factory=ModelConfig)
optim: OptimConfig = field(default_factory=OptimConfig)
fsdp: FSDPConfig = field(default_factory=FSDPConfig)
offload: OffloadConfig = field(default_factory=OffloadConfig)
# below are auto keys
global_batch_size_per_device: int = field(default=-1, init=False)
disable_kl: bool = field(default=False, init=False)
use_kl_loss: bool = field(default=False, init=False)
kl_penalty: str = field(default="kl", init=False)
kl_coef: float = field(default=0.0, init=False)
selection_prune_ratio: float = field(default=0.0, init=False)
"""Auto-propagated from algorithm.selection_prune_ratio. The FSDP worker uses
it to scale global_batch_size by k=floor(n*(1-P)) instead of n, so the
per-rank mini-batch matches the post-selection batch size."""
@dataclass
class RefConfig:
strategy: str = "fsdp"
fsdp: FSDPConfig = field(default_factory=FSDPConfig)
offload: OffloadConfig = field(default_factory=OffloadConfig)
# below are auto keys
micro_batch_size_per_device_for_experience: int = field(default=-1, init=False)
padding_free: bool = field(default=False, init=False)
dynamic_batching: bool = field(default=False, init=False)
max_token_len_per_gpu: Optional[int] = field(default=None, init=False)
ulysses_size: int = field(default=1, init=False)
use_torch_compile: bool = field(default=True, init=False)
|