-
Notifications
You must be signed in to change notification settings - Fork 0
/
ppo_train.py
211 lines (190 loc) · 8.03 KB
/
ppo_train.py
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
import gc
import os
import fire
import torch
from torch.utils.data import DataLoader
from tqdm import tqdm
from src.dataset import JsonDataset
from src.entities import Timer
from src.evaluator import SolverEvaluator
from src.models.llama import LoraLlamaVerifier, LoraLlama
from src.models.modeling_args import LoraLlamaArgs
from src.ppo.buffer import CriticRolloutBuffer, RolloutBuffer, ActorRolloutBuffer
from src.ppo.collector import CriticBufferCollector, ActorBufferCollector
from src.ppo.trainer import ParallelActorTrainerForCausalLM, ParallelCriticTrainerForCausalLM
from src.tokenizers import LlamaTokenizer
from src.utils import setup_model_parallel, set_barrier, json_dump
def run(
actor_ckpt_dir: str,
actor_config_file: str,
actor_save_dir: str,
critic_ckpt_dir: str,
critic_config_file: str,
critic_save_dir: str,
reward_model_ckpt_dir: str,
reward_model_config_file: str,
task: str,
train_file: str,
label_file: str,
log_dir: str,
lora_rank: int = 16,
max_batch_size: int = 4,
max_buffer_size: int = 96,
max_seq_len: int = 512,
epochs: int = 1,
inner_epochs: int = 2,
lr: float = 1e-5,
tokenizer_path: str = None,
):
assert actor_ckpt_dir != critic_save_dir
tokenizer_path = tokenizer_path if tokenizer_path else 'config/tokenizer.model'
dataset = JsonDataset(f=train_file)
dataloader = DataLoader(dataset, batch_size=max_buffer_size)
local_rank, world_size = setup_model_parallel()
tokenizer = LlamaTokenizer(tokenizer_path)
actor_args = LoraLlamaArgs(
max_seq_len=max_seq_len,
local_rank=local_rank,
world_size=world_size,
r=lora_rank
).from_json(actor_config_file)
critic_args = LoraLlamaArgs(
max_seq_len=max_seq_len,
local_rank=local_rank,
world_size=world_size,
r=lora_rank
).from_json(critic_config_file)
reward_model_args = LoraLlamaArgs(
max_seq_len=max_seq_len,
local_rank=local_rank,
world_size=world_size,
r=lora_rank
).from_json(reward_model_config_file)
for epoch in range(epochs):
actor = LoraLlama(actor_args)
actor.init_weights()
actor.load(actor_ckpt_dir if epoch == 0 else os.path.join(actor_save_dir, f"epoch-{epoch}"))
# Evaluation
actor_evaluator = SolverEvaluator(actor, tokenizer, max_buffer_size, max_seq_len)
eval_outputs = actor_evaluator.forward(task, JsonDataset(label_file))
print("Evaluate Accuracy: ", eval_outputs.acc, "Missing: ", eval_outputs.missing)
os.makedirs(log_dir, exist_ok=True)
json_dump(eval_outputs.datalist, os.path.join(
log_dir, f'results-epoch-{epoch}-{round(eval_outputs.acc, 4)}.json'
), indent=4)
actor_buffer_collector = ActorBufferCollector(actor, tokenizer, max_seq_len)
actor_rollout_buffer = ActorRolloutBuffer()
print('Actor buffer collecting ...')
timer = Timer(len(dataloader))
for data in tqdm(dataloader):
timer.step()
actor_rollout_buffer.extend(
actor_buffer_collector.forward(data['instruction'])
)
print(data['instruction'][-1])
print(tokenizer.decode(actor_rollout_buffer.actions[-1][actor_rollout_buffer.action_masks[-1]].tolist()))
actor.cpu()
del actor
del actor_buffer_collector
torch.cuda.empty_cache()
gc.collect()
set_barrier()
critic = LoraLlamaVerifier(critic_args)
critic.init_weights()
critic.load(critic_ckpt_dir if epoch == 0 else os.path.join(critic_save_dir, f"epoch-{epoch}"))
critic_buffer_collector = CriticBufferCollector(critic, tokenizer, max_seq_len)
critic_rollout_buffer = CriticRolloutBuffer()
print('Critic buffer collecting ...')
for data in actor_rollout_buffer.get(max_buffer_size):
critic_rollout_buffer.extend(
critic_buffer_collector.forward(
data.instructions, data.actions, data.action_masks
)
)
critic.cpu()
del critic
del critic_buffer_collector
torch.cuda.empty_cache()
gc.collect()
set_barrier()
reward_model = LoraLlamaVerifier(reward_model_args)
reward_model.init_weights()
reward_model.load(reward_model_ckpt_dir)
reward_buffer_collector = CriticBufferCollector(reward_model, tokenizer, max_seq_len)
reward_rollout_buffer = CriticRolloutBuffer()
print('Reward buffer collecting ...')
for data in actor_rollout_buffer.get(max_buffer_size):
reward_rollout_buffer.extend(
reward_buffer_collector.forward(
data.instructions, data.actions, data.action_masks
)
)
reward_model.cpu()
del reward_model
del reward_buffer_collector
torch.cuda.empty_cache()
gc.collect()
set_barrier()
rollout_buffer = RolloutBuffer(
obs=actor_rollout_buffer.obs,
actions=actor_rollout_buffer.actions,
rewards=reward_rollout_buffer.scores,
values=critic_rollout_buffer.scores,
action_logits=actor_rollout_buffer.action_logits,
action_masks=actor_rollout_buffer.action_masks
)
torch.save({
'obs': rollout_buffer.obs[: max_buffer_size],
'actions': rollout_buffer.actions[: max_buffer_size],
'values': rollout_buffer.values[: max_buffer_size],
'rewards': rollout_buffer.rewards[: max_buffer_size],
'action_masks': rollout_buffer.action_masks[: max_buffer_size],
'advantages': rollout_buffer.advantages[: max_buffer_size],
'returns': rollout_buffer.returns[: max_buffer_size]
}, f'buffer_{epoch}.bin')
actor = LoraLlama(actor_args)
actor.init_weights()
actor_optimizer = torch.optim.Adam(actor.parameters(), lr=0.1 * lr if epoch == 0 else lr)
actor_trainer = ParallelActorTrainerForCausalLM(actor, actor_optimizer)
actor_trainer.load_model(actor_ckpt_dir) if (
epoch == 0
) else actor_trainer.load(os.path.join(actor_save_dir, f"epoch-{epoch}"))
print('Actor training ...')
for inner_epoch in range(inner_epochs):
for data in rollout_buffer.get(max_batch_size):
outputs = actor_trainer.forward(data)
if actor_trainer.step % 100 == 0:
print(f'--------- STEP {actor_trainer.step} OF {len(rollout_buffer) // max_batch_size} ---------')
print('Loss: ', outputs.loss)
actor_trainer.save(os.path.join(actor_save_dir, f"epoch-{epoch + 1}"))
actor.cpu()
del actor
del actor_optimizer
del actor_trainer
torch.cuda.empty_cache()
gc.collect()
set_barrier()
critic = LoraLlamaVerifier(critic_args)
critic.init_weights()
critic_optimizer = torch.optim.Adam(critic.parameters(), lr=lr)
critic_trainer = ParallelCriticTrainerForCausalLM(critic, critic_optimizer)
critic_trainer.load_model(critic_ckpt_dir) if (
epoch == 0
) else critic_trainer.load(os.path.join(critic_save_dir, f"epoch-{epoch}"))
print('Critic training ...')
for inner_epoch in range(inner_epochs):
for data in rollout_buffer.get(max_batch_size):
outputs = critic_trainer.forward(data)
if critic_trainer.step % 100 == 0:
print(f'--------- STEP {critic_trainer.step} OF {len(rollout_buffer) // max_batch_size} ---------')
print('Loss: ', outputs.loss)
critic_trainer.save(os.path.join(critic_save_dir, f"epoch-{epoch + 1}"))
critic.cpu()
del critic
del critic_optimizer
del critic_trainer
torch.cuda.empty_cache()
gc.collect()
set_barrier()
if __name__ == '__main__':
fire.Fire(run)