-
Notifications
You must be signed in to change notification settings - Fork 833
/
Copy pathrun_finetune_with_custom_optim.sh
330 lines (316 loc) · 10 KB
/
run_finetune_with_custom_optim.sh
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
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
#!/bin/bash
# Please run this script under ${project_id} in project directory of
# https://github.com/shizhediao/llm-ft
# COMMIT: d5fecf30ba8011067b10cf51fede53a5ab6574e4
# Parses arguments
model_name_or_path=gpt2
dataset_path=data/alpaca/train_conversation
# Other optional arguments that can improve memory saving
gradient_checkpointing=True
use_flash_attention=0
gradient_accumulation_steps=1
batch_size=1
block_size=256
per_device_train_batch_size=1
conversation_template=llama2
optim=dummy
# Select an optimizer from the following options:
# - 'adamw_torch'
# - 'adafactor'
# - 'sgd'
# - 'lion_8bit'
# - 'lion_32bit'
# - 'rmsprop'
# Additional optimizers are shown below
learning_rate=1e-5
lr_schedule=cosine
beta1=0.9
beta2=0.999
beta3=0.99
weight_decay=0
momentum=0
num_epoch=0.01
use_deepspeed=1
seed=42
# Safety related arguments
trust_remote_code=0
# Enable model parallelism for multiple gpus, modify this if you prefer
# customized deepspeed zero-redundancy optimization settings
num_gpu=$(python -c "import torch; print(torch.cuda.device_count())")
ds_config_file=configs/ds_config_zero0_no_offload.json
if [[ ${num_gpu} -ge 2 ]]; then
ds_config_file=configs/ds_config_zero2_no_offload.json
fi
while [[ $# -ge 1 ]]; do
key="$1"
case ${key} in
-m|--model_name_or_path)
model_name_or_path="$2"
shift
;;
-d|--dataset_path)
dataset_path="$2"
shift
;;
-o|--output_model_path)
output_dir="$2"
shift
;;
--lisa_activated_layers)
lisa_activated_layers="$2"
shift
;;
--lisa_interval_steps)
lisa_interval_steps="$2"
shift
;;
--gradient_checkpointing)
gradient_checkpointing="$2"
shift
;;
--deepspeed)
ds_config_file="$2"
shift
;;
--use_flash_attention)
use_flash_attention="$2"
shift
;;
--gradient_accumulation_steps)
gradient_accumulation_steps="$2"
shift
;;
--block_size)
block_size="$2"
shift
;;
--conversation_template)
conversation_template="$2"
shift
;;
--per_device_train_batch_size|--batch_size)
per_device_train_batch_size="$2"
batch_size="$2"
shift
;;
--trust_remote_code)
trust_remote_code="$2"
shift
;;
--run_name)
run_name="$2"
shift
;;
--optim)
optim="$2"
shift
;;
--lr)
learning_rate="$2"
shift
;;
--beta1)
beta1="$2"
shift
;;
--beta2)
beta2="$2"
shift
;;
--beta3)
beta3="$2"
shift
;;
--weight_decay)
weight_decay="$2"
shift
;;
--momentum)
momentum="$2"
shift
;;
-n|--num_epoch)
num_epoch="$2"
shift
;;
--lr_schedule)
lr_schedule="$2"
shift
;;
--use_deepspeed)
use_deepspeed="$2"
shift
;;
--seed)
seed="$2"
shift
;;
*)
echo "error: unknown option \"${key}\"" 1>&2
exit 1
esac
shift
done
deepspeed_args="--master_port=1103 --hostfile configs/hostfile"
optim_suffix_args=""
if [ "${optim}" == "dummy" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_dummy_beta1 ${beta1}"
optim_suffix_args+=" --optim_dummy_beta2 ${beta2}"
elif [ "${optim}" == "adabelief" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_beta1 ${beta1}"
optim_suffix_args+=" --optim_beta2 ${beta2}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "adabound" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_beta1 ${beta1}"
optim_suffix_args+=" --optim_beta2 ${beta2}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "lars" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_momentum ${momentum}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "lamb" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_beta1 ${beta1}"
optim_suffix_args+=" --optim_beta2 ${beta2}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "adamax" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_beta1 ${beta1}"
optim_suffix_args+=" --optim_beta2 ${beta2}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "nadam" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_beta1 ${beta1}"
optim_suffix_args+=" --optim_beta2 ${beta2}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "radam" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_beta1 ${beta1}"
optim_suffix_args+=" --optim_beta2 ${beta2}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "adamp" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_beta1 ${beta1}"
optim_suffix_args+=" --optim_beta2 ${beta2}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "sgdp" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_momentum ${momentum}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "yogi" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_beta1 ${beta1}"
optim_suffix_args+=" --optim_beta2 ${beta2}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "sophia" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_beta1 ${beta1}"
optim_suffix_args+=" --optim_beta2 ${beta2}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "adan" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_beta1 ${beta1}"
optim_suffix_args+=" --optim_beta2 ${beta2}"
optim_suffix_args+=" --optim_beta3 ${beta3}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "adam" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_beta1 ${beta1}"
optim_suffix_args+=" --optim_beta2 ${beta2}"
elif [ "${optim}" == "novograd" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_beta1 ${beta1}"
optim_suffix_args+=" --optim_beta2 ${beta2}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "adadelta" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
elif [ "${optim}" == "adagrad" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
elif [ "${optim}" == "adamw_schedule_free" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_beta1 ${beta1}"
optim_suffix_args+=" --optim_beta2 ${beta2}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
elif [ "${optim}" == "sgd_schedule_free" ]; then
optim_suffix_args="--use_customized_optim 1"
optim_suffix_args+=" --customized_optim ${optim}"
optim_suffix_args+=" --optim_momentum ${momentum}"
optim_suffix_args+=" --optim_weight_decay ${weight_decay}"
else
optim_suffix_args="--optim ${optim}"
optim_suffix_args+=" --adam_beta1 ${beta1}"
optim_suffix_args+=" --adam_beta2 ${beta2}"
fi
# Finetune
exp_id=alpaca_${optim}_lr-${learning_rate}_beta1-${beta1}_beta2-${beta2}_lr-sched-${lr_schedule}_model-$(basename ${model_name_or_path})_batch-size-${batch_size}x${gradient_accumulation_steps}_seed-${seed}
echo "$(date): ${exp_id}..."
tmp_dir=tmp
mkdir -p ${tmp_dir}
prefix=${exp_id}
if [ -f ${tmp_dir}/${prefix}.mark ]; then
exit 0
fi
trap "rm -f ${tmp_dir}/${prefix}.mark" SIGINT SIGTERM SIGKILL
touch ${tmp_dir}/${prefix}.mark
project_dir=$(cd "$(dirname $0)"/..; pwd)
log_dir=${project_dir}/log/${exp_id}
output_dir=output_models/${exp_id}
mkdir -p ${output_dir} ${log_dir}
exe="deepspeed ${deepspeed_args}"
if [[ ${use_deepspeed} -eq 0 ]]; then
exe=python
fi
${exe} examples/finetune.py \
--model_name_or_path ${model_name_or_path} \
--trust_remote_code ${trust_remote_code} \
--dataset_path ${dataset_path} \
--output_dir ${output_dir} --overwrite_output_dir \
--conversation_template ${conversation_template} \
--num_train_epochs ${num_epoch} \
--learning_rate ${learning_rate} \
--lr_scheduler_type ${lr_schedule} \
--disable_group_texts 1 \
--block_size ${block_size} \
--per_device_train_batch_size ${per_device_train_batch_size} \
--bf16 \
--deepspeed configs/ds_config_zero2_no_offload.json \
--torch_dtype bfloat16 \
--run_name ${exp_id} \
--validation_split_percentage 0 \
--logging_steps 1 \
--do_train \
--ddp_timeout 72000 \
--save_steps 5000 \
--dataloader_num_workers 1 \
--gradient_checkpointing ${gradient_checkpointing} \
--use_flash_attention ${use_flash_attention} \
--gradient_accumulation_steps ${gradient_accumulation_steps} \
--seed ${seed} \
${optim_suffix_args} \
| tee ${log_dir}/train.log \
2> ${log_dir}/train.err
if [[ $? -ne 0 ]]; then
echo "$(date): failed"
rm -f ${tmp_dir}/${prefix}.mark
fi