-
Notifications
You must be signed in to change notification settings - Fork 5
/
slurm_run_ssd_model_decode_fin.sh
204 lines (189 loc) · 7.95 KB
/
slurm_run_ssd_model_decode_fin.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
#!/usr/bin/bash
trap "kill 0" EXIT
script_role="host"
global_seed=$1 # inline param, 2021, 2022, etc
single_device_cuda="0" # inline param, "0", "1", etc
multi_device_cuda=$2 # inline param, "0,1,2,3", "0", etc
hf_cache="/private/home/xhan77/.cache/huggingface"
core_lm_name="roberta-large"
main_log_dir="/private/home/xhan77/ssd-lm/logging"
interpret_dataset_tokenized_path="${main_log_dir}/openwebtext_processed_pct100_blk200"
# # setup accelerate config
accelerate_config="${main_log_dir}/gpu.yaml"
# CUDA_VISIBLE_DEVICES=${multi_device_cuda} HF_HOME=${hf_cache} accelerate config --config_file ${accelerate_config}
# data hyperparameters
global_max_seq_len=200
####
# retrain
retrain_num_train_epochs=10000
retrain_per_device_train_batch_size=1
retrain_per_device_eval_batch_size=1
retrain_learning_rate=$3
retrain_weight_decay=0.01
retrain_gradient_accumulation_steps=1
retrain_num_warmup_steps=2000
retrain_max_train_steps=100000
sigma_num_steps=$4
loss_mode=$5
remove_noise_mode=$6
pa=$7
cs=1 # placeholder
dbs=$8
noise_manual_scale=$9
subdir=${10}
decode_context_size=${11}
decode_truncate_len=${12}
decode_depth=${13}
decode_ctr_lr=${14}
projection_top_p=${15}
out_fn=${16}
################ START ################
available_port=$(python -c 'import socket; s=socket.socket(); s.bind(("", 0)); print(s.getsockname()[1]); s.close()')
CUDA_VISIBLE_DEVICES="0" HF_HOME=${hf_cache} accelerate launch \
--multi_gpu --mixed_precision no \
--num_processes 1 --num_machines 1 --machine_rank 0 \
--main_process_port ${available_port} \
--num_cpu_threads_per_process 4 \
ssd_model_decode_fileio.py \
--max_seq_length ${global_max_seq_len} \
--model_name_or_path ${core_lm_name} \
--num_train_epochs ${retrain_num_train_epochs} \
--per_device_train_batch_size ${retrain_per_device_train_batch_size} \
--per_device_eval_batch_size ${retrain_per_device_eval_batch_size} \
--learning_rate ${retrain_learning_rate} \
--weight_decay ${retrain_weight_decay} \
--gradient_accumulation_steps ${retrain_gradient_accumulation_steps} \
--num_warmup_steps ${retrain_num_warmup_steps} \
--max_train_steps ${retrain_max_train_steps} \
--seed ${global_seed} \
--use_slow_tokenizer \
--output_dir ${main_log_dir}/${subdir} \
--loss_mode ${loss_mode} \
--remove_noise_mode ${remove_noise_mode} \
--hardcoded_pseudo_diralpha ${pa} \
--context_size ${cs} \
--decoding_block_size ${dbs} \
--sigma_num_steps ${sigma_num_steps} \
--tokenized_data_file_path ${interpret_dataset_tokenized_path} \
--if_create_tokenized_data_file "no" \
--decode_context_size ${decode_context_size} \
--decode_truncate_len ${decode_truncate_len} \
--decode_depth ${decode_depth} \
--train_mode decode \
--decode_ctr_lr ${decode_ctr_lr} \
--projection_top_p ${projection_top_p} \
--projection_alg "even" \
--ctr_opt_label_idx 2 \
--out_fn ${out_fn} &
available_port=$(python -c 'import socket; s=socket.socket(); s.bind(("", 0)); print(s.getsockname()[1]); s.close()')
CUDA_VISIBLE_DEVICES="1" HF_HOME=${hf_cache} accelerate launch \
--multi_gpu --mixed_precision no \
--num_processes 1 --num_machines 1 --machine_rank 0 \
--main_process_port ${available_port} \
--num_cpu_threads_per_process 4 \
ssd_model_decode_fileio.py \
--max_seq_length ${global_max_seq_len} \
--model_name_or_path ${core_lm_name} \
--num_train_epochs ${retrain_num_train_epochs} \
--per_device_train_batch_size ${retrain_per_device_train_batch_size} \
--per_device_eval_batch_size ${retrain_per_device_eval_batch_size} \
--learning_rate ${retrain_learning_rate} \
--weight_decay ${retrain_weight_decay} \
--gradient_accumulation_steps ${retrain_gradient_accumulation_steps} \
--num_warmup_steps ${retrain_num_warmup_steps} \
--max_train_steps ${retrain_max_train_steps} \
--seed ${global_seed} \
--use_slow_tokenizer \
--output_dir ${main_log_dir}/${subdir} \
--loss_mode ${loss_mode} \
--remove_noise_mode ${remove_noise_mode} \
--hardcoded_pseudo_diralpha ${pa} \
--context_size ${cs} \
--decoding_block_size ${dbs} \
--sigma_num_steps ${sigma_num_steps} \
--tokenized_data_file_path ${interpret_dataset_tokenized_path} \
--if_create_tokenized_data_file "no" \
--decode_context_size ${decode_context_size} \
--decode_truncate_len ${decode_truncate_len} \
--decode_depth ${decode_depth} \
--train_mode decode \
--decode_ctr_lr ${decode_ctr_lr} \
--projection_top_p ${projection_top_p} \
--projection_alg "even" \
--ctr_opt_label_idx 0 \
--out_fn ${out_fn} &
available_port=$(python -c 'import socket; s=socket.socket(); s.bind(("", 0)); print(s.getsockname()[1]); s.close()')
CUDA_VISIBLE_DEVICES="2" HF_HOME=${hf_cache} accelerate launch \
--multi_gpu --mixed_precision no \
--num_processes 1 --num_machines 1 --machine_rank 0 \
--main_process_port ${available_port} \
--num_cpu_threads_per_process 4 \
ssd_model_decode_fileio.py \
--max_seq_length ${global_max_seq_len} \
--model_name_or_path ${core_lm_name} \
--num_train_epochs ${retrain_num_train_epochs} \
--per_device_train_batch_size ${retrain_per_device_train_batch_size} \
--per_device_eval_batch_size ${retrain_per_device_eval_batch_size} \
--learning_rate ${retrain_learning_rate} \
--weight_decay ${retrain_weight_decay} \
--gradient_accumulation_steps ${retrain_gradient_accumulation_steps} \
--num_warmup_steps ${retrain_num_warmup_steps} \
--max_train_steps ${retrain_max_train_steps} \
--seed ${global_seed} \
--use_slow_tokenizer \
--output_dir ${main_log_dir}/${subdir} \
--loss_mode ${loss_mode} \
--remove_noise_mode ${remove_noise_mode} \
--hardcoded_pseudo_diralpha ${pa} \
--context_size ${cs} \
--decoding_block_size ${dbs} \
--sigma_num_steps ${sigma_num_steps} \
--tokenized_data_file_path ${interpret_dataset_tokenized_path} \
--if_create_tokenized_data_file "no" \
--decode_context_size ${decode_context_size} \
--decode_truncate_len ${decode_truncate_len} \
--decode_depth ${decode_depth} \
--train_mode decode \
--decode_ctr_lr ${decode_ctr_lr} \
--projection_top_p ${projection_top_p} \
--projection_alg "sampling" \
--ctr_opt_label_idx 2 \
--out_fn ${out_fn} &
available_port=$(python -c 'import socket; s=socket.socket(); s.bind(("", 0)); print(s.getsockname()[1]); s.close()')
CUDA_VISIBLE_DEVICES="3" HF_HOME=${hf_cache} accelerate launch \
--multi_gpu --mixed_precision no \
--num_processes 1 --num_machines 1 --machine_rank 0 \
--main_process_port ${available_port} \
--num_cpu_threads_per_process 4 \
ssd_model_decode_fileio.py \
--max_seq_length ${global_max_seq_len} \
--model_name_or_path ${core_lm_name} \
--num_train_epochs ${retrain_num_train_epochs} \
--per_device_train_batch_size ${retrain_per_device_train_batch_size} \
--per_device_eval_batch_size ${retrain_per_device_eval_batch_size} \
--learning_rate ${retrain_learning_rate} \
--weight_decay ${retrain_weight_decay} \
--gradient_accumulation_steps ${retrain_gradient_accumulation_steps} \
--num_warmup_steps ${retrain_num_warmup_steps} \
--max_train_steps ${retrain_max_train_steps} \
--seed ${global_seed} \
--use_slow_tokenizer \
--output_dir ${main_log_dir}/${subdir} \
--loss_mode ${loss_mode} \
--remove_noise_mode ${remove_noise_mode} \
--hardcoded_pseudo_diralpha ${pa} \
--context_size ${cs} \
--decoding_block_size ${dbs} \
--sigma_num_steps ${sigma_num_steps} \
--tokenized_data_file_path ${interpret_dataset_tokenized_path} \
--if_create_tokenized_data_file "no" \
--decode_context_size ${decode_context_size} \
--decode_truncate_len ${decode_truncate_len} \
--decode_depth ${decode_depth} \
--train_mode decode \
--decode_ctr_lr ${decode_ctr_lr} \
--projection_top_p ${projection_top_p} \
--projection_alg "sampling" \
--ctr_opt_label_idx 0 \
--out_fn ${out_fn} &
wait