forked from bigscience-workshop/Megatron-DeepSpeed
-
Notifications
You must be signed in to change notification settings - Fork 0
/
run_fp16.sh
executable file
·163 lines (135 loc) · 3.22 KB
/
run_fp16.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
#!/bin/bash
DIR=`pwd`
DATETIME=`date +'date_%y-%m-%d_time_%H-%M-%S'`
#mkdir -p $DIR/logs
#mkdir -p /tmp/logs
#DATASET_1="<PATH TO THE FIRST DATASET>"
#DATASET_2="<PATH TO THE SECOND DATASET>"
#DATASET_3="<PATH TO THE THIRD DATASET>"
#DATASET="0.2 ${DATASET_1} 0.3 ${DATASET_2} 0.5 ${DATASET_3}"
BASE_DATA_PATH=/data/Megatron-LM/data
DATASET=${BASE_DATA_PATH}/indexed_datasets/megatron
VOCAB_PATH=${BASE_DATA_PATH}/gpt2-vocab.json
MERGE_PATH=${BASE_DATA_PATH}/gpt2-merges.txt
script_path=$(realpath $0)
script_dir=$(dirname $script_path)
#CONFIG_JSON="$script_dir/ds_config.json"
CONFIG_JSON="/tmp/ds_config.json"
USE_DEEPSPEED=1
ZERO_STAGE=0
# Debug
#TP=4
#PP=4
#LAYERS=8
#HIDDEN=512
#SEQ=1024
#GLOBAL_BATCH=128
#WORKER_STR="-i worker-0"
TP=1
PP=1
DP=2
WORLD_SIZE=$((TP*PP*DP))
HIDDEN=1024
LAYERS=24
SEQ=1024
GLOBAL_BATCH=1
WORKER_STR=""
MICRO_BATCH=1
LR=6.0e-4
MIN_LR=6.0e-5
DTYPE="fp16"
EXP_DIR=${HOME}/experiments/results/bf16
LOG_DIR="${EXP_DIR}/tensorboard/tp${TP}_pp${PP}_dp${DP}_hd${HIDDEN}_nl${LAYERS}_gbsz${GLOBAL_BATCH}_mbsz${MICRO_BATCH}_z${ZERO_STAGE}_LR_${LR}_${MIN_LR}_${DTYPE}_fix3"
mkdir -p $LOG_DIR
while [[ $# -gt 0 ]]
do
key="$1"
case $key in
--no-deepspeed)
USE_DEEPSPEED=0;
shift
;;
-z|--zero-stage)
ZERO_STAGE=$2;
shift
;;
*)
echo "Unknown argument(s)"
usage
exit 1
shift
;;
esac
done
options=" \
--tensor-model-parallel-size $TP \
--pipeline-model-parallel-size $PP \
--num-layers $LAYERS \
--hidden-size $HIDDEN \
--num-attention-heads 32 \
--seq-length $SEQ \
--loss-scale 12 \
--max-position-embeddings $SEQ \
--micro-batch-size $MICRO_BATCH \
--global-batch-size $GLOBAL_BATCH \
--train-iters 1000 \
--lr $LR \
--min-lr $MIN_LR \
--lr-decay-style cosine \
--log-interval 1 \
--eval-iters 40 \
--eval-interval 10 \
--data-path ${DATASET} \
--vocab-file ${VOCAB_PATH} \
--merge-file ${MERGE_PATH} \
--save-interval 10000 \
--split 98,2,0 \
--clip-grad 1.0 \
--weight-decay 0.1 \
--adam-beta1 0.9 \
--adam-beta2 0.95 \
--init-method-std 0.006 \
--${DTYPE} \
--checkpoint-activations \
--exit-interval 10000 \
--tensorboard-dir $LOG_DIR
"
if [[ ${USE_DEEPSPEED} -eq 1 ]]; then
echo "Using DeepSpeed"
options="${options} \
--deepspeed \
--deepspeed_config=${CONFIG_JSON} \
--zero-stage=${ZERO_STAGE} \
--deepspeed-activation-checkpointing \
"
fi
cat <<EOT > $CONFIG_JSON
{
"train_batch_size" : $GLOBAL_BATCH,
"train_micro_batch_size_per_gpu": $MICRO_BATCH,
"steps_per_print": 1,
"zero_optimization": {
"stage": $ZERO_STAGE
},
"bf16": {
"enabled": false
},
"fp16": {
"enabled": true,
"loss_scale": 0,
"loss_scale_window": 500,
"hysteresis": 2,
"min_loss_scale": 1,
"initial_scale_power": 8
},
"wall_clock_breakdown" : true
}
EOT
WORKER_STR="--num_nodes 1 --num_gpus $WORLD_SIZE"
#WORKER_STR="-i worker-0:0,1,2,3"
#run_cmd="deepspeed -i worker-0:0,1,2,3 ${DIR}/pretrain_gpt.py $@ ${options}"
#run_cmd="deepspeed -i worker-0 ${DIR}/pretrain_gpt.py $@ ${options}"
run_cmd="deepspeed --master_port 29600 $WORKER_STR ${DIR}/pretrain_gpt.py $@ ${options}"
echo ${run_cmd}
eval ${run_cmd}
set +x