ds_gpt2_test.sh 2.7 KB
Newer Older
Elton Zheng's avatar
Elton Zheng committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
#! /bin/bash

helpFunction()
{
    echo ""
    echo "Usage: $0 -m model-parallelism -g gpu-per-node -n node# -b batch-size -s stpes -l layers -h hidden_size -q seq_length -e heads -c ckpt_num_layers [-d]"
    echo -e "\t-m model parallelism"
    echo -e "\t-g gpus per node"
    echo -e "\t-n node count"
    echo -e "\t-b batch size"
    echo -e "\t-s training steps"
    echo -e "\t-l layers"
    echo -e "\t-h hidden size"
    echo -e "\t-q sequence length"
    echo -e "\t-e attention heads"
    echo -e "\t-c checkpoint num_layers"
    echo -e "\t-o other args"
    echo -e "\t-d DeepSpeed config json file"
    echo -e "\t-z Enable Zero optimization"
    exit 1
}

23
24
layers=2
hidden_size=128
Elton Zheng's avatar
Elton Zheng committed
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
seq_length=1024
ckpt_num_layers=1
other_args=""
ds_opt=""
zero_opt=""

script_path=$(realpath $0)
script_dir=$(dirname $script_path)

while getopts "m:g:n:b:s:l:h:q:e:c:o:d:z" opt
do
    case "$opt" in
        m ) mp="$OPTARG" ;;
        g ) gpus="$OPTARG" ;;
        n ) nodes="$OPTARG" ;;
        b ) bs="$OPTARG" ;;
        s ) steps="$OPTARG" ;;
        l ) layers="$OPTARG" ;;
        h ) hidden_size="$OPTARG" ;;
        q ) seq_length="$OPTARG" ;;
        e ) heads="$OPTARG" ;;
        c ) ckpt_num_layers="$OPTARG" ;;
        o ) other_args="$OPTARG" ;;
        d ) ds_opt="--deepspeed --deepspeed_config $script_dir/$OPTARG" ;;
        z ) zero_opt="--zero_optimization" ;;
        ? ) helpFunction ;;
    esac
done

# Print helpFunction in case parameters are empty
if [ -z "$mp" ] || [ -z "$gpus" ] || [ -z "$nodes" ] || [ -z "$bs" ] || [ -z "$steps" ]
then
    echo "Some or all of the parameters are empty";
    helpFunction
fi

# Change for multinode config
MASTER_ADDR=localhost
MASTER_PORT=6000

gpt_options=" \
       --model-parallel-size ${mp} \
       --num-layers ${layers} \
       --hidden-size ${hidden_size} \
       --num-attention-heads ${heads} \
       --batch-size ${bs} \
       --seq-length ${seq_length} \
       --max-position-embeddings ${seq_length} \
       --train-iters ${steps} \
       --train-data webtext \
       --lazy-loader \
       --tokenizer-type GPT2BPETokenizer \
       --split 949,50,1 \
       --distributed-backend nccl \
       --lr 0.00015 \
       --no-load-optim \
       --lr-decay-style cosine \
       --weight-decay 1e-2 \
       --clip-grad 1.0 \
       --warmup .01 \
       --checkpoint-activations \
       --checkpoint-num-layers ${ckpt_num_layers} \
       --fp16 \
Jeff Rasley's avatar
Jeff Rasley committed
88
       --cache-dir /tmp/cache_dir \
Elton Zheng's avatar
Elton Zheng committed
89
90
91
92
93
94
       --log-interval 1 \
       ${other_args} \
       ${ds_opt} \
       ${zero_opt} \
"

Shaden Smith's avatar
Shaden Smith committed
95
work_dir="../../../DeepSpeedExamples/Megatron-LM/"
Jeff Rasley's avatar
Jeff Rasley committed
96
run_cmd="(cd ${work_dir} && deepspeed --num_nodes $nodes --num_gpus $gpus pretrain_gpt2.py ${gpt_options})"
Elton Zheng's avatar
Elton Zheng committed
97
98
99
100
echo ${run_cmd}
eval ${run_cmd}

set +x