Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
run-train.sh 2.69 KB
Newer Older
hepj987's avatar
GPT2 base on megatron-deepspeed  
hepj987 committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17 
#!/bin/bash

export HSA_FORCE_FINE_GRAIN_PCIE=1
export MIOPEN_FIND_MODE=3

MODEL_NAME=gpt2-4tp
DATA_OUTPUT_PATH=./
LOGS_PATH=$DATA_OUTPUT_PATH/logs
CHECKPOINT_PATH=output-module/$MODEL_NAME
DATA_PATH="my-gpt2_text_document"

TENSORBOARD_PATH=output_dir/tensorboard/$MODEL_NAME
CODECARBON_PATH=output_dir/codecarbon/$MODEL_NAME

N_GPUS=8

TP_SIZE=4   # always fixed to the size of a single node
hepj987's avatar
调整格式  
hepj987 committed
18 
PP_SIZE=1 #128 #96    # NLAYERS must be a multiple of PP_SIZE here
hepj987's avatar
GPT2 base on megatron-deepspeed  
hepj987 committed
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 


MICRO_BATCH_SIZE=2
GLOBAL_BATCH_SIZE=32 #256 #1536
NLAYERS=24
NHIDDEN=1024 #12480
NHEADS=16
SEQ_LEN=1024
SAVE_INTERVAL=200

#rampup-batch-size 16 16 5859375
OPTIMIZER_ARGS=" \
    --optimizer adam \
    --adam-beta1 0.9 \
    --adam-beta2 0.95 \
    --adam-eps 1e-8 \
    --lr 6.0e-5 \
    --min-lr 6.0e-6 \
    --lr-decay-style cosine \
    --clip-grad 1.0 \
    --weight-decay 1e-1 \
    "


GPT_ARGS=" \
    --num-layers $NLAYERS \
    --hidden-size $NHIDDEN \
    --num-attention-heads $NHEADS \
    --seq-length $SEQ_LEN \
    --max-position-embeddings $SEQ_LEN \
    --micro-batch-size $MICRO_BATCH_SIZE \
    --global-batch-size $GLOBAL_BATCH_SIZE \
    --train-samples  3782590 \
    --loss-scale 12 \
    --vocab-file gpt2-vocab.json \
    --merge-file gpt2-merges.txt \
    --clip-grad 1.0 \
    --fp16 \
    --checkpoint-activations \
    --seed 42
    $OPTIMIZER_ARGS \
    "

OUTPUT_ARGS=" \
    --log-interval 1 \
    --save-interval $SAVE_INTERVAL \
    --eval-interval 1000 \
    --eval-iters 40 \
    --tensorboard-dir $TENSORBOARD_PATH \
    --tensorboard-queue-size 5 \
    --log-timers-to-tensorboard \
    --log-batch-size-to-tensorboard \
    --log-validation-ppl-to-tensorboard \
    "

DATA_ARGS=" \
    --save $CHECKPOINT_PATH \
    --load $CHECKPOINT_PATH \
    --data-path $DATA_PATH \
    "
ZERO_STAGE=1
config_json="./${MODEL_NAME}_ds_config.json"

cat <<EOT > $config_json
{
  "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
  "train_batch_size": $GLOBAL_BATCH_SIZE,
  "gradient_clipping": 1.0,
  "zero_optimization": {
    "stage": $ZERO_STAGE
  },
  "fp16": {
    "enabled": true,
    "loss_scale": 0,
    "loss_scale_window": 500,
    "hysteresis": 2,
    "min_loss_scale": 1,
    "initial_scale_power": 12
  },
  "steps_per_print": 2000,
  "wall_clock_breakdown": false
}
EOT

DEEPSPEED_ARGS=" \
    --deepspeed \
    --deepspeed_config ${config_json} \
    --zero-stage ${ZERO_STAGE} \
    --deepspeed-activation-checkpointing \
    "
export CMD="pretrain_gpt.py \
    --tensor-model-parallel-size $TP_SIZE \
    --pipeline-model-parallel-size $PP_SIZE \
    $GPT_ARGS \
    $DATA_ARGS \
    $OUTPUT_ARGS \
    --data-impl mmap \
    --split 949,50,1 \
    --distributed-backend nccl \
     $DEEPSPEED_ARGS \
    "

deepspeed --num_gpus $N_GPUS $CMD