Spaces:
Runtime error
Runtime error
File size: 3,668 Bytes
50f0fbb |
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 |
#!/bin/bash
#SBATCH --job-name=randeng_t5_77M_summary
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=2
#SBATCH --gres=gpu:2 # number of gpus
#SBATCH --cpus-per-task=30
#SBATCH -o %x-%j.log
set -x -e
echo "START TIME: $(date)"
MODEL_NAME=randeng_t5_784M_summary
MICRO_BATCH_SIZE=8
ROOT_DIR=/cognitive_comp/dongxiaoqun/finetune/${MODEL_NAME}
if [ ! -d ${ROOT_DIR} ];then
mkdir ${ROOT_DIR}
echo ${ROOT_DIR} created!!!!!!!!!!!!!!
else
echo ${ROOT_DIR} exist!!!!!!!!!!!!!!!
fi
ZERO_STAGE=1
config_json="${ROOT_DIR}/ds_config.${MODEL_NAME}.json"
# Deepspeed figures out GAS dynamically from dynamic GBS via set_train_batch_size()
cat <<EOT > $config_json
{
"train_micro_batch_size_per_gpu": ${MICRO_BATCH_SIZE},
"steps_per_print": 100,
"gradient_clipping": 1.0,
"zero_optimization": {
"stage": $ZERO_STAGE,
"contiguous_gradients": false,
"overlap_comm": true,
"reduce_scatter": true,
"reduce_bucket_size": 50000000,
"allgather_bucket_size": 500000000
},
"optimizer": {
"type": "Adam",
"params": {
"lr": 1e-4,
"weight_decay": 1e-2
}
},
"scheduler": {
"params": {
"warmup_max_lr": 1e-04,
"warmup_min_lr": 1e-05,
"total_num_steps": 60000,
"warmup_num_steps" : 500
},
"type": "WarmupDecayLR"
},
"zero_allow_untested_optimizer": false,
"fp16": {
"enabled": true,
"loss_scale": 0,
"loss_scale_window": 1000,
"hysteresis": 2,
"min_loss_scale": 1
},
"activation_checkpointing": {
"partition_activations": false,
"contiguous_memory_optimization": false
},
"wall_clock_breakdown": false
}
EOT
export PL_DEEPSPEED_CONFIG_PATH=$config_json
export TORCH_EXTENSIONS_DIR=/cognitive_comp/dongxiaoqun/torch_extendsions
# export MASTER_PORT=$[RANDOM%10000+30000]
# export PL_FAULT_TOLERANT_TRAINING=1
TRAINER_ARGS="
--max_epochs 1 \
--gpus 1 \
--num_nodes 1 \
--strategy deepspeed_stage_${ZERO_STAGE} \
--default_root_dir $ROOT_DIR \
--dirpath $ROOT_DIR/ckpt \
--save_top_k 3 \
--monitor val_loss \
--mode min \
--save_last \
--every_n_train_steps 0 \
--val_check_interval 0.1 \
"
prompt="summary:"
DATA_ARGS="
--datasets_name lcsts \
--num_workers 30 \
--train_batchsize $MICRO_BATCH_SIZE \
--val_batchsize $MICRO_BATCH_SIZE \
--test_batchsize $MICRO_BATCH_SIZE \
--max_enc_length 128 \
--max_dec_length 64 \
--val_datasets_field val \
--prompt $prompt \
"
# --prompt $prompt \
MODEL_ARGS="
--pretrained_model_path /cognitive_comp/ganruyi/experiments/randeng_t5_large_v2/ckpt/hf_pretrained_epoch0_step732500 \
--output_save_path $ROOT_DIR/randeng_t5_784M_predict_lcsts.json \
"
SCRIPTS_PATH=/cognitive_comp/dongxiaoqun/debug/Fengshenbang-LM/fengshen/examples/summary/seq2seq_summary.py
SINGULARITY_PATH=/cognitive_comp/ganruyi/pytorch21_06_py3_docker_image_v2.sif
export CMD=" \
$SCRIPTS_PATH \
$TRAINER_ARGS \
$MODEL_ARGS \
$DATA_ARGS \
"
echo $CMD
source activate
conda activate torchnew
srun --nodes=1 --ntasks-per-node=1 --gres=gpu:1 --cpus-per-task=30 -o ${MODEL_NAME}-%J.log --jobid=229668 bash -c 'python3 $SCRIPT_PATH $CMD'
# source activate base
# python $CMD
# srun --jobid=229668 --nodes=1 --gres=gpu:1 --ntasks-per-node=1 --cpus-per-task=30 -e ${ROOT_DIR}/${MODEL_NAME}-%j.err -o ${ROOT_DIR}/${MODEL_NAME}-%j.log singularity exec --nv -B /cognitive_comp/:/cognitive_comp/ $SINGULARITY_PATH bash -c '/home/ganruyi/anaconda3/bin/python $CMD'
# srun python $CMD
# srun singularity exec --nv -B /cognitive_comp/:/cognitive_comp/ $SINGULARITY_PATH bash -c '/home/ganruyi/anaconda3/bin/python $CMD'
|