File size: 4,596 Bytes
7b58bc5 9149d87 44f990d 9149d87 7b58bc5 9149d87 7b58bc5 9149d87 7b58bc5 44f990d 7b58bc5 44f990d 7b58bc5 |
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 |
#!/bin/bash
#SBATCH --exclude=nid007052,nid005560
#SBATCH --nodes=64
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=40
#SBATCH --mem=256G
#SBATCH -p standard-g
#SBATCH -t 48:00:00
#SBATCH --gpus-per-node=mi250:8
#SBATCH --exclusive=user
#SBATCH --hint=nomultithread
#SBATCH --account=project_462000119
#SBATCH -o logs/%j.out
#SBATCH -e logs/%j.err
VARIANT=8b7178b
KILL_SWITCH_PATH=kill-switch-$VARIANT
CHECKPOINT_PATH=checkpoints_$VARIANT
TENSORBOARD_PATH=tensorboard_$VARIANT
### JEANZAY / NVIDIA ###
set -x -e
source $six_ALL_CCFRWORK/start-tr13f-6B3-ml-t0
GPUS_PER_NODE=8
NNODES=$SLURM_NNODES
TRAIN_DATA_PATH=train55boscar.txt
# "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_55B_text_document"
VALID_DATA_PATH=val.txt
# "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
### LUMI / AMD ###
# if run without sbatch, invoke here
: '
if [ -z $SLURM_JOB_ID ]; then
mkdir -p logs
sbatch "$0"
exit
fi
set -euo pipefail
# symlink logs/latest.out and logs/latest.err
ln -f -s $SLURM_JOB_ID.out logs/latest.out
ln -f -s $SLURM_JOB_ID.err logs/latest.err
DATA_PATH="/scratch/project_462000119/data/oscar_megatron/gpt2tok_oscar_text_document"
'
# Data
VOCAB_FILE="gpt2/vocab.json"
MERGE_FILE="gpt2/merges.txt"
PP_SIZE=2
TP_SIZE=1
MICRO_BATCH_SIZE=2
GRADIENT_ACCUMULATION_STEPS=1
WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
# Model parameters
source model_params.sh
MODEL_PARAM=("${PARAM_9293M[@]}")
NHIDDEN=${MODEL_PARAM[0]}
FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
KV_SIZE=${MODEL_PARAM[2]}
NHEADS=${MODEL_PARAM[3]}
NLAYERS=${MODEL_PARAM[4]}
SEQ_LEN=2048
echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
SAVE_INTERVAL=10000
# Tokens: 178000000000
# -> Samples: 86914062
TRAIN_SAMPLES=86_914_062
OPTIMIZER_ARGS=" \
--optimizer adam \
--adam-beta1 0.9 \
--adam-beta2 0.999 \
--adam-eps 1e-8 \
--lr 2e-4 \
--min-lr 2e-5 \
--lr-decay-style cosine \
--lr-decay-samples $TRAIN_SAMPLES \
--lr-warmup-samples 869_140 \
--clip-grad 1.0 \
--weight-decay 1e-1 \
"
GPT_ARGS=" \
--num-layers $NLAYERS \
--hidden-size $NHIDDEN \
--num-attention-heads $NHEADS \
--kv-channels $KV_SIZE \
--ffn-hidden-size $FFN_HIDDEN_SIZE \
--seq-length $SEQ_LEN \
--max-position-embeddings $SEQ_LEN \
--micro-batch-size $MICRO_BATCH_SIZE \
--global-batch-size $GLOBAL_BATCH_SIZE \
--train-samples $TRAIN_SAMPLES \
--vocab-file $VOCAB_FILE \
--merge-file $MERGE_FILE \
--clip-grad 1.0 \
--kill-switch-path $KILL_SWITCH_PATH \
--bf16 \
$OPTIMIZER_ARGS \
"
OUTPUT_ARGS=" \
--log-interval 10 \
--save-interval $SAVE_INTERVAL \
--eval-interval 1000 \
--eval-iters 1 \
--tensorboard-dir $TENSORBOARD_PATH \
--tensorboard-queue-size 5 \
--log-timers-to-tensorboard \
--log-batch-size-to-tensorboard \
--log-validation-ppl-to-tensorboard \
"
ZERO_STAGE=0
mkdir -p ds_configs
DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
cat <<EOF > $DS_CONFIG_PATH
{
"train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
"train_batch_size": $GLOBAL_BATCH_SIZE,
"gradient_clipping": 1.0,
"zero_optimization": {
"stage": $ZERO_STAGE
},
"bf16": {
"enabled": true
},
"steps_per_print": 2000,
"wall_clock_breakdown": false
}
EOF
DEEPSPEED_ARGS=" \
--deepspeed \
--deepspeed_config $DS_CONFIG_PATH \
--zero-stage $ZERO_STAGE \
"
CMD=" \
Megatron-DeepSpeed/pretrain_gpt.py \
--tensor-model-parallel-size $TP_SIZE \
--pipeline-model-parallel-size $PP_SIZE \
$GPT_ARGS \
$OUTPUT_ARGS \
--save $CHECKPOINT_PATH \
--load $CHECKPOINT_PATH \
--data-path $DATA_PATH \
--data-impl mmap \
--split 949,50,1 \
$DEEPSPEED_ARGS \
"
echo $CMD
echo "START $SLURM_JOBID: $(date)"
### JZ ###
export LAUNCHER="python -u -m torch.distributed.run \
--nproc_per_node $GPUS_PER_NODE \
--nnodes $NNODES \
--rdzv_endpoint $MASTER_ADDR:$MASTER_PORT \
--rdzv_backend c10d \
--max_restarts 0 \
--tee 3 \
"
SRUN_ARGS=" \
--wait=60 \
--kill-on-bad-exit=1 \
"
srun $SRUN_ARGS --jobid $SLURM_JOBID bash -c "$LAUNCHER --node_rank \$SLURM_PROCID $CMD" 2>&1 | tee -a $LOGS_PATH/main_log.txt
### LUMI ###
# bash launch_srun.sh $CMD
# srun --label launch.sh $CMD
echo "END $SLURM_JOBID: $(date)" |