File size: 11,178 Bytes
b757584 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 |
#!/bin/bash
#SBATCH --exclude=nid005159
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=4
#SBATCH --mem=256G
#SBATCH -p small-g
#SBATCH -t 2-0:00:00
#SBATCH --gpus-per-node=mi250:0
#SBATCH --exclusive=user
#SBATCH --hint=nomultithread
#SBATCH --account=project_462000119
#SBATCH -o logs/%j.out
#SBATCH -e logs/%j.err
# 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_eval.out and logs/latest_eval.err
ln -f -s $SLURM_JOB_ID.out logs/latest_eval.out
ln -f -s $SLURM_JOB_ID.err logs/latest_eval.err
source /pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/venv/bin/activate
echo "START TIME: $(date)"
# defining the right environment variables
export HF_DATASETS_OFFLINE=1
export HF_DATASETS_CACHE=/scratch/project_462000119/ds_cache
# Converted transformer checkpoint
# cd /pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/bigscience/lm-evaluation-harness
CKPTS=(
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4-repetitions/4b284b12bc4/global_step80108
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4-repetitions/4b284b17bc4/global_step80108
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4-repetitions/4b284b21bc4/global_step80108
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4-repetitions/4b284b28bc4/global_step80108
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4-repetitions/4b284b42bc4/global_step80108
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4-repetitions/4b284b84bc4/global_step80108
)
CKPTS=(
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-oscar-repetitions/4b284b12boscar/global_step80108
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-oscar-repetitions/4b284b17boscar/global_step80108
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-oscar-repetitions/4b284b21boscar/global_step80108
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-oscar-repetitions/4b284b28boscar/global_step80108
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-oscar-repetitions/4b284b42boscar/global_step80108
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-oscar-repetitions/4b284b84boscar/global_step80108
)
CKPTS=(
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-8b7-178b-c4-repetitions/8b7178b25b/global_step84877
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-8b7-178b-c4-repetitions/8b7178b35b/global_step84877
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-8b7-178b-c4-repetitions/8b7178b88b/global_step84877
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-8b7-178b-oscar-repetitions/8b7178b88b/global_step84877
)
CKPTS=(
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-8b7-178b-c4-repetitions/8b7178b44b/global_step84877
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-8b7-178b-c4-repetitions/8b7178b58b/global_step84877
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-8b7-178b-oscar-repetitions/8b7178b44b/global_step84877
)
CKPTSX=(
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4seed1/global_step80108
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4seed2/global_step80108
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4seed3/global_step80108
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-4b2-84b-c4seeds/4b284b84bc4seed4/global_step80108
)
CKPTS=(
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b1b25c4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b4bc4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b9bc4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b11bc4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b14bc4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b18bc4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b28bc4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b55bc4/global_step52452
)
CKPTSX=(
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-oscar-repetitions/2b855b1b25oscar/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-oscar-repetitions/2b855b4boscar/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-oscar-repetitions/2b855b9boscar/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-oscar-repetitions/2b855b11boscar/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-oscar-repetitions/2b855b14boscar/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-oscar-repetitions/2b855b18boscar/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-oscar-repetitions/2b855b28boscar/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-oscar-repetitions/2b855b55boscar/global_step52452
)
CKPTSX=(
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4seeds/2b855b55bc4seed1/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4seeds/2b855b55bc4seed2/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4seeds/2b855b55bc4seed3/global_step52452
)
CKPTS=(
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b14bc4/global_step52452
)
CKPTS=(
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b1b25c4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b4bc4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b9bc4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b11bc4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b14bc4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b18bc4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b28bc4/global_step52452
/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/lm1-2b8-55b-c4-repetitions/2b855b55bc4/global_step52452
)
CKPTS=(
/pfs/lustrep4/scratch/project_462000119/muennighoff/dec-2022-ul2/lm3-2b8-55b-c4/global_step52452
)
FEWSHOT_CONFIGS=(
0
1
2
3
4
5
)
TOKENIZER=/pfs/lustrep4/scratch/project_462000119/muennighoff/nov-2022-bettercom/gpt2
# Iterate through all possible combinations of data config, model ckpt & fewshot config and run the jobs
for ((i=0; i<${#CKPTS[@]}; i++)); do
for ((j=0; j<${#FEWSHOT_CONFIGS[@]}; j++)); do
#echo "sbatch --export=CKPT=${CKPTS[$i]},FEWSHOT_CONFIG=${FEWSHOT_CONFIGS[$j]},DATASET=${DATASETS[$k]} eval.sh"
MODEL_CKPT=${CKPTS[$i]}
MODEL_CKPT_NO_STEP=${MODEL_CKPT%/*}
MODEL_NAME=${MODEL_CKPT_NO_STEP##*/}
mkdir -p $MODEL_CKPT_NO_STEP/evaluation/rankeval_r_denoiser
#mv $MODEL_CKPT_NO_STEP/evaluation/$MODEL_NAME\_${FEWSHOT_CONFIGS[$j]}.* $MODEL_CKPT_NO_STEP/evaluation/rankeval/
OUTPUT_PATH=$MODEL_CKPT_NO_STEP/evaluation/rankeval_r_denoiser/$MODEL_NAME\_${FEWSHOT_CONFIGS[$j]}.json
eval_script="./eval_$i-$j.slurm"
cat <<EOT > $eval_script
#!/bin/bash
#SBATCH --exclude=nid005159
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=8
#SBATCH --mem=256G
#SBATCH -p small-g
#SBATCH -t 2-0:00:00
#SBATCH --gpus-per-node=mi250:1
#SBATCH --exclusive=user
#SBATCH --hint=nomultithread
#SBATCH --account=project_462000119
#SBATCH -o logs/%j.out
#SBATCH -e logs/%j.err
export HF_DATASETS_OFFLINE=1
export HF_DATASETS_CACHE=/scratch/project_462000119/ds_cache
VOCAB_FILE="gpt2/vocab.json"
MERGE_FILE="gpt2/merges.txt"
PP_SIZE=1
TP_SIZE=1
# different from the training MICRO_BATCH_SIZE - no optim memory, so can do bigger BS
# make as big as it can fit into gpu w/o OOM, but not too close to 100%
EVAL_MICRO_BATCH_SIZE=1
MICRO_BS_MULTIPLIER=1
# Model parameters
SEQ_LEN=2048
# Dummy arguments
MEGATRON_REQUIRED_ARGS=" \
--num-layers -1 \
--hidden-size -1 \
--num-attention-heads -1 \
--seq-length -1 \
--max-position-embeddings -1 \
"
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": 1,
"train_batch_size": 1,
"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/tasks/eval_harness/evaluate.py \
--load $MODEL_CKPT \
--results_path $OUTPUT_PATH \
--tensor-model-parallel-size \$TP_SIZE \
--pipeline-model-parallel-size \$PP_SIZE \
--vocab-file \$VOCAB_FILE \
--merge-file \$MERGE_FILE \
--micro-batch-size \$EVAL_MICRO_BATCH_SIZE \
--no-load-optim \
--no-load-rng \
--bf16 \
--inference \
--seq-length \$SEQ_LEN \
--task_list anli_r1,anli_r2,anli_r3,cb,copa,hellaswag,rte,winogrande,storycloze_2016,boolq,arc_easy,arc_challenge,sciq,piqa \
--intermed_results \
--adaptive_seq_len \
--micro_bs_multiplier \$MICRO_BS_MULTIPLIER \
--fewshots ${FEWSHOT_CONFIGS[$j]} \
\$MEGATRON_REQUIRED_ARGS \
\$DEEPSPEED_ARGS \
"
echo "\$CMD"
echo "START \$SLURM_JOBID: $(date)"
srun --label launch.sh \$CMD
echo "END \$SLURM_JOBID: $(date)"
EOT
sbatch $eval_script
# Sleep for a bit to avoid hitting the job submission limit
sleep 0.1
done
done
|