scripts/yans/baseline_phi2/train.sh

set -eux
NOW=`date +%Y-%m-%d-%H:%M:%S`
LLM_RECIPES_DIR=/project
source $LLM_RECIPES_DIR/scripts/wmt2024/tokens.sh
rm -f /tmp/hffs-*

export WANDB_NOTES="Train sample"
wandb login
NUM_GPU_PER_NODE=$(nvidia-smi --query-gpu=name --format=csv,noheader | wc -l)
NUM_NODES=1
NUM_GPUS=$((${NUM_NODES} * ${NUM_GPU_PER_NODE}))

# training config
SEQ_LENGTH=4096
SLIDING_WINDOW_SIZE=131072
DATA_PARALLEL_SIZE=$NUM_GPUS

MICRO_BATCH_SIZE=1
GLOBAL_BATCH_SIZE=320
TRAIN_STEPS=20000
VALID_MICRO_BATCH_SIZE=1

# optimizer config
LR=2e-5
MIN_LR=1e-6
LR_WARMUP_STEPS=500
LR_DECAY_STEPS=$TRAIN_STEPS
WEIGHT_DECAY=0.1
GRAD_CLIP=1.0

# checkpoint & tokenizer
TOKENIZER_MODEL=/share/pretrained_lm/Phi/Phi-2
BASE_MODEL=$TOKENIZER_MODEL

LOAD_DIR=$BASE_MODEL
SAVE_DIR=/work/llm_recipes/models/yans-baseline-Phi-2
mkdir -p $(dirname $SAVE_DIR)
SAVE_BASE_NAME=$(basename $SAVE_DIR)
LOG_FILE_PATH=$SAVE_DIR/train_${NOW}.log

mkdir -p ${SAVE_DIR}

# data config
TRAIN_DATA_PATH="519177757 /work/llm_recipes/datasets/bin/baseline_phi2/llm_jp_corpus_v1_ja_wiki_train_0/data_text_document"
TRAIN_DATA_PATH="${TRAIN_DATA_PATH} 519177757 /work/llm_recipes/datasets/bin/baseline_phi2/llm_jp_corpus_v1_ja_wiki_train_0/data_text_document"

VALID_DATA_PATH="519177757 /work/llm_recipes/datasets/bin/baseline_phi2/llm_jp_corpus_v1_ja_wiki_train_0/data_text_document"
TEST_DATA_PATH=${VALID_DATA_PATH}


set +e
cd $LLM_RECIPES_DIR
# run

DISTRIBUTED_ARGS="--nproc_per_node $NUM_GPU_PER_NODE --nnodes 1 --node_rank 0 --master_addr localhost --master_port 8000"
torchrun $DISTRIBUTED_ARGS examples/finetuning.py \
  --seq-length ${SEQ_LENGTH} \
  --sliding-window-size ${SLIDING_WINDOW_SIZE} \
  --micro-batch-size ${MICRO_BATCH_SIZE} \
  --valid_micro_batch_size ${VALID_MICRO_BATCH_SIZE} \
  --global-batch-size ${GLOBAL_BATCH_SIZE} \
  --train-iters ${TRAIN_STEPS} \
  --tokenizer-type HFPreTrainedTokenizer \
  --tokenizer-model ${TOKENIZER_MODEL} \
  --train-data-path ${TRAIN_DATA_PATH} \
  --valid-data-path ${VALID_DATA_PATH} \
  --test-data-path ${TEST_DATA_PATH} \
  --lr ${LR} \
  --min-lr ${MIN_LR} \
  --lr-decay-style cosine \
  --lr-warmup-iters ${LR_WARMUP_STEPS} \
  --lr-decay-iters ${LR_DECAY_STEPS} \
  --weight-decay ${WEIGHT_DECAY} \
  --grad-clip-norm ${GRAD_CLIP} \
  --optimizer anyprecision \
  --adam-beta1 0.9 \
  --adam-beta2 0.95 \
  --adam-eps 1e-6 \
  --save-interval 500 \
  --eval-interval 500 \
  --eval-iters 10 \
  --bf16 \
  --mixed-precision \
  --base-model ${BASE_MODEL} \
  --save ${SAVE_DIR} \
  --load ${SAVE_DIR} \
  --fsdp-activation-checkpointing \
  --sharding-strategy FULL_SHARD \
  --checkpoint-type LOCAL_STATE_DICT \
  --save-n-checkpoints 10 \
  --upload-all-checkpoints-to-hf \
  --hf-upload-retry-limit 2 \
  --hf-repo-id shirayukikun/$SAVE_BASE_NAME \
  --wandb-entity "keitokudo" \
  --wandb-project "llm_tutorial" \
  --wandb-name ${SAVE_BASE_NAME}_train_${NOW} 2>&1 | tee $LOG_FILE_PATH

#   --attn-implementation eager \
#  --uploa-all-checkpoints-to-hf

rm -f /tmp/hffs-*