ducdauge / sft-llm Goto Github PK

Hi,

I find the training speed is very slow when we use deepspeed stage 3. Can we use stage 1 or stage 2?

I attempted to replicate the results presented in Table 1 of the paper by executing the codes. However, I encountered discrepancies in achieving the same accuracy. Specifically, I executed MMLU and TyDiQA based on the Flanv2-trained LLaMA2-7b model. I made adjustments to the dataset name within the "finetune_peft_with_accelerate.sh" script and executed it on an A100-40GB GPU. However, the obtained F1 scores for TyDiQA and MMUL were 54.65 and 48.4, which deviates from the anticipated 56.2 and 50.7 as indicated in Table 1.

May I know what the issue is? Are there any configurations I missed during running the training or evaluation?

Below is the content of the "finetune_peft_with_accelerate.sh" script, along with the appended log files detailing the training and evaluation processes.

training_log.txt
mmlu_log.txt
tydiqa_log.txt

#!/bin/bash

MODEL_SIZES=( 7b 13b )
MODEL_SIZE_IDX=0
MODEL_SIZE=${MODEL_SIZES[MODEL_SIZE_IDX]}
PARAMS_PER_RANK_BY_MODEL_SIZE=( 2498560 3911680 )
PARAMS_PER_RANK=${PARAMS_PER_RANK_BY_MODEL_SIZE[MODEL_SIZE_IDX]}

RANKS=( 8 16 32 64 )
RANK_IDX=3
RANK=${RANKS[RANK_IDX]}
NUM_TUNABLE_WEIGHTS=$((RANK * PARAMS_PER_RANK))

WEIGHT_DECAYS=( 1 3 10 30 )
WEIGHT_DECAY_IDX=3
WEIGHT_DECAY=${WEIGHT_DECAYS[WEIGHT_DECAY_IDX]}

PREFIX_LENGTHS=( 10 20 40 80 )
PREFIX_LENGTH_IDX=3
PREFIX_LENGTH=${PREFIX_LENGTHS[PREFIX_LENGTH_IDX]}

LEARNING_RATE=1e-5

DATASET="flan_v2"
if [[ "$DATASET" == "tulu_v2" ]]; then
DATASET_ARG="--dataset_name allenai/tulu-v2-sft-mixture"
elif [[ "$DATASET" == "flan_v2" ]]; then
DATASET_ARG="--dataset_name ostapeno/tulu_v2_flan_v2_subset"
elif [[ "$DATASET" == "gpt4_alpaca" ]]; then
DATASET_ARG="--train_file data/processed/gpt4_alpaca/gpt4_alpaca_data.jsonl"
else
echo "Unsupported dataset "$DATASET"" >&2
exit 1
fi

PEFT_METHOD="sft"
SFT_METHOD="rigl"
if [[ "$PEFT_METHOD" == "sft" ]]; then
METHOD_NAME="sft_${SFT_METHOD}"
HP_STRING="r-${RANK}_lr-${LEARNING_RATE}_wd-${WEIGHT_DECAY}"
elif [[ "$PEFT_METHOD" == "lora" ]]; then
WEIGHT_DECAY=0
METHOD_NAME="lora"
HP_STRING="r-${RANK}_lr-${LEARNING_RATE}"
elif [[ "$PEFT_METHOD" == "ia3" ]]; then
WEIGHT_DECAY=0
METHOD_NAME="ia3"
HP_STRING="lr-${LEARNING_RATE}"
elif [[ "$PEFT_METHOD" == "pt" ]]; then
WEIGHT_DECAY=0
METHOD_NAME="pt"
HP_STRING="lr-${LEARNING_RATE}_plength-${PREFIX_LENGTH}"
else
echo "Unsupported PEFT method $PEFT_METHOD" >&2
exit 1
fi

QPEFT=false
if [[ "$QPEFT" == true ]]; then
QPEFT_ARG="--use_qpeft"
PEFT_DTYPE="bfloat16"
METHOD_NAME="q${METHOD_NAME}"
else
QPEFT_ARG=""
PEFT_DTYPE="float32"
fi
str="raw_codes"
OUTPUT_DIR=${SFT_EXPERIMENT_DIR}/${DATASET}${MODEL_SIZE}${METHOD_NAME}${HP_STRING}${str}
mkdir -p $OUTPUT_DIR

NUM_GPUS=1
BATCH_SIZE_PER_GPU=1
TOTAL_BATCH_SIZE=128
GRADIENT_ACC_STEPS=$(($TOTAL_BATCH_SIZE/$NUM_GPUS/$BATCH_SIZE_PER_GPU))
echo "Training llama model ${MODEL_SIZE} using $NUM_GPUS GPUs, $BATCH_SIZE_PER_GPU batch size per GPU, $GRADIENT_ACC_STEPS gradient accumulation steps" | tee -a $OUTPUT_DIR/log.txt

accelerate launch
--num_machines 1
--num_processes $NUM_GPUS
finetune/finetune.py
--model_name_or_path meta-llama/Llama-2-${MODEL_SIZE}-hf
$DATASET_ARG
--resume_from_checkpoint yes
--use_flash_attn
--seed 42
--peft_tuner $PEFT_METHOD
$QPEFT_ARG
--sft_num_tunable_weights ${NUM_TUNABLE_WEIGHTS}
--peft_dtype $PEFT_DTYPE
--sft_selection $SFT_METHOD
--lora_rank $RANK
--lora_alpha 16
--lora_dropout 0.1
--pt_num_virtual_tokens $PREFIX_LENGTH
--tokenizer_name meta-llama/Llama-2-${MODEL_SIZE}-hf
--use_slow_tokenizer
--max_seq_length 2048
--preprocessing_num_workers 16
--checkpointing_steps 50
--per_device_train_batch_size $BATCH_SIZE_PER_GPU
--gradient_accumulation_steps $GRADIENT_ACC_STEPS
--learning_rate ${LEARNING_RATE}
--lr_scheduler_type linear
--warmup_ratio 0.03
--weight_decay $WEIGHT_DECAY
--num_train_epochs 2
--output_dir $OUTPUT_DIR
--with_tracking
--report_to tensorboard
--logging_steps 1 2>&1 | tee -a $OUTPUT_DIR/log.txt

Hi, I'll put this here just in case someone faces this error after
pip install -r requirements.txt
It's an issue with flash-attention
Dao-AILab/flash-attention#453