mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-10-30 18:26:32 +00:00
8710794223
* init pr * add script * clean config folders * organization * up * tokenizer fixes * Fix dpo script tokenizer setup + minor fixes (#184) * copy tokenizer setup * minor fixes --------- Co-authored-by: Hamish Ivison <[email protected]>
36 lines
1.3 KiB
Bash
Executable File
36 lines
1.3 KiB
Bash
Executable File
export CUDA_VISIBLE_DEVICES=0,1,2,3
|
|
|
|
MODEL_SIZE=7B
|
|
NUM_GPUS=4
|
|
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"
|
|
|
|
# You can also set --gradient_checkpointing or use `stage3_offloading_accelerate.conf` to save memory,
|
|
# but it will trade off speed.
|
|
accelerate launch \
|
|
--mixed_precision bf16 \
|
|
--num_machines 1 \
|
|
--num_processes $NUM_GPUS \
|
|
--use_deepspeed \
|
|
--deepspeed_config_file configs/ds_configs/stage3_no_offloading_accelerate.conf \
|
|
open_instruct/finetune.py \
|
|
--model_name_or_path ../hf_llama2_models/${MODEL_SIZE} \
|
|
--use_flash_attn \
|
|
--tokenizer_name ../hf_llama2_models/${MODEL_SIZE} \
|
|
--use_slow_tokenizer \
|
|
--train_file data/processed/tulu_v2/tulu_v2_data.jsonl \
|
|
--max_seq_length 8192 \
|
|
--preprocessing_num_workers 128 \
|
|
--per_device_train_batch_size $BATCH_SIZE_PER_GPU \
|
|
--gradient_accumulation_steps $GRADIENT_ACC_STEPS \
|
|
--learning_rate 2e-5 \
|
|
--lr_scheduler_type linear \
|
|
--warmup_ratio 0.03 \
|
|
--weight_decay 0. \
|
|
--num_train_epochs 2 \
|
|
--output_dir output/tulu_v2_${MODEL_SIZE}/ \
|
|
--with_tracking \
|
|
--report_to tensorboard \
|
|
--logging_steps 1 |