finetune-i2b2.sh

#!/bin/bash

#source /media/nmitchko/NVME/text-generation-ui/venv/bin/activate
source /media/nmitchko/NVME/text-generation-webui/venv/bin/activate
CURRENTDATEONLY=`date +"%b %d %Y"`

# Change Power limit to 190 for training
sudo nvidia-smi -i 1 -pl 250
sudo nvidia-smi -i 0 -pl 250

export CUDA_VISIBLE_DEVICES=0,1

accelerate launch --num_processes 2 qlora.py \
    --ddp_find_unused_parameters False \
    --model_name_or_path /media/nmitchko/NVME/text-generation-webui/models/codellama_CodeLlama-34b-hf \
    --output_dir /media/ai/blk/loras/i2b2training \
    --logging_steps 100 \
    --save_strategy steps \
    --data_seed 42 \
    --save_steps 200 \
    --save_total_limit 40 \
    --evaluation_strategy steps \
    --eval_dataset_size 1024 \
    --max_eval_samples 1000 \
    --per_device_eval_batch_size 2 \
    --per_device_train_batch_size 2 \
    --trust_remote_code True \
    --use_auth_token False \
    --max_new_tokens 32 \
    --dataloader_num_workers 2 \
    --group_by_length \
    --logging_strategy steps \
    --remove_unused_columns False \
    --do_train \
    --lora_r 64 \
    --lora_alpha 16 \
    --lora_modules all \
    --double_quant \
    --quant_type nf4 \
    --bf16 \
    --bits 4 \
    --legacy=False \
    --warmup_ratio 0.03 \
    --lr_scheduler_type constant \
    --gradient_checkpointing \
    --dataset="i2b2.json" \
    --dataset_format alpaca \
    --trust_remote_code=True \
    --source_max_len 16 \
    --target_max_len 512 \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 16 \
    --max_steps 4500 \
    --eval_steps 1000 \
    --learning_rate 0.0001 \
    --adam_beta2 0.999 \
    --max_grad_norm 0.3 \
    --lora_dropout 0.05 \
    --weight_decay 0.0 \
    --seed 0 > "${CURRENTDATEONLY}-finetune-i2b2.log" &


# Change Power limit to 300 for normal activities training
# Not Needed for non-managed script

deactivate