finetuning_script.txt

FineTune Vicuna Commands
#Rent GPUs with setting of pytorch:latest, ssh connection
#Allocate 80 GB of disk space

ssh -p YOUR_PORT_NUMBER root@YOUR_IP

pip3 install fschat
git clone https://github.com/lm-sys/FastChat.git
cd FastChat
pip3 install --upgrade pip
pip3 install -e .
pip install einops
cd ..
git clone https://github.com/HazyResearch/flash-attention.git
cd flash-attention
conda install -c conda-forge cudatoolkit-dev
sudo apt install g++
python setup.py install
cd ..
pip install protobuf==3.20.3 --upgrade

#ctrl b, then press d

scp -P YOUR_PORT_NUMBER /YOUR_DATA_PATH root@YOUR_IP:FastChat/data/dummy_conversation.json

ssh -p YOUR_PORT_NUMBER root@YOUR_IP

cd FastChat

torchrun --nproc_per_node=YOUR_NUMBER_OF_GPUS --master_port=20001 fastchat/train/train_mem.py \
    --model_name_or_path yahma/llama-7b-hf  \
    --data_path data/dummy_conversation.json \
    --bf16 True \
    --output_dir output_vicuna \
    --num_train_epochs 3 \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 2 \
    --gradient_accumulation_steps 16 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 1200 \
    --save_total_limit 10 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --fsdp "full_shard auto_wrap" \
    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
    --tf32 True \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --lazy_preprocess True

#AFTER TRAINING IS DONE
apt-get install zip
zip -r YOUR_MODEL_NAME.zip output_vicuna/
scp -P YOUR_PORT_NUMBER root@YOUR_IP:FastChat/YOUR_MODEL_NAME /YOUR_MODEL_OUTPUT_PATH