llama-fine-tuning-notes.txt

sudo apt update
sudo apt install -y vim git python3-dev gcc g++ make unzip tree python3.10-venv iftop ripgrep

sudo -v ; curl https://rclone.org/install.sh | sudo bash

mkdir -p ~/.config/rclone/
vim ~/.config/rclone/rclone.conf

# ~/.config/rclone/rclone.conf
[r2]
type = s3
provider = Cloudflare
access_key_id = xxx
secret_access_key = xxx
endpoint = https://xxx.r2.cloudflarestorage.com
acl = private

rclone tree r2:
rclone copy r2:/llama/models/7B ./models/7B
rclone copy r2:/llama/models/tokenizer.model ./models/tokenizer.model

git clone https://github.com/huggingface/transformers
python3 -m venv .venv
source .venv/bin/activate
pip install torch~=1.13
pip install tokenizers==0.13.3
pip install protobuf~=3.20
pip install accelerate sentencepiece

git clone https://github.com/tatsu-lab/stanford_alpaca
cd stanford_alpaca
pip install -r requirements.txt

# on host:
scp output.json [email protected]:/home/ubuntu/

cd transformers
python src/transformers/models/llama/convert_llama_weights_to_hf.py --input_dir ../models --model_size 7B --output_dir ../hf-output
python convert_llama_weights_to_hf.py --input_dir ./models/ --model_size 7B --output_dir ./hf-output

torchrun \
    --nproc_per_node=4 \
    --master_port=55444 \
    train.py \
    --model_name_or_path ../hf-output/ \
    --data_path ../output.json \
    --bf16 True \
    --output_dir ../training-out/ \
    --num_train_epochs 3 \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 8 \
    --evaluation_strategy "no" \
    --save_strategy "no" \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 100000 \
    --fsdp "full_shard auto_wrap" \
    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
    --tf32 True