cog-llama-test / train_model.sh
khoicrtp's picture
Update train_model.sh
45bb876
raw
history blame
703 Bytes
#!/bin/bash
torchrun --nproc_per_node=1 --master_port=9292 train.py \
--tokenizer_name_or_path /src/weights/tokenizer \
--data_path ./alpaca_data.json \
--model_name_or_path /src/weights/llama-7b \
--output_dir alpaca_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 "steps" \
--save_steps 2000 \
--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' \