| num_gpus=1 | |
| deepspeed --num_gpus $num_gpus train.py \ | |
| --deepspeed ./ds_config.json \ | |
| --output_dir="./output/MiniCPM" \ | |
| --per_device_train_batch_size=4 \ | |
| --gradient_accumulation_steps=4 \ | |
| --logging_steps=10 \ | |
| --num_train_epochs=3 \ | |
| --save_steps=500 \ | |
| --learning_rate=1e-4 \ | |
| --save_on_each_node=True \ |