upload_code_run_ppo.sh

#!/bin/bash

deepspeed --num_gpus 1 main.py \
    --actor_model_name_or_path ../step1_supervised_finetuning/output_sft_en \
    --critic_model_name_or_path ../step2_reward_model_finetuning/output_rm_en \
    --data_path /home/ubuntu/DeepSpeedExamples/applications/DeepSpeed-Chat/training/step1_supervised_finetuning/datasets \
    --actor_zero_stage 2 \
    --critic_zero_stage 2 \
    --per_device_generation_batch_size 1 \
    --per_device_training_batch_size 1 \
    --num_padding_at_beginning 1 \
    --gradient_accumulation_steps 1 \
    --max_prompt_seq_len 64 \
    --max_answer_seq_len 128 \
    --actor_lora_dim 64 \
    --critic_lora_dim 64 \
    --dtype bf16 \
    --deepspeed \
    --enable_hybrid_engine \
    --actor_dropout 0.1 \
    --offload \
    --offload_reference_model \
    --offload_reward_model \
    --tensorboard_path ./output \
    --output_dir ./output &> ./output/training.log