export PYTORCH_CUDA_ALLOC_CONF=""
export EXPERIMENT_NAME=llm_guard_3B_10k_v2
export WAND_PROJECT='guard'
export CUDA_DEVICE_ORDER="PCI_BUS_ID"
export CUDA_VISIBLE_DEVICES=1,2
export VLLM_ATTENTION_BACKEND=FLASH_ATTN


PYTHONUNBUFFERED=1 NCCL_P2P_DISABLE=1 NCCL_IB_DISABLE=1 python3 -m verl.trainer.main_ppo \
    data.train_files=/home/mshahidul/readctrl/code/RL_model/verl/Search-R1/dataset/train.parquet \
    data.val_files=/home/mshahidul/readctrl/code/RL_model/verl/Search-R1/dataset/test.parquet \
    data.train_batch_size=64 \
    data.val_batch_size=64 \
    data.max_prompt_length=4096 \
    data.max_response_length=1024 \
    data.shuffle_train_dataloader=True \
    algorithm.adv_estimator=grpo \
    actor_rollout_ref.model.path=Qwen/Qwen3-4B-Instruct-2507 \
    actor_rollout_ref.model.enable_gradient_checkpointing=true \
    actor_rollout_ref.model.use_remove_padding=False \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.actor.ppo_mini_batch_size=64 \
    +actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=16 \
    actor_rollout_ref.actor.fsdp_config.param_offload=true \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=true \
    actor_rollout_ref.rollout.log_prob_micro_batch_size=64 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \
    actor_rollout_ref.ref.log_prob_micro_batch_size=64 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    actor_rollout_ref.actor.kl_loss_coef=0.001 \
    trainer.logger=['wandb'] \
    trainer.n_gpus_per_node=2 \
    trainer.nnodes=1 \
    trainer.save_freq=100 \
    trainer.test_freq=50 \
    trainer.project_name=$WANDB_PROJECT \
    trainer.experiment_name=$EXPERIMENT_NAME \
    trainer.total_epochs=15 \
    trainer.total_training_steps=1005 \
    trainer.default_local_dir=verl_checkpoints/$EXPERIMENT_NAME \
    do_search=false \
    max_turns=1 \
    2>&1 | tee $EXPERIMENT_NAME.log