# cd /home/mshahidul/readctrl/code/RL_model/verl/verl_train
set -x

unset PYTORCH_CUDA_ALLOC_CONF
export EXPERIMENT_NAME=qwen3-4b-instruct-en-h200-optimized
export WAND_PROJECT='readctrl-verl'
export CUDA_DEVICE_ORDER="PCI_BUS_ID"
export CUDA_VISIBLE_DEVICES=2,3 # Ensure these match your H200 indices

PYTHONUNBUFFERED=1 python3 -m verl.trainer.main_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=/home/mshahidul/readctrl/code/RL_model/verl/verl_train/dataset/train.parquet \
    data.val_files=/home/mshahidul/readctrl/code/RL_model/verl/verl_train/dataset/test.parquet \
    custom_reward_function.path=/home/mshahidul/readctrl/code/RL_model/verl/verl_train/reward_func/reward_func/reward_new_v2.py \
    data.train_batch_size=1024 \
    data.max_prompt_length=1024 \
    data.max_response_length=2048 \
    data.filter_overlong_prompts=True \
    data.truncation='error' \
    actor_rollout_ref.model.path=Qwen/Qwen3-4B-Instruct-2507 \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=512 \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=32 \
    actor_rollout_ref.actor.use_kl_loss=True \
    actor_rollout_ref.actor.kl_loss_coef=0.001 \
    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
    actor_rollout_ref.actor.entropy_coeff=0 \
    actor_rollout_ref.model.enable_gradient_checkpointing=False \
    actor_rollout_ref.actor.fsdp_config.param_offload=False \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=32 \
    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.rollout.enforce_eager=False \
    actor_rollout_ref.rollout.max_model_len=8192 \
    actor_rollout_ref.rollout.n=8 \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=64 \
    actor_rollout_ref.ref.fsdp_config.param_offload=False \
    algorithm.use_kl_in_reward=False \
    trainer.critic_warmup=0 \
    trainer.logger='["console"]' \
    trainer.project_name=$WAND_PROJECT \
    trainer.experiment_name=$EXPERIMENT_NAME \
    trainer.n_gpus_per_node=2 \
    trainer.nnodes=1 \
    trainer.save_freq=5 \
    trainer.test_freq=10 \
    +trainer.remove_previous_ckpt_in_save=true \
    trainer.max_actor_ckpt_to_keep=1 \
    trainer.max_critic_ckpt_to_keep=1 \
    trainer.resume_mode=auto \
    trainer.default_local_dir=/home/mshahidul/readctrl/code/RL_model/models/readCtrl_RL_en \
    trainer.total_epochs=15 $@ \
    2>&1 | tee $EXPERIMENT_NAME.log