code/RL_model/verl/verl_train/script/run_qwen3-8b_debug.sh

# 1. Force cleanup
pkill -9 python3
sleep 2

# 2. Set dynamic port to avoid collisions
export MASTER_PORT=$(shuf -i 20000-65000 -n 1)
export MASTER_ADDR=127.0.0.1

# 3. Enable P2P for performance (A100s love NVLink)
# unset NCCL_P2P_DISABLE
# unset NCCL_IB_DISABLE

set -x

# Enable P2P for A100s to leverage NVLink speed
export PYTORCH_CUDA_ALLOC_CONF=""
export EXPERIMENT_NAME=qwen3-4b-instruct-optimized-multiclinsum-gs
export WAND_PROJECT='readctrl-verl'
export CUDA_DEVICE_ORDER="PCI_BUS_ID"
export CUDA_VISIBLE_DEVICES=2,3 
export VLLM_ATTENTION_BACKEND=FLASH_ATTN
export NCCL_DEBUG=INFO
export NCCL_DEBUG_SUBSYS=ALL
export NCCL_P2P_DISABLE=1
export NCCL_IB_DISABLE=1
# export NCCL_NET_GDR_LEVEL=2 # Enable GPUDirect RDMA
# High-performance settings for A100
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.py \
    data.train_batch_size=8 \
    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=4 \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2 \
    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=True \
    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=2 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
    actor_rollout_ref.rollout.max_model_len=8192 \
    actor_rollout_ref.rollout.n=3 \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=32 \
    actor_rollout_ref.ref.fsdp_config.param_offload=False \
    algorithm.use_kl_in_reward=False \
    trainer.critic_warmup=0 \
    trainer.logger='["console","wandb"]' \
    trainer.project_name=$WAND_PROJECT \
    trainer.experiment_name=$EXPERIMENT_NAME \
    trainer.n_gpus_per_node=2 \
    trainer.nnodes=1 \
    trainer.save_freq=100 \
    trainer.test_freq=1 \
    +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/RL_model_subclaim_classifier \
    trainer.total_epochs=15 $@ \
    2>&1 | tee $EXPERIMENT_NAME.log

# python "/home/mshahidul/readctrl/code/readability_control.py"