set -x

export VLLM_ATTENTION_BACKEND=FLASH_ATTN
export PYTORCH_CUDA_ALLOC_CONF="expandable_segments:False"
export VLLM_USE_V1=1
export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
export VLLM_ENGINE_ITERATION_TIMEOUT_S=100000000000

# Find the directory where rllm package is located
RLLM_DIR=$(python3 -c "import rllm; import os; print(os.path.dirname(os.path.dirname(rllm.__file__)))")

# Train over a single node, 8 A100-80GB GPUs.
python3 -m rllm.trainer.verl.train_agent_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=${RLLM_DIR}/data/math_train.parquet \
    data.val_files=${RLLM_DIR}/data/math.parquet \
    data.train_batch_size=64 \
    data.val_batch_size=512 \
    data.max_prompt_length=2048 \
    data.max_response_length=2048 \
    actor_rollout_ref.model.path=deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B  \
    actor_rollout_ref.hybrid_engine=True \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=16 \
    actor_rollout_ref.actor.use_dynamic_bsz=True \
    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=24000 \
    actor_rollout_ref.actor.use_kl_loss=False \
    actor_rollout_ref.actor.kl_loss_coef=0.001 \
    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
    actor_rollout_ref.actor.ulysses_sequence_parallel_size=1 \
    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.tensor_model_parallel_size=1 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.mode="async" \
    actor_rollout_ref.rollout.chat_scheduler=verl.schedulers.completions_scheduler.CompletionsScheduler \
    actor_rollout_ref.rollout.temperature=0.6 \
    actor_rollout_ref.rollout.val_kwargs.temperature=0.6 \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.85 \
    actor_rollout_ref.rollout.n=4 \
    actor_rollout_ref.rollout.val_kwargs.n=1 \
    actor_rollout_ref.rollout.val_kwargs.top_p=0.95 \
    actor_rollout_ref.rollout.enforce_eager=False \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    algorithm.kl_ctrl.kl_coef=0.001 \
    trainer.critic_warmup=0 \
    trainer.logger=['console','wandb'] \
    trainer.project_name='deepscaler-agent' \
    trainer.experiment_name='deepscaler-math-debug-sync' \
    trainer.val_before_train=False \
    trainer.n_gpus_per_node=8 \
    trainer.nnodes=1 \
    trainer.save_freq=10 \
    trainer.test_freq=10 \
    trainer.default_hdfs_dir=null \
    env.name=math \
    agent.name=math_agent \
    agent.max_steps=1 \
    agent.async_engine=True \
    trainer.total_epochs=30 "${@:1}" \