# Tested successfully on the hiyouga/verl:ngc-th2.6.0-cu126-vllm0.8.4-flashinfer0.2.2-cxx11abi0 image.
# It outperforms the Qwen2 7B base model by two percentage points on the test set of GSM8K.

set -x

export PYTHONPATH=/mnt/zhuli_nas/qiqiuyi/verl_test_kl/TravelPlanner:$PYTHONPATH

export HF_ENDPOINT=https://hf-mirror.com

python3 -m verl.trainer.main_ppo_travelplanner \
    algorithm.adv_estimator=grpo \
    data.train_files=/mnt/zhuli_nas/qiqiuyi/deepscaler_planning_test_kl2/data/datasets/travelplanner_rl_train_revision_easy_example_expanded_fined.parquet \
    data.val_files=/mnt/zhuli_nas/qiqiuyi/deepscaler_planning_test_kl2/data/datasets/travelplanner_rl_validation_revision_easy_example.parquet \
    data.train_batch_size=64 \
    data.val_batch_size=512 \
    data.max_prompt_length=16384 \
    data.max_response_length=8192 \
    data.filter_overlong_prompts=True \
    data.truncation='error' \
    data.shuffle=False \
    +data.prompt_key_incomplete=prompt_without_constraint \
    actor_rollout_ref.model.path=/mnt/zhuli_nas/longhan/verl/models/Qwen3-8B \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_mini_batch_size=64 \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \
    actor_rollout_ref.actor.use_dynamic_bsz=True \
    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=32768 \
    actor_rollout_ref.actor.use_kl_loss=True \
    actor_rollout_ref.actor.kl_loss_coef=0.001 \
    +actor_rollout_ref.actor.kl_reward_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.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=4 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.temperature=0.6 \
    actor_rollout_ref.rollout.val_kwargs.temperature=0.6 \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
    actor_rollout_ref.rollout.max_num_batched_tokens=32768 \
    actor_rollout_ref.rollout.n=8 \
    actor_rollout_ref.rollout.val_kwargs.n=1 \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=4 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    algorithm.use_kl_in_reward=False \
    trainer.critic_warmup=0 \
    trainer.logger=['console'] \
    trainer.project_name='travelplanner' \
    trainer.experiment_name='travelplanner-qwen3-8b-8k-mix' \
    trainer.val_before_train=True \
    trainer.n_gpus_per_node=8 \
    trainer.nnodes=2 \
    trainer.resume_mode='auto' \
    trainer.save_freq=25 \
    trainer.test_freq=10 \
    trainer.default_hdfs_dir=null \
    trainer.total_epochs=300 $@ \
    trainer.default_local_dir='/mnt/zhuli_nas/qiqiuyi/trained_models/verl_test_kl/checkpoints/${trainer.project_name}/${trainer.experiment_name}' \
    +reward_model.type='travelplanner'





    # actor_rollout_ref.actor.fsdp_config.grad_offload=False \