#!/usr/bin/env bash
set -euo pipefail

# Reproduces the paper's balanced DualStake setting on Qwen2.5-7B.
# Override paths and hardware settings through environment variables.
PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
DATA_DIR="${DATA_DIR:-${PROJECT_ROOT}/data/nq_hotpotqa}"
MODEL_PATH="${MODEL_PATH:-Qwen/Qwen2.5-7B}"
RETRIEVER_URL="${RETRIEVER_URL:-http://127.0.0.1:8000/retrieve}"
EXPERIMENT_NAME="${EXPERIMENT_NAME:-dualstake-qwen2.5-7b}"
NUM_GPUS="${NUM_GPUS:-8}"
NUM_NODES="${NUM_NODES:-1}"

export VLLM_ATTENTION_BACKEND="${VLLM_ATTENTION_BACKEND:-XFORMERS}"

cd "$PROJECT_ROOT"
PYTHONUNBUFFERED=1 python -m verl.trainer.main_ppo \
    data.train_files="${DATA_DIR}/train.parquet" \
    data.val_files="${DATA_DIR}/test.parquet" \
    data.train_data_num=null \
    data.val_data_num=1024 \
    data.train_batch_size=64 \
    data.val_batch_size=256 \
    data.max_prompt_length=4096 \
    data.max_response_length=3072 \
    data.max_start_length=2048 \
    data.max_obs_length=500 \
    data.shuffle_train_dataloader=true \
    algorithm.adv_estimator=grpo \
    actor_rollout_ref.model.path="$MODEL_PATH" \
    actor_rollout_ref.model.enable_gradient_checkpointing=true \
    actor_rollout_ref.model.use_remove_padding=true \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.actor.optim.lr_warmup_steps_ratio=0.95 \
    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.ppo_mini_batch_size=64 \
    actor_rollout_ref.actor.ppo_micro_batch_size=8 \
    actor_rollout_ref.actor.fsdp_config.param_offload=true \
    actor_rollout_ref.actor.fsdp_config.grad_offload=true \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=true \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.n=5 \
    actor_rollout_ref.rollout.n_agent=5 \
    actor_rollout_ref.rollout.temperature=1.0 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
    actor_rollout_ref.rollout.log_prob_micro_batch_size=8 \
    actor_rollout_ref.ref.log_prob_micro_batch_size=8 \
    actor_rollout_ref.ref.fsdp_config.param_offload=true \
    actor_rollout_ref.actor.state_masking=true \
    algorithm.no_think_rl=false \
    trainer.logger='[console,wandb]' \
    trainer.project_name=dualstake \
    trainer.experiment_name="$EXPERIMENT_NAME" \
    trainer.n_gpus_per_node="$NUM_GPUS" \
    trainer.nnodes="$NUM_NODES" \
    trainer.save_freq=50 \
    trainer.test_freq=50 \
    trainer.total_training_steps=300 \
    trainer.default_hdfs_dir=null \
    trainer.default_local_dir="checkpoints/${EXPERIMENT_NAME}" \
    max_turns=4 \
    retriever.url="$RETRIEVER_URL" \
    retriever.topk=3