scripts/selfplay/7b_single_gpu.sh

#!/bin/bash
set -x

# Force use only GPU 4
export CUDA_VISIBLE_DEVICES=4

export VLLM_ATTENTION_BACKEND=FLASH_ATTN
export RAY_memory_monitor_refresh_ms=0
export RAY_LOGGING_LEVEL=DEBUG
export HYDRA_FULL_ERROR=1
export PYTHONPATH="${PYTHONPATH}:$(pwd)/verl"

OUTPUT_SEED_PATH=${OUTPUT_SEED_PATH:-data/7b_seed_io.jsonl}
OUTPUT_ERROR_SEED_PATH=${OUTPUT_ERROR_SEED_PATH:-data/7b_error_seed_io.jsonl}
OUTPUT_CODE_F_SEED_PATH=${OUTPUT_CODE_F_SEED_PATH:-data/7b_code_f_seed_io.jsonl}

python -m absolute_zero_reasoner.main_azr_ppo \
    data.shuffle=True \
    actor_rollout_ref.ref.include_ref=False \
    algorithm.adv_estimator=reinforce_plus_plus \
    data.train_files=data/code_reason/test_answer.parquet \
    data.val_files=data/code_reason/test_answer.parquet \
    data.train_batch_size=4 \
    data.val_batch_size=656 \
    data.max_prompt_length=4096 \
    data.max_response_length=2048 \
    azr.data_selection_strategy.content_max_length=5600 \
    actor_rollout_ref.model.path=Qwen/Qwen2.5-7B \
    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_kl_loss=False \
    actor_rollout_ref.actor.kl_loss_coef=0.0 \
    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.model.pretrained_tokenizer=True \
    actor_rollout_ref.actor.fsdp_config.param_offload=False \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=32 \
    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.max_num_batched_tokens=16384 \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
    actor_rollout_ref.rollout.enforce_eager=False \
    actor_rollout_ref.rollout.free_cache_engine=False \
    actor_rollout_ref.rollout.n=1 \
    actor_rollout_ref.rollout.temperature=1.0 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    algorithm.kl_ctrl.kl_coef=0.0 \
    trainer.critic_warmup=0 \
    trainer.logger=['console','wandb'] \
    trainer.project_name='azr' \
    trainer.experiment_name='azr_7b_single_gpu6' \
    trainer.n_gpus_per_node=1 \
    trainer.nnodes=1 \
    trainer.save_freq=10 \
    trainer.remove_previous_ckpt_in_save=True \
    trainer.del_local_ckpt_after_load=True \
    trainer.test_freq=10 \
    +trainer.val_before_train=False \
    reward_fn.extraction_type=answer_conditional \
    reward_fn.math_metric=math_verify \
    trainer.log_val_generations=0 \
    azr.data_selection_strategy.update_iteration=1 \
    azr.seed_dataset=$OUTPUT_SEED_PATH \
    azr.output_seed_path=$OUTPUT_SEED_PATH \
    azr.error_seed_dataset=$OUTPUT_ERROR_SEED_PATH \
    azr.output_error_seed_path=$OUTPUT_ERROR_SEED_PATH \
    azr.code_f_seed_dataset=$OUTPUT_CODE_F_SEED_PATH \
    azr.output_code_f_seed_path=$OUTPUT_CODE_F_SEED_PATH \
    azr.pretrain_pred_steps=-1 \
    azr.executor=qwq \
    azr.ast_check=True \
    azr.reward.n_samples=8 \
    azr.problem_types=['code_i','code_o','code_f'] \
    azr.data_selection_strategy.banned_keywords_for_errors_and_exceptions=['raise'] \
    trainer.debug=False \
    azr.reward.generation_reward_config.complexity_reward.coef=0.0 \
    azr.reward.generation_reward_config.complexity_reward.max=0.0 \
    azr.reward.generation_reward_config.complexity_reward.enabled=False \
    azr.reward.generation_reward_config.mean_edit_distance_reward.coef=0.0 \
    azr.reward.generation_reward_config.mean_edit_distance_reward.max=0.0 \
    azr.reward.generation_reward_config.mean_edit_distance_reward.enabled=False \
    azr.reward.generation_reward_config.halstead_reward.coef=0.0 \
    azr.reward.generation_reward_config.halstead_reward.max=0.0 \
    azr.reward.generation_reward_config.halstead_reward.enabled=False \
    azr.reward.generation_reward_config.answer_diversity_reward.coef=0.0 \
    azr.reward.generation_reward_config.answer_diversity_reward.max=0.0 \
    azr.reward.generation_reward_config.answer_diversity_reward.enabled=False \
    azr.reward.generation_reward_config.answer_diversity_reward.hierarchical=False \
    azr.pred_data_mix_strategy=max_new \
    azr.data_selection_strategy.seed_batch_factor=4 \
    azr.data_selection_strategy.valid_program_filter=all \
    azr.data_selection_strategy.max_programs=16384 \
    azr.data_selection_strategy.batched_estimate=False \
    azr.reward.generation_reward_config.intrinsic_combine_method=sum \
    azr.gen_data_probabilities_strategy=uniform \
    trainer.resume_mode=auto \
    azr.data_selection_strategy.composite_start_step=-1 \
    azr.data_selection_strategy.composite_chance=0.0 \
    azr.reward.generation_reward_config.remove_comments=False \
    azr.reward.generation_reward_config.remove_after_return=False \
    azr.reward.generation_reward_config.use_original_code_as_ref=True \
    azr.reward.generation_reward_config.remove_print=False \
    azr.data_selection_strategy.composite_function_n_min=0 \
    azr.data_selection_strategy.composite_function_n_max=0 \
    azr.reward.code_f_reward_type=binary \
    trainer.wandb_run_id=null \
    trainer.total_epochs=30 \
    azr.save_generated_data=True \
    azr.save_data_path="./generated_programs_7b_gpu6" \
    azr.save_valid_data=True \
    azr.save_invalid_data=True \
    azr.save_frequency=10 \
    azr.save_final_datasets=True $@

: <<'COMMENT'
  # 원본 (4개 GPU)                    → 수정 (1개 GPU)
  export CUDA_VISIBLE_DEVICES=6       # GPU 6번만 강제 사용
  trainer.n_gpus_per_node=4           → trainer.n_gpus_per_node=1
  tensor_model_parallel_size=2         → tensor_model_parallel_size=1
  ulysses_sequence_parallel_size=4     → ulysses_sequence_parallel_size=1

  2. 메모리 최적화

  # 배치 크기 조정 (메모리 절약)
  data.train_batch_size=64            → data.train_batch_size=32
  data.val_batch_size=1312            → data.val_batch_size=656
  ppo_mini_batch_size=128             → ppo_mini_batch_size=64

  # GPU 메모리 사용률 증가
  gpu_memory_utilization=0.4          → gpu_memory_utilization=0.8

  # 토큰 수 조정
  max_num_batched_tokens=16384        → max_num_batched_tokens=8192

  3. 실험 구분

  trainer.experiment_name='azr_7b'    → trainer.experiment_name='azr_7b_single_gpu6'
  azr.save_data_path="./generated_programs_7b" → azr.save_data_path="./generated_programs_7b_gpu6"
COMMENT