{
  "hidden_size": 4096,
  "num_head_layers": 1,
  "num_heads": 8,
  "use_adaln": true,
  "use_mask_embeddings": true,
  "use_block_embeddings": false,
  "learning_rate": 1e-05,
  "weight_decay": 0.01,
  "num_epochs": 3,
  "max_grad_norm": 1.0,
  "per_gpu_policy_update_prompt_batch_size": 16,
  "group_size": 16,
  "per_gpu_sampling_mini_batch_size": 8,
  "per_gpu_ref_mini_batch_size": 8,
  "per_gpu_training_mini_batch_size": 2,
  "update_steps_per_rollouts": 2,
  "gradient_accumulation_steps": 32,
  "use_cached_hidden_states": true,
  "use_gradient_checkpointing": false,
  "use_step_probs": true,
  "optimize_denoiser": false,
  "denoiser_learning_rate": null,
  "denoiser_weight_decay": null,
  "kl_coef": 0.1,
  "clip_range": 0.2,
  "advantage_clip_min": null,
  "advantage_clip_max": null,
  "ratio_cap": null,
  "use_ref_kl": true,
  "reward_weights": {
    "judge": 1.0,
    "correctness": 1.0,
    "lookahead": 1.0,
    "format": 0.5
  },
  "alpha_values": [
    1.0
  ],
  "steps": 256,
  "gen_length": 512,
  "block_length": 64,
  "temperature": 1.0,
  "temp_explore": 1.0,
  "use_sampling_new": false,
  "mixed_precision": true,
  "reserve_teacher_gpu": true,
  "teacher_device_index": 7,
  "log_interval": 1,
  "eval_interval": 100,
  "save_interval": 1000,
  "use_wandb": true,
  "wandb_project": "planner-rl-8gpu",
  "wandb_run_name": "my-run-8gpu",
  "log_text_samples": 4,
  "debug_logging": false,
  "debug_log_interval": 1,
  "output_dir": "/accounts/projects/songmei/hengyuf/hengyu_scratch/hengyuf/ete-checkpoints/RL",
  "resume_from_checkpoint": null,
  "compute_logprob_tensor": false
}