# specify the default per-component configs
defaults:

  # <folder_name>@<field_name>.<field_name>: <yaml_file_name>
  # actor_rollout_ref.actor: trainer/config/actor/megatron_actor.yaml
  - actor@actor_rollout_ref.actor: megatron_actor
  # trainer.npu_profile: trainer/config/npu_profile/npu_profile.yaml
  - npu_profile@trainer.npu_profile: npu_profile
  # data: trainer/config/data/legacy_data.yaml
  - data@data: legacy_data
  # load the reference default config, then apply the fields in the current yaml
  # Reference model config.
  # Reference model will be enabled when actor.use_kl_loss or/and algorithm.use_kl_in_reward is/are True.
  - ref@actor_rollout_ref.ref: megatron_ref
  # Rollout model config.
  - rollout@actor_rollout_ref.rollout: rollout
  # Critic model config.
  - critic@critic: megatron_critic
  # Reward model config.
  - reward_model@reward_model: megatron_reward_model
  - _self_

actor_rollout_ref:
  hybrid_engine: True

  nccl_timeout: 600 # seconds, default is 10 minutes for torch, you can set it to a larger value if you have long-running operations like 32B or 72B model using megatron

  model:

    path: ~/models/deepseek-llm-7b-chat

    custom_chat_template: null

    external_lib: null

    override_config:

      model_config: {}

      moe_config:

        freeze_moe_router: False

    use_fused_kernels: False # Whether to use custom fused kernels (PostProcessing, for memory efficiency)

    trust_remote_code: False
    
  rollout:
    # may get higher throughput when set to True. When activated, Please increase max_num_batched_tokens or decrease max_model_len.
    enable_chunked_prefill: False

    load_format: dummy_megatron

    tensor_model_parallel_size: 1

    layer_name_map:
      qkv_layer_name: qkv
      gate_proj_layer_name: gate_up

  profiler:
    _target_: verl.utils.profiler.ProfilerConfig
    discrete: False
    all_ranks: False
    ranks: []

custom_reward_function:
  path: null
  name: compute_score

algorithm:
  # Required when using verl.utils.omega_conf_to_dataclass to instantiate dataclass configs
  _target_: verl.trainer.config.AlgoConfig
  gamma: 1.0
  lam: 1.0
  adv_estimator: gae
  norm_adv_by_std_in_grpo: True
  use_kl_in_reward: False
  kl_penalty: kl  # how to estimate kl divergence
  kl_ctrl:
    # Required when using verl.utils.omega_conf_to_dataclass to instantiate dataclass configs
    _target_: verl.trainer.config.KLControlConfig
    type: fixed
    kl_coef: 0.001
    horizon: 10000
    target_kl: 0.1
  use_pf_ppo: False
  pf_ppo:
    reweight_method: pow  # ["pow", "max_min", "max_random"]
    weight_pow: 2.0

trainer:
  balance_batch: True
  total_epochs: 30
  total_training_steps: null
  profile_steps: null # [1,2,5] or [] or null
  project_name: verl_examples
  experiment_name: gsm8k
  logger: ['console', 'wandb']
  log_val_generations: 0
  nnodes: 1
  n_gpus_per_node: 8
  save_freq: -1
  esi_redundant_time: 0

  # auto: find the last ckpt to resume. If can't find, start from scratch
  resume_mode: auto # or disable or resume_path if resume_from_path is set
  resume_from_path: null
  del_local_ckpt_after_load: False
  val_before_train: True
  test_freq: -1
  critic_warmup: 0
  default_hdfs_dir: null
  default_local_dir: checkpoints/${trainer.project_name}/${trainer.experiment_name}
  max_actor_ckpt_to_keep: null
  max_critic_ckpt_to_keep: null
  # The timeout for ray worker group to wait for the register center to be ready
  ray_wait_register_center_timeout: 300
  device: cuda
  # see ppo_trainer.yaml for more details
  controller_nsight_options:
    trace: "cuda,nvtx,cublas,ucx"
    cuda-memory-usage: "true"
    cuda-graph-trace: "graph"
  worker_nsight_options:
    trace: "cuda,nvtx,cublas,ucx"
    cuda-memory-usage: "true"
    cuda-graph-trace: "graph"
    capture-range: "cudaProfilerApi"
    capture-range-end: null
    kill: none
ray_init:
  num_cpus: null # `None` means using all CPUs, which might cause hang if limited in systems like SLURM. Please set to a number allowed then.
  timeline_json_file: null