defaults:
  - model:
      - qwen25_32b

megatron_training:
  model: qwen25_32b
  use_fused_rmsnorm: true
  use_mcore_models: true
  sequence_parallel: true
  use_flash_attn: true
  no_masked_softmax_fusion: true
  attention_softmax_in_fp32: true
  no_gradient_accumulation_fusion: true
  use_fused_swiglu: true
  use_fused_rotary_pos_emb: true
  bf16: true
  use_distributed_optimizer: true
  tokenizer_type: PretrainedFromHF
  tokenizer_name_or_path: ./Qwen2.5-32B-Instruct
  global_batch_size: 128
  seq_length: 2048
  save_interval: 200
  train_iters: 20000
  stage: ray_ppo
  attention_dropout: 0.0
  init_method_std: 0.01
  hidden_dropout: 0.0
  distributed_backend: nccl
  no_shared_storage: true
  no_pad_to_seq_lengths: true
  dataset_additional_keys: ['labels']
  data_path: ./data
  split: 100,0,0
  no_shuffle: false
  full_shuffle_instruction_dataset: false
  seed: 1

actor_config:
  model: qwen25_32b
  micro_batch_size: 1
  tensor_model_parallel_size: 8
  pipeline_model_parallel_size: 1
  lr: 1e-6
  lr_decay_style: constant
  min_lr: 0
  weight_decay: 0.01
  lr_warmup_fraction: 0.0
  clip_grad: 1.0
  adam_beta1: 0.9
  adam_beta2: 0.95
  finetune: true
  load: ./ckpt
  save: ./ckpt
  no_load_optim: true
  no_load_rng: true

critic_config:
  model: qwen25_32b
  tensor_model_parallel_size: 8
  pipeline_model_parallel_size: 1
  micro_batch_size: 1
  lr: 5e-6
  lr_decay_style: constant
  min_lr: 0
  weight_decay: 0.01
  lr_warmup_fraction: 0.0
  clip_grad: 1.0
  adam_beta1: 0.9
  adam_beta2: 0.95
  finetune: true
  no_load_optim: True
  no_load_rng: True
  load: ./ckpt
  save: ./ckpt

rl_config:
  guarantee_order: true
  use_integrated_worker: true
  blocking: true
  actor_forward_micro_batch_size: 1
  ref_forward_micro_batch_size: 1
  gamma: 1.0
  lam: 1.0
  adv_estimator: gae
  kl_penalty: kl
  kl_ctrl_type: fixed
  init_kl_coef: 0.0
  use_kl_in_reward: False
  mini_batch_size: 128
  max_prompt_length: 2048
  epochs: 1
  clip_ratio: 0.2
  cliprange_value: 0.5
  entropy_coeff: 0
  shuffle_mini_batch: false
  n_samples_per_prompt: 8
  rule_reward: true
  verifier_function: ["acc_for_ppo"]
  verifier_weight: [1.0]
  num_cpus_for_local_task: 1.0
  use_tensorboard: false
  actor_resource:
    num_npus: 32
  critic_resource:
    num_npus: 32

generate_config:
  enforce_eager: True
  trust_remote_code: true
  offload_train_optimizer: true
  offload_train_grad: true
  offload_train_param: true

  # 推理时的并行配置
  infer_tensor_parallel_size: 8
  infer_pipeline_parallel_size: 1
  infer_expert_parallel_size: 1

  # vllm 模型相关设置
  max_num_seqs: 1024
  max_model_len: 4096
  max_num_batched_tokens: 8192
  dtype: "bfloat16"
  gpu_memory_utilization: 0.4

  # 采样配置
  sampling_config:
    seed: 1
    logprobs: 1
    max_tokens: 2048
    top_p: 1
    top_k: -1
    min_p: 0.0
    temperature: 1.0
    detokenize: false