train_dpo.yaml

data:
  tokenizer: null
  train_files: /train.parquet  # useless
  val_files: /test.parquet # useless
  prompt_key: prompt
  max_prompt_length: 1792
  max_response_length: 256
  train_batch_size: 32
  val_batch_size: null
  return_raw_input_ids: False  # This should be set to true when the tokenizer between policy and rm differs
  return_raw_chat: False
  shuffle: True
  filter_overlong_prompts: False # for large-scale dataset, filtering overlong prompts could be timeconsuming. You should disable this and set `truncation='left'
  truncation: error
  image_key: images

actor_rollout_ref:
  hybrid_engine: True
  model:
    path: /PATH/TO/MODEL/CHECKPOINT/
    external_lib: null
    override_config: { }
    enable_gradient_checkpointing: True
    use_remove_padding: False
  actor:
    alg_type: dpo
    strategy: fsdp  # This is for backward-compatibility
    ppo_mini_batch_size: 32
    # ppo_micro_batch_size: 8 # will be deprecated, use ppo_micro_batch_size_per_gpu
    ppo_micro_batch_size_per_gpu: 2 # NOTE
    use_dynamic_bsz: False
    ppo_max_token_len_per_gpu: 16384 # n * ${data.max_prompt_length} + ${data.max_response_length}
    grad_clip: 1.0
    clip_ratio: 0.2
    entropy_coeff: 0.001
    use_kl_loss: True # NOTE
    kl_loss_coef: 0.1 # NOTE: beta for DPO
    kl_loss_type: low_var_kl # for grpo
    ppo_epochs: 1
    shuffle: False
    ulysses_sequence_parallel_size: 1 # sp size
    optim:
      lr: 5e-7
      lr_warmup_steps_ratio: 0.03  # the total steps will be injected during runtime
      min_lr_ratio: 0.1   # only useful for warmup with cosine
      warmup_style: cosine  # select from constant/cosine
      total_training_steps: 783  #
      beta1: 0.9
      beta2: 0.95
    fsdp_config:
      wrap_policy:
        # transformer_layer_cls_to_wrap: None
        min_num_params: 0
      param_offload: False
      optimizer_offload: False
      fsdp_size: -1
  ref:
    fsdp_config:
      param_offload: False
      wrap_policy:
        # transformer_layer_cls_to_wrap: None
        min_num_params: 0
    # log_prob_micro_batch_size: 4 # will be deprecated, use log_prob_micro_batch_size_per_gpu
    log_prob_micro_batch_size_per_gpu: 2 # NOTE
    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
    ulysses_sequence_parallel_size: ${actor_rollout_ref.actor.ulysses_sequence_parallel_size} # sp size
  rollout:
    name: vllm
    temperature: 1.0
    top_k: -1 # 0 for hf rollout, -1 for vllm rollout
    top_p: 1
    use_fire_sampling: False # https://arxiv.org/abs/2410.21236
    prompt_length: ${data.max_prompt_length}  # not use for opensource
    response_length: ${data.max_response_length}
    # for vllm rollout
    dtype: bfloat16 # should align with FSDP
    gpu_memory_utilization: 0.4
    ignore_eos: False
    enforce_eager: True
    free_cache_engine: True
    load_format: dummy_dtensor
    tensor_model_parallel_size: 2
    max_num_batched_tokens: 8192
    max_model_len: null
    max_num_seqs: 1024
    # log_prob_micro_batch_size: 8 # will be deprecated, use log_prob_micro_batch_size_per_gpu
    log_prob_micro_batch_size_per_gpu: 4
    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
    disable_log_stats: True
    enable_chunked_prefill: True # could get higher throughput
    # for hf rollout
    do_sample: True

critic:
  strategy: fsdp
  optim:
    lr: 1e-5
    lr_warmup_steps_ratio: 0.  # the total steps will be injected during runtime
    # min_lr_ratio: null   # only useful for warmup with cosine
    warmup_style: constant  # select from constant/cosine
    total_training_steps: 783  # must be override by program
  model:
    path: /PATH/TO/MODEL/CHECKPOINT/
    tokenizer_path: ${actor_rollout_ref.model.path}
    override_config: { }
    external_lib: ${actor_rollout_ref.model.external_lib}
    enable_gradient_checkpointing: True
    use_remove_padding: False
    fsdp_config:
      param_offload: False
      optimizer_offload: False
      wrap_policy:
        # transformer_layer_cls_to_wrap: None
        min_num_params: 0
      fsdp_size: -1
  ppo_mini_batch_size: ${actor_rollout_ref.actor.ppo_mini_batch_size}
  # ppo_micro_batch_size: 8 # will be deprecated, use ppo_micro_batch_size_per_gpu
  ppo_micro_batch_size_per_gpu: 1
  forward_micro_batch_size_per_gpu: ${critic.ppo_micro_batch_size_per_gpu}
  use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
  ppo_max_token_len_per_gpu: 32768 # (${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}) * 2
  forward_max_token_len_per_gpu: ${critic.ppo_max_token_len_per_gpu}
  ulysses_sequence_parallel_size: 1 # sp size
  ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
  shuffle: ${actor_rollout_ref.actor.shuffle}
  grad_clip: 1.0
  cliprange_value: 0.5

reward_model:
  enable: False
  strategy: fsdp
  model:
    input_tokenizer: ${actor_rollout_ref.model.path}  # set this to null if the chat template is identical
    path: ~/models/FsfairX-LLaMA3-RM-v0.1
    external_lib: ${actor_rollout_ref.model.external_lib}
    use_remove_padding: False
    fsdp_config:
      min_num_params: 0
      param_offload: False
      fsdp_size: -1
  # micro_batch_size: null # will be deprecated, use micro_batch_size_per_gpu
  # micro_batch_size_per_gpu: 2 # set a number
  # max_length: null
  ulysses_sequence_parallel_size: 1 # sp size
  use_dynamic_bsz: ${critic.use_dynamic_bsz}
  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}

custom_reward_function:
  path: null
  name: compute_score

algorithm:
  gamma: 1.0
  lam: 1.0
  adv_estimator: grpo
  kl_penalty: kl
  kl_ctrl:
    type: fixed
    kl_coef: 0.001

trainer:
  balance_batch: False
  total_epochs: 1  #
  total_training_steps: 783 #
  project_name: dpo_example
  experiment_name: trinity_dpo
  logger: [ 'console','wandb' ]
  val_generations_to_log_to_wandb: 0
  nnodes: 1
  n_gpus_per_node: 2
  save_freq: 30
  # auto: find the last ckpt to resume. If can't find, start from scratch
  resume_mode: auto # or auto or resume_path if
  resume_from_path: False
  test_freq: 5
  critic_warmup: 0
  default_hdfs_dir: null
  remove_previous_ckpt_in_save: False
  del_local_ckpt_after_load: False
  default_local_dir: checkpoints/${trainer.project_name}/${trainer.experiment_name}
  val_before_train: False