model:
  pretrained_path: models/AnyFlow-Wan2.1-T2V-1.3B-Diffusers
  generator_init_path: ''
  aux_v_init_path: ''
  model_type: wan
  gradient_checkpointing: true
  dtype: bf16
  generator_lora:
    enabled: true
    rank: 32
    lora_alpha: 64
    target_modules:
    - to_q
    - to_k
    - to_v
    - to_out.0
    - net.0.proj
    - net.2
    init_lora_weights: gaussian
    load_path: ''
    pre_merge_paths: []
    merge_before_training: false
video:
  num_frames: 81
  eval_num_frames: 81
  height: 480
  width: 832
  fps: 16
  max_sequence_length: 512
fewstep_rl:
  num_epochs: 250
  num_batches_per_epoch: 2
  num_inner_epochs: 1
  rl_inner_batch_size: 0
  rl_start_epoch: 0
  num_image_per_prompt: 16
  rl_sample_use_ema: false
  sampling_num_steps: 4
  sampling_guidance_scale: 1.0
  per_prompt_stat_tracking: true
  per_prompt_global_std: true
  weight_advantages: true
  adv_clip_max: 5.0
  beta: 0.1
  kl_weight: 0.0001
  decay_type: 1
  rl_decay_interval: 1
  reward_fn:
    hpsv3_general: 1.0
    hpsv3_percentile: 1.0
    videoalign_mq: 1.0
    videoalign_ta: 1.0
  reward_model_paths:
    videoalign_dir: third_party/VideoAlign
    videoalign_ckpt: models/VideoReward
  log_sample_images: 0
  central_diff_epsilon: 5.0
  cd_velocity_source: noise_minus_x0
  share_cd_with_old: true
  disable_cd: true
  num_training_timesteps_per_sample: 4
  rl_min_timestep: 0
  rl_max_timestep: 1000
  diffusion_ratio: 0.5
  consistency_ratio: 0.25
  use_ref_v_model: false
  guidance_scale: 1.0
  rollout_type: ode
  clip_range: 0.001
  reuse_x_t: false
  aux_v_enabled: false
  max_aux_sample: 0
  renoise_per_epoch: 10
  aux_v_gradient_accum: 1
  aux_v_batch_size: 0
  opsd_enabled: true
  opsd_rho: 0.1
  opsd_query_sigma: 0.278
  opsd_n_ascent: 2
  opsd_eta_tgt: 1.0
  opsd_reward_kind: hpsv3_general
  opsd_target_chunk: 1
  opsd_eps_grad: 1.0e-12
  opsd_eps_gamma: 1.0e-05
  opsd_dual_neg: true
  opsd_state_mode: forward
  opsd_sigma_mode: lognormal
  opsd_logit_mean: -2.053
  opsd_logit_std: 0.6
  opsd_r_mode: given_t
  refl_use_advantage: false
  opsd_dir_mode: grad
  reward_ckpt_path: ''
  grpo_noise_level: 0.7
solver:
  generator:
    lr: 3.0e-06
    warmup_steps: 0
    beta1: 0.9
    beta2: 0.999
    eps: 1.0e-08
    weight_decay: 0.0001
    max_grad_norm: 1.0
  aux_v:
    lr: 1.0e-05
    warmup_steps: 0
    beta1: 0.9
    beta2: 0.999
    eps: 1.0e-08
    weight_decay: 0.0
    max_grad_norm: 1.0
train:
  total_steps: 0
  save_interval: 50
  log_interval: 1
  seed: 42
  batch_size: 2
  gradient_accumulation_steps: 1
  autocast_dtype: bf16
  resume_from: ''
  use_ema: true
  ema_decay: 0.9
  ema_every: 1
distributed:
  strategy: fsdp
  fsdp_sharding: hybrid
  ddp_find_unused_parameters: null
logging:
  project: fewsteprl-wan21
  run_name: meanflowopsd
  entity: ''
  tags:
  - wan2.1
  - video
  - anyflow
  - meanflow-nft
  - 480p
  - 81f
  - rpd
  - opsd
  - differentiable-reward
  - x0-form
  enabled: true
eval:
  enabled: true
  eval_interval: 10
  eval_prompt_source: dataset
  eval_prompt_path: ''
  dataset: wan_dancegrpo
  dataset_path: dataset
  reward_fn:
    hpsv3_general: 1.0
    hpsv3_percentile: 1.0
    videoalign_mq: 1.0
    videoalign_ta: 1.0
  reward_dataset_map: {}
  reward_ckpt_path: ''
  eval_num_steps: 4
  eval_guidance_scale: 1.0
  eval_batch_size: 1
  num_media_images: 4
  eval_seed: 12345
  vbench:
    enabled: false
    aug_info_json: dataset/vbench/VBench_aug_full_info.json
    full_info_json: dataset/vbench/VBench_full_info.json
    num_samples_per_prompt: 5
    num_inference_steps: 4
    guidance_scale: 1.0
    cache_dir: models/vbench
    output_subdir: vbench
    eval_every_epochs: 0
prompt_path: dataset/wan_dancegrpo
output_dir: ./outputs/meanflowopsd