# Distributed training runtime.
infra:
  sequence_parallel_size: 4
  sharding_strategy: hybrid_full
  mixed_precision: true
  gradient_checkpointing: true
  vae_halo_latents: 28
  generator_fsdp_wrap_strategy: size
  text_encoder_fsdp_wrap_strategy: size

# Passed directly to WanDiffusionWrapper(**model_kwargs).
model_kwargs:
  model_name: Wan2.2-TI2V-5B
  timestep_shift: 5.0
  num_frame_per_block: 8
  local_attn_size: -1

# AR diffusion objective.
algorithm:
  trainer: diffusion
  causal: true
  teacher_forcing: true
  num_train_timestep: 1000
  denoising_loss_type: flow

training:
  lr: 1.0e-05
  weight_decay: 0.0
  beta1: 0.0
  beta2: 0.999
  batch_size: 1
  gradient_accumulation_steps: 1
  ema_weight: 0.99
  ema_start_step: 100
  log_iters: 1
  max_checkpoints: 20
  max_iters: 600
  gc_interval: 100
  error_recycling:
    enabled: true
    num_buckets: 50
    buffer_size_per_bucket: 500
    buffer_warmup_iter: 50
    context_inject_prob: 0.9
    latent_inject_prob: 0.9
    noise_inject_prob: 0.01
    clean_prob: 0.2
    clean_buffer_update_prob: 0.1
    modulate_factor: 0.3
    start_step: 0
    replacement_strategy: random

data:
  data_path: /path/to/longlive2/train_dataset
  eval_data_path: /path/to/longlive2/eval_prompts
  image_or_video_shape:
  - 1
  - 96
  - 48
  - 44
  - 80
  load_raw_video: true

# Generation settings used by training-time evaluation.
inference:
  sampling_steps: 50
  sink_size: 0
  multi_shot_sink: true
  multi_shot_rope_offset: 8
  guidance_scale: 3.0
  negative_prompt: '色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走'

# Lightweight validation generation during AR training.
evaluation:
  interval: 1
  num_frames: 16
  use_ema: true
  val_batch_size: 1
  save_latents_only: true

logging:
  seed: 0
  wandb_key: null
  wandb_entity: null
  wandb_project: LongLive2-AR