# Wan2.1-T2V-14B CFG-only LoRA distillation, native 50 -> 50 steps.
#
# CONTINGENCY ONLY: this profile resumes the DP16/GA2 step-92 checkpoint after
# an explicitly receipted, sample-sequence-preserving topology migration.  It
# is scientifically equivalent at global batch 32, but not bitwise exact.
# There is still no SFP/few-step rollout, learned critic, fake-score model, or
# four-step schedule.
infra:
  sequence_parallel_size: 1
  sharding_strategy: full
  mixed_precision: true
  gradient_checkpointing: true
  generator_fsdp_wrap_strategy: size
  real_score_fsdp_wrap_strategy: size
  text_encoder_fsdp_wrap_strategy: size
  text_encoder_cpu_offload: false

model_kwargs:
  model_name: Wan2.1-T2V-14B
  model_dir: wan_models/Wan2.1-T2V-14B
  timestep_shift: 5.0
  num_frame_per_block: 3
  seq_len: 32760

checkpoints:
  generator_ckpt: null
  real_score_ckpt: null
  lora_ckpt: null

algorithm:
  trainer: score_distillation
  distribution_loss: cfg_guidance
  sfp_training: false
  backward_simulation: false
  all_causal: false
  generator_is_causal: false
  real_score_is_causal: false
  cfg_state_source: teacher_trajectory_cache
  teacher_guidance_scale: 5.0
  teacher_sampling_steps: 50
  cfg_distill_loss_weighting: relative_guidance
  cfg_relative_guidance_epsilon: 1.0e-6
  cfg_relative_guidance_max_weight: 16.0
  cfg_verify_cache_hashes: false

training:
  lr: 1.0e-05
  weight_decay: 0.01
  beta1: 0.0
  beta2: 0.999
  batch_size: 1
  # 8 DP ranks x 4 microbatches preserves the source global batch of 32.
  gradient_accumulation_steps: 4
  ema_weight: 0.0
  ema_start_step: 200
  log_iters: 50
  max_checkpoints: 20
  max_iters: 500
  gc_interval: 25
  max_grad_norm_generator: 10.0

data:
  # The launcher overrides this with the source checkpoint's exact data_id.
  data_path: data/cfg_teacher_trajectory_wan21_14b_64/manifest.json
  eval_data_path: null
  data_num_workers: 0
  data_seed: 0
  image_or_video_shape:
  - 1
  - 21
  - 16
  - 60
  - 104
  load_raw_video: false
  uniform_prompt: true
  allow_padding: false

inference:
  sampling_steps: 50
  guidance_scale: 1.0
  sink_size: 0
  multi_shot_rope_offset: 0

evaluation:
  interval: -1
  before_train: false
  num_frames: 21
  use_ema: false
  val_batch_size: 1
  save_latents_only: true

adapter:
  type: lora
  rank: 128
  alpha: 128
  dropout: 0.0
  apply_to_critic: false
  expected_target_modules: 400
  verbose: false

logging:
  seed: 0
  randomize_seed: false
  wandb_key: null
  wandb_entity: null
  wandb_project: LongLive-Wan21-14B-CFG-Only-LoRA128-8GPU-Migration