model:
  checkpoint_path: "./models/aura_flow_0.3.bnb_nf4.safetensors"
  pretrained_model_name_or_path: fal/AuraFlow-v0.3

  dtype: bfloat16

  denoiser:
    use_flash_attn: true

    use_rope: True
    rope_theta: 10000
    rope_dim_sizes: [32, 112, 112]

  noise_prediction_loss: true
  migration_loss: false
  prior_preservation_loss: false

peft:
  type: lora
  rank: 4
  alpha: 1.0
  dropout: 0.0

  dtype: bfloat16

  include_keys:
    - ".mlp."
    - ".attn."

  exclude_keys:
    - "text_encoder"
    - "vae"
    - "t_embedder"
    - "final_linear"
    - regex: .*\.mod[CX]{1,2} # exclude modulation layers (modC, modCX, modX)

dataset:
  folder: "masked"
  num_repeats: 1
  batch_size: 2

  bucket_base_size: 1024
  step: 128
  min_size: 384
  do_upscale: false

  caption_processors: []

optimizer:
  name: "schedulefree.RAdamScheduleFree"
  args:
    lr: 0.03

tracker:
  project_name: "auraflow-rope-1"
  loggers:
    - wandb

saving:
  strategy:
    per_epochs: 1
    per_steps: 500
    # save_last: true

  callbacks:
    - type: "hf_hub" # or "hf_hub" to push to hub
      name: "rope-20"
      save_dir: "./output/rope-20"

      hub_id: "p1atdev/afv03-lora"
      dir_in_repo: "rope-20"

preview:
  strategy:
    per_epochs: 1
    per_steps: 100

  callbacks:
    - type: "discord"
      url: "masked"

  data:
    path: "./projects/rope/preview.yml"

seed: 42
num_train_epochs: 10

trainer:
  # debug_mode: "1step"

  gradient_checkpointing: true
  gradient_accumulation_steps: 8

  torch_compile: true
  torch_compile_args:
    mode: max-autotune
    fullgraph: true
  fp32_matmul_precision: "medium"