runtime:
  platform: "auto"

data:
  data_dir: "./data"
  static_dir: "./data/static"
  stats_dir: "./data/stats"
  train_years: [2000, 2001]
  test_years: [2003]
  train_stride: 1
  test_stride: 1
  reintroduce_sst_nans: true
  # Source total_precipitation must be accumulated over each source interval,
  # in the same units as the official GenCast statistics.
  precipitation_interval_hours: 6

# Full-scale random initialization. For released weights, inference ignores
# this section and loads every architecture field from the official checkpoint.
model:
  mesh_size: 4  # GenCast Mini (2562 nodes); 6=full (40962 nodes, TPU-scale)
  latent_size: 512
  hidden_layers: 1
  radius_query_fraction_edge_length: 0.6
  attention_k_hop: 16
  attention_type: "triblockdiag_mha"  # GPU/CPU; use "splash_mha" only for TPU
  mask_type: "lazy"
  num_layers: 16
  num_heads: 4
  ffw_hidden: 2048

sampler:
  max_noise_level: 80.0
  min_noise_level: 0.03
  num_noise_levels: 20
  rho: 7.0
  stochastic_churn_rate: 2.5
  churn_min_noise_level: 0.75
  churn_max_noise_level: .inf
  noise_level_inflation_factor: 1.05

training:
  max_steps: 10
  learning_rate: 0.0001
  betas: [0.9, 0.999]
  epsilon: 1.0e-8
  seed: 42
  save_interval: 1000

parallel:
  mode: "pmap"
  num_devices: 1
  global_batch_size: 1
  axis_name: "devices"

inference:
  official_checkpoint: null
  # Officially documented GPU substitution for TPU splash attention.
  attention_type_override: null
  prediction_steps: 30
  num_members: 4
  seed: 42
  # Keep full-resolution ensemble memory bounded by writing each lead/member.
  stream_chunks: true

checkpoint:
  trainer: "./data/checkpoints/model_bak.npz"
  resume: null

output:
  prediction: "./result/prediction.nc"
  plot: "./result/gencast_forecast.png"

fake_data:
  height: 9
  width: 16
  timesteps: 8