models/configs/large.yaml

# LILITH-Large Configuration
# ~400M parameters, high accuracy

model:
  variant: large
  hidden_dim: 384
  num_heads: 12
  ffn_dim: 1536

  # Input/Output
  input_features: 7
  output_features: 3
  sequence_length: 30
  forecast_length: 90

  # Component depths
  gat_layers: 4
  temporal_layers: 8
  sfno_layers: 6

  # Grid configuration
  use_grid: true
  nlat: 128
  nlon: 256

  # Features
  use_climate_embed: true
  use_solar_position: true
  use_flash_attention: true
  use_rope: true

  # Ensemble
  ensemble_method: diffusion
  ensemble_members: 20

  # Regularization
  dropout: 0.1

  # Memory optimization
  gradient_checkpointing: true

training:
  learning_rate: 5e-5
  weight_decay: 0.01
  max_grad_norm: 1.0
  warmup_steps: 2000
  max_steps: 200000

  batch_size: 4
  gradient_accumulation_steps: 8

  use_amp: true
  amp_dtype: bfloat16

  use_deepspeed: true
  deepspeed_stage: 2

  curriculum_enabled: true
  curriculum_stages: [7, 14, 30, 60, 90]
  curriculum_switch_steps: [20000, 60000, 120000, 160000]

inference:
  quantization: static_int8
  batch_size: 8
  max_stations: 500