# LILITH-Tiny Configuration
# ~50M parameters, suitable for RTX 3060 inference

model:
  variant: tiny
  hidden_dim: 128
  num_heads: 4
  ffn_dim: 512

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

  # Component depths
  gat_layers: 2
  temporal_layers: 4
  sfno_layers: 2

  # Grid configuration
  use_grid: true
  nlat: 32
  nlon: 64

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

  # Ensemble
  ensemble_method: gaussian
  ensemble_members: 10

  # Regularization
  dropout: 0.1

  # Memory optimization
  gradient_checkpointing: false

training:
  learning_rate: 2e-4
  weight_decay: 0.01
  max_grad_norm: 1.0
  warmup_steps: 500
  max_steps: 50000

  batch_size: 16
  gradient_accumulation_steps: 2

  use_amp: true
  amp_dtype: float16

  curriculum_enabled: true
  curriculum_stages: [7, 14, 30, 60, 90]
  curriculum_switch_steps: [5000, 15000, 30000, 40000]

inference:
  quantization: dynamic_int8
  batch_size: 32
  max_stations: 100