config.json

{
  "input_patch_len": 32,
  "input_transform": "identity",
  "linear_detrending_threshold": 0.5,
  "output_patch_len": 64,
  "quantiles": [
    0.1,
    0.2,
    0.3,
    0.4,
    0.5,
    0.6,
    0.7,
    0.8,
    0.9
  ],
  "residual_block_config": {
    "activation": "relu",
    "dropout": 0.0,
    "hidden_dims": 1280,
    "identity_skip": false,
    "output_dims": 1280,
    "prenorm": "none",
    "use_bias": false
  },
  "transformer_config": {
    "num_layers": 20,
    "transformer": {
      "attention_norm": "rms",
      "causal_attention": true,
      "debug_no_masking": false,
      "deterministic": true,
      "feedforward_norm": "rms",
      "ff_activation": "relu",
      "hidden_dims": 1280,
      "max_variates": 32,
      "model_dims": 1280,
      "num_heads": 16,
      "paired_token_skip_second": false,
      "qk_norm": "rms",
      "training": true,
      "use_bias": false,
      "use_memory_efficient_attention": true,
      "use_rope_seq": true,
      "use_rope_var": false,
      "use_sdpa": true,
      "v_norm": "none"
    },
    "use_remat": true
  },
  "use_frozen_running_stats": false,
  "use_iterative_cpm_revin": true,
  "use_linear_detrending": true,
  "use_stitching": true,
  "use_variate_attention": true,
  "value_clip": 1e+20
}