model_config.json

{
  "architecture": "recurrent_transformer",
  "bias": false,
  "block_size": 1024,
  "dropout": 0.0,
  "mlp_type": "swiglu",
  "n_embd": 768,
  "n_head": 12,
  "n_kv_head": 4,
  "n_layer": 16,
  "norm_type": "rmsnorm",
  "physical_layers": 16,
  "position_encoding": "rope",
  "recurrent_steps": 2,
  "tie_embeddings": true,
  "unrolled_depth": 32,
  "vocab_size": 4096,
  "zero_init_residual": true
}