generation_config.json

{
  "bos_token_id": 128000,
  "do_sample": true,
  "eos_token_id": 128001,
  "exit_attention": false,
  "exit_decoder_layer": true,
  "exit_layer_indices": [
    3,
    6,
    12,
    17
  ],
  "exit_mlp": false,
  "max_length": 4096,
  "output_exit_layers": [
    3,
    6,
    12,
    17
  ],
  "output_full_model": true,
  "temperature": 0.6,
  "tie_exit_lm_head": true,
  "top_p": 0.9,
  "transformers_version": "4.48.0"
}