generation_config.json

{
  "bos_token_id": 128000,
  "depth_decoder__from_model_config": false,
  "depth_decoder_do_sample": true,
  "depth_decoder_output_attentions": false,
  "depth_decoder_output_hidden_states": false,
  "depth_decoder_temperature": 0.9,
  "depth_decoder_top_k": 50,
  "depth_decoder_use_cache": true,
  "do_sample": true,
  "max_new_tokens": 125,
  "pad_token_id": 128002,
  "temperature": 0.9,
  "transformers_version": "5.0.0"
}