generation_config.json

{
  "attn_implementation": "sdpa",
  "bos_token_id": 128000,
  "depth_decoder_do_sample": true,
  "depth_decoder_max_new_tokens": 31,
  "depth_decoder_min_new_tokens": 31,
  "depth_decoder_temperature": 0.9,
  "do_sample": true,
  "max_length": 2048,
  "max_new_tokens": 125,
  "pad_token_id": 128256,
  "temperature": 0.9,
  "transformers_version": "4.52.3"
}