generation_config.json

{
  "bos_token_id": 1,
  "eos_token_id": 2,
  "max_length": 2048,
  "max_memory": {
    "mps": 5850611712
  },
  "no_split_module_classes": [
    "LlamaDecoderLayer"
  ],
  "pad_token_id": 0,
  "special_dtypes": {},
  "transformers_version": "4.31.0"
}