generation_config.json

{
  "_from_model_config": true,
  "attn_softmax_bf16": null,
  "bos_token_id": 1,
  "bucket_size": -1,
  "eos_token_id": 2,
  "flash_attention_recompute": false,
  "ignore_eos": null,
  "kv_cache_fp8": null,
  "limit_hpu_graphs": null,
  "reduce_recompile": null,
  "reuse_cache": null,
  "static_shapes": null,
  "transformers_version": "4.37.2",
  "trim_logits": null,
  "use_flash_attention": true
}