openvino_config.json

{
  "compression": {
    "algorithm": "quantization",
    "export_to_onnx_standard_ops": false,
    "ignored_scopes": [
      "{re}.*Embedding.*",
      "{re}.*add___.*",
      "{re}.*norm.*"
    ],
    "initializer": {
      "batchnorm_adaptation": {
        "num_bn_adaptation_samples": 0
      },
      "range": {
        "num_init_samples": 16,
        "type": "min_max"
      }
    },
    "overflow_fix": "disable",
    "preset": "performance"
  },
  "input_info": [
    {
      "keyword": "input_ids",
      "sample_size": [
        1,
        32
      ],
      "type": "long"
    },
    {
      "keyword": "attention_mask",
      "sample_size": [
        1,
        32
      ],
      "type": "long"
    }
  ],
  "log_dir": "./logs/llama-2-13b-w8a8-unstructured50/",
  "optimum_version": "1.13.2",
  "save_onnx_model": false,
  "transformers_version": "4.34.0"
}