openvino_config.json

{
  "compression": {
    "algorithm": "NoCompressionAlgorithm"
  },
  "input_info": [
    {
      "keyword": "input_ids",
      "sample_size": [
        8,
        1024
      ],
      "type": "long"
    },
    {
      "keyword": "attention_mask",
      "sample_size": [
        8,
        1024
      ],
      "type": "long"
    }
  ],
  "log_dir": "/data1/vchua/run/hf-model/gpt2-fp32-ov-kv-cache",
  "optimum_version": "1.8.8",
  "save_onnx_model": false,
  "transformers_version": "4.30.2"
}