ort_config.json

{
  "one_external_file": true,
  "opset": null,
  "optimization": {
    "disable_attention": false,
    "disable_attention_fusion": false,
    "disable_bias_gelu": false,
    "disable_bias_gelu_fusion": false,
    "disable_bias_skip_layer_norm": false,
    "disable_bias_skip_layer_norm_fusion": false,
    "disable_embed_layer_norm": false,
    "disable_embed_layer_norm_fusion": false,
    "disable_gelu": false,
    "disable_gelu_fusion": false,
    "disable_group_norm_fusion": false,
    "disable_layer_norm": false,
    "disable_layer_norm_fusion": false,
    "disable_packed_kv": false,
    "disable_rotary_embeddings": false,
    "disable_shape_inference": false,
    "disable_skip_layer_norm": false,
    "disable_skip_layer_norm_fusion": false,
    "enable_gelu_approximation": true,
    "enable_gemm_fast_gelu_fusion": false,
    "enable_transformers_specific_optimizations": true,
    "fp16": true,
    "no_attention_mask": false,
    "optimization_level": 99,
    "optimize_for_gpu": true,
    "optimize_with_onnxruntime_only": false,
    "use_mask_index": false,
    "use_multi_head_attention": true,
    "use_raw_attention_mask": false
  },
  "quantization": {},
  "use_external_data_format": true
}