| { |
| "format": "openmed-onnx", |
| "format_version": 1, |
| "formats": [ |
| "onnx-android", |
| "onnx-int8", |
| "ort-android" |
| ], |
| "task": "token-classification", |
| "family": "roberta", |
| "source_model_id": "OpenMed/OpenMed-NER-PathologyDetect-TinyMed-82M", |
| "config_path": "config.json", |
| "label_map_path": "id2label.json", |
| "artifacts": [ |
| { |
| "format": "onnx-android", |
| "path": "model.onnx", |
| "precision": "float32", |
| "metadata": { |
| "profile": "android", |
| "opset": 18, |
| "execution_providers": [ |
| "NNAPI", |
| "XNNPACK" |
| ], |
| "inputs": [ |
| { |
| "name": "input_ids", |
| "dtype": "int64", |
| "axes": [ |
| "batch", |
| "sequence" |
| ], |
| "shape": [ |
| { |
| "kind": "dynamic", |
| "name": "batch" |
| }, |
| { |
| "kind": "dynamic", |
| "name": "sequence" |
| } |
| ] |
| }, |
| { |
| "name": "attention_mask", |
| "dtype": "int64", |
| "axes": [ |
| "batch", |
| "sequence" |
| ], |
| "shape": [ |
| { |
| "kind": "dynamic", |
| "name": "batch" |
| }, |
| { |
| "kind": "dynamic", |
| "name": "sequence" |
| } |
| ] |
| } |
| ], |
| "outputs": [ |
| { |
| "name": "logits", |
| "dtype": "float32", |
| "axes": [ |
| "batch", |
| "sequence", |
| "labels" |
| ], |
| "shape": [ |
| { |
| "kind": "dynamic", |
| "name": "batch" |
| }, |
| { |
| "kind": "dynamic", |
| "name": "sequence" |
| }, |
| { |
| "kind": "static", |
| "value": 3 |
| } |
| ] |
| } |
| ], |
| "operators": [ |
| "Constant", |
| "Equal", |
| "Not", |
| "Cast", |
| "CumSum", |
| "Mul", |
| "Add", |
| "Shape", |
| "Gather", |
| "Unsqueeze", |
| "Concat", |
| "Reshape", |
| "ConstantOfShape", |
| "Where", |
| "Expand", |
| "GatherElements", |
| "LayerNormalization", |
| "Range", |
| "GreaterOrEqual", |
| "And", |
| "Flatten", |
| "MatMul", |
| "Transpose", |
| "Softmax", |
| "IsNaN", |
| "Div", |
| "Erf" |
| ], |
| "unsupported_ops": [ |
| "CumSum", |
| "GreaterOrEqual", |
| "Flatten", |
| "IsNaN" |
| ], |
| "warnings": [ |
| "CumSum is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set", |
| "GreaterOrEqual is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set", |
| "Flatten is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set", |
| "IsNaN is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set" |
| ] |
| } |
| }, |
| { |
| "format": "onnx-android", |
| "path": "model_fp16.onnx", |
| "precision": "float16", |
| "metadata": { |
| "profile": "android", |
| "opset": 18, |
| "execution_providers": [ |
| "NNAPI", |
| "XNNPACK" |
| ], |
| "inputs": [ |
| { |
| "name": "input_ids", |
| "dtype": "int64", |
| "axes": [ |
| "batch", |
| "sequence" |
| ], |
| "shape": [ |
| { |
| "kind": "dynamic", |
| "name": "batch" |
| }, |
| { |
| "kind": "dynamic", |
| "name": "sequence" |
| } |
| ] |
| }, |
| { |
| "name": "attention_mask", |
| "dtype": "int64", |
| "axes": [ |
| "batch", |
| "sequence" |
| ], |
| "shape": [ |
| { |
| "kind": "dynamic", |
| "name": "batch" |
| }, |
| { |
| "kind": "dynamic", |
| "name": "sequence" |
| } |
| ] |
| } |
| ], |
| "outputs": [ |
| { |
| "name": "logits", |
| "dtype": "float32", |
| "axes": [ |
| "batch", |
| "sequence", |
| "labels" |
| ], |
| "shape": [ |
| { |
| "kind": "dynamic", |
| "name": "batch" |
| }, |
| { |
| "kind": "dynamic", |
| "name": "sequence" |
| }, |
| { |
| "kind": "static", |
| "value": 3 |
| } |
| ] |
| } |
| ], |
| "operators": [ |
| "Constant", |
| "Equal", |
| "Not", |
| "Cast", |
| "CumSum", |
| "Mul", |
| "Add", |
| "Shape", |
| "Gather", |
| "Unsqueeze", |
| "Concat", |
| "Reshape", |
| "ConstantOfShape", |
| "Where", |
| "Expand", |
| "GatherElements", |
| "LayerNormalization", |
| "Range", |
| "GreaterOrEqual", |
| "And", |
| "Flatten", |
| "MatMul", |
| "Transpose", |
| "Softmax", |
| "IsNaN", |
| "Div", |
| "Erf" |
| ], |
| "unsupported_ops": [ |
| "CumSum", |
| "GreaterOrEqual", |
| "Flatten", |
| "IsNaN" |
| ], |
| "warnings": [ |
| "CumSum is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set", |
| "GreaterOrEqual is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set", |
| "Flatten is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set", |
| "IsNaN is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set" |
| ] |
| } |
| }, |
| { |
| "format": "onnx-int8", |
| "path": "model_int8.onnx", |
| "precision": "int8", |
| "metadata": { |
| "format": "onnx-int8", |
| "quantization": { |
| "scheme": "dynamic", |
| "weight_type": "qint8", |
| "op_types": [ |
| "MatMul", |
| "Gemm" |
| ], |
| "calibration": "none", |
| "target": "arm64-cpu", |
| "output": "model_int8.onnx" |
| }, |
| "profile": "android", |
| "opset": 18, |
| "execution_providers": [ |
| "NNAPI", |
| "XNNPACK" |
| ], |
| "inputs": [ |
| { |
| "name": "input_ids", |
| "dtype": "int64", |
| "axes": [ |
| "batch", |
| "sequence" |
| ], |
| "shape": [ |
| { |
| "kind": "dynamic", |
| "name": "batch" |
| }, |
| { |
| "kind": "dynamic", |
| "name": "sequence" |
| } |
| ] |
| }, |
| { |
| "name": "attention_mask", |
| "dtype": "int64", |
| "axes": [ |
| "batch", |
| "sequence" |
| ], |
| "shape": [ |
| { |
| "kind": "dynamic", |
| "name": "batch" |
| }, |
| { |
| "kind": "dynamic", |
| "name": "sequence" |
| } |
| ] |
| } |
| ], |
| "outputs": [ |
| { |
| "name": "logits", |
| "dtype": "float32", |
| "axes": [ |
| "batch", |
| "sequence", |
| "labels" |
| ], |
| "shape": [ |
| { |
| "kind": "dynamic", |
| "name": "batch" |
| }, |
| { |
| "kind": "dynamic", |
| "name": "sequence" |
| }, |
| { |
| "kind": "static", |
| "value": 3 |
| } |
| ] |
| } |
| ], |
| "operators": [ |
| "Constant", |
| "Cast", |
| "Shape", |
| "Gather", |
| "Equal", |
| "ConstantOfShape", |
| "Flatten", |
| "Not", |
| "Unsqueeze", |
| "Concat", |
| "CumSum", |
| "Range", |
| "Reshape", |
| "Mul", |
| "Add", |
| "Where", |
| "Expand", |
| "GatherElements", |
| "LayerNormalization", |
| "DynamicQuantizeLinear", |
| "MatMulInteger", |
| "Transpose", |
| "GreaterOrEqual", |
| "MatMul", |
| "And", |
| "Softmax", |
| "IsNaN", |
| "Div", |
| "Erf" |
| ], |
| "unsupported_ops": [ |
| "Flatten", |
| "CumSum", |
| "DynamicQuantizeLinear", |
| "MatMulInteger", |
| "GreaterOrEqual", |
| "IsNaN" |
| ], |
| "warnings": [ |
| "Flatten is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set", |
| "CumSum is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set", |
| "DynamicQuantizeLinear is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set", |
| "MatMulInteger is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set", |
| "GreaterOrEqual is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set", |
| "IsNaN is not in the Android ONNX Runtime Mobile NNAPI/XNNPACK safe set" |
| ] |
| } |
| }, |
| { |
| "format": "ort-android", |
| "path": "model.ort", |
| "precision": "float32", |
| "metadata": { |
| "format": "ort-android", |
| "profile": "android", |
| "source_onnx_path": "model.onnx", |
| "optimization_style": "Fixed", |
| "target_platform": "arm", |
| "op_config_type": "required_operators_and_types", |
| "enable_type_reduction": true, |
| "ort_path": "model.ort", |
| "op_config_path": "model.required_operators_and_types.config", |
| "operators": [ |
| "Constant", |
| "Equal", |
| "Not", |
| "Cast", |
| "CumSum", |
| "Mul", |
| "Add", |
| "Shape", |
| "Gather", |
| "Unsqueeze", |
| "Concat", |
| "Reshape", |
| "ConstantOfShape", |
| "Where", |
| "Expand", |
| "GatherElements", |
| "LayerNormalization", |
| "Range", |
| "GreaterOrEqual", |
| "And", |
| "Flatten", |
| "MatMul", |
| "Transpose", |
| "Softmax", |
| "IsNaN", |
| "Div", |
| "Erf" |
| ], |
| "opset": 18 |
| } |
| } |
| ], |
| "minimum_opset": 18, |
| "max_sequence_length": 512, |
| "dynamic_shapes": { |
| "batch_axis": "dynamic", |
| "sequence_axis": "dynamic", |
| "shape_buckets": { |
| "strategy": "next_configured_bucket_or_exact", |
| "axes": { |
| "input": [ |
| "batch", |
| "sequence" |
| ], |
| "output": [ |
| "batch", |
| "sequence" |
| ] |
| }, |
| "buckets": [ |
| 8, |
| 16, |
| 32, |
| 64, |
| 128, |
| 256, |
| 512, |
| 1024, |
| 2048 |
| ], |
| "min_length": 8, |
| "max_length": 2048, |
| "overflow": "exact_length" |
| } |
| }, |
| "optimization": { |
| "enabled": false, |
| "minimum_opset": 18 |
| }, |
| "validation": {}, |
| "operator_fallbacks": [], |
| "tokenizer": { |
| "path": ".", |
| "files": [ |
| "tokenizer.json", |
| "tokenizer_config.json" |
| ] |
| } |
| } |
