{
  "prefill": {
    "input": "higgs_audio_v3_ar_prefill.onnx",
    "output": "higgs_audio_v3_ar_prefill_matmul4.onnx",
    "bits": 4,
    "block_size": 128,
    "quantizer": "onnxruntime.quantization.matmul_nbits_quantizer.MatMulNBitsQuantizer",
    "op_types_requested": [
      "MatMul",
      "Gather"
    ],
    "op_counts_before": {
      "Shape": 38,
      "Squeeze": 1,
      "Gather": 2,
      "Range": 1,
      "Cast": 294,
      "Unsqueeze": 79,
      "LessOrEqual": 1,
      "And": 2,
      "Max": 1,
      "Expand": 73,
      "Concat": 150,
      "GatherND": 1,
      "MatMul": 325,
      "Transpose": 181,
      "Cos": 1,
      "Sin": 1,
      "Pow": 145,
      "ReduceMean": 145,
      "Add": 325,
      "Sqrt": 145,
      "Reciprocal": 145,
      "Mul": 578,
      "Reshape": 289,
      "Slice": 252,
      "Neg": 72,
      "Where": 37,
      "Softmax": 36,
      "IsNaN": 36,
      "Sigmoid": 36,
      "Gemm": 1
    },
    "op_counts_after": {
      "Cast": 294,
      "Shape": 38,
      "GatherBlockQuantized": 1,
      "Squeeze": 1,
      "Concat": 150,
      "Range": 1,
      "Pow": 145,
      "Unsqueeze": 79,
      "ReduceMean": 145,
      "Add": 325,
      "LessOrEqual": 1,
      "Max": 1,
      "MatMul": 73,
      "Sqrt": 145,
      "And": 2,
      "Transpose": 181,
      "Reciprocal": 145,
      "Expand": 73,
      "Mul": 578,
      "Cos": 1,
      "Sin": 1,
      "GatherND": 1,
      "MatMulNBits": 252,
      "Reshape": 289,
      "Where": 37,
      "Slice": 252,
      "Neg": 72,
      "Softmax": 36,
      "IsNaN": 36,
      "Sigmoid": 36,
      "Gather": 1,
      "Gemm": 1
    },
    "input_bytes": 8094757353,
    "output_bytes": 2123593945,
    "onnx_checker": "pass"
  },
  "decode": {
    "input": "higgs_audio_v3_ar_decode.onnx",
    "output": "higgs_audio_v3_ar_decode_matmul4.onnx",
    "bits": 4,
    "block_size": 128,
    "quantizer": "onnxruntime.quantization.matmul_nbits_quantizer.MatMulNBitsQuantizer",
    "op_types_requested": [
      "MatMul",
      "Gather"
    ],
    "op_counts_before": {
      "Shape": 37,
      "Squeeze": 1,
      "Add": 328,
      "Gather": 2,
      "Cast": 295,
      "ReduceSum": 1,
      "Unsqueeze": 80,
      "Range": 1,
      "LessOrEqual": 1,
      "Reshape": 290,
      "Concat": 220,
      "Expand": 73,
      "MatMul": 325,
      "Transpose": 181,
      "Cos": 1,
      "Sin": 1,
      "Pow": 145,
      "ReduceMean": 145,
      "Sqrt": 145,
      "Reciprocal": 145,
      "Mul": 578,
      "Slice": 252,
      "Neg": 72,
      "Where": 37,
      "Softmax": 36,
      "IsNaN": 36,
      "Sigmoid": 36,
      "Gemm": 1
    },
    "op_counts_after": {
      "Add": 328,
      "Shape": 37,
      "Unsqueeze": 80,
      "GatherBlockQuantized": 1,
      "Squeeze": 1,
      "Cast": 295,
      "MatMul": 73,
      "ReduceSum": 1,
      "Range": 1,
      "Reshape": 290,
      "Transpose": 181,
      "Concat": 220,
      "Cos": 1,
      "Sin": 1,
      "LessOrEqual": 1,
      "Pow": 145,
      "Expand": 73,
      "ReduceMean": 145,
      "Where": 37,
      "Sqrt": 145,
      "Reciprocal": 145,
      "Mul": 578,
      "MatMulNBits": 252,
      "Slice": 252,
      "Neg": 72,
      "Softmax": 36,
      "IsNaN": 36,
      "Sigmoid": 36,
      "Gather": 1,
      "Gemm": 1
    },
    "input_bytes": 7317033532,
    "output_bytes": 1934059497,
    "onnx_checker": "pass"
  },
  "ort_cpu_quantized_validation": {
    "session_create": "pass",
    "prefill_run": "pass",
    "prefill_logits_shape": [
      1,
      8,
      1026
    ],
    "decode_run": "pass",
    "decode_logits_shape": [
      1,
      8,
      1026
    ]
  }
}