vocoder_onnx_matmul4_report.json

{
  "input": "higgs_audio_v3_vocoder_decode.onnx",
  "output": "higgs_audio_v3_vocoder_decode_matmul4.onnx",
  "bits": 4,
  "block_size": 128,
  "quantizer": "onnxruntime.quantization.matmul_nbits_quantizer.MatMulNBitsQuantizer",
  "op_types_requested": [
    "MatMul",
    "Gather"
  ],
  "op_counts_before": {
    "Shape": 2,
    "Squeeze": 9,
    "Transpose": 11,
    "Slice": 8,
    "Gather": 8,
    "MatMul": 9,
    "Add": 67,
    "Conv": 32,
    "Concat": 11,
    "Reshape": 76,
    "Mul": 77,
    "Sin": 36,
    "Pow": 36,
    "ConvTranspose": 5
  },
  "op_counts_after": {
    "Shape": 2,
    "Transpose": 11,
    "Concat": 11,
    "Squeeze": 9,
    "Slice": 8,
    "Mul": 77,
    "Reshape": 76,
    "GatherBlockQuantized": 8,
    "MatMulNBits": 9,
    "Add": 67,
    "Conv": 32,
    "Sin": 36,
    "Pow": 36,
    "ConvTranspose": 5
  },
  "input_bytes": 87197263,
  "output_bytes": 82877763,
  "test_frames": 64,
  "ort_check": "pass",
  "max_abs_diff_vs_fp_onnx": 0.1914266049861908,
  "mean_abs_diff_vs_fp_onnx": 0.006282582879066467,
  "mean_rel_diff_vs_fp_onnx": 1.8227379322052002
}