| """Quantization settings for the published decoder; this script never retrains it.""" | |
| import argparse | |
| import hashlib | |
| from pathlib import Path | |
| from onnxruntime.quantization import QuantType, quantize_dynamic | |
| def main(): | |
| parser = argparse.ArgumentParser(description=__doc__) | |
| parser.add_argument("input", type=Path, help="Original FP32 decoder_model.onnx") | |
| parser.add_argument("output", type=Path, help="New INT8 decoder output path") | |
| args = parser.parse_args() | |
| expected = "6b1fb216d542c4b2a4fa5b9d7ae3522081eb85fb959d2cecd28055af956a8a5e" | |
| actual = hashlib.sha256(args.input.read_bytes()).hexdigest() | |
| if actual != expected: | |
| parser.error(f"Input SHA-256 mismatch: {actual}") | |
| if args.output.exists(): | |
| parser.error("Output already exists; choose a new path") | |
| quantize_dynamic( | |
| str(args.input), str(args.output), | |
| op_types_to_quantize=["MatMul"], | |
| weight_type=QuantType.QInt8, | |
| per_channel=True, | |
| ) | |
| print(hashlib.sha256(args.output.read_bytes()).hexdigest(), args.output.name) | |
| if __name__ == "__main__": | |
| main() | |