Reference for ultralytics/utils/export/paddle.py#

def torch2paddle(
    model: torch.nn.Module,
    im: torch.Tensor,
    output_dir: Path | str,
    metadata: dict | None = None,
    prefix: str = "",
) -> str:
    """Export a PyTorch model to PaddlePaddle format using X2Paddle.

    Args:
        model (torch.nn.Module): The PyTorch model to export.
        im (torch.Tensor): Example input tensor for tracing.
        output_dir (Path | str): Directory to save the exported PaddlePaddle model.
        metadata (dict | None): Optional metadata saved as ``metadata.yaml``.
        prefix (str): Prefix for log messages.

    Returns:
        (str): Path to the exported ``_paddle_model`` directory.
    """
    assert not IS_JETSON, "Jetson Paddle exports not supported yet"
    from ultralytics.utils.checks import check_requirements

    check_requirements(
        (
            # Interchangeable candidates so an installed variant is never dual-installed over (both ship 'paddle')
            (
                "paddlepaddle-gpu>=3.0.0,<3.3.0"  # pin <3.3.0 https://github.com/PaddlePaddle/Paddle/issues/77340
                if im.device.type == "cuda"
                else "paddlepaddle==3.0.0"  # pin 3.0.0 for ARM64
                if ARM64
                else "paddlepaddle>=3.0.0,<3.3.0",  # pin <3.3.0 https://github.com/PaddlePaddle/Paddle/issues/77340
                "paddlepaddle==3.0.0" if ARM64 else "paddlepaddle>=3.0.0,<3.3.0",
                "paddlepaddle-gpu>=3.0.0,<3.3.0",
            ),
            "x2paddle",
        )
    )

    import x2paddle
    from x2paddle.convert import pytorch2paddle
    from x2paddle.op_mapper.pytorch2paddle import prim2code

    # x2paddle 1.6.0 codegen for pooling asserts reads exec() results from locals(), which PEP 667 broke in Python
    # 3.13. The assert only re-checks pooling args that torch already validated during tracing, so skip emitting it.
    prim_assert = prim2code.prim_assert
    prim2code.prim_assert = lambda layer, **kwargs: None

    LOGGER.info(f"\n{prefix} starting export with X2Paddle {x2paddle.__version__}...")

    try:
        pytorch2paddle(module=model, save_dir=output_dir, jit_type="trace", input_examples=[im])  # export
    finally:
        prim2code.prim_assert = prim_assert
    if metadata:
        YAML.save(Path(output_dir) / "metadata.yaml", metadata)  # add metadata.yaml
    return str(output_dir)