Function _compiled_unsupported

monai/transforms/spatial/functional.py:57–95

Return True if ``monai._C`` (the compiled C extension providing ``grid_pull``) is not compiled with support for the given CUDA device's compute capability. Args: device: The torch device to check for compiled extension support. Returns: True if the device is CUDA w

(device: torch.device)

def _compiled_unsupported(device: torch.device) -> bool:
    """
    Return True if ``monai._C`` (the compiled C extension providing ``grid_pull``) is not
    compiled with support for the given CUDA device's compute capability.

    Args:
        device: The torch device to check for compiled extension support.

    Returns:
        True if the device is CUDA with compute capability major >= 12 (Blackwell+),
        False otherwise. Always returns False for CPU devices.

    Note:
        ``monai._C`` is built at install time against a fixed set of CUDA architectures.
        NVIDIA Blackwell GPUs (sm_120, compute capability 12.x) and newer were not included in
        the default ``TORCH_CUDA_ARCH_LIST`` when the MONAI slim image was originally built,
        so executing ``grid_pull`` on those devices produces incorrect results.  Falling back to
        the PyTorch-native ``affine_grid`` + ``grid_sample`` path (``USE_COMPILED=False``) gives
        correct output on all architectures.

        The threshold (``major >= 12``) matches the first architecture family (Blackwell, sm_120)
        that shipped after the highest sm supported in the current default build list (sm_90,
        Hopper).  Adjust this constant when ``monai._C`` is rebuilt with sm_120+ support.
    """
    if device.type != "cuda":
        return False
    try:
        from monai._C import max_compute_capability as _max_cc_func

        max_cc = _max_cc_func()
        if max_cc == 0:
            # No architecture info embedded (older build), fall back to heuristic
            return bool(torch.cuda.get_device_properties(device).major >= 12)
        device_cc = (
            torch.cuda.get_device_properties(device).major * 100 + torch.cuda.get_device_properties(device).minor
        )
        return bool(device_cc > max_cc)
    except (ImportError, AttributeError):
        return bool(torch.cuda.get_device_properties(device).major >= 12)


def _maybe_new_metatensor(img, dtype=None, device=None):