load_fused_fft_conv2d

load_fused_fft_conv2d()

Import the fused CUDA kernel now instead of on first call.

Importing subquadratic_ops_torch.fused_fft_conv2d is what registers the torch.ops.subquadratic_ops_torch.fused_* operators. The wrappers here import it lazily, which is fine in eager mode — but a torch.compile artifact restored from the on-disk FX cache references those operators without any Python call happening first, and fails with a confusing AttributeError on the op namespace. Callers that compile ahead of the first eager call should invoke this to force registration up front.

Raises:

ImportError – If subquadratic_ops_torch is not installed.

Return type:

None