Torch-mlir support unknown rank tensor?


I am trying to convert torch script model to torch-mlir for FCN and DeepLabV3. those 2 models are for semantic segmentation and can take arbitrary image size as input.

I got error like below complaining unknown rank during torch-lower-to-backend-contract .
unsupported by backend contract: tensor with unknown rank
note: see current operation: %251 = “torch.tensor_static_info_cast”(%250) : (!torch.vtensor<[1,21,28,28],f32>) → !torch.vtensor<*,f32>

by taking look at the description for LowerToBackendContract pass, it explains the backend contract requires tensor have at least a known rank.
// For value-semantic tensors, we require at least a known rank and dtype.
// We are not aware of a situation where our backends can handle an unranked
// tensor type or a tensor with a dynamic dtype.

does it mean currently torch-mlir doesnt support model like FCN and DeepLabV3 which tensor with arbitrary shape? any suggestions to get around this?