empty_memory_format_test.mlir

func.func private @forward(%arg0: !torch.vtensor<[20,100,35,45],f32>) -> !torch.vtensor<[20,100,35,45],f32> {
    %int0 = torch.constant.int 0
    %0 = torch.prim.ListConstruct %int0 : (!torch.int) -> !torch.list<int>
    %int0_0 = torch.constant.int 0
    %int0_1 = torch.constant.int 0
    %cpu = torch.constant.device "cpu"
    %none = torch.constant.none
    %none_2 = torch.constant.none
    %1 = torch.aten.aten.empty.memory_format %0, %int0_0, %int0_1, %cpu, %none, %none_2 : !torch.list<int>, !torch.int, !torch.int, !torch.Device, !torch.none, !torch.none -> !torch.vtensor<[0],ui8>
    return %arg0 : !torch.vtensor<[20,100,35,45],f32>
}

This issue is found from nod-ai/SHARK-ModelDev#110.
torch.empty((0))
torch-mlir-opt empty-memory_format_test.mlir -convert-torch-to-linalg --debug

// *** IR Dump After Pattern Application ***
mlir-asm-printer: Verifying operation: func.func
ImplicitTypeIDRegistry::lookupOrInsert(mlir::OpTrait::OneTypedResult<mlir::TensorType>::Impl<Empty>)
'tensor.cast' op operand type 'tensor<?xui8>' and result type 'tensor<0xi8>' are cast incompatible
mlir-asm-printer: 'func.func' failed to verify and will be printed in generic form
"func.func"() <{function_type = (!torch.vtensor<[20,100,35,45],f32>) -> !torch.vtensor<[20,100,35,45],f32>, sym_name = "forward", sym_visibility = "private"}> ({
^bb0(%arg0: !torch.vtensor<[20,100,35,45],f32>):
  %0 = "torch.constant.int"() <{value = 0 : i64}> : () -> !torch.int
  %1 = "torch.prim.ListConstruct"(%0) : (!torch.int) -> !torch.list<int>
  %2 = "torch.constant.int"() <{value = 0 : i64}> : () -> !torch.int
  %3 = "builtin.unrealized_conversion_cast"(%2) : (!torch.int) -> i64
  %4 = "torch.constant.int"() <{value = 0 : i64}> : () -> !torch.int
  %5 = "builtin.unrealized_conversion_cast"(%4) : (!torch.int) -> i64
  %6 = "torch.constant.device"() <{value = "cpu"}> : () -> !torch.Device
  %7 = "torch.constant.none"() : () -> !torch.none
  %8 = "torch.constant.none"() : () -> !torch.none
  %9 = "torch_c.to_i64"(%0) : (!torch.int) -> i64
  %10 = "arith.index_cast"(%9) : (i64) -> index
  %11 = "tensor.empty"(%10) : (index) -> tensor<?xui8>
  %12 = "tensor.cast"(%11) : (tensor<?xui8>) -> tensor<0xi8>
  %13 = "torch.aten.empty.memory_format"(%1, %2, %4, %6, %7, %8) : (!torch.list<int>, !torch.int, !torch.int, !torch.Device, !torch.none, !torch.none) -> !torch.vtensor<[0],ui8>
  "func.return"(%arg0) : (!torch.vtensor<[20,100,35,45],f32>) -> ()
}) : () -> ()

The bug is because of the tensor.cast ui8 to i8 is not supported, which is probably because the mis-use of tensor.cast input/result element types.
In the definition of tensor.cast op:

Convert a tensor from one type to an equivalent type without changing any
data elements. The source and destination types must both be tensor types
with the same element type.

As we can see in the input mlir file, the result type should be ui8. As we look at the code,
The wrong i8 issue arises from the typeConverter, it wrongly converts the ui8 to i8.

    op.getType().dump();// !torch.vtensor<[0],ui8>
    auto resultType =
        typeConverter->convertType(op.getType()).cast<RankedTensorType>();
    resultType.dump();// tensor<0xi8>

The code link is here https://github.com/llvm/torch-mlir/blob/d50d3aa5e77117fbb7078c25831ea2913a1c5566/lib/Conversion/TorchToLinalg/TensorConstructors.cpp#L226C15-L226C15