convert-torch-onnx-to-torch generates invalid IR for onnx.Resize where scaling is in the first two dimensions

The code in that pass seems to silently assume the the first two dimensions are not scaled, but ONNX has no such restriction.

With input

func.func @test_resize_middle(%arg0: !torch.vtensor<[1,36,42,384],f32>) -> !torch.vtensor<[1,72,84,384],f32> 
  attributes {torch.onnx_meta.ir_version = 7 : si64, torch.onnx_meta.opset_version = 19 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
 %none = torch.constant.none
 %12 = torch.vtensor.literal(dense<[1.000000e+00, 2.000000e+00, 2.000000e+00, 1.000000e+00]> : tensor<4xf32>) : !torch.vtensor<[4],f32>
 %19 = torch.operator "onnx.Resize"(%arg0, %none, %12) {
  torch.onnx.coordinate_transformation_mode = "half_pixel",
  torch.onnx.mode = "nearest",
  torch.onnx.nearest_mode = "round_prefer_floor"} : (!torch.vtensor<[1,36,42,384],f32>, !torch.none, !torch.vtensor<[4],f32>) -> !torch.vtensor<[1,72,84,384],f32>
 return %19 : !torch.vtensor<[1,72,84,384],f32>
}

we get

module {
  func.func @test_resize_middle(%arg0: !torch.vtensor<[1,36,42,384],f32>) -> !torch.vtensor<[1,72,84,384],f32> attributes {torch.onnx_meta.ir_version = 7 : si64, torch.onnx_meta.opset_version = 19 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
    %none = torch.constant.none
    %0 = torch.vtensor.literal(dense<[1.000000e+00, 2.000000e+00, 2.000000e+00, 1.000000e+00]> : tensor<4xf32>) : !torch.vtensor<[4],f32>
    %none_0 = torch.constant.none
    %int0 = torch.constant.int 0
    %false = torch.constant.bool false
    %true = torch.constant.bool true
    %str = torch.constant.str "nearest_half_pixel,round_prefer_floor"
    %int2 = torch.constant.int 2
    %1 = torch.aten.select.int %0, %int0, %int2 : !torch.vtensor<[4],f32>, !torch.int, !torch.int -> !torch.vtensor<[1],f32>
    %2 = torch.aten.item %1 : !torch.vtensor<[1],f32> -> !torch.float
    %int3 = torch.constant.int 3
    %3 = torch.aten.select.int %0, %int0, %int3 : !torch.vtensor<[4],f32>, !torch.int, !torch.int -> !torch.vtensor<[1],f32>
    %4 = torch.aten.item %3 : !torch.vtensor<[1],f32> -> !torch.float
    %5 = torch.prim.ListConstruct %2, %4 : (!torch.float, !torch.float) -> !torch.list<float>
    %6 = torch.aten.__interpolate.size_list_scale_list %arg0, %none_0, %5, %str, %false, %none_0, %false : !torch.vtensor<[1,36,42,384],f32>, !torch.none, !torch.list<float>, !torch.str, !torch.bool, !torch.none, !torch.bool -> !torch.vtensor<[1,72,84,384],f32>
    return %6 : !torch.vtensor<[1,72,84,384],f32>
  }
}

Here, %1 and %3 only read the last two elements out of %0. When lowering this IR to linalg (-convert-torch-to-linalg), we get error: unexpected error: 'tensor.cast' op operand type 'tensor<1x36x?x?xf32>' and result type 'tensor<1x72x84x384xf32>' are cast incompatible because the scales used by torch.aten.__interpolate.size_list_scale_list are not matching the output shape anymore.

llvm / torch-mlir

convert-torch-onnx-to-torch generates invalid IR for onnx.Resize where scaling is in the first two dimensions #3453