"onnx.Resize" failed to legalize operation 'torch.operator' that was explicitly marked illegal

[afalkenberg1]

pytorch-3dunet_vaiq_int8
fails at iree-compile

pytorch-3dunet_vaiq_int8.default.onnx.torch.mlir:2346:13: error: failed to legalize operation 'torch.operator' that was explicitly marked illegal
%1777 = torch.operator "onnx.Resize"(%1716, %none, %none, %1776) {torch.onnx.coordinate_transformation_mode = "asymmetric", torch.onnx.cubic_coeff_a = -7.500000e-01 : f32, torch.onnx.mode = "nearest", torch.onnx.nearest_mode = "floor"} : (!torch.vtensor<[?,256,?,?,?],f32>, !torch.none, !torch.none, !torch.vtensor<[5],si64>) -> !torch.vtensor<[?,?,?,?,?],f32>
^
pytorch-3dunet_vaiq_int8.default.onnx.torch.mlir:2346:13: note: see current operation: %864 = "torch.operator"(%859, %76, %76, %863) <{name = "onnx.Resize"}> {torch.onnx.coordinate_transformation_mode = "asymmetric", torch.onnx.cubic_coeff_a = -7.500000e-01 : f32, torch.onnx.mode = "nearest", torch.onnx.nearest_mode = "floor"} : (!torch.vtensor<[?,256,?,?,?],f32>, !torch.none, !torch.none, !torch.vtensor<[5],si64>) -> !torch.vtensor<[?,?,?,?,?],f32>

[pdhirajkumarprasad]

we have 11 models failing in CI.

crossvit_15_240
crossvit_15_dagger_240
crossvit_15_dagger_408
crossvit_18_240
crossvit_18_dagger_240
crossvit_18_dagger_408
crossvit_9_240
crossvit_9_dagger_240
crossvit_base_240
crossvit_small_240
crossvit_tiny_240

Simple IR to reproduce the issue:

module {
  func.func @torch_jit(%arg0: !torch.vtensor<[1,3,240,240],f32>) -> !torch.vtensor<[1,3,224,224],f32>  attributes {torch.onnx_meta.ir_version = 7 : si64, torch.onnx_meta.opset_version = 21 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "1.12.1"} {
    %none = torch.constant.none
    %1 = torch.operator "onnx.Shape"(%arg0) : (!torch.vtensor<[1,3,240,240],f32>) -> !torch.vtensor<[4],si64> 
    %2 = torch.operator "onnx.Resize"(%arg0, %none, %none, %1) {torch.onnx.coordinate_transformation_mode = "pytorch_half_pixel", torch.onnx.cubic_coeff_a = -7.500000e-01 : f32, torch.onnx.mode = "cubic", torch.onnx.nearest_mode = "floor"} : (!torch.vtensor<[1,3,240,240],f32>, !torch.none, !torch.none, !torch.vtensor<[4],si64>) -> !torch.vtensor<[1,3,224,224],f32> 
    return %2 : !torch.vtensor<[1,3,224,224],f32>
  }
}

[ScottTodd]

Many resize unit tests have been failing, so full models using those ops have no hope of working yet:

• Failing tests: https://github.com/iree-org/iree/blob/fa752ae1e491a1f8fde8967bf04473c6a6c1ca18/tests/external/iree-test-suites/onnx_ops/onnx_ops_cpu_llvm_sync.json#L302-L325
• Sample logs: https://github.com/iree-org/iree-test-suites/actions/runs/11520424883/job/32071702693#step:6:13080
• Sample test case: https://github.com/iree-org/iree-test-suites/blob/main/onnx_ops/onnx/node/generated/test_resize_downsample_scales_cubic/model.mlir

[pdhirajkumarprasad]

@ScottTodd, most of these models were working but recent change in front-end, broke some of these and we are looking into that