iree-compile doesn't support converting "onnx.Unsqueeze"

[phoebesv]

I obtained model.onnx from MLPerf in the first step and used

1. iree-import-onnx model.onnx -o model.mlir convert success Then I used
2. iree-compile --iree-input-type=onnx model.mlir --compile-to=input I've got error like this

   error: failed to legalize operation 'torch.operator' that was explicitly marked illegal
   %393 = torch.operator "onnx.Unsqueeze"(%arg1) {torch.onnx.axes = [1 : si64]} : (!torch.vtensor<[?,384],si64>) -> !torch.vtensor<[?,1,384],si64>
          ^

   I found torch.operator "onnx.Unsqueeze" doesn't support convert in step 2.

On the tracking page, I noticed that the Unsqueeze operation was marked as completed. https://github.com/nod-ai/SHARK-Turbine/issues/215

What component(s) does this issue relate to?

MLIR, Compiler

Version information

protobuf>=3.20.2 in /usr/local/lib/python3.10/dist-packages (from onnx>=1.15.0->iree-compiler[onnx]) (3.20.3) installed iree-compiler-20240206.793

[stellaraccident]

@rsuderman looks like was marked completed but may need more work

[ScottTodd]

cc @saienduri, https://github.com/llvm/torch-mlir/pull/2601 added a conversion for unsqueeze

[renxida]

Picking this up. This interferes with one of the test cases I'm making for https://github.com/nod-ai/SHARK-TestSuite/

[ScottTodd]

Would you happen to have the model.mlir on hand for this? I just hit a similar error where my program has torch.onnx_meta.ir_version = 7 but the lowering pattern requires IR version 13 or higher. Unsqueeze has versions in 1, 11, 13, and 21: https://github.com/onnx/onnx/blob/main/docs/Operators.md#Unsqueeze.

[ScottTodd]

Here is the program (test case) I'm using:

module {
  func.func @test_loop13_seq(%arg0: !torch.vtensor<[],si64>, %arg1: !torch.vtensor<[],i1>, %arg2: !torch.list<vtensor<[],f32>>) -> !torch.list<vtensor<[],f32>> attributes {torch.onnx_meta.ir_version = 7 : si64, torch.onnx_meta.opset_version = 13 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
    %none = torch.constant.none
    %0 = torch.operator "onnx.Loop"(%arg0, %arg1, %arg2) : (!torch.vtensor<[],si64>, !torch.vtensor<[],i1>, !torch.list<vtensor<[],f32>>) -> !torch.list<vtensor<[],f32>> {
    ^bb0(%arg3: !torch.vtensor<[],si64>, %arg4: !torch.vtensor<[],i1>, %arg5: !torch.list<vtensor<[],f32>>):
      %1 = torch.operator "onnx.Identity"(%arg4) : (!torch.vtensor<[],i1>) -> !torch.vtensor<[],i1> 
      %2 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<[1.000000e+00, 2.000000e+00, 3.000000e+00, 4.000000e+00, 5.000000e+00]> : tensor<5xf32>} : () -> !torch.vtensor<[5],f32> 
      %3 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<1> : tensor<si64>} : () -> !torch.vtensor<[],si64> 
      %4 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<0> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64> 
      %5 = torch.operator "onnx.Add"(%arg3, %3) : (!torch.vtensor<[],si64>, !torch.vtensor<[],si64>) -> !torch.vtensor<[],si64> 
      %6 = torch.operator "onnx.Constant"() {torch.onnx.value = dense<0> : tensor<si64>} : () -> !torch.vtensor<[],si64> 
      %7 = torch.operator "onnx.Unsqueeze"(%5, %6) : (!torch.vtensor<[],si64>, !torch.vtensor<[],si64>) -> !torch.vtensor<[1],si64> 
      %8 = torch.operator "onnx.Slice"(%2, %4, %7) : (!torch.vtensor<[5],f32>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[?],f32> 
      %9 = torch.operator "onnx.SequenceInsert"(%arg5, %8) : (!torch.list<vtensor<[],f32>>, !torch.vtensor<[?],f32>) -> !torch.list<vtensor<[],f32>> 
      torch.operator_terminator %1, %9 : !torch.vtensor<[],i1>, !torch.list<vtensor<[],f32>>
    }
    return %0 : !torch.list<vtensor<[],f32>>
  }
}

Note that Identity has a similar issue - the pattern requires a minimum version of 14, but the op has versions in 21, 19, 16, 14, 13, 1 (https://github.com/onnx/onnx/blob/main/docs/Operators.md#Identity)

[renxida]

@ScottTodd last time i checked with @stellaraccident , it's fine to just edit DefaultDomainQtoZ.cpp and set since to 1

It's been a while since i looked at this but i think i dealt with this by just setting my tiny simple model to have a higher opset, vs the actual solution is DefaultDomainQtoZ.cpp. I later went through and fixed a bunch of them but this one sneaked past.

[phoebesv]

@ScottTodd The Onnx model was obtained from https://zenodo.org/records/3733910, thanks.