[compile][cpu]: failed to legalize operation 'torch.operator' that was explicitly marked illegal(onnx.Gather)

[pdhirajkumarprasad]

What happened?

For the given IR

module {
  func.func @"torch-jit-export"(%arg0: !torch.vtensor<[?,?],si64>, %arg1:!torch.vtensor<[28996,768],f16>) -> !torch.vtensor<[?,?,768],f16> attributes {torch.onnx_meta.ir_version = 7 : si64, torch.onnx_meta.opset_version = 12 : si64, torch.onnx_meta.producer_name = "onnxruntime.transformers", torch.onnx_meta.producer_version = "1.13.0"} {
    %362 = torch.operator "onnx.Gather"(%arg1, %arg0) : (!torch.vtensor<[28996,768],f16>, !torch.vtensor<[?,?],si64>) -> !torch.vtensor<[?,?,768],f16> 
    return %362 : !torch.vtensor<[?,?,768],f16>
  }
}

getting error as

model.torch_onnx.mlir:3:12: error: failed to legalize operation 'torch.operator' that was explicitly marked illegal
    %362 = torch.operator "onnx.Gather"(%arg1, %arg0) : (!torch.vtensor<[28996,768],f16>, !torch.vtensor<[?,?],si64>) -> !torch.vtensor<[?,?,768],f16> 
           ^
model.torch_onnx.mlir:3:12: note: see current operation: %0 = "torch.operator"(%arg1, %arg0) <{name = "onnx.Gather"}> : (!torch.vtensor<[28996,768],f16>, !torch.vtensor<[?,?],si64>) -> !torch.vtensor<[?,?,768],f16>
// -----// IR Dump After ConvertTorchOnnxToTorch Failed (convert-torch-onnx-to-torch) //----- //
func.func @"torch-jit-export"(%arg0: !torch.vtensor<[?,?],si64>, %arg1: !torch.vtensor<[28996,768],f16>) -> !torch.vtensor<[?,?,768],f16> attributes {torch.onnx_meta.ir_version = 7 : si64, torch.onnx_meta.opset_version = 12 : si64, torch.onnx_meta.producer_name = "onnxruntime.transformers", torch.onnx_meta.producer_version = "1.13.0"} {
  %0 = torch.operator "onnx.Gather"(%arg1, %arg0) : (!torch.vtensor<[28996,768],f16>, !torch.vtensor<[?,?],si64>) -> !torch.vtensor<[?,?,768],f16> 
  return %0 : !torch.vtensor<[?,?,768],f16>
}

// -----// IR Dump After AutoInputConversionPipelinePass Failed (iree-auto-input-conversion) //----- //
module {
  func.func @"torch-jit-export"(%arg0: !torch.vtensor<[?,?],si64>, %arg1: !torch.vtensor<[28996,768],f16>) -> !torch.vtensor<[?,?,768],f16> attributes {torch.onnx_meta.ir_version = 7 : si64, torch.onnx_meta.opset_version = 12 : si64, torch.onnx_meta.producer_name = "onnxruntime.transformers", torch.onnx_meta.producer_version = "1.13.0"} {
    %0 = torch.operator "onnx.Gather"(%arg1, %arg0) : (!torch.vtensor<[28996,768],f16>, !torch.vtensor<[?,?],si64>) -> !torch.vtensor<[?,?,768],f16> 
    return %0 : !torch.vtensor<[?,?,768],f16>
  }
}

Steps to reproduce your issue

Step to reproduce:

iree-compile model.torch_onnx.mlir --iree-hal-target-backends=llvm-cpu --iree-input-demote-i64-to-i32 --mlir-print-ir-after-all

IREE version:
IREE compiler version 20240819.990 @ aeda149
LLVM version 20.0.0git

What component(s) does this issue relate to?

Compiler

Version information

No response

Additional context

No response

[nirvedhmeshram]

I think you might want to file this in torch-mlir as this is failing in convert-torch-onnx-to-torch

[pdhirajkumarprasad]

closing this as open in torch-mlir nod-ai/SHARK-ModelDev#853