torch.aten._scaled_dot_product_efficient_attention：Failed to legalize operation 'torch.operator' that was explicitly marked illegal

[bilu-1]

The convert cmd：

iree-opt --torch-to-iree --mlir-print-ir-after-all {torch_ir} -o {linalg_ir}

Torch IR：

module {
  func.func @main(%arg0: !torch.vtensor<[1,32,7,128],f32>, %arg1: !torch.vtensor<[1,32,7,128],f32>, %arg2: !torch.vtensor<[1,32,7,128],f32>, %arg3: !torch.vtensor<[1,32,7,7],f32>, %arg4: !torch.int, %arg5: !torch.float, %arg6: !torch.int) -> (!torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,32],f32>, !torch.vtensor<[],si64>, !torch.vtensor<[],si64>) {
    %true = torch.constant.bool true
    %float8.000000e-02 = torch.constant.float 8.000000e-02
    %true_0 = torch.constant.bool true
    %none = torch.constant.none
    %0:4 = torch.operator "torch.aten._scaled_dot_product_efficient_attention"(%arg0, %arg1, %arg2, %arg3, %true, %float8.000000e-02, %true_0, %none) : (!torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,7,7],f32>, !torch.bool, !torch.float, !torch.bool, !torch.none) -> (!torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,32],f32>, !torch.vtensor<[],si64>, !torch.vtensor<[],si64>) 
    return %0#0, %0#1, %0#2, %0#3 : !torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,32],f32>, !torch.vtensor<[],si64>, !torch.vtensor<[],si64>
  }
}

The error info printed：

/workspace/iree/tests/e2e/dlc_specific/torch-ir/ScaledDotProductEfficientAttention/ScaledDotProductEfficientAttention.mlir:7:12: error: failed to legalize operation 'torch.operator' that was explicitly marked illegal
    %0:4 = torch.operator "torch.aten._scaled_dot_product_efficient_attention"(%arg0, %arg1, %arg2, %arg3, %true, %float8.000000e-02, %true_0, %none) : (!torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,7,7],f32>, !torch.bool, !torch.float, !torch.bool, !torch.none) -> (!torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,32],f32>, !torch.vtensor<[],si64>, !torch.vtensor<[],si64>) 
           ^
/workspace/iree/tests/e2e/dlc_specific/torch-ir/ScaledDotProductEfficientAttention/ScaledDotProductEfficientAttention.mlir:7:12: note: see current operation: %3:4 = "torch.operator"(%arg0, %arg1, %arg2, %arg3, %0, %1, %0, %2) <{name = "torch.aten._scaled_dot_product_efficient_attention"}> : (!torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,7,7],f32>, !torch.bool, !torch.float, !torch.bool, !torch.none) -> (!torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,32],f32>, !torch.vtensor<[],si64>, !torch.vtensor<[],si64>)
// -----// IR Dump After ReduceOpVariants Failed (torch-reduce-op-variants) //----- //
func.func @main(%arg0: !torch.vtensor<[1,32,7,128],f32>, %arg1: !torch.vtensor<[1,32,7,128],f32>, %arg2: !torch.vtensor<[1,32,7,128],f32>, %arg3: !torch.vtensor<[1,32,7,7],f32>, %arg4: !torch.int, %arg5: !torch.float, %arg6: !torch.int) -> (!torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,32],f32>, !torch.vtensor<[],si64>, !torch.vtensor<[],si64>) attributes {torch.assume_strict_symbolic_shapes} {
  %true = torch.constant.bool true
  %float8.000000e-02 = torch.constant.float 8.000000e-02
  %none = torch.constant.none
  %0:4 = torch.operator "torch.aten._scaled_dot_product_efficient_attention"(%arg0, %arg1, %arg2, %arg3, %true, %float8.000000e-02, %true, %none) : (!torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,7,7],f32>, !torch.bool, !torch.float, !torch.bool, !torch.none) -> (!torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,32],f32>, !torch.vtensor<[],si64>, !torch.vtensor<[],si64>) 
  return %0#0, %0#1, %0#2, %0#3 : !torch.vtensor<[1,32,7,128],f32>, !torch.vtensor<[1,32,32],f32>, !torch.vtensor<[],si64>, !torch.vtensor<[],si64>
}

At present, it seems that efficient_attention does not support falling. Hope that the support of this op can be considered in the future. Thanks a lot!