Tried to visualize mlir mhlo - got No graphs

[apivovarov]

How to visualize mlir mhlo functions. e.g.

func.func @main(%arg0: tensor<3x4x2xf32>) -> tensor<3x4x2xf32> {
  %0 = mhlo.constant dense<1.000000e+04> : tensor<f32>
  %1 = "mhlo.broadcast_in_dim"(%0) {broadcast_dimensions = dense<> : tensor<0xi64>} : (tensor<f32>) -> tensor<3x4x2xf32>
  %2 = mhlo.add %arg0, %1 : tensor<3x4x2xf32>
  return %2 : tensor<3x4x2xf32>
}

Currently I see just "No graphs" in the middle of the page.

v.0.1.8

[pkgoogle]

Hi @apivovarov, we don't currently support mhlo, but it shouldn't be too hard to add -- though it won't have any special formatting for now.

[apivovarov]

Hi @pkgoogle. Thank you for your reply. Can I ask you couple additional questions:

1. Does model-explorer support stablehlo mlir?
2. What mlir is currently supported? Any examples I can look? builtin_mlir_adapter.py says "A built-in adapter that converts MLIR file to Model Explorer format."

[yijie-yang]

Hi @apivovarov,

I can answer the questions. Currently model-explorer supports tf, tfl and stablehlo dialect MLIR files. Here is an example of stablehlo mlir: https://github.com/openxla/stablehlo/blob/main/stablehlo/tests/interpret/collective_broadcast.mlir

For mhlo dialect, I can make a quick fix to use stablehlo adapter to visualize it since they are all part of hlo family and have similar syntax.

[ptoulme-aws]

You can convert from hlo/mhlo to stablehlo with the hlo translate tool in OpenXLA

[yijie-yang]

The mhlo mlir is now supported in model-explorer-v0.1.10. Please give it a try!