[onnx_importer.py] Fix dim_value None not correctly processed and missing Float8E4M3FNUZType.

[giacs-epic]

As per title. Changes tested on SHARK-TestSuite's alt_e2eshark.

[giacs-epic]

This is never None (protobuf default initializes), but rather an empty TensorShapeProto, which corresponds to a valid shape.

[zjgarvey]

I'm also going to take a look at the CI failure and let you know what I figure out.

[zjgarvey]

This section precisely occurs when the provided model is large, so saving a temp file would be expensive. I'd prefer to only do this if it had actually been modified, so perhaps add a bool to track if the model got modified by previous arg specifications, and only do this if so.

I'm also concerned about not saving external data in this case, since this is exactly when we would be exceeding the 2gb protobuf limit.

[zjgarvey]

Running

projects/pt1/tools/e2e_test.sh -c onnx -f ElementwiseFlattenBroadcastModule_basic -v

With this patch:

Crashes with:

projects/pt1/tools/e2e_test.sh: line 13: 38611 Floating point exception(core dumped) python -m e2e_testing.main "$@"

IR dump:

module {
  func.func @main_graph(%arg0: !torch.vtensor<[?],f32>, %arg1: !torch.vtensor<[],f32>) -> !torch.vtensor<[?],f32> attributes {torch.onnx_meta.ir_version = 9 : si64, torch.onnx_meta.opset_version = 20 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "2.7.0"} {
    %none = torch.constant.none
    %0 = torch.operator "onnx.Shape"(%arg1) : (!torch.vtensor<[],f32>) -> !torch.vtensor<[?],si64> 
    %1 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<_> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64> 
    %2 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__1> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64> 
    %3 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__2> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64> 
    %4 = torch.operator "onnx.Slice"(%0, %2, %3, %1) : (!torch.vtensor<[?],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[?],si64> 
    %5 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__3> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64> 
    %6 = torch.operator "onnx.Concat"(%4, %5) {torch.onnx.axis = 0 : si64} : (!torch.vtensor<[?],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1],si64> 
    %7 = torch.operator "onnx.Reshape"(%arg1, %6) {torch.onnx.allowzero = 0 : si64} : (!torch.vtensor<[],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[?],f32> 
    %8 = torch.operator "onnx.Mul"(%arg0, %7) : (!torch.vtensor<[?],f32>, !torch.vtensor<[?],f32>) -> !torch.vtensor<[?],f32> 
    return %8 : !torch.vtensor<[?],f32>
  }
}

{-#
  dialect_resources: {
    builtin: {
      _: "0x080000000000000000000000",
      __1: "0x080000000000000000000000",
      __2: "0x08000000FFFFFFFFFFFFFFFF",
      __3: "0x08000000FFFFFFFFFFFFFFFF"
    }
  }
#-}

On main:

module {
  func.func @main_graph(%arg0: !torch.vtensor<[?],f32>, %arg1: !torch.vtensor<[],f32>) -> !torch.vtensor<[?],f32> attributes {torch.onnx_meta.ir_version = 9 : si64, torch.onnx_meta.opset_version = 20 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "2.7.0"} {
    %none = torch.constant.none
    %0 = torch.operator "onnx.Shape"(%arg1) : (!torch.vtensor<[],f32>) -> !torch.vtensor<[0],si64> 
    %1 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<_> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64> 
    %2 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__1> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64> 
    %3 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__2> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64> 
    %4 = torch.operator "onnx.Slice"(%0, %2, %3, %1) : (!torch.vtensor<[0],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[0],si64> 
    %5 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__3> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64> 
    %6 = torch.operator "onnx.Concat"(%4, %5) {torch.onnx.axis = 0 : si64} : (!torch.vtensor<[0],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[1],si64> 
    %7 = torch.operator "onnx.Reshape"(%arg1, %6) {torch.onnx.allowzero = 0 : si64} : (!torch.vtensor<[],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[?],f32> 
    %8 = torch.operator "onnx.Mul"(%arg0, %7) : (!torch.vtensor<[?],f32>, !torch.vtensor<[?],f32>) -> !torch.vtensor<[?],f32> 
    return %8 : !torch.vtensor<[?],f32>
  }
}

{-#
  dialect_resources: {
    builtin: {
      _: "0x080000000000000000000000",
      __1: "0x080000000000000000000000",
      __2: "0x08000000FFFFFFFFFFFFFFFF",
      __3: "0x08000000FFFFFFFFFFFFFFFF"
    }
  }
#-}

I'm going to print out the onnx model and take a look at the shape information to see what is going on.

[zjgarvey]

Looking at a printout of the value info for the first shape node:

  value_info {
    name: "/Shape_output_0"
    type {
      tensor_type {
        elem_type: 7
        shape {
          dim {
            dim_value: 0
          }
        }
      }
    }
  }

So it looks like since the dim_value is actually 0, and not because it is unspecified, but because it is supposed to be zero.

I'm not sure if I have a good idea of how to fix this.

Do you know if shape inference will fill the value info fields with dim_params for actual dynamic dims?

Or is there python protobuf api for checking if a field is not explicitly set?

`