Setting `num_elements_threshold` <= 2 during import fails compilation for ONNX Zoo models

[vinayakdsci]

After applying the following patch to helper_classes.py:

diff --git a/alt_e2eshark/onnx_tests/helper_classes.py b/alt_e2eshark/onnx_tests/helper_classes.py
index af30cfe..41651c9 100644
--- a/alt_e2eshark/onnx_tests/helper_classes.py
+++ b/alt_e2eshark/onnx_tests/helper_classes.py
@@ -21,6 +21,7 @@ from e2e_testing import azutils
 from e2e_testing.framework import (
     ExtraOptions,
     ImporterOptions,
+    RuntimeOptions,
     OnnxModelInfo,
     TestTensors,
 )
@@ -164,10 +165,16 @@ class OnnxModelZooDownloadableModel(OnnxModelInfo):
         return yaml_path
 
     def update_extra_options(self):
+        param_str = f"parameters=model={Path(self.model).parent}" + "/model.torch_onnx_params.irpa"
         # called in __init__
         # default to using opset version 21 for all ONNX Model Zoo models.
         self.extra_options = ExtraOptions(
-            import_model_options=ImporterOptions(opset_version=21)
+            import_model_options=ImporterOptions(opset_version=21, externalize_params=True, num_elements_threshold=2),
+            compiled_inference_options=RuntimeOptions(
+                common_extra_args=(
+                    param_str,
+                ),
+            )
         )
 
     def update_input_name_to_shape_map(self):

All the MobileNet models fail compilation with the following error:

SHARK-TestSuite/alt_e2eshark/test-run/mobilenetv3_small_100_Opset16_timm/model.torch_onnx.mlir:343:12: error: failed to legalize operation 'torch.aten.sum.dim_IntList' that was explicitly marked illegal
    %113 = torch.operator "onnx.ReduceMean"(%111, %112) {torch.onnx.keepdims = 1 : si64} : (!torch.vtensor<[1,16,56,56],f32>, !torch.vtensor<[2],si64>) -> !torch.vtensor<[1,16,1,1],f32> 
           ^
/home/vinaydev/SHARK-TestSuite/alt_e2eshark/test-run/mobilenetv3_small_100_Opset16_timm/model.torch_onnx.mlir:343:12: note: see current operation: %636 = "torch.aten.sum.dim_IntList"(%509, %635, %21, %32) : (!torch.vtensor<[1,16,56,56],f32>, !torch.list<int>, !torch.bool, !torch.none) -> !torch.vtensor<[1,16,1,1],f32>

This issue is observable in a newly created environment when setting num_elements_threshold to any value <= 2 in ImporterOptions.

The failing models pass once the threshold is increased to more than 2.

[vinayakdsci]

cc @zjgarvey @pdhirajkumarprasad @pravg-amd @amd-vivekag.

[zjgarvey]

Yeah, most models will have things like static shapes as constant tensors in the IR. We can't scalarize those if they are externalized.