onnxscript: ORT Flaky Segmentation Fault during model load

models.zip contains two models, and they are only different on aten_sym_size, but results in differently. I tested them on the following code:

import onnx
import onnxruntime as ort

model = onnx.load("model_works.onnx")
model = onnx.load("model_segfault.onnx")

model_string = model.SerializeToString()
print(model_string)
sess_options = ort.SessionOptions()
sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL
sess = ort.InferenceSession(model_string, sess_options)

The one with Gather node flaky fails when InferenceSession is initialized, while the other one works fine. Even tried with ORT_DISABLE_ALL, but not worked either.

I am not sure if it’s some corner case of local function that isn’t handled well in ORT. I suspect this only happens when dynamic shape is involved, as I have seen it in other cases which I tried to enable dynamic shapes.

version: ONNX 1.13.1 ONNXRUNTIME 1.14.1

cc @gramalingam @justinchuby @BowenBao