When converting Tensorflow models to Tensorflow Lite (.tflite files) format, it is very strongly advised to use the exact same Tensorflow Lite version for the conversion and for the inference on the device.
I have myself had issues in the past where converting a model with a version mismatch wouldn't work. One typical error with image labeling is "Didn't find op for builtin opcode ..." when trying to load a model on the device, after a successful conversion.
What is the Tensorflow Lite runtime version internally used by ML Kit? This piece of information is nowhere to be found.