Following up on this. The MIL op does work on the Neural Engine. It runs in float16, which can overflow and yield the incorrect-looking results.

I've filed FB12150787 for this.