Internal x-ref
We also see this with vLLM. internal tlparse.
vLLM's custom backend invokes AOTAutograd+Inductor multiple times (see image), some way to customize the behavior of that, like tag each AOTAutograd+Inductor run, would make things a bit easier to grok.
