Why isn't there a binding for clearing the XLAGraphExecutor::ComputationCache?

[jameszianxuTT]

We have exposed this function in our tenstorrent fork and found that it works for post-test cleanup.

My assumption was that TPU runtime does not require such a feature because it does not bind scarce device resources to PJRTComputation lifetime. So, implementers did not find it necessary to implement such a function. Is that correct? Were there any other reasons to avoid exposing this to the user?

[ysiraichi]

Thank you for your question.
Unfortunately, I don't know the historical reasons around that.
@JackCaoG any thoughts?

[JackCaoG]

I did not remember the details, but we were not expecting user to need to manually flush the computation cache. I think it is fair to just give it a try and check if it will work(I think it will).

[jameszianxuTT]

I have found that this works without issue and we've had it in our CI @ tenstorrent for a while now.

I've put up a PR to upstream this small change, if you think it would be acceptable.