[MetaSchedule] Fix XGBoost Import Issue

[zxybazh]

Previous upgrade introduced a import of xgboost in meta_schedule, removed in current version by using a function to return the call back class.

We've recently introduced a XGBoost Model upgrade to support new xgboost version of callback class in #12141. However, in this PR it uses a function called optional_xgboost_callback that works to avoid compatibility issue (xgboost 1.5.2 v.s. 1.6.0). In this specific function, it tries to import the newly introduced xgboost callback class and create a new class using it as base class. This actually imported xgboost when meta_schedule is imported, which is not ideal because xgboost is not a dependency of tvm and meta_schedule, it should only be required when xgboost cost model is employed. This PR fixes the problem by moving the class and the function mentioened above under a function that returns this class when needed. In this way we avoided unwanted import of xgboost in meta_schedule.

CC @shingjan @junrushao

[shingjan]

Thanks for sending this! Moving xgboost out as a top-level import makes sense here.

[junrushao]

LGTM!

[leandron]

Hi @zxybazh, thanks for the PR. Can you clarify what this change is actually doing, in the description of the PR? I see some functions moved around, but I can't understand what is actually being introduced.

Apart from that, can you also add a unit test to make sure your fix don't get regress in future, as it seems this is fixing something that stopped working, and we didn't notice because we don't have a test to assert that behaviour. Thanks.

[zxybazh]

Hi @leandron, thanks for looking into this PR! Let me explain more context of what's going on in this PR. We've recently introduced a XGBoost Model upgrade to support new xgboost version of callback class in #12141. However, in this PR it uses a function called optional_xgboost_callback that works to avoid compatibility issue (xgboost 1.5.2 v.s. 1.6.0). In this specific function, it tries to import the newly introduced xgboost callback class and create a new class using it as base class. This actually imported xgboost when meta_schedule is imported, which is not ideal because xgboost is not a dependency of tvm and meta_schedule, it should only be required when xgboost cost model is employed. This PR fixes the problem by moving the class and the function mentioened above under a function that returns this class when needed. In this way we avoided unwanted import of xgboost in meta_schedule.

In this case, it's a bit tricky to create a unit test for that, I've tested locally though. Would really appreciate it if you know how to create a test case for unwanted import in meta_schedule.

[zxybazh]

Hi, I made some changes to the PR so that the test for legacy callback function is only run for certain xgboost version, and fixed the CI issues. Would you please take another look, thanks!

[junrushao]

I think the main thing @leandron cares about is the documentation for each PR, i.e. would you mind if you add your explanation to the git commit?

[junrushao]

i added the detailed explanation to the commit message