Revert "Static alloc for hybridblock"

[marcoabreu]

Reverts #11313

My veto has been ignored and thus I'm allowed to revert this PR. See https://www.apache.org/foundation/voting.html for reference.

[tqchen]

it seems you have been using veto too often lately, please @marcoabreu we are not being unreasonable here, your concern about testcases have been addressed.

[marcoabreu]

My as well as Thomas' concerns have not been addressed.

[tqchen]

If you have specific concerns, raise them. Your concern of lack of test coverage has been addressed, and you are abusing your right as committer to revert good changes back...

[marcoabreu]

#11171 has not been addressed

Pedros review in #10817 has not been addressed

Thomas comment #10817 (comment) has not been addressed

[ThomasDelteil]

I think I have raised specific concerns (twice) that have gone unanswered, just in case they got lost in the flow, let me write them here again:

• "good changes", I'm sure they are, but the problem with the previous PR was that it was actually making things slower due to a bug, hopefully now fixed. I think we cannot just rely on blind trust, being data driven is beneficial for everybody. Could we have a runtime of maybe just a MNIST training to show that the speed is actually improved and rejoice accordingly?
• The new API says that static_shape cannot be used without static_alloc. Can we assert that static_alloc is set if users set static_shape to True to avoid undefined behavior?

Bonus ask:

• a description of the PR to help reviewers, and users alike to appreciate the changes and guide them through the 1400 lines of code, that have barely any comments.
• better documentation on the impact of the bulk_size arguments in the hybridize() function

[tqchen]

This PR is already fixing the problems and added test coverages. If we just block this, how can we move forward?

[zheng-da]

@marcoabreu I have been intensively using CachedOp with static memory allocation this week. I also read the code this week. Other than the small bug reported previously, I don't find problems and it works very well with me in my PR. #10451
I observe the memory allocation behavior and it works exactly as I expected.

What are you looking for? If you are looking for test code, my entire PR can serve as test code. If you are looking for review, I have been reading the code for the entire week. I don't think it's reasonable to block other people's work for no reason.

[zheng-da]

@marcoabreu I think you should at least give people who understand the PR the opportunity to review it.

[marcoabreu]

Exactly, that's all I'm asking for. Give the community enough time to review the PR and address their concerns before merging it.