[Bug] PaddlePaddle integration tests busted at main

[areusch]

We temporarily had a problem with our post-merge CI builds, so apologies this wasn't caught quite as quickly as we normally would. Seems that 5 integration tests related to PaddlePaddle are busted at main likely due to a merge conflict.

https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/main/1883/pipeline

FAILED tests/python/driver/tvmc/test_compiler.py::test_compile_paddle_module
FAILED tests/python/driver/tvmc/test_compiler.py::test_cross_compile_aarch64_paddle_module
FAILED tests/python/driver/tvmc/test_compiler.py::test_cross_compile_options_aarch64_paddle_module
FAILED tests/python/driver/tvmc/test_frontends.py::test_load_model__paddle - ...
FAILED tests/python/driver/tvmc/test_tvmc_common.py::test_compile_paddle_module_nchw_to_nhwc

I suspect #9083 might be related. Is anyone more familiar with it able to look? cc @jiangjiajun @masahi @leandron

[masahi]

I think this paddle PR merged yesterday is related #9126

Since #9083 was merged three days ago and we had no problem until yesterday. And there was about three days between the last CI job for 9126 ran before it was merged yesterday.

[masahi]

This is also blocking PR jobs, https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-9230/1/pipeline, so I'm going to send a revert. The question is, which of #9083 or #9126 we should revert?

cc @jiangjiajun @junrushao1994

[junrushao]

I have no idea why it happens, and no objection for a revert

[junrushao]

What about we try to revert #9083 first? Looks like the unittests came from that PR

[masahi]

Yeah, #9083 was merged three days ago and as far as I know, there have been no problem with that tests until yesterday. My hunch is that #9126, which apparently didn't run CI job after #9083 was merged, broke it.

[masahi]

Sent a revert for #9126 for now. We can decide later if we want to merge that or not.

[junrushao]

Sounds good to me!

[masahi]

#9232 is looking good

[jiangjiajun]

I just saw this issue, I'll try to reproduce this problem now

[jiangjiajun]

I have reproduced the problem, and send a pull request #9236 to fix the bug, let's see if this could solve the ci problem @masahi @junrushao1994 @areusch

[jiangjiajun]

I have reproduced the problem, and send a pull request #9236 to fix the bug, let's see if this could solve the ci problem @masahi @junrushao1994 @areusch

Now all the ci test passed, sorry for all the blocking PRs, we can merge #9236 and retrigger ci for the blocked pull request.

The reason this bug not detected by ci is,

1. Add TVMC Frontend for PaddlePaddle #9083 was merged at 10/05
2. [Frontend][PaddlePaddle] Add 10+ operators for PaddlePaddle #9126 passed all the ci test at 10/04, merged at 10/08

this caused #9126 didn't test with #9083