[FIX] Fix threadpool reset by killing threads before destroying their shared queue

[tkonolige]

@electriclilies@mbs-octoml@junrushao1994

[electriclilies]

Thanks for the fix! Just curious, why wasn't this causing a segfault on linux?

[tkonolige]

No clue. Maybe the thread kill implementation on linux worked faster than the one on macOS?

[tqchen]

LGTM, thanks @tkonolige !

[junrushao]

Thanks Tristan!

[tkonolige]

I will note that I introduced this bug, so don't give me too much thanks :)

[tqchen]

I would separate the thanks from past events as we work together to make the code base better together.

[tmoreau89]

LGTM

[masahi]

The problem is still there https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-8658/1/pipeline/41

[tkonolige]

@mashi, I'm not sure that's related (although it is hanging on test_runtime_profiling.py). I believe #8656 fixed that issue. I've rebased, lets see if it has been fixed.

[jcf94]

Thanks! So is it ready to merge? @tkonolige@masahi

[masahi]

I've seen a number of recent CI jobs killed at test_runtime_profiling.py::test_vm[llvm -device=arm_cpu], and I thought this PR was the fix for it. No idea if this is related to the RPC socket error. Let's see if the issue is gone after merging #8656