[AutoTVM] Temporary fix to the stack overflow issue in autotvm task extraction

[icemelon]

Thanks for contributing to TVM! Please refer to guideline https://docs.tvm.ai/contribute/ for useful information and tips. After the pull request is submitted, please request code reviews from Reviewers by @ them in the pull request thread.

[icemelon]

cc @comaniac @merrymercy @kevinthesun

[comaniac]

LGTM

[jmorrill]

Thanks for looking into this. I have hit the stackoverflow before. I've been working around it by starting up my autotvm script on another thread after I set a huge stack size.

threading.stack_size(10 * 1024 * 1024)
t = threading.Thread(...)
t.start()
t.join()

[comaniac]

Thanks for looking into this. I have hit the stackoverflow before. I've been working around it by starting up my autotvm script on another thread after I set a huge stack size.

threading.stack_size(10 * 1024 * 1024)
t = threading.Thread(...)
t.start()
t.join()

We did try this solution but found two problems. First, it is not always working (at least not working for some of us). Second, it is not scalable. The ultimate goal should be a non-recursive graph traversing, but this workaround should solve >90% of current use cases.

btw, CI seems failed due to inconsistent return behaviors.

[icemelon]

@kevinthesun ci is green now. Could you take a look at the pr?