Add flag torch_jit_disable_warning_prints to allow disabling all warnings.warn

[gmagogsfm]

Adding a flag torch_jit_disable_warning_prints to optimize interpreter performance by suppressing (potentially large amount) of warnings.warn.

This is to work around TorchScript's warning behavior mismatch with Python. Python by default triggers a warning once per location but TorchScript doesn't support it. This causes same warning to trigger and print once per inference run, hurting performance.

[facebook-github-bot]

@gmagogsfm has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[facebook-github-bot]

@gmagogsfm has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[facebook-github-bot]

@gmagogsfm has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[facebook-github-bot]

@gmagogsfm has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[codecov]

Codecov Report

Merging #49313 (3ec71b1) into master (94a3d4b) will increase coverage by 0.00%.
The diff coverage is 50.00%.

@@           Coverage Diff           @@
##           master   #49313   +/-   ##
=======================================
  Coverage   80.56%   80.56%           
=======================================
  Files        1874     1874           
  Lines      202776   202777    +1     
=======================================
+ Hits       163362   163363    +1     
  Misses      39414    39414           

[facebook-github-bot]

@gmagogsfm has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[SplitInfinity]

Python by default triggers a warning once per location but TorchScript doesn't support it.

Why not?

[gmagogsfm]

Python by default triggers a warning once per location but TorchScript doesn't support it.

Why not?

It's a bit of long story, there are a couple reasons:

1. Python relies on the callstack to tell different "warnings.warn" apart. TorchScript can't accurately mimic this behavior because all the inlining and unrolling we do in optimization.
2. Even if we accept the fact that we don't match Python fully, and use the less good option of making each warning node fire only once. it is still very difficult because there is no good place to store the information of "this warning node has triggered before". Two places I can think of are "Code", which is supposed to represent the program. Or "InterpreterState", which is supposed to capture the execution state. "Code" doesn't work because it is supposed to be unchanged during execution, thus not representing execution state. "InterpreterState" doesn't work either because caller can (and does) destroy/recreate it between inference calls, thus the information of "warning has fired before" is lost between inference calls.

I am not entirely happy with current approach because it completely suppresses all warnings, but couldn't come up with a good alternative Let me know if you know of good options to implement the functionality of "limiting number of warnings fired per process regardless of how many independent inference calls are made".

[SplitInfinity]

Hm, okay.

[facebook-github-bot]

@gmagogsfm merged this pull request in 7518f54.