Make check_isfinite, check_scale optional in clip_global_norm

[leezu]

Description

Make check_isfinite, check_scale optional in clip_global_norm. If both are set to false, clip_global_norm does not force any synchronization and throughput can be increased. Note if check_scale=False, this requires multiplying all arrays with 1 in cases where the multiplication could be skipped by doing a blocking check.

While this PR preserves the old default behavior of using blocking calls, we may want to change the default behavior to improve throughput.

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• For new C++ functions in header files, their functionalities and arguments are documented.
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• Make check_isfinite, check_scale optional in clip_global_norm

Comments

• This improves throughput in Add BIG LSTM + CNN INPUTS to large_word_language_model.py dmlc/gluon-nlp#253 by around 10%.
  @eric-haibin-lin

[lupesko]

@leezu can you please follow the community's convention to file a JIRA and update the issue title?
@eric-haibin-lin bouncing for review.

[leezu]

@lupesko I considered this to fall into "PRs with tiny changes" which don't require JIRA. Currently tests are failing, I'll need to add some fixes before review can proceed. Sorry for the delay

[szha]

rescale is not blocking

[leezu]

Yes. But previously there is a blocking call to check if re-scale is necessary.
If check_scale is False, we always rescale, possibly by 1.

[szha]

I see. Maybe we can use contrib.cond for this

[leezu]

contrib.cond would only help when working with the symbolic interface. For ndarray, contrib.cond also uses a blocking .asscalar() call. clip_global_norm always works with the ndarray API.

[leezu]

For now the PR preserves the old default behavior. I'm happy to remove the check_scale argument and to always rescale, trading off computation against avoiding blocking calls. That would assume that it is always cheaper/better do the rescaling than to wait for a blocking asscalar() and potentially avoid rescaling.

[eric-haibin-lin]

+1 for removing check_scale. It doesn't seem necessary. We can always perform a scaling.