-
Notifications
You must be signed in to change notification settings - Fork 6.8k
Fix for handling negative indices in the fusion of slice #17937
Conversation
Hey @ptrendx , Thanks for submitting the PR
CI supported jobs: [miscellaneous, website, centos-cpu, unix-cpu, windows-cpu, unix-gpu, centos-gpu, windows-gpu, sanity, edge, clang] Note: |
Thanks @ptrendx. Do you plan to backport this change to 1.6 and perhaps have a 1.6.1 release? |
unix-gpu and unix-cpu failed with gcc segfaulting - @ChaiBapchya have you seen this before? |
Nope.. does it have something to do with your change? |
I don't believe so - my change does not even touch files that are compiled in CPU only build. |
Let's test the bot :-) @mxnet-bot run ci [unix-cpu] |
Jenkins CI successfully triggered : [unix-cpu] |
@ptrendx Thank you for the ping. @ciyongch will help to manage the release of 1.7 and will send out the plan and timeline to dev@ soon. |
@mxnet-bot run ci [unix-gpu] |
Jenkins CI successfully triggered : [unix-gpu] |
@mxnet-bot run ci [unix-gpu] |
Jenkins CI successfully triggered : [unix-gpu] |
@szha @ChaiBapchya Is anybody looking at this hang after e.g. |
Yes. Getting a stacktrace as it hangs needs to be done before the instance is shut down by the autoscaler [as it is not feasible to disable autoscaler]. Right now only 1 unix-gpu instance is running and the python3 tests have passed. So we really have to time it such that the gpu test fail and the node is alive so that we get the stacktrace. @leezu found that this might be operator related bug that can be tested locally.
If I can reproduce it locally then we can ID the issue with the operator test. |
@ChaiBapchya the hang occurs after ~10 minutes but the instance will remain available until timeout happens (3 hours). So @ptrendx's suggestion should also work. |
@mxnet-bot run ci [centos-gpu, unix-gpu] |
Jenkins CI successfully triggered : [unix-gpu, centos-gpu] |
@mxnet-bot run ci [centos-gpu, unix-gpu] |
Jenkins CI successfully triggered : [unix-gpu, centos-gpu] |
Added the v1.x label even though this PR goes to master in order to track that we want this included in 1.7 @ciyongch FYI. |
Yes, I'm going to backport my prs. |
* Fix for handling of negative axis, begin and end in fusion of slice ops * Added test
* Fix for handling of negative axis, begin and end in fusion of slice ops * Added test
Thanks for the fix. I'm wondering when are you planning to release a fix for 1.6? |
@ptrendx not everyone is using the binary wheels from pypi. Having the fix in the 1.6 branch can still be useful despite not having a 1.6.1 release in the short term |
I see. Ok, will push it there. |
* Fix for handling of negative axis, begin and end in fusion of slice ops * Added test
Description
Fixes #17914
@leezu @vafl
Checklist
Essentials
Please feel free to remove inapplicable items for your PR.
Changes
build_tuple
helper function and added tests