Update SYNC_UPSTREAM.md #2544
Update SYNC_UPSTREAM.md #2544
Conversation
i-chaochen
requested review from
hsharsha,
draganmladjenovic,
zoranjovanovic-ns and
jayfurmanek
| @@ -4,6 +4,8 @@ How to synchonize TensorFlow ROCm port with upstream TensorFlow | |||
| This article shows the step to synchronize TensorFlow with upstream. The | |||
| process is currently carried out manually but it could and should be automated. | |||
|
|
|||
| **Important:** Synchronization should be performed on a weekly basis. Even if it takes more than a week to merge the PR, the next weekly sync must not be postponed. This way, long breaks (e.g. one month) between syncs will be avoided. | |||
|
|
|||
There was a problem hiding this comment.
As we discussed, I think we should have something like this:
Synchronization should be performed on a weekly basis. Even if it has taken more than 2 weeks (>2) to merge the PR, the next weekly sync should first sync in the middle week between the last one and current one, and then sync the current week.
For example, if the last weekly-sync PR is 3 weeks ago (2024-04-29), and this weekly-sync (supposed to be 2024-05-13) should start from 2024-05-06, and it needs another weekly-sync at 2024-05-13.
This way, long breaks (e.g. one month) between syncs will be avoided.
| * Merge to that commit point (i.e. perform a "mini sync") | ||
| * Resolve merge conflicts based on what was done in that weekly sync | ||
| * Test | ||
| * Repeat until the required commit is found |
There was a problem hiding this comment.
You can also provide reference to https://confluence.amd.com/display/~cchen104/XLA+hotfix+track to cherry-pick XLA hot fixes if the build breaks in xla
|
retest Ubuntu-GPU-single please |
Added instructions for the event when weekly sync takes too long to merge, as well as debugging tips.