Fix synchronization issues in RNG checkpointing utils #5273
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: Szymon Karpiński <skarpinski@nvidia.com>
Signed-off-by: Szymon Karpiński <skarpinski@nvidia.com>
Signed-off-by: Szymon Karpiński <skarpinski@nvidia.com>
Signed-off-by: Szymon Karpiński <skarpinski@nvidia.com>
mzient
reviewed
Jan 12, 2024
Comment on lines
+62
to
+64
| CUDA_CALL(cudaMemcpyAsync(states_, other.states_, sizeof(curandState) * len_, | ||
| cudaMemcpyDeviceToDevice, cudaStreamDefault)); | ||
| CUDA_CALL(cudaStreamSynchronize(cudaStreamDefault)); |
There was a problem hiding this comment.
It will do for a quick bug fix, but it would probably be better to use a proper stream.
mzient
approved these changes
Jan 12, 2024
banasraf
approved these changes
Jan 15, 2024
|
!build |
|
CI MESSAGE: [12105479]: BUILD STARTED |
|
CI MESSAGE: [12105479]: BUILD FAILED |
|
CI MESSAGE: [12105479]: BUILD PASSED |
stiepan
pushed a commit
that referenced
this pull request
Jan 16, 2024
This PR fixes two synchronization bugs. Both problems happen when you're restoring from checkpoint in random GPU operators and are due to lack of synchronization after H2D & D2D copies. --------- Signed-off-by: Szymon Karpiński <skarpinski@nvidia.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Category:
Bug fix (non-breaking change which fixes an issue)
Description:
This PR fixes two synchronization bugs pointed out by @mzient . Both problems happen when you're restoring from checkpoint in random GPU operators and are due to lack of synchronization after H2D & D2D copies.
PS. Another problem pointed out was that synchronization between
SaveStateandSerializeis handled by the pipeline and not by those functions. This is a design problem and I added a TODO to fix it. I'm currently reworking how synchronization works when a checkpoint is saved, so I'll fix it there.Additional information:
Affected modules and functionalities:
RngCheckpointingUtils
Key points relevant for the review:
Tests:
Checklist
Documentation
DALI team only
Requirements
REQ IDs: N/A
JIRA TASK: DALI-3777