Repository navigation
Conversation
stage index arrays beside their cache paths and replace them before publishing the description. interrupted writes leave the final paths readable or absent, so the next normal build can regenerate an incomplete cache. cover interrupted gpt, blended, and masked dataset cache writes through dataset construction, sample reads, and cache reloads. fixes NVIDIA#8189 Signed-off-by: ZenAlexa <zimingwang945@gmail.com>
Signed-off-by: ZenAlexa <zimingwang945@gmail.com>
svcnvidia-nemo-ci
added
the
waiting-on-maintainers
This branch has not been deployed
This file contains hidden or 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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
what changed
stage dataset index arrays on the same filesystem, replace their final paths after all writes finish, and publish the description last. gpt, blended, and masked datasets use the same save path. cache names, formats, and existing read paths stay unchanged.
fixes #8189.
why
an interrupted
numpy.savecurrently leaves a truncated file that passes the cache hit check on restart. staging keeps incomplete writes away from the final paths. a normal restart rebuilds an incomplete cache, and existing readers keep their original mappings during replacement.this prevents new torn writes. caches already damaged by earlier runs still need rebuilding. power-loss durability and fast-cache-load synchronization are outside this change.
validation
tools/autoformat.shcompleted in check mode. black, isort, pylint, and ruff pass. its non-blocking mypy step reports existing type errors and missing imports in the lint environment.Project CI: Please run the project GPU dataset test runner and its supported shared-filesystem cache cases. The Linux CPU/Gloo recovery, process-interruption, and mmap-reader results above are ready for review.