Repository navigation
Prefill remote Webshart shards from the training cursor - #3410
Merged
Merged
Conversation
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Sequential Webshart training now starts cache prefill at the physical shard window the sampler opens, including on checkpoint resume. Consuming a window frees space for more upcoming shards. A local main process manages prefill and closes the cursor with the epoch. Older Webshart builds continue to support on-demand reads. The cache cursor is provided by Webshart PR #53.
Set
SIMPLETUNER_WEBSHART_PREFILL_DOWNLOADSto tune the background reader independently of the saved dataset configuration. On the live job this allowed 16 background downloads with the original four-way foreground I/O and the original eight rank cursors.Performance
On four L40S GPUs, two 160-step HDiT+ runs used identical inputs, separate cold 1 GiB shard caches, and 32 images per rank. From the end of step 16 through step 159, mean wall time fell from 0.483 to 0.313 seconds per step. Throughput rose from 265 to 408 images/s. Elapsed time from the first shard-window scan through step 160 fell from 100.6 to 78.6 seconds. Sample order, pixels, and captions matched for all 640 rank-batches.
The live eight-H200 run averaged 0.798 seconds per step over 400 steps with on-demand reads and 0.347 over 600 steps with a full prefill buffer. These were different sequential shard windows, excluding checkpoint and validation pauses. The buffer reached 484 shards within 15 MiB of its 32 GiB limit and continued refilling as training advanced.
Verification
The Webshart epoch and Parquet backend suites passed 51 SimpleTuner tests. Coverage includes the resumed first window, reproduced batch order, cursor cleanup, and independent background concurrency.