Repository navigation
optimizer: keep the probe's margin when the whole job fits - #43
Merged
niklashoelter merged 1 commit intoOct 5, 2026
Merged
Conversation
The whole-job shortcut ran a job as one batch whenever its forward-pass peak fit in max_memory_padding of free memory. Optimization needs more than one forward pass (optimizer state, fragmentation), so jobs at 0.78-0.80 of the budget passed the check and then ran out of memory mid-run (UMA-s, 878 structures, peak 18.1-18.8 of 23.4 GB). Require peak * memory_scaling_factor to fit instead. The probe returns a batch one growth step below a size that fit, so this gives the shortcut the same margin. Jobs over the threshold fall back to the probe.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The whole-job shortcut ran a job as one batch whenever its forward-pass peak fit in max_memory_padding of free memory. Optimization needs more than one forward pass (optimizer state, fragmentation), so jobs at 0.78-0.80 of the budget passed the check and then ran out of memory mid-run (UMA-s, 878 structures, peak 18.1-18.8 of 23.4 GB).
Require peak * memory_scaling_factor to fit instead. The probe returns a batch one growth step below a size that fit, so this gives the shortcut the same margin. Jobs over the threshold fall back to the probe.