Expose Consecutive Hysteresis to Users #3553
Merged
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Outdated suggestions cannot be applied.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.
There's already a nice
consecutive_hysteresis
feature in the DynamicLossScaler that replenishes thehysteresis
whenever a non-overflowing iteration is encountered. This is useful for training runs that periodically face instabilities or bad samples.This PR exposes this existing feature to users via a new config option
consecutive_hysteresis
, which is added to the fp16 dict like:So if
"consecutive_hysteresis": true
,"hysteresis: 2"
, and every 2k iterations (outside the loss window) there's a single loss overflow, we'd restore thehysteresis
to 2 after each overflow and won't update the loss scale. Alternatively, if"consecutive_hysteresis": false
, we'd update the loss scale after everyhysteresis=2
loss overflows.