fix: skip duplicate B/C RMSNorm in checkpoint_lvl=1 backward#964
Open
Chessing234 wants to merge 1 commit into
Open
fix: skip duplicate B/C RMSNorm in checkpoint_lvl=1 backward#964Chessing234 wants to merge 1 commit into
Chessing234 wants to merge 1 commit into
Conversation
When checkpoint_lvl==1, B and C are already RMS-normalized in forward and saved via save_for_backward. Re-applying rms_norm_forward in backward double-normalizes them before selective_scan_cuda.bwd. Fixes state-spaces#885 Co-authored-by: Cursor <cursoragent@cursor.com>
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fixes #885
Bug
With
checkpoint_lvl=1andb_rms_weight/c_rms_weight,MambaInnerFn.backwardapplies RMSNorm to B and C a second time even though forward already normalized them and saved the post-norm tensors.Root cause
deltais cleared beforesave_for_backwardso it is recomputed once in backward; B and C are saved post-norm but backward still re-ranrms_norm_forwardon those saved values.Why this fix is correct
Drop the redundant B/C RMSNorm block in the
checkpoint_lvl == 1path so backward uses the same post-norm B/C that forward passed toselective_scan_cuda.fwd, matching the intended checkpointing behavior.Made with Cursor