Make OL gradient A2A pipeline-triggered (#4498) - #4498
Closed
spmex wants to merge 1 commit into
Closed
Conversation
Contributor
|
@spmex has exported this pull request. If you are a Meta employee, you can view the originating Diff in D110148587. |
Summary: ## Context `PECAll2AllSeqWait.backward` eagerly fired the OL gradient AllToAll, so gradient communication was not fully under pipeline control -- which the compute-ahead pipeline needs for overlap and for the deferred-NOL schedule. ## Approach - Remove the eager OL `dist()` from `PECAll2AllSeqWait.backward` so backward only splits the gradient and creates both OL/NOL `PECGradientApply` instances. The pipeline now starts both AllToAlls via `grad_dist`, making gradient communication fully pipeline-controlled (OL before optimizer, NOL deferred to the next batch). - No test change needed: the existing `_pec_backward` already calls `grad_dist(is_overlapped=True)` before waiting OL, so the OL path still triggers `dist()`. ___ Differential Revision: D110148587
spmex
force-pushed
the
export-D110148587
branch
from
August 6, 2026 17:00
fd8af09 to
66bb5aa
Compare
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary:
Context
PECAll2AllSeqWait.backwardeagerly fired the OL gradient AllToAll, so gradient communication was not fully under pipeline control -- which the compute-ahead pipeline needs for overlap and for the deferred-NOL schedule.Approach
dist()fromPECAll2AllSeqWait.backwardso backward only splits the gradient and creates both OL/NOLPECGradientApplyinstances. The pipeline now starts both AllToAlls viagrad_dist, making gradient communication fully pipeline-controlled (OL before optimizer, NOL deferred to the next batch)._pec_backwardalready callsgrad_dist(is_overlapped=True)before waiting OL, so the OL path still triggersdist().Differential Revision: D110148587