[graph_trainer] Save CUDA-to-CPU copies in SAC pass to match core behavior#2811
Closed
SherlockNoMad wants to merge 1 commit intogh/SherlockNoMad/16/basefrom
Closed
[graph_trainer] Save CUDA-to-CPU copies in SAC pass to match core behavior#2811SherlockNoMad wants to merge 1 commit intogh/SherlockNoMad/16/basefrom
SherlockNoMad wants to merge 1 commit intogh/SherlockNoMad/16/basefrom
Conversation
…avior Core's _apply_op_sac always marks aten._to_copy CUDA->CPU transfers as MUST_SAVE to avoid wastefully recomputing device transfers during backward (e.g., MoE D2H sync for all-to-all metadata). Graph_trainer's apply_sac_pass was missing this check, causing these transfers to fall through to PREFER_RECOMPUTE. Add the same logic at the FX graph level by inspecting the source node's fake tensor device and the target device kwarg. [ghstack-poisoned]
This was referenced Apr 3, 2026
meta-cla
bot
added
the
CLA Signed
SherlockNoMad
added a commit
that referenced
this pull request
Apr 9, 2026
…core behavior Port the fix from PR #2811 to graph_trainer's `apply_sac_pass`. Core's `_apply_op_sac` always marks `aten._to_copy` CUDA->CPU transfers as MUST_SAVE to avoid wastefully recomputing device transfers during backward (e.g., MoE D2H sync for all-to-all metadata). Graph_trainer's `apply_sac_pass` was missing this check.
4 tasks
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Stack from ghstack (oldest at bottom):
Core's _apply_op_sac always marks aten._to_copy CUDA->CPU transfers as
MUST_SAVE to avoid wastefully recomputing device transfers during
backward (e.g., MoE D2H sync for all-to-all metadata).
Graph_trainer's apply_sac_pass was missing this check, causing these
transfers to fall through to PREFER_RECOMPUTE. Add the same logic at
the FX graph level by inspecting the source node's fake tensor device
and the target device kwarg.