Repository navigation
Add CSA2 chunked scoring and selective recompute compatibility - #7761
Closed
HollowMan6 wants to merge 3 commits into
Closed
HollowMan6 wants to merge 3 commits into
HollowMan6 wants to merge 3 commits into
Conversation
Keep the existing eager attention contract while bounding no-grad CSA2 score memory and allowing non-attention selective recompute. Recompute single-pass mHC mappings to reduce activation retention. Signed-off-by: Hollow Man <hollowman@opensuse.org>
HollowMan6
force-pushed
the
csa2-mhc-recompute-memory-fixes
branch
from
September 30, 2026 22:31
ae0bebc to
ef16913
Compare
Contributor
Author
|
@guihong-nv This is a follow-up PR on top of #7500. It reduces memory by chunking the no-grad CSA2 indexer scoring and recomputing the single-pass mHC mappings, and it allows selective recompute with CSA2 when no attention modules are recomputed. Verified together with NVIDIA-NeMo/Megatron-Bridge#6199. Please let me know if you have any comments. |
svcnvidia-nemo-ci
added
the
waiting-on-maintainers
Contributor
|
/ok to test ef16913 |
svcnvidia-nemo-ci
removed
the
waiting-on-maintainers
Signed-off-by: Hollow Man <hollowman@opensuse.org>
Connor-XY
approved these changes
Oct 5, 2026
svcnvidia-nemo-ci
added
the
waiting-on-maintainers
Signed-off-by: Hollow Man <hollowman@opensuse.org>
This was referenced Oct 11, 2026
Open
Contributor
Author
|
Merged into #8018 |
svcnvidia-nemo-ci
removed
the
waiting-on-maintainers
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Stacked on #7500.
Summary
_chunked_csa2_indexer_scores), so the full per-head score tensor is never materialized at once.core_attn,gdn_norm_out,gdp_qkv,gdp_in_proj,shortcut_pre_mlp_layernorm) are recomputed. Full recompute is still rejected because it would re-publish cross-layer KV.