Deterministic fused cross-CTA dW reduction in RMSNorm backward#110
Closed
santoshmo wants to merge 1 commit into
Closed
Deterministic fused cross-CTA dW reduction in RMSNorm backward#110santoshmo wants to merge 1 commit into
santoshmo wants to merge 1 commit into
Conversation
Eliminates the separate .sum(dim=0) kernel for dw_partial reduction by fusing a deterministic last-CTA-reduces pattern into the backward kernel. Each CTA writes its partial to dw_partial[bidx, :] as before, then does a threadfence + atomic increment of a global counter. The last CTA to arrive loads all partials in fixed order 0..sm_count-1 and accumulates into dw_final, ensuring deterministic results across runs. Only enabled for N <= 8192 (cluster_n == 1). For larger N, falls back to the existing host-side .sum(dim=0) reduction. Based on the approach discussed in Dao-AILab#101. Co-authored-by: Aaron Wang <aaronwang04@users.noreply.github.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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Eliminates the separate .sum(dim=0) kernel for dw_partial reduction by fusing a deterministic last-CTA-reduces pattern into the backward kernel.
Each CTA writes its partial to dw_partial[bidx, :] as before, then does a threadfence + atomic increment of a global counter. The last CTA to arrive loads all partials in fixed order 0..sm_count-1 and accumulates into dw_final, ensuring deterministic results across runs.
Only enabled for N <= 8192 (cluster_n == 1). For larger N, falls back to the existing host-side .sum(dim=0) reduction.
Based on the approach discussed in #101 by @AaronWang04