Layernorm bwd OPT #1880
Open
Layernorm bwd OPT #1880
Conversation
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
This comment was marked as outdated.
This comment was marked as outdated.
Sorry, something went wrong.
|
Maybe we can follow recent cuda change here pytorch/pytorch@73b4938 |
There was a problem hiding this comment.
Pull Request Overview
This PR optimizes the backward pass computation for layer normalization's gamma and beta gradients by implementing a two-stage column reduction approach to improve parallelism when the matrix dimension M is much larger than N.
- Introduces a new
GammaBetaReduceFunctorkernel that uses tiled computation with local memory for better occupancy - Adds logic to automatically select between the optimized two-stage reduction and the existing simple kernel based on occupancy thresholds
- Implements separate code paths for different combinations of gamma and beta gradient computations
xytintel
approved these changes
Aug 8, 2025
liangan1
approved these changes
Aug 19, 2025
Co-authored-by: Copilot <[email protected]>
Co-authored-by: Copilot <[email protected]>
jianyizh
force-pushed
the
jianyi/ln_bwd
branch
from
e8bef72 to
0bfee0f
Compare
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Outdated suggestions cannot be applied.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.
I noticed layer norm backward on gamma and beta is very slow when column is much longer. i.e. [M,N] column reduction and M>>N.
For example, in timm tnt_s_patch16_224 training, layernorm bwd shape [25088,16,24], normalized shape [24]. it will only launch one workgroup. I use a two staged column reduction to increase parallelism. GammaBetaBackwardSimpleKernelFunctor takes 9 ms on PVC, 8.5ms on BMG. After opt, we use GammaBetaReduceFunctor and two sum to do column reduction, they will take 0.09ms + 0.06ms x2 on PVC and 0.19ms + 0.04ms x 2 on BMG