Add backward RMSNorm+Add fusion #2028
Open
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
Signed-off-by: Jan Bielak <[email protected]>
Signed-off-by: Jan Bielak <[email protected]>
Signed-off-by: Jan Bielak <[email protected]>
for more information, see https://pre-commit.ci
Signed-off-by: Jan Bielak <[email protected]>
|
/te-ci |
timmoon10
reviewed
Aug 6, 2025
Comment on lines
+20
to
+21
| auto kernel = launch_params.params.add ? &rmsnorm_bwd_tuned_kernel<Kernel_traits, true> | ||
| : &rmsnorm_bwd_tuned_kernel<Kernel_traits, false>; |
There was a problem hiding this comment.
We are doubling the number of kernel compilations for RMSNorm. This is fine for now, but it makes an NVRTC impl more compelling.
| @@ -111,6 +118,16 @@ __global__ __launch_bounds__(Ktraits::THREADS_PER_CTA) void rmsnorm_bwd_tuned_ke | |||
| } | |||
| } | |||
|
|
|||
| maybe_t<Ovec[LDGS], FusedAdd> add; | |||
There was a problem hiding this comment.
Future optimization: Load add directly into dx to avoid extra register usage. However, this would require that Ivec and Ovec are identical types.
transformer_engine/common/include/transformer_engine/normalization.h
Outdated
Show resolved
Hide resolved
Co-authored-by: Tim Moon <[email protected]> Signed-off-by: Jan Bielak <[email protected]> Signed-off-by: Jan Bielak <[email protected]>
janekb04
force-pushed
the
normalization-backward-add-fusion
branch
from
22d91b0 to
65aeee7
Compare
timmoon10
reviewed
Aug 7, 2025
Signed-off-by: Tim Moon <[email protected]>
|
/te-ci core pytorch |
Signed-off-by: Jan Bielak <[email protected]>
Signed-off-by: Jan Bielak <[email protected]>
|
/te-ci pytorch core |
Signed-off-by: Jan Bielak <[email protected]>
|
/te-ci core |
…error in `test_normalization.cu` Signed-off-by: Jan Bielak <[email protected]>
|
/te-ci core |
Signed-off-by: Jan Bielak <[email protected]>
|
/te-ci core |
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.
Description
Fuse
MakeExtraOutputwithRMSNormin the backward pass. Works through addingnvte_rmsnorm_bwd_add.Type of change
Changes
nvte_rmsnorm_bwd_addas well as PyTorch binding, test, and update kernel to support the fusionBackwardAddRMSNormfused operation, including testChecklist: