Reland "Support fusing broadcast transposes with attention" #19962
Conversation
This file contains 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
IanWood1
force-pushed
the
reland_bcast_attn
branch
from
8b3ba8d to
d624520
Compare
| @@ -37,12 +37,10 @@ void AttentionOpDetail::inferFromIndexingMaps(AffineMap qMap, AffineMap kMap, | |||
| llvm::SmallDenseSet<int64_t> vSet = findPermutationsIndexingOperand(vMap); | |||
| llvm::SmallDenseSet<int64_t> oSet = findPermutationsIndexingOperand(oMap); | |||
|
|
|||
| // B = (Q & K & O) U (K & V & O) | |||
| // B = (Q & V) U (K & O) | |||
There was a problem hiding this comment.
Interesting. Is this true. cc @Groverkss
There was a problem hiding this comment.
Yes. We can actually make it just (Q & V) or just (K & O) also because they are the same.
K := batch x k1 x k2
O := batch x m x n
so their intersection has to be the batch dimensions.
Similarily,
Q := batch x m x k1
V := batch x k2 x n
so their intersection has to be the batch dimension
Interestingly, this actually made me realize we are doing GQA wrong. GQA is a broadcast on the M dimension, which is why it's only on Q.
…ree-org#19835)" This reverts commit 9870a6d. Signed-off-by: Ian Wood <ianwood2024@u.northwestern.edu>
Signed-off-by: Ian Wood <ianwood2024@u.northwestern.edu>
MaheshRavishankar
approved these changes
Feb 12, 2025
|
I'll merge after double checking that this doesn't break 405b |
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Outdated suggestions cannot be applied.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.
Reland the changes to fold attention ops with broadcasts with a small tweak to
AttentionOpDetailso that the batch dimensions are properly computed when an operand is broadcasted.Original PR #19828
Revert PR #19835
Issue causing revert #19833