Attempt to improve FlashMLA on the CPU #277
Merged
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Outdated suggestions cannot be applied.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.
@saood06 Can you try if this works better for your setup with
-mla 3 -fa
? Thanks.There is a faster path for TG with FA and
mla=1,3
. But it only gets taken if some values are a multiple of the number of threads. This PR changes the implementation to also take the fast path when this is not the case. On a 32-coreAVX2
system I observe some speedup with 24 and 48 threads compared to main, so would be curious to know if this also improves things on a dual-socket system.