[WIP] rafactor: make gqa_group_size
a function argument instead of template parameter
#262
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Outdated suggestions cannot be applied.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.
Our previous implementation treats
gqa_group_size
as a template parameter.We also have an implicit assumption that 8 should be divisible by
gqa_group_size
, this is not documented and cause lots of confusion for developers.#223 could be a workaround but this would quickly increase the binary size of this library if we keep adding more and more gqa group sizes.
This PR refactors the code so that
gqa_group_size
becomes a runtime function argument, and we don't need to compile one kernel for each of them. This would be bring very slight performance degradation togqa_group_size=1
but should work good in general. This PR also removes the requirement that "8 should be divisible bygqa_group_size
", so that any gqa group size should be supported.This PR should resolve #142 #181 #246 #254 #258 and so on.
Still working in progress, estimated finish time:
May 27th at noon (PST time).May 30thJune 3rd.