Add support for non-power-of-two heads with Alibi #1611
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Outdated suggestions cannot be applied.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.
This change adds support for non-power-of-two number of heads in Alibi.
The code for context Alibi was likely derived from FasterTransformer, which produces wrong (not monotonically decreasing) slopes on the half-interval [h_pow_2, h). The code for generation Alibi is open and we rewrote
generate_alibi_slopes()
so that it is way simpler while returning the exactly same result for power of two and correctly handling the others. Besides, we added support for padded heads, e.g. 66 -> 72 which is required by tp=8. Very few attempted to train Alibi models with non-power-of-two number of heads; we discovered that this change was essential for producing correct results in ours.We have tested this change for precise backward compatibility with powers of two.