Enable assymetric quantization for all MultiHeadAttention qdq layers#2468
Merged
Conversation
wpietka
force-pushed
the
dev/wpietkax/always-asymmetric-qdq-for-int8
branch
from
e07f851 to
5a2d00b
Compare
Signed-off-by: Wojciech Piętka <wojciechx.pietka@intel.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Type of Change
Improvement
Description
Currently QDynamicMultiHeadAttention creates 4 separate qdq layers. Two of them are potentially asymmetrical - depending on activation dtype - and two others are always symmetrical. There are two problems here: firstly the symmetrical/asymmetrical policy differs from Static version which allows asymmetrical computation for all qdq layers and secondly dynamic version doesn't need 4 separate qdq layers. Since scale is computed in runtime and not preserved in the layer itself a single qdq layer can be reused for queries, keys and values. Attention qdq stays separate due to fixed range.
Expected Behavior & Potential Risk
Slightly increased dynamic layers accuracy
How has this PR been tested?
Vit benchmark has been run with different configurations, and the results show better accuracy with asymmetric layers
Dependency Change?
No dependency changes