CUDA: fix insufficient buffer clearing for MMQ #10032
Merged
+3
−2
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
JohannesGaessler
added
the
Review Complexity : Low
slaren
approved these changes
Oct 24, 2024
4 tasks
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Outdated suggestions cannot be applied.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.
Fixes #10011 .
Followup to #10021 .
As of right now a quantized K cache works correctly with an FP16 model but not with a quantized model. The problem is that the clearing I added in #10021 is insufficient. It only cleared the last bit of the buffer but because multiple matrices are copied to the buffer at different regions this is insufficient. For an FP16 model this is not an issue because the initial buffer data is never NaN unless one of the FP16 values is NaN. For quantized data however the integer data can encode a non-finite value.