Remove hardcoded value from softmax in flat_pa #280
Merged
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
kzawora-intel
approved these changes
Sep 12, 2024
We had without the workaround to .max() . With it there is a slight drop, but I don't have exact numbers at hand. |
|
I tested this change on mixtral and accuracy is still fine. Merging then. |
xuechendi
added a commit
to xuechendi/vllm-fork
that referenced
this pull request
Sep 12, 2024
This reverts commit 35a4a98.
zhouyu5
pushed a commit
to zhouyu5/vllm-fork
that referenced
this pull request
Sep 13, 2024
This PR removes the hardcoded value used to normalize softmax in flat_pa . Current approach is to use the global maximum as it is very easy to compute, but it has the drawback that other samples in a batch might slightly affect numerical stability. This is a first step to eliminated some of the INF/NaN issues we see in certain configurations and by no means this is a complete solutions. This needs to be revised in the future.
zhouyu5
pushed a commit
to zhouyu5/vllm-fork
that referenced
this pull request
Sep 20, 2024
This PR removes the hardcoded value used to normalize softmax in flat_pa . Current approach is to use the global maximum as it is very easy to compute, but it has the drawback that other samples in a batch might slightly affect numerical stability. This is a first step to eliminated some of the INF/NaN issues we see in certain configurations and by no means this is a complete solutions. This needs to be revised in the future.
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Outdated suggestions cannot be applied.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.
This PR removes the hardcoded value used to normalize softmax in flat_pa . Current approach is to use the global maximum as it is very easy to compute, but it has the drawback that other samples in a batch might slightly affect numerical stability.
This is a first step to eliminated some of the INF/NaN issues we see in certain configurations and by no means this is a complete solutions. This needs to be revised in the future.