Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

llama:use F32 precision in GLM4 attention and no FA #9130

Merged
merged 1 commit into from
Aug 23, 2024
Merged

llama:use F32 precision in GLM4 attention and no FA #9130

merged 1 commit into from
Aug 23, 2024

Conversation

piDack
Copy link
Contributor

@piDack piDack commented Aug 22, 2024 •
edited
Loading

There have been few reports of glm4 models generating "GGGG" when FA is disabled. This should it.
I used the following command line, and after modification, there was no GGG in output

#!/bin/bashit.
../build/bin/llama-cli -m "/mnt/edge/yhl/model/glm4-9b-chat-Q4_K_M.gguf" -c 8192 -ngl 41 -np 2 -p "hello how are you?"

before
截屏2024-08-22 16 59 29

after
截屏2024-08-22 17 01 55

@piDack piDack changed the title fix glm GGG err Fix glm4 GGG err Aug 22, 2024
@piDack piDack changed the title Fix glm4 GGG err llama:use F32 precision in GLM4 attention and no FA Aug 22, 2024
ggerganov
ggerganov approved these changes Aug 23, 2024
View reviewed changes
Copy link
Owner

@ggerganov ggerganov left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I guess we should switch to F32 precision by default for the KQ multiplication

@ggerganov ggerganov merged commit a07c32e into ggerganov:master Aug 23, 2024
52 checks passed
@piDack
Copy link
Contributor Author

piDack commented Aug 23, 2024

I guess we should switch to F32 precision by default for the KQ multiplication

Good Idea

@cosmic-snow cosmic-snow mentioned this pull request Aug 25, 2024
Open
@wszgrcy wszgrcy mentioned this pull request Sep 2, 2024
Closed
@ThiloteE
Copy link

I guess this could be classified as a follow-up to #8031

arthw pushed a commit to arthw/llama.cpp that referenced this pull request Nov 15, 2024
@piDack @arthw
llama : use F32 precision in GLM4 attention and no FA (ggerganov#9130)
5ff6c1e
arthw pushed a commit to arthw/llama.cpp that referenced this pull request Nov 18, 2024
@piDack @arthw
llama : use F32 precision in GLM4 attention and no FA (ggerganov#9130)
c68b570
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@ggerganov ggerganov ggerganov approved these changes

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
3 participants
@piDack @ThiloteE @ggerganov