Skip to content

accuracy: Q4 matrix multiply error is very bad for small K #998

New issue

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

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

Already on GitHub? # to your account

Closed
jon-chuang opened this issue Apr 15, 2023 · 1 comment
Closed

accuracy: Q4 matrix multiply error is very bad for small K #998

jon-chuang opened this issue Apr 15, 2023 · 1 comment
Labels

Comments

@jon-chuang
Copy link
Contributor

jon-chuang commented Apr 15, 2023

In benchmark/benchmark-q4_0-matmult.c:

Set sizey=sizez=N,sizex=K

For K=128,N=2, the deviation is expected 1020.00, got 1280.00
For K=128,N=32, the deviation is expected 262144.00, got 508160.03 
For K=64,N=32 the deviation is expected 131072.00, got 476928.03 
For K=32,N=32 the deviation is expected 65536.00, got 418560.03 

This is rather worrying, although it does not practically affect any known use-case.

@jon-chuang jon-chuang changed the title accuracy: Q4 matrix multiply is very bad for small K accuracy: Q4 matrix multiply error is very bad for small K Apr 15, 2023
@github-actions github-actions bot added the stale label Mar 25, 2024
Copy link
Contributor

github-actions bot commented Apr 9, 2024

This issue was closed because it has been inactive for 14 days since being marked as stale.

# for free to join this conversation on GitHub. Already have an account? # to comment
Labels
Projects
None yet
Development

No branches or pull requests

1 participant