-
Notifications
You must be signed in to change notification settings - Fork 12.2k
ggml: Add epsilon as a parameter for group_norm #8818
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.
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
Conversation
return ggml_group_norm_impl(ctx, a, n_groups, false); | ||
int n_groups, | ||
float eps) { | ||
return ggml_group_norm_impl(ctx, a, n_groups, eps, false); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
(Side question) Is ggml_group_norm
equivalent to ggml_norm
with ggml_reshape
to group rows together beforehand and split them back afterwards?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I guess yes?
6b20811
to
3603bb8
Compare
Signed-off-by: Molly Sophia <[email protected]>
3603bb8
to
2bdcb7b
Compare
Hmm... It seems that one of the ci checks is failing |
Seems like a timeout. I've seen that test fail before for the same reason, and usually it passes when retried. The timeout for the slot erase in the server test suite is likely too short when all is run at the same time. Not sure why it only times out in the Release build and not also in the slower builds with sanitizers. (I've restarted the failing check) (EDIT: still fails. Hmm.) |
Yeah, that test started failing recently for some reason - definitely not related to this PR. I tried to figure out where to increase the timeout for slot erasure, but could figure out. Would be nice if fix this eventually |
Signed-off-by: Molly Sophia <[email protected]>
Hi! I'm working on adding RWKV implementation in llama.cpp. The graph needs a custom eps value for group_norm.
This PR adds epsilon as a parameter for group_norm operators.