-
Notifications
You must be signed in to change notification settings - Fork 6.4k
Use SDP on BF16 in GPU/HPU migration #12310
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
feb4827
to
de8e8ef
Compare
Signed-off-by: Daniel Socek <[email protected]>
de8e8ef
to
8e73b58
Compare
LGTM |
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.
LGTM
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
@dsocek It seems there is a formatting error: https://github.com/huggingface/diffusers/actions/runs/17622299913/job/50216966858?pr=12310 |
@regisss fixed the minor formatting issue |
Signed-off-by: Daniel Socek <[email protected]>
77b888f
to
b05259a
Compare
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.
thanks!
What does this PR do?
The latest Gaudi stack now defaults to eager mode, where HPU/GPU Migration uses the default SDP kernel in FP32 precision. Previously, the default lazy mode used SDP in BF16 precision.
To improve performance on HPU devices, this PR changes the default to SDP with BF16 precision when running on HPU, provided that the dtype is BF16 or unset.