-
Notifications
You must be signed in to change notification settings - Fork 2.6k
[GPU] GroupNormalization opt kernels #25417
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
236ed6b
to
bd8702a
Compare
bd8702a
to
98bd39c
Compare
@@ -294,6 +296,8 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) { | |||
return !is_type<ov::op::v0::MatMul>(next_node); | |||
}); | |||
|
|||
manager.register_pass<ov::intel_gpu::GroupNormComposition>(); |
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.
GroupNormDecomposition is being disabled, but still need Composition too?
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.
Current OV stable diffusion models does not have group_normalization
layers. They have decomposed group_normalization layers instead. So we need a composition pass to apply group_normalization
in the current OV SD models.
namespace ov { | ||
namespace intel_gpu { | ||
|
||
GroupNormComposition::GroupNormComposition() { |
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.
- That can be a common pass.
- Are new IRs generated w/ decomposed GroupNorm?
- Rename to
GroupNormFusion
to be aligned with other passes
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.
- Okay, I'll do it in a separate PR.
- Yes, new IRs have decomposed GroupNorm layers.
- Okay, I'll apply it too.
### Details: - This PR adds two group_normalization opt kernels, group_normalization_bfyx and group_normalization_b_fs_yx_fsv16. ### Tickets: - 144094
Details:
Tickets: