Skip to content

[GPU] Add Softmax/Eltwise Dropout attr support #4501

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
Merged
h-sadia merged 6 commits into from
Jan 8, 2026
Merged

[GPU] Add Softmax/Eltwise Dropout attr support #4501

h-sadia merged 6 commits into from
Jan 8, 2026

Conversation

@h-sadia
Copy link
Contributor

@h-sadia h-sadia commented Jan 7, 2026 •
edited
Loading

This PR is part 2 of adding dropout support for SDPA. This PR covers support for simple softmax kernel and ref eltwise kernel
Part 1 was here: #4494

@h-sadia h-sadia requested review from a team as code owners January 7, 2026 05:33
@github-actions github-actions bot added platform:gpu-intel Codeowner: @oneapi-src/onednn-gpu-intel component:tests Codeowner: @oneapi-src/onednn-arch labels Jan 7, 2026
@h-sadia h-sadia force-pushed the hsadia/dropout_eltwise_sftmx branch from c2451bf to eb011f5 Compare January 7, 2026 06:10
@h-sadia h-sadia requested a review from a team as a code owner January 7, 2026 06:10
@h-sadia h-sadia force-pushed the hsadia/dropout_eltwise_sftmx branch from eb011f5 to a638f8c Compare January 7, 2026 06:11
@h-sadia h-sadia changed the title [WIP][GPU] Add Softmax Dropout attr support [GPU] Add Softmax Dropout attr support Jan 7, 2026
@h-sadia h-sadia force-pushed the hsadia/dropout_eltwise_sftmx branch from a638f8c to 85b4869 Compare January 7, 2026 06:17
@h-sadia
Copy link
Contributor Author

h-sadia commented Jan 7, 2026

make test
set test_scope=NIGHTLY
disable test_device_cpu
disable benchdnn_all
enable benchdnn_softmax
enable arch_gpu_xe-hpc
enable arch_gpu_xe-hpg-atsm
enable arch_gpu_xe-hpg-dg2
enable arch_gpu_xe-lp
enable arch_gpu_xe-lpg
enable arch_gpu_xe-lpg+
enable arch_gpu_xe2-hpg-bmg
enable arch_gpu_xe2-lpg
enable arch_gpu_xe3-lpg

@h-sadia h-sadia force-pushed the hsadia/dropout_eltwise_sftmx branch 3 times, most recently from 403c679 to 6bc2241 Compare January 7, 2026 21:51
@h-sadia h-sadia changed the title [GPU] Add Softmax Dropout attr support [GPU] Add Softmax/Eltwise Dropout attr support Jan 7, 2026
@h-sadia h-sadia force-pushed the hsadia/dropout_eltwise_sftmx branch 3 times, most recently from 3e335af to 3aa6cba Compare January 7, 2026 22:05
@h-sadia
Copy link
Contributor Author

h-sadia commented Jan 7, 2026

make test
set test_scope=NIGHTLY
disable test_device_cpu
disable benchdnn_all
enable benchdnn_softmax
enable benchdnn_eltwise
enable arch_gpu_xe-hpc
enable arch_gpu_xe-hpg-atsm
enable arch_gpu_xe-hpg-dg2
enable arch_gpu_xe-lp
enable arch_gpu_xe-lpg
enable arch_gpu_xe-lpg+
enable arch_gpu_xe2-hpg-bmg
enable arch_gpu_xe2-lpg
enable arch_gpu_xe3-lpg

@h-sadia h-sadia force-pushed the hsadia/dropout_eltwise_sftmx branch from 3aa6cba to 45a485b Compare January 8, 2026 01:09
@h-sadia
Copy link
Contributor Author

h-sadia commented Jan 8, 2026

make test
set test_scope=NIGHTLY
disable test_device_cpu
disable benchdnn_all
enable benchdnn_softmax
enable benchdnn_eltwise
enable arch_gpu_xe-hpc

@h-sadia
Copy link
Contributor Author

h-sadia commented Jan 8, 2026

make test
set test_scope=NIGHTLY
disable test_device_cpu
disable benchdnn_all
enable benchdnn_softmax
enable benchdnn_eltwise
enable benchdnn_matmul

kealan-barbieri
kealan-barbieri reviewed Jan 8, 2026
View reviewed changes
src/gpu/intel/matmul/ref.hpp
status_t init(impl::engine_t *engine) override {
compute::kernel_ctx_t kernel_ctx;

bool with_seed_s64
Copy link
Contributor

@kealan-barbieri kealan-barbieri Jan 8, 2026

Choose a reason for hiding this comment

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

what is the status of non-s64 seed? it is still supported via API but not through benchdnn?

Copy link
Contributor Author

@h-sadia h-sadia Jan 8, 2026 •
edited
Loading

Choose a reason for hiding this comment

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

Yes, thats true. the seed was previously s32 in the API.

Copy link
Contributor

@kealan-barbieri kealan-barbieri Jan 8, 2026

Choose a reason for hiding this comment

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

Seems like something that should be enabled through benchdnn so we can test full range of supported features.

Copy link
Contributor

@dzarukin dzarukin Jan 8, 2026

Choose a reason for hiding this comment

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

It was a deliberate choice not to test s32 in favor of s64. It's highly unlikely anybody will use this functionality besides PyTorch which has a requirement for int64 seed.

Copy link
Contributor Author

@h-sadia h-sadia Jan 8, 2026

Choose a reason for hiding this comment

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

keeping it as a keep sake for a while @kealan-barbieri ;)

dzarukin
dzarukin approved these changes Jan 8, 2026
View reviewed changes
@h-sadia h-sadia force-pushed the hsadia/dropout_eltwise_sftmx branch from cf0e687 to 4e652b8 Compare January 8, 2026 21:38
@h-sadia h-sadia force-pushed the hsadia/dropout_eltwise_sftmx branch from 4e652b8 to d1594ed Compare January 8, 2026 21:46
@h-sadia h-sadia merged commit 042355e into main Jan 8, 2026
16 of 22 checks passed
@h-sadia h-sadia deleted the hsadia/dropout_eltwise_sftmx branch January 8, 2026 21:54
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@kealan-barbieri kealan-barbieri kealan-barbieri approved these changes

@umar456 umar456 umar456 approved these changes

@dzarukin dzarukin dzarukin approved these changes

Assignees

No one assigned

Labels

component:tests Codeowner: @oneapi-src/onednn-arch platform:gpu-intel Codeowner: @oneapi-src/onednn-gpu-intel

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@h-sadia @umar456 @dzarukin @kealan-barbieri