Skip to content

[TRANSFORMATIONS] Fix sliding window handling for gpt-oss in SDPAToPA #32939

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

Merged
merged 1 commit into from
Nov 20, 2025
Merged

[TRANSFORMATIONS] Fix sliding window handling for gpt-oss in SDPAToPA #32939

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 3 additions & 6 deletions .../transformations/src/transformations/sdpa_to_paged_attention/state_management_pattern.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -207,22 +207,19 @@ static std::tuple<std::shared_ptr<ov::Node>, std::shared_ptr<ov::Node>> gpt_oss_
auto q_idx = pattern::any_input();
auto kv_idx = pattern::any_input();

auto kv_idx_opt_conv_0 = pattern::optional<v0::Convert>();
auto kv_idx_opt_conv_1 = pattern::optional<v0::Convert>(kv_idx_opt_conv_0);
auto less_eq = pattern::wrap_type<v1::LessEqual>({q_idx, kv_idx_opt_conv_1});
auto kv_idx_opt_conv = pattern::optional<v0::Convert>(kv_idx);

auto offset = wrap_type<v0::Constant>();

auto add = wrap_type<v1::Add>({q_idx, offset});
auto opt_conv_2 = pattern::optional<v0::Convert>(add);
auto greater = pattern::wrap_type<v1::Greater>({kv_idx_opt_conv_1, opt_conv_2});
auto greater = pattern::wrap_type<v1::Greater>({kv_idx_opt_conv, add});
auto bitwise_and = pattern::wrap_type<v13::BitwiseAnd>({any_input(), greater});
auto bitwise_and_1 = pattern::wrap_type<v13::BitwiseAnd>({bitwise_and, any_input()});
auto bitwise_and_2 = pattern::wrap_type<v13::BitwiseAnd>({any_input(), bitwise_and_1});
auto bitwise_and_3 = pattern::wrap_type<v13::BitwiseAnd>({bitwise_and_2, any_input()});
auto broadcast = pattern::wrap_type<v3::Broadcast>({bitwise_and_3, any_input()});
auto select = pattern::wrap_type<v1::Select>({broadcast, any_input(), any_input()});
auto mask = pattern::wrap_type<v1::StridedSlice>({select, any_input(), any_input(), any_input()});
auto mask = pattern::wrap_type<v8::Slice>({select, any_input(), any_input(), any_input(), any_input()});

return {mask, offset};
}
Expand Down
7 changes: 5 additions & 2 deletions src/common/transformations/tests/op_conversions/sdpa_to_paged_attention_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2646,8 +2646,11 @@ TEST_F(SDPAToPATest, SDPAToPA_gpt_oss_General) {
}),
MOCK_VALUE);

auto scale = v0::Constant::create(element::f32, {}, {0.125000f});
auto sliding_window = v0::Constant::create(element::i32, {}, {0});
auto sliding_window_neg = makeConst(element::f32, ov::Shape({1, 1, 1, 1}), {-128.0f});
auto Squeeze2 = makeOP<v15::Squeeze>({sliding_window_neg}, {{"allow_axis_skip", false}});
auto Convert16 = makeOP<v0::Convert>({Squeeze2}, {{"destination_type", "i32"}});
auto sliding_window = makeOP<v1::Multiply>({Convert16, -1}, {{"auto_broadcast", "numpy"}});
auto scale = v0::Constant::create(element::f32, {}, {0.1250f});
auto alibi_slopes_stub = v0::Constant::create(element::f32, Shape{0}, {});
auto PagedAttentionExtension =
std::make_shared<ov::op::PagedAttentionExtension>(OutputVector{Reshape1,
Expand Down
Loading