Skip to content

[GPU] Add MVN eltwise fusion tests for AdaLayerNorm pattern #35027

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

Open
MDerazNasr wants to merge 3 commits into
base: master
Choose a base branch
from
Open

[GPU] Add MVN eltwise fusion tests for AdaLayerNorm pattern #35027

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c3c9247
[GPU] Add MVN eltwise fusion tests for AdaLayerNorm pattern
MDerazNasr Mar 30, 2026
df5de4b
Merge branch 'master' into mvn-ada-layer-norm-fusion-tests
p-durandin Apr 23, 2026
5858767
Merge branch 'master' into mvn-ada-layer-norm-fusion-tests
p-durandin May 4, 2026
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
88 changes: 87 additions & 1 deletion src/plugins/intel_gpu/tests/unit/fusions/mvn_fusion_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -298,5 +298,91 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, mvn_eltwise_f16, ::testing::ValuesIn(std::
mvn_test_params{ CASE_MVN_I8_6, 2, 2, 3 },
mvn_test_params{ CASE_MVN_I8_8, 3, 3, 3 },
mvn_test_params{ CASE_MVN_U8_2, 2, 2, 3 },
mvn_test_params{ CASE_MVN_F16_1, 2, 2, 3},
mvn_test_params{ CASE_MVN_F16_1, 2, 2, 3 },
}));

// AdaLayerNorm decomposes in OpenVINO IR to MVN followed by eltwise_prod(scale) and eltwise_sum(shift).
// Existing tests cover a single eltwise post-op on MVN or eltwise followed by quantize.
// These tests cover the chained scale and shift pattern without quantize.
class mvn_scale_shift : public MVNFusingTest {};
TEST_P(mvn_scale_shift, basic) {
auto p = GetParam();
create_topologies(
input_layout("input", get_input_layout(p)),
mvn("mvn", input_info("input"), p.normalize_variance, 1e-10f, false, p.reduction_axes),
data("scale_data", get_mem(get_per_channel_layout(p))),
eltwise("scale", { input_info("mvn"), input_info("scale_data") }, eltwise_mode::prod, p.default_type),
data("shift_data", get_mem(get_per_channel_layout(p))),
eltwise("shift", { input_info("scale"), input_info("shift_data") }, eltwise_mode::sum, p.default_type),
reorder("reorder_bfyx", input_info("shift"), format::bfyx, data_types::f32)
);

tolerance = (p.input_type == data_types::f32) ? 1e-5f : 0.1f;
execute(p);
}

INSTANTIATE_TEST_SUITE_P(fusings_gpu, mvn_scale_shift, ::testing::ValuesIn(std::vector<mvn_test_params>{
mvn_test_params{ CASE_MVN_F32_1, 2, 2, 4 },
mvn_test_params{ CASE_MVN_F32_2, 2, 2, 4 },
mvn_test_params{ CASE_MVN_3D_F32_1, 2, 2, 4 },
mvn_test_params{ CASE_MVN_3D_F32_2, 2, 2, 4 },
mvn_test_params{ CASE_MVN_F16_1, 2, 2, 4 },
mvn_test_params{ CASE_MVN_F16_2, 2, 2, 4 },
mvn_test_params{ CASE_MVN_3D_F16_1, 2, 2, 4 },
mvn_test_params{ CASE_MVN_3D_F16_2, 2, 2, 4 },
}));

// Same pattern as mvn_scale_shift but scale and shift are runtime inputs, not constants.
// This matches the actual AdaLayerNorm conditioning path where scale and shift are
// produced by a linear layer on the timestep embedding rather than stored as weights.
class MVNDynamicAffineTest : public MVNFusingTest {
public:
void execute(mvn_test_params& p) {
if (engine.get_device_info().supports_immad)
p.expected_fused_primitives = p.expected_fused_primitives_onednn;

auto input_prim = get_mem(get_input_layout(p));
auto scale_prim = get_mem(get_per_channel_layout(p));
auto shift_prim = get_mem(get_per_channel_layout(p));

network network_not_fused(this->engine, this->topology_non_fused, cfg_not_fused);
network network_fused(this->engine, this->topology_fused, cfg_fused);

network_fused.set_input_data("input", input_prim);
network_fused.set_input_data("scale_data", scale_prim);
network_fused.set_input_data("shift_data", shift_prim);
network_not_fused.set_input_data("input", input_prim);
network_not_fused.set_input_data("scale_data", scale_prim);
network_not_fused.set_input_data("shift_data", shift_prim);

compare(network_not_fused, network_fused, p);
}
};

class mvn_dynamic_affine : public MVNDynamicAffineTest {};
TEST_P(mvn_dynamic_affine, basic) {
auto p = GetParam();
create_topologies(
input_layout("input", get_input_layout(p)),
input_layout("scale_data", get_per_channel_layout(p)),
input_layout("shift_data", get_per_channel_layout(p)),
mvn("mvn", input_info("input"), p.normalize_variance, 1e-10f, false, p.reduction_axes),
eltwise("scale", { input_info("mvn"), input_info("scale_data") }, eltwise_mode::prod, p.default_type),
eltwise("shift", { input_info("scale"), input_info("shift_data") }, eltwise_mode::sum, p.default_type),
reorder("reorder_bfyx", input_info("shift"), format::bfyx, data_types::f32)
);

tolerance = (p.input_type == data_types::f32) ? 1e-5f : 0.1f;
execute(p);
}

INSTANTIATE_TEST_SUITE_P(fusings_gpu, mvn_dynamic_affine, ::testing::ValuesIn(std::vector<mvn_test_params>{
mvn_test_params{ CASE_MVN_F32_1, 2, 2, 4 },
mvn_test_params{ CASE_MVN_F32_2, 2, 2, 4 },
mvn_test_params{ CASE_MVN_3D_F32_1, 2, 2, 4 },
mvn_test_params{ CASE_MVN_3D_F32_2, 2, 2, 4 },
mvn_test_params{ CASE_MVN_F16_1, 2, 2, 4 },
mvn_test_params{ CASE_MVN_F16_2, 2, 2, 4 },
mvn_test_params{ CASE_MVN_3D_F16_1, 2, 2, 4 },
mvn_test_params{ CASE_MVN_3D_F16_2, 2, 2, 4 },
}));
Loading