Skip to content

[GPU] Recognize parameters as valid inputs for compressed weights #32276

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
mdvoretc-intel wants to merge 20 commits into
base: master
Choose a base branch
from
+310 −34
Open

[GPU] Recognize parameters as valid inputs for compressed weights #32276

Show file tree
Hide file tree
Changes from 11 commits
Commits
Show all changes
20 commits
Select commit Hold shift + click to select a range
f557a05
[gpu] Recognize parameters as valid inputs for compressed weights
mdvoretc-intel Oct 2, 2025
1c45696
Adjust ConvertMatMulToFullyConnectedExceptionTest_sibling_matmul
mdvoretc-intel Oct 30, 2025
a714620
Address review comments
mdvoretc-intel Nov 3, 2025
53d9074
Merge branch 'master' into param_quant_weight
mdvoretc-intel Nov 11, 2025
ea48b32
Restore convert_matmul_to_fc change after refactor
mdvoretc-intel Nov 11, 2025
29d0d89
Merge branch 'master' into param_quant_weight
mdvoretc-intel Nov 14, 2025
e32bf59
Merge branch 'master' into param_quant_weight
mdvoretc-intel Nov 21, 2025
8f8becf
Merge branch 'master' into param_quant_weight
mdvoretc-intel Nov 27, 2025
30dd0ef
Add unit tests per review comments
mdvoretc-intel Nov 28, 2025
d80ce82
Merge branch 'master' into param_quant_weight
mdvoretc-intel Nov 28, 2025
515f0d0
Remove redundant code
mdvoretc-intel Dec 1, 2025
83a0b55
Generalize immediate weight reshape
mdvoretc-intel Dec 1, 2025
d8d9e22
Add test cases
mdvoretc-intel Dec 4, 2025
6f49a0d
Skip cases that would require 4-bit transpose
mdvoretc-intel Dec 4, 2025
9831c47
Remove unused variable
mdvoretc-intel Jan 6, 2026
9f75ca4
Merge branch 'master' into param_quant_weight
mdvoretc-intel Jan 6, 2026
ae5c996
Cover the added test case
mdvoretc-intel Jan 6, 2026
9f96b7f
Remove incorrect assertions
mdvoretc-intel Jan 7, 2026
d443c08
Fix test API issue
mdvoretc-intel Jan 8, 2026
b510b91
Update logic for squeezing parameter weights
mdvoretc-intel Jan 8, 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
14 changes: 10 additions & 4 deletions src/plugins/intel_gpu/src/plugin/transformations/compressed_weights_pattern.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,18 +14,24 @@ using namespace ov::pass::pattern;
((in_ps.size() == 3 && out_ps.size() == 2) || (in_ps.size() == 4 && out_ps.size() == 3));\
};\
\
auto compressed_weights_m = wrap_type<ov::op::v0::Constant>(compressed_constant);\
auto weights_const_m = wrap_type<ov::op::v0::Constant>(compressed_constant);\
auto weights_param_m = wrap_type<ov::op::v0::Parameter>(compressed_constant);\
auto weights_param_reshape_m = wrap_type<ov::op::v1::Reshape>({weights_param_m, any_input()});\
auto compressed_weights_m = std::make_shared<ov::pass::pattern::op::Or>(OutputVector{weights_const_m, weights_param_m, weights_param_reshape_m});\
auto convert_m = wrap_type<ov::op::v0::Convert>({compressed_weights_m});\
auto weights_param_convert_m = wrap_type<ov::op::v0::Convert>({weights_param_m});\
auto weights_convert_reshape_m = wrap_type<ov::op::v1::Reshape>({weights_param_convert_m, any_input()});\
auto decompressed_weights_m = std::make_shared<ov::pass::pattern::op::Or>(OutputVector{convert_m, weights_convert_reshape_m});\
\
auto sub_const_m = wrap_type<ov::op::v0::Constant>();\
auto sub_convert_const_m = wrap_type<ov::op::v0::Convert>({sub_const_m});\
auto sub_with_convert_m = wrap_type<ov::op::v1::Subtract>({convert_m, sub_convert_const_m});\
auto sub_no_convert_m = wrap_type<ov::op::v1::Subtract>({convert_m, sub_const_m});\
auto sub_with_convert_m = wrap_type<ov::op::v1::Subtract>({decompressed_weights_m, sub_convert_const_m});\
auto sub_no_convert_m = wrap_type<ov::op::v1::Subtract>({decompressed_weights_m, sub_const_m});\
auto subtract_m = std::make_shared<ov::pass::pattern::op::Or>(OutputVector{sub_with_convert_m, sub_no_convert_m});\
\
auto mul_const_m = wrap_type<ov::op::v0::Constant>();\
auto mul_with_sub_m = wrap_type<ov::op::v1::Multiply>({subtract_m, mul_const_m});\
auto mul_no_sub_m = wrap_type<ov::op::v1::Multiply>({convert_m, mul_const_m});\
auto mul_no_sub_m = wrap_type<ov::op::v1::Multiply>({decompressed_weights_m, mul_const_m});\
auto mul_m = std::make_shared<ov::pass::pattern::op::Or>(OutputVector{mul_with_sub_m, mul_no_sub_m});\
\
auto reshape_const_m = wrap_type<ov::op::v0::Constant>();\
Expand Down
45 changes: 26 additions & 19 deletions src/plugins/intel_gpu/src/plugin/transformations/convert_fc_to_compressed.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,9 +40,8 @@ ConvertFullyConnectedToFullyConnectedCompressed::ConvertFullyConnectedToFullyCon
const auto& pattern_map = m.get_pattern_value_map();
OPENVINO_ASSERT(pattern_map.count(fully_connected_m));
OPENVINO_ASSERT(pattern_map.count(mul_const_m));
OPENVINO_ASSERT(pattern_map.count(compressed_weights_m));
OPENVINO_ASSERT(pattern_map.count(decompressed_weights_m));
OPENVINO_ASSERT(pattern_map.count(bias_m));
OPENVINO_ASSERT(pattern_map.count(convert_m));
auto fc = ov::as_type_ptr<op::FullyConnected>(pattern_map.at(fully_connected_m).get_node_shared_ptr());
if (!fc || transformation_callback(fc)) {
return false;
Expand All @@ -55,8 +54,9 @@ ConvertFullyConnectedToFullyConnectedCompressed::ConvertFullyConnectedToFullyCon
auto weight_shape = fc->get_input_shape(1);
bool is_weight_3d = (std::count_if(weight_shape.begin(), weight_shape.end(), [](size_t d) { return d > 1; }) == 3);

auto weight_ptr = ov::as_type_ptr<ov::op::v0::Constant>(pattern_map.at(compressed_weights_m).get_node_shared_ptr());
bool weight_u8 = false;
std::shared_ptr<ov::Node> weight_ptr =
pattern_map.count(weights_const_m) ? pattern_map.at(weights_const_m).get_node_shared_ptr() : pattern_map.at(weights_param_m).get_node_shared_ptr();
if (weight_ptr->get_element_type() == ov::element::u8 || weight_ptr->get_element_type() == ov::element::i8)
weight_u8 = true;

Expand Down Expand Up @@ -102,7 +102,6 @@ ConvertFullyConnectedToFullyConnectedCompressed::ConvertFullyConnectedToFullyCon
return result;
};


const ov::Output<Node>& fc_input_a = fc->input(0).get_source_output();
const auto& scale = reshape_const(pattern_map.at(mul_const_m).get_node_shared_ptr());
std::shared_ptr<ov::Node> optional_zero_point = nullptr;
Expand All @@ -112,12 +111,28 @@ ConvertFullyConnectedToFullyConnectedCompressed::ConvertFullyConnectedToFullyCon
optional_zero_point = convert_const_to_u8(reshape_const(pattern_map.at(sub_const_m).get_node_shared_ptr()));
}

std::shared_ptr<ov::Node> fc_input_b = reshape_const(pattern_map.at(compressed_weights_m).get_node_shared_ptr());
std::shared_ptr<ov::Node> fc_input_b =
pattern_map.count(weights_const_m) ? reshape_const(pattern_map.at(weights_const_m).get_node_shared_ptr())
: (pattern_map.count(weights_param_reshape_m) ? pattern_map.at(weights_param_reshape_m).get_node_shared_ptr()
: pattern_map.at(weights_param_m).get_node_shared_ptr());
std::shared_ptr<ov::Node> fc_input_scale = scale;
std::shared_ptr<ov::Node> fc_input_zp = optional_zero_point;
std::shared_ptr<ov::Node> fc_input_bias = pattern_map.at(bias_m).get_node_shared_ptr();
std::vector<std::shared_ptr<ov::Node>> result_nodes = {};

if (fc_input_b->get_output_partial_shape(0).size() != weight_shape.size()) {
OPENVINO_ASSERT(weight_shape.size() < 3);
if (has_transpose) {
OPENVINO_ASSERT(weight_shape.size() == 2);
std::swap(weight_shape[0], weight_shape[1]);
}
std::shared_ptr<ov::Node> weight_shape_const =
std::make_shared<ov::op::v0::Constant>(ov::element::i32, ov::Shape{weight_shape.size()}, weight_shape);
fc_input_b = std::make_shared<ov::op::v1::Reshape>(fc_input_b, weight_shape_const, false);
result_nodes.push_back(weight_shape_const);
result_nodes.push_back(fc_input_b);
}

if (has_transpose) {
const auto& transpose = pattern_map.at(transpose_m).get_node_shared_ptr();
std::shared_ptr<ov::Node> transpose_const = pattern_map.at(transpose_const_m).get_node_shared_ptr();
Expand All @@ -128,16 +143,16 @@ ConvertFullyConnectedToFullyConnectedCompressed::ConvertFullyConnectedToFullyCon
transpose_const = std::make_shared<ov::op::v0::Constant>(ov::element::i32, ov::Shape{new_order.size()}, new_order);
}

fc_input_b = transpose->clone_with_new_inputs({ fc_input_b->output(0), transpose_const });
fc_input_b = transpose->clone_with_new_inputs({fc_input_b->output(0), transpose_const});
result_nodes.push_back(fc_input_b);

if (ov::shape_size(scale->output(0).get_shape()) > 1) {
fc_input_scale = transpose->clone_with_new_inputs({ scale->output(0), transpose_const });
fc_input_scale = transpose->clone_with_new_inputs({scale->output(0), transpose_const});
result_nodes.push_back(fc_input_scale);
}

if (with_zero_point && ov::shape_size(optional_zero_point->output(0).get_shape()) > 1) {
fc_input_zp = transpose->clone_with_new_inputs({ optional_zero_point->output(0), transpose_const });
fc_input_zp = transpose->clone_with_new_inputs({optional_zero_point->output(0), transpose_const});
result_nodes.push_back(fc_input_zp);
}
}
Expand All @@ -149,18 +164,10 @@ ConvertFullyConnectedToFullyConnectedCompressed::ConvertFullyConnectedToFullyCon

std::shared_ptr<ov::Node> new_fc = nullptr;
if (with_zero_point) {
new_fc = std::make_shared<op::FullyConnectedCompressed>(fc_input_a,
fc_input_b,
fc_input_bias,
fc_input_scale,
fc_input_zp,
fc->get_output_type());
new_fc =
std::make_shared<op::FullyConnectedCompressed>(fc_input_a, fc_input_b, fc_input_bias, fc_input_scale, fc_input_zp, fc->get_output_type());
} else {
new_fc = std::make_shared<op::FullyConnectedCompressed>(fc_input_a,
fc_input_b,
fc_input_bias,
fc_input_scale,
fc->get_output_type());
new_fc = std::make_shared<op::FullyConnectedCompressed>(fc_input_a, fc_input_b, fc_input_bias, fc_input_scale, fc->get_output_type());
}

result_nodes.push_back(new_fc);
Expand Down
2 changes: 1 addition & 1 deletion src/plugins/intel_gpu/src/plugin/transformations/convert_matmul_to_fc.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,7 @@ ConvertMatMulToFullyConnected::ConvertMatMulToFullyConnected(bool supports_immad
};
auto weights_path = [&static_rank_gt_1](const ov::Output<ov::Node>& output) {
const auto& pshape = output.get_partial_shape();
return ov::op::util::is_on_path<ov::op::v0::Constant>(output) &&
return ov::op::util::is_on_path<ov::op::v0::Constant, ov::op::v0::Parameter>(output) &&
static_rank_gt_1(output) &&
pshape.is_static();
};
Expand Down
141 changes: 141 additions & 0 deletions src/plugins/intel_gpu/tests/unit/transformations/convert_fc_to_compressed_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -463,6 +463,147 @@ TEST_F(TransformationTestsF, ConvertFCToCompressed10) {
}
}

TEST_F(TransformationTestsF, ConvertFCToCompressed11) {
Copy link
Contributor

@Lyamin-Roman Lyamin-Roman Nov 28, 2025

Choose a reason for hiding this comment

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

[random spot] Please add functional accuracy tests
It looks like you can extend an existing test with additional parameters
src/plugins/intel_gpu/tests/functional/subgraph_tests/dynamic/matmul_weights_decompression.cpp

Copy link
Contributor Author

@mdvoretc-intel mdvoretc-intel Dec 1, 2025

Choose a reason for hiding this comment

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

Attempted this. The tests have a singular parameter for input precision which prevents proper creation of weight parameters in compressed types.

Copy link
Contributor Author

@mdvoretc-intel mdvoretc-intel Dec 4, 2025

Choose a reason for hiding this comment

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

Tests added with a configure_model() override to ensure weight parameters are provided in appropriate type. Tests currently fail due to the lack of u4 transposition support, WIP to fix.

{
auto input1 = std::make_shared<ov::op::v0::Parameter>(ov::element::f16, ov::PartialShape{-1, 16});
auto weights_param = std::make_shared<ov::op::v0::Parameter>(ov::element::u4, ov::Shape{32, 16});
auto convert = std::make_shared<ov::op::v0::Convert>(weights_param, ov::element::f16);
auto scale_const = ov::op::v0::Constant::create(ov::element::f16, ov::Shape{32, 1}, {1});
auto scale = std::make_shared<ov::op::v1::Multiply>(convert, scale_const);
auto no_bias = std::make_shared<ov::intel_gpu::op::Placeholder>();

auto fc = std::make_shared<ov::intel_gpu::op::FullyConnected>(input1, scale, no_bias);

model = std::make_shared<ov::Model>(ov::OutputVector{fc}, ov::ParameterVector{input1, weights_param});
manager.register_pass<ConvertFullyConnectedToFullyConnectedCompressed>();
}
{
auto input1 = std::make_shared<ov::op::v0::Parameter>(ov::element::f16, ov::PartialShape{-1, 16});
auto weights_param = std::make_shared<ov::op::v0::Parameter>(ov::element::u4, ov::Shape{32, 16});
auto no_bias = std::make_shared<ov::intel_gpu::op::Placeholder>();
auto scale_const = ov::op::v0::Constant::create(ov::element::f16, ov::Shape{32, 1}, {1});

auto fc_compressed = std::make_shared<ov::intel_gpu::op::FullyConnectedCompressed>(input1, weights_param, no_bias, scale_const);

model_ref = std::make_shared<ov::Model>(ov::OutputVector{fc_compressed}, ov::ParameterVector{input1, weights_param});
}
}

TEST_F(TransformationTestsF, ConvertFCToCompressed12) {
{
auto input1 = std::make_shared<ov::op::v0::Parameter>(ov::element::f16, ov::PartialShape{-1, 16});
auto weights_param = std::make_shared<ov::op::v0::Parameter>(ov::element::u4, ov::Shape{4, 4, 32});
auto reshape_const = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {16, 32});
auto reshape = std::make_shared<ov::op::v1::Reshape>(weights_param, reshape_const, false);
auto convert = std::make_shared<ov::op::v0::Convert>(reshape, ov::element::f16);
auto scale_const = ov::op::v0::Constant::create(ov::element::f16, ov::Shape{1, 32}, {1});
auto scale = std::make_shared<ov::op::v1::Multiply>(convert, scale_const);
auto transpose_const = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {1, 0});
auto transpose = std::make_shared<ov::op::v1::Transpose>(scale, transpose_const);
auto no_bias = std::make_shared<ov::intel_gpu::op::Placeholder>();

auto fc = std::make_shared<ov::intel_gpu::op::FullyConnected>(input1, transpose, no_bias);

model = std::make_shared<ov::Model>(ov::OutputVector{fc}, ov::ParameterVector{input1, weights_param});
manager.register_pass<ConvertFullyConnectedToFullyConnectedCompressed>();
}
{
auto input1 = std::make_shared<ov::op::v0::Parameter>(ov::element::f16, ov::PartialShape{-1, 16});
auto weights_param = std::make_shared<ov::op::v0::Parameter>(ov::element::u4, ov::Shape{4, 4, 32});
auto reshape_const = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {16, 32});
auto reshape = std::make_shared<ov::op::v1::Reshape>(weights_param, reshape_const, false);
auto transpose_weights_param = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {1, 0});
auto transpose_weights = std::make_shared<ov::op::v1::Transpose>(reshape, transpose_weights_param);
auto no_bias = std::make_shared<ov::intel_gpu::op::Placeholder>();
auto scale_const = ov::op::v0::Constant::create(ov::element::f16, ov::Shape{1, 32}, {1});
auto transpose_scale_const = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {1, 0});
auto transpose_scale = std::make_shared<ov::op::v1::Transpose>(scale_const, transpose_scale_const);

auto fc_compressed = std::make_shared<ov::intel_gpu::op::FullyConnectedCompressed>(input1, transpose_weights, no_bias, transpose_scale);

model_ref = std::make_shared<ov::Model>(ov::OutputVector{fc_compressed}, ov::ParameterVector{input1, weights_param});
}
}

TEST_F(TransformationTestsF, ConvertFCToCompressed13) {
{
auto input1 = std::make_shared<ov::op::v0::Parameter>(ov::element::f16, ov::PartialShape{-1, 16});
auto weights_param = std::make_shared<ov::op::v0::Parameter>(ov::element::u4, ov::Shape{4, 4, 32});
auto convert = std::make_shared<ov::op::v0::Convert>(weights_param, ov::element::f16);
auto reshape_const = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {16, 32});
auto reshape = std::make_shared<ov::op::v1::Reshape>(convert, reshape_const, false);
auto scale_const = ov::op::v0::Constant::create(ov::element::f16, ov::Shape{1, 32}, {1});
auto scale = std::make_shared<ov::op::v1::Multiply>(reshape, scale_const);
auto transpose_const = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {1, 0});
auto transpose = std::make_shared<ov::op::v1::Transpose>(scale, transpose_const);
auto no_bias = std::make_shared<ov::intel_gpu::op::Placeholder>();

auto fc = std::make_shared<ov::intel_gpu::op::FullyConnected>(input1, transpose, no_bias);

model = std::make_shared<ov::Model>(ov::OutputVector{fc}, ov::ParameterVector{input1, weights_param});
manager.register_pass<ConvertFullyConnectedToFullyConnectedCompressed>();
}
{
auto input1 = std::make_shared<ov::op::v0::Parameter>(ov::element::f16, ov::PartialShape{-1, 16});
auto weights_param = std::make_shared<ov::op::v0::Parameter>(ov::element::u4, ov::Shape{4, 4, 32});
auto reshape_const = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {16, 32});
auto reshape = std::make_shared<ov::op::v1::Reshape>(weights_param, reshape_const, false);
auto transpose_weights_param = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {1, 0});
auto transpose_weights = std::make_shared<ov::op::v1::Transpose>(reshape, transpose_weights_param);
auto no_bias = std::make_shared<ov::intel_gpu::op::Placeholder>();
auto scale_const = ov::op::v0::Constant::create(ov::element::f16, ov::Shape{1, 32}, {1});
auto transpose_scale_const = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {1, 0});
auto transpose_scale = std::make_shared<ov::op::v1::Transpose>(scale_const, transpose_scale_const);

auto fc_compressed = std::make_shared<ov::intel_gpu::op::FullyConnectedCompressed>(input1, transpose_weights, no_bias, transpose_scale);

model_ref = std::make_shared<ov::Model>(ov::OutputVector{fc_compressed}, ov::ParameterVector{input1, weights_param});
}
}

TEST_F(TransformationTestsF, ConvertFCToCompressed14) {
{
auto input1 = std::make_shared<ov::op::v0::Parameter>(ov::element::f16, ov::PartialShape{-1, 16});
auto weights_param = std::make_shared<ov::op::v0::Parameter>(ov::element::u4, ov::Shape{4, 4, 32});
auto convert = std::make_shared<ov::op::v0::Convert>(weights_param, ov::element::f16);
auto zp_const = ov::op::v0::Constant::create(ov::element::u8, ov::Shape{4, 1, 32}, {1});
auto zp_convert = std::make_shared<ov::op::v0::Convert>(zp_const, ov::element::f16);
auto sub = std::make_shared<ov::op::v1::Subtract>(convert, zp_convert);
auto scale_const = ov::op::v0::Constant::create(ov::element::f16, ov::Shape{4, 1, 32}, {1});
auto scale = std::make_shared<ov::op::v1::Multiply>(sub, scale_const);
auto reshape_const = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {16, 32});
auto reshape = std::make_shared<ov::op::v1::Reshape>(scale, reshape_const, false);
auto transpose_const = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {1, 0});
auto transpose = std::make_shared<ov::op::v1::Transpose>(reshape, transpose_const);
auto no_bias = std::make_shared<ov::intel_gpu::op::Placeholder>();

auto fc = std::make_shared<ov::intel_gpu::op::FullyConnected>(input1, transpose, no_bias);

model = std::make_shared<ov::Model>(ov::OutputVector{fc}, ov::ParameterVector{input1, weights_param});
manager.register_pass<ConvertFullyConnectedToFullyConnectedCompressed>();
}
{
auto input1 = std::make_shared<ov::op::v0::Parameter>(ov::element::f16, ov::PartialShape{-1, 16});
auto weights_param = std::make_shared<ov::op::v0::Parameter>(ov::element::u4, ov::Shape{4, 4, 32});
auto reshape_const = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {16, 32});
auto reshape = std::make_shared<ov::op::v1::Reshape>(weights_param, reshape_const, false);
auto transpose_weights_param = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {1, 0});
auto transpose_weights = std::make_shared<ov::op::v1::Transpose>(reshape, transpose_weights_param);
auto no_bias = std::make_shared<ov::intel_gpu::op::Placeholder>();
auto scale_const = ov::op::v0::Constant::create(ov::element::f16, ov::Shape{4, 32}, {1});
auto transpose_scale_const = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {1, 0});
auto transpose_scale = std::make_shared<ov::op::v1::Transpose>(scale_const, transpose_scale_const);
auto zp_const = ov::op::v0::Constant::create(ov::element::u8, ov::Shape{4, 32}, {1});
auto zp_convert = std::make_shared<ov::op::v0::Convert>(zp_const, ov::element::f16);
auto transpose_zp_const = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{2}, {1, 0});
auto transpose_zp = std::make_shared<ov::op::v1::Transpose>(zp_const, transpose_zp_const);

auto fc_compressed = std::make_shared<ov::intel_gpu::op::FullyConnectedCompressed>(input1, transpose_weights, no_bias, transpose_scale, transpose_zp);

model_ref = std::make_shared<ov::Model>(ov::OutputVector{fc_compressed}, ov::ParameterVector{input1, weights_param});
}
}

} // namespace intel_gpu
} // namespace test
} // namespace ov
Loading