[LPT] QDQ stripping

src/common/low_precision_transformations/include/low_precision/qdq_stripping.hpp

@@ -0,0 +1,31 @@
+// Copyright (C) 2018-2025 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <memory>
+#include <set>
+
+#include "lpt_visibility.hpp"
+#include "openvino/pass/matcher_pass.hpp"
+#include "quantization_details.hpp"
+
+namespace ov {
+namespace pass {
+namespace low_precision {
+
+/**
+ * @ingroup ov_transformation_common_api
+ * @brief FQStrippingTransformation strips FakeQuantize operations with specified levels
+ * by replacing them with Clamp operations.
+ */
+class LP_TRANSFORMATIONS_API FQStrippingTransformation : public ov::pass::MatcherPass {
+public:
+    OPENVINO_RTTI("FQStrippingTransformation", "0", MatcherPass);
+    FQStrippingTransformation(const std::set<size_t>& levels_to_strip, bool replace_with_clamp);
+};
+
+} // namespace low_precision
+} // namespace pass
+} // namespace ov

src/common/low_precision_transformations/src/low_precision.cpp

@@ -323,7 +323,7 @@ bool LowPrecision::isFunctionQuantized(const std::shared_ptr<const ov::Model>& m
         } else if (const auto multiSubGraph = ov::as_type_ptr<ov::op::util::MultiSubGraphOp>(node)) {
             // Look inside subraph operations, such as TensorIterator, Loop, If, etc
             for (size_t i = 0; i < multiSubGraph->get_internal_subgraphs_size(); i++) {
-                if (isFunctionQuantized(multiSubGraph->get_function(i))) {
+                if (isFunctionQuantized(multiSubGraph->get_function(i), supported_levels, check_fake_convert)) {
                     return true;
                 }
             }

src/common/low_precision_transformations/src/qdq_stripping.cpp

@@ -0,0 +1,92 @@
+// Copyright (C) 2018-2025 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "low_precision/qdq_stripping.hpp"
+
+#include <memory>
+
+#include "itt.hpp"
+#include "low_precision/common/ie_lpt_exception.hpp"
+#include "low_precision/lpt_itt.hpp"
+#include "low_precision/network_helper.hpp"
+#include "openvino/core/except.hpp"
+#include "openvino/core/type.hpp"
+#include "openvino/op/clamp.hpp"
+#include "openvino/op/constant.hpp"
+#include "openvino/op/equal.hpp"
+#include "openvino/op/fake_quantize.hpp"
+#include "openvino/pass/pattern/op/wrap_type.hpp"
+#include "openvino/util/log.hpp"
+#include "transformations/utils/utils.hpp"
+
+namespace ov {
+namespace pass {
+namespace low_precision {
+
+FQStrippingTransformation::FQStrippingTransformation(const std::set<size_t>& levels_to_strip, bool replace_with_clamp) {
+    MATCHER_SCOPE(FQStrippingTransformation);
+    auto is_scalar = [](const Output<Node>& output) -> bool {
+        return ov::shape_size(output.get_shape()) == 1;
+    };
+    auto input_low_m = pattern::wrap_type<ov::op::v0::Constant>(is_scalar);
+    auto input_high_m = pattern::wrap_type<ov::op::v0::Constant>(is_scalar);
+    auto output_low_m = pattern::wrap_type<ov::op::v0::Constant>(is_scalar);
+    auto output_high_m = pattern::wrap_type<ov::op::v0::Constant>(is_scalar);
+    auto fq_m = pattern::wrap_type<ov::op::v0::FakeQuantize>(
+        {pattern::any_input(), input_low_m, input_high_m, output_low_m, output_high_m});
+
+    ov::graph_rewrite_callback callback = [OV_CAPTURE_CPY_AND_THIS](pattern::Matcher& m) {
+        const auto& pattern_map = m.get_pattern_value_map();
+        auto node = ov::as_type_ptr<ov::op::v0::FakeQuantize>(pattern_map.at(fq_m).get_node_shared_ptr());
+        if (!node) {
+            return false;
+        }
+
+        const size_t levels = node->get_levels();
+        if (!levels_to_strip.count(levels)) {
+            return false;
+        }
+
+        auto input = node->get_input_node_shared_ptr(0);
+        auto input_low = ov::as_type_ptr<ov::op::v0::Constant>(pattern_map.at(input_low_m).get_node_shared_ptr());
+        auto input_high = ov::as_type_ptr<ov::op::v0::Constant>(pattern_map.at(input_high_m).get_node_shared_ptr());
+        auto output_low = ov::as_type_ptr<ov::op::v0::Constant>(pattern_map.at(output_low_m).get_node_shared_ptr());
+        auto output_high = ov::as_type_ptr<ov::op::v0::Constant>(pattern_map.at(output_high_m).get_node_shared_ptr());
+
+        if (!input_low || !input_high || !output_low || !output_high) {
+            return false;
+        }
+        auto constants_are_equal = [](const std::shared_ptr<ov::op::v0::Constant>& lhs,
+                                      const std::shared_ptr<ov::op::v0::Constant>& rhs) {
+            auto equal =
+                ov::as_type_ptr<ov::op::v0::Constant>(ov::op::util::make_try_fold<ov::op::v1::Equal>(lhs, rhs));
+            OPENVINO_ASSERT(equal && ov::shape_size(equal->get_shape()) == 1,
+                            "constants_are_equal expects scalar constant as a comparison result");
+            return equal->get_vector<bool>()[0] == true;
+        };
+
+        if (!constants_are_equal(input_low, output_low) || !constants_are_equal(input_high, output_high)) {
+            return false;
+        }
+
+        bool res = false;
+        if (replace_with_clamp) {
+            auto clamp = std::make_shared<ov::op::v0::Clamp>(input->output(0),
+                                                             output_low->cast_vector<double>()[0],
+                                                             output_high->cast_vector<double>()[0]);
+            res = replace_node_update_name(node, clamp);
+        } else {
+            res = replace_output_update_name(node->output(0), node->input_value(0));
+        }
+        OPENVINO_ASSERT(res, "FQ stripping failed");
+        return res;
+    };
+
+    auto m = std::make_shared<ov::pass::pattern::Matcher>(fq_m, matcher_name);
+    this->register_matcher(m, callback);
+}
+
+}  // namespace low_precision
+}  // namespace pass
+}  // namespace ov

src/common/transformations/include/transformations/common_optimizations/convert_quantize_dequantize.hpp

@@ -7,6 +7,7 @@
 #include <memory>
 #include <vector>
 
+#include "openvino/core/type/element_type.hpp"
 #include "openvino/pass/matcher_pass.hpp"
 #include "transformations_visibility.hpp"
 
@@ -32,5 +33,10 @@ class TRANSFORMATIONS_API ConvertQuantizeDequantize;
 class ov::pass::ConvertQuantizeDequantize : public ov::pass::MatcherPass {
 public:
     OPENVINO_MATCHER_PASS_RTTI("ConvertQuantizeDequantize");
-    ConvertQuantizeDequantize();
+    ConvertQuantizeDequantize(const ov::element::TypeVector& supported_low_precisions = {ov::element::i8,
+                                                                                         ov::element::u8,
+                                                                                         ov::element::i16,
+                                                                                         ov::element::u16},
+                              const ov::element::TypeVector& supported_original_precisions = {ov::element::f32},
+                              const bool ignore_consumers_count_check = false);
 };

src/common/transformations/src/transformations/common_optimizations/convert_quantize_dequantize.cpp

@@ -16,6 +16,7 @@
 #include "openvino/op/fake_quantize.hpp"
 #include "openvino/op/multiply.hpp"
 #include "openvino/op/subtract.hpp"
+#include "openvino/pass/pattern/op/optional.hpp"
 #include "openvino/pass/pattern/op/wrap_type.hpp"
 #include "transformations/utils/utils.hpp"
 
@@ -62,100 +63,97 @@
 //                                        v
 //
 
-ov::pass::ConvertQuantizeDequantize::ConvertQuantizeDequantize() {
+ov::pass::ConvertQuantizeDequantize::ConvertQuantizeDequantize(
+    const ov::element::TypeVector& supported_low_precisions,
+    const ov::element::TypeVector& supported_original_precisions,
+    const bool ignore_consumers_count_check) {
     MATCHER_SCOPE(ConvertQuantizeDequantize);
-    auto data_pattern = pass::pattern::any_input();
-    auto input_low_pattern = pass::pattern::any_input();
-    auto input_high_pattern = pass::pattern::any_input();
-    auto output_low_pattern = ov::pass::pattern::wrap_type<ov::op::v0::Constant>();
-    auto output_high_pattern = ov::pass::pattern::wrap_type<ov::op::v0::Constant>();
-    auto fq_pattern = ov::pass::pattern::wrap_type<ov::op::v0::FakeQuantize>(
+
+    using namespace ov::pass::pattern;
+    using namespace ov::op;
+
+    auto data_pattern = any_input(type_matches_any(supported_original_precisions));
+    auto input_low_pattern = any_input();
+    auto input_high_pattern = any_input();
+    auto output_low_pattern = wrap_type<v0::Constant>();
+    auto output_high_pattern = wrap_type<v0::Constant>();
+    auto fq_pattern = wrap_type<v0::FakeQuantize>(
         {data_pattern, input_low_pattern, input_high_pattern, output_low_pattern, output_high_pattern});
-    auto convert1_pattern = ov::pass::pattern::wrap_type<ov::op::v0::Convert>(
-        {fq_pattern},
-        pattern::type_matches_any({element::i8, element::u8, element::i16, element::u16}));
-    auto convert2_pattern =
-        ov::pass::pattern::wrap_type<ov::op::v0::Convert>({convert1_pattern}, pattern::type_matches(element::f32));
-    auto zero_point_pattern = pass::pattern::any_input();
-    auto sub_pattern = ov::pass::pattern::wrap_type<ov::op::v1::Subtract>({convert2_pattern, zero_point_pattern},
-                                                                          pattern::consumers_count(1));
-    auto scale_pattern = pass::pattern::any_input();
-    auto mul_pattern = ov::pass::pattern::wrap_type<ov::op::v1::Multiply>({sub_pattern, scale_pattern});
-
-    ov::matcher_pass_callback callback = [OV_CAPTURE_CPY_AND_THIS](pattern::Matcher& m) {
+    ov::pass::pattern::op::Predicate convert1_predicate =
+        ignore_consumers_count_check ? type_matches_any(supported_low_precisions)
+                                     : type_matches_any(supported_low_precisions) && consumers_count(1);
+    auto convert1_pattern = wrap_type<v0::Convert>({fq_pattern}, convert1_predicate);
+    ov::pass::pattern::op::Predicate convert2_predicate =
+        ignore_consumers_count_check ? type_matches_any(supported_original_precisions)
+                                     : type_matches_any(supported_original_precisions) && consumers_count(1);
+    auto convert2_pattern = wrap_type<v0::Convert>({convert1_pattern}, convert2_predicate);
+
+    auto zero_point_pattern = any_input();
+    ov::pass::pattern::op::Predicate sub_predicate =
+        ignore_consumers_count_check ? ov::pass::pattern::op::Predicate() : consumers_count(1);
+    auto sub_pattern = optional<v1::Subtract>({convert2_pattern, zero_point_pattern}, sub_predicate);
+    auto scale_pattern = any_input();
+    auto mul_pattern = wrap_type<v1::Multiply>({sub_pattern, scale_pattern});
+
+    ov::matcher_pass_callback callback = [OV_CAPTURE_CPY_AND_THIS](Matcher& m) {
         auto pattern_map = m.get_pattern_value_map();
 
         if (transformation_callback(m.get_match_root())) {
             return false;
         }
 
-        auto data = pattern_map[data_pattern];
-        auto input_low = pattern_map[input_low_pattern];
-        auto input_high = pattern_map[input_high_pattern];
-        auto output_low = ov::as_type_ptr<ov::op::v0::Constant>(pattern_map[output_low_pattern].get_node_shared_ptr());
+        auto data = pattern_map.at(data_pattern);
+        auto input_low = pattern_map.at(input_low_pattern);
+        auto input_high = pattern_map.at(input_high_pattern);
+        auto output_low =
+            ov::as_type_ptr<ov::op::v0::Constant>(pattern_map.at(output_low_pattern).get_node_shared_ptr());
         if (!output_low)
             return false;
         auto output_high =
-            ov::as_type_ptr<ov::op::v0::Constant>(pattern_map[output_high_pattern].get_node_shared_ptr());
+            ov::as_type_ptr<ov::op::v0::Constant>(pattern_map.at(output_high_pattern).get_node_shared_ptr());
         if (!output_high)
             return false;
-        auto fq = ov::as_type_ptr<ov::op::v0::FakeQuantize>(pattern_map[fq_pattern].get_node_shared_ptr());
+        auto fq = ov::as_type_ptr<ov::op::v0::FakeQuantize>(pattern_map.at(fq_pattern).get_node_shared_ptr());
         if (!fq)
             return false;
-        auto zero_point = pattern_map[zero_point_pattern];
-        auto scale = pattern_map[scale_pattern];
-        auto convert1 = pattern_map[convert1_pattern];
-        auto convert2 = pattern_map[convert2_pattern];
-        auto mul = pattern_map[mul_pattern].get_node_shared_ptr();
-
-        // convert1 and convert2 should have only one input
-        if (convert1.get_target_inputs().size() != 1)
-            return false;
-        if (convert2.get_target_inputs().size() != 1)
+        auto scale = pattern_map.at(scale_pattern);
+        auto convert1 = pattern_map.at(convert1_pattern);
+        auto convert2 = pattern_map.at(convert2_pattern);
+        auto mul = pattern_map.at(mul_pattern).get_node_shared_ptr();
+
+        static const std::unordered_set<size_t> supported_levels{256, 65536};
+        const auto levels = fq->get_levels();
+        if (!supported_levels.count(levels))
             return false;
 
-        // we support:
-        // i8 or u8: 'levels' attribute must be 256
-        // i16 or u16: 'levels' attribute must be 65536
-        size_t levels = fq->get_levels();
-        if (levels != 256 && levels != 65536)
-            return false;
-
-        // check if (out_low_val, out_high_val) is (-128, 127) or (0, 255) or (-32768, 32767) or (0, 65535)
         float out_low_val;
-        if (!op::util::get_single_value(output_low, out_low_val))
+        if (!ov::op::util::get_single_value(output_low, out_low_val))
             return false;
         float out_high_val;
-        if (!op::util::get_single_value(output_high, out_high_val))
+        if (!ov::op::util::get_single_value(output_high, out_high_val))
             return false;
+
+        static const std::unordered_map<ov::element::Type_t, std::pair<float, float>> supported_intervals{
+            {ov::element::i8, {-128.f, 127.f}},
+            {ov::element::u8, {0.f, 255.f}},
+            {ov::element::i16, {-32768.f, 32767.f}},
+            {ov::element::u16, {0.f, 65535.f}}};
         const auto& type = convert1.get_element_type();
-        switch (type) {
-        case element::Type_t::i8:
-            if (out_low_val != -128 || out_high_val != 127)
-                return false;
-            break;
-        case element::Type_t::u8:
-            if (out_low_val != 0 || out_high_val != 255)
-                return false;
-            break;
-        case element::Type_t::i16:
-            if (out_low_val != -32768 || out_high_val != 32767)
-                return false;
-            break;
-        case element::Type_t::u16:
-            if (out_low_val != 0 || out_high_val != 65535)
-                return false;
-            break;
-        default:
+        // check if (out_low_val, out_high_val) pair is mapped on the expected precision ranges
+        if (supported_intervals.count(type) == 0 ||
+            supported_intervals.at(type) != std::make_pair(out_low_val, out_high_val)) {
             return false;
         }
 
-        std::shared_ptr<Node> new_out_low =
-            std::make_shared<ov::op::v1::Multiply>(std::make_shared<ov::op::v1::Subtract>(output_low, zero_point),
-                                                   scale);
-        std::shared_ptr<Node> new_out_high =
-            std::make_shared<ov::op::v1::Multiply>(std::make_shared<ov::op::v1::Subtract>(output_high, zero_point),
-                                                   scale);
+        const bool has_zero_point = pattern_map.count(zero_point_pattern);
+        std::shared_ptr<Node> new_out_low = output_low, new_out_high = output_high;
+        if (has_zero_point) {
+            const auto& zero_point = pattern_map.at(zero_point_pattern);
+            new_out_low = std::make_shared<ov::op::v1::Subtract>(new_out_low, zero_point);
+            new_out_high = std::make_shared<ov::op::v1::Subtract>(new_out_high, zero_point);
+        }
+        new_out_low = std::make_shared<ov::op::v1::Multiply>(new_out_low, scale);
+        new_out_high = std::make_shared<ov::op::v1::Multiply>(new_out_high, scale);
 
         // check if new_out_low/high shapes are broadcastable to FQ's input
         auto data_shape = data.get_partial_shape();

src/plugins/intel_gpu/src/plugin/transformations_pipeline.cpp

@@ -24,6 +24,7 @@
 #include "low_precision/fold_convert.hpp"
 #include "low_precision/fuse_convert.hpp"
 #include "low_precision/group_convolution.hpp"
+#include "low_precision/qdq_stripping.hpp"
 #include "low_precision/low_precision.hpp"
 #include "low_precision/mat_mul.hpp"
 #include "low_precision/multiply_to_group_convolution.hpp"
@@ -387,8 +388,22 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
             ov::disable_keep_const_precision(node);
         }
 
-        auto is_model_quantized = ov::pass::low_precision::LowPrecision::isFunctionQuantized(func);
+        using namespace ov::pass::low_precision;
+        auto is_model_quantized = LowPrecision::isFunctionQuantized(func, std::set<levels>{levels::int8, levels::int8_narrow_range});
         enableInt8 = config.get_enable_lp_transformations() && is_model_quantized;
+        {
+            using namespace ov::element;
+            // QDQ stripping pipeline
+            // 1. Transform DQ part to canonicalized form: Multiply->Add => Subtract->Multiply
+            manager.register_pass<AddTransformation>();
+            // 2. Fuse FQ->Convert->DQ to a single FQ
+            manager.register_pass<ov::pass::ConvertQuantizeDequantize>(TypeVector{i16, u16}, TypeVector{f32}, true);
+            // 3. Strip FQ layers with unsupported levels
+            bool replace_with_clamp = ov::util::getenv_bool("REPLACE_QDQ_WITH_CLAMP", true);
+            std::cout << "[ QDQ STRIPPING INFO ] replace_with_clamp = " << replace_with_clamp << std::endl;
+            manager.register_pass<FQStrippingTransformation>(std::set<size_t>{levels::int16}, replace_with_clamp);
+            manager.register_pass<ov::pass::Validate>();
+        }
 
         manager.register_pass<ov::pass::MarkDequantization>(
             std::vector<ov::element::Type>{ ov::element::i8, ov::element::u8, ov::element::i4, ov::element::u4 },

src/plugins/intel_gpu/tests/functional/shared_tests_instances/low_precision_transformations/fq_transformation.cpp

@@ -44,13 +44,14 @@ const std::vector<FakeQuantizeTransformationParam> fakeQuantizeOnDataValues = {
         { 256ul, {}, { -127.5f }, { 0.f }, { -127.5f }, { 0.f } },
         "Pooling", "u8"
     },
+    // Not expected FQ levels
     {
         { 16ul, {}, { 0.f }, { 1.5f }, { 0.f }, { 1.5f } },
-        "Pooling", "u8"
+        "Pooling", "f32"
     },
     {
         { 16ul, {}, { -8.f }, { 7.f }, { -0.8f }, { 0.7f } },
-        "Pooling", "i8"
+        "Pooling", "f32"
     },
     // nGraph: I8->FP32 Convert is not supported
     // { 256ul, {}, { -1.28f} , { 1.27f }, { -1.28f} , { 1.27f } },