disabled scaling for quantized models

openvinotoolkit · Jan 13, 2025 · 0874b17 · 0874b17
1 parent 8122cde
commit 0874b17
Show file tree

Hide file tree

Showing 4 changed files with 10 additions and 10 deletions.
diff --git a/src/common/low_precision_transformations/src/add.cpp b/src/common/low_precision_transformations/src/add.cpp
@@ -216,11 +216,11 @@ bool AddTransformation::transform(TransformationContext& context, ov::pass::patt
 
         auto output_type = scalingMode ? add->get_output_element_type(0) : element::f32;
         newAddOrSubtract = std::make_shared<ov::op::TypeRelaxed<ov::opset1::Add>>(
-            std::vector<element::Type>{output_type, output_type}, std::vector<element::Type>{ output_type },
+            std::vector<element::Type>{output_type, output_type}, std::vector<element::Type>{output_type},
             ov::op::TemporaryReplaceOutputType(inputs[0], output_type).get(),
             ov::op::TemporaryReplaceOutputType(inputs[1], output_type).get());
         newMultiply = std::make_shared<ov::op::TypeRelaxed<ov::opset1::Multiply>>(
-            std::vector<element::Type>{output_type, output_type}, std::vector<element::Type>{ add->get_output_element_type(0) },
+            std::vector<element::Type>{output_type, output_type}, std::vector<element::Type>{add->get_output_element_type(0)},
             ov::op::TemporaryReplaceOutputType(newAddOrSubtract, output_type).get(),
             ov::op::TemporaryReplaceOutputType(multiplyEmptyPathValues, output_type).get());
 

diff --git a/src/common/low_precision_transformations/src/multiply_partial.cpp b/src/common/low_precision_transformations/src/multiply_partial.cpp
@@ -79,17 +79,17 @@ bool MultiplyPartialTransformation::transform(TransformationContext& context, ov
         auto constParent = multiply->input_value(multiplyBranch.first == 0 ? 1 : 0);
         auto multiplyParentParent = multiplyParent.get_node_shared_ptr()->input_value(multiplyBranch.second);
         auto multiplyParentConst = multiplyParent.get_node_shared_ptr()->input_value(multiplyBranch.second == 0 ? 1 : 0);
-        auto input_data_type = scalingMode ? multiply->get_output_element_type(0) : element::f32;
+        auto inputDataType = scalingMode ? multiply->get_output_element_type(0) : element::f32;
 
         newMultiply = std::make_shared<ov::op::TypeRelaxed<ov::opset1::Multiply>>(
-            std::vector<ov::element::Type>{ input_data_type, input_data_type },
+            std::vector<ov::element::Type>{ inputDataType, inputDataType },
             std::vector<ov::element::Type>{ multiply->get_output_element_type(0) },
-            ov::op::TemporaryReplaceOutputType(multiplyParentParent, input_data_type).get(),
+            ov::op::TemporaryReplaceOutputType(multiplyParentParent, inputDataType).get(),
             ov::op::TemporaryReplaceOutputType(
                 fold<ov::opset1::Multiply>(
-                    foldConvert(multiplyParentConst, input_data_type),
-                    foldConvert(constParent, input_data_type)),
-                input_data_type).get());
+                    foldConvert(multiplyParentConst, inputDataType),
+                    foldConvert(constParent, inputDataType)),
+                inputDataType).get());
 
         NetworkHelper::copyInfo(multiplyParent.get_node_shared_ptr(), newMultiply);
         NetworkHelper::copyInfo(multiply, newMultiply);

diff --git a/src/plugins/intel_gpu/src/plugin/transformations_pipeline.cpp b/src/plugins/intel_gpu/src/plugin/transformations_pipeline.cpp
@@ -929,7 +929,7 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
 
         float activations_scale_factor = config.get_property(ov::hint::activations_scale_factor);
 
-        if (activations_scale_factor > 0.f && infer_precision == ov::element::f16) {
+        if (activations_scale_factor > 0.f && infer_precision == ov::element::f16 && !enableInt8) {
             using namespace ov::pass::low_precision;
 
             auto supportedPrecisions = std::vector<PrecisionsRestriction>({});

diff --git a/src/plugins/intel_gpu/src/runtime/execution_config.cpp b/src/plugins/intel_gpu/src/runtime/execution_config.cpp
@@ -275,8 +275,8 @@ void ExecutionConfig::apply_user_properties(const cldnn::device_info& info) {
 void ExecutionConfig::apply_rt_info(const cldnn::device_info& info, const ov::RTMap& rt_info) {
     if (!info.supports_immad) {
         apply_rt_info_property(ov::hint::kv_cache_precision, rt_info);
+        apply_rt_info_property(ov::hint::activations_scale_factor, rt_info);
     }
-    apply_rt_info_property(ov::hint::activations_scale_factor, rt_info);
     apply_rt_info_property(ov::hint::dynamic_quantization_group_size, rt_info);
 }