Disabled dynamic quantization by default

Author: dmitry-gorokhov

src/plugins/intel_cpu/src/config.h

@@ -55,7 +55,7 @@ struct Config {
     std::string dumpToDot = {};
     std::string device_id = {};
     float fcSparseWeiDecompressionRate = 1.0f;
-    uint64_t fcDynamicQuantizationGroupSize = 32;
+    uint64_t fcDynamicQuantizationGroupSize = 0;
 #if defined(OPENVINO_ARCH_X86_64)
     size_t rtCacheCapacity = 5000ul;
 #else

src/plugins/intel_cpu/src/nodes/executors/dnnl/dnnl_fullyconnected_primitive.cpp

@@ -116,7 +116,7 @@ bool DnnlFCPrimitive::useWeightsDecompressionImpl(const ov::element::Type inputT
            one_of(weightsType, u8, nf4, u4, i4);
 }
 
-bool DnnlFCPrimitive::useDynamicQuantizationImpl(size_t dqGroupSize, const MemoryDescPtr weightsDesc,
+bool DnnlFCPrimitive::useDynamicQuantizationImpl(size_t dqGroupSize, const MemoryDescPtr srcDesc, const MemoryDescPtr weightsDesc,
                                                  MemoryCPtr scalesPtr, MemoryCPtr zpPtr, bool needTranspose) {
     if (dqGroupSize == 0)
         return false;
@@ -125,6 +125,9 @@ bool DnnlFCPrimitive::useDynamicQuantizationImpl(size_t dqGroupSize, const Memor
         !dnnl::impl::cpu::x64::mayiuse(dnnl::impl::cpu::x64::avx512_core_vnni))
         return false;
 
+    if (srcDesc->getPrecision() != ov::element::f32)
+        return false;
+
     if (!one_of(weightsDesc->getPrecision(), ov::element::u8, ov::element::u4))
         return false;
 
@@ -315,7 +318,7 @@ DnnlShapeAgnosticDataPtr DnnlFCPrimitive::createShapeAgnosticData(const FCAttrs&
 
     const auto useWeightsDecompression = useWeightsDecompressionImpl(srcDesc->getPrecision(), weiDesc->getPrecision());
     const auto useDynamicQuantization = useWeightsDecompression &&
-        useDynamicQuantizationImpl(attrs.dynamicQuantizationGroupSize, weiDesc,
+        useDynamicQuantizationImpl(attrs.dynamicQuantizationGroupSize, srcDesc, weiDesc,
                                    attrs.decompressionMultiplyPtr, attrs.decompressionSubtractPtr, !attrs.weightsNonTransposed);
 
     const auto postOpData = createPrimitiveAttrs(attrs, postOps, memory, context, useDynamicQuantization);

src/plugins/intel_cpu/src/nodes/executors/dnnl/dnnl_fullyconnected_primitive.hpp

@@ -70,7 +70,7 @@ class DnnlFCPrimitive {
                                                    const DnnlShapeAgnosticDataPtr& shapeAgnosticData);
 
 private:
-    static bool useDynamicQuantizationImpl(size_t dqGroupSize, const MemoryDescPtr weightsDesc,
+    static bool useDynamicQuantizationImpl(size_t dqGroupSize, const MemoryDescPtr srcDesc, const MemoryDescPtr weightsDesc,
                                            MemoryCPtr scalesPtr, MemoryCPtr zpPtr, bool needTranspose);
 
     dnnl::stream m_stream;

src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp

@@ -288,7 +288,8 @@ void Transformations::PreLpt(const std::vector<ov::element::Type>& defaultPrecis
                                                      ov::element::u4,
                                                      ov::element::i4,
                                                      ov::element::nf4};
-    CPU_REGISTER_PASS_X64(decompression_handling_manager, ov::pass::MarkDequantizationSubgraph, decompression_precisions, false);
+    bool fold_subtract_const = config.fcDynamicQuantizationGroupSize != 0;
+    CPU_REGISTER_PASS_X64(decompression_handling_manager, ov::pass::MarkDequantizationSubgraph, decompression_precisions, fold_subtract_const);
     CPU_SET_CALLBACK_X64(decompression_handling_manager, [&](const_node_ptr &node) -> bool {
         return !is_decompression_multiply(node);
     }, ov::pass::MarkDequantizationSubgraph);