Only LLMs scalar constant nodes need to be checked for saturation to further optimize model compile time

Author: liubo-intel

src/plugins/intel_cpu/src/nodes/input.cpp

@@ -402,8 +402,9 @@ void Input::cloneBlobIfRequired() {
             if (!size)
                 return;
 
+            // Only LLMs scalar constant nodes with bf16 inferencePrecision need to be checked for saturation
             const bool do_bf16_saturation_check =
-                (context->getConfig().inferencePrecision == ov::element::bf16) ? true : false;
+                (context->getConfig().inferencePrecision == ov::element::bf16 && size == 1) ? true : false;
 
 #if defined(OPENVINO_ARCH_X86_64)
             auto fn = jit_has_subnormals_function();

src/plugins/intel_cpu/tests/functional/custom/subgraph_tests/src/x64/bf16_convert_saturation.cpp

@@ -58,7 +58,7 @@ class BF16ConvertSaturation : public testing::WithParamInterface<selectParams>,
         in_data.resolution = 1;
         auto thenTensor = ov::test::utils::create_and_fill_tensor(precision, ov::Shape{1}, in_data);
 
-        in_data.start_from = 3.40282e+38;
+        in_data.start_from = 1;
         in_data.range = 10;
         in_data.resolution = 2;
         auto elseTensor = ov::test::utils::create_and_fill_tensor(precision, ov::Shape{2, 1, 32, 32}, in_data);