release useless inferreq

Signed-off-by: xipingya <xiping.yan@intel.com>

Author: xipingyan

src/cpp/src/module_genai/modules/md_text_to_speech/models/qwen3_omni.cpp

@@ -207,6 +207,11 @@ void TextToSpeechImpl_Qwen3Omni::load_code_predictor_models(const ov::AnyMap& tt
         GENAI_INFO("TextToSpeechModule[" + module_desc->name +
                    "]: sample_codec_token_greedy_search is enabled, will use greedy decoding in sample_codec_token");
         merge_code_predictor_ov_models(ar_models, sce_models);
+        if (m_enable_merge_ov_models) {
+            // Release original infer requests to save memory since they won't be used anymore.
+            m_code_predictor_ar_infers.clear();
+            m_code_predictor_single_codec_embed_infers.clear();
+        }
     }
 }