diff --git a/src/llm/language_model/continuous_batching/servable.cpp b/src/llm/language_model/continuous_batching/servable.cpp
index 90a2342fcf..9734adf4d8 100644
--- a/src/llm/language_model/continuous_batching/servable.cpp
+++ b/src/llm/language_model/continuous_batching/servable.cpp
@@ -17,7 +17,6 @@
 #include <memory>
 #include <stdexcept>
 #include <string>
-#include <thread>
 #include <vector>
 
 #include "../../../logging.hpp"
@@ -57,20 +56,6 @@ void ContinuousBatchingServable::logPerfMetrics(ov::genai::PerfMetrics& perfMetr
         prefillSpeedTps);
 }
 
-// CB stepping thread writes metrics in _free_non_running_requests() slightly after
-// pushing the final output. Yield briefly to close the race window.
-// TODO: remove once GenAI's get_perf_metrics() blocks instead of asserting (fix in generation_stream.hpp)
-static std::optional<ov::genai::PerfMetrics> tryGetPerfMetrics(const ov::genai::GenerationHandle& handle) {
-    for (int i = 0; i < 1000; ++i) {
-        try {
-            return handle->get_perf_metrics();
-        } catch (const ov::Exception&) {
-            std::this_thread::yield();
-        }
-    }
-    return std::nullopt;
-}
-
 void ContinuousBatchingServable::notifyExecutorThread() {
     SPDLOG_LOGGER_TRACE(llm_calculator_logger, "Notifying executor thread");
     if (properties->llmExecutorWrapper == nullptr) {
@@ -177,9 +162,8 @@ absl::Status ContinuousBatchingServable::prepareCompleteResponse(std::shared_ptr
     auto status = GenAiServable::prepareCompleteResponse(executionContext);
     if (status.ok() && llm_calculator_logger->should_log(spdlog::level::debug)) {
         auto cbExecutionContext = std::static_pointer_cast<ContinuousBatchingServableExecutionContext>(executionContext);
-        auto perfMetrics = tryGetPerfMetrics(cbExecutionContext->generationHandle);
-        if (perfMetrics)
-            logPerfMetrics(*perfMetrics);
+        auto perfMetrics = cbExecutionContext->generationHandle->get_perf_metrics();
+        logPerfMetrics(perfMetrics);
     }
     return status;
 }
@@ -190,9 +174,8 @@ absl::Status ContinuousBatchingServable::preparePartialResponse(std::shared_ptr<
         !executionContext->sendLoopbackSignal &&
         llm_calculator_logger->should_log(spdlog::level::debug)) {
         auto cbExecutionContext = std::static_pointer_cast<ContinuousBatchingServableExecutionContext>(executionContext);
-        auto perfMetrics = tryGetPerfMetrics(cbExecutionContext->generationHandle);
-        if (perfMetrics)
-            logPerfMetrics(*perfMetrics);
+        auto perfMetrics = cbExecutionContext->generationHandle->get_perf_metrics();
+        logPerfMetrics(perfMetrics);
     }
     return status;
 }
diff --git a/src/llm/visual_language_model/continuous_batching/servable.cpp b/src/llm/visual_language_model/continuous_batching/servable.cpp
index d459d2dba9..e3237f125b 100644
--- a/src/llm/visual_language_model/continuous_batching/servable.cpp
+++ b/src/llm/visual_language_model/continuous_batching/servable.cpp
@@ -19,7 +19,6 @@
 #include <memory>
 #include <stdexcept>
 #include <string>
-#include <thread>
 #include <vector>
 
 #include "src/port/rapidjson_document.hpp"
@@ -36,21 +35,6 @@
 
 namespace ovms {
 
-// CB stepping thread writes base perf metrics in _free_non_running_requests() slightly
-// after pushing the final output; get_vlm_perf_metrics() calls get_perf_metrics() internally.
-// Yield briefly to close the race window.
-// TODO: remove once GenAI's get_perf_metrics() blocks instead of asserting (fix in generation_stream.hpp)
-static std::optional<ov::genai::VLMPerfMetrics> tryGetVlmPerfMetrics(const ov::genai::GenerationHandle& handle) {
-    for (int i = 0; i < 1000; ++i) {
-        try {
-            return handle->get_vlm_perf_metrics();
-        } catch (const ov::Exception&) {
-            std::this_thread::yield();
-        }
-    }
-    return std::nullopt;
-}
-
 void VisualLanguageModelServable::logPerfMetrics(ov::genai::VLMPerfMetrics& perfMetrics) {
     const size_t inputTokenCount = perfMetrics.get_num_input_tokens();
     const size_t outputTokenCount = perfMetrics.get_num_generated_tokens();
@@ -100,9 +84,8 @@ absl::Status VisualLanguageModelServable::prepareCompleteResponse(std::shared_pt
     auto status = GenAiServable::prepareCompleteResponse(executionContext);
     if (status.ok() && llm_calculator_logger->should_log(spdlog::level::debug)) {
         auto vlmExecutionContext = std::static_pointer_cast<VisualLanguageModelServableExecutionContext>(executionContext);
-        auto perfMetrics = tryGetVlmPerfMetrics(vlmExecutionContext->generationHandle);
-        if (perfMetrics)
-            logPerfMetrics(*perfMetrics);
+        auto perfMetrics = vlmExecutionContext->generationHandle->get_vlm_perf_metrics();
+        logPerfMetrics(perfMetrics);
     }
     return status;
 }
@@ -113,9 +96,8 @@ absl::Status VisualLanguageModelServable::preparePartialResponse(std::shared_ptr
         !executionContext->sendLoopbackSignal &&
         llm_calculator_logger->should_log(spdlog::level::debug)) {
         auto vlmExecutionContext = std::static_pointer_cast<VisualLanguageModelServableExecutionContext>(executionContext);
-        auto perfMetrics = tryGetVlmPerfMetrics(vlmExecutionContext->generationHandle);
-        if (perfMetrics)
-            logPerfMetrics(*perfMetrics);
+        auto perfMetrics = vlmExecutionContext->generationHandle->get_vlm_perf_metrics();
+        logPerfMetrics(perfMetrics);
     }
     return status;
 }