Skip to content

Remove perfmetrics synchronization workaround #4619

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
mzegla merged 1 commit into from
Oct 2, 2026
Merged

Remove perfmetrics synchronization workaround #4619

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 4 additions & 21 deletions src/llm/language_model/continuous_batching/servable.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,6 @@
#include <memory>
#include <stdexcept>
#include <string>
#include <thread>
#include <vector>

#include "../../../logging.hpp"
Expand Down Expand Up @@ -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) {
Expand Down Expand Up @@ -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;
}
Expand All @@ -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;
}
Expand Down
26 changes: 4 additions & 22 deletions src/llm/visual_language_model/continuous_batching/servable.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,7 +19,6 @@
#include <memory>
#include <stdexcept>
#include <string>
#include <thread>
#include <vector>

#include "src/port/rapidjson_document.hpp"
Expand All @@ -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();
Expand Down Expand Up @@ -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;
}
Expand All @@ -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;
}
Expand Down
Loading