Skip to content

Info -> warning: Override performance mode to THROUGHPUT for compilation #34666

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

Open
DariaMityagina wants to merge 2 commits into
base: master
Choose a base branch
from
Open

Info -> warning: Override performance mode to THROUGHPUT for compilation #34666

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
48f74cd
Info -> warning: Override performance mode to THROUGHPUT for compilation
DariaMityagina Mar 12, 2026
b1b9bcb
Info -> warning: Override performance mode to THROUGHPUT for compilat…
DariaMityagina Mar 13, 2026
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
35 changes: 25 additions & 10 deletions src/plugins/intel_npu/src/plugin/src/plugin.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -674,29 +674,44 @@ std::shared_ptr<ov::ICompiledModel> Plugin::compile_model(const std::shared_ptr<
std::shared_ptr<intel_npu::IGraph> graph;

auto compileWithConfig = [&](auto&& modelToCompile, const auto& config) {
if (!localConfig.get<WEIGHTLESS_BLOB>() && !localConfig.get<ENABLE_WEIGHTLESS>() ) {
if (!localConfig.get<WEIGHTLESS_BLOB>() && !localConfig.get<ENABLE_WEIGHTLESS>()) {
return compiler->compile(modelToCompile, config);
} else {
check_weightless_cache_attribute_occurrence(model);
return compiler->compileWS(std::move(modelToCompile), config);
}
};

const bool performanceHintSetByUser = properties.find(ov::hint::performance_mode.name()) != properties.end();

try {
_logger.debug("performing compile");

// Determine which model to use
auto modelToCompile = successfullyDebatched ? batchedModel : model->clone();

if (successfullyDebatched && localConfig.get<PERFORMANCE_HINT>() == ov::hint::PerformanceMode::LATENCY) {
_logger.info("Override performance mode to THROUGHPUT for compilation");

auto modifiedConfig = localConfig; // Copy only when needed
std::stringstream strStream;
strStream << ov::hint::PerformanceMode::THROUGHPUT;
modifiedConfig.update({{ov::hint::performance_mode.name(), strStream.str()}});

graph = compileWithConfig(std::move(modelToCompile), modifiedConfig);
if (successfullyDebatched) {
if (!performanceHintSetByUser) {
_logger.info("Setting performance mode to THROUGHPUT for batched model compilation.");

auto modifiedConfig = localConfig; // Copy only when needed
std::stringstream strStream;
strStream << ov::hint::PerformanceMode::THROUGHPUT;
modifiedConfig.update({{ov::hint::performance_mode.name(), strStream.str()}});

graph = compileWithConfig(std::move(modelToCompile), modifiedConfig);
} else if (localConfig.get<PERFORMANCE_HINT>() == ov::hint::PerformanceMode::LATENCY) {
_logger.warning("PERFORMANCE_HINT is explicitly set to LATENCY mode, but batch dimension (N) is "
"detected in the model. The NPU Plugin will reshape the model to batch size 1 and "
"process each batch slice separately.");
_logger.warning("For optimal performance with batched models, THROUGHPUT mode is highly recommended, "
"as LATENCY mode prevents parallel batch processing.");
_logger.warning("If batch detection appears incorrect, verify that the input and output layouts are "
"configured properly.");
graph = compileWithConfig(std::move(modelToCompile), localConfig);
} else {
graph = compileWithConfig(std::move(modelToCompile), localConfig);
}
} else {
graph = compileWithConfig(std::move(modelToCompile), localConfig); // No copy
}
Expand Down
Loading