#include <TritonClient.h>

Inheritance diagram for TritonClient:

Classes
struct	ServerSideStats

Public Member Functions
unsigned	batchSize () const

bool	noBatch () const

void	reset () override

TritonServerType	serverType () const

bool	setBatchSize (unsigned bsize)

void	setUseSharedMemory (bool useShm)

	TritonClient (const edm::ParameterSet &params, const std::string &debugName)

bool	useSharedMemory () const

bool	verbose () const

	~TritonClient () override

Public Member Functions inherited from SonicClient< TritonInputMap, TritonOutputMap >
	SonicClient (const edm::ParameterSet &params, const std::string &debugName, const std::string &clientName)

Public Member Functions inherited from SonicClientBase
const std::string &	clientName () const

const std::string &	debugName () const

virtual void	dispatch (edm::WaitingTaskWithArenaHolder holder)

virtual void	dispatch ()

SonicMode	mode () const

	SonicClientBase (const edm::ParameterSet &params, const std::string &debugName, const std::string &clientName)

virtual	~SonicClientBase ()=default

Public Member Functions inherited from SonicClientTypes< TritonInputMap, TritonOutputMap >
Input &	input ()

const Output &	output () const

virtual	~SonicClientTypes ()=default

Static Public Member Functions
static void	fillPSetDescription (edm::ParameterSetDescription &iDesc)

Static Public Member Functions inherited from SonicClientBase
static void	fillBasePSetDescription (edm::ParameterSetDescription &desc, bool allowRetry=true)

Protected Member Functions
void	evaluate () override

void	getResults (std::shared_ptr< triton::client::InferResult > results)

inference::ModelStatistics	getServerSideStatus () const

template<typename F >
bool	handle_exception (F &&call)

void	reportServerSideStats (const ServerSideStats &stats) const

ServerSideStats	summarizeServerStats (const inference::ModelStatistics &start_status, const inference::ModelStatistics &end_status) const

Protected Member Functions inherited from SonicClientBase
void	finish (bool success, std::exception_ptr eptr=std::exception_ptr{})

void	setMode (SonicMode mode)

void	start (edm::WaitingTaskWithArenaHolder holder)

void	start ()

Protected Attributes
unsigned	batchSize_

std::unique_ptr< triton::client::InferenceServerGrpcClient >	client_

grpc_compression_algorithm	compressionAlgo_

triton::client::Headers	headers_

std::vector< triton::client::InferInput * >	inputsTriton_

unsigned	maxBatchSize_

bool	noBatch_

triton::client::InferOptions	options_

std::vector< const triton::client::InferRequestedOutput * >	outputsTriton_

TritonServerType	serverType_

bool	useSharedMemory_

bool	verbose_

Protected Attributes inherited from SonicClientBase
unsigned	allowedTries_

std::string	clientName_

std::string	debugName_

std::unique_ptr< SonicDispatcher >	dispatcher_

std::string	fullDebugName_

std::optional< edm::WaitingTaskWithArenaHolder >	holder_

SonicMode	mode_

unsigned	tries_

bool	verbose_

Protected Attributes inherited from SonicClientTypes< TritonInputMap, TritonOutputMap >
Input	input_

Output	output_

Private Member Functions
auto	client ()

Private Attributes
friend	TritonInputData

friend	TritonOutputData

Additional Inherited Members
Public Types inherited from SonicClientTypes< TritonInputMap, TritonOutputMap >
typedef TritonInputMap	Input

typedef TritonOutputMap	Output

Detailed Description

Definition at line 19 of file TritonClient.h.

Constructor & Destructor Documentation

◆ TritonClient()

TritonClient::TritonClient	(	const edm::ParameterSet &	params,
		const std::string &	debugName
	)

Definition at line 39 of file TritonClient.cc.

References client_, SonicClientBase::fullDebugName_, SonicClientTypes< TritonInputMap, TritonOutputMap >::input_, inputsTriton_, LocalCPU, SiStripPI::max, maxBatchSize_, mps_check::msg, noBatch_, heppy_report::oname, options_, SonicClientTypes< TritonInputMap, TritonOutputMap >::output_, outputsTriton_, submitPVValidationJobs::params, TritonService::pid(), triton_utils::printColl(), contentValuesFiles::server, TritonService::serverInfo(), serverType_, setBatchSize(), SonicClientBase::setMode(), AlCaHLTBitMon_QueryRunRegistry::string, summarizeEdmComparisonLogfiles::success, Sync, TRITON_THROW_IF_ERROR, and verbose_.

     : SonicClient(params, debugName, "TritonClient"),
       verbose_(params.getUntrackedParameter<bool>("verbose")),
       useSharedMemory_(params.getUntrackedParameter<bool>("useSharedMemory")),
       compressionAlgo_(getCompressionAlgo(params.getUntrackedParameter<std::string>("compression"))),
       options_(params.getParameter<std::string>("modelName")) {
   //get appropriate server for this model
   edm::Service<TritonService> ts;
   const auto& server =
       ts->serverInfo(options_.model_name_, params.getUntrackedParameter<std::string>("preferredServer"));
   serverType_ = server.type;
   if (verbose_)
     edm::LogInfo(fullDebugName_) << "Using server: " << server.url;
   //enforce sync mode for fallback CPU server to avoid contention
   //todo: could enforce async mode otherwise (unless mode was specified by user?)
   if (serverType_ == TritonServerType::LocalCPU)
     setMode(SonicMode::Sync);
 
   //connect to the server
   TRITON_THROW_IF_ERROR(
       tc::InferenceServerGrpcClient::Create(&client_, server.url, false, server.useSsl, server.sslOptions),
       "TritonClient(): unable to create inference context");
 
   //set options
   options_.model_version_ = params.getParameter<std::string>("modelVersion");
   //convert seconds to microseconds
   options_.client_timeout_ = params.getUntrackedParameter<unsigned>("timeout") * 1e6;
 
   //config needed for batch size
   inference::ModelConfigResponse modelConfigResponse;
   TRITON_THROW_IF_ERROR(client_->ModelConfig(&modelConfigResponse, options_.model_name_, options_.model_version_),
                         "TritonClient(): unable to get model config");
   inference::ModelConfig modelConfig(modelConfigResponse.config());
 
   //check batch size limitations (after i/o setup)
   //triton uses max batch size = 0 to denote a model that does not support batching
   //but for models that do support batching, a given event may set batch size 0 to indicate no valid input is present
   //so set the local max to 1 and keep track of "no batch" case
   maxBatchSize_ = modelConfig.max_batch_size();
   noBatch_ = maxBatchSize_ == 0;
   maxBatchSize_ = std::max(1u, maxBatchSize_);
 
   //get model info
   inference::ModelMetadataResponse modelMetadata;
   TRITON_THROW_IF_ERROR(client_->ModelMetadata(&modelMetadata, options_.model_name_, options_.model_version_),
                         "TritonClient(): unable to get model metadata");
 
   //get input and output (which know their sizes)
   const auto& nicInputs = modelMetadata.inputs();
   const auto& nicOutputs = modelMetadata.outputs();
 
   //report all model errors at once
   std::stringstream msg;
   std::string msg_str;
 
   //currently no use case is foreseen for a model with zero inputs or outputs
   if (nicInputs.empty())
     msg << "Model on server appears malformed (zero inputs)\n";
 
   if (nicOutputs.empty())
     msg << "Model on server appears malformed (zero outputs)\n";
 
   //stop if errors
   msg_str = msg.str();
   if (!msg_str.empty())
     throw cms::Exception("ModelErrors") << msg_str;
 
   //setup input map
   std::stringstream io_msg;
   if (verbose_)
     io_msg << "Model inputs: "
            << "\n";
   inputsTriton_.reserve(nicInputs.size());
   for (const auto& nicInput : nicInputs) {
     const auto& iname = nicInput.name();
     auto [curr_itr, success] = input_.emplace(std::piecewise_construct,
                                               std::forward_as_tuple(iname),
                                               std::forward_as_tuple(iname, nicInput, this, ts->pid()));
     auto& curr_input = curr_itr->second;
     inputsTriton_.push_back(curr_input.data());
     if (verbose_) {
       io_msg << "  " << iname << " (" << curr_input.dname() << ", " << curr_input.byteSize()
              << " b) : " << triton_utils::printColl(curr_input.shape()) << "\n";
     }
   }
 
   //allow selecting only some outputs from server
   const auto& v_outputs = params.getUntrackedParameter<std::vector<std::string>>("outputs");
   std::unordered_set s_outputs(v_outputs.begin(), v_outputs.end());
 
   //setup output map
   if (verbose_)
     io_msg << "Model outputs: "
            << "\n";
   outputsTriton_.reserve(nicOutputs.size());
   for (const auto& nicOutput : nicOutputs) {
     const auto& oname = nicOutput.name();
     if (!s_outputs.empty() and s_outputs.find(oname) == s_outputs.end())
       continue;
     auto [curr_itr, success] = output_.emplace(std::piecewise_construct,
                                                std::forward_as_tuple(oname),
                                                std::forward_as_tuple(oname, nicOutput, this, ts->pid()));
     auto& curr_output = curr_itr->second;
     outputsTriton_.push_back(curr_output.data());
     if (verbose_) {
       io_msg << "  " << oname << " (" << curr_output.dname() << ", " << curr_output.byteSize()
              << " b) : " << triton_utils::printColl(curr_output.shape()) << "\n";
     }
     if (!s_outputs.empty())
       s_outputs.erase(oname);
   }
 
   //check if any requested outputs were not available
   if (!s_outputs.empty())
     throw cms::Exception("MissingOutput")
         << "Some requested outputs were not available on the server: " << triton_utils::printColl(s_outputs);
 
   //propagate batch size to inputs and outputs
   setBatchSize(1);
 
   //print model info
   std::stringstream model_msg;
   if (verbose_) {
     model_msg << "Model name: " << options_.model_name_ << "\n"
               << "Model version: " << options_.model_version_ << "\n"
               << "Model max batch size: " << (noBatch_ ? 0 : maxBatchSize_) << "\n";
     edm::LogInfo(fullDebugName_) << model_msg.str() << io_msg.str();
   }
 }

◆ ~TritonClient()

TritonClient::~TritonClient ( )

override

Definition at line 169 of file TritonClient.cc.

References SonicClientTypes< TritonInputMap, TritonOutputMap >::input_, and SonicClientTypes< TritonInputMap, TritonOutputMap >::output_.

                             {
   //by default: members of this class destroyed before members of base class
   //in shared memory case, TritonMemResource (member of TritonData) unregisters from client_ in its destructor
   //but input/output objects are member of base class, so destroyed after client_ (member of this class)
   //therefore, clear the maps here
   input_.clear();
   output_.clear();
 }

Member Function Documentation

◆ batchSize()

unsigned TritonClient::batchSize ( ) const

inline

Definition at line 39 of file TritonClient.h.

References batchSize_.

39 { return batchSize_; }

TritonClient::batchSize_

unsigned batchSize_

Definition: TritonClient.h:66

◆ client()

auto TritonClient::client ( )

inlineprivate

Definition at line 87 of file TritonClient.h.

References client_.

87 { return client_.get(); }

TritonClient::client_

std::unique_ptr< triton::client::InferenceServerGrpcClient > client_

Definition: TritonClient.h:78

◆ evaluate()

void TritonClient::evaluate ( )

overrideprotectedvirtual

Implements SonicClientBase.

Definition at line 242 of file TritonClient.cc.

References Async, batchSize_, client_, compressionAlgo_, SonicClientBase::finish(), getResults(), getServerSideStatus(), handle_exception(), headers_, inputsTriton_, SonicClientBase::mode_, options_, SonicClientTypes< TritonInputMap, TritonOutputMap >::output_, outputsTriton_, reportServerSideStats(), mysort::results, dqmMemoryStats::stats, summarizeEdmComparisonLogfiles::success, summarizeServerStats(), TRITON_THROW_IF_ERROR, and verbose().

                             {
   //in case there is nothing to process
   if (batchSize_ == 0) {
     finish(true);
     return;
   }
 
   //set up shared memory for output
   auto success = handle_exception([&]() {
     for (auto& element : output_) {
       element.second.prepare();
     }
   });
   if (!success)
     return;
 
   // Get the status of the server prior to the request being made.
   inference::ModelStatistics start_status;
   success = handle_exception([&]() {
     if (verbose())
       start_status = getServerSideStatus();
   });
   if (!success)
     return;
 
   if (mode_ == SonicMode::Async) {
     //non-blocking call
     success = handle_exception([&]() {
       TRITON_THROW_IF_ERROR(
           client_->AsyncInfer(
               [start_status, this](tc::InferResult* results) {
                 //get results
                 std::shared_ptr<tc::InferResult> results_ptr(results);
                 auto success = handle_exception(
                     [&]() { TRITON_THROW_IF_ERROR(results_ptr->RequestStatus(), "evaluate(): unable to get result"); });
                 if (!success)
                   return;
 
                 if (verbose()) {
                   inference::ModelStatistics end_status;
                   success = handle_exception([&]() { end_status = getServerSideStatus(); });
                   if (!success)
                     return;
 
                   const auto& stats = summarizeServerStats(start_status, end_status);
                   reportServerSideStats(stats);
                 }
 
                 //check result
                 success = handle_exception([&]() { getResults(results_ptr); });
                 if (!success)
                   return;
 
                 //finish
                 finish(true);
               },
               options_,
               inputsTriton_,
               outputsTriton_,
               headers_,
               compressionAlgo_),
           "evaluate(): unable to launch async run");
     });
     if (!success)
       return;
   } else {
     //blocking call
     tc::InferResult* results;
     success = handle_exception([&]() {
       TRITON_THROW_IF_ERROR(
           client_->Infer(&results, options_, inputsTriton_, outputsTriton_, headers_, compressionAlgo_),
           "evaluate(): unable to run and/or get result");
     });
     if (!success)
       return;
 
     if (verbose()) {
       inference::ModelStatistics end_status;
       success = handle_exception([&]() { end_status = getServerSideStatus(); });
       if (!success)
         return;
 
       const auto& stats = summarizeServerStats(start_status, end_status);
       reportServerSideStats(stats);
     }
 
     std::shared_ptr<tc::InferResult> results_ptr(results);
     success = handle_exception([&]() { getResults(results_ptr); });
     if (!success)
       return;
 
     finish(true);
   }
 }

◆ fillPSetDescription()

void TritonClient::fillPSetDescription ( edm::ParameterSetDescription & iDesc )

static

Definition at line 406 of file TritonClient.cc.

References edm::ParameterSetDescription::add(), edm::ParameterSetDescription::addUntracked(), SonicClientBase::fillBasePSetDescription(), and AlCaHLTBitMon_QueryRunRegistry::string.

Referenced by DeepMETSonicProducer::fillDescriptions(), ParticleNetSonicJetTagsProducer::fillDescriptions(), SCEnergyCorrectorDRNProducer::fillDescriptions(), and DRNCorrectionProducerT< T >::fillDescriptions().

                                                                         {
   edm::ParameterSetDescription descClient;
   fillBasePSetDescription(descClient);
   descClient.add<std::string>("modelName");
   descClient.add<std::string>("modelVersion", "");
   descClient.add<edm::FileInPath>("modelConfigPath");
   //server parameters should not affect the physics results
   descClient.addUntracked<std::string>("preferredServer", "");
   descClient.addUntracked<unsigned>("timeout");
   descClient.addUntracked<bool>("useSharedMemory", true);
   descClient.addUntracked<std::string>("compression", "");
   descClient.addUntracked<std::vector<std::string>>("outputs", {});
   iDesc.add<edm::ParameterSetDescription>("Client", descClient);
 }

◆ getResults()

void TritonClient::getResults ( std::shared_ptr< triton::client::InferResult > results )

protected

Definition at line 225 of file TritonClient.cc.

References noBatch_, heppy_report::oname, SonicClientTypes< TritonInputMap, TritonOutputMap >::output_, mysort::results, and TRITON_THROW_IF_ERROR.

Referenced by evaluate().

                                                                   {
   for (auto& [oname, output] : output_) {
     //set shape here before output becomes const
     if (output.variableDims()) {
       std::vector<int64_t> tmp_shape;
       TRITON_THROW_IF_ERROR(results->Shape(oname, &tmp_shape), "getResults(): unable to get output shape for " + oname);
       if (!noBatch_)
         tmp_shape.erase(tmp_shape.begin());
       output.setShape(tmp_shape);
       output.computeSizes();
     }
     //extend lifetime
     output.setResult(results);
   }
 }

◆ getServerSideStatus()

inference::ModelStatistics TritonClient::getServerSideStatus ( ) const

protected

Definition at line 395 of file TritonClient.cc.

References client_, options_, TRITON_THROW_IF_ERROR, and verbose_.

Referenced by evaluate().

                                                                  {
   if (verbose_) {
     inference::ModelStatisticsResponse resp;
     TRITON_THROW_IF_ERROR(client_->ModelInferenceStatistics(&resp, options_.model_name_, options_.model_version_),
                           "getServerSideStatus(): unable to get model statistics");
     return *(resp.model_stats().begin());
   }
   return inference::ModelStatistics{};
 }

◆ handle_exception()

template<typename F >

bool TritonClient::handle_exception ( F && call )

protected

Definition at line 206 of file TritonClient.cc.

References CMS_SA_ALLOW, MillePedeFileConverter_cfg::e, and SonicClientBase::finish().

Referenced by evaluate().

                                             {
   //caught exceptions will be propagated to edm::WaitingTaskWithArenaHolder
   CMS_SA_ALLOW try {
     call();
     return true;
   }
   //TritonExceptions are intended/expected to be recoverable, i.e. retries should be allowed
   catch (TritonException& e) {
     e.convertToWarning();
     finish(false);
     return false;
   }
   //other exceptions are not: execution should stop if they are encountered
   catch (...) {
     finish(false, std::current_exception());
     return false;
   }
 }

◆ noBatch()

bool TritonClient::noBatch ( ) const

inline

Definition at line 45 of file TritonClient.h.

References noBatch_.

45 { return noBatch_; }

TritonClient::noBatch_

bool noBatch_

Definition: TritonClient.h:67

◆ reportServerSideStats()

void TritonClient::reportServerSideStats ( const ServerSideStats & stats ) const

protected

Definition at line 337 of file TritonClient.cc.

References submitPVResolutionJobs::count, SonicClientBase::debugName_, SonicClientBase::fullDebugName_, mps_check::msg, and dqmMemoryStats::stats.

Referenced by evaluate().

                                                                                        {
   std::stringstream msg;
 
   // https://github.com/triton-inference-server/server/blob/v2.3.0/src/clients/c++/perf_client/inference_profiler.cc
   const uint64_t count = stats.success_count_;
   msg << "  Inference count: " << stats.inference_count_ << "\n";
   msg << "  Execution count: " << stats.execution_count_ << "\n";
   msg << "  Successful request count: " << count << "\n";
 
   if (count > 0) {
     auto get_avg_us = [count](uint64_t tval) {
       constexpr uint64_t us_to_ns = 1000;
       return tval / us_to_ns / count;
     };
 
     const uint64_t cumm_avg_us = get_avg_us(stats.cumm_time_ns_);
     const uint64_t queue_avg_us = get_avg_us(stats.queue_time_ns_);
     const uint64_t compute_input_avg_us = get_avg_us(stats.compute_input_time_ns_);
     const uint64_t compute_infer_avg_us = get_avg_us(stats.compute_infer_time_ns_);
     const uint64_t compute_output_avg_us = get_avg_us(stats.compute_output_time_ns_);
     const uint64_t compute_avg_us = compute_input_avg_us + compute_infer_avg_us + compute_output_avg_us;
     const uint64_t overhead =
         (cumm_avg_us > queue_avg_us + compute_avg_us) ? (cumm_avg_us - queue_avg_us - compute_avg_us) : 0;
 
     msg << "  Avg request latency: " << cumm_avg_us << " usec"
         << "\n"
         << "  (overhead " << overhead << " usec + "
         << "queue " << queue_avg_us << " usec + "
         << "compute input " << compute_input_avg_us << " usec + "
         << "compute infer " << compute_infer_avg_us << " usec + "
         << "compute output " << compute_output_avg_us << " usec)" << std::endl;
   }
 
   if (!debugName_.empty())
     edm::LogInfo(fullDebugName_) << msg.str();
 }

◆ reset()

void TritonClient::reset ( void )

overridevirtual

Reimplemented from SonicClientBase.

Definition at line 196 of file TritonClient.cc.

References SonicClientTypes< TritonInputMap, TritonOutputMap >::input_, and SonicClientTypes< TritonInputMap, TritonOutputMap >::output_.

                          {
   for (auto& element : input_) {
     element.second.reset();
   }
   for (auto& element : output_) {
     element.second.reset();
   }
 }

◆ serverType()

TritonServerType TritonClient::serverType ( ) const

inline

Definition at line 46 of file TritonClient.h.

References serverType_.

46 { return serverType_; }

TritonClient::serverType_

TritonServerType serverType_

Definition: TritonClient.h:70

◆ setBatchSize()

bool TritonClient::setBatchSize ( unsigned bsize )

Definition at line 178 of file TritonClient.cc.

References batchSize_, SonicClientBase::fullDebugName_, SonicClientTypes< TritonInputMap, TritonOutputMap >::input_, maxBatchSize_, and SonicClientTypes< TritonInputMap, TritonOutputMap >::output_.

Referenced by TritonClient().

                                               {
   if (bsize > maxBatchSize_) {
     edm::LogWarning(fullDebugName_) << "Requested batch size " << bsize << " exceeds server-specified max batch size "
                                     << maxBatchSize_ << ". Batch size will remain as" << batchSize_;
     return false;
   } else {
     batchSize_ = bsize;
     //set for input and output
     for (auto& element : input_) {
       element.second.setBatchSize(bsize);
     }
     for (auto& element : output_) {
       element.second.setBatchSize(bsize);
     }
     return true;
   }
 }

◆ setUseSharedMemory()

void TritonClient::setUseSharedMemory ( bool useShm )

inline

Definition at line 42 of file TritonClient.h.

References useSharedMemory_.

42 { useSharedMemory_ = useShm; }

TritonClient::useSharedMemory_

bool useSharedMemory_

Definition: TritonClient.h:69

◆ summarizeServerStats()

TritonClient::ServerSideStats TritonClient::summarizeServerStats	(	const inference::ModelStatistics &	start_status,
		const inference::ModelStatistics &	end_status
	)		const

protected

Definition at line 374 of file TritonClient.cc.

References TritonClient::ServerSideStats::compute_infer_time_ns_, TritonClient::ServerSideStats::compute_input_time_ns_, TritonClient::ServerSideStats::compute_output_time_ns_, TritonClient::ServerSideStats::cumm_time_ns_, TritonClient::ServerSideStats::execution_count_, TritonClient::ServerSideStats::inference_count_, TritonClient::ServerSideStats::queue_time_ns_, and TritonClient::ServerSideStats::success_count_.

Referenced by evaluate().

                                                                                                                    {
   TritonClient::ServerSideStats server_stats;
 
   server_stats.inference_count_ = end_status.inference_count() - start_status.inference_count();
   server_stats.execution_count_ = end_status.execution_count() - start_status.execution_count();
   server_stats.success_count_ =
       end_status.inference_stats().success().count() - start_status.inference_stats().success().count();
   server_stats.cumm_time_ns_ =
       end_status.inference_stats().success().ns() - start_status.inference_stats().success().ns();
   server_stats.queue_time_ns_ = end_status.inference_stats().queue().ns() - start_status.inference_stats().queue().ns();
   server_stats.compute_input_time_ns_ =
       end_status.inference_stats().compute_input().ns() - start_status.inference_stats().compute_input().ns();
   server_stats.compute_infer_time_ns_ =
       end_status.inference_stats().compute_infer().ns() - start_status.inference_stats().compute_infer().ns();
   server_stats.compute_output_time_ns_ =
       end_status.inference_stats().compute_output().ns() - start_status.inference_stats().compute_output().ns();
 
   return server_stats;
 }

◆ useSharedMemory()

bool TritonClient::useSharedMemory ( ) const

inline

Definition at line 41 of file TritonClient.h.

References useSharedMemory_.

41 { return useSharedMemory_; }

TritonClient::useSharedMemory_

bool useSharedMemory_

Definition: TritonClient.h:69

◆ verbose()

bool TritonClient::verbose ( ) const

inline

Definition at line 40 of file TritonClient.h.

References verbose_.

Referenced by evaluate().

40 { return verbose_; }

TritonClient::verbose_

bool verbose_

Definition: TritonClient.h:68

Member Data Documentation

◆ batchSize_

unsigned TritonClient::batchSize_

protected

Definition at line 66 of file TritonClient.h.

Referenced by batchSize(), evaluate(), and setBatchSize().

◆ client_

std::unique_ptr<triton::client::InferenceServerGrpcClient> TritonClient::client_

protected

Definition at line 78 of file TritonClient.h.

Referenced by client(), evaluate(), getServerSideStatus(), and TritonClient().

◆ compressionAlgo_

grpc_compression_algorithm TritonClient::compressionAlgo_

protected

Definition at line 71 of file TritonClient.h.

Referenced by evaluate().

◆ headers_

triton::client::Headers TritonClient::headers_

protected

Definition at line 72 of file TritonClient.h.

Referenced by evaluate().

◆ inputsTriton_

std::vector<triton::client::InferInput*> TritonClient::inputsTriton_

protected

Definition at line 75 of file TritonClient.h.

Referenced by evaluate(), and TritonClient().

◆ maxBatchSize_

unsigned TritonClient::maxBatchSize_

protected

Definition at line 65 of file TritonClient.h.

Referenced by setBatchSize(), and TritonClient().

◆ noBatch_

bool TritonClient::noBatch_

protected

Definition at line 67 of file TritonClient.h.

Referenced by getResults(), noBatch(), and TritonClient().

◆ options_

triton::client::InferOptions TritonClient::options_

protected

Definition at line 80 of file TritonClient.h.

Referenced by batchmanager.BatchManager::CheckBatchScript(), evaluate(), getServerSideStatus(), batchmanager.BatchManager::ManageOutputDir(), batchmanager.BatchManager::ParseOptions(), batchmanager.BatchManager::SubmitJob(), batchmanager.BatchManager::SubmitJobs(), and TritonClient().

◆ outputsTriton_

std::vector<const triton::client::InferRequestedOutput*> TritonClient::outputsTriton_

protected

Definition at line 76 of file TritonClient.h.

Referenced by evaluate(), and TritonClient().

◆ serverType_

TritonServerType TritonClient::serverType_

protected

Definition at line 70 of file TritonClient.h.

Referenced by serverType(), and TritonClient().

◆ TritonInputData

friend TritonClient::TritonInputData

private

Definition at line 83 of file TritonClient.h.

◆ TritonOutputData

friend TritonClient::TritonOutputData

private

Definition at line 84 of file TritonClient.h.

◆ useSharedMemory_

bool TritonClient::useSharedMemory_

protected

Definition at line 69 of file TritonClient.h.

Referenced by setUseSharedMemory(), and useSharedMemory().

◆ verbose_

bool TritonClient::verbose_

protected

Definition at line 68 of file TritonClient.h.

Referenced by getServerSideStatus(), TritonClient(), and verbose().

Classes

Public Member Functions

Static Public Member Functions

Protected Member Functions

Protected Attributes

Private Member Functions

Private Attributes

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

◆ TritonClient()

◆ ~TritonClient()

Member Function Documentation

◆ batchSize()

◆ client()

◆ evaluate()

◆ fillPSetDescription()

◆ getResults()

◆ getServerSideStatus()

◆ handle_exception()

◆ noBatch()

◆ reportServerSideStats()

◆ reset()

◆ serverType()

◆ setBatchSize()

◆ setUseSharedMemory()

◆ summarizeServerStats()

◆ useSharedMemory()

◆ verbose()

Member Data Documentation

◆ batchSize_

◆ client_

◆ compressionAlgo_

◆ headers_

◆ inputsTriton_

◆ maxBatchSize_

◆ noBatch_

◆ options_

◆ outputsTriton_

◆ serverType_

◆ TritonInputData

◆ TritonOutputData

◆ useSharedMemory_

◆ verbose_