feat: Extend histogram metrics to non-decoupled models (#464)

Author: yinggeh

src/backend_model_instance.cc

@@ -1,4 +1,4 @@
-// Copyright 2020-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// Copyright 2020-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 //
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions
@@ -192,8 +192,8 @@ TritonModelInstance::TritonModelInstance(
         model_->Server()->ResponseCacheEnabled();
     MetricModelReporter::Create(
         model_->ModelId(), model_->Version(), id, response_cache_enabled,
-        model_->IsDecoupled(), model_->Config().metric_tags(),
-        model_->Config().model_metrics(), &reporter_);
+        model_->Config().metric_tags(), model_->Config().model_metrics(),
+        &reporter_);
   }
 #endif  // TRITON_ENABLE_METRICS
 }

src/ensemble_scheduler/ensemble_scheduler.cc

@@ -1648,13 +1648,12 @@ EnsembleScheduler::EnsembleScheduler(
   }
 #endif  // TRITON_ENABLE_GPU
 
-  const bool is_decoupled = config.model_transaction_policy().decoupled();
 #ifdef TRITON_ENABLE_METRICS
   if (Metrics::Enabled()) {
     // Ensemble scheduler doesn't currently support response cache at top level.
     MetricModelReporter::Create(
         model_id, 1 /* model_version */, METRIC_REPORTER_ID_CPU,
-        false /* response_cache_enabled */, is_decoupled, config.metric_tags(),
+        false /* response_cache_enabled */, config.metric_tags(),
         config.model_metrics(), &metric_reporter_);
   }
 #endif  // TRITON_ENABLE_METRICS
@@ -1665,7 +1664,7 @@ EnsembleScheduler::EnsembleScheduler(
   info_->ensemble_name_ = config.name();
 
   // This config field is filled internally for ensemble models
-  info_->is_decoupled_ = is_decoupled;
+  info_->is_decoupled_ = config.model_transaction_policy().decoupled();
 
   // field to check if response cache enabled in the ensemble model config.
   info_->is_cache_enabled_ =

src/metric_model_reporter.cc

@@ -1,4 +1,4 @@
-// Copyright 2019-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// Copyright 2019-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 //
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions
@@ -42,8 +42,7 @@ namespace triton { namespace core {
 //
 void
 MetricReporterConfig::ParseConfig(
-    bool response_cache_enabled, bool is_decoupled,
-    const inference::ModelMetrics& model_metrics)
+    bool response_cache_enabled, const inference::ModelMetrics& model_metrics)
 {
   // Global config only for now in config map
   auto metrics_config_map = Metrics::ConfigMap();
@@ -74,7 +73,6 @@ MetricReporterConfig::ParseConfig(
 
   // Set flag to signal to stats aggregator if caching is enabled or not
   cache_enabled_ = response_cache_enabled;
-  is_decoupled_ = is_decoupled;
 
   // Override default histogram options if set in model_metrics.
   for (const auto& metric_control : model_metrics.metric_control()) {
@@ -139,7 +137,7 @@ const std::map<FailureReason, std::string>
 Status
 MetricModelReporter::Create(
     const ModelIdentifier& model_id, const int64_t model_version,
-    const int device, bool response_cache_enabled, bool is_decoupled,
+    const int device, bool response_cache_enabled,
     const triton::common::MetricTagsMap& model_tags,
     const inference::ModelMetrics& model_metrics,
     std::shared_ptr<MetricModelReporter>* metric_model_reporter)
@@ -169,23 +167,23 @@ MetricModelReporter::Create(
   }
 
   metric_model_reporter->reset(new MetricModelReporter(
-      model_id, model_version, device, response_cache_enabled, is_decoupled,
-      model_tags, model_metrics));
+      model_id, model_version, device, response_cache_enabled, model_tags,
+      model_metrics));
   reporter_map.insert({hash_labels, *metric_model_reporter});
   return Status::Success;
 }
 
 MetricModelReporter::MetricModelReporter(
     const ModelIdentifier& model_id, const int64_t model_version,
-    const int device, bool response_cache_enabled, bool is_decoupled,
+    const int device, bool response_cache_enabled,
     const triton::common::MetricTagsMap& model_tags,
     const inference::ModelMetrics& model_metrics)
 {
   std::map<std::string, std::string> labels;
   GetMetricLabels(&labels, model_id, model_version, device, model_tags);
 
   // Parse metrics config to control metric setup and behavior
-  config_.ParseConfig(response_cache_enabled, is_decoupled, model_metrics);
+  config_.ParseConfig(response_cache_enabled, model_metrics);
 
   // Initialize families and metrics
   InitializeCounters(labels);
@@ -306,12 +304,9 @@ MetricModelReporter::InitializeHistograms(
     const std::map<std::string, std::string>& labels)
 {
   // Update MetricReporterConfig::metric_map_ for new histograms.
-  // Only create response metrics if decoupled model to reduce metric output
   if (config_.latency_histograms_enabled_) {
-    if (config_.is_decoupled_) {
-      histogram_families_[kFirstResponseHistogram] =
-          &Metrics::FamilyFirstResponseDuration();
-    }
+    histogram_families_[kFirstResponseHistogram] =
+        &Metrics::FamilyFirstResponseDuration();
   }
 
   for (auto& iter : histogram_families_) {

src/metric_model_reporter.h

@@ -1,4 +1,4 @@
-// Copyright 2019-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// Copyright 2019-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 //
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions
@@ -47,7 +47,7 @@ struct MetricReporterConfig {
 #ifdef TRITON_ENABLE_METRICS
   // Parses Metrics::ConfigMap and sets fields if specified
   void ParseConfig(
-      bool response_cache_enabled, bool is_decoupled,
+      bool response_cache_enabled,
       const inference::ModelMetrics& model_metrics);
   // Parses pairs of quantiles "quantile1:error1, quantile2:error2, ..."
   // and overwrites quantiles_ field if successful.
@@ -76,8 +76,6 @@ struct MetricReporterConfig {
   // This helps handle infer_stats aggregation for summaries on cache misses.
   bool cache_enabled_ = false;
 
-  bool is_decoupled_ = false;
-
  private:
   // Maps the metric family fullname to its lookup key. This field is required
   // because the users are expected to configure metric configuration
@@ -97,7 +95,7 @@ class MetricModelReporter {
   static Status Create(
       const triton::core::ModelIdentifier& model_id,
       const int64_t model_version, const int device,
-      bool response_cache_enabled, bool is_decoupled,
+      bool response_cache_enabled,
       // FIXME: [DLIS-7497] Merge model_tags with model_metrics
       const triton::common::MetricTagsMap& model_tags,
       const inference::ModelMetrics& model_metrics,
@@ -125,7 +123,7 @@ class MetricModelReporter {
  private:
   MetricModelReporter(
       const ModelIdentifier& model_id, const int64_t model_version,
-      const int device, bool response_cache_enabled, bool is_decoupled,
+      const int device, bool response_cache_enabled,
       const triton::common::MetricTagsMap& model_tags,
       const inference::ModelMetrics& model_metrics);
 

src/model.cc

@@ -1,4 +1,4 @@
-// Copyright 2018-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// Copyright 2018-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 //
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions
@@ -135,8 +135,7 @@ Model::Init(const bool is_config_provided)
 #ifdef TRITON_ENABLE_METRICS
   MetricModelReporter::Create(
       ModelId(), Version(), METRIC_REPORTER_ID_UTILITY, ResponseCacheEnabled(),
-      IsDecoupled(), Config().metric_tags(), Config().model_metrics(),
-      &reporter_);
+      Config().metric_tags(), Config().model_metrics(), &reporter_);
 #endif  // TRITON_ENABLE_METRICS
 
   return Status::Success;

src/model.h

@@ -1,4 +1,4 @@
-// Copyright 2018-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// Copyright 2018-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 //
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions
@@ -148,12 +148,6 @@ class Model {
     return config_.response_cache().enable();
   }
 
-  // Get whether the model is decoupled.
-  bool IsDecoupled() const
-  {
-    return config_.model_transaction_policy().decoupled();
-  }
-
   // Get the number of required inputs
   size_t RequiredInputCount() const { return required_input_count_; }