Rebuilding heaps after profile completes (#514)

* Redoing heaps after profile completes. Fixing incorrect config usage

* Adding story number

Author: nv-braf

model_analyzer/analyzer.py

@@ -113,6 +113,7 @@ def profile(self, client: TritonClient, gpus: List[GPUDevice], mode: str,
         logger.info("")
 
         if not self._config.skip_summary_reports:
+            self._result_manager.add_results_to_heaps()
             self._create_summary_tables(verbose)
             self._create_summary_reports(mode)
             logger.info(self._get_report_command_help_string())

model_analyzer/config/generate/quick_plus_concurrency_sweep_run_config_generator.py

@@ -106,7 +106,7 @@ def _create_quick_run_config_generator(self) -> QuickRunConfigGenerator:
 
     def _sweep_concurrency_over_top_results(self):
         top_results = self._result_manager.top_n_results(
-            n=self._config.num_top_model_configs)
+            n=self._config.num_configs_per_model)
 
         for count, result in enumerate(top_results):
             new_config = self._create_new_config_command_profile(result)

model_analyzer/result/result_manager.py

@@ -233,6 +233,16 @@ def _update_stats(statistics, result_heap, stats_key):
 
         return result_stats
 
+    # TODO: TMA-898 This is the big hammer to fix the issues that arise because we
+    # can create new measurements inside existing RCR's after the heaps have been
+    # created either when re-profiling or running quick + brute w/ concurrency sweep
+    def add_results_to_heaps(self):
+        # Delete the current heaps
+        self._per_model_sorted_results = defaultdict(ResultHeap)
+        self._across_model_sorted_results = ResultHeap()
+
+        self._add_results_to_heaps()
+
     def _init_state(self):
         """
         Sets ResultManager object managed