fixed quick search

Author: tgerdesnv

model_analyzer/config/generate/quick_plus_concurrency_sweep_run_config_generator.py

@@ -16,7 +16,7 @@
 
 import logging
 from copy import deepcopy
-from typing import Generator, List, Optional
+from typing import Dict, Generator, List, Optional
 
 from model_analyzer.config.generate.model_profile_spec import ModelProfileSpec
 from model_analyzer.config.generate.model_variant_name_manager import (
@@ -139,7 +139,11 @@ def _sweep_concurrency_over_top_results(self) -> Generator[RunConfig, None, None
 
             for result in top_results:
                 run_config = deepcopy(result.run_config())
-                inference_load_search = InferenceLoadSearch(self._config)
+                perf_analyzer_flags = self._get_model_perf_analyzer_flags(model_name)
+
+                inference_load_search = InferenceLoadSearch(
+                    config=self._config, perf_analyzer_flags=perf_analyzer_flags
+                )
                 for concurrency in inference_load_search.search_inference_loads():
                     run_config = self._set_concurrency(run_config, concurrency)
                     yield run_config
@@ -153,3 +157,8 @@ def _set_concurrency(self, run_config: RunConfig, concurrency: int) -> RunConfig
             perf_config.update_config({"concurrency-range": concurrency})
 
         return run_config
+
+    def _get_model_perf_analyzer_flags(self, model_name: str) -> Dict:
+        for model in self._models:
+            if model_name == model.model_name():
+                return model.perf_analyzer_flags()

model_analyzer/config/generate/quick_run_config_generator.py

@@ -511,9 +511,12 @@ def _get_next_perf_analyzer_config(
 
         perf_analyzer_config.update_config_from_profile_config(model_name, self._config)
 
-        concurrency = self._calculate_concurrency(dimension_values)
+        perf_config_params = {"batch-size": 1}
+
+        if not "request-intervals" in model.perf_analyzer_flags():
+            concurrency = self._calculate_concurrency(dimension_values)
+            perf_config_params["concurrency-range"] = concurrency
 
-        perf_config_params = {"batch-size": 1, "concurrency-range": concurrency}
         perf_analyzer_config.update_config(perf_config_params)
 
         perf_analyzer_config.update_config(model.perf_analyzer_flags())
@@ -704,7 +707,11 @@ def _create_default_perf_analyzer_config(
             model_config.get_field("name"), self._config
         )
 
-        perf_config_params = {"batch-size": 1, "concurrency-range": 1}
+        perf_config_params = {"batch-size": 1}
+
+        if not "request-intervals" in model.perf_analyzer_flags():
+            perf_config_params["concurrency-range"] = 1
+
         default_perf_analyzer_config.update_config(perf_config_params)
 
         default_perf_analyzer_config.update_config(model.perf_analyzer_flags())

model_analyzer/result/inference_load_search.py

@@ -37,6 +37,7 @@ class InferenceLoadSearch:
       - Will sweep from by powers of two from min to max inference load
       - If the user specifies a constraint, the algorithm will perform a binary search
         around the boundary if the constraint is violated
+      - Will not sweep at all if custom stimulus is provided by the user (via the "request-intervals" perf analyzer flag)
 
     Invariant: It is necessary for the user to add new measurements as they are taken
     """