fixes for handling OutputFormat

triton-inference-server · Jul 4, 2024 · 60b658a · 60b658a
1 parent f0a87f0
commit 60b658a
Show file tree

Hide file tree

Showing 2 changed files with 8 additions and 3 deletions.
diff --git a/src/c++/perf_analyzer/genai-perf/genai_perf/profile_data_parser/profile_data_parser.py b/src/c++/perf_analyzer/genai-perf/genai_perf/profile_data_parser/profile_data_parser.py
@@ -59,7 +59,11 @@ def _get_profile_metadata(self, data: dict) -> None:
         self._service_kind = data["service_kind"]
         if self._service_kind == "openai":
             if data["endpoint"] == "v1/chat/completions":
-                self._response_format = ResponseFormat.OPENAI_CHAT_COMPLETIONS
+                self._response_format = (
+                    ResponseFormat.OPENAI_VISION
+                    if self.output_format == OutputFormat.OPENAI_VISION
+                    else ResponseFormat.OPENAI_CHAT_COMPLETIONS
+                )
             elif data["endpoint"] == "v1/completions":
                 self._response_format = ResponseFormat.OPENAI_COMPLETIONS
             elif data["endpoint"] == "v1/embeddings":

diff --git a/src/c++/perf_analyzer/genai-perf/tests/test_profile_data_parser.py b/src/c++/perf_analyzer/genai-perf/tests/test_profile_data_parser.py
@@ -33,6 +33,7 @@
 import pytest
 from genai_perf.metrics import Metrics
 from genai_perf.profile_data_parser import ProfileDataParser
+from genai_perf.llm_inputs.llm_inputs import OutputFormat
 
 
 def ns_to_sec(ns: int) -> Union[int, float]:
@@ -132,7 +133,7 @@ def test_embedding_profile_data(self, mock_read_write: pytest.MonkeyPatch) -> No
         * request throughputs
             - [2 / (5e-9 - 1e-9)] = [5e8]
         """
-        pd = ProfileDataParser(filename=Path("embedding_profile_export.json"))
+        pd = ProfileDataParser(filename=Path("embedding_profile_export.json"), output_format=OutputFormat.OPENAI_EMBEDDINGS)
 
         # experiment 1 statistics
         stats = pd.get_statistics(infer_mode="concurrency", load_level="10")
@@ -202,7 +203,7 @@ def test_ranking_profile_data(self, mock_read_write: pytest.MonkeyPatch) -> None
         * request throughputs
             - [2 / (5e-9 - 1e-9)] = [5e8]
         """
-        pd = ProfileDataParser(filename=Path("ranking_profile_export.json"))
+        pd = ProfileDataParser(filename=Path("ranking_profile_export.json"), output_format=OutputFormat.RANKINGS)
 
         # experiment 1 statistics
         stats = pd.get_statistics(infer_mode="concurrency", load_level="10")