Add tensorrtllm_engine option to service-kind and update testing (#700)

* Add tensorrtllm_engine option to service-kind and update testing

* Add output format check for tensorrtllm_engine

src/c++/perf_analyzer/genai-perf/genai_perf/llm_inputs/llm_inputs.py

@@ -53,6 +53,7 @@ class OutputFormat(Enum):
     RANKINGS = auto()
     TENSORRTLLM = auto()
     VLLM = auto()
+    TENSORRTLLM_ENGINE = auto()
 
     def to_lowercase(self):
         return self.name.lower()

src/c++/perf_analyzer/genai-perf/genai_perf/parser.py

@@ -177,6 +177,9 @@ def _check_conditional_args(
         args = _convert_str_to_enum_entry(args, "backend", OutputFormat)
         args.output_format = args.backend
 
+    if args.service_kind == "tensorrtllm_engine":
+        args.output_format = OutputFormat.TENSORRTLLM_ENGINE
+
     # Output token distribution checks
     if args.output_tokens_mean == LlmInputs.DEFAULT_OUTPUT_TOKENS_MEAN:
         if args.output_tokens_stddev != LlmInputs.DEFAULT_OUTPUT_TOKENS_STDDEV:
@@ -268,6 +271,8 @@ def _set_artifact_paths(args: argparse.Namespace) -> argparse.Namespace:
             name += [f"{args.service_kind}-{args.endpoint_type}"]
         elif args.service_kind == "triton":
             name += [f"{args.service_kind}-{args.backend.to_lowercase()}"]
+        elif args.service_kind == "tensorrtllm_engine":
+            name += [f"{args.service_kind}"]
         else:
             raise ValueError(f"Unknown service kind '{args.service_kind}'.")
 
@@ -528,25 +533,6 @@ def _add_profile_args(parser):
 def _add_endpoint_args(parser):
     endpoint_group = parser.add_argument_group("Endpoint")
 
-    endpoint_group.add_argument(
-        "-m",
-        "--model",
-        nargs="+",
-        default=[],
-        help=f"The name of the model(s) to benchmark.",
-    )
-    endpoint_group.add_argument(
-        "--model-selection-strategy",
-        type=str,
-        choices=utils.get_enum_names(ModelSelectionStrategy),
-        default="round_robin",
-        required=False,
-        help=f"When multiple model are specified, this is how a specific model "
-        "should be assigned to a prompt.  round_robin means that ith prompt in the "
-        "list gets assigned to i mod len(models).  random means that assignment is "
-        "uniformly random",
-    )
-
     endpoint_group.add_argument(
         "--backend",
         type=str,
@@ -576,10 +562,29 @@ def _add_endpoint_args(parser):
         'server. This is only used with the "openai" service-kind.',
     )
 
+    endpoint_group.add_argument(
+        "-m",
+        "--model",
+        nargs="+",
+        default=[],
+        help=f"The name of the model(s) to benchmark.",
+    )
+    endpoint_group.add_argument(
+        "--model-selection-strategy",
+        type=str,
+        choices=utils.get_enum_names(ModelSelectionStrategy),
+        default="round_robin",
+        required=False,
+        help=f"When multiple model are specified, this is how a specific model "
+        "should be assigned to a prompt.  round_robin means that ith prompt in the "
+        "list gets assigned to i mod len(models).  random means that assignment is "
+        "uniformly random",
+    )
+
     endpoint_group.add_argument(
         "--service-kind",
         type=str,
-        choices=["triton", "openai"],
+        choices=["triton", "openai", "tensorrtllm_engine"],
         default="triton",
         required=False,
         help="The kind of service perf_analyzer will "

src/c++/perf_analyzer/genai-perf/tests/test_cli.py

@@ -203,6 +203,10 @@ def test_help_version_arguments_output_and_exit(
             (["--request-rate", "9.0"], {"request_rate": 9.0}),
             (["-s", "99.5"], {"stability_percentage": 99.5}),
             (["--service-kind", "triton"], {"service_kind": "triton"}),
+            (
+                ["--service-kind", "tensorrtllm_engine"],
+                {"service_kind": "tensorrtllm_engine"},
+            ),
             (
                 ["--service-kind", "openai", "--endpoint-type", "chat"],
                 {"service_kind": "openai", "endpoint": "v1/chat/completions"},
@@ -654,6 +658,7 @@ def test_conditional_errors(self, args, expected_output, monkeypatch, capsys):
                 OutputFormat.TENSORRTLLM,
             ),
             (["--service-kind", "triton", "--backend", "vllm"], OutputFormat.VLLM),
+            (["--service-kind", "tensorrtllm_engine"], OutputFormat.TENSORRTLLM_ENGINE),
         ],
     )
     def test_inferred_output_format(self, monkeypatch, args, expected_format):