Allow extra request inputs

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

@@ -86,6 +86,7 @@ def create_llm_inputs(
         add_model_name: bool = False,
         add_stream: bool = False,
         tokenizer: AutoTokenizer = DEFAULT_TOKENIZER,
+        extra_inputs: Dict = {},
     ) -> Dict:
         """
         Given an input type, input format, and output type. Output a string of LLM Inputs
@@ -109,9 +110,11 @@ def create_llm_inputs(
         length:
             Number of entries to gather
         add_model_name:
-            If true adds a model name field to each payload
+            If true, adds a model name field to each payload
         add_stream:
-            If true adds a steam field to each payload
+            If true, adds a steam field to each payload
+        extra_inputs:
+            If provided, append these inputs to every request
 
         Required Synthetic Prompt Generation Parameters
         -----------------------------------------------
@@ -164,7 +167,12 @@ def create_llm_inputs(
             )
 
         json_in_pa_format = LlmInputs._convert_generic_json_to_output_format(
-            output_format, generic_dataset_json, add_model_name, add_stream, model_name
+            output_format,
+            generic_dataset_json,
+            add_model_name,
+            add_stream,
+            model_name,
+            extra_inputs,
         )
         LlmInputs._write_json_to_file(json_in_pa_format)
 
@@ -309,24 +317,29 @@ def _convert_generic_json_to_output_format(
         add_model_name: bool,
         add_stream: bool,
         model_name: str = "",
+        extra_inputs: Dict = {},
     ) -> Dict:
         if output_format == OutputFormat.OPENAI_CHAT_COMPLETIONS:
             output_json = (
                 LlmInputs._convert_generic_json_to_openai_chat_completions_format(
-                    generic_dataset, add_model_name, add_stream, model_name
+                    generic_dataset,
+                    add_model_name,
+                    add_stream,
+                    model_name,
+                    extra_inputs,
                 )
             )
         elif output_format == OutputFormat.OPENAI_COMPLETIONS:
             output_json = LlmInputs._convert_generic_json_to_openai_completions_format(
-                generic_dataset, add_model_name, add_stream, model_name
+                generic_dataset, add_model_name, add_stream, model_name, extra_inputs
             )
         elif output_format == OutputFormat.VLLM:
             output_json = LlmInputs._convert_generic_json_to_vllm_format(
-                generic_dataset, add_model_name, add_stream, model_name
+                generic_dataset, add_model_name, add_stream, model_name, extra_inputs
             )
         elif output_format == OutputFormat.TRTLLM:
             output_json = LlmInputs._convert_generic_json_to_trtllm_format(
-                generic_dataset, add_model_name, add_stream, model_name
+                generic_dataset, add_model_name, add_stream, model_name, extra_inputs
             )
         else:
             raise GenAIPerfException(
@@ -342,6 +355,7 @@ def _convert_generic_json_to_openai_chat_completions_format(
         add_model_name: bool,
         add_stream: bool,
         model_name: str = "",
+        extra_inputs: Dict = {},
     ) -> Dict:
         # TODO (TMA-1757): Implement a way to select a role for `text_input`
         (
@@ -356,6 +370,7 @@ def _convert_generic_json_to_openai_chat_completions_format(
             add_model_name,
             add_stream,
             model_name,
+            extra_inputs,
         )
 
         return pa_json
@@ -367,6 +382,7 @@ def _convert_generic_json_to_openai_completions_format(
         add_model_name: bool,
         add_stream: bool,
         model_name: str = "",
+        extra_inputs: Dict = {},
     ) -> Dict:
         (
             system_role_headers,
@@ -381,6 +397,7 @@ def _convert_generic_json_to_openai_completions_format(
             add_model_name,
             add_stream,
             model_name,
+            extra_inputs,
         )
 
         return pa_json
@@ -392,6 +409,7 @@ def _convert_generic_json_to_vllm_format(
         add_model_name: bool,
         add_stream: bool,
         model_name: str = "",
+        extra_inputs: Dict = {},
     ) -> Dict:
         (
             system_role_headers,
@@ -407,6 +425,7 @@ def _convert_generic_json_to_vllm_format(
             add_model_name,
             add_stream,
             model_name,
+            extra_inputs,
         )
 
         return pa_json
@@ -418,6 +437,7 @@ def _convert_generic_json_to_trtllm_format(
         add_model_name: bool,
         add_stream: bool,
         model_name: str = "",
+        extra_inputs: Dict = {},
     ) -> Dict:
         (
             system_role_headers,
@@ -433,6 +453,7 @@ def _convert_generic_json_to_trtllm_format(
             add_model_name,
             add_stream,
             model_name,
+            extra_inputs,
         )
 
         return pa_json
@@ -480,6 +501,7 @@ def _populate_openai_chat_completions_output_json(
         add_model_name: bool,
         add_stream: bool,
         model_name: str = "",
+        extra_inputs: Dict = {},
     ) -> Dict:
         pa_json = LlmInputs._create_empty_openai_pa_json()
 
@@ -497,7 +519,7 @@ def _populate_openai_chat_completions_output_json(
                 )
 
             pa_json = LlmInputs._add_optional_tags_to_openai_json(
-                pa_json, index, add_model_name, add_stream, model_name
+                pa_json, index, add_model_name, add_stream, model_name, extra_inputs
             )
 
         return pa_json
@@ -512,6 +534,7 @@ def _populate_openai_completions_output_json(
         add_model_name: bool,
         add_stream: bool,
         model_name: str = "",
+        extra_inputs: Dict = {},
     ) -> Dict:
         pa_json = LlmInputs._create_empty_openai_pa_json()
 
@@ -531,7 +554,7 @@ def _populate_openai_completions_output_json(
                 pa_json = LlmInputs._add_new_prompt_to_json(pa_json, index, new_prompt)
 
             pa_json = LlmInputs._add_optional_tags_to_openai_json(
-                pa_json, index, add_model_name, add_stream, model_name
+                pa_json, index, add_model_name, add_stream, model_name, extra_inputs
             )
 
         return pa_json
@@ -546,6 +569,7 @@ def _populate_vllm_output_json(
         add_model_name: bool,
         add_stream: bool,
         model_name: str = "",
+        extra_inputs: Dict = {},
     ) -> Dict:
         pa_json = LlmInputs._create_empty_vllm_pa_json()
 
@@ -566,7 +590,7 @@ def _populate_vllm_output_json(
                 )
 
             pa_json = LlmInputs._add_optional_tags_to_vllm_json(
-                pa_json, index, add_model_name, add_stream, model_name
+                pa_json, index, add_model_name, add_stream, model_name, extra_inputs
             )
 
         return pa_json
@@ -581,6 +605,7 @@ def _populate_trtllm_output_json(
         add_model_name: bool,
         add_stream: bool,
         model_name: str = "",
+        extra_inputs: Dict = {},
     ) -> Dict:
         pa_json = LlmInputs._create_empty_trtllm_pa_json()
 
@@ -602,7 +627,7 @@ def _populate_trtllm_output_json(
 
             pa_json = LlmInputs._add_required_tags_to_trtllm_json(pa_json, index)
             pa_json = LlmInputs._add_optional_tags_to_trtllm_json(
-                pa_json, index, add_model_name, add_stream, model_name
+                pa_json, index, add_model_name, add_stream, model_name, extra_inputs
             )
 
         return pa_json
@@ -737,11 +762,14 @@ def _add_optional_tags_to_openai_json(
         add_model_name: bool,
         add_stream: bool,
         model_name: str = "",
+        extra_inputs: Dict = {},
     ) -> Dict:
         if add_model_name:
             pa_json["data"][index]["payload"][0]["model"] = model_name
         if add_stream:
             pa_json["data"][index]["payload"][0]["stream"] = True
+        for key, value in extra_inputs.items():
+            pa_json["data"][index]["payload"][0][key] = value
 
         return pa_json
 
@@ -753,11 +781,14 @@ def _add_optional_tags_to_vllm_json(
         add_model_name: bool,
         add_stream: bool,
         model_name: str = "",
+        extra_inputs: Dict = {},
     ) -> Dict:
         if add_model_name:
             pa_json["data"][index]["model"] = model_name
         if add_stream:
             pa_json["data"][index]["stream"] = [True]
+        for key, value in extra_inputs.items():
+            pa_json["data"][index][key] = value
 
         return pa_json
 
@@ -769,11 +800,14 @@ def _add_optional_tags_to_trtllm_json(
         add_model_name: bool,
         add_stream: bool,
         model_name: str = "",
+        extra_inputs: Dict = {},
     ) -> Dict:
         if add_model_name:
             pa_json["data"][index]["model"] = model_name
         if add_stream:
             pa_json["data"][index]["stream"] = [True]
+        for key, value in extra_inputs.items():
+            pa_json["data"][index][key] = value
 
         return pa_json
 

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

@@ -61,6 +61,7 @@ def generate_inputs(args: ArgumentParser, tokenizer: AutoTokenizer) -> None:
         add_model_name=add_model_name,
         add_stream=args.streaming,
         tokenizer=tokenizer,
+        extra_inputs=parser.get_extra_inputs_as_dict(args),
     )
 
 

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

@@ -147,6 +147,13 @@ def _add_input_args(parser):
         help="The seed used to generate random values.",
     )
 
+    parser.add_argument(
+        "--synthetic-extra-inputs",
+        action="append",
+        help="Provide additional inputs to include with every request when prompt-source is synthetic. "
+        "You can repeat this flag for multiple inputs. Inputs should be in a key:value format.",
+    )
+
     input_group.add_argument(
         "--synthetic-requested-output-tokens",
         type=int,
@@ -318,6 +325,21 @@ def _add_other_args(parser):
     )
 
 
+def get_extra_inputs_as_dict(args: argparse.ArgumentParser) -> dict:
+    request_inputs = {}
+    if hasattr(args, "synthetic_extra_inputs"):
+        for input_str in args.synthetic_extra_inputs:
+            try:
+                key, value = input_str.split(":", 1)
+                request_inputs[key] = value
+            except ValueError:
+                args.error(
+                    f"Invalid input format for --synthetic--extra-inputs: {input_str}"
+                    "Expected input format: 'key:value'"
+                )
+    return request_inputs
+
+
 ### Entrypoint ###
 
 

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

@@ -57,6 +57,7 @@ def build_cmd(args, extra_args):
             "input_format",
             "model",
             "backend",
+            "extra_inputs",
             "output_format",
             # The 'streaming' passed in to this script is to determine if the
             # LLM response should be streaming. That is different than the

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

@@ -75,6 +75,19 @@ def test_help_version_arguments_output_and_exit(
                 ["--endpoint", "v1/chat/completions"],
                 {"endpoint": "v1/chat/completions"},
             ),
+            (
+                ["--synthetic-extra-inputs", "random_key:random_value"],
+                {"synthetic_extra_inputs": ["random_key:random_value"]},
+            ),
+            (
+                [
+                    "--synthetic-extra-inputs",
+                    "random_key:5",
+                    "--synthetic-extra-inputs",
+                    "another_random_key:6",
+                ],
+                {"synthetic_extra_inputs": ["random_key:5", "another_random_key:6"]},
+            ),
             (
                 ["--synthetic-requested-output-tokens", "5"],
                 {"synthetic_requested_output_tokens": 5},

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

@@ -358,3 +358,60 @@ def test_synthetic_to_openai_completions(self, default_tokenizer):
 
         assert pa_json is not None
         assert len(pa_json["data"]) == 5
+
+    @pytest.mark.parametrize(
+        "output_format",
+        [
+            (OutputFormat.OPENAI_CHAT_COMPLETIONS),
+            (OutputFormat.OPENAI_COMPLETIONS),
+            (OutputFormat.TRTLLM),
+            (OutputFormat.VLLM),
+        ],
+    )
+    def test_llm_inputs_extra_inputs(self, default_tokenizer, output_format) -> None:
+        # Simulate --request-input arguments
+        request_inputs = {"additional_key": "additional_value"}
+
+        # Generate input data with the additional request inputs
+        pa_json = LlmInputs.create_llm_inputs(
+            input_type=PromptSource.SYNTHETIC,
+            output_format=output_format,
+            num_of_output_prompts=5,  # Generate a small number of prompts for the test
+            add_model_name=False,
+            add_stream=True,
+            tokenizer=default_tokenizer,
+            extra_inputs=request_inputs,  # Pass the simulated --request-input arguments here
+        )
+
+        assert len(pa_json["data"]) == 5
+
+        if (
+            output_format == OutputFormat.OPENAI_CHAT_COMPLETIONS
+            or output_format == OutputFormat.OPENAI_COMPLETIONS
+        ):
+            # Verify that each entry in the generated JSON includes the additional key-value pairs
+            for entry in pa_json.get("data", []):
+                assert "payload" in entry, "Payload is missing in the request"
+                payload = entry["payload"]
+                for item in payload:
+                    assert (
+                        "additional_key" in item
+                    ), "The additional_key is not present in the request"
+                    assert (
+                        item["additional_key"] == "additional_value"
+                    ), "The value of additional_key is incorrect"
+        elif output_format == OutputFormat.TRTLLM or output_format == OutputFormat.VLLM:
+            # Verify that each entry in the generated JSON includes the additional key-value pairs
+            for entry in pa_json.get("data", []):
+                assert (
+                    "additional_key" in entry
+                ), "The additional_key is not present in the request"
+                assert (
+                    entry["additional_key"] == "additional_value"
+                ), "The value of additional_key is incorrect"
+        else:
+            assert False, f"Unsupported output format: {output_format}"
+
+    print(
+        "Test passed: --request-input key:value is correctly added to every request in input-data.json"
+    )