Accept extra request inputs in GenAi-Perf (#552)

Author: the-david-oy

src/c++/perf_analyzer/genai-perf/README.md

@@ -131,6 +131,11 @@ options:
 * `--num-prompts`: The number of unique prompts to generate.
 * `--dataset`: HuggingFace dataset to use for benchmarking.
 
+You can optionally set additional model inputs with the following option:
+* `--extra-inputs {input_name}:{value}`: An additional input for use with the model with a singular value,
+such as `stream:true` or `max_tokens:5`. This flag can be repeated to supply multiple extra inputs.
+
+
 # Metrics
 
 GenAI-Perf collects a diverse set of metrics that captures the performance of
@@ -241,6 +246,13 @@ both infer per second and latency.
 
 Enables the use of the streaming API.
 
+##### `--extra-inputs`
+
+Provides an additional input for use with the model with a singular value,
+such as `stream:true` or `max_tokens:5`. This flag can be repeated to supply multiple extra inputs.
+
+
+
 ##### `--endpoint {v1/completions,v1/chat/completions}`
 
 Describes what endpoint to send requests to on the server. This is required when

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

@@ -45,6 +45,11 @@ def generate_inputs(args: ArgumentParser, tokenizer: AutoTokenizer) -> None:
     input_file_name = ""
     # TODO (TMA-1759): review if add_model_name is always true
     add_model_name = True
+    try:
+        extra_input_dict = parser.get_extra_inputs_as_dict(args)
+    except ValueError as e:
+        raise GenAIPerfException(e)
+
     LlmInputs.create_llm_inputs(
         input_type=args.prompt_source,
         output_format=args.output_format,
@@ -61,6 +66,7 @@ def generate_inputs(args: ArgumentParser, tokenizer: AutoTokenizer) -> None:
         add_model_name=add_model_name,
         add_stream=args.streaming,
         tokenizer=tokenizer,
+        extra_inputs=extra_input_dict,
     )
 
 

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

@@ -55,8 +55,8 @@ def _check_conditional_args(
         elif args.endpoint == "v1/completions":
             args.output_format = OutputFormat.OPENAI_COMPLETIONS
     elif args.endpoint is not None:
-        logger.warning(
-            "The --endpoint option is ignored when not using the 'openai' service-kind."
+        parser.error(
+            "The --endpoint option should only be used when using the 'openai' service-kind."
         )
     if args.service_kind == "triton":
         args = _convert_str_to_enum_entry(args, "backend", OutputFormat)
@@ -113,6 +113,13 @@ def handler(args, extra_args):
 def _add_input_args(parser):
     input_group = parser.add_argument_group("Input")
 
+    input_group.add_argument(
+        "--extra-inputs",
+        action="append",
+        help="Provide additional inputs to include with every request. "
+        "You can repeat this flag for multiple inputs. Inputs should be in an input_name:value format.",
+    )
+
     input_group.add_argument(
         "--input-dataset",
         type=str.lower,
@@ -242,9 +249,8 @@ def _add_endpoint_args(parser):
         type=str,
         choices=["v1/chat/completions", "v1/completions"],
         required=False,
-        help="The endpoint to send requests to on the "
-        'server. This is required when using the "openai" service-kind. '
-        "This is ignored in other cases.",
+        help=f"The endpoint to send requests to on the "
+        'server. This is only used with the "openai" service-kind. ',
     )
 
     endpoint_group.add_argument(
@@ -318,6 +324,46 @@ def _add_other_args(parser):
     )
 
 
+def get_extra_inputs_as_dict(args: argparse.ArgumentParser) -> dict:
+    request_inputs = {}
+    if hasattr(args, "extra_inputs"):
+        for input_str in args.extra_inputs:
+            semicolon_count = input_str.count(":")
+            if semicolon_count != 1:
+                raise ValueError(
+                    f"Invalid input format for --extra-inputs: {input_str}\n"
+                    "Expected input format: 'input_name:value'"
+                )
+            input_name, value = input_str.split(":", 1)
+
+            if not input_name or not value:
+                raise ValueError(
+                    f"Input name or value is empty in --extra-inputs: {input_str}\n"
+                    "Expected input format: 'input_name:value'"
+                )
+
+            is_bool = value.lower() in ["true", "false"]
+            is_int = value.isdigit()
+            is_float = value.count(".") == 1 and (
+                value[0] == "." or value.replace(".", "").isdigit()
+            )
+
+            if is_bool:
+                value = value.lower() == "true"
+            elif is_int:
+                value = int(value)
+            elif is_float:
+                value = float(value)
+
+            if input_name in request_inputs:
+                raise ValueError(
+                    f"Input name already exists in request_inputs dictionary: {input_name}"
+                )
+            request_inputs[input_name] = 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