Add CLI options for synthetic image generation

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

@@ -121,6 +121,11 @@ class LlmInputs:
     DEFAULT_OUTPUT_TOKENS_STDDEV = 0
     DEFAULT_NUM_PROMPTS = 100
 
+    DEFAULT_IMAGE_WIDTH_MEAN = 100
+    DEFAULT_IMAGE_WIDTH_STDDEV = 0
+    DEFAULT_IMAGE_HEIGHT_MEAN = 100
+    DEFAULT_IMAGE_HEIGHT_STDDEV = 0
+
     EMPTY_JSON_IN_VLLM_PA_FORMAT: Dict = {"data": []}
     EMPTY_JSON_IN_TENSORRTLLM_PA_FORMAT: Dict = {"data": []}
     EMPTY_JSON_IN_OPENAI_PA_FORMAT: Dict = {"data": []}
@@ -143,6 +148,11 @@ def create_llm_inputs(
         output_tokens_deterministic: bool = False,
         prompt_tokens_mean: int = DEFAULT_PROMPT_TOKENS_MEAN,
         prompt_tokens_stddev: int = DEFAULT_PROMPT_TOKENS_STDDEV,
+        image_width_mean: int = DEFAULT_IMAGE_WIDTH_MEAN,
+        image_width_stddev: int = DEFAULT_IMAGE_WIDTH_STDDEV,
+        image_height_mean: int = DEFAULT_IMAGE_HEIGHT_MEAN,
+        image_height_stddev: int = DEFAULT_IMAGE_HEIGHT_STDDEV,
+        image_format: ImageFormat = ImageFormat.PNG,
         random_seed: int = DEFAULT_RANDOM_SEED,
         num_of_output_prompts: int = DEFAULT_NUM_PROMPTS,
         add_model_name: bool = False,
@@ -185,6 +195,16 @@ def create_llm_inputs(
             The standard deviation of the length of the output to generate. This is only used if output_tokens_mean is provided.
         output_tokens_deterministic:
             If true, the output tokens will set the minimum and maximum tokens to be equivalent.
+        image_width_mean:
+            The mean width of images when generating synthetic image data.
+        image_width_stddev:
+            The standard deviation of width of images when generating synthetic image data.
+        image_height_mean:
+            The mean height of images when generating synthetic image data.
+        image_height_stddev:
+            The standard deviation of height of images when generating synthetic image data.
+        image_format:
+            The compression format of the images.
         batch_size:
             The number of inputs per request (currently only used for the embeddings and rankings endpoints)
 
@@ -221,6 +241,11 @@ def create_llm_inputs(
             prompt_tokens_mean,
             prompt_tokens_stddev,
             num_of_output_prompts,
+            image_width_mean,
+            image_width_stddev,
+            image_height_mean,
+            image_height_stddev,
+            image_format,
             batch_size,
             input_filename,
         )
@@ -256,6 +281,11 @@ def get_generic_dataset_json(
         prompt_tokens_mean: int,
         prompt_tokens_stddev: int,
         num_of_output_prompts: int,
+        image_width_mean: int,
+        image_width_stddev: int,
+        image_height_mean: int,
+        image_height_stddev: int,
+        image_format: ImageFormat,
         batch_size: int,
         input_filename: Optional[Path],
     ) -> Dict:
@@ -282,6 +312,16 @@ def get_generic_dataset_json(
             The standard deviation of the length of the prompt to generate
         num_of_output_prompts:
             The number of synthetic output prompts to generate
+        image_width_mean:
+            The mean width of images when generating synthetic image data.
+        image_width_stddev:
+            The standard deviation of width of images when generating synthetic image data.
+        image_height_mean:
+            The mean height of images when generating synthetic image data.
+        image_height_stddev:
+            The standard deviation of height of images when generating synthetic image data.
+        image_format:
+            The compression format of the images.
         batch_size:
             The number of inputs per request (currently only used for the embeddings and rankings endpoints)
         input_filename:
@@ -361,7 +401,7 @@ def get_generic_dataset_json(
                 input_filename = cast(Path, input_filename)
                 input_file_dataset = cls._get_input_dataset_from_file(input_filename)
                 input_file_dataset = cls._encode_images_in_input_dataset(
-                    input_file_dataset
+                    input_file_dataset, image_format
                 )
                 generic_dataset_json = (
                     cls._convert_input_synthetic_or_file_dataset_to_generic_json(
@@ -648,13 +688,14 @@ def _convert_to_openai_multi_modal_content(
         return generic_dataset_json
 
     @classmethod
-    def _encode_images_in_input_dataset(cls, input_file_dataset: Dict) -> Dict:
+    def _encode_images_in_input_dataset(
+        cls, input_file_dataset: Dict, image_format: ImageFormat
+    ) -> Dict:
         for row in input_file_dataset["rows"]:
             filename = row["row"].get("image")
             if filename:
                 img = Image.open(filename)
-                # (TMA-1985) Support multiple image formats
-                img_base64 = cls._encode_image(img, ImageFormat.PNG)
+                img_base64 = cls._encode_image(img, image_format)
                 row["row"]["image"] = f"data:image/png;base64,{img_base64}"
 
         return input_file_dataset

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

@@ -76,6 +76,11 @@ def generate_inputs(args: Namespace, tokenizer: Tokenizer) -> None:
         output_tokens_mean=args.output_tokens_mean,
         output_tokens_stddev=args.output_tokens_stddev,
         output_tokens_deterministic=args.output_tokens_mean_deterministic,
+        image_width_mean=args.image_width_mean,
+        image_width_stddev=args.image_width_stddev,
+        image_height_mean=args.image_height_mean,
+        image_height_stddev=args.image_height_stddev,
+        image_format=args.image_format,
         random_seed=args.random_seed,
         num_of_output_prompts=args.num_prompts,
         add_model_name=add_model_name,

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

@@ -41,6 +41,7 @@
     OPEN_ORCA,
 )
 from genai_perf.llm_inputs.llm_inputs import (
+    ImageFormat,
     LlmInputs,
     ModelSelectionStrategy,
     OutputFormat,
@@ -116,6 +117,25 @@ def _check_compare_args(
     return args
 
 
+def _check_image_input_args(
+    parser: argparse.ArgumentParser, args: argparse.Namespace
+) -> argparse.Namespace:
+    """
+    Sanity check the image input args
+    """
+    if args.image_width_mean <= 0 or args.image_height_mean <= 0:
+        parser.error(
+            "Both --image-width-mean and --image-height-mean values must be positive."
+        )
+    if args.image_width_stddev < 0 or args.image_height_stddev < 0:
+        parser.error(
+            "Both --image-width-stddev and --image-height-stddev values must be non-negative."
+        )
+
+    args = _convert_str_to_enum_entry(args, "image_format", ImageFormat)
+    return args
+
+
 def _check_conditional_args(
     parser: argparse.ArgumentParser, args: argparse.Namespace
 ) -> argparse.Namespace:
@@ -417,6 +437,51 @@ def _add_input_args(parser):
     )
 
 
+def _add_image_input_args(parser):
+    input_group = parser.add_argument_group("Image Input")
+
+    input_group.add_argument(
+        "--image-width-mean",
+        type=int,
+        default=LlmInputs.DEFAULT_IMAGE_WIDTH_MEAN,
+        required=False,
+        help=f"The mean width of images when generating synthetic image data.",
+    )
+
+    input_group.add_argument(
+        "--image-width-stddev",
+        type=int,
+        default=LlmInputs.DEFAULT_IMAGE_WIDTH_STDDEV,
+        required=False,
+        help=f"The standard deviation of width of images when generating synthetic image data.",
+    )
+
+    input_group.add_argument(
+        "--image-height-mean",
+        type=int,
+        default=LlmInputs.DEFAULT_IMAGE_HEIGHT_MEAN,
+        required=False,
+        help=f"The mean height of images when generating synthetic image data.",
+    )
+
+    input_group.add_argument(
+        "--image-height-stddev",
+        type=int,
+        default=LlmInputs.DEFAULT_IMAGE_HEIGHT_STDDEV,
+        required=False,
+        help=f"The standard deviation of height of images when generating synthetic image data.",
+    )
+
+    input_group.add_argument(
+        "--image-format",
+        type=str,
+        choices=utils.get_enum_names(ImageFormat),
+        default="png",
+        required=False,
+        help=f"The compression format of the images.",
+    )
+
+
 def _add_profile_args(parser):
     profile_group = parser.add_argument_group("Profiling")
     load_management_group = profile_group.add_mutually_exclusive_group(required=False)
@@ -664,6 +729,7 @@ def _parse_profile_args(subparsers) -> argparse.ArgumentParser:
     )
     _add_endpoint_args(profile)
     _add_input_args(profile)
+    _add_image_input_args(profile)
     _add_profile_args(profile)
     _add_output_args(profile)
     _add_other_args(profile)
@@ -743,6 +809,7 @@ def refine_args(
         args = _infer_prompt_source(args)
         args = _check_model_args(parser, args)
         args = _check_conditional_args(parser, args)
+        args = _check_image_input_args(parser, args)
         args = _check_load_manager_args(args)
         args = _set_artifact_paths(args)
     elif args.subcommand == Subcommand.COMPARE.to_lowercase():

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

@@ -31,6 +31,7 @@
 import pytest
 from genai_perf import __version__, parser
 from genai_perf.llm_inputs.llm_inputs import (
+    ImageFormat,
     ModelSelectionStrategy,
     OutputFormat,
     PromptSource,
@@ -40,7 +41,7 @@
 
 class TestCLIArguments:
     # ================================================
-    # GENAI-PERF COMMAND
+    # PROFILE COMMAND
     # ================================================
     expected_help_output = (
         "CLI to profile LLMs and Generative AI models with Perf Analyzer"
@@ -215,6 +216,23 @@ def test_help_version_arguments_output_and_exit(
                 ["--synthetic-input-tokens-stddev", "7"],
                 {"synthetic_input_tokens_stddev": 7},
             ),
+            (
+                ["--image-width-mean", "123"],
+                {"image_width_mean": 123},
+            ),
+            (
+                ["--image-width-stddev", "123"],
+                {"image_width_stddev": 123},
+            ),
+            (
+                ["--image-height-mean", "456"],
+                {"image_height_mean": 456},
+            ),
+            (
+                ["--image-height-stddev", "456"],
+                {"image_height_stddev": 456},
+            ),
+            (["--image-format", "png"], {"image_format": ImageFormat.PNG}),
             (["-v"], {"verbose": True}),
             (["--verbose"], {"verbose": True}),
             (["-u", "test_url"], {"u": "test_url"}),
@@ -732,6 +750,26 @@ def test_prompt_source_assertions(self, monkeypatch, mocker, capsys):
         captured = capsys.readouterr()
         assert expected_output in captured.err
 
+    @pytest.mark.parametrize(
+        "args",
+        [
+            # negative numbers
+            ["--image-width-mean", "-123"],
+            ["--image-width-stddev", "-34"],
+            ["--image-height-mean", "-123"],
+            ["--image-height-stddev", "-34"],
+            # zeros
+            ["--image-width-mean", "0"],
+            ["--image-height-mean", "0"],
+        ],
+    )
+    def test_positive_image_input_args(self, monkeypatch, args):
+        combined_args = ["genai-perf", "profile", "-m", "test_model"] + args
+        monkeypatch.setattr("sys.argv", combined_args)
+
+        with pytest.raises(SystemExit) as excinfo:
+            parser.parse_args()
+
     # ================================================
     # COMPARE SUBCOMMAND
     # ================================================

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

@@ -249,6 +249,11 @@ def test_generate_json(self, monkeypatch) -> None:
           "random_seed": 0,
           "synthetic_input_tokens_mean": 550,
           "synthetic_input_tokens_stddev": 0,
+          "image_width_mean": 100,
+          "image_width_stddev": 0,
+          "image_height_mean": 100,
+          "image_height_stddev": 0,
+          "image_format": "png",
           "concurrency": 1,
           "measurement_interval": 10000,
           "request_rate": null,