Add vision genai inference path for multi-file VLM evaluation (microsoft#2488)

## Describe your changes

Adds `_inference_vision_genai` method to `OnnxEvaluator` that enables
`olive run` to evaluate multi-file ONNX vision-language models (e.g.,
Qwen3-VL) using `onnxruntime-genai`.

### Problem

Vision-language models exported via `onnxruntime-genai` produce multiple
ONNX files (`vision.onnx`, `text.onnx`, `embedding.onnx`) with a
`genai_config.json`. The existing `_inference_vision` method only
supports single-file ONNX models with classification-style forward pass.
This prevented using `olive run --config` for evaluation of
autoregressive VLMs.

### Solution

- Auto-detect genai vision models by checking if `genai_config.json`
contains a `vision` field
- Route to new `_inference_vision_genai` method which uses `og.Model`,
`multimodal_processor`, and `og.Generator` for autoregressive text
generation
- Follows the same pattern as the existing speech genai inference
(`_inference_text_genai` for Whisper, `_inference_text_genai_streaming`
for Nemotron)
- Falls back to existing `_inference_vision` for single-file ONNX VQA
models

### Usage

```json
{
    "input_model": {
        "type": "OnnxModel",
        "model_path": "path/to/models",
        "onnx_file_name": "text.onnx"
    }
}
```

The evaluator will auto-detect `genai_config.json` in the model
directory and use the genai path.

### Related

- Depends on: microsoft#2476 (vision metrics: exact_match, relaxed_accuracy,
word_sort_ratio)
- Tested with: Qwen3-VL-2B-Instruct on AI2D benchmark in olive-recipes

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: jiafatom

olive/data/component/pre_process_data.py

@@ -387,6 +387,8 @@ def vision_vqa_pre_process(
     image_col: str = "image",
     question_col: str = "question",
     answer_col: str = "answer",
+    options_col: str = "",
+    system_prompt: str = "",
     max_samples: Optional[int] = None,
     limit: Optional[float] = None,
     seed: int = 42,
@@ -408,6 +410,10 @@ def vision_vqa_pre_process(
         image_col: Name of the image column. Defaults to "image".
         question_col: Name of the question column. Defaults to "question".
         answer_col: Name of the answer column. Defaults to "answer".
+        options_col: Name of the options column for multiple-choice questions. If specified,
+            options are formatted as numbered choices and appended to the question. Defaults to "".
+        system_prompt: System prompt to guide model responses (e.g., "Reply with only the
+            option number"). Passed through to the evaluator. Defaults to "".
         max_samples: Maximum number of samples (deprecated, use limit). Defaults to None.
         limit: Sampling limit following Olive convention:
             If >= 1: use first N samples.
@@ -438,11 +444,13 @@ class VisionVQADataset:
         Note: Use batch_size=1 in dataloader config as images have variable sizes.
         """
 
-        def __init__(self, hf_dataset, image_column, question_column, answer_column):
+        def __init__(self, hf_dataset, image_column, question_column, answer_column, options_column="", sys_prompt=""):
             self.dataset = hf_dataset
             self.image_column = image_column
             self.question_column = question_column
             self.answer_column = answer_column
+            self.options_column = options_column
+            self.system_prompt = sys_prompt
 
         def __len__(self):
             return len(self.dataset)
@@ -452,11 +460,27 @@ def __getitem__(self, idx):
             image = item[self.image_column]
             question = item[self.question_column]
             answer = item[self.answer_column]
+
+            # Format options into the question if options_col is specified
+            has_options = False
+            if self.options_column and self.options_column in item:
+                options = item[self.options_column]
+                if isinstance(options, (list, tuple)):
+                    options_text = "\n".join(f"{i}. {opt}" for i, opt in enumerate(options))
+                    question = f"{question}\n{options_text}"
+                    has_options = True
+
             # Handle list/tuple answers (some datasets have multiple valid answers)
             # Join with | separator so metrics can match against any valid answer
             if isinstance(answer, (list, tuple)):
                 answer = "|".join(str(a) for a in answer) if answer else ""
-            return {"image": image, "question": question}, str(answer)
+            input_dict = {
+                "image": image,
+                "question": question,
+                "system_prompt": self.system_prompt,
+                "extract_number": has_options,
+            }
+            return input_dict, str(answer)
 
         @staticmethod
         def collate_fn(batch):
@@ -472,4 +496,4 @@ def collate_fn(batch):
             answers = [item[1] for item in batch]
             return (inputs, answers)
 
-    return VisionVQADataset(dataset, image_col, question_col, answer_col)
+    return VisionVQADataset(dataset, image_col, question_col, answer_col, options_col, system_prompt)

olive/data/config.py

@@ -213,12 +213,11 @@ def to_data_container(self) -> "DataContainer":
         return dc_cls(config=self)
 
     def _update_default_component_type_with_task_type(self, dc_cls, default_components_type):
-        for component_name, config in self.components.items():
+        for config in self.components.values():
             if config and config.params:
                 task_type = config.params.get("task")
                 if task_type:
-                    task_specific_override = dc_cls.task_type_components_map.get(
-                        task_type.replace("-with-past", ""), {}
-                    ).get(component_name)
-                    if task_specific_override:
-                        default_components_type[component_name] = task_specific_override
+                    task_overrides = dc_cls.task_type_components_map.get(task_type.replace("-with-past", ""), {})
+                    # Apply all component overrides for this task type
+                    for override_component, override_type in task_overrides.items():
+                        default_components_type[override_component] = override_type

olive/evaluator/olive_evaluator.py

@@ -582,6 +582,20 @@ def _inference(
             dump_tuning_result(session.session, tuning_result_file)
         return OliveModelOutput(preds=preds, logits=logits), targets
 
+    @staticmethod
+    def _load_genai_config(model: ONNXModelHandler) -> Optional[dict]:
+        """Load genai_config.json from the model directory, or return None if not found."""
+        genai_config_path = Path(model.model_path).parent / "genai_config.json"
+        if not genai_config_path.exists():
+            return None
+        import json
+
+        try:
+            with genai_config_path.open(encoding="utf-8") as f:
+                return json.load(f)
+        except json.JSONDecodeError as e:
+            raise ValueError(f"Invalid JSON in genai config file: {genai_config_path}") from e
+
     def _evaluate_onnx_accuracy(
         self,
         model: ONNXModelHandler,
@@ -593,18 +607,21 @@ def _evaluate_onnx_accuracy(
     ) -> MetricResult:
         if _is_vision_metric(metric):
             _validate_vision_task_metric(metric)
-            inference_output, targets = self._inference_vision(
-                model, metric, dataloader, post_func, device, execution_providers
-            )
+            # Auto-detect genai vision model by checking for genai_config.json with vision field
+            genai_cfg = self._load_genai_config(model)
+            use_genai_vision = genai_cfg is not None and "vision" in genai_cfg.get("model", {})
+
+            if use_genai_vision:
+                inference_output, targets = self._inference_vision_genai(model, dataloader, device)
+            else:
+                inference_output, targets = self._inference_vision(
+                    model, metric, dataloader, post_func, device, execution_providers
+                )
         elif _is_text_based_metric(metric):
             # Auto-detect genai model by checking for genai_config.json
-            genai_config_path = Path(model.model_path).parent / "genai_config.json"
-            if genai_config_path.exists():
-                import json
-
-                with genai_config_path.open() as f:
-                    genai_config = json.load(f)
-                model_type = genai_config.get("model", {}).get("type", "")
+            genai_cfg = self._load_genai_config(model)
+            if genai_cfg:
+                model_type = genai_cfg.get("model", {}).get("type", "")
 
                 if model_type == "whisper":
                     inference_output, targets = self._inference_text_genai(
@@ -795,6 +812,142 @@ def _inference_vision(
 
         return OliveModelOutput(preds=all_preds, logits=None), all_targets
 
+    def _inference_vision_genai(
+        self,
+        model: ONNXModelHandler,
+        dataloader: "DataLoader",
+        device: Device = Device.CPU,
+    ) -> tuple[OliveModelOutput, Any]:
+        """Vision-based inference for VQA/OCR metrics using onnxruntime-genai.
+
+        Auto-detected when the model directory contains genai_config.json with a vision field.
+        Uses og.Model with multimodal processor for vision-language models (e.g., Qwen3-VL).
+        The dataloader must yield (input_dict, labels) where input_dict contains
+        'image' (PIL Image) and 'question' (str), and labels are reference answer strings.
+
+        Note: GPU/CPU selection is driven by the `device` parameter. onnxruntime-genai uses
+        short provider names internally (e.g., "cuda") which differ from ORT-style EP names.
+        """
+        try:
+            import onnxruntime_genai as og
+        except ImportError as e:
+            raise ImportError(
+                "onnxruntime-genai is required for genai-based vision evaluation. "
+                "Install it with: pip install onnxruntime-genai"
+            ) from e
+
+        import json
+        import re
+        import tempfile
+
+        try:
+            from PIL import Image
+        except ImportError as e:
+            raise ImportError("Pillow is required for vision evaluation. Install it with: pip install Pillow") from e
+
+        model_dir = str(Path(model.model_path).parent)
+
+        # max_length in genai is total sequence length (input + output).
+        # Default to 1028 which accommodates image/prompt tokens (~200-500) plus answer tokens.
+        # Note: genai_config.json's search.max_length is typically the full context window
+        # (e.g., 262144) which is too large — the model will stop at EOS well before this cap.
+        max_length = 1028
+
+        # Build og.Model with appropriate execution provider
+        # Note: onnxruntime-genai uses CPU by default when no provider is appended.
+        # Only non-CPU providers need to be explicitly added using short names (e.g., "cuda").
+        # This follows the same pattern as _inference_text_genai and _inference_text_genai_streaming.
+        config = og.Config(model_dir)
+        config.clear_providers()
+        if device == Device.GPU:
+            config.append_provider("cuda")
+        og_model = og.Model(config)
+        processor = og_model.create_multimodal_processor()
+        tokenizer = og.Tokenizer(og_model)
+
+        all_preds = []
+        all_targets = []
+
+        # Use a temporary directory for image files to avoid per-file create/delete overhead
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            tmp_img_path = Path(tmp_dir) / "input.png"
+
+            for batch in dataloader:
+                input_data, labels = OliveEvaluator.unpack_batch_for_accuracy(batch)
+
+                # input_data is a dict with 'image' (PIL) and 'question' (str)
+                # or a list of such dicts for batch_size > 1
+                items = [input_data] if isinstance(input_data, dict) else input_data
+
+                for item in items:
+                    pil_image = item.get("image")
+                    question = item.get("question", "")
+                    sys_prompt = item.get("system_prompt", "")
+                    extract_number = item.get("extract_number", False)
+
+                    if pil_image is None:
+                        # Append empty pred to maintain alignment with targets
+                        all_preds.append("")
+                        continue
+
+                    # Ensure PIL Image
+                    if not isinstance(pil_image, Image.Image):
+                        with Image.open(pil_image) as img:
+                            pil_image = img.convert("RGB")
+
+                    # Build chat messages for the vision-language model
+                    messages = []
+                    if sys_prompt:
+                        messages.append({"role": "system", "content": sys_prompt})
+                    messages.append(
+                        {
+                            "role": "user",
+                            "content": [
+                                {"type": "image"},
+                                {"type": "text", "text": question},
+                            ],
+                        }
+                    )
+                    messages_json = json.dumps(messages)
+
+                    # Save image to temp file for og.Images (reuse same path to minimize I/O)
+                    pil_image.save(str(tmp_img_path), format="PNG")
+                    images = og.Images.open(str(tmp_img_path))
+
+                    prompt = tokenizer.apply_chat_template(messages_json, add_generation_prompt=True)
+                    inputs = processor(prompt, images=images)
+
+                    params = og.GeneratorParams(og_model)
+                    params.set_search_options(max_length=max_length, do_sample=False)
+
+                    generator = og.Generator(og_model, params)
+                    generator.set_inputs(inputs)
+
+                    tokens = []
+                    while not generator.is_done():
+                        generator.generate_next_token()
+                        tokens.append(generator.get_next_tokens()[0])
+                    del generator
+
+                    pred = tokenizer.decode(tokens).strip()
+                    # For multiple-choice tasks, extract leading number from responses
+                    # like "1. D" or "0. krill" to match the expected answer format
+                    if extract_number:
+                        num_match = re.match(r"^(\d+)", pred)
+                        if num_match:
+                            pred = num_match.group(1)
+                    all_preds.append(pred)
+
+                # Collect reference texts (aligned with preds including empty ones for None images)
+                if isinstance(labels, (list, tuple)):
+                    all_targets.extend(labels)
+                else:
+                    all_targets.append(labels)
+
+        del og_model
+
+        return OliveModelOutput(preds=all_preds, logits=None), all_targets
+
     def _inference_text_genai(
         self,
         model: ONNXModelHandler,