Update capture onnx graph cli exporting logic (microsoft#2496)

## Describe your changes
1. mobius exporter always default to genai.
2. remove import error catch from mb pass, it can swallow the real
error.
## Checklist before requesting a review
- [ ] Add unit tests for this change.
- [ ] Make sure all tests can pass.
- [ ] Update documents if necessary.
- [ ] Lint and apply fixes to your code by running `lintrunner -a`
- [ ] Is this a user-facing change? If yes, give a description of this
change to be included in the release notes.

## (Optional) Issue link

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>

Author: 3 people

olive/cli/capture_onnx.py

@@ -130,7 +130,7 @@ def register_subcommand(parser: ArgumentParser):
             type=str,
             default="fp16",
             choices=["fp16", "fp32", "int4", "bf16"],
-            help="The precision of the ONNX model. This is used by Model Builder",
+            help="The precision of the ONNX model. Used by Model Builder and Mobius Builder.",
         )
         mb_group.add_argument(
             "--int4_block_size",
@@ -178,7 +178,12 @@ def register_subcommand(parser: ArgumentParser):
         )
 
         sub_parser.add_argument(
-            "--use_ort_genai", action="store_true", help="Use OnnxRuntime generate() API to run the model"
+            "--use_ort_genai",
+            action="store_true",
+            help=(
+                "Use OnnxRuntime generate() API to run the model. "
+                "Ignored when --use_mobius_builder is set (mobius always emits ORT GenAI artifacts)."
+            ),
         )
 
         add_logging_options(sub_parser)
@@ -241,15 +246,7 @@ def _get_run_config(self, tempdir: str) -> dict:
                 )
             del config["passes"]["c"]
             del config["passes"]["m"]
-            to_replace.extend(
-                [
-                    (("passes", "b", "precision"), self.args.precision),
-                    (
-                        ("passes", "b", "runtime"),
-                        "ort-genai" if self.args.use_ort_genai else "none",
-                    ),
-                ]
-            )
+            to_replace.append((("passes", "b", "precision"), self.args.precision))
         elif is_diffusers_model:
             del config["passes"]["m"]
             del config["passes"]["b"]

olive/passes/onnx/mobius_model_builder.py

@@ -54,12 +54,6 @@ class MobiusBuilder(Pass):
     See https://github.com/onnxruntime/mobius
     """
 
-    class MobiusRuntime(StrEnumBase):
-        """Target runtimes for genai config generation."""
-
-        NONE = "none"
-        ORT_GENAI = "ort-genai"
-
     class MobiusEP(StrEnumBase):
         """Execution providers supported by mobius."""
 
@@ -97,16 +91,6 @@ def _default_config(cls, accelerator_spec: AcceleratorSpec) -> dict[str, PassCon
                     "quantization pass (e.g. OnnxMatMulNBits) after this pass."
                 ),
             ),
-            "runtime": PassConfigParam(
-                type_=MobiusBuilder.MobiusRuntime,
-                required=False,
-                default_value=MobiusBuilder.MobiusRuntime.ORT_GENAI,
-                description=(
-                    "Target runtime. 'ort-genai' (default) generates "
-                    "genai_config.json, tokenizer files, and processor "
-                    "configs alongside the ONNX models. 'none' to skip."
-                ),
-            ),
         }
 
     def _run_for_config(
@@ -169,10 +153,8 @@ def _run_for_config(
         pkg.save(str(output_dir))
 
         # Generate ORT GenAI config artifacts (genai_config.json, tokenizer
-        # files, processor configs) when runtime is set to ort-genai.
-        genai_artifacts = {}
-        if config.runtime == self.MobiusRuntime.ORT_GENAI:
-            genai_artifacts = self._write_genai_config(pkg, str(output_dir), model_id, ep_str)
+        # files, processor configs) alongside the ONNX models.
+        genai_artifacts = self._write_genai_config(pkg, str(output_dir), model_id, ep_str)
 
         package_keys = list(pkg.keys())
         logger.info("MobiusBuilder: saved components %s to '%s'", package_keys, output_dir)

olive/passes/onnx/model_builder.py

@@ -214,14 +214,8 @@ def _run_for_config(
         config: type[BasePassConfig],
         output_model_path: str,
     ) -> ONNXModelHandler:
-        try:
-            from onnxruntime_genai.models.builder import create_model
-        except ImportError as e:
-            raise ImportError(
-                "onnxruntime-genai package is required to run ModelBuilder pass. Please install the package"
-                " corresponding to your onnxruntime installation using pip. cpu: onnxruntime-genai, cuda:"
-                " onnxruntime-genai-cuda, directml: onnxruntime-genai-directml"
-            ) from e
+        from onnxruntime_genai.models.builder import create_model
+
         self.maybe_patch_quant()
 
         precision = config.precision

test/cli/test_cli.py

@@ -485,7 +485,6 @@ def test_capture_onnx_command_use_mobius_builder(_, mock_run, precision, use_ort
     assert "m" not in config["passes"]
     assert config["passes"]["b"]["type"] == "MobiusBuilder"
     assert config["passes"]["b"]["precision"] == precision
-    assert config["passes"]["b"]["runtime"] == ("ort-genai" if use_ort_genai else "none")
     assert mock_run.call_count == 1
 
 

test/cli/test_cli_test_model_smoke.py

@@ -282,7 +282,6 @@ def _assert_discrepancy_mobius(self, tmp_path: Path, model_id: str):
                 "mobius_builder": {
                     "type": "MobiusBuilder",
                     "precision": "fp32",
-                    "runtime": "none",
                 },
                 "discrepancy_check": {
                     "type": "OnnxDiscrepancyCheck",

test/passes/onnx/test_mobius_model_builder.py

@@ -98,7 +98,7 @@ def _save(directory: str, **_kwargs):
 def _patch_build(pkg: MagicMock):
     # Patch mobius.build directly — lazy import inside _run_for_config means
     # patching the module attribute, not the local binding.
-    # Also patch _write_genai_config since the default runtime is ort-genai.
+    # Also patch _write_genai_config since mobius always emits ORT GenAI artifacts.
     return _CombinePatches(
         patch("mobius.build", return_value=pkg),
         patch.object(MobiusBuilder, "_write_genai_config"),
@@ -127,13 +127,12 @@ def __exit__(self, *args):
 
 
 def test_default_config_params():
-    """MobiusBuilder must declare precision and runtime, and must not declare execution_provider or trust_remote_code."""
+    """MobiusBuilder must declare precision, and must not declare execution_provider or trust_remote_code."""
     accelerator_spec = AcceleratorSpec(
         accelerator_type=Device.CPU, execution_provider=ExecutionProvider.CPUExecutionProvider
     )
     config = MobiusBuilder._default_config(accelerator_spec)  # pylint: disable=protected-access
     assert "precision" in config
-    assert "runtime" in config
     assert "execution_provider" not in config
     assert "trust_remote_code" not in config