feat: add MobiusModelBuilder Olive pass for mobius-backed ONNX export

examples/gemma4/gemma4_fp32_cpu.json

@@ -0,0 +1,12 @@
+{
+    "comment": "Build google/gemma-4-E2B-it as a float32 ONNX model using mobius, targeting CPU execution. E2B and E4B are Any-to-Any (vision + audio + text). For Image-Text-to-Text only models (no audio encoder), use google/gemma-4-26B-A4B-it or google/gemma-4-31B-it.",
+    "input_model": { "type": "HfModel", "model_path": "google/gemma-4-E2B-it", "task": "text-generation" },
+    "systems": {
+        "local_system": {
+            "type": "LocalSystem",
+            "accelerators": [ { "device": "cpu", "execution_providers": [ "CPUExecutionProvider" ] } ]
+        }
+    },
+    "passes": { "mobius_build": { "type": "MobiusModelBuilder", "precision": "fp32" } },
+    "engine": { "target": "local_system", "output_dir": "models/gemma4-e2b-fp32-cpu", "log_severity_level": 1 }
+}

examples/gemma4/gemma4_int4_cuda.json

@@ -0,0 +1,15 @@
+{
+    "comment": "Build google/gemma-4-E2B-it as a float16 ONNX model using mobius, then apply INT4 weight-only quantization for compact deployment. E2B and E4B are Any-to-Any (vision + audio + text). For Image-Text-to-Text only models (no audio encoder), use google/gemma-4-26B-A4B-it or google/gemma-4-31B-it.",
+    "input_model": { "type": "HfModel", "model_path": "google/gemma-4-E2B-it", "task": "text-generation" },
+    "systems": {
+        "local_system": {
+            "type": "LocalSystem",
+            "accelerators": [ { "device": "gpu", "execution_providers": [ "CUDAExecutionProvider" ] } ]
+        }
+    },
+    "passes": {
+        "mobius_build": { "type": "MobiusModelBuilder", "precision": "fp16" },
+        "int4_quantize": { "type": "GptqQuantizer", "bits": 4, "group_size": 128, "sym": true }
+    },
+    "engine": { "target": "local_system", "output_dir": "models/gemma4-e2b-int4-cuda", "log_severity_level": 1 }
+}

olive/olive_config.json

@@ -196,6 +196,15 @@
             "supported_algorithms": [  ],
             "supported_quantization_encodings": [  ]
         },
+        "MobiusModelBuilder": {
+            "module_path": "olive.passes.onnx.mobius_model_builder.MobiusModelBuilder",
+            "supported_providers": [ "*" ],
+            "supported_accelerators": [ "*" ],
+            "supported_precisions": [ "fp32", "fp16", "bf16" ],
+            "supported_algorithms": [  ],
+            "supported_quantization_encodings": [  ],
+            "extra_dependencies": [ "mobius-genai", "onnx-ir" ]
+        },
         "LoftQ": {
             "module_path": "olive.passes.pytorch.lora.LoftQ",
             "supported_providers": [ "*" ],
@@ -665,6 +674,8 @@
         "inc": [ "neural-compressor" ],
         "lora": [ "accelerate>=0.30.0", "peft", "scipy" ],
         "diffusers": [ "accelerate>=0.30.0", "peft", "diffusers" ],
+        "mobius-genai": [ "mobius-genai" ],
+        "onnx-ir": [ "onnx-ir" ],
         "nvmo": [ "nvidia-modelopt[onnx]" ],
         "openvino": [
             "openvino>=2025.4.1",

olive/passes/onnx/mobius_model_builder.py

@@ -0,0 +1,196 @@
+# -------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+# --------------------------------------------------------------------------
+"""Build ONNX models from HuggingFace model IDs using the mobius package."""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import TYPE_CHECKING, ClassVar
+
+from olive.constants import Precision
+from olive.hardware.constants import ExecutionProvider
+from olive.model import HfModelHandler, ONNXModelHandler
+from olive.model.handler.composite import CompositeModelHandler
+from olive.passes import Pass
+from olive.passes.olive_pass import PassConfigParam
+
+if TYPE_CHECKING:
+    from olive.hardware.accelerator import AcceleratorSpec
+    from olive.passes.pass_config import BasePassConfig
+
+logger = logging.getLogger(__name__)
+
+# Maps Olive Precision values to mobius dtype strings.
+# "f32" = 32-bit float (torch.float32), standard full precision.
+# "f16" = 16-bit float (torch.float16), half precision — good for GPU inference.
+# "bf16" = bfloat16 (torch.bfloat16), brain float — preferred over f16 on newer hardware.
+# For INT4/INT8 quantization, use a downstream Olive quantization pass (e.g. OnnxMatMulNBits)
+# after this pass rather than setting precision here.
+_PRECISION_TO_DTYPE: dict[str, str] = {
+    Precision.FP32: "f32",
+    Precision.FP16: "f16",
+    Precision.BF16: "bf16",
+}
+
+
+class MobiusModelBuilder(Pass):
+    """Olive pass that uses mobius to build ONNX models from HuggingFace model IDs.
+
+    Supports all model architectures registered in mobius (LLMs, VLMs, speech
+    models, diffusion models).  For multi-component models (e.g. vision-language
+    models that produce ``model``, ``vision``, and ``embedding`` sub-graphs) the
+    pass returns a :class:`~olive.model.handler.composite.CompositeModelHandler`
+    whose components are individual :class:`~olive.model.ONNXModelHandler` objects.
+    Single-component models return a plain :class:`~olive.model.ONNXModelHandler`.
+
+    Requires ``mobius-genai`` to be installed::
+
+        pip install mobius-genai
+
+    See https://github.com/microsoft/mobius
+    """
+
+    # Maps Olive ExecutionProvider enum values to mobius EP names.
+    EP_MAP: ClassVar[dict[ExecutionProvider, str]] = {
+        ExecutionProvider.CPUExecutionProvider: "cpu",
+        ExecutionProvider.CUDAExecutionProvider: "cuda",
+        ExecutionProvider.DmlExecutionProvider: "dml",
+        ExecutionProvider.WebGpuExecutionProvider: "webgpu",
+    }
+
+    @classmethod
+    def is_accelerator_agnostic(cls, accelerator_spec: AcceleratorSpec) -> bool:
+        # EP selection determines which fused ops are emitted, so this pass is
+        # EP-specific.
+        return False
+
+    @classmethod
+    def _default_config(cls, accelerator_spec: AcceleratorSpec) -> dict[str, PassConfigParam]:
+        return {
+            "precision": PassConfigParam(
+                type_=Precision,
+                required=False,
+                default_value=Precision.FP32,
+                description=(
+                    "Model weight / compute precision. One of: fp32, fp16, bf16. "
+                    "Defaults to fp32. For INT4 quantization, run an Olive "
+                    "quantization pass (e.g. OnnxMatMulNBits) after this pass."
+                ),
+            ),
+            "execution_provider": PassConfigParam(
+                type_=str,
+                required=False,
+                default_value=None,
+                description=(
+                    "Override the mobius execution provider (cpu, cuda, dml, webgpu). "
+                    "When None (default), the EP is auto-detected from the Olive "
+                    "accelerator spec."
+                ),
+            ),
+            "trust_remote_code": PassConfigParam(
+                type_=bool,
+                required=False,
+                default_value=False,
+                description="Pass trust_remote_code=True to the HuggingFace config loader.",
+            ),
+        }
+
+    def _run_for_config(
+        self,
+        model: HfModelHandler,
+        config: type[BasePassConfig],
+        output_model_path: str,
+    ) -> ONNXModelHandler | CompositeModelHandler:
+        try:
+            from mobius import build  # noqa: PLC0415
+        except ImportError as exc:
+            raise ImportError(
+                "mobius-genai is required to run MobiusModelBuilder. Install with: pip install mobius-genai"
+            ) from exc
+
+        if not isinstance(model, HfModelHandler):
+            raise ValueError(f"MobiusModelBuilder requires an HfModelHandler input, got {type(model).__name__}.")
+
+        # Resolve EP: explicit config override > accelerator spec > fallback to cpu.
+        ep_str: str = config.execution_provider or self.EP_MAP.get(self.accelerator_spec.execution_provider, "cpu")
+
+        dtype_str: str = _PRECISION_TO_DTYPE.get(config.precision, "f32")
+        model_id: str = model.model_name_or_path
+
+        logger.info(
+            "MobiusModelBuilder: building '%s' (ep=%s, dtype=%s)",
+            model_id,
+            ep_str,
+            dtype_str,
+        )
+
+        if config.trust_remote_code:
+            logger.warning("MobiusModelBuilder: trust_remote_code=True — only use with trusted model sources.")
+
+        output_dir = Path(output_model_path)
+        output_dir.mkdir(parents=True, exist_ok=True)
+
+        pkg = build(
+            model_id,
+            dtype=dtype_str,
+            execution_provider=ep_str,
+            load_weights=True,
+            trust_remote_code=config.trust_remote_code,
+        )
+
+        # ModelPackage.save() handles both single and multi-component layouts:
+        #   single component  → <output_dir>/model.onnx
+        #   multi-component   → <output_dir>/<name>/model.onnx  for each key
+        pkg.save(str(output_dir))
+
+        package_keys = list(pkg.keys())
+        logger.info("MobiusModelBuilder: saved components %s to '%s'", package_keys, output_dir)
+
+        if len(package_keys) == 1:
+            # Single-component model (most LLMs): return a plain ONNXModelHandler.
+            onnx_path = output_dir / "model.onnx"
+            if not onnx_path.exists():
+                raise RuntimeError(
+                    f"MobiusModelBuilder: expected output file not found: {onnx_path}. "
+                    "mobius.build() may have failed silently or saved to an unexpected path."
+                )
+            return ONNXModelHandler(
+                model_path=str(output_dir),
+                onnx_file_name="model.onnx",
+                model_attributes={
+                    "mobius_package_keys": package_keys,
+                    **(model.model_attributes or {}),
+                },
+            )
+
+        # Multi-component model (VLMs, encoder-decoders, diffusion pipelines):
+        # mobius saves each component to <output_dir>/<key>/model.onnx.
+        components = []
+        for key in package_keys:
+            component_dir = output_dir / key
+            onnx_path = component_dir / "model.onnx"
+            if not onnx_path.exists():
+                raise RuntimeError(
+                    f"MobiusModelBuilder: expected output file not found: {onnx_path}. "
+                    f"mobius.build() may have failed silently for component '{key}'."
+                )
+            components.append(
+                ONNXModelHandler(
+                    model_path=str(component_dir),
+                    onnx_file_name="model.onnx",
+                    model_attributes={"mobius_component": key},
+                )
+            )
+
+        return CompositeModelHandler(
+            model_components=components,
+            model_component_names=package_keys,
+            model_path=str(output_dir),
+            model_attributes={
+                "mobius_package_keys": package_keys,
+                **(model.model_attributes or {}),
+            },
+        )