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

[justinchuby]

Summary

Adds a new MobiusModelBuilder Olive pass that wraps the mobius package's build() function to produce ONNX models directly from HuggingFace model IDs.

What this does

• Single-component models (LLMs): returns ONNXModelHandler
• Multi-component models (VLMs, encoder-decoders): returns CompositeModelHandler with one ONNXModelHandler per component
• EP auto-detection: maps Olive accelerator spec → mobius EP (cpu/cuda/dml/webgpu)
• Precision: fp32 (default), fp16, bf16
• trust_remote_code passthrough

Files

File	Description
olive/passes/onnx/mobius_model_builder.py	Pass implementation
olive/olive_config.json	Registered as MobiusModelBuilder
examples/gemma4/gemma4_int4_pipeline.json	Example: Gemma4-E2B CUDA/fp16 → INT4
examples/gemma4/gemma4_fp32_cpu.json	Example: Gemma4-E2B CPU/fp32
test/passes/onnx/test_mobius_model_builder.py	10 unit tests

E2E Test

Verified with HuggingFaceTB/SmolLM2-135M-Instruct (135M llama-type, locally cached):

olive run --config /tmp/olive_e2e_test.json

Result:

• Pipeline executed without errors
• model.onnx (289KB) + model.onnx.data (1.2GB) produced
• onnxruntime.InferenceSession loaded successfully
• Correct causal-LM I/O: input_ids/attention_mask/position_ids + 30×past_kv → logits + 30×present_kv

Test coverage

10 passed in 2.0s

Tests cover: single-component, multi-component, EP auto-detection, EP override, precision mapping, non-HF model rejection, missing mobius error, all 4 EP map entries.

Notes

• E2B/E4B are Any-to-Any (vision+audio+text); 26B-A4B/31B are Image-Text-to-Text only

[Copilot]

Pull request overview

Adds a new ONNX pass (MobiusModelBuilder) that uses the mobius package to build ONNX models directly from HuggingFace model IDs, returning either a single ONNXModelHandler or a CompositeModelHandler for multi-component exports.

Changes:

• Introduces olive/passes/onnx/mobius_model_builder.py implementing the new pass (EP mapping, precision mapping, trust_remote_code passthrough).
• Registers the pass in olive/olive_config.json and adds two Gemma4 example run configs.
• Adds unit tests for single-component, multi-component, EP selection, and error paths.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
olive/passes/onnx/mobius_model_builder.py	New pass wrapping mobius.build() and emitting Olive model handlers.
olive/olive_config.json	Registers MobiusModelBuilder and declares extras for its dependencies.
examples/gemma4/gemma4_int4_pipeline.json	Example pipeline: mobius export (fp16 CUDA) then INT4 quantization.
examples/gemma4/gemma4_fp32_cpu.json	Example pipeline: mobius export (fp32 CPU).
test/passes/onnx/test_mobius_model_builder.py	New unit tests for config, handler types, EP mapping, and missing dependency behavior.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 4 comments.

[jambayk]

we don't keep examples in this repo anymore. can you create an accompanying PR in microsoft/olive-recipes?

[justinchuby]

I am just using this PR to iterate the files. Could you comment on whether there are errors or changes needed? I will move the files over once stable.

[jambayk]

some comments:

• have never tried adding an unused field like "comment" in the config. don't remember if the config accepts unknown field.
• engine level nesting is optional. most examples we have now don't use it for cleaner configs
• cpu ep is the default the system doesn't need to be added explicitly

[jambayk]

we could create an enum of the supported eps for automatic validation like in

Olive/olive/passes/pytorch/autoawq.py

Line 27 in 8b1957e

class ModelDtype(StrEnumBase):

.
unless you think the options might keep growing and it would be hard to keep it in sync across versions

[jambayk]

the quantization pass works in a pytorch model and should be run before mobius

[justinchuby]

Is there a pass I can use to quantize after the model? I would like to use that as a an example for now

[jambayk]

you can use the rtn pass

{ 
      "type": "rtn", 
      "bits": 4, 
      "sym": false, 
      "group_size": 32,
      "embeds": true,
      "lm_head": true
  }

[jambayk]

oh, i misread the comment. you can use the blockwise quantizer pass:

 {
      "type": "OnnxBlockWiseRtnQuantization",
      "block_size": 128,
      "is_symmetric": true,
      "accuracy_level": 4,
      "save_as_external_data": true
  }

[justinchuby]

Thanks. And it will process all components together?

[jambayk]

if the model is composite it should run the quantizer on each component and return a new composite model.

[jambayk]

nit: maybe we could do this removal in a different PR? these make the PR seem bigger than it actually is.

[justinchuby]

sorry I ran lintrunner in the wrong repo 😅 Looks like Olive's ruff version can be updated