Add execution provider selection for builder capture

[David Fan (jiafatom)]

Summary

• add --execution_provider to olive capture-onnx-graph for Model Builder and Mobius Builder exports
• configure a separate Olive target system when the option is provided
• keep --conversion_device scoped to exporters that need a conversion host device
• preserve automatic CUDA routing for FP16/BF16 when the new option is omitted
• add coverage for legacy INT4/FP16 defaults and CPU-host/CUDA-target exports through both builders

Why

Model Builder and Mobius Builder use the target execution provider to choose the generated graph format. In particular, a CUDA INT4 QMoE export uses CUDA-specific expert-weight layouts and FP16 I/O. The existing capture CLI derives that argument from Olive's target accelerator but cannot independently select a CUDA target for INT4.

This change uses Olive's existing host/target separation:

olive capture-onnx-graph \
  --use_model_builder \
  --precision int4 \
  --execution_provider CUDAExecutionProvider \
  -m <model> -o <output>

The export host uses the existing CPU default and does not need a GPU. The capture workflow does not run on or evaluate the target, so Olive skips the local supported-EP check, and ONNX Runtime GenAI CUDA weight prepacking has a CPU fallback.

--execution_provider currently requires --use_model_builder or --use_mobius_builder.

Testing

• pytest test/cli/test_cli.py -k "capture_onnx_command_builder_execution_provider or capture_onnx_command_model_builder_accelerator or capture_onnx_command_use_mobius_builder or capture_onnx_command" -q (14 passed)
• ruff check olive/cli/capture_onnx.py test/cli/test_cli.py
• ruff format --check olive/cli/capture_onnx.py test/cli/test_cli.py

All tests were run in the jiafa-dev container.

[Copilot]

Copilot review overview

🟡 Changes recommended

FP16/BF16 CPU routing regression remains unresolved; MobiusBuilder GPU coverage is also requested.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 1

Open (1)

• Preserve automatic GPU routing for FP16 and BF16 · New

What changed in this PR

Fixes GPU routing for INT4 Model Builder and Mobius Builder exports.

Changes:

• Routes explicit GPU conversions through CUDA.
• Adds INT4 Model Builder GPU regression coverage.

File	Description
test/​cli/​test_cli.py	Adds INT4 Model Builder GPU configuration coverage.
olive/​cli/​capture_onnx.py	Updates builder accelerator and execution-provider routing.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.