Add EP- and device-specific graph surgery support to capture-onnx-graph CLI

[Xiaoyu Z (xiaoyu-work)]

Describe your changes

This PR changes olive capture-onnx-graph --use_mobius_builder to use a two-stage graph flow:

1. Mobius exports its canonical graph without automatically applying graph rewrites. Function-backed custom operators and their standard ONNX fallback bodies are preserved.
2. Olive decides the final graph representation after export.

When --execution_provider and --device are supplied, Olive applies the selected EP surgery profile. For a multi-component model, builds selects only the exported decoder, leaving vision, embedding, and other components unchanged.

When --onnx_standard is supplied, Olive runs InlineModelLocalFunctions across every exported component. All model-local functions are inlined, function definitions are removed, and Olive rejects any remaining non-standard-domain operator that has no standard fallback body. --onnx_standard may be combined with EP/device so structural requirements such as OpenVINO Gemma4 rank-4 per_layer_inputs and runtime configuration are retained.

Olive now owns every graph rewrite removed from Mobius. In addition to the migrated fusion/lowering surgeons, ConvertGroupQueryAttentionKVCacheToFp8 carries the former Mobius FP8 GQA KV-cache transformation, including calibrated per-layer scale inputs.

The accelerator EP/device is forwarded to Mobius only as a structural build contract; it never enables Mobius graph rewrites. This depends on onnxruntime/mobius#743.

Execution provider \ Device	cpu	gpu	npu
cuda	Invalid	AttentionToGroupQueryAttention (FP16/BF16) → PackQKVForGroupQueryAttention → FuseSkipRMSNormalization → FuseSkipLayerNormalization	Invalid
openvino	None	None	None
qnn	Invalid	AttentionToGroupQueryAttention → PackQKVForGroupQueryAttention → FuseSkipRMSNormalization → AttentionMaskToSequenceLengths	AttentionToGroupQueryAttention → PackQKVForGroupQueryAttention → FuseSkipRMSNormalization → AttentionMaskToSequenceLengths → SimplifiedLayerNormToL2Norm
trt-rtx	Invalid	AttentionToGroupQueryAttention (FP16/BF16) → PackQKVForGroupQueryAttention	Invalid
vitisai	Invalid	Invalid	None

Invalid denotes an unsupported provider/device combination. None denotes a valid combination with no automatic EP surgeries; users may still request strict expansion with --onnx_standard.

The EP profiles intentionally cover ordinary decoder graphs. Model-specific paths remain explicit rather than silently applying an incompatible generic rewrite.

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all relevant tests pass.
• Update directly related documentation.
• Run lintrunner on changed files.
• Include a user-facing release note.

Release note: capture-onnx-graph now keeps Mobius export canonical and lets Olive explicitly apply EP graph surgeries or expand all function-backed custom operators with --onnx_standard.

(Optional) Issue link

N/A

[Copilot]

🟡 Changes recommended

The QNN mapping passes an unsupported Mobius runtime EP instead of retaining the existing standard-export fallback.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

This PR updates Mobius ONNX capture to preserve source precision and optionally apply execution-provider-specific graph surgeries.

Changes:

• Standardizes Mobius export and removes precision overrides.
• Adds CLI execution-provider/device surgery profiles.
• Adds dtype-aware GQA surgery filtering and tests.

File summaries

File	Description
olive/cli/capture_onnx.py	Adds EP profiles and CLI validation.
olive/passes/onnx/mobius_model_builder.py	Uses standard ONNX export and runtime packaging.
olive/passes/onnx/graph_surgery/attention.py	Filters GQA rewrites by dtype.
olive/olive_config.json	Allows all source precisions.
test/cli/test_cli.py	Tests CLI profiles and validation.
test/passes/onnx/test_mobius_model_builder.py	Tests standard Mobius export behavior.
test/passes/onnx/test_graph_surgeries_attention.py	Tests dtype filtering.

Review details

• Files reviewed: 7/7 changed files
• Comments generated: 1
• Review effort level: Lite

---

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