feat: add ExecuTorch runtime

[leeclemnet]

What does this PR do?

ExecuTorch runtime

What it does: Adds ExecuTorchRuntime (XNNPACK, device cpu) on the Runtime contract. It loads a .pte program and times only Method.execute(). No script rows use it yet.
Why: A .pte comes from torch.export on the PyTorch model, not from ONNX, and the exported artifacts are not in the bucket yet. This PR lands the runtime, so the rows are only a config change when the artifacts exist. Stacked on #16.

Change map

• Runtime (sab/runtimes/executorch.py) — inputs are cast to the program's dtypes before the timed call. Outputs are cloned after it, because ExecuTorch can reuse its output buffers on the next call. Inputs are named by position (input0, input1), because a .pte has no input names.
• Tests (tests/test_executorch.py) — build tiny .pte programs with torch.export and XnnpackPartitioner, and check run() against the eager model.
• Packaging and CI (pyproject.toml, uv.lock, .github/workflows/tests.yml, README.md) — extra executorch. It is not in nvidia, because the NVIDIA host runs no ExecuTorch rows. CI installs it.

Behavior changes

• None. No script row uses the runtime.

Watch out for

• sab/runtimes/executorch.py:81 — the outputs are cloned after the timed call. Without the clone, a later call overwrites the outputs of an earlier one.
• pyproject.toml:34 — executorch pulls a large dependency set (uv.lock grows by about 800 lines). The lock keeps torch at 2.14.
• Only the XNNPACK backend on cpu is in this PR. The Core ML backend (npu) comes with the Mac PR.

Related Issue(s): n/a

Type of Change

• New feature (non-breaking change that adds functionality)

Testing

• Unit tests
• ExecuTorch XNNPACK vs ORT CPU on an Apple M4 Max (50 coco val2017 images)

CPU: ExecuTorch (XNNPACK) on an Apple M4 Max (fp32, on battery). The artifacts are the official exports in gs://rfdetr/export-2026-09-30. ORT reads the .onnx of the same export. YOLO26 has no official exports, so it has no rows.

Model	ORT CPU mAP50:95	ExecuTorch fp32 mAP50:95	ORT CPU median ms	ExecuTorch fp32 median ms
rfdetr-nano	53.9	53.9	34.5	51.7
rfdetr-small	58.3	58.3	58.3	84.9
rfdetr-medium	57.1	57.1	68.3	120.9
rfdetr-large	60.2	60.2	123.4	192.8
rfdetr-xlarge	61.1	61.1	234.9	342.8
rfdetr-2xlarge	62.3	62.3	322.3	416.9

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary, particularly in hard-to-understand areas
• My changes generate no new warnings or errors
• I have updated the documentation accordingly (if applicable)

Additional Context

• Stacked on feat: add OpenVINO runtime and all RF-DETR detection sizes #16. Merge refactor: compose runtimes and processors, add run control and uv packaging #15 and feat: add OpenVINO runtime and all RF-DETR detection sizes #16 first.

🤖 Generated with Claude Code

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	executorch@​1.5.1

View full report