[Rui Ren (rui-ren)]

Describe your changes

Add a config-driven, reference-only native-preservation example for an existing local three-component NVIDIA Nemotron 3.5 Bolt CUDA export.

• Reuse existing CompositeModel and ONNXModel handlers; no framework changes or new pass.
• Validate the standard RunConfig, resolve decoder/vision_encoder/embedding and bounded package-local external-data references, and round-trip source/GenAI/graph metadata through existing handlers.
• Stream SHA-256 hashes and write only a reference model config and receipt. Never copy/rewrite/dequantize/requantize native IQ2_XXS, raw Q8_0 or A16 NVFP4 weights.
• Reject passes/builds, unsupported CPU/stock GenAI/ort-extensions routes, absent CUDA, missing custom ops and CUDA initialization fallback. Default runtime checks create sessions only; --inspect-only explicitly records NOT_CHECKED.
• Document the prepared-feature/source-bound host prompt-compaction prerequisite, which this example does not implement.

This is not an Olive engine optimization workflow, accuracy evaluator, optimized package, or model-readiness claim. It intentionally does not call olive run with a dummy pass. The source package remains in place and the output is not relocatable.

Validation

• python -m pytest examples/bolt_native_cuda/test_preserve.py --basetemp=./bolt-recipe-test-artifacts -q: 22 passed (network-free IR-built opaque fixtures; runtime-success branches are mocked, not numerical fixtures).
• lintrunner: ok No lint issues, including pylint.
• Package import and olive --help: pass.
• Read-only native v8 inspection: all 13 files resolved, 24,200,581,839 referenced bytes; all three external-payload SHA-256 values match the export receipt. No model artifact copied.
• Existing tiny IQ2 fixture executed on assigned GPU3 with corrected ORT 1.32; profiling confirms MatMulGGMLIQ2XXS on CUDA and finite output. This is kernel availability evidence, not full-model accuracy evidence.
• Stock ORT 1.28 explicitly rejects the tiny fixture because MatMulGGMLIQ2XXS is not registered.
• Full-package session creation was attempted and bounded/stopped after more than four minutes without a completion result. It is not reported as passing; the final source was separately verified by native inspection and all 22 fixture tests.

Source pin: nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB at c6cc2e8ffa353403213d6b9b18aaf8eb684d705a.
Local native export source: d3ac98d5aa571b0c4c25fd288bd0a36741c342ee.
Corrected CUDA provider SHA-256: 394cc8fadb5006635368e96c8cd0fe124462d1a2a6eb67c3b7fc953f11d19c5d.

Model readiness remains NOT_ESTABLISHED. Earlier hidden/cache failures are not waived. Separate qualification must retain finite outputs, NRMSE <= 0.03, cosine >= 0.999 and exact required-token gates.

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass. (22 focused tests pass; full repository/model qualification not claimed.)
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.

Release note: Add a read-only, zero-pass native Bolt CUDA package-consumption example using existing Olive model handlers, with preservation receipts and explicit unsupported-runtime/route errors.

(Optional) Issue link

Independent example-only draft; no framework compatibility fix was warranted.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Rui Ren (rui-ren)]

Exact-head CI follow-up: 8e0278ad391fe1a6a6d703d19c590984095878cb fixes one real documentation issue from 2fcfff67: Docs linkcheck returned 404 for the not-yet-merged main/examples/bolt_native_cuda URL. The docs now reference the existing source-checkout README path, without suppressing linkcheck or changing framework/runtime/quantization behavior. Failure log: https://github.com/microsoft/Olive/actions/runs/38829738766/job/116745510608

Existing focused fixtures on the follow-up commit: 22 passed, and lintrunner passed. Validation clarification: the initial GPU-hidden local attempt failed during inherited Torch/Transformers accelerator-discovery imports; rerunning the same mocked/IR-only fixtures with GPU3 visible passed in 3.91 seconds. No fixture executed CUDA kernels and no additional full-model/session run was performed. The follow-up commit message's “with GPUs hidden” phrase is incorrect; this comment and the retained failed/successful logs correct that validation detail.

New-head workflows are running: Docs, Lint, fast tests. CLA passed. No actionable review/inline comments were present. The Azure bot note says authorized /azp run may be needed; no authorization bypass attempted.

The original bounded full-session attempt/evidence is retained unchanged. It produced no numerical result, and must not be treated as a numerical failure or successful inference. Full model readiness remains NOT_ESTABLISHED.