Skip to content

ci: release 1.29.1 - #124

Merged
kibae merged 2 commits into
mainfrom
release/1.29.1
Sep 10, 2026
Merged

kibae merged 2 commits into
mainfrom
release/1.29.1

Conversation

@kibae

@kibae kibae commented Sep 10, 2026

Copy link
Copy Markdown
Owner

Summary

  • Sync onnxruntime-server with upstream ONNX Runtime 1.29.1.
  • Bumps version strings via deploy/update-version.sh 1.29.0 -> 1.29.1.

Upstream surface review (1.29.0 -> 1.29.1)

  • onnxruntime_c_api.h, onnxruntime_cxx_api.h and onnxruntime_cxx_inline.h are byte-identical
    between v1.29.0 and v1.29.1. No new element data types, no new GraphOptimizationLevel /
    ExecutionMode members, no new Ort::SessionOptions setter. Nothing to mirror.
  • kOrtSessionOptionsGqaValueLayout ("session.gqa_value_layout", values BNSH / BNHS) — new in
    1.29.1 for the GroupQueryAttention BNHS Value KV-cache layout. Not implemented as a typed option:
    apply_session_options' config_entries passthrough already forwards arbitrary keys verbatim and
    echoes them back through GetSessionConfigEntry, so the key is reachable over the API today. Left
    out of the docs/swagger/openapi.yaml example list because that example documents the passthrough
    mechanism, not a catalogue of ORT keys.
  • kOrtEpDevice_EpMetadataKey_GqaPreferredValueLayout ("gqa_preferred_value_layout") — OrtEpDevice
    metadata key. Not applicable: the server does not expose EP device metadata.
  • Bundled ONNX version unchanged (patch release, no opset bump), so the
    test/sample-onnx-generator/*.py opset pins stay as they are.
  • Release asset naming unchanged (onnxruntime-linux-x64-gpu_cuda12-<VER>.tgz); no install-script,
    dockerfile, or .github/actions/download-onnxruntime-*.sh change needed.

Version-pinned assumptions re-validated (still true at 1.29.1)

  • src/onnx/value_info.cpp:203 / unit_test_value_info.cpp:249 — uint2 still stored unpacked, one
    value per byte; the sample-datatypes.py packed-uint2 tripwire did not fire in ctest.
  • src/onnx/value_info.cpp:275 / unit_test_value_info.cpp:284 — complex64/128 tensors still not
    allocatable by ORT; the defensive decoder branches stay defensive.

Test plan

The dependencies-apt cache is keyed on the runner label alone, so a build job
can restore a deb set that prepare-apt downloaded on a different runner image.
During an image rollout the two disagree on preinstalled library versions and
'dpkg -i' aborts on a versioned dependency -- ubuntu-22.04 image 20260831.284
carries libevent-core 2.1.12-stable-1build3 while the cached libevent-dev from
image 20260907.292 requires 2.1.12-stable-1ubuntu0.1.

Follow the dpkg install with 'apt-get install -f' so apt fetches whatever the
cached debs are missing, keeping the cache's speed without depending on the two
jobs landing on identical images.
@kibae
kibae merged commit e6da84b into main Sep 10, 2026
13 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant