Repository navigation
Tags
Tags
ci(onnxruntime-release): drop the -full variant
The full variant has no consumers left: since the EE DeepDoc migration
rebase, both EE and OSS link the minimal build (EE rec's HardSwish is
decomposed to Mul/Clip/Div and every EE weight ships as .ort). Its smoke
test ("full build loads .onnx with HardSwish") could never pass anyway —
vanilla ORT v1.29.0 has no HardSwish CPU kernel — which is why every
patch2 CI run failed and the release job never ran (the patch2 artifact
ended up being built locally instead).
Keep the four minimal targets (linux-x86_64 / linux-aarch64 / osx-arm64 /
osx-universal) with their weight-load smoke test and nm portability gate.
onnxruntime: add shared-initializer C APIs for in-process weight sharing Add three custom C API entries (SessionGetInitializerCount/Name/GetInitializer) that extract already-loaded constant initializers from a session so they can be re-injected into other sessions via AddInitializer, sharing one copy of the weights in memory instead of deserializing them per session. The change is shipped as a vendored patch (onnxruntime/patches/) applied by init-submodules.sh on top of the pinned upstream tag, because the release CI re-fetches pristine microsoft/onnxruntime and ignores the submodule commit. Also add tools/verify_shared_weights/verify.cpp, a standalone proof-of-concept that loads a .ort model, extracts its weights once, injects them into N sessions, and asserts bitwise-identical inference output while showing the per-session marginal memory stays flat (weights are shared). The release workflow now builds and runs it on linux-x86_64 to gate the published artifact.