[kibae]

Aligns server version references, CI/Docker runtime downloads, and the runtime version test with ONNX Runtime v1.31.1. This patch requires no server API adaptation; the seven version files are updated, and the macOS downloader handles the upstream archive-root change.

Upstream review: 1.31.0 → 1.31.1

• All public session headers are byte-identical: no new/removed header, enum member/value, SessionOptions setter/signature, or config key. Existing config-entry and CUDA-provider passthrough behavior remains applicable; no OpenAPI additions are needed.
• Upstream fixes an OpenVINO/Python provider-info namespace mismatch (#33327) and WASM Debug external-data tests (#33226). Neither requires a change to this C++ CPU/CUDA server.
• ONNX remains 1.22.0; datatype fixtures remain at opset 23. Existing packed UINT2, INT2, float8, and complex-allocation assumptions pass local tests; historical comments are preserved.
• Exact CPU amd64/arm64, CUDA12/CUDA13 amd64, Windows x64, and macOS arm64 archives are present. The local CUDA12 archive passed its published SHA256 checksum. The macOS archive no longer prefixes its root with ./; the downloader now preserves the root during extraction instead of stripping a fixed number of path components. No dependency changes are needed.

macOS archive compatibility

The initial macOS CI attempt failed before configuration: stripping two path components removed include/ and lib/ from the 1.31.1 archive. The 1.31.0 archive had a leading ./ component; 1.31.1 does not. Extracting the archive unchanged and selecting its exact versioned root handles both. Both real releases passed download, checksum, extraction, required-file checks and installation into isolated local paths. Native macOS build/test validation is provided by final-head CI.

CUDA alignment

Retain Ubuntu 24.04, CUDA 12.8.1/cuDNN 9.8.0.87 and CUDA 13.0.3/cuDNN 9.14.0.64. These are the highest published matching CUDA-minor cuDNN image patches on this Ubuntu base. CUDA 12.8.2 has no cuDNN-tagged image. Full NVIDIA_REQUIRE_* values for both builder/runtime stages match the published 1.31.0 server images.

The immutable Linux packaging stage, build command, cuda12 build image, cuda13 build image, and pipeline variables are unchanged from v1.31.0. No driver-support-floor change is introduced.

Local arm64 build environment

The initial CPU arm64 publication build failed in tar before any upload. The same archive extracted successfully on native x86_64 but failed under the local arm64 emulator with openat2(... O_LARGEFILE ... O_PATH ...) = EINVAL, matching tonistiigi/binfmt#285. Registering the fixed QEMU 10.2.3 arm64 emulator resolved the exact extraction reproducer and the complete arm64 image build. No repository change or new CI run was needed for this environment repair.

Validation and publication

• Fresh local Debug build with explicit ORT 1.31.1 paths; all 9 CTest executables passed, including CUDA, datatypes, TCP, HTTP, HTTPS, Swagger, and runtime version.
• CPU/CUDA12/CUDA13 amd64 Docker builds, version/session checks and inference.
• Linux/Windows/macOS/CodeQL CI on the final release commit.
• Publish normal 1.31.1 image tags and verify remote digests/platforms.
• Merge, tag the actual merge commit, and publish GitHub release.
• Manual Docker Hub description update.

CPU publication includes amd64 and arm64; image execution tests cover amd64 only. GPU results apply to the measured RTX 5080 host rather than establishing a universal minimum driver.