Skip to content

v1.32.0

Latest

Choose a tag to compare

@kibae kibae released this 14 Nov 09:20
7bb2fb0

Aligns with upstream ONNX Runtime v1.32.0.

Updates the pinned runtime and recognizes the new float6 tensor type names. The official runtime still rejects float6 tensor allocation; this release does not claim float6 inference support.

🔧 Chores

  • fix: recognize upstream float6 tensor type names — map FLOAT6E2M3/FLOAT6E3M2 to their names and add type-name coverage plus a runtime-allocation tripwire.
  • ci: build CodeQL against the pinned runtime — replace automatic dependency discovery, which selected system ORT 1.23.2, with an explicit build using the downloaded runtime. The equivalent Release build passed locally.
  • ci: release 1.32.0 — update the seven version files, including ORT_VERSION and the runtime version test.

📋 Upstream 1.32.0 review

  • The ONNX 1.23.2 integration restores public float6 mappings and adds opset28 capability. Float6 decoding and fixture generation are deferred because the exact 1.32.0 runtime cannot allocate these tensors. Existing opset23 fixtures remain unchanged.
  • Existing tensor enum numbers, graph optimization/execution enums and SessionOptions setters are unchanged. Device-memory capacity and weightless compilation APIs are outside the existing JSON server API.
  • Weightless-mode, MoE offload, external-loader and CPU/CUDA tuning keys remain reachable through session_options.config_entries. Provider options retain existing forwarding behavior.
  • Packed UINT2/INT2, float8, complex-allocation and string-access assumptions passed local tests. Removed ACL and legacy TensorRT RTX bridges, DLPack and language-binding changes require no adaptation to the server's integrations.
  • Exact-version release asset names remain supported by existing downloaders.

🐳 Docker Images

  • 1.32.0-linux-cpu — amd64 and arm64.
  • 1.32.0-linux-cuda12 — amd64; CUDA 12.8.1, cuDNN 9.8.0.87.
  • 1.32.0-linux-cuda13 — amd64; CUDA 13.0.3, cuDNN 9.14.0.64.

Both GPU images retain their Ubuntu 24.04 builder/runtime bases and NVIDIA_REQUIRE constraints. Upstream's immutable packaging stage and build command select CUDA 12.8/13.0. The new upstream producer image is private and returned HTTP401, so its exact cuDNN provenance could not be verified. Existing bases were retained by explicit user decision and validated with the released runtime. No driver-support-floor change is introduced.

Local validation passed all nine Debug CTest executables including CUDA. The final PR head passed Linux (Ubuntu 22.04/24.04, GCC/Clang), Windows, macOS and CodeQL workflows. All three amd64 image tests confirmed ORT 1.32.0, session creation and inference returning [13, 14, 11]. Both CUDA sessions reported device ID 0 on the RTX 5080 host. These measurements do not establish a universal minimum driver.

Published image digests:

Tag Digest
1.32.0-linux-cpu sha256:a0ed0317da454d82d12e8f99e1a6b6e674288e7f97ac4bbb23b44cf44075a63f
1.32.0-linux-cuda12 sha256:6248d0503bc47e14e994beeaec1944fd7b49886527003c7eaa32e620e882c09d
1.32.0-linux-cuda13 sha256:47994d7c804844830fb35d65025ab5aa0020e96f3d972c61149375bbf1ef1294

CPU arm64 publication does not include an arm64 execution test. No new opset28 model execution coverage is claimed.

Note: the telemetry guidance documented since v1.29.0 continues to apply: use ORT_DISABLE_TELEMETRY=1 to opt out.

Full Changelog: v1.31.1...v1.32.0