Skip to content

Tool Solutions: ML Frameworks r26.02

Latest

Choose a tag to compare

@danwhittaker-arm danwhittaker-arm released this 12 Feb 09:42
01bf1d3

PyTorch build

  • The pytorch-aarch64 r26.02 README.md is available here.
  • A Docker image is available on Docker Hub under armlimited/pytorch-arm-neoverse.
  • Attached as artefacts are:
    • A full Software Bill of Materials (SBOM) as Tool-Solutions_r26.02_SPDX_SBOM.json, and;
    • License texts as Tool-Solutions_r26.02_licenses.zip which cover all packages included in the built image, in addition to the base Ubuntu image.

Added

  • Adds PyTorch PR #170600, to patch incremental build support.
  • Adds PyTorch PR #170062, to add ccache support to ACL/OpenBLAS and manywheel build script.

Changed

  • Updates hashes for:
    • PYTORCH_HASH to 77da53a7356e033e3fc1e03fdd960fc4ad117882, 2.11.0.dev20260129 from viable/strict, Jan 29th.
    • IDEEP_HASH to bbb9ffb9e0c401ca058b7f35a6ebe7d0e08ffd34, from ideep_pytorch, Jan 30th.
    • ONEDNN_HASH to 804f364c04ad8a763d534abaabc99bf99c2754e0, from main, Jan 30th.
    • TORCH_AO_HASH to 30fcb156945ecacd515775414d37c09bfe60727e, from main, Jan 30th.
    • KLEIDIAI_HASH to 5addaad73ebbb02e7dde6c50fff3bdb2ae8c407f, v1.20.0 from main, Jan 30th.
  • Replaces ACL_VERSION=v52.6.0 with ACL_VERSION=v52.8.0, from main, Jan 30th.
  • Updates torchvision from 0.25.0.dev20251104 to 0.25.0.dev20260129.
  • Updates PyTorch and TorchAO manylinux AArch64 builder image from cpu-aarch64-d8be0384e085f551506bd739678109fa0f5ee7ac
    to cpu-aarch64-69d4c1f80b5e7da224d4f9c2170ef100e75dfe03.
  • Updates the URL used to download weights for ResNet50 used by classify_image.py.
  • Updates Ubuntu and Python version to 24.04 and 3.12, respectively.

Removed

  • Removes patches which are no longer required or have landed upstream.

TensorFlow build

Changed

  • Updates TensorFlow hash to 3d4958d107b77a2e4398c066971a099a369ef739, from nightly, Feb 2nd
  • Updates Ubuntu and Python version to 24.04 and 3.12, respectively.

Note

Running tensorflow-aarch64/examples/run_unit_tests.sh with ONEDNN_DEFAULT_FPMATH_MODE=BF16 can lead to some failing tests. Addressing these failures is a WIP. However, all testing is expected to pass with the default setting ONEDNN_DEFAULT_FPMATH_MODE=FP32.

What's Changed

New Contributors

Full Changelog: r25.12...r26.02