Tool Solutions: ML Frameworks r25.07

PyTorch build

https://github.com/ARM-software/Tool-Solutions/blob/r25.07/ML-Frameworks/pytorch-aarch64/README.md

Changed

• Updates Dockerfile to minimise layers in the final image.
• Updates hashes for:
  • PyTorch to 5dfd8a9c7a464bb42e81b8594eefd2fa865e5423, to 2.9.0.dev20250703 from viable/strict
  • ideep to 6eb12eaad5e0f7d8c8613c744ac8ba5a0843cb99, from ideep_pytorch, July 3rd
  • oneDNN to 0abfca1947b53c03ee74207e4710941ab6456f3b, from main, July 3rd
  • Compute Library to f69b48afcc59f1b3b0d4544289249bebba489f0a, from main, June 26th

TensorFlow build

https://github.com/ARM-software/Tool-Solutions/blob/r25.07/ML-Frameworks/tensorflow-aarch64/README.md

Changed

• Updates Dockerfile to minimise layers in the final image.
• Updates TensorFlow hash to 70d137e8794f97728e094608ec9ea1c6c9fd42af, 2.20.0-dev0 from nightly, July 1st

Full list of PRs

• Bump PyTorch & Tensorflow versions for 25.07 release; update PR commits where necessary by @murste01 in #338
• Add initial precommit ci by @Ryo-not-rio in #340
• Updates for 25.07 release. by @nSircombe in #339
• Fixes missing docker build arg in tensorflow build. by @nSircombe in #342
• Removes unwanted DOCKER_BUILDKIT=1 setting by @nSircombe in #343

Full Diff: r25.06...r25.07

Tool Solutions: ML Frameworks r25.06

PyTorch build

https://github.com/ARM-software/Tool-Solutions/blob/r25.06/ML-Frameworks/pytorch-aarch64/README.md

Added

• Adds WIP patch to update the OpenBLAS commit - pytorch/pytorch#151547
  • Speedups across most of the HF models via significant boost to SDPA layers.
  • Overall torch-bench pass-rate increased
• Adds support in github-apply-patch for local caching of patches.

Changed

• Updates protobuf from 5.29.2 to 5.29.5 in response to CVE-2025-4565
• Updates hashes for:
  • PyTorch to 3040ca6d0f8558e39919b14eebeacc34ddf980f5 2.8.0.dev20250611 from viable/strict
  • ideep to 2ef932a861439e4cc9bb8baee8424b57573de023 from ideep_pytorch, June 10.
  • oneDNN to 106a7b41bc4156297b8a88cd1951304b739cc427 form main, June 10th
  • Compute Library to 6bc1c7b8d0756272e2a97a7489e13de90f864326 from main, June 9th

Removed

• Removes WIP patches which have now landed in the upstream nightly PyTorch builds.

TensorFlow build

https://github.com/ARM-software/Tool-Solutions/blob/r25.06/ML-Frameworks/tensorflow-aarch64/README.md

Added

• Adds support in github-apply-patch for local caching of patches.
• Adds WIP patch to tensorflow/tensorflow#93951
  to update to oneDNN-3.7 + ACL-24.12, and address unit test failures

Changed

• Updates TensorFlow hash to 65781570c55d2338106767de200323f123c3f91f
  tensorflow-2.20.0.dev0 from main

Complete List of PRs

• Bump commits and update patches for 25.06 by @murste01 in #332
• Bump protobuf from 5.29.2 to 5.29.5 in /ML-Frameworks/pytorch-aarch64 by @dependabot in #333
• Re-add patch which pins PyTorch requirements by @jondea in #334
• Add Update OpenBLAS Commit patch to PyTorch by @murste01 in #336
• Add patch cache to github-apply-patch by @jondea in #335
• Update CHANGELOGs for 25.06 release. by @nSircombe in #337

New Contributors

• @murste01 made their first contribution in #332

Full Diff: r25.05...r25.06

Tool Solutions: ML Frameworks r25.05

PyTorch build

https://github.com/ARM-software/Tool-Solutions/blob/r25.05/ML-Frameworks/pytorch-aarch64/README.md

Added

• WIP patch to build a newer version of libgomp from source - pytorch/pytorch#152361
  Improve scaling for >16 threads.
• Support for authenticated GitHub access in apply-github-patch

Changed

• Updates hashes for:
  • PyTorch e872bf8f888bdbb27a03e03935db61babf7180b8, 2.8.0.dev20250430 from viable/strict
  • ideep to 2ef932a861439e4cc9bb8baee8424b57573de023 from ideep_pytorch
  • oneDNN to 69150ce5fe1f453af9125ca42a921e017092ccf7 from main
  • Compute Library to 334108c0efc512efdc9576ba957dbcf5b7ee168a, rc_25_04_29_0
• Updates existing WIP patches.

Removed

• Removes WIP patches which have now landed in the upstream nightly PyTorch builds.

Fixed

• Reverted this commit to add back --compile option. The issue was fixed upstream and is included in the bumped hash

TensorFlow build

https://github.com/ARM-software/Tool-Solutions/blob/r25.05/ML-Frameworks/tensorflow-aarch64/README.md

Added

• Support for authenticated GitHub access in apply-github-patch

Complete List of PRs

• Bump sources for 25.05 by @karmeh01 in #325
• Revert "Removes compile option from genai PyTorch examples" by @karmeh01 in #326
• Modify CHANGELOG.md to update unreleased section by @karmeh01 in #327
• Update tag naming scheme in CHANGELOGs by @nSircombe in #324
• Adds patch to build libgomp from sources by @nSircombe in #328
• Add option for auth. GitHub access and update CHANGELOGS for 25.05 by @nSircombe in #329
• Catches unbound GITHUB_TOKEN. by @nSircombe in #330
• Sets r25.05 release date. by @nSircombe in #331

New Contributors

• @karmeh01 made their first contribution in #325

Full Changelog: r25.04...r25.05

Tool Solutions: ML Frameworks r25.04

PyTorch build

https://github.com/ARM-software/Tool-Solutions/blob/r25.04/ML-Frameworks/pytorch-aarch64/README.md

Added

• Work in progress oneDNN patch, Enable jit conv for 128 with ~30% speed up for backward convolutions
• Add --wheel-only flag for only building the torch wheel
• Work in progress PyTorch patch for pinning the major versions of build dependencies - pytorch/pytorch#150833

Changed

• Updates hashes for:
  • PyTorch fc674b45d4d8edfd4c630d89f71ea9f85a2f61f2, 2.8.0.dev20250403 from viable/strict
  • ideep to 719d8e6cd7f7a0e01b155657526d693acf97c2b3 from ideep_pytorch
  • oneDNN to 5de25f354afee38bf2db61f485c729d30f62c611 from main
  • Compute Library to 9033bdacdc3840c80762bc56e8facb87b0e1048e, 25.03 release
  • OpenBLAS to edef2e4441e50e3a2da1920fdbde09101087c43d from main
• Updates existing WIP patches.
• Update torchvision from 0.22.0.dev20250305 to a more recent nightly build, 0.22.0.dev20250403

Removed

• Removes WIP patches which have now landed in the upstream nightly PyTorch builds.
• Removes --tags --force from git clone command, and adds --depth=1 to speedup the checkout.
• Temporarily removes --compile option from some examples due to an issue with pytorch/pytorch#147151
  the compile path does not work as expected in these cases.

TensorFlow build

https://github.com/ARM-software/Tool-Solutions/blob/r25.04/ML-Frameworks/tensorflow-aarch64/README.md

Added

• Enables patching of build outside of Bazel build.
• default num_threads to max for acl_threadpool, see www.github.com/tensorflow/uxlfoundation/oneDNN/2958

Changed

• Updates TensorFlow build to use oneDNN 3.7 + ACL 24.12, see www.github.com/tensorflow/tensorflow/pull/84975

Complete list of Pls

• fix: typo in stdout to stderr redirection by @Sqvid in #312
• Bump hashes ahead of r25.04 release. by @nSircombe in #313
• Adds back BF16 SVE intrinsics via new patch by @nSircombe in #314
• simplify apply-github-patch to implicitly use GitHub URL by @Sqvid in #315
• Update torchvision version and PyTorch hash to fix vision test failure. by @nSircombe in #316
• add common dirs and files to gitignore by @Sqvid in #317
• Add oneDNN sve_128 jit conv patch for PyTorch by @jondea in #319
• build TensorFlow against + oneDNN 3.7 + ACL 24.12 by @Sqvid in #318
• More dev improvements by @jondea in #320
• Pin major versions of PyTorch build deps by @jondea in #321
• update pytorch src's by @aditew01 in #322
• Update examples and CHANGELOG for r25.04 by @nSircombe in #323

New Contributors

• @Sqvid made their first contribution in #312

Full Diff: r25.03.1...r25.04

Tool Solutions: ML Frameworks r25.03.1

PyTorch build

https://github.com/ARM-software/Tool-Solutions/blob/r25.03.1/ML-Frameworks/pytorch-aarch64/README.md

Changed

• Move torch* packages to latest stable to fix build error
  • torchaudio==2.6.0.dev20250305 -> 2.6.0
  • torchdata~=0.7.1 -> 0.11.0
  • torchvision~=0.22.0.dev20250305 -> 0.21.0
  • torchtune==0.6.0.dev20250115 -> 0.5.0

Fixed

• Fix build error due to missing torchtune nightly build
• Fix build error due to auditwheel version change by applying commits from pytorch/pytorch#149471

TensorFlow build

https://github.com/ARM-software/Tool-Solutions/blob/r25.03.1/ML-Frameworks/tensorflow-aarch64/README.md
No changes to the TensorFlow build with this release.

Complete list of PRs

• Fix missing torch* nightlies by using latest stable by @jondea in #307
• Fix build error due to auditwheel version change by @jondea in #308
• Update CHANGELOGs for patch release. by @nSircombe in #309
• Add missing deps for torchchat example by @jondea in #310
• Update CHANGELOGs for patch release. by @nSircombe in #311

Full Diff: r25.03...r25.03.1

Tool Solutions: ML Frameworks r25.03

PyTorch build

https://github.com/ARM-software/Tool-Solutions/blob/r25.03/ML-Frameworks/pytorch-aarch64/README.md

Added

• Adds work-in-progress PyTorch PRs:
  • pytorch/pytorch#148542 - Enables direct use Compute Library in ATen.
  • pytorch/pytorch#147337 - Enables a fast path for static qlinear via Compute Library directly.
  • pytorch/pytorch#146620 - Enables qint8 and quint8 add via Compute Library directly. Speedup for OMP_NUM_THREADS=1 is ~15x, and ~5.4x for 32 threads.
  • pytorch/pytorch#148197 - Enables oneDNN dispatch for GEMM bf16bf16->bf16.
  • pytorch/pytorch#140159 - Enables gemm-bf16f32/
• Adds work-in-progress oneDNN PRs:
  • uxlfoundation/oneDNN#2838 - Dispatches fpmath_mode::bf16 conv to Compute Library.

Changed

• Updates hashes for:
  • PyTorch to e555c4d (2.7.0.dev20250305) from viable/strict branch.
  • ideep to 719d8e6 from ideep_pytorch branch.
  • oneDNN to 321c452 from main branch.
  • Compute Library to v25.02.1.
  • OpenBLAS to ef9e3f7 from main.
• Updates work-in-progress PyTorch PRs.
• Updates torchaudio to 2.6.0.dev20250305.
• Updates torchvision to 0.22.0.dev20250305.
• Dockerfile now upgrades pip before installing Python packages.
• git-shallow-clone function now supports cloning by tag as well as hash.

Removed

• Removes patches which have now been merged into the upstream branches.
• Removes static_quantize_conv example since pytorch/pytorch#141127 is no longer included in the build.

TensorFlow build

https://github.com/ARM-software/Tool-Solutions/blob/r25.03/ML-Frameworks/tensorflow-aarch64/README.md
This release establishes a new TensorFlow build based on the approach taken with PyTorch builds.

Changed

• Updated TensorFlow to 2.20.dev-6506da6

Removed

• Removed MLCommons examples and patches
• Removed cpp examples
• Removed all inline patches, work in progress features should now be applied
  from PRs using wget in ./get-source.sh.

Complete list of PRs:

• TensorFlow refactor: Build wheel + install in docker by @davsva01 in #293
• Update openblas hash by @aditew01 in #294
• Add test script for TensorFlow by @davsva01 in #295
• Remove bumping of oneDNN and ACL by @davsva01 in #296
• Updates PyTorch build. by @nSircombe in #297
• Fixes PyTorch build. by @nSircombe in #298
• Fixes mistake in OpenBLAS build by @nSircombe in #299
• Fixes OpenBLAS PR5157 hash. by @nSircombe in #300
• Updates for r25.03 by @nSircombe in #302
• Set weights_only=False in torch.load by @nSircombe in #303
• Fix broken links and incorrect docs by @davsva01 in #304
• Removes static_quatize_conv example and adds fpmath_mode::bf16 Compute Library dispatch patch by @nSircombe in #305
• Adds --force to git fetch command by @nSircombe in #306

Full Diff: pytorch-aarch64--r25.02...r25.03