Patch build and update hashes.

Co-authored-by: Robert Hardwick <Robert.Hardwick@arm.com>

Author: puneetmatharu

ML-Frameworks/pytorch-aarch64/CHANGELOG.md

@@ -8,12 +8,28 @@ where `YY` is the year, and `MM` the month of the increment.
 ## [unreleased]
 
 ### Added
- - Adds PyTorch [PR][https://github.com/pytorch/pytorch/pull/144992] Enable fp16 linear layers in PyTorch via ACL
- - Adds PyTorch [PR](https://github.com/pytorch/pytorch/pull/167328) Build cpuinfo into c10 shared library
+ - Adds PyTorch [PR #167720](https://github.com/pytorch/pytorch/pull/167720), to allow missing cutlass file 
+   if CUDA disabled.
+ - Adds oneDNN [PR #4237](https://github.com/uxlfoundation/oneDNN/pull/4237), which improves cache
+   blocking 4/8 inner blocks.
+ - Adds PyTorch [PR](https://github.com/pytorch/pytorch/pull/167328) Build cpuinfo into c10 shared library.
+ - Adds PyTorch [PR](https://github.com/pytorch/pytorch/pull/144992) Enable fp16 linear layers in PyTorch via ACL.
 
 ### Changed
+ - Now building ACL with PyTorch `install_acl.sh` script.
+ - PyTorch manylinux AArch64 builder image from `cpu-aarch64-a040006da76a51c4f660331e9abd3affe5a4bd81`
+   to `cpu-aarch64-d8be0384e085f551506bd739678109fa0f5ee7ac`
+ - Updates hashes for:
+   - `PYTORCH_HASH` to `5ce4a8b49f9986b050a9f6fcc7dd4cf999baa509`, 2.10.0.dev20251112 from viable/strict, Nov 12th.
+   - `IDEEP_HASH` to `927570638b237b0e39fb0626a868adffdbf70bbb`, from ideep_pytorch, October 20th.
+   - `ONEDNN_HASH` to `80886d0559482dfe2019c2ae83eebd6d0d3a17d4`, from main, Nov 9th.
+   - `TORCH_AO_HASH` to `17867e6788e4889b294449770f0275045384eab2`, from main, Nov 8th.
+   - `KLEIDIAI_HASH` to `7bf4de9a56106f0fb0d57dfabeb4c7a2668deaf6`, v1.16.0 from main, Nov 10th.
+ - Replaced `ACL_HASH=007264fa740de5723ebddef16b7bb3657692c088` with `ACL_VERSION=v52.6.0`, from main, October 31st.
+ - Update `torchvision` from 0.25.0.dev20250923 to 0.25.0.dev20251104.
 
 ### Removed
+ - Removes patches which are no longer required.
 
 ### Fixed
 

ML-Frameworks/pytorch-aarch64/Dockerfile

@@ -102,7 +102,7 @@ COPY $TORCH_AO_WHEEL /home/$DOCKER_USER/
 # and we usually don't need features from the nightlies.
 # Note: torchvision is pinned to a nightly build, this can be updated
 # at the next vision release, and the `--extra-index-url` removed.
-RUN pip install --pre torchvision==0.25.0.dev20250923 --index-url https://download.pytorch.org/whl/nightly/cpu --no-deps
+RUN pip install --pre torchvision==0.25.0.dev20251104 --index-url https://download.pytorch.org/whl/nightly/cpu --no-deps
 
 # We need --no-deps because the torch version won't match the versions on torch*
 RUN pip install "$(basename "$TORCH_WHEEL")" --no-deps \

ML-Frameworks/pytorch-aarch64/build-wheel.sh

@@ -30,21 +30,34 @@
 set -eux -o pipefail
 
 PYTHON_VERSION="3.10"
+OPENBLAS_VERSION="v0.3.30"
+ACL_VERSION="v52.6.0"
 
 # Specify DOCKER_IMAGE_MIRROR if you want to use a mirror of hub.docker.com
-IMAGE_NAME="${DOCKER_IMAGE_MIRROR:-}pytorch/manylinux2_28_aarch64-builder:cpu-aarch64-a040006da76a51c4f660331e9abd3affe5a4bd81"
+IMAGE_NAME="${DOCKER_IMAGE_MIRROR:-}pytorch/manylinux2_28_aarch64-builder:cpu-aarch64-d8be0384e085f551506bd739678109fa0f5ee7ac"
 TORCH_BUILD_CONTAINER_ID_FILE="${PWD}/.torch_build_container_id"
 
 # Output dir for PyTorch wheel and other artifacts
 OUTPUT_DIR=${OUTPUT_DIR:-"${PWD}/results"}
-PYTORCH_FINAL_PACKAGE_DIR=$OUTPUT_DIR
 
+# Where folders sit locally
 PYTORCH_HOST_DIR="${PWD}/pytorch"
 ACL_HOST_DIR="${PWD}/ComputeLibrary"
+OPENSSL_HOST_DIR="/opt/openssl"
+PYTORCH_FINAL_PACKAGE_DIR="${OUTPUT_DIR}"
 
+# Where folders sit mounted in the container
 PYTORCH_ROOT=/pytorch
-UTILS="/utils"
-COMMON_UTILS="/common_utils"
+ACL_ROOT=/acl
+UTILS=/utils
+COMMON_UTILS=/common_utils
+
+# Want a CPU build
+DESIRED_CUDA=cpu
+GPU_ARCH_TYPE=cpu-aarch64
+
+# Affects the number of jobs used in install_acl.sh and install_openblas.sh
+NPROC=${NPROC:-$(nproc --ignore=2)}
 
 if [ -f "$TORCH_BUILD_CONTAINER_ID_FILE" ]; then
     TORCH_BUILD_CONTAINER=$(cat $TORCH_BUILD_CONTAINER_ID_FILE)
@@ -55,26 +68,28 @@ else
 fi
 
 if ! docker container inspect $TORCH_BUILD_CONTAINER >/dev/null 2>&1 ; then
-
     # Based on environment used in pytorch/.github/workflows/_binary-build-linux.yml
     # and pytorch/.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
     TORCH_BUILD_CONTAINER=$(docker run -t -d \
+        -e NPROC=${NPROC} \
+        -e OPENBLAS_VERSION=${OPENBLAS_VERSION} \
+        -e ACL_VERSION=${ACL_VERSION} \
         -e BINARY_ENV_FILE=/tmp/env \
         -e BUILD_ENVIRONMENT=linux-aarch64-binary-manywheel \
-        -e DESIRED_CUDA=cpu \
-        -e DESIRED_PYTHON=$PYTHON_VERSION \
+        -e DESIRED_CUDA=${DESIRED_CUDA} \
+        -e DESIRED_PYTHON=${PYTHON_VERSION} \
         -e GITHUB_ACTIONS=0 \
-        -e GPU_ARCH_TYPE=cpu-aarch64 \
+        -e GPU_ARCH_TYPE=${GPU_ARCH_TYPE} \
         -e PACKAGE_TYPE=manywheel \
-        -e PYTORCH_FINAL_PACKAGE_DIR=$PYTORCH_FINAL_PACKAGE_DIR \
-        -e PYTORCH_ROOT=$PYTORCH_ROOT \
+        -e PYTORCH_FINAL_PACKAGE_DIR="${PYTORCH_FINAL_PACKAGE_DIR}" \
+        -e PYTORCH_ROOT="${PYTORCH_ROOT}" \
         -e SKIP_ALL_TESTS=1 \
         -e PYTORCH_EXTRA_INSTALL_REQUIREMENTS="nvidia-cuda-nvrtc-cu12==12.1.105; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.1.105; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.1.105; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==8.9.2.26; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.1.3.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.0.2.54; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.2.106; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.4.5.107; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.1.0.106; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.19.3; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.1.105; platform_system == 'Linux' and platform_machine == 'x86_64'" \
-        -e OPENSSL_ROOT_DIR=/opt/openssl \
-        -e CMAKE_INCLUDE_PATH="/opt/openssl/include" \
+        -e OPENSSL_ROOT_DIR="${OPENSSL_HOST_DIR}" \
+        -e CMAKE_INCLUDE_PATH="${OPENSSL_HOST_DIR}/include" \
         -v "${PYTORCH_HOST_DIR}:${PYTORCH_ROOT}" \
         -v "${PYTORCH_FINAL_PACKAGE_DIR}:/artifacts" \
-        -v "${ACL_HOST_DIR}:/ComputeLibrary" \
+        -v "${ACL_HOST_DIR}:${ACL_ROOT}" \
         -v "${PWD}/utils:${UTILS}" \
         -v "${PWD}/../utils:${COMMON_UTILS}" \
         -w / \
@@ -83,14 +98,12 @@ if ! docker container inspect $TORCH_BUILD_CONTAINER >/dev/null 2>&1 ; then
     # Currently changes in these scripts will not be applied without a clean
     # build, which is not ideal for dev work. But we have to balance this with
     # extra time/network traffic when rebuilding many times.
-    docker exec -t $TORCH_BUILD_CONTAINER bash -c $PYTORCH_ROOT/.circleci/scripts/binary_populate_env.sh
+    docker exec -t $TORCH_BUILD_CONTAINER bash -c "$PYTORCH_ROOT/.circleci/scripts/binary_populate_env.sh"
     docker exec -t $TORCH_BUILD_CONTAINER bash -c "$PYTORCH_ROOT/.ci/aarch64_linux/aarch64_ci_setup.sh"
 
-    # Build a newer version of libgomp from source (see https://github.com/pytorch/pytorch/pull/152361)
-    docker exec -t $TORCH_BUILD_CONTAINER bash -c "bash $PYTORCH_ROOT/.ci/docker/common/install_libgomp.sh"
-
     # This must be in this if block because it cannot handle being called twice
-    docker exec -t $TORCH_BUILD_CONTAINER bash -c "bash $UTILS/build_openblas.sh"
+    docker exec -t $TORCH_BUILD_CONTAINER bash -c "bash $PYTORCH_ROOT/.ci/docker/common/install_acl.sh"
+    docker exec -t $TORCH_BUILD_CONTAINER bash -c "bash $PYTORCH_ROOT/.ci/docker/common/install_openblas.sh"
 
     echo "Storing torch build container id in $TORCH_BUILD_CONTAINER_ID_FILE for reuse: $TORCH_BUILD_CONTAINER"
     echo $TORCH_BUILD_CONTAINER > "$TORCH_BUILD_CONTAINER_ID_FILE"

ML-Frameworks/pytorch-aarch64/bump-sources.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
 # *******************************************************************************
-# Copyright 2024 Arm Limited and affiliates.
+# Copyright 2024-2025 Arm Limited and affiliates.
 # SPDX-License-Identifier: Apache-2.0
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -42,13 +42,7 @@ function git-bump {
     )
 )
 
-(
-    cd ComputeLibrary
-    git-bump main
-)
-
 echo "Put this into your get-sources.sh file"
 echo PYTORCH_HASH=$(cd pytorch && git rev-parse HEAD)
 echo IDEEP_HASH=$(cd pytorch/third_party/ideep && git rev-parse HEAD)
 echo ONEDNN_HASH=$(cd pytorch/third_party/ideep/mkl-dnn && git rev-parse HEAD)
-echo ACL_HASH=$(cd ComputeLibrary && git rev-parse HEAD)

ML-Frameworks/pytorch-aarch64/get-source.sh

@@ -20,62 +20,41 @@
 source ../utils/git-utils.sh
 
 set -eux -o pipefail
-PYTORCH_HASH=45d9dcccc5bfb46a9eaca037270278bc3d7c95ea  # 2.10.0.dev20250923 from viable/strict, September 23rd
-IDEEP_HASH=fd11055f4800ac89291e30b5387a79a1e6496aa6    # From ideep_pytorch, September 10th
-ONEDNN_HASH=9e8f619477469ed75d323d4915bf7a2513f01713   # From main, September 23rd
-ACL_HASH=531a4968cecb7b4fc0a3b65482e2c524289e087e      # From main, September 23rd
-TORCH_AO_HASH=8e2ca35ea603349e71c2467e10fd371e34bf52bc # From main, September 23rd
-KLEIDIAI_HASH=bd2e6ae060014035e25bf4986be682762c446c2d # v1.14 from main
+
+PYTORCH_HASH=5ce4a8b49f9986b050a9f6fcc7dd4cf999baa509  # 2.10.0.dev20251112 from viable/strict, Nov 12th
+IDEEP_HASH=927570638b237b0e39fb0626a868adffdbf70bbb    # From ideep_pytorch, October 20th
+ONEDNN_HASH=80886d0559482dfe2019c2ae83eebd6d0d3a17d4   # From main, Nov 9th
+TORCH_AO_HASH=17867e6788e4889b294449770f0275045384eab2 # From main, Nov 8th
+KLEIDIAI_HASH=7bf4de9a56106f0fb0d57dfabeb4c7a2668deaf6 # v1.16.0 from main, Nov 10th
 
 git-shallow-clone https://github.com/pytorch/pytorch.git $PYTORCH_HASH
 (
     # Apply patches to PyTorch build
     cd pytorch
 
-    # https://github.com/pytorch/pytorch/pull/152361 - Build libgomp (gcc-11) from source
-    apply-github-patch pytorch/pytorch 7c54b6b07558c330ee2f95b4793edb3bfbb814c9
-    apply-github-patch pytorch/pytorch 3e17ce1619b2d02543a619f6217919b5adb36123
-    apply-github-patch pytorch/pytorch 2c884c2b580a93cd0b1e5eea36aa24e3acab91a9
-    apply-github-patch pytorch/pytorch c4c280eb27859221159108356b7c91376202cdd8
-
     # https://github.com/pytorch/pytorch/pull/160184 - Draft: separate reqs for manywheel build and pin
     # Note: as part of this patch, setuptools is pinned to ~= 78.1.1 which is not affected by
     # CVE-2025-47273 and CVE-2024-6345
     apply-github-patch pytorch/pytorch 4d344570e5a114fa522e3370c5d59161e2ed8619
 
-    # https://github.com/pytorch/pytorch/pull/158250 - Ingtegrate INT4→BF16 via KleidiAI, with fallback
-    apply-github-patch pytorch/pytorch 7c55f2af0adf9ce62c2226e739a3c84902fe0048
-    apply-github-patch pytorch/pytorch 8c27947566c85d44bc7dcd7189db5da608453bbb
-    apply-github-patch pytorch/pytorch 15d78c833b032d3c76b70b12a5f2762fa87d2640
-    apply-github-patch pytorch/pytorch 186cbcf641f99a301cb26013e8d74d444ad1dcb9
-    apply-github-patch pytorch/pytorch a6128ce3a0d2080d80e6fa59061d6c085865376c
-    apply-github-patch pytorch/pytorch 52ee4ddc9a5a9cec8793b1ffeb0d74113e3da417
-    apply-github-patch pytorch/pytorch ab2a6760e4a4891accbacb9187cf3782cb4b55c3
-    apply-github-patch pytorch/pytorch 93384233d166dccab5724f9d2e50b6eb3f47cbe6
-    apply-github-patch pytorch/pytorch 9f6d435629dd251620a1e17b8baa6bc18997f8ab
-    apply-github-patch pytorch/pytorch b68b7867a72fe2ef4c38f9a3cdd93693700a182e
-
-    # https://github.com/pytorch/pytorch/pull/161049 - optimised SVE exp_u20 implementation
-    # based on Arm Optimised Routines - https://github.com/ARM-software/optimized-routines
-    apply-github-patch pytorch/pytorch 3de5651bafcdabbc52d5205c0de3976188eba7fb
-
-    # https://github.com/pytorch/pytorch/pull/160080 - VLA Vectorized POC
-    apply-github-patch pytorch/pytorch e84eabd4f9761362ba081512b2922b4f18c97d41
-    apply-github-patch pytorch/pytorch a49982892480af69fae8bb19505b31b3304cda7a
-    apply-github-patch pytorch/pytorch 6ca9dc026d8d65c575c880ebe8b678f724d609a1
-    apply-github-patch pytorch/pytorch 3b92a1adfe40ca9c37e7db523eccaad4358d949c
-    apply-github-patch pytorch/pytorch 0384f48daa4b27d155632329521128212dd6fda3
-    apply-github-patch pytorch/pytorch bf4b0e8c41c75d9106e2e432c6b9a00319295930
-    apply-github-patch pytorch/pytorch dae9a71d99faa19764c47c602cb92bbf72ca7260
-    apply-github-patch pytorch/pytorch 8ac81dba2155808427ec3943c4d057f6b05b23d6
+    # https://github.com/pytorch/pytorch/pull/167720 - Allow missing cutlass file if CUDA disabled
+    apply-github-patch pytorch/pytorch 18f9ef2fe29b10b385f25eb6c98e3ac06227d2d9
+
+    # https://github.com/pytorch/pytorch/pull/158250 - Integrate INT4→BF16 via KleidiAI, with fallback
+    apply-github-patch pytorch/pytorch a9ec9d509167bfd33cbcd168cb40d183acf9c13a
+    apply-github-patch pytorch/pytorch 67f1076366b88c6617256236020b58da00665ed4
+    apply-github-patch pytorch/pytorch 99c57644d5d8a9359b6b98ac7bb96787ac594606
+    apply-github-patch pytorch/pytorch a770fb9a9786d7ce39a3b066809fa8c0de7d47d5
+    apply-github-patch pytorch/pytorch 30dd7406155c51b033b5e8a9c5a453fa59599db8
+    apply-github-patch pytorch/pytorch 00b919af8e7bb50f52ec45fdad09304d4104464a
+    apply-github-patch pytorch/pytorch fe40a60d7ad506aab016e66b53fdf0fc4f83b7a1
+    apply-github-patch pytorch/pytorch 89fc01183127da738fc3723747f7bf0721fe9e09
+    apply-github-patch pytorch/pytorch 23b4c39348426914cf3e6770dfaff0745245976c
+    apply-github-patch pytorch/pytorch c5e778f5d4cac56b9d96f666c3082aab244e662f
 
     # https://github.com/pytorch/pytorch/pull/159859 - PoC LUT optimisation for GELU bf16 operators
     apply-github-patch pytorch/pytorch ebcc874e317f9563ab770fc5c27df969e0438a5e
 
-    # https://github.com/pytorch/pytorch/pull/164741 - Enable mimalloc on non-Windows platforms and
-    # make default for AArch64 builds
-    apply-github-patch pytorch/pytorch 9f6a4018f6e7d77d4ac974a38f68fbd7c8eef25c
-
     # https://github.com/pytorch/pytorch/pull/144992 - Enable fp16 linear layers in PyTorch via ACL
     apply-github-patch pytorch/pytorch 00076d21ed6cd7df2a61165b1fb1d0a436f4e403
     apply-github-patch pytorch/pytorch 850db41fe6d33c6460740da781b40e009f04a47c
@@ -120,8 +99,9 @@ git-shallow-clone https://github.com/pytorch/pytorch.git $PYTORCH_HASH
         (
             cd mkl-dnn
             git fetch origin $ONEDNN_HASH && git clean -f && git checkout -f FETCH_HEAD
-            # https://github.com/uxlfoundation/oneDNN/pull/3022 - cpu: aarch64: enable jit conv for 128
-            apply-github-patch uxlfoundation/oneDNN b43cc9c4526c16a292860dadf34b3585b1f33531
+
+            # https://github.com/uxlfoundation/oneDNN/pull/4237 - cpu: aarch64: jit_reorder: cache blocking 4/8 inner blocks
+            apply-github-patch uxlfoundation/oneDNN 8bdff1a2a6625432701363185a9bd34f7c22f241
         )
     )
     (
@@ -136,6 +116,4 @@ git-shallow-clone https://github.com/pytorch/pytorch.git $PYTORCH_HASH
     fi
 )
 
-git-shallow-clone https://github.com/ARM-software/ComputeLibrary.git $ACL_HASH
-
 git-shallow-clone https://github.com/pytorch/ao.git $TORCH_AO_HASH

ML-Frameworks/tensorflow-aarch64/CHANGELOG.md

@@ -11,6 +11,7 @@ where `YY` is the year, and `MM` the month of the increment.
  -  Adds patch fixing CPUIDInfo initialization, re-enabling oneDNN+ACL path
 
 ### Changed
+ - Updates TensorFlow hash to f40a10e064561ed66c03359c7751ccc5ae03dc4d, from nightly, Nov 10th
 
 ### Removed
 

ML-Frameworks/tensorflow-aarch64/get-source.sh

@@ -21,7 +21,7 @@ source ../utils/git-utils.sh
 
 set -eux -o pipefail
 
-TENSORFLOW_HASH=6aa8fd07270293b918255af5988aef45b844c5b4 # from nightly, September 22nd
+TENSORFLOW_HASH=f40a10e064561ed66c03359c7751ccc5ae03dc4d # from nightly, Nov 10th
 
 git-shallow-clone https://github.com/tensorflow/tensorflow.git $TENSORFLOW_HASH