ARM-software · puneetmatharu · Apr 3, 2026 · Mar 23, 2026 · Mar 31, 2026
diff --git a/ML-Frameworks/pytorch-aarch64/CHANGELOG.md b/ML-Frameworks/pytorch-aarch64/CHANGELOG.md
@@ -21,6 +21,27 @@ where `YY` is the year, and `MM` the month of the increment.
 
 ### Fixed
 
+## [r26.04] 2026-04-10
+
+### Added
+ - Adds PyTorch [PR #177009](https://github.com/pytorch/pytorch/pull/177009), to accelerate SDPA on Arm CPUs.
+ - Adds oneDNN [PR #4911](https://github.com/uxlfoundation/oneDNN/pull/4911), to add LUT-based `gelu_erf` JIT implementations.
+
+### Changed
+ - Updates hashes for:
+   - `PYTORCH_HASH` to `fb69aa6b76c7ddbcac13ac3dd8b14625f4352ffd`, 2.12.0.dev20260328 from viable/strict, Mar 28th.
+   - `IDEEP_HASH` to `cbbfd4ad7c5ac6d7683af571055e95c948d8cf54`, from ideep_pytorch, Mar 17th.
+   - `ONEDNN_HASH` to `abc14842394f985313191bc1e3c69bb7f8cecd23`, from main, Mar 27th.
+   - `TORCH_AO_HASH` to `d5814ae2edf434bb7927d3d992b33239e6e2aa1b`, from main, Mar 29th.
+   - `KLEIDIAI_HASH` to `6c544373d1731d2b74435fc02f8bad5f4631b0b1`, v1.23.0 from main, Mar 25th.
+ - Updates `OPENBLAS_VERSION` from v0.3.30 to `d26960a21ec5da7f77377f28bd6e230060841ae0` from main, Mar 27th.
+ - Updates `torchvision` from 0.25.0.dev20260130 to 0.26.0.dev20260329.
+
+### Removed
+ - Removes PyTorch [PR #159859](https://github.com/pytorch/pytorch/pull/159859), which added PoC LUT optimisation for GELU bf16 operators.
+
+### Fixed
+
 ## [r26.03] 2026-03-16
 https://github.com/ARM-software/Tool-Solutions/tree/r26.03
 

diff --git a/ML-Frameworks/pytorch-aarch64/build-torch-ao-wheel.sh b/ML-Frameworks/pytorch-aarch64/build-torch-ao-wheel.sh
@@ -60,7 +60,10 @@ if ! docker container inspect "$TORCH_BUILD_CONTAINER" >/dev/null 2>&1 ; then
     docker_exec rm -rf /acl
 
     docker_exec python${PYTHON_VERSION} -m venv ${TEST_VENV_CONTAINER_DIR}
-    docker_exec bash -c "source ${TEST_VENV_CONTAINER_DIR}/bin/activate && pip install typing_extensions torch wheel numpy --no-deps"
+    docker_exec bash -c "
+        source ${TEST_VENV_CONTAINER_DIR}/bin/activate && \
+        pip install typing_extensions wheel numpy --no-deps && \
+        pip install torch --index-url https://download.pytorch.org/whl/cpu"
 
     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"

diff --git a/ML-Frameworks/pytorch-aarch64/build-wheel.sh b/ML-Frameworks/pytorch-aarch64/build-wheel.sh
@@ -123,7 +123,7 @@ if ! docker container inspect "$TORCH_BUILD_CONTAINER" >/dev/null 2>&1 ; then
     # Currently changes in these scripts will not be applied without a clean
     # build, which is not ideal for dev work. But we have to balance this with
     # extra time/network traffic when rebuilding many times.
-    docker_exec bash "${PYTORCH_CONTAINER_DIR}/.circleci/scripts/binary_populate_env.sh"
+    docker_exec bash "${PYTORCH_CONTAINER_DIR}/.ci/pytorch/binary_populate_env.sh"
 
     # Install scons for the Compute Library (ACL) build
     docker_exec ${PYTHON_CONTAINER_BIN}/python3 -m pip install scons==4.7.0

diff --git a/ML-Frameworks/pytorch-aarch64/get-source.sh b/ML-Frameworks/pytorch-aarch64/get-source.sh
@@ -24,20 +24,15 @@ git-shallow-clone https://github.com/pytorch/pytorch.git $PYTORCH_HASH
     # https://github.com/pytorch/pytorch/pull/170600 - Gate deletion of clean-up steps in build_common.sh
     apply-github-patch pytorch/pytorch e368ec2693b8b2b8ba35d0913f1d663ba2fdc804
 
-    # FIXME: Temporarily disabled; to be updated in a later PR
-    # # 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/159859 - PoC LUT optimisation for GELU bf16 operators
-    apply-github-patch pytorch/pytorch ebcc874e317f9563ab770fc5c27df969e0438a5e
-
     # https://github.com/pytorch/pytorch/pull/167328 - Build cpuinfo into c10 shared library
     apply-github-patch pytorch/pytorch 7c053dd1582b778c81101dd452708c4ec6e58233
     apply-github-patch pytorch/pytorch b1782bbe0eda5957870e2f6e95b8f167e04843cb
     apply-github-patch pytorch/pytorch 337925aed2babb3ef7808f78536bbbc9df346a4f
 
+    # https://github.com/pytorch/pytorch/pull/177009 - Accelerate SDPA on Arm CPUs: Unroll exp_sum and max_mul kernels
+    apply-github-patch pytorch/pytorch f6a6cc94e92636902acbbe443d00cec4a8efa12d
+    apply-github-patch pytorch/pytorch 253a1ea6db210c01178dd365922eb49fe4892fe7
+
     # Remove deps that we don't need for manylinux AArch64 CPU builds before fetching.
     # Only used when jni.h is present (see .ci/pytorch/build.sh:116), which is not the case for manylinux
     git rm android/libs/fbjni
@@ -78,6 +73,9 @@ 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/4911 - cpu: aarch64: add LUT-based gelu_erf JIT implementations #4911
+            apply-github-patch uxlfoundation/oneDNN 86661869d4267b49db667ed7946eefba4e44d2c0
         )
     )
     (

diff --git a/ML-Frameworks/pytorch-aarch64/versions.sh b/ML-Frameworks/pytorch-aarch64/versions.sh
@@ -9,15 +9,15 @@
 # For information on how to update the versions below, read the README.md.
 
 # get-source.sh deps
-PYTORCH_HASH=434b8cfdfac8ef7c741385561e340781ad76512f   # 2.12.0.dev20260227 from viable/strict, Feb 27th
-IDEEP_HASH=bbb9ffb9e0c401ca058b7f35a6ebe7d0e08ffd34     # From ideep_pytorch, Jan 16th
-ONEDNN_HASH=a83f8b4c7ca45fb339f3ecdf82895bc496941808    # From main, Feb 27th
-TORCH_AO_HASH=9bdc0ca87c1134b7c1dedaa9512233b726f22955  # From main, Feb 27th
-KLEIDIAI_HASH=98a6df72bdbb566bc7d8ba13d71991bcd94a8393  # v1.22.0 from main, Feb 19th
+PYTORCH_HASH=fb69aa6b76c7ddbcac13ac3dd8b14625f4352ffd   # 2.12.0.dev20260328 from viable/strict, Mar 28th
+IDEEP_HASH=cbbfd4ad7c5ac6d7683af571055e95c948d8cf54     # From ideep_pytorch, Mar 17th
+ONEDNN_HASH=abc14842394f985313191bc1e3c69bb7f8cecd23    # From main, Mar 27th
+TORCH_AO_HASH=d5814ae2edf434bb7927d3d992b33239e6e2aa1b  # From main, Mar 29th
+KLEIDIAI_HASH=6c544373d1731d2b74435fc02f8bad5f4631b0b1  # v1.23.0 from main, Mar 25th
 
 # build-wheel.sh deps
-ACL_VERSION="v52.8.0"
-OPENBLAS_VERSION="v0.3.30"
+ACL_VERSION="v52.8.0"   # Jan 23rd
+OPENBLAS_VERSION="d26960a21ec5da7f77377f28bd6e230060841ae0"  # Mar 27th
 
 # Dockerfile deps
-TORCHVISION_NIGHTLY="0.25.0.dev20260130"
+TORCHVISION_NIGHTLY="0.26.0.dev20260329"
diff --git a/ML-Frameworks/tensorflow-aarch64/CHANGELOG.md b/ML-Frameworks/tensorflow-aarch64/CHANGELOG.md
@@ -21,6 +21,18 @@ where `YY` is the year, and `MM` the month of the increment.
 
 ### Fixed
 
+## [r26.04] 2026-04-10
+
+### Added
+
+### Changed
+ - Replaces patch to bump ComputeLibrary version [from v24.12 to v52.4.0](https://github.com/tensorflow/tensorflow/pull/100882) with [one to bump to v52.8.0](https://github.com/tensorflow/tensorflow/pull/113368)
+ - Updates TensorFlow hash to 25cf6546b44bdb3d3f17277a4b49b027dcdc0ae7 from nightly, Mar 23rd
+
+### Removed
+
+### Fixed
+
 ## [r26.03] 2026-03-16
 https://github.com/ARM-software/Tool-Solutions/tree/r26.03
 

diff --git a/ML-Frameworks/tensorflow-aarch64/get-source.sh b/ML-Frameworks/tensorflow-aarch64/get-source.sh
@@ -8,7 +8,7 @@ source ../utils/git-utils.sh
 
 set -eux -o pipefail
 
-TENSORFLOW_HASH=535fc05dbac76cec4a446eec0bed866394167c05 # from nightly, Feb 25th
+TENSORFLOW_HASH=25cf6546b44bdb3d3f17277a4b49b027dcdc0ae7 # from nightly, Mar 23rd
 
 git-shallow-clone https://github.com/tensorflow/tensorflow.git $TENSORFLOW_HASH
 
@@ -17,8 +17,8 @@ git-shallow-clone https://github.com/tensorflow/tensorflow.git $TENSORFLOW_HASH
 
     # Apply TensorFlow WIP patches here
 
-    # https://github.com/tensorflow/tensorflow/pull/100882 - build(aarch64): Update Compute Library to 52.4.0
-    apply-github-patch tensorflow/tensorflow f15bec785f25dacaf9ae18250e499274e4ec7fb1
+    # https://github.com/tensorflow/tensorflow/pull/113368 - Bump Compute Library version from v24.12 to v52.8.0
+    apply-github-patch tensorflow/tensorflow 189698cf3b3dab5375508ff2e02f6aba3a586323
     # https://github.com/tensorflow/tensorflow/pull/102272 - Fix AArch64 CPUIDInfo init
     apply-github-patch tensorflow/tensorflow b6a9e1c1173d675533ffbb71c5eb36c7060ae2d0
 )