Update base for Update on "[llm_patch] Fix out-of-bounds access in pad2d function"

#15623

Add checks for pad1d and pad3d, as well as pad2d.

I think what happens is, we 'broadcast' the input tensor into the padded regions, depending on the pad algorithm (replication, reflection).

The pad algorithm takes an output tensor index, and returns an input tensor index. We need to check that this input tensor index is valid/within bounds.

---
The crash is a "wild-addr-read" that occurs in the `pad2d` function, which is part of the Executorch library. This type of crash typically indicates that the program is attempting to read from an invalid or uninitialized memory address.

The root cause of the crash is an out-of-bounds access in the `pad2d` function. The function uses a `padding_ix` function to calculate indices for the input tensor `in`, but it does not perform sufficient bounds checking to ensure that these indices are valid. As a result, the program may attempt to read from memory outside the bounds of the `in` tensor, leading to the crash.

The patch fixes the crash by adding bounds checking to the `pad2d` function. Specifically, it adds two `ET_CHECK` statements to verify that the indices calculated by `padding_ix` are within the valid range of the `in` tensor. The checks are performed using the following code: `ET_CHECK(in_h_idx < in_height)` and `ET_CHECK(in_w_idx < in_width)`. By adding these checks, the patch ensures that the program will not attempt to read from invalid memory addresses, preventing the "wild-addr-read" crash.

Other considerations that reviewers should take into account when validating the patch include the potential impact on performance. The added `ET_CHECK` statements may introduce a small performance overhead, particularly if the `pad2d` function is called frequently. Reviewers should verify that the performance impact is acceptable and that the patch does not introduce any other unintended consequences. Additionally, reviewers should test the patch with a variety of input tensors and padding configurations to ensure that it correctly handles different edge cases. They should also verify that the `ET_CHECK` statements are triggered correctly when invalid indices are encountered, and that the program behaves as expected in these cases.

NOTE: This diff is entirely auto-generated by LLM-based patch generator.
Reviewer should carefully examine this diff as Lionhead does not guarrantee the
correctnesss of the patch beyond fixing the crash and passing existing tests.
Please commandeer this diff and revise as needed. Our bot does not respond to
comments or revision requests (yet).

Differential Revision: [D80831697](https://our.internmc.facebook.com/intern/diff/D80831697/)

[ghstack-poisoned]

Author: lucylq

.ci/docker/common/install_arm.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+# Copyright 2025 Arm Limited and/or its affiliates.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+set -ex
+
+install_arm_prerequiresites() {
+    apt-get update -y
+    apt-get install -y --no-install-recommends \
+            mesa-vulkan-drivers libvulkan1
+    rm -rf /var/lib/apt/lists/*
+}
+
+install_arm_prerequiresites

.ci/docker/ubuntu/Dockerfile

@@ -83,6 +83,9 @@ RUN if [ -n "${ANDROID_NDK_VERSION}" ]; then bash ./install_android.sh; fi
 RUN rm install_android.sh
 
 ARG ARM_SDK
+COPY ./common/install_arm.sh install_arm.sh
+RUN if [ -n "${ARM_SDK}" ]; then bash ./install_arm.sh; fi
+RUN rm install_arm.sh
 
 ARG ZEPHYR_SDK
 COPY ./common/install_zephyr.sh install_zephyr.sh

.ci/scripts/test_llama.sh

@@ -171,15 +171,14 @@ cmake_build_llama_runner() {
     git submodule update --init
     popd
     dir="examples/models/llama"
-    retry cmake \
-        -DEXECUTORCH_BUILD_TESTS=ON \
-        -DBUILD_TESTING=OFF \
-        -DCMAKE_INSTALL_PREFIX=cmake-out \
-        -DCMAKE_BUILD_TYPE="$CMAKE_BUILD_TYPE" \
-        -Bcmake-out/${dir} \
-        ${dir}
-    cmake --build cmake-out/${dir} -j9 --config "$CMAKE_BUILD_TYPE"
-
+    if [[ "$CMAKE_BUILD_TYPE" == "Debug" ]]; then
+        PRESET="llama-debug"
+    else
+        PRESET="llama-release"
+    fi
+    pushd "${dir}"
+    cmake --workflow --preset "${PRESET}"
+    popd
 }
 
 cleanup_files() {

.ci/scripts/test_llama_lora.sh

@@ -12,10 +12,7 @@ source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 cmake_install_executorch_libraries() {
     echo "Installing libexecutorch.a, libextension_module.so, libportable_ops_lib.a"
     rm -rf cmake-out
-    retry cmake --preset llm \
-        -DCMAKE_INSTALL_PREFIX=cmake-out \
-        -DCMAKE_BUILD_TYPE=Release
-    cmake --build cmake-out -j9 --target install --config Release
+    cmake --workflow llm-release
 }
 
 cmake_build_llama_runner() {

.ci/scripts/test_model_e2e.sh

@@ -156,27 +156,13 @@ echo "::endgroup::"
 
 echo "::group::Build $MODEL_NAME Runner"
 
-if [ "$DEVICE" = "cuda" ]; then
-  BUILD_BACKEND="EXECUTORCH_BUILD_CUDA"
-elif [ "$DEVICE" = "metal" ]; then
-  BUILD_BACKEND="EXECUTORCH_BUILD_METAL"
-else
+if [ "$DEVICE" != "cuda" ] && [ "$DEVICE" != "metal" ]; then
   echo "Error: Unsupported device '$DEVICE'. Must be 'cuda' or 'metal'."
   exit 1
 fi
 
-cmake --preset llm \
-      -D${BUILD_BACKEND}=ON \
-      -DCMAKE_INSTALL_PREFIX=cmake-out \
-      -DCMAKE_BUILD_TYPE=Release \
-      -Bcmake-out -S.
-cmake --build cmake-out -j$(nproc) --target install --config Release
-
-cmake -D${BUILD_BACKEND}=ON \
-      -DCMAKE_BUILD_TYPE=Release \
-      -Sexamples/models/$RUNNER_PATH \
-      -Bcmake-out/examples/models/$RUNNER_PATH/
-cmake --build cmake-out/examples/models/$RUNNER_PATH --target $RUNNER_TARGET --config Release
+MAKE_TARGET="${RUNNER_PATH}-${DEVICE}"
+make "${MAKE_TARGET}"
 echo "::endgroup::"
 
 echo "::group::Run $MODEL_NAME Runner"

.ci/scripts/test_phi_3_mini.sh

@@ -23,8 +23,16 @@ if hash nproc &> /dev/null; then NPROC=$(nproc); fi
 
 cmake_install_executorch_libraries() {
   rm -rf cmake-out
-  cmake --preset llm -DCMAKE_INSTALL_PREFIX=cmake-out -DCMAKE_BUILD_TYPE=${BUILD_TYPE}
-  cmake --build cmake-out -j16 --target install --config ${BUILD_TYPE}
+
+  # Select workflow preset based on BUILD_TYPE
+  if [[ "${BUILD_TYPE}" == "Debug" ]]; then
+    WORKFLOW_PRESET="llm-debug"
+  else
+    WORKFLOW_PRESET="llm-release"
+  fi
+
+  echo "Using workflow preset: ${WORKFLOW_PRESET}"
+  cmake --workflow --preset ${WORKFLOW_PRESET}
 }
 
 cmake_build_phi_3_mini() {

.ci/scripts/utils.sh

@@ -84,8 +84,8 @@ dedupe_macos_loader_path_rpaths() {
 
 install_domains() {
   echo "Install torchvision and torchaudio"
-  pip install --no-use-pep517 --user "git+https://github.com/pytorch/audio.git@${TORCHAUDIO_VERSION}"
-  pip install --no-use-pep517 --user "git+https://github.com/pytorch/vision.git@${TORCHVISION_VERSION}"
+  pip install --no-build-isolation --user "git+https://github.com/pytorch/audio.git@${TORCHAUDIO_VERSION}"
+  pip install --no-build-isolation --user "git+https://github.com/pytorch/vision.git@${TORCHVISION_VERSION}"
 }
 
 install_pytorch_and_domains() {

.github/workflows/pull.yml

@@ -340,6 +340,7 @@ jobs:
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       timeout: 90
       script: |
+        set -eux
         # The generic Linux job chooses to use base env, not the one setup by the image
         CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
         conda activate "${CONDA_ENV}"

.github/workflows/trunk.yml

@@ -347,7 +347,7 @@ jobs:
         elif [[ ${{ matrix.os}} == "zephyr-preset" ]]; then
           setup_script_args="--target-toolchain zephyr"
           toolchain_prefix=arm-zephyr-eabi-
-          threshold="135656" # 132 KiB
+          threshold="135768" # 136 KiB
           toolchain_cmake=examples/zephyr/x86_64-linux-arm-zephyr-eabi-gcc.cmake
         else
           echo "Fail unsupport OS selection ${{ matrix.os }}"