Merge branch 'main' into baris_mletorch-2055_dynamic-pad-guard

Author: bdemirb

.ci/docker/build.sh

@@ -84,6 +84,11 @@ case "${IMAGE_NAME}" in
     CUDA_VERSION=12.8
     SKIP_PYTORCH=yes
     ;;
+  executorch-ubuntu-24.04-gcc14)
+    LINTRUNNER=""
+    OS_VERSION=24.04
+    GCC_VERSION=14
+    ;;
   *)
     echo "Invalid image name ${IMAGE_NAME}"
     exit 1
@@ -92,6 +97,10 @@ esac
 TORCH_VERSION=$(cat ci_commit_pins/pytorch.txt)
 BUILD_DOCS=1
 
+if [[ "${GCC_VERSION:-}" == "11" && -z "${SKIP_PYTORCH:-}" ]]; then
+  PYTORCH_BUILD_MAX_JOBS=6
+fi
+
 # Copy requirements-lintrunner.txt from root to here
 cp ../../requirements-lintrunner.txt ./
 
@@ -104,6 +113,7 @@ docker build \
   --build-arg "PYTHON_VERSION=${PYTHON_VERSION}" \
   --build-arg "MINICONDA_VERSION=${MINICONDA_VERSION}" \
   --build-arg "TORCH_VERSION=${TORCH_VERSION}" \
+  --build-arg "PYTORCH_BUILD_MAX_JOBS=${PYTORCH_BUILD_MAX_JOBS:-}" \
   --build-arg "BUCK2_VERSION=${BUCK2_VERSION}" \
   --build-arg "LINTRUNNER=${LINTRUNNER:-}" \
   --build-arg "BUILD_DOCS=${BUILD_DOCS}" \

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-release/2.11
+release/2.12

.ci/docker/common/install_android.sh

@@ -43,10 +43,10 @@ install_ndk() {
   ARCH=$(uname -m)
   if [ "${ARCH}" = "aarch64" ]; then
     # aarch64 NDK is not cached on S3, download from Google directly
-    curl -Os --retry 3 "https://dl.google.com/android/repository/android-ndk-${ANDROID_NDK_VERSION}-linux.zip"
+    curl -Os --retry 3 --retry-all-errors "https://dl.google.com/android/repository/android-ndk-${ANDROID_NDK_VERSION}-linux.zip"
   else
     # The NDK installation is cached on ossci-android S3 bucket
-    curl -Os --retry 3 "https://ossci-android.s3.amazonaws.com/android-ndk-${ANDROID_NDK_VERSION}-linux.zip"
+    curl -Os --retry 3 --retry-all-errors "https://ossci-android.s3.amazonaws.com/android-ndk-${ANDROID_NDK_VERSION}-linux.zip"
   fi
   unzip -qo "android-ndk-${ANDROID_NDK_VERSION}-linux.zip"
 
@@ -62,7 +62,7 @@ install_cmdtools() {
 
   pushd /tmp
   # The file is cached on ossci-android S3 bucket
-  curl -Os --retry 3 "https://ossci-android.s3.us-west-1.amazonaws.com/${CMDTOOLS_FILENAME}"
+  curl -Os --retry 3 --retry-all-errors "https://ossci-android.s3.us-west-1.amazonaws.com/${CMDTOOLS_FILENAME}"
   unzip -qo "${CMDTOOLS_FILENAME}" -d /opt
 
   ls -lah /opt/cmdline-tools/bin

.ci/docker/common/install_cache.sh

@@ -34,7 +34,7 @@ install_ubuntu() {
 
 install_binary() {
   echo "Downloading sccache binary from S3 repo"
-  curl --retry 3 https://s3.amazonaws.com/ossci-linux/sccache -o /opt/cache/bin/sccache
+  curl --retry 3 --retry-all-errors https://s3.amazonaws.com/ossci-linux/sccache -o /opt/cache/bin/sccache
   chmod +x /opt/cache/bin/sccache
 }
 
@@ -76,6 +76,9 @@ init_sccache() {
   # This is the remote cache bucket
   export SCCACHE_BUCKET=ossci-compiler-cache-circleci-v2
   export SCCACHE_S3_KEY_PREFIX=executorch
+  export SCCACHE_REGION=us-east-1
+  export AWS_REGION=us-east-1
+  export AWS_DEFAULT_REGION=us-east-1
   export SCCACHE_IDLE_TIMEOUT=0
   export SCCACHE_ERROR_LOG=/tmp/sccache_error.log
   export RUST_LOG=sccache::server=error

.ci/docker/common/install_docs_reqs.sh

@@ -12,10 +12,10 @@ if [ -n "$BUILD_DOCS" ]; then
   # Ignore error if gpg-agent doesn't exist (for Ubuntu 16.04)
   apt-get install -y gpg-agent || :
 
-  curl --retry 3 -sL https://deb.nodesource.com/setup_16.x | sudo -E bash -
+  curl --retry 3 --retry-all-errors -sL https://deb.nodesource.com/setup_16.x | sudo -E bash -
   sudo apt-get install -y nodejs
 
-  curl --retry 3 -sS https://dl.yarnpkg.com/debian/pubkey.gpg | sudo apt-key add -
+  curl --retry 3 --retry-all-errors -sS https://dl.yarnpkg.com/debian/pubkey.gpg | sudo apt-key add -
   echo "deb https://dl.yarnpkg.com/debian/ stable main" | sudo tee /etc/apt/sources.list.d/yarn.list
 
   apt-get update

.ci/docker/common/install_linter.sh

@@ -15,5 +15,5 @@ source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
 pip_install -r requirements-lintrunner.txt
 
 # Install google-java-format
-curl -L --retry 3 https://github.com/google/google-java-format/releases/download/v1.23.0/google-java-format_linux-x86-64 > /opt/google-java-format
+curl -L --retry 3 --retry-all-errors https://github.com/google/google-java-format/releases/download/v1.23.0/google-java-format_linux-x86-64 > /opt/google-java-format
 chmod +x /opt/google-java-format

.ci/docker/common/install_pytorch.sh

@@ -27,14 +27,20 @@ install_pytorch_and_domains() {
   chown -R ci-user .
 
   export _GLIBCXX_USE_CXX11_ABI=1
+  if [[ "$(uname -m)" == "aarch64" ]]; then
+    export BUILD_IGNORE_SVE_UNAVAILABLE=1
+  fi
+  if [[ -n "${PYTORCH_BUILD_MAX_JOBS:-}" ]]; then
+    export MAX_JOBS="${PYTORCH_BUILD_MAX_JOBS}"
+  fi
   # Then build and install PyTorch
   conda_run python setup.py bdist_wheel
   pip_install "$(echo dist/*.whl)"
 
   # Grab the pinned audio and vision commits from PyTorch
   TORCHAUDIO_VERSION=release/2.11
   export TORCHAUDIO_VERSION
-  TORCHVISION_VERSION=release/0.26
+  TORCHVISION_VERSION=release/0.27
   export TORCHVISION_VERSION
 
   install_domains

.ci/docker/common/install_user.sh

@@ -7,6 +7,11 @@
 
 set -ex
 
+# On Ubuntu 24.04, there is a `ubuntu` user with id=1000
+if id ubuntu >/dev/null && [[ "$(id -u ubuntu)" == 1000 ]]; then
+    sudo userdel --remove ubuntu;
+fi
+
 # Same as ec2-user
 echo "ci-user:x:1000:1000::/var/lib/ci-user:" >> /etc/passwd
 echo "ci-user:x:1000:" >> /etc/group

.ci/docker/ubuntu/Dockerfile

@@ -62,9 +62,12 @@ RUN bash ./install_cache.sh && rm install_cache.sh utils.sh
 ENV SCCACHE_BUCKET ossci-compiler-cache-circleci-v2
 ENV SCCACHE_S3_KEY_PREFIX executorch
 ENV SCCACHE_REGION us-east-1
+ENV AWS_REGION us-east-1
+ENV AWS_DEFAULT_REGION us-east-1
 
 ARG TORCH_VERSION
 ARG SKIP_PYTORCH
+ARG PYTORCH_BUILD_MAX_JOBS
 COPY ./common/install_pytorch.sh install_pytorch.sh
 COPY ./common/utils.sh utils.sh
 RUN if [ -z "${SKIP_PYTORCH}" ]; then bash ./install_pytorch.sh; fi && rm install_pytorch.sh utils.sh

.ci/scripts/export_model_artifact.sh

@@ -195,9 +195,17 @@ case "$HF_MODEL" in
     PREPROCESSOR_FEATURE_SIZE=""
     PREPROCESSOR_OUTPUT=""
     ;;
+  SocialLocalMobile/gemma-4-31B-it-HQQ-INT4)
+    MODEL_NAME="gemma4_31b"
+    TASK=""
+    MAX_SEQ_LEN=""
+    EXTRA_PIP=""
+    PREPROCESSOR_FEATURE_SIZE=""
+    PREPROCESSOR_OUTPUT=""
+    ;;
   *)
     echo "Error: Unsupported model '$HF_MODEL'"
-    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, openai/whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}, google/gemma-3-4b-it, Qwen/Qwen3-0.6B, nvidia/diar_streaming_sortformer_4spk-v2, nvidia/parakeet-tdt, facebook/dinov2-small-imagenet1k-1-layer, SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4"
+    echo "Supported models: mistralai/Voxtral-Mini-3B-2507, mistralai/Voxtral-Mini-4B-Realtime-2602, openai/whisper-{small, medium, large, large-v2, large-v3, large-v3-turbo}, google/gemma-3-4b-it, Qwen/Qwen3-0.6B, nvidia/diar_streaming_sortformer_4spk-v2, nvidia/parakeet-tdt, facebook/dinov2-small-imagenet1k-1-layer, SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4, SocialLocalMobile/gemma-4-31B-it-HQQ-INT4"
     exit 1
     ;;
 esac
@@ -459,6 +467,50 @@ if [ "$MODEL_NAME" = "qwen3_5_moe" ]; then
   exit 0
 fi
 
+# Gemma 4 31B uses a prequantized checkpoint and custom export script
+if [ "$MODEL_NAME" = "gemma4_31b" ]; then
+  pip install safetensors huggingface_hub gguf
+
+  # Download prequantized model outside OUTPUT_DIR to avoid uploading on failure
+  LOCAL_MODEL_DIR=$(mktemp -d)
+  INDUCTOR_CACHE=$(mktemp -d)
+  trap 'rm -rf "$LOCAL_MODEL_DIR" "$INDUCTOR_CACHE"' EXIT
+
+  python -c "from huggingface_hub import snapshot_download; snapshot_download('${HF_MODEL}', local_dir='${LOCAL_MODEL_DIR}')"
+
+  # Sanity check: run inference on the prequantized model
+  echo "::group::Inference sanity check"
+  INFERENCE_OUTPUT=$(python -m executorch.examples.models.gemma4_31b.inference \
+      --prequantized "$LOCAL_MODEL_DIR" \
+      --prompt "What is the capital of France?" \
+      --max-new-tokens 32 \
+      --temperature 0 \
+      --no-compile 2>&1)
+  echo "$INFERENCE_OUTPUT"
+  if ! echo "$INFERENCE_OUTPUT" | grep -q "Paris"; then
+    echo "ERROR: Inference sanity check failed — expected 'Paris' in output"
+    exit 1
+  fi
+  echo "::endgroup::"
+
+  # Copy tokenizer for the runner
+  cp "$LOCAL_MODEL_DIR/tokenizer.json" "${OUTPUT_DIR}/tokenizer.json"
+
+  # Export to .pte/.ptd (short cache dir avoids objcopy symbol length issues)
+  echo "::group::Export"
+  TORCHINDUCTOR_CACHE_DIR="$INDUCTOR_CACHE" \
+  python -m executorch.examples.models.gemma4_31b.export \
+      --prequantized "$LOCAL_MODEL_DIR" \
+      --output-dir "${OUTPUT_DIR}"
+  echo "::endgroup::"
+
+  test -f "${OUTPUT_DIR}/model.pte"
+  test -f "${OUTPUT_DIR}/aoti_cuda_blob.ptd"
+  ls -al "${OUTPUT_DIR}"
+
+  exit 0
+fi
+
 MAX_SEQ_LEN_ARG=""
 if [ -n "$MAX_SEQ_LEN" ]; then
   MAX_SEQ_LEN_ARG="--max_seq_len $MAX_SEQ_LEN"