Enable ARM (aarch64) support for cuvs-java and cuvs-lucene

.github/workflows/build.yaml

@@ -190,7 +190,8 @@ jobs:
       file_to_upload: "java/cuvs-java/target/"
       sha: ${{ inputs.sha }}
   lucene-build:
-    # Depends on the Java job for the cuvs-java artifact that cuvs-lucene builds against.
+    # Depends on the Java job for the cuvs-java artifact that cuvs-lucene builds against, and
+    # reuses its matrix so the two always agree on the set of CUDA versions.
     needs: [java-build-matrix, java-build]
     permissions:
       actions: read

.github/workflows/pr.yaml

@@ -26,7 +26,10 @@ jobs:
       - rocky8-clib-tests
       - conda-java-build-and-tests-matrix
       - conda-java-build-and-tests
+      - conda-java-tests-other-arch-matrix
+      - conda-java-tests-other-arch
       - conda-lucene-build-and-tests
+      - conda-lucene-tests-other-arch
       - rust-build-matrix
       - rust-build
       - go-build-matrix
@@ -613,6 +616,39 @@ jobs:
       script: "ci/test_java.sh"
       artifact-name: "cuvs-java-cuda${{ matrix.CUDA_VER }}"
       file_to_upload: "java/cuvs-java/target/"
+  conda-java-tests-other-arch-matrix:
+    # All architectures the C++ core builds for, except amd64.
+    needs: [conda-cpp-build, changed-files]
+    permissions:
+      contents: read
+    uses: rapidsai/shared-workflows/.github/workflows/compute-matrix.yaml@main
+    if: fromJSON(needs.changed-files.outputs.changed_file_groups).test_java || fromJSON(needs.changed-files.outputs.changed_file_groups).test_cpp
+    with:
+      build_type: pull-request
+      matrix_name: conda-cpp-build
+      matrix_filter: map(select(.ARCH != "amd64"))
+  conda-java-tests-other-arch:
+    # Verifies that the amd64-built cuvs-java jar (jextract-generated Panama bindings and
+    # all) also works correctly on every non-amd64 architecture, by running its own IT test
+    # suite against the amd64-built classes without recompiling -- see ci/test_java_prebuilt.sh.
+    needs: [conda-java-tests-other-arch-matrix, conda-java-build-and-tests]
+    permissions:
+      actions: read
+      contents: read
+      id-token: write
+      packages: read
+      pull-requests: read
+    secrets: inherit # zizmor: ignore[secrets-inherit]
+    uses: rapidsai/shared-workflows/.github/workflows/custom-job.yaml@main
+    strategy:
+      fail-fast: false
+      matrix: ${{ fromJSON(needs.conda-java-tests-other-arch-matrix.outputs.matrix) }}
+    with:
+      build_type: pull-request
+      node_type: "gpu-l4-latest-1"
+      arch: "${{ matrix.ARCH }}"
+      container_image: "rapidsai/ci-conda:26.10-cuda${{ matrix.CUDA_VER }}-${{ matrix.LINUX_VER }}-py${{ matrix.PY_VER }}"
+      script: "ci/test_java_prebuilt.sh ${{ matrix.ARCH }} cuvs-java-cuda${{ matrix.CUDA_VER }}"
   conda-lucene-build-and-tests:
     # Depends on the Java job for the cuvs-java artifact that cuvs-lucene builds against, and
     # reuses its matrix so the two always agree on the set of CUDA versions.
@@ -638,6 +674,29 @@ jobs:
       script: "ci/test_lucene.sh cuvs-java-cuda${{ matrix.CUDA_VER }}"
       artifact-name: "cuvs-lucene-cuda${{ matrix.CUDA_VER }}"
       file_to_upload: "java/cuvs-lucene/target/"
+  conda-lucene-tests-other-arch:
+    # Verifies that cuvs-lucene, built against the amd64-built cuvs-java jar, also works
+    # correctly on every non-amd64 architecture, by running its test suite against the
+    # amd64-built classes without recompiling -- see ci/test_lucene_prebuilt.sh. Reuses
+    # conda-java-tests-other-arch-matrix so the two always agree on which architectures to test.
+    needs: [conda-java-tests-other-arch-matrix, conda-java-build-and-tests, conda-lucene-build-and-tests]
+    permissions:
+      actions: read
+      contents: read
+      id-token: write
+      packages: read
+      pull-requests: read
+    secrets: inherit # zizmor: ignore[secrets-inherit]
+    uses: rapidsai/shared-workflows/.github/workflows/custom-job.yaml@main
+    strategy:
+      fail-fast: false
+      matrix: ${{ fromJSON(needs.conda-java-tests-other-arch-matrix.outputs.matrix) }}
+    with:
+      build_type: pull-request
+      node_type: "gpu-l4-latest-1"
+      arch: "${{ matrix.ARCH }}"
+      container_image: "rapidsai/ci-conda:26.10-cuda${{ matrix.CUDA_VER }}-${{ matrix.LINUX_VER }}-py${{ matrix.PY_VER }}"
+      script: "ci/test_lucene_prebuilt.sh ${{ matrix.ARCH }} cuvs-java-cuda${{ matrix.CUDA_VER }} cuvs-lucene-cuda${{ matrix.CUDA_VER }}"
   rust-build-matrix:
     needs: [conda-cpp-build, changed-files]
     permissions:

.github/workflows/test.yaml

@@ -104,6 +104,41 @@ jobs:
       script: "ci/test_java.sh"
       artifact-name: "cuvs-java-cuda${{ matrix.CUDA_VER }}"
       file_to_upload: "java/cuvs-java/target/"
+  conda-java-tests-other-arch-matrix:
+    # All architectures the C++ core builds for, except amd64 -- so a newly added
+    # architecture is automatically picked up here without a workflow change.
+    permissions:
+      contents: read
+    uses: rapidsai/shared-workflows/.github/workflows/compute-matrix.yaml@main
+    with:
+      build_type: ${{ inputs.build_type }}
+      matrix_name: conda-cpp-build
+      matrix_filter: map(select(.ARCH != "amd64"))
+  conda-java-tests-other-arch:
+    # Verifies that the amd64-built cuvs-java jar (jextract-generated Panama bindings and
+    # all) also works correctly on every non-amd64 architecture, by running its own IT test
+    # suite against the amd64-built classes without recompiling -- see ci/test_java_prebuilt.sh.
+    needs: [conda-java-tests-other-arch-matrix, conda-java-build-and-tests]
+    permissions:
+      actions: read
+      contents: read
+      id-token: write
+      packages: read
+      pull-requests: read
+    secrets: inherit # zizmor: ignore[secrets-inherit]
+    uses: rapidsai/shared-workflows/.github/workflows/custom-job.yaml@main
+    strategy:
+      fail-fast: false
+      matrix: ${{ fromJSON(needs.conda-java-tests-other-arch-matrix.outputs.matrix) }}
+    with:
+      build_type: ${{ inputs.build_type }}
+      branch: ${{ inputs.branch }}
+      date: ${{ inputs.date }}
+      sha: ${{ inputs.sha }}
+      node_type: "gpu-l4-latest-1"
+      arch: "${{ matrix.ARCH }}"
+      container_image: "rapidsai/ci-conda:26.10-cuda${{ matrix.CUDA_VER }}-${{ matrix.LINUX_VER }}-py${{ matrix.PY_VER }}"
+      script: "ci/test_java_prebuilt.sh ${{ matrix.ARCH }} cuvs-java-cuda${{ matrix.CUDA_VER }}"
   conda-lucene-build-and-tests:
     # Depends on the Java job for the cuvs-java artifact that cuvs-lucene builds against, and
     # reuses its matrix so the two always agree on the set of CUDA versions.
@@ -132,6 +167,32 @@ jobs:
       script: "ci/test_lucene.sh cuvs-java-cuda${{ matrix.CUDA_VER }}"
       artifact-name: "cuvs-lucene-cuda${{ matrix.CUDA_VER }}"
       file_to_upload: "java/cuvs-lucene/target/"
+  conda-lucene-tests-other-arch:
+    # Verifies that cuvs-lucene, built against the amd64-built cuvs-java jar, also works
+    # correctly on every non-amd64 architecture, by running its test suite against the
+    # amd64-built classes without recompiling -- see ci/test_lucene_prebuilt.sh. Reuses
+    # conda-java-tests-other-arch-matrix so the two always agree on which architectures to test.
+    needs: [conda-java-tests-other-arch-matrix, conda-java-build-and-tests, conda-lucene-build-and-tests]
+    permissions:
+      actions: read
+      contents: read
+      id-token: write
+      packages: read
+      pull-requests: read
+    secrets: inherit # zizmor: ignore[secrets-inherit]
+    uses: rapidsai/shared-workflows/.github/workflows/custom-job.yaml@main
+    strategy:
+      fail-fast: false
+      matrix: ${{ fromJSON(needs.conda-java-tests-other-arch-matrix.outputs.matrix) }}
+    with:
+      build_type: ${{ inputs.build_type }}
+      branch: ${{ inputs.branch }}
+      date: ${{ inputs.date }}
+      sha: ${{ inputs.sha }}
+      node_type: "gpu-l4-latest-1"
+      arch: "${{ matrix.ARCH }}"
+      container_image: "rapidsai/ci-conda:26.10-cuda${{ matrix.CUDA_VER }}-${{ matrix.LINUX_VER }}-py${{ matrix.PY_VER }}"
+      script: "ci/test_lucene_prebuilt.sh ${{ matrix.ARCH }} cuvs-java-cuda${{ matrix.CUDA_VER }} cuvs-lucene-cuda${{ matrix.CUDA_VER }}"
   wheel-tests-cuvs:
     permissions:
       actions: read

ci/test_java_prebuilt.sh

@@ -0,0 +1,83 @@
+#!/bin/bash
+# SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+set -euo pipefail
+
+# Runs cuvs-java's own IT test suite against the classes from a prior amd64 build job,
+# without recompiling (or running jextract) on this host. cuvs-java is always built on
+# amd64; this verifies that the resulting jextract-generated Panama bindings also work
+# correctly against a native libcuvs_c.so on the given (non-amd64) architecture.
+#
+# Takes the target architecture (as used by the GitHub Actions job matrix, e.g. "arm64")
+# and the name of the cuvs-java artifact uploaded by the amd64 Java job.
+
+ARCH="${1:?Usage: $0 <arch> <cuvs-java-artifact-name>}"
+CUVS_JAVA_ARTIFACT="${2:?Usage: $0 <arch> <cuvs-java-artifact-name>}"
+
+rapids-logger "Testing the amd64-built cuvs-java artifact on ${ARCH}"
+
+case "${ARCH}" in
+  amd64) CONDA_ARCH="x86_64" ;;
+  arm64) CONDA_ARCH="aarch64" ;;
+  *) CONDA_ARCH="${ARCH}" ;;
+esac
+
+if [ -e "/opt/conda/etc/profile.d/conda.sh" ]; then
+  . /opt/conda/etc/profile.d/conda.sh
+fi
+
+rapids-logger "Check GPU usage"
+nvidia-smi
+
+rapids-logger "Configuring conda strict channel priority"
+conda config --set channel_priority strict
+
+rapids-logger "Downloading artifacts from previous jobs"
+CPP_CHANNEL=$(rapids-download-from-github "$(rapids-artifact-name conda_cpp libcuvs cuvs --cuda "$RAPIDS_CUDA_VERSION")")
+CUVS_JAVA_DIR=$(rapids-download-from-github "${CUVS_JAVA_ARTIFACT}")
+
+rapids-logger "Generate Java testing dependencies"
+
+ENV_YAML_DIR="$(mktemp -d)"
+
+rapids-dependency-file-generator \
+  --output conda \
+  --file-key java \
+  --prepend-channel "${CPP_CHANNEL}" \
+  --matrix "cuda=${RAPIDS_CUDA_VERSION%.*};arch=${CONDA_ARCH}" | tee "${ENV_YAML_DIR}/env.yaml"
+
+rapids-mamba-retry env create --yes -f "${ENV_YAML_DIR}/env.yaml" -n java
+
+# Temporarily allow unbound variables for conda activation.
+set +u
+conda activate java
+set -u
+
+rapids-print-env
+
+# libcuvs comes from the conda environment here, matching the architecture this script is
+# running on.
+export LD_LIBRARY_PATH="${CONDA_PREFIX}/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
+
+rapids-logger "Restore the amd64-built target/ directory so Maven can run the tests without recompiling"
+
+rm -rf java/cuvs-java/target
+mkdir -p java/cuvs-java/target
+cp -a "${CUVS_JAVA_DIR}/." java/cuvs-java/target/
+
+EXITCODE=0
+trap "EXITCODE=1" ERR
+set +e
+
+rapids-logger "Run cuvs-java IT tests against the amd64-built classes"
+
+# -Dskip.compile activates the pom's "skip-compile" profile, which disables all
+# compilation for this run, forcing test to use amd64-compiled jar instead of
+# local code.
+pushd java/cuvs-java
+mvn --batch-mode verify -Dskip.compile=true
+popd
+
+rapids-logger "Test script exiting with value: $EXITCODE"
+exit ${EXITCODE}

ci/test_lucene_prebuilt.sh

@@ -0,0 +1,112 @@
+#!/bin/bash
+# SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+set -euo pipefail
+
+# Runs cuvs-lucene's test suite against the classes from a prior amd64 build job, without
+# recompiling on this host. cuvs-lucene is always built on amd64, against the amd64-built
+# cuvs-java jar; this verifies that pairing also works correctly on the given (non-amd64)
+# architecture. Since cuvs-lucene depends on the plain (no bundled natives) cuvs-java jar,
+# this doubles as a cross-arch check that the jextract-generated Panama bindings baked
+# into the amd64 jar work unmodified against a native libcuvs_c.so on that architecture.
+#
+# Takes the target architecture (as used by the GitHub Actions job matrix, e.g. "arm64")
+# and the names of the cuvs-java and cuvs-lucene artifacts uploaded by the amd64 jobs.
+
+ARCH="${1:?Usage: $0 <arch> <cuvs-java-artifact-name> <cuvs-lucene-artifact-name>}"
+CUVS_JAVA_ARTIFACT="${2:?Usage: $0 <arch> <cuvs-java-artifact-name> <cuvs-lucene-artifact-name>}"
+CUVS_LUCENE_ARTIFACT="${3:?Usage: $0 <arch> <cuvs-java-artifact-name> <cuvs-lucene-artifact-name>}"
+
+rapids-logger "Testing the amd64-built cuvs-java/cuvs-lucene artifacts on ${ARCH}"
+
+case "${ARCH}" in
+  amd64) CONDA_ARCH="x86_64" ;;
+  arm64) CONDA_ARCH="aarch64" ;;
+  *) CONDA_ARCH="${ARCH}" ;;
+esac
+
+if [ -e "/opt/conda/etc/profile.d/conda.sh" ]; then
+  . /opt/conda/etc/profile.d/conda.sh
+fi
+
+rapids-logger "Check GPU usage"
+nvidia-smi
+
+rapids-logger "Configuring conda strict channel priority"
+conda config --set channel_priority strict
+
+rapids-logger "Downloading artifacts from previous jobs"
+CPP_CHANNEL=$(rapids-download-from-github "$(rapids-artifact-name conda_cpp libcuvs cuvs --cuda "$RAPIDS_CUDA_VERSION")")
+CUVS_JAVA_DIR=$(rapids-download-from-github "${CUVS_JAVA_ARTIFACT}")
+CUVS_LUCENE_DIR=$(rapids-download-from-github "${CUVS_LUCENE_ARTIFACT}")
+
+rapids-logger "Generate Java testing dependencies"
+
+ENV_YAML_DIR="$(mktemp -d)"
+
+rapids-dependency-file-generator \
+  --output conda \
+  --file-key java \
+  --prepend-channel "${CPP_CHANNEL}" \
+  --matrix "cuda=${RAPIDS_CUDA_VERSION%.*};arch=${CONDA_ARCH}" | tee "${ENV_YAML_DIR}/env.yaml"
+
+rapids-mamba-retry env create --yes -f "${ENV_YAML_DIR}/env.yaml" -n java
+
+# Temporarily allow unbound variables for conda activation.
+set +u
+conda activate java
+set -u
+
+rapids-print-env
+
+# libcuvs comes from the conda environment here, matching the architecture this script is
+# running on. cuvs-lucene depends on the plain cuvs-java jar, which bundles no native
+# libraries, so the JVM resolves libcuvs_c.so through the dynamic loader.
+export LD_LIBRARY_PATH="${CONDA_PREFIX}/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
+
+rapids-logger "Install the amd64-built cuvs-java artifact into the local Maven repository"
+
+# cuvs-lucene resolves cuvs-java from the local Maven repository. Its pom.xml travels with
+# the artifact and supplies the coordinates, so no version needs to be hardcoded.
+CUVS_JAVA_POM="${CUVS_JAVA_DIR}/pom.xml"
+if [ ! -f "${CUVS_JAVA_POM}" ]; then
+  echo "Could not find pom.xml in the cuvs-java artifact at ${CUVS_JAVA_DIR}" >&2
+  exit 1
+fi
+
+# The artifact also carries the per-architecture native jar and the sources/javadoc/test jars;
+# cuvs-lucene depends on the plain one.
+mapfile -t CUVS_JAVA_JARS < <(find "${CUVS_JAVA_DIR}" -maxdepth 1 -name 'cuvs-java-*.jar' \
+  ! -name '*-sources.jar' ! -name '*-javadoc.jar' ! -name '*-tests.jar' ! -name '*-cuda*.jar')
+if [ "${#CUVS_JAVA_JARS[@]}" -ne 1 ]; then
+  echo "Expected exactly one cuvs-java jar in ${CUVS_JAVA_DIR}, found: ${CUVS_JAVA_JARS[*]:-none}" >&2
+  exit 1
+fi
+
+# Install cuvs-java jar into .m2, cd is needed to pick up pom.xml in order to avoid rate limit of main maven repo
+pushd java/cuvs-lucene
+mvn --batch-mode install:install-file -Dfile="${CUVS_JAVA_JARS[0]}" -DpomFile="${CUVS_JAVA_POM}"
+popd
+
+rapids-logger "Restore the amd64-built cuvs-lucene target/ directory so Maven can run the tests without recompiling"
+
+rm -rf java/cuvs-lucene/target
+mkdir -p java/cuvs-lucene/target
+cp -a "${CUVS_LUCENE_DIR}/." java/cuvs-lucene/target/
+
+EXITCODE=0
+trap "EXITCODE=1" ERR
+set +e
+
+rapids-logger "Run cuvs-lucene tests against the amd64-built classes"
+
+# -Dskip.compile activates the pom's "skip-compile" profile, which disables all
+# compilation for this run, forcing test to use amd64-compiled jar instead of
+# local code.
+pushd java/cuvs-lucene
+mvn --batch-mode test -Dskip.compile=true
+popd
+
+rapids-logger "Test script exiting with value: $EXITCODE"
+exit ${EXITCODE}