Build vLLM aarch64 nightly wheels (#162664)

PyTorch has published its aarch64 nightly wheels for all CUDA version after https://github.com/pytorch/pytorch/pull/162364
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162664
Approved by: https://github.com/atalman

.github/actionlint.yaml

@@ -21,6 +21,7 @@ self-hosted-runner:
     - linux.arm64.2xlarge.ephemeral
     - linux.arm64.m7g.4xlarge
     - linux.arm64.m7g.4xlarge.ephemeral
+    - linux.arm64.r7g.12xlarge.memory
     - linux.4xlarge.nvidia.gpu
     - linux.8xlarge.nvidia.gpu
     - linux.16xlarge.nvidia.gpu

.github/ci_configs/vllm/Dockerfile.tmp_vllm

@@ -82,16 +82,10 @@ RUN if command -v apt-get >/dev/null; then \
         apt-get update -y \
         && apt-get install -y ccache software-properties-common git curl wget sudo vim; \
     else \
-        dnf install -y git curl wget sudo vim; \
+        dnf install -y git curl wget sudo; \
     fi \
     && python3 --version && python3 -m pip --version
 
-# Workaround for https://github.com/openai/triton/issues/2507 and
-# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
-# this won't be needed for future versions of this docker image
-# or future versions of triton.
-RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
-
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
     if ! python3 -m uv --version >/dev/null 2>&1; then \
@@ -220,11 +214,16 @@ ARG SCCACHE_S3_NO_CREDENTIALS=0
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \
     if [ "$USE_SCCACHE" = "1" ]; then \
-        echo "Installing sccache..." \
-        && curl -L -o sccache.tar.gz https://github.com/mozilla/sccache/releases/download/v0.8.1/sccache-v0.8.1-x86_64-unknown-linux-musl.tar.gz \
+        echo "Installing sccache..."; \
+        if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
+            SCCACHE_ARCHIVE="sccache-v0.8.1-aarch64-unknown-linux-musl"; \
+        else \
+            SCCACHE_ARCHIVE="sccache-v0.8.1-x86_64-unknown-linux-musl"; \
+        fi; \
+        curl -L -o sccache.tar.gz "https://github.com/mozilla/sccache/releases/download/v0.8.1/${SCCACHE_ARCHIVE}.tar.gz" \
         && tar -xzf sccache.tar.gz \
-        && sudo mv sccache-v0.8.1-x86_64-unknown-linux-musl/sccache /usr/bin/sccache \
-        && rm -rf sccache.tar.gz sccache-v0.8.1-x86_64-unknown-linux-musl \
+        && sudo mv "${SCCACHE_ARCHIVE}"/sccache /usr/bin/sccache \
+        && rm -rf sccache.tar.gz "${SCCACHE_ARCHIVE}" \
         && export SCCACHE_BUCKET=${SCCACHE_BUCKET_NAME} \
         && export SCCACHE_REGION=${SCCACHE_REGION_NAME} \
         && export SCCACHE_S3_NO_CREDENTIALS=${SCCACHE_S3_NO_CREDENTIALS} \
@@ -285,7 +284,7 @@ RUN if command -v apt-get >/dev/null; then \
         && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
         && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION}; \
     else \
-        dnf install -y git curl wget sudo vim; \
+        dnf install -y git curl wget sudo; \
     fi \
     && python3 --version && python3 -m pip --version
 
@@ -298,12 +297,6 @@ RUN echo "[INFO] Listing current directory before torch install step:" && \
     echo "[INFO] Showing torch_build_versions.txt content:" && \
     cat torch_build_versions.txt
 
-# Workaround for https://github.com/openai/triton/issues/2507 and
-# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
-# this won't be needed for future versions of this docker image
-# or future versions of triton.
-RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
-
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
     if ! python3 -m uv --version > /dev/null 2>&1; then \

.github/scripts/prepare_vllm_wheels.sh

@@ -84,6 +84,9 @@ repackage_wheel() {
   rm -rf $package
 }
 
+# Require to re-package the wheel
+${PYTHON_EXECUTABLE} -mpip install wheel==0.45.1
+
 pushd externals/vllm/wheels
 for package in xformers flashinfer-python vllm; do
   repackage_wheel $package

.github/workflows/build-vllm-wheel.yml

@@ -12,6 +12,9 @@ on:
     paths:
       - .github/workflows/build-vllm-wheel.yml
       - .github/ci_commit_pins/vllm.txt
+  schedule:
+    # every morning at 01:30PM UTC, 9:30AM EST, 6:30AM PST
+    - cron: 30 13 * * *
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
@@ -24,21 +27,33 @@ jobs:
       fail-fast: false
       matrix:
         python-version: [ '3.12' ]
-        # TODO (huydhn): Add cu130 https://github.com/pytorch/pytorch/pull/162000#issuecomment-3261541554
+        # TODO (huydhn): Add cu130 after https://github.com/vllm-project/vllm/issues/24464 is resolved
+        platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
         device: [ 'cu128', 'cu129' ]
-        runner: [ 'linux.12xlarge.memory' ]
         include:
-          - device: cu128
+          - platform: manylinux_2_28_x86_64
+            device: cu128
             manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.8'
-          - device: cu129
+            runner: linux.12xlarge.memory
+          - platform: manylinux_2_28_x86_64
+            device: cu129
             manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.9'
-    name: "Build ${{ matrix.device }} vLLM wheel"
+            runner: linux.12xlarge.memory
+          - platform: manylinux_2_28_aarch64
+            device: cu128
+            manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.8'
+            runner: linux.arm64.r7g.12xlarge.memory
+          - platform: manylinux_2_28_aarch64
+            device: cu129
+            manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.9'
+            runner: linux.arm64.r7g.12xlarge.memory
+    name: "Build ${{ matrix.device }} vLLM wheel on ${{ matrix.platform }}"
     runs-on: ${{ matrix.runner }}
     timeout-minutes: 480
     env:
       PY_VERS: ${{ matrix.python-version }}
       MANYLINUX_IMAGE: ${{ matrix.manylinux-image }}
-      PLATFORM: 'manylinux_2_28_x86_64'
+      PLATFORM: ${{ matrix.platform }}
       BUILD_DEVICE: ${{ matrix.device }}
     steps:
       - name: Setup SSH (Click me for login details)
@@ -136,7 +151,7 @@ jobs:
 
       - uses: actions/upload-artifact@50769540e7f4bd5e21e526ee35c689e35e0d6874 # v4.4.0
         with:
-          name: vllm-wheel-${{ matrix.device }}-${{ matrix.python-version }}-${{ env.PLATFORM }}
+          name: vllm-wheel-${{ matrix.device }}-${{ matrix.platform }}-${{ matrix.python-version }}
           if-no-files-found: error
           path: ${{ runner.temp }}/artifacts/externals/vllm/wheels/*.whl
 
@@ -146,15 +161,17 @@ jobs:
 
   # Copied from build-triton-wheel workflow (mostly)
   upload-wheel:
-    name: "Upload ${{ matrix.device }} vLLM wheel"
+    name: "Upload ${{ matrix.device }} vLLM wheel on ${{ matrix.platform }}"
     needs:
       - build-wheel
     runs-on: ubuntu-latest
     strategy:
       fail-fast: false
       matrix:
+        platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
         device: [ 'cu128', 'cu129' ]
     env:
+      PLATFORM: ${{ matrix.platform }}
       BUILD_DEVICE: ${{ matrix.device }}
     permissions:
       id-token: write
@@ -190,15 +207,15 @@ jobs:
         run: |
           set -eux
           mkdir -p "${RUNNER_TEMP}/artifacts/"
-          mv "${RUNNER_TEMP}"/artifacts-all/vllm-wheel-"${BUILD_DEVICE}"-*/* "${RUNNER_TEMP}/artifacts/"
+          mv "${RUNNER_TEMP}"/artifacts-all/vllm-wheel-"${BUILD_DEVICE}"-"${PLATFORM}"-*/* "${RUNNER_TEMP}/artifacts/"
 
-      - name: Set DRY_RUN (only for tagged pushes)
-        if: ${{ github.event_name == 'push' && (github.event.ref == 'refs/heads/main' || startsWith(github.event.ref, 'refs/tags/v')) }}
+      - name: Set DRY_RUN
+        if: ${{ (github.event_name == 'push' && (github.event.ref == 'refs/heads/main' || startsWith(github.event.ref, 'refs/tags/v'))) || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }}
         shell: bash
         run: |
           echo "DRY_RUN=disabled" >> "$GITHUB_ENV"
 
-      - name: Set UPLOAD_CHANNEL (only for tagged pushes)
+      - name: Set UPLOAD_CHANNEL
         if: ${{ github.event_name == 'push' && startsWith(github.event.ref, 'refs/tags/v') }}
         shell: bash
         run: |