Add jetson build on CI (#3524)

Co-authored-by: Naren Dasan <naren@narendasan.com>

Author: lanluo-nvidia

.github/scripts/filter-matrix.py

@@ -9,6 +9,13 @@
 # currently we don't support python 3.13t due to tensorrt does not support 3.13t
 disabled_python_versions: List[str] = ["3.13t"]
 
+# jetpack 6.2 only officially supports python 3.10 and cu126
+jetpack_python_versions: List[str] = ["3.10"]
+jetpack_cuda_versions: List[str] = ["cu126"]
+
+jetpack_container_image: str = "nvcr.io/nvidia/l4t-jetpack:r36.4.0"
+sbsa_container_image: str = "quay.io/pypa/manylinux_2_34_aarch64"
+
 
 def main(args: list[str]) -> None:
     parser = argparse.ArgumentParser()
@@ -19,8 +26,23 @@ def main(args: list[str]) -> None:
         default="",
     )
 
-    options = parser.parse_args(args)
+    parser.add_argument(
+        "--jetpack",
+        help="is jetpack",
+        type=str,
+        choices=["true", "false"],
+        default="false",
+    )
 
+    parser.add_argument(
+        "--limit-pr-builds",
+        help="If it is a PR build",
+        type=str,
+        choices=["true", "false"],
+        default=os.getenv("LIMIT_PR_BUILDS", "false"),
+    )
+
+    options = parser.parse_args(args)
     if options.matrix == "":
         raise Exception("--matrix needs to be provided")
 
@@ -30,14 +52,29 @@ def main(args: list[str]) -> None:
     for item in includes:
         if item["python_version"] in disabled_python_versions:
             continue
-        if item["gpu_arch_type"] == "cuda-aarch64":
-            # pytorch image:pytorch/manylinuxaarch64-builder:cuda12.8 comes with glibc2.28
-            # however, TensorRT requires glibc2.31 on aarch64 platform
-            # TODO: in future, if pytorch supports aarch64 with glibc2.31, we should switch to use the pytorch image
-            item["container_image"] = "quay.io/pypa/manylinux_2_34_aarch64"
-            filtered_includes.append(item)
+        if options.jetpack == "true":
+            if options.limit_pr_builds == "true":
+                # limit pr build, matrix passed in from test-infra is cu128, python 3.9, change to cu126, python 3.10
+                item["desired_cuda"] = "cu126"
+                item["python_version"] = "3.10"
+                item["container_image"] = jetpack_container_image
+                filtered_includes.append(item)
+            else:
+                if (
+                    item["python_version"] in jetpack_python_versions
+                    and item["desired_cuda"] in jetpack_cuda_versions
+                ):
+                    item["container_image"] = jetpack_container_image
+                    filtered_includes.append(item)
         else:
-            filtered_includes.append(item)
+            if item["gpu_arch_type"] == "cuda-aarch64":
+                # pytorch image:pytorch/manylinuxaarch64-builder:cuda12.8 comes with glibc2.28
+                # however, TensorRT requires glibc2.31 on aarch64 platform
+                # TODO: in future, if pytorch supports aarch64 with glibc2.31, we should switch to use the pytorch image
+                item["container_image"] = sbsa_container_image
+                filtered_includes.append(item)
+            else:
+                filtered_includes.append(item)
     filtered_matrix_dict = {}
     filtered_matrix_dict["include"] = filtered_includes
     print(json.dumps(filtered_matrix_dict))

.github/workflows/build-test-linux-aarch64-jetpack.yml

@@ -0,0 +1,86 @@
+name: Build and test Linux aarch64 wheels for Jetpack
+
+on:
+  pull_request:
+  push:
+    branches:
+      - main
+      - nightly
+      - release/*
+    tags:
+      # NOTE: Binary build pipelines should only get triggered on release candidate builds
+      # Release candidate tags look like: v1.11.0-rc1
+      - v[0-9]+.[0-9]+.[0-9]+-rc[0-9]+
+  workflow_dispatch:
+
+jobs:
+  generate-matrix:
+    uses: pytorch/test-infra/.github/workflows/generate_binary_build_matrix.yml@main
+    with:
+      package-type: wheel
+      os: linux-aarch64
+      test-infra-repository: pytorch/test-infra
+      test-infra-ref: main
+      with-rocm: false
+      with-cpu: false
+
+  filter-matrix:
+    needs: [generate-matrix]
+    outputs:
+      matrix: ${{ steps.filter.outputs.matrix }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - uses: actions/checkout@v4
+        with:
+          repository: pytorch/tensorrt
+      - name: Filter matrix
+        id: filter
+        env:
+           LIMIT_PR_BUILDS: ${{ github.event_name == 'pull_request' && !contains( github.event.pull_request.labels.*.name, 'ciflow/binaries/all') }}
+        run: |
+          set -eou pipefail
+          echo "LIMIT_PR_BUILDS=${LIMIT_PR_BUILDS}"
+          MATRIX_BLOB=${{ toJSON(needs.generate-matrix.outputs.matrix) }}
+          MATRIX_BLOB="$(python3 .github/scripts/filter-matrix.py --matrix "${MATRIX_BLOB}" --jetpack true)"
+          echo "${MATRIX_BLOB}"
+          echo "matrix=${MATRIX_BLOB}" >> "${GITHUB_OUTPUT}"
+
+  build:
+    needs: filter-matrix
+    permissions:
+      id-token: write
+      contents: read
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - repository: pytorch/tensorrt
+            pre-script: packaging/pre_build_script.sh
+            env-var-script: packaging/env_vars.txt
+            post-script: packaging/post_build_script.sh
+            smoke-test-script: packaging/smoke_test_script.sh
+            package-name: torch_tensorrt
+    name: Build torch-tensorrt whl package
+    uses: ./.github/workflows/build_wheels_linux_aarch64.yml
+    with:
+      repository: ${{ matrix.repository }}
+      ref: ""
+      test-infra-repository: pytorch/test-infra
+      test-infra-ref: main
+      build-matrix: ${{ needs.filter-matrix.outputs.matrix }}
+      pre-script: ${{ matrix.pre-script }}
+      env-var-script: ${{ matrix.env-var-script }}
+      post-script: ${{ matrix.post-script }}
+      package-name: ${{ matrix.package-name }}
+      smoke-test-script: ${{ matrix.smoke-test-script }}
+      trigger-event: ${{ github.event_name }}
+      architecture: "aarch64"
+      is-jetpack: true
+
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ inputs.repository }}-${{ github.event_name == 'workflow_dispatch' }}-${{ inputs.job-name }}
+  cancel-in-progress: true

.github/workflows/build-test-linux-aarch64.yml

@@ -27,7 +27,7 @@ jobs:
   filter-matrix:
     needs: [generate-matrix]
     outputs:
-      matrix: ${{ steps.generate.outputs.matrix }}
+      matrix: ${{ steps.filter.outputs.matrix }}
     runs-on: ubuntu-latest
     steps:
       - uses: actions/setup-python@v5
@@ -36,8 +36,10 @@ jobs:
       - uses: actions/checkout@v4
         with:
           repository: pytorch/tensorrt
-      - name: Generate matrix
-        id: generate
+      - name: Filter matrix
+        id: filter
+        env:
+          LIMIT_PR_BUILDS: ${{ github.event_name == 'pull_request' && !contains( github.event.pull_request.labels.*.name, 'ciflow/binaries/all') }}
         run: |
           set -eou pipefail
           MATRIX_BLOB=${{ toJSON(needs.generate-matrix.outputs.matrix) }}

.github/workflows/build-test-linux-x86_64.yml

@@ -1,4 +1,4 @@
-name: Build and test Linux wheels
+name: Build and test Linux x86_64 wheels
 
 on:
   pull_request:

.github/workflows/build_wheels_linux_aarch64.yml

@@ -88,6 +88,11 @@ on:
         required: false
         default: "python -m build --wheel"
         type: string
+      is-jetpack:
+        description: Set to true if the build is for jetpack
+        required: false
+        default: false
+        type: boolean
       pip-install-torch-extra-args:
         # NOTE: Why does this exist?
         # Well setuptools / python packaging doesn't actually allow you to specify dependencies
@@ -128,7 +133,7 @@ jobs:
       UPLOAD_TO_BASE_BUCKET: ${{ matrix.upload_to_base_bucket }}
       ARCH: ${{ inputs.architecture }}
       BUILD_TARGET: ${{ inputs.build-target }}
-    name: build-${{ matrix.build_name }}
+    name: build-wheel-${{ matrix.python_version }}-${{ matrix.desired_cuda }}-${{ matrix.gpu_arch_type }}
     runs-on: ${{ matrix.validation_runner }}
     environment: ${{(inputs.trigger-event == 'schedule' || (inputs.trigger-event == 'push' && (startsWith(github.event.ref, 'refs/heads/nightly') || startsWith(github.event.ref, 'refs/tags/v')))) && 'pytorchbot-env' || ''}}
     container:
@@ -170,6 +175,11 @@ jobs:
           # when using Python version, less than the conda latest
           ###############################################################################
           echo 'Installing conda-forge'
+          if [[ ${{ inputs.is-jetpack }} == true ]]; then
+            # jetpack base image is ubuntu 22.04, does not have curl installed
+            apt-get update
+            apt-get install -y curl git
+          fi
           curl -L -o /mambaforge.sh https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-aarch64.sh
           chmod +x /mambaforge.sh
           /mambaforge.sh -b -p /opt/conda
@@ -195,12 +205,11 @@ jobs:
           python-version: ${{ env.PYTHON_VERSION }}
           cuda-version: ${{ env.CU_VERSION }}
           arch: ${{ env.ARCH }}
-
       - name: Combine Env Var and Build Env Files
         if: ${{ inputs.env-var-script != '' }}
         working-directory: ${{ inputs.repository }}
         run: |
-          set -euxo pipefail
+          set -x
           cat "${{ inputs.env-var-script }}" >> "${BUILD_ENV_FILE}"
       - name: Add XPU Env Vars in Build Env File
         if: ${{ matrix.gpu_arch_type == 'xpu' }}
@@ -211,6 +220,7 @@ jobs:
             echo "source /opt/intel/oneapi/pti/latest/env/vars.sh"
           } >> "${BUILD_ENV_FILE}"
       - name: Install torch dependency
+        if: ${{ inputs.is-jetpack == false }}
         run: |
           set -euxo pipefail
           # shellcheck disable=SC1090
@@ -241,12 +251,17 @@ jobs:
         working-directory: ${{ inputs.repository }}
         shell: bash -l {0}
         run: |
-          set -euxo pipefail
+          #set -euxo pipefail
+          set -x
           source "${BUILD_ENV_FILE}"
           export PYTORCH_VERSION="$(${CONDA_RUN} pip show torch | grep ^Version: | sed 's/Version: *//' | sed 's/+.\+//')"
           ${CONDA_RUN} python setup.py clean
           echo "Successfully ran `python setup.py clean`"
-          ${CONDA_RUN} python setup.py bdist_wheel
+          if [[ ${{ inputs.is-jetpack }} == false ]]; then
+            ${CONDA_RUN} python setup.py bdist_wheel
+          else
+            ${CONDA_RUN} python setup.py bdist_wheel --jetpack --plat-name=linux_tegra_aarch64
+          fi
       - name: Repair Manylinux_2_28 Wheel
         shell: bash -l {0}
         env:
@@ -272,6 +287,7 @@ jobs:
           script: ${{ inputs.post-script }}
       - name: Smoke Test
         shell: bash -l {0}
+        if: ${{ inputs.is-jetpack == false }}
         env:
           PACKAGE_NAME: ${{ inputs.package-name }}
           SMOKE_TEST_SCRIPT: ${{ inputs.smoke-test-script }}
@@ -316,7 +332,8 @@ jobs:
   upload:
     needs: build
     uses: pytorch/test-infra/.github/workflows/_binary_upload.yml@main
-    if: always()
+    # only upload to pytorch index for non jetpack builds
+    if: ${{ inputs.is-jetpack == false }}
     with:
       repository: ${{ inputs.repository }}
       ref: ${{ inputs.ref }}

MODULE.bazel

@@ -87,6 +87,15 @@ http_archive(
     urls = ["https://download.pytorch.org/libtorch/nightly/cu128/libtorch-win-shared-with-deps-latest.zip"],
 )
 
+http_archive(
+    name = "torch_l4t",
+    build_file = "@//third_party/libtorch:BUILD",
+    strip_prefix = "torch",
+    type = "zip",
+    urls = ["https://pypi.jetson-ai-lab.dev/jp6/cu126/+f/6ef/f643c0a7acda9/torch-2.7.0-cp310-cp310-linux_aarch64.whl"],
+    sha256 = "6eff643c0a7acda92734cc798338f733ff35c7df1a4434576f5ff7c66fc97319"
+)
+
 # Download these tarballs manually from the NVIDIA website
 # Either place them in the distdir directory in third_party and use the --distdir flag
 # or modify the urls to "file:///<PATH TO TARBALL>/<TARBALL NAME>.tar.gz

core/BUILD

@@ -66,6 +66,7 @@ cc_library(
     }) + select({
         ":windows": ["@libtorch_win//:libtorch"],
         ":use_torch_whl": ["@torch_whl//:libtorch"],
+        ":jetpack": ["@torch_l4t//:libtorch"],
         "//conditions:default": ["@libtorch"],
     }),
     alwayslink = True,

core/conversion/BUILD

@@ -61,6 +61,7 @@ cc_library(
     }) + select({
         ":windows": ["@libtorch_win//:libtorch"],
         ":use_torch_whl": ["@torch_whl//:libtorch"],
+        ":jetpack": ["@torch_l4t//:libtorch"],
         "//conditions:default": ["@libtorch"],
     }),
     alwayslink = True,

core/conversion/conversionctx/BUILD

@@ -56,6 +56,7 @@ cc_library(
     }) + select({
         ":windows": ["@libtorch_win//:libtorch"],
         ":use_torch_whl": ["@torch_whl//:libtorch"],
+        ":jetpack": ["@torch_l4t//:libtorch"],
         "//conditions:default": ["@libtorch"],
     }),
     alwayslink = True,

core/conversion/converters/BUILD

@@ -56,6 +56,7 @@ cc_library(
     }) + select({
         ":windows": ["@libtorch_win//:libtorch"],
         ":use_torch_whl": ["@torch_whl//:libtorch"],
+        ":jetpack": ["@torch_l4t//:libtorch"],
         "//conditions:default": ["@libtorch"],
     }),
     alwayslink = True,
@@ -81,6 +82,7 @@ cc_library(
     }) + select({
         ":windows": ["@libtorch_win//:libtorch"],
         ":use_torch_whl": ["@torch_whl//:libtorch"],
+        ":jetpack": ["@torch_l4t//:libtorch"],
         "//conditions:default": ["@libtorch"],
     }),
     alwayslink = True,
@@ -143,6 +145,7 @@ cc_library(
     }) + select({
         ":windows": ["@libtorch_win//:libtorch"],
         ":use_torch_whl": ["@torch_whl//:libtorch"],
+        ":jetpack": ["@torch_l4t//:libtorch"],
         "//conditions:default": ["@libtorch"],
     }),
     alwayslink = True,