pytorch · lanluo-nvidia · May 9, 2026 · Mar 18, 2026 · Mar 18, 2026 · Mar 24, 2026
diff --git a/.github/scripts/verify-executorch-reference-runner.sh b/.github/scripts/verify-executorch-reference-runner.sh
@@ -0,0 +1,283 @@
+#!/usr/bin/env bash
+set +x
+
+# Verifies the documented end-user flow for the ExecuTorch reference runner:
+#
+#   1. Build //:libtorchtrt first so bazel-bin/libtorchtrt.tar.gz exists.
+#   2. Provide an ExecuTorch source checkout with EXECUTORCH_SOURCE_DIR.
+#   3. This script exports a small Torch-TensorRT ExecuTorch .pte model,
+#      unpacks libtorchtrt.tar.gz, configures the packaged CMake runner,
+#      builds my_runner, and runs one inference.
+#
+# Required:
+#   EXECUTORCH_SOURCE_DIR=/path/to/executorch
+#
+# Optional:
+#   TensorRT_ROOT=/path/to/extracted/TensorRT
+#     If unset, the script reuses Bazel's fetched TensorRT SDK when available
+#     and otherwise downloads the archive pinned in MODULE.bazel.
+#   RUNNER_TEMP=/path/to/temp-root
+#     Parent directory for the temporary verification workspace.
+#   MAX_JOBS=N
+#     Parallelism passed to cmake --build.
+#   TORCHTRT_TENSORRT_DISTDIR=/path/to/cache
+#   TORCHTRT_TENSORRT_EXTRACT_DIR=/path/to/extracted-sdk
+#     Override locations used only by the TensorRT download fallback.
+
+repo_root="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+cd "${repo_root}"
+
+python_executable="${PYTHON_EXECUTABLE:-}"
+if [[ -z "${python_executable}" ]]; then
+  python_executable="$(command -v python || true)"
+fi
+if [[ -z "${python_executable}" ]]; then
+  echo "Could not find python on PATH" >&2
+  exit 1
+fi
+export PYTHON_EXECUTABLE="${python_executable}"
+
+: "${EXECUTORCH_SOURCE_DIR:?Set EXECUTORCH_SOURCE_DIR to an ExecuTorch source checkout}"
+
+if [[ ! -f "${EXECUTORCH_SOURCE_DIR}/CMakeLists.txt" ]]; then
+  echo "EXECUTORCH_SOURCE_DIR must point to an ExecuTorch source checkout: ${EXECUTORCH_SOURCE_DIR}" >&2
+  exit 1
+fi
+
+tarball="${repo_root}/bazel-bin/libtorchtrt.tar.gz"
+if [[ ! -f "${tarball}" ]]; then
+  echo "Missing ${tarball}; build //:libtorchtrt before running this check" >&2
+  exit 1
+fi
+
+verify_parent="${RUNNER_TEMP:-/tmp}"
+mkdir -p "${verify_parent}"
+verify_root="$(mktemp -d "${verify_parent%/}/torchtrt_executorch_readme_verify.XXXXXX")"
+
+# Prefer the TensorRT SDK that Bazel already fetched for //:libtorchtrt. This
+# keeps CI from downloading the same SDK twice and keeps CMake linked against
+# the same TensorRT version used to build the release tarball.
+find_bazel_tensorrt_root() {
+  local repo_name="$1"
+  local output_base
+  local trt_header
+  local tensorrt_root
+
+  if ! command -v bazel >/dev/null 2>&1; then
+    return 1
+  fi
+
+  output_base="$(bazel info output_base 2>/dev/null)" || return 1
+  trt_header="$(
+    find -L "${output_base}/external" \
+      \( -path "*/+*${repo_name}/include/NvInfer.h" -o -path "*/${repo_name}/include/NvInfer.h" \) \
+      -print -quit 2>/dev/null || true
+  )"
+
+  if [[ -z "${trt_header}" ]]; then
+    return 1
+  fi
+
+  tensorrt_root="$(dirname "$(dirname "${trt_header}")")"
+  echo "Using Bazel TensorRT SDK: ${tensorrt_root}" >&2
+  printf '%s\n' "${tensorrt_root}"
+}
+
+# The archive repo is architecture-specific in MODULE.bazel.
+select_tensorrt_archive_repo() {
+  case "$(uname -m)" in
+    aarch64|arm64)
+      echo "tensorrt_sbsa"
+      ;;
+    x86_64|amd64)
+      echo "tensorrt"
+      ;;
+    *)
+      return 1
+      ;;
+  esac
+}
+
+# Extract the pinned TensorRT URL and strip_prefix from MODULE.bazel so the
+# verifier follows the same dependency pin as Bazel without duplicating it here.
+read_tensorrt_archive_metadata() {
+  local repo_name="$1"
+
+  "${python_executable}" - "${repo_name}" <<'PY'
+import re
+import sys
+from pathlib import Path
+
+repo_name = sys.argv[1]
+module_bazel = Path("MODULE.bazel").read_text()
+
+for match in re.finditer(r"http_archive\((?P<body>.*?)\n\)", module_bazel, re.DOTALL):
+    body = match.group("body")
+    name = re.search(r'name\s*=\s*"([^"]+)"', body)
+    if name is None or name.group(1) != repo_name:
+        continue
+
+    url = re.search(r'urls\s*=\s*\[\s*"([^"]+)"', body, re.DOTALL)
+    strip_prefix = re.search(r'strip_prefix\s*=\s*"([^"]+)"', body)
+    if url is None:
+        raise SystemExit(f"Could not find urls[] for {repo_name} in MODULE.bazel")
+
+    print(url.group(1), strip_prefix.group(1) if strip_prefix else "")
+    break
+else:
+    raise SystemExit(f'Could not find http_archive(name = "{repo_name}") in MODULE.bazel')
+PY
+}
+
+# Download TensorRT only when it cannot be found in Bazel's external repo cache
+# and TensorRT_ROOT was not provided by the caller.
+download_tensorrt_root() {
+  local repo_name="$1"
+  local tensorrt_url
+  local tensorrt_strip_prefix
+  local tensorrt_distdir
+  local tensorrt_extract_dir
+  local tensorrt_archive
+  local tensorrt_root
+
+  echo "Downloading TensorRT SDK for ${repo_name}" >&2
+
+  read -r tensorrt_url tensorrt_strip_prefix < <(read_tensorrt_archive_metadata "${repo_name}") || return 1
+
+  tensorrt_distdir="${TORCHTRT_TENSORRT_DISTDIR:-${verify_root}/tensorrt-distdir}"
+  tensorrt_extract_dir="${TORCHTRT_TENSORRT_EXTRACT_DIR:-${verify_root}/tensorrt-sdk}"
+  tensorrt_archive="${tensorrt_distdir}/$(basename "${tensorrt_url}")"
+
+  mkdir -p "${tensorrt_distdir}" "${tensorrt_extract_dir}"
+  if [[ ! -f "${tensorrt_archive}" ]]; then
+    curl -fL "${tensorrt_url}" -o "${tensorrt_archive}" || return 1
+  fi
+  tar -xzf "${tensorrt_archive}" -C "${tensorrt_extract_dir}" || return 1
+
+  if [[ -n "${tensorrt_strip_prefix}" ]]; then
+    tensorrt_root="${tensorrt_extract_dir}/${tensorrt_strip_prefix}"
+  else
+    tensorrt_root="$(find "${tensorrt_extract_dir}" -mindepth 1 -maxdepth 1 -type d -print -quit)"
+  fi
+
+  if [[ ! -f "${tensorrt_root}/include/NvInfer.h" ]]; then
+    echo "TensorRT_ROOT does not contain include/NvInfer.h: ${tensorrt_root}" >&2
+    return 1
+  fi
+
+  echo "Using downloaded TensorRT SDK: ${tensorrt_root}" >&2
+  printf '%s\n' "${tensorrt_root}"
+}
+
+if [[ -z "${TensorRT_ROOT:-}" ]]; then
+  tensorrt_repo_name="$(select_tensorrt_archive_repo)" || {
+    echo "Unsupported TensorRT archive platform: $(uname -m)" >&2
+    exit 1
+  }
+  TensorRT_ROOT="$(find_bazel_tensorrt_root "${tensorrt_repo_name}" || download_tensorrt_root "${tensorrt_repo_name}" || true)"
+  if [[ -n "${TensorRT_ROOT}" ]]; then
+    export TensorRT_ROOT
+  fi
+elif [[ ! -f "${TensorRT_ROOT}/include/NvInfer.h" ]]; then
+  echo "TensorRT_ROOT must point to an extracted TensorRT SDK with include/NvInfer.h: ${TensorRT_ROOT}" >&2
+  exit 1
+fi
+
+# torch_tensorrt and the native runner both need TensorRT/PyTorch shared
+# libraries on the runtime path. Set this before importing torch_tensorrt or
+# exporting the .pte model.
+torch_lib_dir="$("${python_executable}" - <<'PY'
+import os
+import torch
+
+print(os.path.join(os.path.dirname(torch.__file__), "lib"))
+PY
+)"
+
+if [[ -n "${TensorRT_ROOT:-}" && -d "${TensorRT_ROOT}/lib" ]]; then
+  export LD_LIBRARY_PATH="${TensorRT_ROOT}/lib:${torch_lib_dir}${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
+else
+  export LD_LIBRARY_PATH="${torch_lib_dir}${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}"
+fi
+
+# Fail early if the Python environment cannot export a Torch-TensorRT
+# ExecuTorch model or run ExecuTorch's CMake codegen.
+if ! "${python_executable}" - <<'PY'
+import importlib
+import importlib.util
+
+missing = [
+    name
+    for name in ("yaml", "torch", "torch_tensorrt", "executorch.exir")
+    if importlib.util.find_spec(name) is None
+]
+if missing:
+    raise SystemExit(
+        "Missing Python package(s) required to export the .pte and build the runner: "
+        + ", ".join(missing)
+    )
+
+for name in ("yaml", "torch", "torch_tensorrt", "executorch.exir"):
+    importlib.import_module(name)
+PY
+then
+  exit 1
+fi
+
+model_path="${verify_root}/model.pte"
+"${python_executable}" examples/torchtrt_executorch_example/export_static_shape.py --model_path="${model_path}"
+test -f "${model_path}"
+
+tar -xzf "${tarball}" -C "${verify_root}"
+
+# Check the release tarball contract used by the README: libtorchtrt_executorch
+# must be a top-level source package, separate from torch_tensorrt/.
+tar_entries="${verify_root}/libtorchtrt_tar_entries.txt"
+tar -tf "${tarball}" > "${tar_entries}"
+grep -qx "libtorchtrt_executorch/CMakeLists.txt" "${tar_entries}"
+grep -qx "libtorchtrt_executorch/examples/executorch_reference_runner/CMakeLists.txt" "${tar_entries}"
+grep -qx "torch_tensorrt/BUILD" "${tar_entries}"
+
+if grep -q "^torch_tensorrt/libtorchtrt_executorch/" "${tar_entries}"; then
+  echo "libtorchtrt_executorch must be a top-level tar entry, not nested under torch_tensorrt/" >&2
+  exit 1
+fi
+
+export TORCHTRT_EXECUTORCH_SOURCE_DIR="${verify_root}/libtorchtrt_executorch"
+
+if [[ -z "${CMAKE_PREFIX_PATH:-}" ]]; then
+  CMAKE_PREFIX_PATH="$("${python_executable}" -c "import torch; print(torch.utils.cmake_prefix_path)")"
+  export CMAKE_PREFIX_PATH
+fi
+
+# Configure the example exactly as an end user would after unpacking
+# libtorchtrt.tar.gz.
+cmake_args=(
+  -S "${TORCHTRT_EXECUTORCH_SOURCE_DIR}/examples/executorch_reference_runner"
+  -B "${verify_root}/build-executorch-reference-runner"
+  -DEXECUTORCH_SOURCE_DIR="${EXECUTORCH_SOURCE_DIR}"
+  -DTORCHTRT_EXECUTORCH_SOURCE_DIR="${TORCHTRT_EXECUTORCH_SOURCE_DIR}"
+  -DCMAKE_PREFIX_PATH="${CMAKE_PREFIX_PATH}"
+  -DPYTHON_EXECUTABLE="${python_executable}"
+)
+
+if [[ -n "${TensorRT_ROOT:-}" ]]; then
+  cmake_args+=(-DTensorRT_ROOT="${TensorRT_ROOT}")
+fi
+
+cmake "${cmake_args[@]}"
+
+cmake --build "${verify_root}/build-executorch-reference-runner" \
+  --target my_runner \
+  -j"${MAX_JOBS:-$(nproc)}"
+
+runner_log="${verify_root}/my_runner.log"
+"${verify_root}/build-executorch-reference-runner/my_runner" \
+  --model_path="${model_path}" \
+  --num_runs=1 2>&1 | tee "${runner_log}"
+
+# The sample model is x + 1, and my_runner fills inputs with 1.0f, so the
+# output sample should contain 2.0000.
+grep -q "Inference completed" "${runner_log}"
+grep -q "output\\[0\\] shape=" "${runner_log}"
+grep -Eq "first [0-9]+ values:.* 2\\.0000" "${runner_log}"
diff --git a/.github/workflows/build-test-linux-x86_64.yml b/.github/workflows/build-test-linux-x86_64.yml
@@ -79,7 +79,18 @@
             use-rtx: false
             pip-install-torch-extra-args: "--extra-index-url https://pypi.org/simple"
 
+    executorch-static-build:
+        needs: [filter-matrix, build]
+        uses: ./.github/workflows/executorch-static-linux.yml
+        with:
+          repository: "pytorch/tensorrt"
+          ref: ""
+          test-infra-repository: pytorch/test-infra
+          test-infra-ref: main
+          build-matrix: ${{ needs.filter-matrix.outputs.matrix }}
+
+
     L0-dynamo-converter-tests:
         name: ${{ matrix.display-name }}
         needs: [filter-matrix, build]
        if: ${{ (github.ref_name == 'main' || github.ref_name == 'nightly' || contains(github.event.pull_request.labels.*.name, 'Force All Tests[L0+L1+L2]')) && always() || success() }}
@@ -482,6 +493,7 @@
                 pushd .
                 cd tests/py/dynamo
                 python -m pytest -ra --junitxml=${RUNNER_TEST_RESULTS_DIR}/l2_dynamo_core_tests_results.xml  -k "not test_000_ and not test_001_" runtime/*
+                python -m pytest -ra --junitxml=${RUNNER_TEST_RESULTS_DIR}/l2_dynamo_executorch_tests_results.xml executorch/
                 popd
 
     L2-dynamo-plugin-tests: