add torch-tensorrt-executorch-delegate wheel build

.github/scripts/verify-executorch-reference-runner.sh

@@ -6,11 +6,12 @@ set +x
 #
 #   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 example_executorch_runner, and runs one inference.
+#   3. Provide a Torch-TensorRT ExecuTorch .pte model.
+#   4. This script unpacks libtorchtrt.tar.gz, configures and builds the
+#      packaged CMake runner, and runs one inference.
 #
 # Required:
+#   First argument: path to an existing .pte model.
 #   EXECUTORCH_SOURCE_DIR=/path/to/executorch
 #
 # Optional:
@@ -28,6 +29,16 @@ set +x
 repo_root="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
 cd "${repo_root}"
 
+if [[ $# -ne 1 ]]; then
+  echo "Usage: $0 PATH_TO_MODEL.pte" >&2
+  exit 1
+fi
+model_path="$1"
+if [[ ! -f "${model_path}" ]]; then
+  echo "ExecuTorch model not found: ${model_path}" >&2
+  exit 1
+fi
+
 python_executable="${PYTHON_EXECUTABLE:-}"
 if [[ -z "${python_executable}" ]]; then
   python_executable="$(command -v python || true)"
@@ -201,14 +212,13 @@ 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.
+# Fail early if the Python environment cannot run ExecuTorch's CMake codegen.
 if ! "${python_executable}" - <<'PY'
 import importlib
 import importlib.util
 
 missing = []
-for name in ("yaml", "torch", "torch_tensorrt", "executorch.exir"):
+for name in ("yaml", "torch", "executorch.exir"):
     try:
         spec = importlib.util.find_spec(name)
     except ModuleNotFoundError:
@@ -217,56 +227,17 @@ for name in ("yaml", "torch", "torch_tensorrt", "executorch.exir"):
         missing.append(name)
 if missing:
     raise SystemExit(
-        "Missing Python package(s) required to export the .pte and build the runner: "
+        "Missing Python package(s) required to build the runner: "
         + ", ".join(missing)
     )
 
-for name in ("yaml", "torch", "torch_tensorrt", "executorch.exir"):
+for name in ("yaml", "torch", "executorch.exir"):
     importlib.import_module(name)
 PY
 then
   exit 1
 fi
 
-model_path="${verify_root}/model.pte"
-"${python_executable}" - "${model_path}" <<'PY'
-import importlib.util
-import runpy
-import sys
-from pathlib import Path
-
-model_path = sys.argv[1]
-repo_root = Path.cwd()
-
-# Use the installed package for native extensions and the in-tree ExecuTorch
-# route for the serializer/backend under test.
-import torch_tensorrt  # noqa: F401
-
-
-def overlay_module(name: str, path: Path) -> None:
-    spec = importlib.util.spec_from_file_location(name, path)
-    if spec is None or spec.loader is None:
-        raise RuntimeError(f"Could not load {name} from {path}")
-    module = importlib.util.module_from_spec(spec)
-    sys.modules[name] = module
-    spec.loader.exec_module(module)
-
-
-overlay_module(
-    "torch_tensorrt.executorch.serialization",
-    repo_root / "py/torch_tensorrt/executorch/serialization.py",
-)
-overlay_module(
-    "torch_tensorrt.executorch.backend",
-    repo_root / "py/torch_tensorrt/executorch/backend.py",
-)
-
-export_script = repo_root / "examples/torchtrt_executorch_example/export_static_shape.py"
-sys.argv = [str(export_script), "--model_path", model_path]
-runpy.run_path(str(export_script), run_name="__main__")
-PY
-test -f "${model_path}"
-
 if [[ -n "${TensorRT_ROOT:-}" && -d "${TensorRT_ROOT}/lib" ]]; then
   export LD_LIBRARY_PATH="${TensorRT_ROOT}/lib${original_ld_library_path:+:${original_ld_library_path}}"
 else

.github/workflows/_linux-x86_64-core.yml

@@ -98,11 +98,22 @@ jobs:
       use-rtx: ${{ inputs.use-rtx }}
       pip-install-torch-extra-args: "--extra-index-url https://pypi.org/simple"
 
-  # Standard-TRT only: ExecuTorch static build is not part of the RTX matrix.
-  executorch-static-build:
+  # Standard-TRT only: ExecuTorch runtime build is not part of the RTX matrix.
+  executorch-runtime-build:
     needs: [filter-matrix, build]
     if: ${{ !inputs.use-rtx }}
-    uses: ./.github/workflows/executorch-static-linux.yml
+    uses: ./.github/workflows/executorch-build-linux.yml
+    with:
+      repository: "pytorch/tensorrt"
+      ref: ""
+      test-infra-repository: pytorch/test-infra
+      test-infra-ref: main
+      build-matrix: ${{ needs.filter-matrix.outputs.matrix }}
+
+  executorch-runtime-test:
+    needs: [filter-matrix, executorch-runtime-build]
+    if: ${{ !inputs.use-rtx && needs.executorch-runtime-build.result == 'success' }}
+    uses: ./.github/workflows/executorch-test-linux.yml
     with:
       repository: "pytorch/tensorrt"
       ref: ""
@@ -639,6 +650,8 @@ jobs:
     needs:
       [
         build,
+        executorch-runtime-build,
+        executorch-runtime-test,
         L0-dynamo-converter-tests,
         L0-dynamo-core-tests,
         L0-py-core-tests,

.github/workflows/executorch-static-linux.yml → .github/workflows/executorch-build-linux.yml

@@ -1,4 +1,4 @@
-name: ExecuTorch Static Linux Build
+name: ExecuTorch Delegate Linux Build
 
 on:
   workflow_call:
@@ -60,14 +60,15 @@ jobs:
     needs: select-matrix
     uses: ./.github/workflows/linux-test.yml
     with:
-      job-name: executorch-static-build
+      job-name: executorch-runtime-build
       repository: ${{ inputs.repository }}
       ref: ${{ inputs.ref }}
       test-infra-repository: ${{ inputs.test-infra-repository }}
       test-infra-ref: ${{ inputs.test-infra-ref }}
       build-matrix: ${{ needs.select-matrix.outputs.matrix }}
       pre-script: packaging/pre_build_script.sh
       fail-on-empty: false
+      upload-artifact: torch-tensorrt-executorch-runtime
       script: |
         set -euo pipefail
         BAZELISK_VERSION="1.26.0"
@@ -90,11 +91,8 @@ jobs:
         export PATH="${RUNNER_TEMP}/bin:${PATH}"
         bazel --version
 
-        # this is to build the libtorchtrt.tar.gz
-        bazel build //:libtorchtrt --compilation_mode opt --config=linux
-        executorch_cmake_location="$(bazel query @executorch//:executorch/CMakeLists.txt --output=location)"
-        export EXECUTORCH_SOURCE_DIR="$(dirname "${executorch_cmake_location%%:*}")"
-        export EXECUTORCH_ROOT="${EXECUTORCH_SOURCE_DIR}"
-        # this is to verify the end user's workflow
-        python -m pip install pyyaml "executorch>=1.3.1"
-        .github/scripts/verify-executorch-reference-runner.sh
+        python -m pip install pyyaml "executorch==1.3.1"
+        export TORCH_TENSORRT_EXECUTORCH_RUNTIME_VERSION="$(python -c 'import importlib.metadata; print(importlib.metadata.version("torch-tensorrt"))')"
+
+        # Build the no-compile-for-users Python runtime wheel.
+        python -m pip wheel --no-build-isolation --no-deps --wheel-dir dist py/torch-tensorrt-executorch-runtime

.github/workflows/executorch-test-linux.yml

@@ -0,0 +1,94 @@
+name: ExecuTorch Delegate Linux Test
+
+on:
+  workflow_call:
+    inputs:
+      build-matrix:
+        description: "Build matrix to utilize"
+        default: ""
+        type: string
+      repository:
+        description: 'Repository to checkout, defaults to ""'
+        default: ""
+        type: string
+      ref:
+        description: 'Reference to checkout, defaults to ""'
+        default: ""
+        type: string
+      test-infra-repository:
+        description: "Test infra repository to use"
+        default: "pytorch/test-infra"
+        type: string
+      test-infra-ref:
+        description: "Test infra reference to use"
+        default: ""
+        type: string
+
+jobs:
+  select-matrix:
+    runs-on: ubuntu-latest
+    outputs:
+      matrix: ${{ steps.pick.outputs.matrix }}
+    steps:
+      - id: pick
+        env:
+          FULL_MATRIX: ${{ inputs.build-matrix }}
+        run: |
+          set -euo pipefail
+          python - <<'PY' >> "${GITHUB_OUTPUT}"
+          import json
+          import os
+
+          matrix = json.loads(os.environ["FULL_MATRIX"])
+          include = matrix.get("include", [])
+          if not include:
+              raise SystemExit("build-matrix include[] is empty")
+
+          preferred = None
+          for entry in include:
+              if entry.get("python_version") == "3.11":
+                  preferred = entry
+                  break
+
+          if preferred is None:
+              preferred = include[0]
+
+          print("matrix=" + json.dumps({"include": [preferred]}))
+          PY
+
+  test:
+    needs: select-matrix
+    uses: ./.github/workflows/linux-test.yml
+    with:
+      job-name: executorch-runtime-test
+      repository: ${{ inputs.repository }}
+      ref: ${{ inputs.ref }}
+      test-infra-repository: ${{ inputs.test-infra-repository }}
+      test-infra-ref: ${{ inputs.test-infra-ref }}
+      build-matrix: ${{ needs.select-matrix.outputs.matrix }}
+      pre-script: packaging/pre_build_script.sh
+      additional-artifact: torch-tensorrt-executorch-runtime
+      script: |
+        set -euo pipefail
+        mkdir -p "${RUNNER_TEMP}/bin"
+        curl -L "https://github.com/bazelbuild/bazelisk/releases/download/v1.26.0/bazelisk-linux-amd64" \
+          -o "${RUNNER_TEMP}/bin/bazel"
+        chmod +x "${RUNNER_TEMP}/bin/bazel"
+        export PATH="${RUNNER_TEMP}/bin:${PATH}"
+
+        python -m pip install pyyaml executorch
+        python -c 'from torch_tensorrt_executorch_runtime import get_runtime; assert get_runtime().backend_registry.is_available("TensorRTBackend")'
+
+        source tests/py/utils/ci_helpers.sh
+        trt_tier_executorch
+
+        bazel build //:libtorchtrt --compilation_mode opt --config=linux
+        executorch_cmake_location="$(bazel query @executorch//:executorch/CMakeLists.txt --output=location)"
+        export EXECUTORCH_SOURCE_DIR="$(dirname "${executorch_cmake_location%%:*}")"
+        export EXECUTORCH_ROOT="${EXECUTORCH_SOURCE_DIR}"
+        python examples/torchtrt_executorch_example/export_static_shape.py \
+          --model_path="${RUNNER_TEMP}/torchtrt-python.pte"
+        .github/scripts/verify-executorch-reference-runner.sh \
+          "${RUNNER_TEMP}/torchtrt-python.pte"
+        python examples/executorch_reference_runner/load_model.py \
+          --model_path="${RUNNER_TEMP}/torchtrt-python.pte" --num_runs=1

.github/workflows/linux-test.yml

@@ -50,9 +50,13 @@ on:
         default: false
         type: boolean
       upload-artifact:
-        description: 'Name to give artifacts uploaded from ${RUNNER_ARTIFACT_DIR}'
+        description: 'Name to give wheel artifacts uploaded from the repository dist directory'
         default: ''
         type: string
+      additional-artifact:
+        description: "Optional additional artifact to install alongside the core wheel"
+        default: ""
+        type: string
       use-rtx:
         description: "Whether to use TensorRT-RTX"
         default: false
@@ -146,6 +150,12 @@ jobs:
       #   with:
       #     repository: ${{ inputs.repository  }}
       #     script: .github/scripts/install-torch-tensorrt.sh
+      - name: Download additional artifacts
+        if: ${{ inputs.additional-artifact != '' }}
+        uses: actions/download-artifact@v7
+        with:
+          name: ${{ inputs.additional-artifact }}
+          path: /opt/torch-tensorrt-builds/
       - name: Pack script
         continue-on-error: ${{ inputs.continue-on-error }}
         working-directory: ${{ inputs.repository }}
@@ -174,23 +184,10 @@ jobs:
           display-options: fEs
           fail-on-empty: ${{ inputs.fail-on-empty }}
 
-      - name: Prepare artifacts for upload
+      - name: Check documentation for upload
         working-directory: ${{ inputs.repository }}
         id: check-artifacts
-        env:
-          UPLOAD_ARTIFACT_NAME: ${{ inputs.upload-artifact }}
         run: |
-          # Only do these steps if we actually want to upload an artifact
-          if [[ -n "${UPLOAD_ARTIFACT_NAME}" ]]; then
-            # If the default execution path is followed then we should get a wheel in the dist/ folder
-            # attempt to just grab whatever is in there and scoop it all up
-            if find "dist/" -name "*.whl" >/dev/null 2>/dev/null; then
-              mv -v dist/*.whl "${RUNNER_ARTIFACT_DIR}/"
-            fi
-            # Set to fail upload step if there are no files for upload and expected files for upload
-            echo 'if-no-files-found=error' >> "${GITHUB_OUTPUT}"
-          fi
-
           upload_docs=0
           # Check if there are things in the documentation folder to upload
           if find "${RUNNER_DOCS_DIR}" -mindepth 1 -maxdepth 1 | read -r; then
@@ -205,8 +202,8 @@ jobs:
         if: ${{ inputs.upload-artifact != '' }}
         with:
           name: ${{ inputs.upload-artifact }}
-          path: ${{ runner.temp }}/artifacts/
-          if-no-files-found: ${{ steps.check-artifacts.outputs.if-no-files-found }}
+          path: ${{ inputs.repository }}/dist/
+          if-no-files-found: error
 
       - name: Upload documentation to S3 (if any)
         uses: seemethere/upload-artifact-s3@v5

.gitignore

@@ -94,3 +94,5 @@ CLAUDE.md
 /compile_commands.json
 /external/
 build-executorch/
+py/torch-tensorrt-executorch-runtime/build/
+py/torch-tensorrt-executorch-runtime/dist/

BUILD.bazel

@@ -53,6 +53,14 @@ config_setting(
     },
 )
 
+filegroup(
+    name = "torchtrt_cmake_modules",
+    srcs = [
+        "cmake/Modules/FindTensorRT.cmake",
+    ],
+    visibility = ["//visibility:public"],
+)
+
 pkg_tar(
     name = "include_core",
     package_dir = "include/torch_tensorrt",

MODULE.bazel

@@ -32,8 +32,8 @@ git_override(
 bazel_dep(name = "hedron_compile_commands", dev_dependency = True)
 git_override(
     module_name = "hedron_compile_commands",
-    remote = "https://github.com/hedronvision/bazel-compile-commands-extractor.git",
     commit = "0e990032f3c5a866e72615cf67e5ce22186dcb97",
+    remote = "https://github.com/hedronvision/bazel-compile-commands-extractor.git",
 )
 
 new_local_repository = use_repo_rule("@bazel_tools//tools/build_defs/repo:local.bzl", "new_local_repository")
@@ -42,12 +42,13 @@ new_git_repository = use_repo_rule("@bazel_tools//tools/build_defs/repo:git.bzl"
 
 local_torch = use_repo_rule("//toolchains:local_torch.bzl", "local_torch")
 
-# Pinned to the ExecuTorch release/1.3 branch head.
+# Keep this pin synchronized with the ExecuTorch release installed in
+# py/torch-tensorrt-executorch-runtime/README.md.
 new_git_repository(
     name = "executorch",
     build_file = "@//third_party/executorch:BUILD",
-    # latest commit in release/1.3 branch
-    commit = "6118688a095fd9697224f5cad72ce42db641c9cd",
+    # executorch==1.3.1
+    commit = "e2f18eb23c45bd22ca332b0b8b49a81de304b472",
     patch_cmds = [
         "find . -mindepth 2 \\( -name BUILD -o -name BUILD.bazel \\) -delete",
         "mkdir executorch && find . -mindepth 1 -maxdepth 1 ! -name executorch ! -name BUILD ! -name BUILD.bazel ! -name REPO.bazel -exec cp -a {} executorch/ \\;",