bring back cu126 support

.github/scripts/filter-matrix.py

@@ -13,8 +13,10 @@
 # jetpack 6.2 only officially supports python 3.10 and cu126
 jetpack_python_versions: List[str] = ["3.10"]
 jetpack_cuda_versions: List[str] = ["cu126"]
-rtx_cuda_versions: List[str] = ["cu130", "cu132"]
-trt_cuda_versions: List[str] = ["cu130", "cu132"]
+# CUDA 12.6 wheels are published for x86_64 only. Keep the Arm matrices on
+# CUDA 13, including Windows Arm/AArch64.
+x86_cuda_versions: List[str] = ["cu126", "cu130", "cu132"]
+arm_cuda_versions: List[str] = ["cu130", "cu132"]
 
 # For PRs we build/test a single representative config to keep cycle time short.
 # Full matrix runs on main / nightly / release branches.
@@ -74,12 +76,16 @@ def filter_matrix_item(
             return True
         return False
     else:
+        cuda_versions = (
+            arm_cuda_versions
+            if item["gpu_arch_type"] in {"cuda-aarch64", "cuda-arm64"}
+            else x86_cuda_versions
+        )
         if use_rtx:
-            if item["desired_cuda"] not in rtx_cuda_versions:
-                return False
-        else:
-            if item["desired_cuda"] not in trt_cuda_versions:
+            if item["desired_cuda"] not in cuda_versions:
                 return False
+        elif item["desired_cuda"] not in cuda_versions:
+            return False
         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

.github/scripts/generate-release-matrix.py

@@ -5,8 +5,8 @@
 import sys
 
 RELEASE_CUDA_VERSION = {
-    "wheel": ["cu130", "cu132"],
-    "tarball": ["cu130", "cu132"],
+    "wheel": ["cu126", "cu130", "cu132"],
+    "tarball": ["cu126", "cu130", "cu132"],
 }
 RELEASE_PYTHON_VERSION = {
     "wheel": ["3.10", "3.11", "3.12", "3.13", "3.14"],
@@ -49,6 +49,13 @@ def main(args: list[str]) -> None:
     includes = matrix_dict["include"]
     filtered_includes = []
     for item in includes:
+        # CUDA 12.6 release artifacts are available only for x86_64. This
+        # excludes both Linux AArch64 and Windows Arm/AArch64 matrix entries.
+        if item["desired_cuda"] == "cu126" and item["gpu_arch_type"] in {
+            "cuda-aarch64",
+            "cuda-arm64",
+        }:
+            continue
         if (
             item["desired_cuda"] in cuda_versions
             and item["python_version"] in python_versions

.github/workflows/release-linux-x86_64.yml

@@ -124,6 +124,27 @@ jobs:
       smoke-test-script: ${{ matrix.smoke-test-script }}
       trigger-event: ${{ github.event_name }}
       is-release-wheel: true
+
+  release-executorch-runtime-wheel-artifacts:
+    name: Release x86_64 torch-tensorrt-executorch-runtime wheel artifacts
+    needs: [generate-release-wheel-matrix]
+    uses: ./.github/workflows/build_linux.yml
+    with:
+      repository: pytorch/tensorrt
+      ref: ""
+      test-infra-repository: pytorch/test-infra
+      test-infra-ref: main
+      build-matrix: ${{ needs.generate-release-wheel-matrix.outputs.matrix }}
+      pre-script: packaging/pre_build_script.sh
+      package-name: torch_tensorrt_executorch_runtime
+      build-platform: python-build-package
+      build-command: >-
+        bash -c 'python -m pip install pyyaml "executorch==1.4.1" && python -m pip wheel --no-build-isolation --no-deps --wheel-dir dist py/torch-tensorrt-executorch-runtime'
+      artifact-name-prefix: torch-tensorrt-executorch-runtime
+      trigger-event: ${{ github.event_name }}
+      is-release-wheel: true
+      run-smoke-test: false
+
 concurrency:
   group: ${{ github.workflow }}-x86_64-release-${{ github.event.pull_request.number || github.ref_name }}-${{ inputs.repository }}-${{ github.event_name == 'workflow_dispatch' }}-${{ inputs.job-name }}
-  cancel-in-progress: true
+  cancel-in-progress: true

.gitignore

@@ -1,4 +1,6 @@
 bazel
+/node_modules/
+/package.json
 bazel-bazel-test
 bazel-bin
 bazel-genfiles
@@ -95,4 +97,4 @@ CLAUDE.md
 /external/
 build-executorch/
 py/torch-tensorrt-executorch-runtime/build/
-py/torch-tensorrt-executorch-runtime/dist/
+py/torch-tensorrt-executorch-runtime/dist/

packaging/pre_build_script.sh

@@ -72,6 +72,23 @@ if [[ ! -d "${TORCH_INSTALL_PATH}/include/c10" ]]; then
     exit 1
 fi
 
+# TensorRT archives have different CUDA compatibility ceilings. CI provides
+# CU_VERSION in the PyTorch wheel format (for example, cu132).
+case "${CU_VERSION}" in
+    cu12*)
+        export TENSORRT_CUDA_VERSION_UPPER_BOUND="12.9"
+        export TENSORRT_RTX_CUDA_VERSION_UPPER_BOUND="12.9"
+        ;;
+    cu13*)
+        export TENSORRT_CUDA_VERSION_UPPER_BOUND="13.3"
+        export TENSORRT_RTX_CUDA_VERSION_UPPER_BOUND="13.4"
+        ;;
+    *)
+        echo "ERROR: Unsupported CUDA version '${CU_VERSION}' for TensorRT archive selection."
+        exit 1
+        ;;
+esac
+
 cat toolchains/ci_workspaces/MODULE.bazel.tmpl | envsubst > MODULE.bazel
 
 if [[ ${TENSORRT_VERSION} != "" ]]; then

packaging/pre_build_script_windows.sh

@@ -42,6 +42,23 @@ unzip fmt.zip
 cp -r fmt-12.0.0/include/fmt/ $TORCH_INSTALL_PATH/include/
 ls -lart $TORCH_INSTALL_PATH/include/fmt/
 
+# TensorRT archives have different CUDA compatibility ceilings. CI provides
+# CU_VERSION in the PyTorch wheel format (for example, cu132).
+case "${CU_VERSION}" in
+    cu12*)
+        export TENSORRT_CUDA_VERSION_UPPER_BOUND="12.9"
+        export TENSORRT_RTX_CUDA_VERSION_UPPER_BOUND="12.9"
+        ;;
+    cu13*)
+        export TENSORRT_CUDA_VERSION_UPPER_BOUND="13.3"
+        export TENSORRT_RTX_CUDA_VERSION_UPPER_BOUND="13.4"
+        ;;
+    *)
+        echo "ERROR: Unsupported CUDA version '${CU_VERSION}' for TensorRT archive selection."
+        exit 1
+        ;;
+esac
+
 cat toolchains/ci_workspaces/MODULE.bazel.tmpl | envsubst > MODULE.bazel
 
 if [[ ${TENSORRT_VERSION} != "" ]]; then

py/torch-tensorrt-executorch-runtime/setup.py

@@ -21,7 +21,6 @@
 REPO_ROOT = HERE.parents[1]
 BAZEL_TARGET = "//py/torch-tensorrt-executorch-runtime/native:delegate_native"
 BUILD_NONCE = os.getenv("TORCH_TENSORRT_EXECUTORCH_BUILD_NONCE", uuid.uuid4().hex)
-TENSORRT_DISTRIBUTION = "tensorrt-cu13"
 CUDA_RUNTIME_DISTRIBUTION = "nvidia-cuda-runtime"
 
 
@@ -50,14 +49,16 @@ def installed_version(distribution: str) -> str:
         ) from error
 
 
-def require_cuda_13() -> None:
-    """Reject builds whose native dependencies do not use supported CUDA 13."""
+def tensorrt_distribution() -> str:
+    """Return the TensorRT distribution matching the PyTorch CUDA build."""
     cuda_version = torch.version.cuda
-    if cuda_version is None or not cuda_version.startswith("13."):
-        raise RuntimeError(
-            "CUDA 13-enabled PyTorch is required to build this wheel "
-            f"(found CUDA {cuda_version or 'None'})"
-        )
+    if cuda_version is None:
+        raise RuntimeError("CUDA-enabled PyTorch is required to build this wheel")
+    if cuda_version.startswith("12.6"):
+        return "tensorrt-cu12"
+    if cuda_version.startswith("13."):
+        return "tensorrt-cu13"
+    raise RuntimeError(f"Unsupported CUDA version: {cuda_version}")
 
 
 class BazelExtension(Extension):
@@ -139,7 +140,7 @@ def build_extension(self, ext: Extension) -> None:
             shutil.copy2(source, output.parent / dependency)
 
 
-require_cuda_13()
+TENSORRT_DISTRIBUTION = tensorrt_distribution()
 executorch_version = installed_version("executorch")
 tensorrt_version = installed_version(TENSORRT_DISTRIBUTION)
 cuda_runtime_version = installed_version(CUDA_RUNTIME_DISTRIBUTION)