lanluo-nvidia · Oct 6, 2026 · Oct 3, 2026
diff --git a/.github/scripts/filter-matrix.py b/.github/scripts/filter-matrix.py
@@ -17,8 +17,8 @@
 jetpack_python_versions: List[str] = ["3.10"]
 jetpack_cuda_versions: List[str] = ["cu126"]
 # CUDA 13 wheels are supported on x86_64 and Arm, including Windows Arm/AArch64.
-x86_cuda_versions: List[str] = ["cu130", "cu132", "cu134"]
-arm_cuda_versions: List[str] = ["cu130", "cu132", "cu134"]
+x86_cuda_versions: List[str] = ["cu132", "cu134"]
+arm_cuda_versions: List[str] = ["cu132", "cu134"]
 
 # For PRs we build/test a single representative config to keep cycle time short.
 # Full matrix runs on main / nightly / release branches.

diff --git a/.github/scripts/generate-release-matrix.py b/.github/scripts/generate-release-matrix.py
@@ -7,7 +7,8 @@
 import sys
 
 RELEASE_CUDA_VERSION = {
-    "wheel": ["cu130", "cu132", "cu134"],
+    # Wheels follow the CUDA rows in filter-matrix.py; libtorch still ships cu130 tarballs.
+    "wheel": ["cu132", "cu134"],
     "tarball": ["cu130", "cu132", "cu134"],
 }
 RELEASE_PYTHON_VERSION = {

diff --git a/.github/workflows/docgen.yml b/.github/workflows/docgen.yml
@@ -19,11 +19,11 @@ jobs:
         if: ${{ ! contains(github.actor, 'pytorchbot') }}
         environment: pytorchbot-env
         container:
-            image: docker.io/pytorch/manylinux2_28-builder:cuda13.0
+            image: docker.io/pytorch/manylinux2_28-builder:cuda13.2
         env:
-            CUDA_HOME: /usr/local/cuda-13.0
-            VERSION_SUFFIX: cu130
-            CU_VERSION: cu130
+            CUDA_HOME: /usr/local/cuda-13.2
+            VERSION_SUFFIX: cu132
+            CU_VERSION: cu132
             CHANNEL: nightly
             CI_BUILD: 1
         steps:
@@ -39,7 +39,7 @@ jobs:
             - name: Install base deps
               run: |
                   python3 -m pip install pip --upgrade
-                  python3 -m pip install pyyaml numpy torch --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+                  python3 -m pip install pyyaml numpy torch --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu132
                   ./packaging/pre_build_script.sh
             - name: Get HEAD SHA
               id: vars
@@ -49,7 +49,7 @@ jobs:
                   # Use the source-paired wheel, not a newer member of the authoring range.
                   python3 -m pip install ".[executorch]" \
                     "executorch==$(python3 -c 'import yaml;print(yaml.safe_load(open("dev_dep_versions.yml"))["__executorch_version__"])')" \
-                    --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+                    --extra-index-url https://download.pytorch.org/whl/nightly/cu132
             - name: Install uv
               run: |
                   curl -LsSf https://astral.sh/uv/install.sh | sh

diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
@@ -62,7 +62,7 @@ We use the PyTorch Slack for communication about core development, integration w
 
 ### Controlling CI scope via PR labels
 
-**PR default:** build + L0 + L1 on a single representative config (Python 3.12 × CUDA 13.4). L2 (slow model-level suites) is opt-in on PRs. The full matrix ({Python 3.10–3.13} × {CUDA 13.0, 13.2, 13.4}) runs on main / nightly / release branches.
+**PR default:** build + L0 + L1 on a single representative config (Python 3.12 × CUDA 13.4). L2 (slow model-level suites) is opt-in on PRs. The full matrix ({Python 3.10–3.13} × {CUDA 13.2, 13.4}) runs on main / nightly / release branches.
 
 Apply labels in the PR's right sidebar and re-push (or close/reopen) to re-trigger:
 

diff --git a/docsrc/getting_started/installation.rst b/docsrc/getting_started/installation.rst
@@ -46,7 +46,7 @@ Torch-TensorRT distributed nightlies targeting the PyTorch nightly. These can be
 
 .. code-block:: sh
 
-    python -m pip install --pre torch torch-tensorrt tensorrt --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+    python -m pip install --pre torch torch-tensorrt tensorrt --extra-index-url https://download.pytorch.org/whl/nightly/cu132
 
 
 
@@ -153,7 +153,7 @@ Once the WORKSPACE has been configured properly, all that is required to build t
 
     .. code-block:: sh
 
-        python -m pip install --pre . --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+        python -m pip install --pre . --extra-index-url https://download.pytorch.org/whl/nightly/cu132
 
 
 If you use the ``uv`` (`https://docs.astral.sh/uv/ <https://docs.astral.sh/uv/>`_) tool to manage python and your projects, the command is slightly simpler
@@ -168,7 +168,7 @@ To build the wheel file
 
     .. code-block:: sh
 
-        python -m pip wheel --no-deps --pre . --extra-index-url https://download.pytorch.org/whl/nightly/cu130 -w dist
+        python -m pip wheel --no-deps --pre . --extra-index-url https://download.pytorch.org/whl/nightly/cu132 -w dist
 
 Additional Build Options
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -186,7 +186,7 @@ which has implications for features like serialization.
 
 .. code-block:: sh
 
-    PYTHON_ONLY=1 python -m pip install --pre . --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+    PYTHON_ONLY=1 python -m pip install --pre . --extra-index-url https://download.pytorch.org/whl/nightly/cu132
 
 
 No TorchScript Frontend
@@ -197,7 +197,7 @@ of C++ code that is no longer necessary for most users. Therefore you can exclud
 
 .. code-block:: sh
 
-    NO_TORCHSCRIPT=1 python -m pip install --pre . --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+    NO_TORCHSCRIPT=1 python -m pip install --pre . --extra-index-url https://download.pytorch.org/whl/nightly/cu132
 
 
 Building the C++ Library Standalone (TorchScript Only)
@@ -277,7 +277,7 @@ Build steps
 
 * Open the app "x64 Native Tools Command Prompt for VS 2022" - note that Admin privileges may be necessary
 * Ensure Bazelisk (Bazel launcher) is installed on your machine and available from the command line. Package installers such as Chocolatey can be used to install Bazelisk
-* Install latest version of Torch (i.e. with ``pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu130``)
+* Install latest version of Torch (i.e. with ``pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu132``)
 * Clone the Torch-TensorRT repository and navigate to its root directory
 * Run ``pip install ninja wheel setuptools``
 * Run ``pip install --pre -r py/requirements.txt``

diff --git a/docsrc/getting_started/tensorrt_rtx.rst b/docsrc/getting_started/tensorrt_rtx.rst
@@ -55,7 +55,7 @@ CUDA version):
 
 .. code-block:: sh
 
-    python -m pip install --pre torch torch_tensorrt_rtx --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+    python -m pip install --pre torch torch_tensorrt_rtx --extra-index-url https://download.pytorch.org/whl/nightly/cu132
 
 
 Import Test

diff --git a/docsrc/getting_started/windows_arm64.rst b/docsrc/getting_started/windows_arm64.rst
@@ -173,7 +173,7 @@ and the resulting Torch-TensorRT wheel target CPython 3.13.
       python -m pip install numpy packaging pyyaml setuptools==72.1.0 wheel fmt build
 
       # only for run the setup.py, not use this torch wheelto build windows on arm
-      python -m pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu130
+      python -m pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu132
 
       # cross-build the ARM64 TensorRT-RTX wheel
       python setup.py bdist_wheel --use-rtx --windows-on-arm

diff --git a/docsrc/user_guide/runtime_performance/saving_models.rst b/docsrc/user_guide/runtime_performance/saving_models.rst
@@ -229,12 +229,12 @@ The ``executorch`` output format lowers the compiled module to an ExecuTorch
 ``.pte`` program, delegating the TensorRT engines to the Torch-TensorRT ExecuTorch
 backend. This CUDA integration links CUDA 13 libraries, so it supports Linux with
 any CUDA 13 PyTorch build and needs an ExecuTorch nightly wheel. Install from the
-channel matching your CUDA, for example a CUDA 13.0 environment:
+channel matching your CUDA, for example a CUDA 13.2 environment:
 
 .. code-block:: bash
 
     pip install --pre "torch_tensorrt[executorch]" \
-      --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+      --extra-index-url https://download.pytorch.org/whl/nightly/cu132
 
 Substitute the channel for your CUDA, such as ``cu134`` for CUDA 13.4. ExecuTorch
 publishes a channel later than PyTorch does, so a very new CUDA minor may not have

diff --git a/examples/executorch_reference_runner/README.md b/examples/executorch_reference_runner/README.md
@@ -102,7 +102,7 @@ same channel:
 
 ```bash
 pip install --pre "torch-tensorrt[executorch]" \
-  --extra-index-url https://download.pytorch.org/whl/nightly/cu130 \
+  --extra-index-url https://download.pytorch.org/whl/nightly/cu132 \
   --extra-index-url https://pypi.nvidia.com
 ```
 

diff --git a/examples/torchtrt_executorch_example/export_coalesced.py b/examples/torchtrt_executorch_example/export_coalesced.py
@@ -45,7 +45,7 @@
 Install Torch-TensorRT with the ExecuTorch extra before running this example::
 
     pip install -e ".[executorch]" \
-        --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+        --extra-index-url https://download.pytorch.org/whl/nightly/cu132
 
 ExecuTorch's CUDA backend also needs a CUDA toolkit (``nvcc``) at export time,
 for the AOTInductor compile.

diff --git a/examples/torchtrt_executorch_example/export_device_resident.py b/examples/torchtrt_executorch_example/export_device_resident.py
@@ -60,7 +60,7 @@
 Install Torch-TensorRT with the ExecuTorch extra before running this example::
 
     pip install -e ".[executorch]" \
-        --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+        --extra-index-url https://download.pytorch.org/whl/nightly/cu132
 
 ExecuTorch's CUDA backend also needs a CUDA toolkit (``nvcc``) at export time,
 for the AOTInductor compile.

diff --git a/examples/torchtrt_executorch_example/export_dynamic_shape.py b/examples/torchtrt_executorch_example/export_dynamic_shape.py
@@ -20,7 +20,7 @@
 Install Torch-TensorRT with the ExecuTorch extra before running this example::
 
     pip install -e ".[executorch]" \
-        --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+        --extra-index-url https://download.pytorch.org/whl/nightly/cu132
 
 See https://pytorch.org/executorch/stable/getting-started-setup.html for details.
 """

diff --git a/examples/torchtrt_executorch_example/export_kv_cache_decode.py b/examples/torchtrt_executorch_example/export_kv_cache_decode.py
@@ -32,7 +32,7 @@
 Install Torch-TensorRT with the ExecuTorch extra before running this example::
 
     pip install -e ".[executorch]" \
-        --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+        --extra-index-url https://download.pytorch.org/whl/nightly/cu132
 """
 
 import argparse

diff --git a/examples/torchtrt_executorch_example/export_static_shape.py b/examples/torchtrt_executorch_example/export_static_shape.py
@@ -16,7 +16,7 @@
 Install Torch-TensorRT with the ExecuTorch extra before running this example::
 
     pip install -e ".[executorch]" \
-        --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+        --extra-index-url https://download.pytorch.org/whl/nightly/cu132
 
 See https://pytorch.org/executorch/stable/getting-started-setup.html for details.
 """

diff --git a/justfile b/justfile
@@ -90,18 +90,18 @@ summary *args:
 
 # Added without a rebuild via `uv pip install --group` (sidesteps the
 # test-ext↔quantization `uv sync` lockfile conflict). Run before `just lane full`.
-# Install optional test deps for Linux with the project's CUDA 13.0 (cu130) default.
+# Install optional test deps for Linux with the project's CUDA 13.2 (cu132) default.
 install-test-ext:
     uv pip install --group test-ext --group kernels --group quantization
     # ExecuTorch's CUDA wheels are only on the PyTorch nightly index, so the channel is needed.
     # No --pre: a specifier naming a prerelease admits prereleases by itself, and --pre would
-    # apply to pyyaml here too. cu130 matches the torch index this project resolves against by
+    # apply to pyyaml here too. cu132 matches the torch index this project resolves against by
     # default.
     #
     # Exact, not a range: the nightly channel gains a member every day, and the delegate is
     # compiled from the commit this version pairs with.
     uv pip install pyyaml patchelf \
-      --extra-index-url https://download.pytorch.org/whl/nightly/cu130 \
+      --extra-index-url https://download.pytorch.org/whl/nightly/cu132 \
       "executorch==1.6.0.dev20260925"
 
 # ── Linting ───────────────────────────────────────────────────────────────────

diff --git a/py/torch-tensorrt-executorch-runtime/README.md b/py/torch-tensorrt-executorch-runtime/README.md
@@ -109,12 +109,12 @@ and leave the backend unregistered.
 
 The wheel must use the same Python, PyTorch, ExecuTorch, CUDA, TensorRT, and
 C++ ABI as its matching Torch-TensorRT wheel. This delegate requires CUDA 13;
-the build matrix currently covers `cu130`, `cu132` and `cu134`, on both architectures. Ordinary Torch-TensorRT
+the build matrix currently covers `cu132` and `cu134`, on both architectures. Ordinary Torch-TensorRT
 release and JetPack builds retain their separate CUDA 12 support.
 
 ## Install
 One command. The `executorch` extra brings this wheel and a CUDA build of ExecuTorch, and
-Torch-TensorRT brings PyTorch. Swap `cu132` for the CUDA 13 channel you run, `cu130` or `cu134`:
+Torch-TensorRT brings PyTorch. Swap `cu132` for `cu134` if you run CUDA 13.4:
 ```bash
 python -m pip install --pre "torch-tensorrt[executorch]" \
   --index-url https://download.pytorch.org/whl/nightly/cu132 \
@@ -384,7 +384,7 @@ index applies to the whole dependency solve.
 
 ```bash
 python -m pip install dist/torch_tensorrt_executorch_runtime-*.whl \
-  --extra-index-url https://download.pytorch.org/whl/nightly/cu130
+  --extra-index-url https://download.pytorch.org/whl/nightly/cu132
 ```
 
 ```python

diff --git a/pyproject.toml b/pyproject.toml
@@ -217,15 +217,15 @@ constraint-dependencies = [
 
 [tool.uv.sources]
 torch = [
-    { index = "pytorch-nightly-cu130" },
+    { index = "pytorch-nightly-cu132" },
 ]
 torchvision = [
-    { index = "pytorch-nightly-cu130" },
+    { index = "pytorch-nightly-cu132" },
 ]
 
 [[tool.uv.index]]
-name = "pytorch-nightly-cu130"
-url = "https://download.pytorch.org/whl/nightly/cu130"
+name = "pytorch-nightly-cu132"
+url = "https://download.pytorch.org/whl/nightly/cu132"
 explicit = false
 
 [[tool.uv.index]]
@@ -234,8 +234,8 @@ url = "https://download.pytorch.org/whl/nightly/cu128"
 explicit = false
 
 [[tool.uv.index]]
-name = "pytorch-test-cu130"
-url = "https://download.pytorch.org/whl/test/cu130"
+name = "pytorch-test-cu132"
+url = "https://download.pytorch.org/whl/test/cu132"
 explicit = false
 
 [[tool.uv.index]]
@@ -244,8 +244,8 @@ url = "https://download.pytorch.org/whl/test/cu128"
 explicit = false
 
 [[tool.uv.index]]
-name = "pytorch-release-cu130"
-url = "https://download.pytorch.org/whl/release/cu130"
+name = "pytorch-release-cu132"
+url = "https://download.pytorch.org/whl/release/cu132"
 explicit = false
 
 [[tool.uv.index]]

diff --git a/tests/py/dynamo/executorch/test_shared_runtime_workflow.py b/tests/py/dynamo/executorch/test_shared_runtime_workflow.py
@@ -609,9 +609,9 @@ def kept(rows, *flags):
 
     rows = [
         {"desired_cuda": cuda, "python_version": "3.10", "gpu_arch_type": "cuda"}
-        for cuda in ("cu126", "cu130", "cu134")
+        for cuda in ("cu126", "cu132", "cu134")
     ]
-    assert kept(rows, "--executorch-runtime") == {"cu130", "cu134"}
+    assert kept(rows, "--executorch-runtime") == {"cu132", "cu134"}
     # A row the other rules keep, so only the flag can drop it. Every row those rules keep on x86
     # and on Arm is already CUDA 13, so the assertion above held with the flag's branch deleted and
     # could not see it.

diff --git a/tests/py/dynamo/executorch/test_update_executorch_pin.py b/tests/py/dynamo/executorch/test_update_executorch_pin.py
@@ -175,7 +175,7 @@ def test_upper_bound(version, expected):
     assert updater._upper_bound(version) == expected
 
 
-@pytest.mark.parametrize("channel", ["cu126", "cu128", "cu14", "cpu", "cu13"])
+@pytest.mark.parametrize("channel", ["cu126", "cu128", "cu130", "cu14", "cpu", "cu13"])
 @pytest.mark.unit
 def test_main_rejects_unsupported_channels(monkeypatch, channel):
     monkeypatch.setattr(
@@ -186,7 +186,7 @@ def test_main_rejects_unsupported_channels(monkeypatch, channel):
     assert error.value.code == 2
 
 
-@pytest.mark.parametrize("channel", ["cu130", "cu132", "cu134"])
+@pytest.mark.parametrize("channel", ["cu132", "cu134"])
 @pytest.mark.unit
 def test_main_uses_selected_channel(monkeypatch, channel):
     calls = []
@@ -236,8 +236,8 @@ def test_main_refuses_a_version_missing_from_an_accepted_channel(monkeypatch):
 def test_main_adopts_a_version_present_in_every_channel_despite_its_label(monkeypatch):
     """The channels label the same build differently, so comparison must ignore the label.
 
-    This is the case the index actually produces: one version, three spellings. Comparing the
-    raw strings finds no overlap at all and rejects every candidate.
+    This is the case the index actually produces: one version, one spelling per channel. Comparing
+    the raw strings finds no overlap at all and rejects every candidate.
     """
     accepted = updater.delegate_channels()
     per_channel = {ch: [f"1.0.dev1+{ch}", f"1.0.dev2+{ch}"] for ch in accepted}