Drop CUDA 12.6 support

.ci/docker/common/install_cuda_windows_cross_compile.sh

@@ -14,7 +14,6 @@ INSTALL_DIR="${WINDOWS_CUDA_INSTALL_DIR:-/opt/cuda-windows}"
 # Mapping of CUDA versions to their corresponding driver versions for Windows installers
 # Source: https://developer.nvidia.com/cuda-toolkit-archive
 declare -A CUDA_DRIVER_MAP=(
-    ["12.6"]="12.6.3:561.17"
     ["12.8"]="12.8.1:572.61"
     ["12.9"]="12.9.1:576.57"
     ["13.0"]="13.0.2:"

.ci/scripts/cuda_benchmark.py

@@ -588,7 +588,7 @@ def to_v3_format(
             workflow_run_id: GitHub workflow run ID
             workflow_run_url: GitHub workflow run URL
             gpu_name: GPU device name (e.g., "Tesla V100", "A100")
-            cuda_driver_version: CUDA driver version (e.g., "12.6", "535.104.05")
+            cuda_driver_version: CUDA driver version (e.g., "13.0", "535.104.05")
 
         Returns:
             List of benchmark records in v3 format
@@ -873,7 +873,7 @@ def main():
         "--cuda_driver_version",
         type=str,
         default=None,
-        help="CUDA driver version (e.g., '12.6', '535.104.05') - optional for v3 format",
+        help="CUDA driver version (e.g., '13.0', '535.104.05') - optional for v3 format",
     )
     parser.add_argument("--verbose", action="store_true", help="Print verbose output")
 

.ci/scripts/tests/test_cu134_dependencies.py

@@ -103,7 +103,7 @@ def test_all_install_steps_preserve_exact_cu134_selection(self):
                         )
 
     def test_other_cuda_trains_keep_existing_pins(self):
-        for cuda in ((12, 6), (13, 0), (13, 2)):
+        for cuda in ((13, 0), (13, 2)):
             for machine in ("x86_64", "aarch64"):
                 with self.subTest(cuda=cuda, machine=machine):
                     core, local, domains, examples = self.install_commands(
@@ -159,7 +159,7 @@ def test_cpu_release_dependencies_are_opt_in(self):
                     )
 
     def test_cpu_release_option_keeps_cuda_pins(self):
-        for cuda in ((12, 6), (13, 0), (13, 2), (13, 4)):
+        for cuda in ((13, 0), (13, 2), (13, 4)):
             for machine in ("x86_64", "aarch64"):
                 with self.subTest(cuda=cuda, machine=machine):
                     self.assertEqual(

.ci/scripts/tests/test_cuda_workflow.py

@@ -34,9 +34,7 @@ def _model_quant(entry):
 class CudaWorkflowTest(unittest.TestCase):
     def test_build_matrix_preserves_existing_cuda_versions(self):
         job = WORKFLOW["jobs"]["test-cuda-builds"]
-        self.assertEqual(
-            job["strategy"]["matrix"]["cuda-version"], ["12.6", "13.0", "13.4"]
-        )
+        self.assertEqual(job["strategy"]["matrix"]["cuda-version"], ["13.0", "13.4"])
         self.assertEqual(job["with"]["gpu-arch-version"], "${{ matrix.cuda-version }}")
 
     def test_cuda_builds_take_the_node_driver_on_an_unpinned_v3(self):
@@ -58,9 +56,7 @@ def test_cuda_runtime_update_precedes_build_and_propagates_failure(self):
 source() { printf 'BUILD %s\n' "$*"; printf 'LIBRARY PATH %s\n' "$LD_LIBRARY_PATH"; }
 """
         for version, conda_status in (
-            (version, status)
-            for version in ("12.6", "13.0", "13.4")
-            for status in (0, 1)
+            (version, status) for version in ("13.0", "13.4") for status in (0, 1)
         ):
             with self.subTest(version=version, conda_status=conda_status):
                 result = subprocess.run(

.ci/scripts/wheel/cuda_arch_list.sh

@@ -41,29 +41,6 @@ _cuda_arch_aarch64_cu130="8.0 9.0 10.0 11.0 12.0"
 _cuda_arch_aarch64_cu132="${_cuda_arch_aarch64_cu130}"
 _cuda_arch_aarch64_cu134="${_cuda_arch_aarch64_cu130}"
 
-# The older CUDA train.
-#
-# The two architectures do not carry identical lists, because each covers what the published PyTorch
-# build for that architecture covers, and those differ. Matching them to each other instead would mean
-# advertising a GPU on one architecture that PyTorch cannot serve there.
-#
-# The smaller embedded modules are deliberately absent, with one exception. An embedded-only
-# architecture in a generic wheel would advertise a device the row cannot otherwise serve, since
-# those devices also need the CUDA, TensorRT and PyTorch pinned by their own software release
-# rather than the ones a generic wheel resolves.
-#
-# 8.7 is that exception. This is the only row whose CUDA major matches what that module's software
-# release ships, and the wheel declares no PyTorch, so the user supplies the build that carries
-# their architecture. Omitting it does not protect them from a bad pairing, it only removes the
-# device code they need.
-#
-# The floor is 8.0 rather than the oldest architecture PyTorch still carries. One of these sources compiles
-# an integer matrix-multiply path only at 8.0 and newer, so an older architecture would get a delegate that
-# loads, runs most models, and fails on one that needs that operator. Claiming hardware the delegate only
-# partly serves is the same problem the embedded modules have, so the row leaves it out for the same reason.
-_cuda_arch_x86_64_cu126="8.0 8.6 8.9 9.0"
-_cuda_arch_aarch64_cu126="8.0 8.7 9.0"
-
 # A CUDA train with no architecture list would leave the build detecting the builder's GPU, which is
 # the failure this file exists to prevent. Adding a train to the release matrix without adding its
 # architectures should fail loudly rather than silently produce a single-GPU wheel.
@@ -103,7 +80,6 @@ executorch_cuda_arch_list() {
   case "${machine}" in
     aarch64 | arm64)
       case "${train}" in
-        126) printf '%s' "${_cuda_arch_aarch64_cu126}" ;;
         130) printf '%s' "${_cuda_arch_aarch64_cu130}" ;;
         132) printf '%s' "${_cuda_arch_aarch64_cu132}" ;;
         134) printf '%s' "${_cuda_arch_aarch64_cu134}" ;;
@@ -112,7 +88,6 @@ executorch_cuda_arch_list() {
       ;;
     x86_64)
       case "${train}" in
-        126) printf '%s' "${_cuda_arch_x86_64_cu126}" ;;
         130) printf '%s' "${_cuda_arch_x86_64_cu130}" ;;
         132) printf '%s' "${_cuda_arch_x86_64_cu132}" ;;
         134) printf '%s' "${_cuda_arch_x86_64_cu134}" ;;

.ci/scripts/wheel/test_shared_libraries.py

@@ -1217,7 +1217,7 @@ def _wheel_cuda_train() -> str:
     """The CUDA train the installed wheel was built for, or "" for a CPU wheel.
 
     Read from the local version segment, which is the only place the wheel states what
-    it was built for. `1.5.0+cu126` gives "126".
+    it was built for. `1.5.0+cu132` gives "132".
     """
     local = importlib.metadata.version("executorch").partition("+")[2]
     return local[2:] if local.startswith("cu") else ""
@@ -1360,7 +1360,7 @@ def _resolve_required(required):
     ),
     # The CUDA components. Required exactly when the wheel says it is a CUDA wheel,
     # which is decided at check time rather than here: a fixed False meant a wheel
-    # tagged +cu126 carrying no CUDA library at all passed every check in this file,
+    # tagged +cu132 carrying no CUDA library at all passed every check in this file,
     # while a fixed True would fail every CPU wheel. The marker is the string these
     # rows are keyed on below.
     (
@@ -1457,7 +1457,7 @@ def test_each_component_has_one_owner() -> None:
     shipped = {path.name for path in _shipped_shared_objects(_installed_package_dir())}
     for what, symbols, owner, required in _OWNED_COMPONENTS:
         # Resolved here rather than in the table, because it depends on the installed
-        # wheel. A fixed False let a wheel tagged +cu126 ship with no CUDA library at
+        # wheel. A fixed False let a wheel tagged +cu132 ship with no CUDA library at
         # all and still pass, which is the whole point of the conditional rows.
         required = _resolve_required(required)
         present = any(name.startswith(owner) for name in shipped)

.github/scripts/filter_cuda_matrix.py

@@ -53,12 +53,6 @@
 # Skip wholly absent trains so an upstream removal cannot block the remaining releases.
 # Offered trains must still cover every supported Python version.
 #
-# cu126 was the floor until PyTorch stopped offering it on the nightly channel. The generator
-# still offers it when a release is cut, so keeping it here would publish a train that no
-# nightly has built since, and a release is the wrong place to find out that it no longer
-# builds. A machine on CUDA 12.6 can still build from source, where the pinned torch comes
-# from a channel that carries 12.6.
-#
 # cu130 is not published. PyTorch keeps it on its nightly builds only as a temporary hold for
 # other projects, and its release-candidate builds do not carry it, so a release would have
 # no cu130 train even while nightlies do.
@@ -91,9 +85,8 @@
 # version. Kept empty on purpose today, so no Jetson row is emitted.
 #
 # Published PyTorch stopped shipping sm_87 device code after 2.8.0, so a Jetson row today
-# would produce a wheel whose PyTorch dependency cannot execute on the device. The aarch64
-# CUDA 12.6 row, which was the one that compiled sm_87 device code for an embedded module,
-# is no longer published either. Populate this when both change.
+# would produce a wheel whose PyTorch dependency cannot execute on the device, and no
+# published row compiles sm_87 device code. Populate this when both change.
 #
 # Because both lists are empty, asking for the JetPack rows can only produce an empty result.
 # No workflow asks, and the request is rejected up front with that reason rather than left to

.github/workflows/cuda.yml

@@ -1,6 +1,6 @@
 # Test ExecuTorch CUDA Build Compatibility
 # This workflow tests whether ExecuTorch can be successfully built with CUDA support
-# across different CUDA versions (12.6, 13.0, 13.4) using the command:
+# across different CUDA versions (13.0, 13.4) using the command:
 #   ./install_executorch.sh
 #
 # It also runs model accuracy (e2e EXPECTED_OUTPUT checks) and, for the perf
@@ -142,7 +142,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        cuda-version: ["12.6", "13.0", "13.4"]
+        cuda-version: ["13.0", "13.4"]
 
     name: test-executorch-cuda-build-${{ matrix.cuda-version }}
     uses: pytorch/test-infra/.github/workflows/linux_job_v3.yml@main
@@ -159,8 +159,8 @@ jobs:
       script: |
         set -eux
 
-        # AOTI model libraries need a newer C++ runtime than the 12.6 image's
-        # system library; the 13.4 image also needs the thread-local cleanup fix.
+        # AOTI model libraries need a newer C++ runtime than the image's system
+        # library; the 13.4 image also needs the thread-local cleanup fix.
         conda install -y -n base -c conda-forge 'libstdcxx-ng=16.2.0' 'libgcc-ng=16.2.0'
         export LD_LIBRARY_PATH="/opt/conda/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
         source .ci/scripts/test-cuda-build.sh "${{ matrix.cuda-version }}"

.github/workflows/vulkan.yml

@@ -79,7 +79,7 @@ jobs:
       timeout: 120
       runner: mt-l-x86aavx2-11-41-a10g
       gpu-arch-type: cuda
-      gpu-arch-version: "12.6"
+      gpu-arch-version: "13.2"
       submodules: recursive
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       script: |
@@ -171,7 +171,7 @@ jobs:
       timeout: 180
       runner: mt-l-x86aavx2-11-41-a10g
       gpu-arch-type: cuda
-      gpu-arch-version: "12.6"
+      gpu-arch-version: "13.2"
       submodules: recursive
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       upload-artifact: vulkan-llm-e2e

docs/source/backends/cuda/cuda-overview.md

@@ -17,7 +17,7 @@ Below are the requirements for running a CUDA-delegated ExecuTorch model:
 - **Hardware**: NVIDIA GPU with compute capability 8.0 (Ampere) or newer. This is the floor the
   prebuilt packages are built for.
 - **CUDA**: 13 for the prebuilt packages. On Linux they bring the CUDA runtime with them. On
-  Windows it comes from a CUDA 13 toolkit install. A build from source also supports CUDA 12.6.
+  Windows it comes from a CUDA 13 toolkit install. A build from source also supports CUDA 13.0.
 - **Operating System**: Linux, or Windows from C++ for programs exported on Linux (see
   [CUDA on Windows](../../using-executorch-cpp.md#cuda-on-windows))
 - **Drivers**: An NVIDIA driver that supports the CUDA version you use: CUDA 13 for the prebuilt

install_utils.py

@@ -18,7 +18,6 @@
 # Supported CUDA versions - modify this to add/remove supported versions
 # Format: tuple of (major, minor) version numbers
 SUPPORTED_CUDA_VERSIONS = (
-    (12, 6),
     (13, 0),
     (13, 2),
     (13, 4),
@@ -114,7 +113,7 @@ def _get_cuda_version():
             _selected_nvcc(), capture_output=True, text=True, check=True
         )
         # Parse nvcc output for CUDA version
-        # Output contains line like "Cuda compilation tools, release 12.6, V12.6.68"
+        # Output contains line like "Cuda compilation tools, release 13.0, V13.0.88"
         match = re.search(r"release (\d+)\.(\d+)", nvcc_result.stdout)
         if match:
             major, minor = int(match.group(1)), int(match.group(2))
@@ -301,11 +300,11 @@ def _cuda_version_to_pytorch_suffix(major, minor):
     Generate PyTorch CUDA wheel suffix from CUDA version numbers.
 
     Args:
-        major: CUDA major version (e.g., 12)
-        minor: CUDA minor version (e.g., 6)
+        major: CUDA major version (e.g., 13)
+        minor: CUDA minor version (e.g., 2)
 
     Returns:
-        PyTorch wheel suffix string (e.g., "cu126")
+        PyTorch wheel suffix string (e.g., "cu132")
     """
     return f"cu{major}{minor}"
 

setup.py

@@ -735,9 +735,9 @@ def _cuda_train() -> str:
 
     Raises when a release row names a train the installed toolkit does not provide. The
     declared packages and the loader paths both come from this value, so disagreeing with
-    the toolkit that compiled the libraries produces a wheel that installs cleanly and then
-    cannot load: a cu126 row built against a 13.0 toolkit declares the CUDA 12 runtime for
-    binaries that need libcudart.so.13.
+    the toolkit that compiled the libraries produces a wheel that installs cleanly but
+    declares a train it was not compiled for, such as a cu134 row built against a 13.2
+    toolkit.
     """
     # An explicit OFF first, ahead of the release field. A CPU row on a builder that has a
     # toolkit installed sets both, so reading the row field first would declare a runtime