Adjust tol

.ci/docker/common/install_conda.sh

@@ -92,6 +92,9 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
   # Install some other packages, including those needed for Python test reporting
   pip_install -r /opt/conda/requirements-ci.txt
 
+  # Installed spmd-types with --no-deps to avoid pulling torch dependency at this point
+  pip_install --no-deps spmd-types==0.2.1
+
   if [ -n "$DOCS" ]; then
     apt-get update
     apt-get -y install expect-dev

.ci/libtorch/extract_libtorch_from_wheel.py

@@ -234,16 +234,23 @@ def create_libtorch_zip(
     return zip_path
 
 
-def compute_zip_prefix(platform: str, desired_cuda: str, libtorch_variant: str) -> str:
+def compute_zip_prefix(
+    platform: str, desired_cuda: str, libtorch_variant: str, arch: str
+) -> str:
     """Compute the zip filename prefix matching existing naming conventions.
 
     Linux:  libtorch-shared-with-deps
     macOS:  libtorch-macos-arm64
     Windows: libtorch-win-shared-with-deps (or libtorch-win-arm64-shared-with-deps)
+
+    The arch component keeps the Windows x86_64 and arm64 packages from
+    sharing a filename and overwriting each other in the upload bucket.
     """
     if platform == "macos":
         return "libtorch-macos-arm64"
     elif platform == "windows":
+        if arch == "arm64":
+            return f"libtorch-win-arm64-{libtorch_variant}"
         return f"libtorch-win-{libtorch_variant}"
     else:
         return f"libtorch-{libtorch_variant}"
@@ -273,6 +280,12 @@ def main() -> None:
         default="shared-with-deps",
         help="Libtorch variant (shared-with-deps, etc.)",
     )
+    parser.add_argument(
+        "--arch",
+        default="x86_64",
+        choices=["x86_64", "arm64"],
+        help="Target architecture (used to disambiguate Windows package names)",
+    )
     parser.add_argument(
         "--git-hash",
         default="",
@@ -316,7 +329,7 @@ def main() -> None:
 
         # Compute zip prefix
         zip_prefix = compute_zip_prefix(
-            args.platform, args.desired_cuda, args.libtorch_variant
+            args.platform, args.desired_cuda, args.libtorch_variant, args.arch
         )
 
         # Split debug symbols on Linux

.ci/manywheel/build_env_setup.py

@@ -74,6 +74,10 @@
         "x86_64": {50, 60, 70, 75, 80, 86, 90},
         "aarch64": {80, 90},
     },
+    "12.9": {
+        "x86_64": {75, 80, 86, 90, 100, 120},
+        "aarch64": {80, 90, 100, 120},
+    },
     "13.0": {
         "x86_64": {75, 80, 86, 90, 100, 120},
         "aarch64": {80, 90, 100, 110, 120},
@@ -84,14 +88,35 @@
     },
 }
 
-# Architectures we additionally emit PTX for (forward-compat for newer GPUs).
+# Architectures we additionally emit PTX for on nightly/dev builds
+# (forward-compat for newer GPUs). Release/RC wheels ship SASS-only to keep
+# libtorch_cuda.so size down; see _ptx_arches().
 _PTX_ARCHES: set[int] = {120}
 
 
+def _is_release_build() -> bool:
+    """True for release / RC binary builds (vs nightly / dev builds).
+
+    Binary builds off a git tag (releases and RCs, e.g. v2.13.0 / v2.13.0-rc1)
+    get a ``PYTORCH_BUILD_VERSION`` with no ``.dev<date>`` suffix, while
+    nightlies do -- see .ci/pytorch/binary_populate_env.sh, which relies on the
+    same ``dev`` check for Triton pinning. A missing/empty version (local or
+    non-binary builds) is treated as non-release, so PTX is kept.
+    """
+    version = os.environ.get("PYTORCH_BUILD_VERSION", "")
+    return bool(version) and "dev" not in version
+
+
+def _ptx_arches() -> set[int]:
+    """PTX arches for this build: empty on release/RC, forward-compat set otherwise."""
+    return set() if _is_release_build() else _PTX_ARCHES
+
+
 def torch_cuda_arch_list(cuda_version: str, arch: str) -> str:
     """Format TORCH_CUDA_ARCH_LIST for the wheel build (";"-separated).
 
-    Returns e.g. "8.0;9.0;10.0;11.0;12.0+PTX" for cuda 13.x aarch64.
+    Returns e.g. "8.0;9.0;10.0;11.0;12.0+PTX" for cuda 13.x aarch64 on
+    nightly builds; release/RC builds omit the +PTX suffix (see _ptx_arches()).
 
     CUDA 13.x dropped sm_50/60/70, so we must NOT leave this empty --
     CMake's defaults still include compute_50 which nvcc 13 rejects with
@@ -102,8 +127,9 @@ def torch_cuda_arch_list(cuda_version: str, arch: str) -> str:
     archs = TORCH_CUDA_ARCH_LIST_TABLE[cuda_version].get(arch)
     if not archs:
         raise SystemExit(f"no TORCH_CUDA_ARCH_LIST for cuda {cuda_version} on {arch}")
+    ptx_arches = _ptx_arches()
     return ";".join(
-        f"{cc // 10}.{cc % 10}" + ("+PTX" if cc in _PTX_ARCHES else "")
+        f"{cc // 10}.{cc % 10}" + ("+PTX" if cc in ptx_arches else "")
         for cc in sorted(archs)
     )
 

.ci/pytorch/binary_populate_env.sh

@@ -112,15 +112,6 @@ if [[ "$PACKAGE_TYPE" =~ .*wheel.* && -n "$PYTORCH_BUILD_VERSION" && "$PYTORCH_B
     fi
 fi
 
-if [[ "$PACKAGE_TYPE" =~ .*wheel.* ]]; then
-  SPMD_TYPES_REQUIREMENT="spmd-types==0.2.1"
-  if [[ -z "${PYTORCH_EXTRA_INSTALL_REQUIREMENTS:-}" ]]; then
-    export PYTORCH_EXTRA_INSTALL_REQUIREMENTS="${SPMD_TYPES_REQUIREMENT}"
-  else
-    export PYTORCH_EXTRA_INSTALL_REQUIREMENTS="${PYTORCH_EXTRA_INSTALL_REQUIREMENTS} | ${SPMD_TYPES_REQUIREMENT}"
-  fi
-fi
-
 USE_GLOO_WITH_OPENSSL="OFF"
 if [[ "$GPU_ARCH_TYPE" =~ .*aarch64.* ]]; then
   USE_GOLD_LINKER="OFF"

.ci/pytorch/build.sh

@@ -279,6 +279,20 @@ if [[ "$BUILD_ENVIRONMENT" != *libtorch* ]]; then
     python -m build --wheel --no-isolation
   fi
   pip_install_whl "$(echo dist/*.whl)"
+
+  # Smoke-test tools/build_with_debinfo.py against the real build tree: it must
+  # still emit a debug-rebuild plan with a -g compile and the libtorch_python
+  # relink. This guards against build-system changes (e.g. a new
+  # CONFIGURE_DEPENDS glob scheme) silently breaking the tool, which only works
+  # on a from-source build that test jobs don't have. --dry-run reads the tree
+  # without rebuilding, so it leaves the checkout clean (assert_git_not_dirty).
+  if [[ -f build/compile_commands.json ]] && command -v ninja > /dev/null && grep -q "csrc/Module.cpp" build/compile_commands.json; then
+    debinfo_plan="$(python tools/build_with_debinfo.py --dry-run torch/csrc/Module.cpp)"
+    echo "${debinfo_plan}"
+    grep -qE ' -g( |$)' <<< "$debinfo_plan" || { echo "ERROR: build_with_debinfo --dry-run emitted no -g debug compile flag"; exit 1; }
+    grep -q 'libtorch_python' <<< "$debinfo_plan" || { echo "ERROR: build_with_debinfo --dry-run emitted no libtorch_python link command"; exit 1; }
+  fi
+
   if [[ "$BUILD_ENVIRONMENT" == *full-debug* ]]; then
     # Regression test for https://github.com/pytorch/pytorch/issues/164297
     # Torch should be importable and that's about it

.ci/pytorch/common_utils.sh

@@ -279,7 +279,7 @@ function install_torchrec_and_fbgemm() {
 
 function clone_pytorch_xla() {
   if [[ ! -d ./xla ]]; then
-    git clone --recursive --quiet https://github.com/pytorch/xla.git
+    git clone --recursive -b r2.13 https://github.com/pytorch/xla.git
     pushd xla
     # pin the xla hash so that we don't get broken by changes to xla
     git checkout "$(cat ../.github/ci_commit_pins/xla.txt)"

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-41398bfff334fc8d3b1c00be6ea8cc5411f6d6bf
+r2.13

.github/scripts/filter_test_configs.py

@@ -39,9 +39,9 @@
 }
 
 # The link to the published list of disabled jobs
-DISABLED_JOBS_URL = "https://ossci-metrics.s3.amazonaws.com/disabled-jobs.json"
+DISABLED_JOBS_URL = "https://ossci-metrics.s3.amazonaws.com/disabled-jobs.json?versionId=Oekxne_7ooz8C5nmynVmrc5UzuZ875Tr"
 # and unstable jobs
-UNSTABLE_JOBS_URL = "https://ossci-metrics.s3.amazonaws.com/unstable-jobs.json"
+UNSTABLE_JOBS_URL = "https://ossci-metrics.s3.amazonaws.com/unstable-jobs.json?versionId=V37EHozs4nE3_iE8GfJ8sctdQzmU9zPH"
 
 # Some constants used to handle disabled and unstable jobs
 JOB_NAME_SEP = "/"

.github/scripts/generate_binary_build_matrix.py

@@ -24,15 +24,17 @@
 REPO_ROOT = SCRIPT_DIR.parent.parent
 
 
-CUDA_ARCHES = ["12.6", "13.0", "13.2"]
+CUDA_ARCHES = ["12.6", "12.9", "13.0", "13.2"]
 CUDA_STABLE = "13.0"
 CUDA_ARCHES_FULL_VERSION = {
     "12.6": "12.6.3",
-    "13.0": "13.0.2",
+    "12.9": "12.9.1",
+    "13.0": "13.0.3",
     "13.2": "13.2.1",
 }
 CUDA_ARCHES_CUDNN_VERSION = {
     "12.6": "9",
+    "12.9": "9",
     "13.0": "9",
     "13.2": "9",
 }
@@ -47,14 +49,11 @@
 
 CUDA_AARCH64_ARCHES = [
     "12.6-aarch64",
+    "12.9-aarch64",
     "13.0-aarch64",
     "13.2-aarch64",
 ]
 
-
-# WARNING: For CUDA 13.0, cublas is pinned to a version range rather
-# than an exact version. A broken cublas release within that range will be
-# silently pulled in.
 PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
     "12.6": (
         "cuda-toolkit[nvrtc,cudart,cupti,cufft,curand,cusolver,cusparse,cublas,cufile,nvjitlink,nvtx]==12.6.3; platform_system == 'Linux' | "
@@ -64,9 +63,16 @@
         "nvidia-nccl-cu12==2.29.3; platform_system == 'Linux' | "
         "nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux'"
     ),
+    "12.9": (
+        "cuda-toolkit[nvrtc,cudart,cupti,cufft,curand,cusolver,cusparse,cublas,cufile,nvjitlink,nvtx]==12.9.1; platform_system == 'Linux' | "
+        "cuda-bindings>=12.9.4,<13; platform_system == 'Linux' and python_version < '3.15' | "
+        "nvidia-cudnn-cu12==9.20.0.48; platform_system == 'Linux' | "
+        "nvidia-cusparselt-cu12==0.8.1; platform_system == 'Linux' | "
+        "nvidia-nccl-cu12==2.29.7; platform_system == 'Linux' | "
+        "nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux'"
+    ),
     "13.0": (
-        "cuda-toolkit[nvrtc,cudart,cupti,cufft,curand,cusolver,cusparse,cufile,nvjitlink,nvtx]==13.0.2; platform_system == 'Linux' | "
-        "nvidia-cublas>=13.1.0.3,<=13.1.1.3; platform_system == 'Linux' | "
+        "cuda-toolkit[nvrtc,cudart,cupti,cufft,curand,cusolver,cusparse,cublas,cufile,nvjitlink,nvtx]==13.0.3; platform_system == 'Linux' | "
         "cuda-bindings>=13.0.3,<14; platform_system == 'Linux' and python_version < '3.15' | "
         "nvidia-cudnn-cu13==9.20.0.48; platform_system == 'Linux' | "
         "nvidia-cusparselt-cu13==0.8.1; platform_system == 'Linux' | "
@@ -304,6 +310,24 @@ def arch_type(arch_version: str) -> str:
     "cpu-s390x": "manylinuxs390x-builder:cpu-s390x",
 }
 
+# RELEASE-ONLY: pin the manywheel builder images to a fixed build so the release
+# uses a reproducible toolchain instead of main's floating tags. The suffix is
+# the .ci/docker tree hash (`git rev-parse HEAD:.ci/docker`), i.e. the same tag
+# .github/actions/binary-docker-build publishes. Only linux manywheel builds run
+# inside these containers, so only those images are pinned. s390x is excluded:
+# its builder images are built locally on self-hosted runners and never published
+# to docker.io under the pinned tag, so pinning it breaks the image pull.
+DOCKER_IMAGE_PIN = "78e737ad29420ffc4800e677c51e2a852caf8359"
+MANYWHEEL_OSES = ("linux", "linux-aarch64")
+
+
+def wheel_container_image_tag_prefix(arch_version: str, os: str) -> str:
+    tag_prefix = WHEEL_CONTAINER_IMAGES[arch_version].split(":")[1]
+    if os in MANYWHEEL_OSES:
+        return f"{tag_prefix}-{DOCKER_IMAGE_PIN}"
+    return tag_prefix
+
+
 RELEASE = "release"
 DEBUG = "debug"
 
@@ -344,7 +368,9 @@ def generate_libtorch_matrix(
     if arches is None:
         arches = ["cpu"]
         if os == "windows":
-            arches += CUDA_ARCHES
+            # CUDA 12.9 is only built for Linux
+            windows_cuda_arches = list_without(CUDA_ARCHES, ["12.9"])
+            arches += windows_cuda_arches
     if libtorch_variants is None:
         libtorch_variants = [
             "shared-with-deps",
@@ -397,7 +423,9 @@ def generate_wheels_matrix(
         if os == "linux":
             arches += CUDA_ARCHES + ROCM_ARCHES + XPU_ARCHES
         elif os == "windows":
-            arches += CUDA_ARCHES + XPU_ARCHES
+            # CUDA 12.9 is only built for Linux
+            windows_cuda_arches = list_without(CUDA_ARCHES, ["12.9"])
+            arches += windows_cuda_arches + XPU_ARCHES
         elif os == "linux-aarch64":
             # Separate new if as the CPU type is different and
             # uses different build/test scripts
@@ -439,7 +467,7 @@ def generate_wheels_matrix(
             # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
 
             if (
-                arch_version in ["13.2", "13.0", "12.6"]
+                arch_version in ["13.2", "13.0", "12.9", "12.6"]
                 and os == "linux"
                 or arch_version in CUDA_AARCH64_ARCHES
             ):
@@ -453,9 +481,9 @@ def generate_wheels_matrix(
                         "container_image": WHEEL_CONTAINER_IMAGES[arch_version].split(
                             ":"
                         )[0],
-                        "container_image_tag_prefix": WHEEL_CONTAINER_IMAGES[
-                            arch_version
-                        ].split(":")[1],
+                        "container_image_tag_prefix": wheel_container_image_tag_prefix(
+                            arch_version, os
+                        ),
                         "package_type": package_type,
                         "pytorch_extra_install_requirements": (
                             PYTORCH_EXTRA_INSTALL_REQUIREMENTS[
@@ -484,9 +512,9 @@ def generate_wheels_matrix(
                         "container_image": WHEEL_CONTAINER_IMAGES[arch_version].split(
                             ":"
                         )[0],
-                        "container_image_tag_prefix": WHEEL_CONTAINER_IMAGES[
-                            arch_version
-                        ].split(":")[1],
+                        "container_image_tag_prefix": wheel_container_image_tag_prefix(
+                            arch_version, os
+                        ),
                         "package_type": package_type,
                         "build_name": f"{package_type}-py{python_version}-{gpu_arch_type}{gpu_arch_version}".replace(
                             ".", "_"
@@ -516,6 +544,7 @@ def generate_libtorch_extraction_configs(
     uses to add an extraction job that depends on that wheel's build job.
     """
     preferred_python = "3.11" if os == "windows-arm64" else "3.10"
+    arch = "arm64" if os == "windows-arm64" else "x86_64"
 
     # Group wheel configs by (gpu_arch_type, gpu_arch_version)
     arch_to_config: dict[tuple[str, str], dict[str, str]] = {}
@@ -532,7 +561,10 @@ def generate_libtorch_extraction_configs(
 
         desired_cuda = source_config["desired_cuda"]
         libtorch_variant = "shared-with-deps"
-        build_name = f"libtorch-{gpu_arch_type}{gpu_arch_version}-{libtorch_variant}-release".replace(
+        # Include arch in the build name so windows x86_64 and arm64 libtorch
+        # packages don't share a name and overwrite each other on upload.
+        arch_tag = f"{arch}-" if os == "windows-arm64" else ""
+        build_name = f"libtorch-{arch_tag}{gpu_arch_type}{gpu_arch_version}-{libtorch_variant}-release".replace(
             ".", "_"
         )
 
@@ -546,6 +578,7 @@ def generate_libtorch_extraction_configs(
                 "desired_cuda": desired_cuda,
                 "gpu_arch_type": gpu_arch_type,
                 "gpu_arch_version": gpu_arch_version,
+                "arch": arch,
             }
         )
 

.github/templates/common.yml.j2

@@ -3,7 +3,7 @@
 {%- set upload_artifact_action = "actions/upload-artifact@v4.4.0" -%}
 {%- set download_artifact_action = "actions/download-artifact@v4.1.7" -%}
 
-{%- set timeout_minutes = 280 -%}
+{%- set timeout_minutes = 400 -%}
 {%- set timeout_minutes_windows_binary = 360 -%}
 
 {%- macro concurrency(build_environment) -%}
@@ -32,7 +32,7 @@ concurrency:
 {%- macro setup_ec2_windows() -%}
       !{{ display_ec2_information() }}
       - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
-        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        uses: pytorch/test-infra/.github/actions/setup-ssh@release/2.13
         continue-on-error: true
         with:
           github-secret: ${{ secrets.GITHUB_TOKEN }}

.github/templates/linux_binary_build_workflow.yml.j2

@@ -113,7 +113,7 @@ jobs:
   get-label-type:
     if: github.repository_owner == 'pytorch'
     name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
+    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@release/2.13
     with:
       triggering_actor: ${{ github.triggering_actor }}
       issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
@@ -173,7 +173,10 @@ jobs:
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
         with:
           ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-          fetch-depth: 2
+          # Full fetch on tag pushes so the release tag is visible to
+          # `git describe --tags --exact`; nightlies stay shallow for speed.
+          fetch-depth: ${{ github.ref_type != 'tag' && 2 || 0 }}
+          fetch-tags: ${{ github.ref_type == 'tag' }}
           submodules: recursive
           show-progress: false
       - name: Populate binary env