Update

[ghstack-poisoned]
Update (base update)
2025-11-02 14:34:54 +08:00 · 2025-10-31 14:16:01 +08:00 · 2025-10-31 14:16:01 +08:00 · 2025-10-31 00:49:09 +00:00 · 2025-10-31 00:48:01 +00:00 · 2025-10-30 23:59:27 +00:00
749 changed files with 16544 additions and 6914 deletions
--- a/.ci/docker/common/install_conda.sh
+++ b/.ci/docker/common/install_conda.sh
@ -49,12 +49,20 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
    export SYSROOT_DEP="sysroot_linux-64=2.17"
  fi

+# Install correct Python version
+# Also ensure sysroot is using a modern GLIBC to match system compilers
+if [ "$ANACONDA_PYTHON_VERSION" = "3.14" ]; then
+  as_jenkins conda create -n py_$ANACONDA_PYTHON_VERSION -y\
+             python="3.14.0" \
+             ${SYSROOT_DEP} \
+             -c conda-forge
+else
  # Install correct Python version
  # Also ensure sysroot is using a modern GLIBC to match system compilers
  as_jenkins conda create -n py_$ANACONDA_PYTHON_VERSION -y\
             python="$ANACONDA_PYTHON_VERSION" \
             ${SYSROOT_DEP}
-
+fi
  # libstdcxx from conda default channels are too old, we need GLIBCXX_3.4.30
  # which is provided in libstdcxx 12 and up.
  conda_install libstdcxx-ng=12.3.0 --update-deps -c conda-forge
--- a/.ci/docker/common/install_cuda.sh
+++ b/.ci/docker/common/install_cuda.sh
@ -10,7 +10,7 @@ else
  arch_path='sbsa'
 fi

-NVSHMEM_VERSION=3.3.24
+NVSHMEM_VERSION=3.4.5

 function install_cuda {
  version=$1
--- a/.ci/docker/common/install_rocm.sh
+++ b/.ci/docker/common/install_rocm.sh
@ -40,11 +40,7 @@ EOF

    # Default url values
    rocm_baseurl="http://repo.radeon.com/rocm/apt/${ROCM_VERSION}"
-    amdgpu_baseurl="https://repo.radeon.com/amdgpu/${ROCM_VERSION}/ubuntu"
-
-    # Add amdgpu repository
    UBUNTU_VERSION_NAME=`cat /etc/os-release | grep UBUNTU_CODENAME | awk -F= '{print $2}'`
-    echo "deb [arch=amd64] ${amdgpu_baseurl} ${UBUNTU_VERSION_NAME} main" > /etc/apt/sources.list.d/amdgpu.list

    # Add rocm repository
    wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add -
--- a/.ci/docker/requirements-ci.txt
+++ b/.ci/docker/requirements-ci.txt
@ -138,10 +138,12 @@ numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
 #test_binary_ufuncs.py
 numpy==1.22.4; python_version == "3.10"
 numpy==1.26.2; python_version == "3.11" or python_version == "3.12"
-numpy==2.1.2; python_version >= "3.13"
+numpy==2.1.2; python_version >= "3.13" and python_version < "3.14"
+numpy==2.3.4; python_version >= "3.14"

 pandas==2.0.3; python_version < "3.13"
-pandas==2.2.3; python_version >= "3.13"
+pandas==2.2.3; python_version >= "3.13" and python_version < "3.14"
+pandas==2.3.3; python_version >= "3.14"

 #onnxruntime
 #Description: scoring engine for Open Neural Network Exchange (ONNX) models
@ -153,7 +155,8 @@ opt-einsum==3.3
 #Pinned versions: 3.3
 #test that import: test_linalg.py

-optree==0.13.0
+optree==0.13.0 ; python_version < "3.14"
+optree==0.17.0 ; python_version >= "3.14"
 #Description: A library for tree manipulation
 #Pinned versions: 0.13.0
 #test that import: test_vmap.py, test_aotdispatch.py, test_dynamic_shapes.py,
@ -252,7 +255,8 @@ scikit-image==0.22.0
 #test that import:

 scipy==1.10.1 ; python_version <= "3.11"
-scipy==1.14.1 ; python_version >= "3.12"
+scipy==1.14.1 ; python_version > "3.11" and python_version < "3.14"
+scipy==1.16.2 ; python_version >= "3.14"
 # Pin SciPy because of failing distribution tests (see #60347)
 #Description: scientific python
 #Pinned versions: 1.10.1
@ -324,7 +328,8 @@ pywavelets==1.7.0 ; python_version >= "3.12"
 #Pinned versions: 1.4.1
 #test that import:

-lxml==5.3.0
+lxml==5.3.0 ; python_version < "3.14"
+lxml==6.0.2 ; python_version >= "3.14"
 #Description: This is a requirement of unittest-xml-reporting

 PyGithub==2.3.0
@ -334,7 +339,9 @@ sympy==1.13.3
 #Pinned versions:
 #test that import:

-onnx==1.19.1
+onnx==1.19.1 ; python_version < "3.14"
+# Unpin once Python 3.14 is supported. See  onnxruntime issue 26309.
+onnx==1.18.0 ; python_version == "3.14"
 #Description: Required by onnx tests, and mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
@ -359,7 +366,7 @@ pwlf==2.2.1
 #test that import: test_sac_estimator.py

 # To build PyTorch itself
-pyyaml==6.0.2
+pyyaml==6.0.3
 pyzstd
 setuptools==78.1.1
 packaging==23.1
--- a/.ci/lumen_cli/pyproject.toml
+++ b/.ci/lumen_cli/pyproject.toml
@ -6,7 +6,7 @@ dependencies = [
    "GitPython==3.1.45",
    "docker==7.1.0",
    "pytest==7.3.2",
-    "uv==0.9.5"
+    "uv==0.9.6"
 ]

 [tool.setuptools]
--- a/.github/actions/diskspace-cleanup/action.yml
+++ b/.github/actions/diskspace-cleanup/action.yml
@ -27,7 +27,9 @@ runs:
            docker system prune -af
            diskspace_new=$(df -H --output=pcent ${docker_root_dir} | sed -n 2p | sed 's/%//' | sed 's/ //')
            if [[ "$diskspace_new" -gt "$diskspace_cutoff" ]] ; then
-                echo "Error: Available diskspace is less than $diskspace_cutoff percent. Not enough diskspace."
+                diskspace_cutoff_int=$((diskspace_cutoff + 0))
+                difference=$((100 - diskspace_cutoff_int))
+                echo "Error: Available diskspace is less than $difference percent. Not enough diskspace."
                echo "$msg"
                exit 1
            else
--- a/.github/ci_commit_pins/audio.txt
+++ b/.github/ci_commit_pins/audio.txt
@ -1 +1 @@
-69bbe7363897764f9e758d851cd0340147d27f94
+3b0e7a6f192ca2715e7e6cbe5db007aea7165fe2
--- a/.github/ci_commit_pins/vision.txt
+++ b/.github/ci_commit_pins/vision.txt
@ -1 +1 @@
-1752fe6809b74921644866275ab80244b96e80bc
+218d2ab791d437309f91e0486eb9fa7f00badc17
--- a/.github/merge_rules.yaml
+++ b/.github/merge_rules.yaml
@ -540,6 +540,26 @@
  - Lint
  - pull

+- name: PrivateUse1
+  patterns:
+  - torch/accelerator/**
+  - torch/utils/backend_registration.py
+  - torch/csrc/acc/**
+  - torch/csrc/DeviceAccelerator.*
+  - torch/csrc/profiler/standalone/privateuse1_observer.*
+  - aten/src/ATen/DeviceAccelerator.*
+  - aten/src/ATen/core/GeneratorForPrivateuseone.*
+  - aten/src/ATen/detail/PrivateUse1HooksInterface.*
+  - docs/source/accelerator/**
+  - test/cpp_extensions/open_registration_extension/torch_openreg/**
+  approved_by:
+  - albanD
+  - fffrog
+  mandatory_checks_name:
+  - EasyCLA
+  - Lint
+  - pull
+
 - name: superuser
  patterns:
  - '*'
--- a/.github/pytorch-probot.yml
+++ b/.github/pytorch-probot.yml
@ -26,6 +26,7 @@ ciflow_push_tags:
 - ciflow/nightly
 - ciflow/op-benchmark
 - ciflow/periodic
+- ciflow/periodic-rocm-mi200
 - ciflow/periodic-rocm-mi300
 - ciflow/pull
 - ciflow/quantization-periodic
--- a/.github/scripts/generate_binary_build_matrix.py
+++ b/.github/scripts/generate_binary_build_matrix.py
@ -56,7 +56,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
        "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | "
        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
+        "nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | "
        "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | "
        "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | "
        "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'"
@ -73,7 +73,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
        "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | "
        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
+        "nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | "
        "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | "
        "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
        "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
@ -90,7 +90,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
        "nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | "
        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | "
+        "nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | "
        "nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | "
        "nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | "
        "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'"
@ -107,7 +107,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
        "nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | "
        "nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | "
        "nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | "
+        "nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | "
        "nvidia-nvtx==13.0.85; platform_system == 'Linux' | "
        "nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | "
        "nvidia-cufile==1.15.1.6; platform_system == 'Linux'"
--- a/.github/workflows/docker-builds.yml
+++ b/.github/workflows/docker-builds.yml
@ -57,6 +57,7 @@ jobs:
          pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11,
          pytorch-linux-jammy-py3.10-clang12,
          pytorch-linux-jammy-py3.13-clang12,
+          pytorch-linux-jammy-py3.14-clang12,
          pytorch-linux-jammy-rocm-n-py3,
          pytorch-linux-noble-rocm-n-py3,
          pytorch-linux-jammy-rocm-n-py3-benchmarks,
--- a/.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
+++ b/.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
@ -132,7 +132,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_10-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -178,7 +178,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_10-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -224,7 +224,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_10-cuda-aarch64-12_9
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -270,7 +270,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_10-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -381,7 +381,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_11-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -427,7 +427,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_11-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -473,7 +473,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_11-cuda-aarch64-12_9
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -519,7 +519,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_11-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -630,7 +630,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_12-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -676,7 +676,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_12-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -722,7 +722,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_12-cuda-aarch64-12_9
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -768,7 +768,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_12-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -879,7 +879,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -925,7 +925,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -971,7 +971,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13-cuda-aarch64-12_9
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1017,7 +1017,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1128,7 +1128,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13t-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1174,7 +1174,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13t-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1220,7 +1220,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13t-cuda-aarch64-12_9
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1266,7 +1266,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_13t-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1377,7 +1377,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1423,7 +1423,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1469,7 +1469,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14-cuda-aarch64-12_9
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1515,7 +1515,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1626,7 +1626,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14t-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1672,7 +1672,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14t-cuda-aarch64-12_8
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1718,7 +1718,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14t-cuda-aarch64-12_9
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1764,7 +1764,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_14t-cuda-aarch64-13_0
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
--- a/.github/workflows/generated-linux-binary-manywheel-nightly.yml
+++ b/.github/workflows/generated-linux-binary-manywheel-nightly.yml
@ -127,7 +127,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_10-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_10-cuda12_6-test:  # Testing
@ -193,7 +193,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_10-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_10-cuda12_8-test:  # Testing
@ -259,7 +259,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_10-cuda12_9
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_10-cuda12_9-test:  # Testing
@ -325,7 +325,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_10-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_10-cuda13_0-test:  # Testing
@ -793,7 +793,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_11-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_11-cuda12_6-test:  # Testing
@ -859,7 +859,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_11-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_11-cuda12_8-test:  # Testing
@ -925,7 +925,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_11-cuda12_9
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_11-cuda12_9-test:  # Testing
@ -991,7 +991,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_11-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_11-cuda13_0-test:  # Testing
@ -1459,7 +1459,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_12-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_12-cuda12_6-test:  # Testing
@ -1525,7 +1525,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_12-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_12-cuda12_8-test:  # Testing
@ -1591,7 +1591,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_12-cuda12_9
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_12-cuda12_9-test:  # Testing
@ -1657,7 +1657,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_12-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_12-cuda13_0-test:  # Testing
@ -2125,7 +2125,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13-cuda12_6-test:  # Testing
@ -2191,7 +2191,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13-cuda12_8-test:  # Testing
@ -2257,7 +2257,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13-cuda12_9
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13-cuda12_9-test:  # Testing
@ -2323,7 +2323,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13-cuda13_0-test:  # Testing
@ -2791,7 +2791,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13t-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13t-cuda12_6-test:  # Testing
@ -2857,7 +2857,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13t-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13t-cuda12_8-test:  # Testing
@ -2923,7 +2923,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13t-cuda12_9
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13t-cuda12_9-test:  # Testing
@ -2989,7 +2989,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_13t-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_13t-cuda13_0-test:  # Testing
@ -3457,7 +3457,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14-cuda12_6-test:  # Testing
@ -3523,7 +3523,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14-cuda12_8-test:  # Testing
@ -3589,7 +3589,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14-cuda12_9
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14-cuda12_9-test:  # Testing
@ -3655,7 +3655,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14-cuda13_0-test:  # Testing
@ -4123,7 +4123,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14t-cuda12_6
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14t-cuda12_6-test:  # Testing
@ -4189,7 +4189,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14t-cuda12_8
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14t-cuda12_8-test:  # Testing
@ -4255,7 +4255,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14t-cuda12_9
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14t-cuda12_9-test:  # Testing
@ -4321,7 +4321,7 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      build_name: manywheel-py3_14t-cuda13_0
      build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
  manywheel-py3_14t-cuda13_0-test:  # Testing
--- a/.github/workflows/periodic-rocm-mi200.yml
+++ b/.github/workflows/periodic-rocm-mi200.yml
@ -0,0 +1,84 @@
+name: periodic-rocm-mi200
+
+on:
+  schedule:
+    # We have several schedules so jobs can check github.event.schedule to activate only for a fraction of the runs.
+    # Also run less frequently on weekends.
+    - cron: 45 0,8,16 * * 1-5
+    - cron: 45 4 * * 0,6
+    - cron: 45 4,12,20 * * 1-5
+    - cron: 45 12 * * 0,6
+    - cron: 29 8 * * *  # about 1:29am PDT, for mem leak check and rerun disabled tests
+  push:
+    tags:
+      - ciflow/periodic/*
+      - ciflow/periodic-rocm-mi200/*
+    branches:
+      - release/*
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}-${{ github.event.schedule }}
+  cancel-in-progress: true
+
+permissions:
+  id-token: write
+  contents: read
+
+jobs:
+  llm-td:
+    if: github.repository_owner == 'pytorch'
+    name: before-test
+    uses: ./.github/workflows/llm_td_retrieval.yml
+    permissions:
+      id-token: write
+      contents: read
+
+  target-determination:
+    name: before-test
+    uses: ./.github/workflows/target_determination.yml
+    needs: llm-td
+    permissions:
+      id-token: write
+      contents: read
+
+  get-label-type:
+    name: get-label-type
+    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
+    if: (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch'
+    with:
+      triggering_actor: ${{ github.triggering_actor }}
+      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
+      curr_branch: ${{ github.head_ref || github.ref_name }}
+      curr_ref_type: ${{ github.ref_type }}
+
+  linux-jammy-rocm-py3_10-build:
+    name: linux-jammy-rocm-py3.10
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
+      test-matrix: |
+        { include: [
+          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
+        ]}
+    secrets: inherit
+
+  linux-jammy-rocm-py3_10-test:
+    permissions:
+      id-token: write
+      contents: read
+    name: linux-jammy-rocm-py3.10
+    uses: ./.github/workflows/_rocm-test.yml
+    needs:
+      - linux-jammy-rocm-py3_10-build
+      - target-determination
+    with:
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
+    secrets: inherit
--- a/.github/workflows/periodic.yml
+++ b/.github/workflows/periodic.yml
@ -204,37 +204,6 @@ jobs:
      test-matrix: ${{ needs.linux-jammy-cuda13_0-py3_10-gcc11-build.outputs.test-matrix }}
    secrets: inherit

-  linux-jammy-rocm-py3_10-build:
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
-      test-matrix: |
-        { include: [
-          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
-        ]}
-    secrets: inherit
-
-  linux-jammy-rocm-py3_10-test:
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-jammy-rocm-py3_10-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
-    secrets: inherit
-
  linux-jammy-cuda12_8-py3-gcc11-slow-gradcheck-build:
    name: linux-jammy-cuda12.8-py3-gcc11-slow-gradcheck
    uses: ./.github/workflows/_linux-build.yml
--- a/.github/workflows/upload-test-stats.yml
+++ b/.github/workflows/upload-test-stats.yml
@ -6,6 +6,7 @@ on:
      - pull
      - trunk
      - periodic
+      - periodic-rocm-mi200
      - periodic-rocm-mi300
      - inductor
      - unstable
--- a/.github/workflows/xpu.yml
+++ b/.github/workflows/xpu.yml
@ -59,14 +59,18 @@ jobs:
      runner: linux.c7i.12xlarge
      test-matrix: |
        { include: [
-          { config: "default", shard: 1, num_shards: 8, runner: "linux.idc.xpu" },
-          { config: "default", shard: 2, num_shards: 8, runner: "linux.idc.xpu" },
-          { config: "default", shard: 3, num_shards: 8, runner: "linux.idc.xpu" },
-          { config: "default", shard: 4, num_shards: 8, runner: "linux.idc.xpu" },
-          { config: "default", shard: 5, num_shards: 8, runner: "linux.idc.xpu" },
-          { config: "default", shard: 6, num_shards: 8, runner: "linux.idc.xpu" },
-          { config: "default", shard: 7, num_shards: 8, runner: "linux.idc.xpu" },
-          { config: "default", shard: 8, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 1, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 2, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 3, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 4, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 5, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 6, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 7, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 8, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 9, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 10, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 11, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 12, num_shards: 12, runner: "linux.idc.xpu" },
        ]}
    secrets: inherit

--- a/.lintrunner.toml
+++ b/.lintrunner.toml
@ -1198,12 +1198,6 @@ exclude_patterns = [
    'torch/_inductor/fx_passes/serialized_patterns/**',
    'torch/_inductor/autoheuristic/artifacts/**',
    'torch/utils/model_dump/preact.mjs',
-    # These files are all grandfathered in, feel free to remove from this list
-    # as necessary
-    # NOTE: remove the patterns in the order they are listed
-    'aten/src/ATen/native/[a-pA-P]*/**',
-    'aten/src/ATen/[a-mA-M]*/**',
-    'test/**',
 ]
 init_command = [
    'python3',
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -374,7 +374,7 @@ cmake_dependent_option(
  "Build the lazy Torchscript backend, not compatible with mobile builds" ON
  "NOT INTERN_BUILD_MOBILE" OFF)
 cmake_dependent_option(BUILD_FUNCTORCH "Build Functorch" ON "BUILD_PYTHON" OFF)
-cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin fodler"
+cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin folder"
                       OFF "USE_CUDA" OFF)
 cmake_dependent_option(USE_KLEIDIAI "Use KleidiAI for the ARM CPU & AARCH64 architecture." ON
                        "CPU_AARCH64" OFF)
--- a/aten/src/ATen/CMakeLists.txt
+++ b/aten/src/ATen/CMakeLists.txt
@ -260,7 +260,7 @@ IF(USE_FBGEMM_GENAI)
  if(USE_CUDA)
    # To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
    # If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
-    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*mx8mx8bf16_grouped.*")
+    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*(mx8mx8bf16_grouped|f4f4bf16_grouped).*")
    file(GLOB_RECURSE fbgemm_genai_native_cuda_cu
      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/*.cu"
      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/**/*.cu")
@ -291,6 +291,7 @@ IF(USE_FBGEMM_GENAI)

    set(fbgemm_genai_cuh
      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
+      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/f4f4bf16_grouped/"
      "${FBGEMM_GENAI_SRCS}/"
    )

--- a/aten/src/ATen/core/CachingHostAllocator.h
+++ b/aten/src/ATen/core/CachingHostAllocator.h
@ -677,8 +677,8 @@ struct CachingHostAllocatorImpl {
  // size. This allows us to quickly find a free block of the right size.
  // We use deque to store per size free list and guard the list with its own
  // mutex.
-  alignas(hardware_destructive_interference_size) std::vector<FreeBlockList<B>> free_list_ =
-      std::vector<FreeBlockList<B>>(MAX_SIZE_INDEX);
+  alignas(hardware_destructive_interference_size) std::vector<FreeBlockList<B>>
+      free_list_{MAX_SIZE_INDEX};

  alignas(hardware_destructive_interference_size) std::mutex events_mutex_;
  std::deque<std::pair<E, B*>> events_; // event queue paired with block
--- a/aten/src/ATen/core/DimVector.h
+++ b/aten/src/ATen/core/DimVector.h
@ -3,7 +3,7 @@

 namespace at {

-// Re-declaring 'DimVector' type and size inside 'at' namespace.
+// Redeclaring 'DimVector' type and size inside 'at' namespace.
 // This is done to avoid modifying every use into their 'c10'
 // equivalent.

--- a/aten/src/ATen/core/GeneratorForPrivateuseone.cpp
+++ b/aten/src/ATen/core/GeneratorForPrivateuseone.cpp
@ -16,7 +16,7 @@ _GeneratorRegister::_GeneratorRegister(const GeneratorFuncType& func) {

  TORCH_WARN_DEPRECATION(
      "REGISTER_GENERATOR_PRIVATEUSE1 is deprecated. \
-      Please derive PrivateUse1HooksInterface to implememt getNewGenerator instead.")
+      Please derive PrivateUse1HooksInterface to implement getNewGenerator instead.")

  TORCH_CHECK(
      !GetGeneratorPrivate().has_value(),
--- a/aten/src/ATen/core/IListRef.h
+++ b/aten/src/ATen/core/IListRef.h
@ -149,7 +149,7 @@
 *    First, keep in mind that we assume that boxed containers will
 *    have to deal with `IValue` (e.g. `c10::List`). In this context,
 *    what may be happening is that `IValue` doesn't store internally
- *    your type `T`. Instead, it constructs a type new `T` everytime
+ *    your type `T`. Instead, it constructs a type new `T` every time
 *    you try to get `T` for it (see `IListRef<at::OptinalTensorRef>`).
 */

@ -186,7 +186,7 @@ class IListRef;
 * This macro is useful because it allows us to handle different
 * types (that correspond to different tags) to be implemented
 * only once. We can do it even when the implementation of the
- * different tags aren't syntatically the same, by dispatching
+ * different tags aren't syntactically the same, by dispatching
 * it to a function (e.g. `ImplT::<dispatch-function>(this_)`).
 */
 #define TORCH_ILISTREF_UNWRAP(TAG, BODY)                         \
--- a/aten/src/ATen/core/IListRef_inl.h
+++ b/aten/src/ATen/core/IListRef_inl.h
@ -42,7 +42,7 @@ class IListRefTagImplBase<IListRefTag::Unboxed, T, ListElemT> {
  /*
   * We have these function (besides the `unwrap`s above) because the
   * implementation for both `IListRef::operator[]` and `IListRefIterator::operator*`
-   * weren't syntatically equal for the existing tags at the time
+   * weren't syntactically equal for the existing tags at the time
   * (`Unboxed` and `Boxed`).
   */
  static IListRefConstRef<T> front(const list_type& lst) {
--- a/aten/src/ATen/core/Variadic.h
+++ b/aten/src/ATen/core/Variadic.h
@ -12,7 +12,7 @@ namespace at {
 // in order.  This is most commonly used in autogenerated code,
 // where it is convenient to have a function that can uniformly
 // take arguments of different types.  If your arguments
-// are homogenous consider using a std::initializer_list instead.
+// are homogeneous consider using a std::initializer_list instead.
 //
 // For examples of this in use, see torch/csrc/utils/variadic.h
 template <typename F>
--- a/aten/src/ATen/core/dispatch/Dispatcher.cpp
+++ b/aten/src/ATen/core/dispatch/Dispatcher.cpp
@ -111,7 +111,7 @@ void Dispatcher::waitForDef(const FunctionSchema& schema) {
  TORCH_INTERNAL_ASSERT(r,
    "Expected main interpreter to define ", schema.operator_name(),
    ", but this didn't happen within timeout.  Are you trying to load "
-    "different models in the same torchdeploy/multipy instance?  You "
+    "different models in the same torchdeploy/multipy instance?  You " // codespell:ignore
    "must warmup each interpreter identically, e.g., import all "
    "the same dependencies.");
 }
@ -129,7 +129,7 @@ void Dispatcher::waitForImpl(const OperatorName& op_name, std::optional<c10::Dis
  TORCH_INTERNAL_ASSERT(r,
    "Expected main interpreter to implement ", dk, " for ", op_name,
    ", but this didn't happen within timeout.  Are you trying to load "
-    "different models in the same torchdeploy/multipy instance?  You "
+    "different models in the same torchdeploy/multipy instance?  You " // codespell:ignore
    "must warmup each interpreter identically, e.g., import all "
    "the same dependencies.");
 }
@ -442,8 +442,8 @@ RegistrationHandleRAII Dispatcher::registerFallback(DispatchKey dispatchKey, Ker

  auto idx = getDispatchTableIndexForDispatchKey(dispatchKey);
  TORCH_CHECK(idx >= 0 && static_cast<uint64_t>(idx) < backendFallbackKernels_.size(), "idx=", idx);
-  // NB: Perserve BC for registering fallback for AutogradPrivateUse1 multiple time,
-  // refer to https://github.com/pytorch/pytorch/issues/163979 for more informations.
+  // NB: Preserve BC for registering fallback for AutogradPrivateUse1 multiple time,
+  // refer to https://github.com/pytorch/pytorch/issues/163979 for more information.
  TORCH_CHECK(
      dispatchKey == DispatchKey::AutogradPrivateUse1 ||
          !backendFallbackKernels_[idx].kernel.isValid(),
--- a/aten/src/ATen/core/dispatch/Dispatcher.h
+++ b/aten/src/ATen/core/dispatch/Dispatcher.h
@ -222,7 +222,8 @@ class TORCH_API Dispatcher final {
    return backendFallbackKernels_[dispatch_ix].kernel.isValid();
  }

-  // Used by torchdeploy/multipy for multiple interpreters racing.
+  // Used by torchdeploy/multipy for multiple  // codespell:ignore: multipy
+  // interpreters racing.
  void waitForDef(const FunctionSchema& schema);
  void waitForImpl(
      const OperatorName& op_name,
@ -414,7 +415,7 @@ class TORCH_API Dispatcher final {
  std::unique_ptr<detail::RegistrationListenerList> listeners_;

  // This condition variable gets notified whenever we add a new def/impl to the
-  // dispatch table.  This is primarily used by multipy/torchdeploy, when
+  // dispatch table.  This is primarily used by multiply/torchdeploy, when
  // we have multiple interpreters trying to register to the dispatch table.
  // In this situation, whenever the non-primary interpreter would have tried
  // to register to the dispatch table, instead it will check to see if the
--- a/aten/src/ATen/core/ivalue_inl.h
+++ b/aten/src/ATen/core/ivalue_inl.h
@ -990,7 +990,7 @@ struct C10_EXPORT ivalue::Future final : c10::intrusive_ptr_target {
    std::unique_lock<std::mutex> lock(mutex_);
    if (completed_) {
      // This should be rare and shouldn't cause log spew. Its important to
-      // log errors and thats why we have this log here.
+      // log errors and that's why we have this log here.
      std::string msg = c10::str(
          "Skipping setting following error on the Future since "
          "it is already marked completed (this is not necessarily "
--- a/aten/src/ATen/core/jit_type.h
+++ b/aten/src/ATen/core/jit_type.h
@ -887,7 +887,7 @@ struct TORCH_API ListType
  // this function will return the global singleton type pointer
  // the type List<T>.
  // The extra "identifier" argument is needed because we have multiple container types
-  // that all re-use this function (List<T>, array<T, N>, etc.)
+  // that all reuse this function (List<T>, array<T, N>, etc.)
  static TypePtr get(const std::string& identifier, TypePtr inner);

  // common cast List[Tensor]
@ -983,7 +983,7 @@ struct TORCH_API DictType : public SharedType {
  // this function will return the global singleton type pointer
  // the type List<T>.
  // The extra "identifier" argument is needed because we have multiple container types
-  // that all re-use this function (Dict<K, V> and unordered_map<K, V>)
+  // that all reuse this function (Dict<K, V> and unordered_map<K, V>)
  static TypePtr get(const std::string& identifier, TypePtr key, TypePtr val);

 private:
--- a/aten/src/ATen/core/type.cpp
+++ b/aten/src/ATen/core/type.cpp
@ -680,7 +680,7 @@ TORCH_API bool elementTypeCanBeInferredFromMembers(const TypePtr& elem_type) {
    return false;
  }
  if (elem_type->kind() == AnyType::Kind) {
-    // List of Any can contains heterogenous types
+    // List of Any can contains heterogeneous types
    return false;
  }
  return true;
--- a/aten/src/ATen/cpu/vec/vec128/vec128_bfloat16_neon.h
+++ b/aten/src/ATen/cpu/vec/vec128/vec128_bfloat16_neon.h
@ -354,47 +354,9 @@ class Vectorized<c10::BFloat16> : public Vectorized16<

  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(abs)
  Vectorized frac() const;
+  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(neg)
  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(trunc)
  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(sqrt)
-
-#ifdef __ARM_FEATURE_BF16
-  Vectorized<c10::BFloat16> neg() const {
-    return -values;
-  }
-  Vectorized<c10::BFloat16> reciprocal() const {
-    return 1.0f / values;
-  }
-  Vectorized<c10::BFloat16> operator==(
-      const Vectorized<c10::BFloat16>& other) const {
-    return values == other.values;
-  }
-
-  Vectorized<c10::BFloat16> operator!=(
-      const Vectorized<c10::BFloat16>& other) const {
-    return values != other.values;
-  }
-
-  Vectorized<c10::BFloat16> operator<(
-      const Vectorized<c10::BFloat16>& other) const {
-    return values < other.values;
-  }
-
-  Vectorized<c10::BFloat16> operator<=(
-      const Vectorized<c10::BFloat16>& other) const {
-    return values <= other.values;
-  }
-
-  Vectorized<c10::BFloat16> operator>(
-      const Vectorized<c10::BFloat16>& other) const {
-    return values > other.values;
-  }
-
-  Vectorized<c10::BFloat16> operator>=(
-      const Vectorized<c10::BFloat16>& other) const {
-    return values >= other.values;
-  }
-#else
-  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(neg)
  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(reciprocal)
  DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator==)
  DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator!=)
@ -402,7 +364,6 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
  DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator<=)
  DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator>)
  DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator>=)
-#endif

 #undef DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD
 #undef DEFINE_BINARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD
@ -451,52 +412,28 @@ template <>
 Vectorized<c10::BFloat16> inline operator+(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  return x + y;
-#else
  return binary_operator_via_float(std::plus<Vectorized<float>>(), a, b);
-#endif
 }

 template <>
 Vectorized<c10::BFloat16> inline operator-(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  return x - y;
-#else
  return binary_operator_via_float(std::minus<Vectorized<float>>(), a, b);
-#endif
 }

 template <>
 Vectorized<c10::BFloat16> inline operator*(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  return x * y;
-#else
  return binary_operator_via_float(std::multiplies<Vectorized<float>>(), a, b);
-#endif
 }

 template <>
 Vectorized<c10::BFloat16> inline operator/(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  return x / y;
-#else
  return binary_operator_via_float(std::divides<Vectorized<float>>(), a, b);
-#endif
 }

 // frac. Implement this here so we can use subtraction
@ -607,19 +544,12 @@ Vectorized<c10::BFloat16> inline fmadd(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b,
    const Vectorized<c10::BFloat16>& c) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  bfloat16x8_t z = c;
-  return x * y + z;
-#else
  // NOTE [BF16 FMA]: There isn't an FMA that accumulates into BF16!  Also,
  // vbfmlalbq_f32 and vbfmlaltq_f32 take the even and odd-numbered
  // elements, not the bottom and top half, so they don't seem
  // particularly useful here. Ideally we would include dot product in
  // the Vectorized interface...
  return a * b + c;
-#endif
 }

 template <>
@ -627,15 +557,8 @@ Vectorized<c10::BFloat16> inline fnmadd(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b,
    const Vectorized<c10::BFloat16>& c) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  bfloat16x8_t z = c;
-  return (-x) * y + z;
-#else
  // See NOTE [BF16 FMA] above.
  return -a * b + c;
-#endif
 }

 template <>
@ -643,15 +566,8 @@ Vectorized<c10::BFloat16> inline fmsub(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b,
    const Vectorized<c10::BFloat16>& c) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  bfloat16x8_t z = c;
-  return x * y - z;
-#else
  // See NOTE [BF16 FMA] above.
  return a * b - c;
-#endif
 }

 template <>
@ -659,15 +575,8 @@ Vectorized<c10::BFloat16> inline fnmsub(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b,
    const Vectorized<c10::BFloat16>& c) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  bfloat16x8_t z = c;
-  return (-x) * y - z;
-#else
  // See NOTE [BF16 FMA] above.
  return -a * b - c;
-#endif
 }

 #endif // !defined(C10_MOBILE) && defined(__aarch64__)
--- a/aten/src/ATen/cpu/vec/vec128/vec128_convert.h
+++ b/aten/src/ATen/cpu/vec/vec128/vec128_convert.h
@ -21,12 +21,46 @@ inline void convertImpl(
  }
 }

+template <typename to_type>
+inline void convertFromBool(
+    const bool* __restrict src,
+    to_type* __restrict dst,
+    int64_t n) {
+  const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
+  uint64_t len = static_cast<uint64_t>(n);
+  for (uint64_t i = 0; i < len; i++) {
+    dst[i] = srcPtr[i] != 0 ? static_cast<to_type>(1) : static_cast<to_type>(0);
+  }
+}
+
+template <typename from_type>
+inline void convertToBool(
+    const from_type* __restrict src,
+    bool* __restrict dst,
+    int64_t n) {
+  uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
+  uint64_t len = static_cast<uint64_t>(n);
+  for (uint64_t i = 0; i < len; i++) {
+    dstPtr[i] = src[i] != static_cast<from_type>(0) ? 1 : 0;
+  }
+}
+
 #define CONVERT_TEMPLATE(from_type, to_type)                           \
  template <>                                                          \
  inline void convert(const from_type* src, to_type* dst, int64_t n) { \
    return convertImpl<from_type, to_type>(src, dst, n);               \
  }

+#define CONVERT_FROM_BOOL_TEMPLATE(to_type)                       \
+  inline void convert(const bool* src, to_type* dst, int64_t n) { \
+    return convertFromBool<to_type>(src, dst, n);                 \
+  }
+
+#define CONVERT_TO_BOOL_TEMPLATE(from_type)                         \
+  inline void convert(const from_type* src, bool* dst, int64_t n) { \
+    return convertToBool<from_type>(src, dst, n);                   \
+  }
+
 CONVERT_TEMPLATE(uint8_t, uint8_t)
 CONVERT_TEMPLATE(uint8_t, int8_t)
 CONVERT_TEMPLATE(uint8_t, int16_t)
@ -34,6 +68,7 @@ CONVERT_TEMPLATE(uint8_t, int32_t)
 CONVERT_TEMPLATE(uint8_t, int64_t)
 CONVERT_TEMPLATE(uint8_t, float)
 CONVERT_TEMPLATE(uint8_t, double)
+CONVERT_TO_BOOL_TEMPLATE(uint8_t)
 CONVERT_TEMPLATE(int8_t, uint8_t)
 CONVERT_TEMPLATE(int8_t, int8_t)
 CONVERT_TEMPLATE(int8_t, int16_t)
@ -41,6 +76,7 @@ CONVERT_TEMPLATE(int8_t, int32_t)
 CONVERT_TEMPLATE(int8_t, int64_t)
 CONVERT_TEMPLATE(int8_t, float)
 CONVERT_TEMPLATE(int8_t, double)
+CONVERT_TO_BOOL_TEMPLATE(int8_t)
 CONVERT_TEMPLATE(int16_t, uint8_t)
 CONVERT_TEMPLATE(int16_t, int8_t)
 CONVERT_TEMPLATE(int16_t, int16_t)
@ -48,6 +84,7 @@ CONVERT_TEMPLATE(int16_t, int32_t)
 CONVERT_TEMPLATE(int16_t, int64_t)
 CONVERT_TEMPLATE(int16_t, float)
 CONVERT_TEMPLATE(int16_t, double)
+CONVERT_TO_BOOL_TEMPLATE(int16_t)
 CONVERT_TEMPLATE(int32_t, uint8_t)
 CONVERT_TEMPLATE(int32_t, int8_t)
 CONVERT_TEMPLATE(int32_t, int16_t)
@ -55,6 +92,7 @@ CONVERT_TEMPLATE(int32_t, int32_t)
 CONVERT_TEMPLATE(int32_t, int64_t)
 CONVERT_TEMPLATE(int32_t, float)
 CONVERT_TEMPLATE(int32_t, double)
+CONVERT_TO_BOOL_TEMPLATE(int32_t)
 CONVERT_TEMPLATE(int64_t, uint8_t)
 CONVERT_TEMPLATE(int64_t, int8_t)
 CONVERT_TEMPLATE(int64_t, int16_t)
@ -62,6 +100,7 @@ CONVERT_TEMPLATE(int64_t, int32_t)
 CONVERT_TEMPLATE(int64_t, int64_t)
 CONVERT_TEMPLATE(int64_t, float)
 CONVERT_TEMPLATE(int64_t, double)
+CONVERT_TO_BOOL_TEMPLATE(int64_t)
 CONVERT_TEMPLATE(float, uint8_t)
 CONVERT_TEMPLATE(float, int8_t)
 CONVERT_TEMPLATE(float, int16_t)
@ -69,6 +108,7 @@ CONVERT_TEMPLATE(float, int32_t)
 CONVERT_TEMPLATE(float, int64_t)
 CONVERT_TEMPLATE(float, float)
 CONVERT_TEMPLATE(float, double)
+CONVERT_TO_BOOL_TEMPLATE(float)
 CONVERT_TEMPLATE(double, uint8_t)
 CONVERT_TEMPLATE(double, int8_t)
 CONVERT_TEMPLATE(double, int16_t)
@ -76,6 +116,14 @@ CONVERT_TEMPLATE(double, int32_t)
 CONVERT_TEMPLATE(double, int64_t)
 CONVERT_TEMPLATE(double, float)
 CONVERT_TEMPLATE(double, double)
+CONVERT_TO_BOOL_TEMPLATE(double)
+CONVERT_FROM_BOOL_TEMPLATE(uint8_t)
+CONVERT_FROM_BOOL_TEMPLATE(int8_t)
+CONVERT_FROM_BOOL_TEMPLATE(int16_t)
+CONVERT_FROM_BOOL_TEMPLATE(int32_t)
+CONVERT_FROM_BOOL_TEMPLATE(int64_t)
+CONVERT_FROM_BOOL_TEMPLATE(float)
+CONVERT_FROM_BOOL_TEMPLATE(double)
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC

 #define CONVERT_FROM_FP16_TEMPLATE(to_type)                            \
@ -107,6 +155,41 @@ CONVERT_TO_FP16_TEMPLATE(int32_t)
 CONVERT_TO_FP16_TEMPLATE(int64_t)
 CONVERT_TO_FP16_TEMPLATE(float)
 CONVERT_TO_FP16_TEMPLATE(double)
+
+inline void convertBoolToFp16Impl(
+    const bool* __restrict src,
+    at::Half* __restrict dst,
+    int64_t n) {
+  const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
+  float16_t* dstPtr = reinterpret_cast<float16_t*>(dst);
+  uint64_t len = static_cast<uint64_t>(n);
+  for (uint64_t i = 0; i < len; i++) {
+    dstPtr[i] = srcPtr[i] != 0 ? 1.0 : 0;
+  }
+}
+
+template <>
+inline void convert(const bool* src, at::Half* dst, int64_t n) {
+  return convertBoolToFp16Impl(src, dst, n);
+}
+
+inline void convertFp16ToBoolImpl(
+    const at::Half* __restrict src,
+    bool* __restrict dst,
+    int64_t n) {
+  const float16_t* srcPtr = reinterpret_cast<const float16_t*>(src);
+  uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
+  uint64_t len = static_cast<uint64_t>(n);
+  for (uint64_t i = 0; i < len; i++) {
+    dstPtr[i] = srcPtr[i] != 0.0 ? 1 : 0;
+  }
+}
+
+template <>
+inline void convert(const at::Half* src, bool* dst, int64_t n) {
+  return convertFp16ToBoolImpl(src, dst, n);
+}
+
 #endif
 #ifdef __ARM_FEATURE_BF16
 CONVERT_TEMPLATE(bfloat16_t, uint8_t)
@ -124,6 +207,44 @@ CONVERT_TEMPLATE(int32_t, bfloat16_t)
 CONVERT_TEMPLATE(int64_t, bfloat16_t)
 CONVERT_TEMPLATE(float, bfloat16_t)
 CONVERT_TEMPLATE(double, bfloat16_t)
+
+inline void convertBoolToBfloat16Impl(
+    const bool* __restrict src,
+    c10::BFloat16* __restrict dst,
+    int64_t n) {
+  const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
+  uint16_t* dstPtr = reinterpret_cast<uint16_t*>(dst);
+  uint64_t len = static_cast<uint64_t>(n);
+  constexpr uint16_t kBf16One = 0x3f80; // 1.0 in bfloat16
+  for (uint64_t i = 0; i < len; i++) {
+    dstPtr[i] = srcPtr[i] != 0 ? kBf16One : 0;
+  }
+}
+
+template <>
+inline void convert(const bool* src, c10::BFloat16* dst, int64_t n) {
+  return convertBoolToBfloat16Impl(src, dst, n);
+}
+
+inline void convertBfloat16ToBoolImpl(
+    const c10::BFloat16* __restrict src,
+    bool* __restrict dst,
+    int64_t n) {
+  uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
+  const uint16_t* srcPtr = reinterpret_cast<const uint16_t*>(src);
+  uint64_t len = static_cast<uint64_t>(n);
+  for (uint64_t i = 0; i < len; i++) {
+    // Check if all non-sign bits are 0
+    bool isBf16Zero = (srcPtr[i] & 0x7fff) == 0;
+    dstPtr[i] = isBf16Zero ? 0 : 1;
+  }
+}
+
+template <>
+inline void convert(const c10::BFloat16* src, bool* dst, int64_t n) {
+  return convertBfloat16ToBoolImpl(src, dst, n);
+}
+
 #endif

 #endif
--- a/aten/src/ATen/cpu/vec/vec128/vec128_float_neon.h
+++ b/aten/src/ATen/cpu/vec/vec128/vec128_float_neon.h
@ -309,7 +309,7 @@ class Vectorized<float> {
  DEFINE_SLEEF_COMPATIBLE_UNARY_ELEMENTWISE_FUNC(expm1)
  // Implementation copied from Arm Optimized Routine
  // https://github.com/ARM-software/optimized-routines/blob/master/math/aarch64/advsimd/expf.c
-  Vectorized<float> exp_u20() const {
+  inline Vectorized<float> vexpq_f32_u20() const {
    // bail out to sleef if it's a special case:
    // i.e. there's an input s.t. |input| > 87.3....
    const float32x4_t special_bound = vdupq_n_f32(0x1.5d5e2ap+6f);
@ -348,6 +348,9 @@ class Vectorized<float> {

    return vfmaq_f32(scale, poly, scale);
  }
+  Vectorized<float> exp_u20() const {
+    return vexpq_f32_u20();
+  }
  Vectorized<float> fexp_u20() const {
    return exp_u20();
  }
@ -634,7 +637,7 @@ inline Vectorized<float> Vectorized<float>::erf() const {
  // - exp(- x * x)
  auto pow_2 = (*this) * (*this);
  auto neg_pow_2 = pow_2 ^ neg_zero_vec;
-  auto tmp4 = neg_pow_2.exp();
+  auto tmp4 = neg_pow_2.vexpq_f32_u20();
  auto tmp5 = tmp4 ^ neg_zero_vec;
  // erf(x) = sign(x) * (1 - r * t * exp(- x * x))
  auto tmp6 = t * tmp5;
--- a/aten/src/ATen/cpu/vec/vec512/vec512_float8.h
+++ b/aten/src/ATen/cpu/vec/vec512/vec512_float8.h
@ -498,8 +498,8 @@ static inline Vectorized<T> binary_fp8_op_as_fp32(

 // Refer to
 // https://github.com/pytorch/pytorch/pull/153364#discussion_r2086509353 FP8 +,
-// -, *, /, planed to be deleted in the future and here is just to make compiler
-// happy
+// -, *, /, planned to be deleted in the future and here is just to make
+// compiler happy
 Vectorized<Float8_e4m3fn> inline operator+(
    const Vectorized<Float8_e4m3fn>& a,
    const Vectorized<Float8_e4m3fn>& b) {
@ -585,8 +585,8 @@ class Vectorized<Float8_e5m2> : public Vectorizedf8<Float8_e5m2> {

 // Refer to
 // https://github.com/pytorch/pytorch/pull/153364#discussion_r2086509353 FP8 +,
-// -, *, /, planed to be deleted in the future and here is just to make compiler
-// happy
+// -, *, /, planned to be deleted in the future and here is just to make
+// compiler happy
 Vectorized<Float8_e5m2> inline operator+(
    const Vectorized<Float8_e5m2>& a,
    const Vectorized<Float8_e5m2>& b) {
--- a/aten/src/ATen/cuda/CUDAGreenContext.cpp
+++ b/aten/src/ATen/cuda/CUDAGreenContext.cpp
@ -1,78 +1,90 @@
 #include <ATen/cuda/CUDAGreenContext.h>

-namespace at::cuda {
-  GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
-#if CUDA_HAS_GREEN_CONTEXT
-    int driver_version;
-    C10_CUDA_CHECK(cudaDriverGetVersion(&driver_version));
-    TORCH_CHECK(
-        driver_version >= 12080, "cuda driver too old to use green context!");
-    CUcontext pctx = nullptr;
-    C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuCtxGetCurrent_(&pctx));
-    if (C10_UNLIKELY(!pctx)) {
-      TORCH_WARN(
-          "Attempted to create a green context but"
-          " there was no primary context! Creating a primary context...");
-
-      cudaFree(0);
-    }
-
-    CUdevice device;
-    device_id_ = device_id;
-    C10_CUDA_DRIVER_CHECK(
-        c10::cuda::DriverAPI::get()->cuDeviceGet_(&device, device_id));
-
-    // Get device resources
-    CUdevResource device_resource;
-    C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuDeviceGetDevResource_(
-        device, &device_resource, CU_DEV_RESOURCE_TYPE_SM));
-
-    // Split resources
-    std::vector<CUdevResource> result(1);
-    auto result_data = result.data();
-    unsigned int nb_groups = 1;
-    CUdevResource remaining;
-
-    C10_CUDA_DRIVER_CHECK(
-        c10::cuda::DriverAPI::get()->cuDevSmResourceSplitByCount_(
-            result_data,
-            &nb_groups,
-            &device_resource,
-            &remaining,
-            0, // default flags
-            num_sms));
-
-    TORCH_CHECK(nb_groups == 1, "Failed to create single resource group");
-
-    // Generate resource descriptor
-    CUdevResourceDesc desc;
-    C10_CUDA_DRIVER_CHECK(
-        c10::cuda::DriverAPI::get()->cuDevResourceGenerateDesc_(
-            &desc, result_data, 1));
-
-    // Create green context
-    // CU_GREEN_CTX_DEFAULT_STREAM is required per docs:
-    // https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__GREEN__CONTEXTS.html
-    C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuGreenCtxCreate_(
-        &green_ctx_, desc, device, CU_GREEN_CTX_DEFAULT_STREAM));
-
-    // Convert to regular context
-    C10_CUDA_DRIVER_CHECK(
-        c10::cuda::DriverAPI::get()->cuCtxFromGreenCtx_(&context_, green_ctx_));
-    TORCH_CHECK(context_, "Green ctx conversion to regular ctx failed!");
+#if defined(CUDA_VERSION) && !defined(USE_ROCM) && defined(PYTORCH_C10_DRIVER_API_SUPPORTED)
+#include <c10/cuda/driver_api.h>
+#include <stdexcept>
+#include <vector>
+#define HAS_CUDA_GREEN_CONTEXT() 1
 #else
-    TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
+#define HAS_CUDA_GREEN_CONTEXT() 0
+// Suppress unused private field warnings as this class is not supposed to be called
+C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-private-field")
+#endif
+
+namespace at::cuda {
+
+GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
+#if HAS_CUDA_GREEN_CONTEXT()
+  int driver_version;
+  C10_CUDA_CHECK(cudaDriverGetVersion(&driver_version));
+  TORCH_CHECK(
+      driver_version >= 12080, "cuda driver too old to use green context!");
+  CUcontext pctx = nullptr;
+  C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuCtxGetCurrent_(&pctx));
+  if (C10_UNLIKELY(!pctx)) {
+    TORCH_WARN(
+        "Attempted to create a green context but"
+        " there was no primary context! Creating a primary context...");
+
+    cudaFree(0);
+  }
+
+   CUdevice device;
+  device_id_ = device_id;
+  C10_CUDA_DRIVER_CHECK(
+      c10::cuda::DriverAPI::get()->cuDeviceGet_(&device, device_id));
+
+  // Get device resources
+  CUdevResource device_resource;
+  C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuDeviceGetDevResource_(
+      device, &device_resource, CU_DEV_RESOURCE_TYPE_SM));
+
+  // Split resources
+  std::vector<CUdevResource> result(1);
+  auto result_data = result.data();
+  unsigned int nb_groups = 1;
+  CUdevResource remaining;
+
+  C10_CUDA_DRIVER_CHECK(
+      c10::cuda::DriverAPI::get()->cuDevSmResourceSplitByCount_(
+          result_data,
+          &nb_groups,
+          &device_resource,
+          &remaining,
+          0, // default flags
+          num_sms));
+
+  TORCH_CHECK(nb_groups == 1, "Failed to create single resource group");
+
+  // Generate resource descriptor
+  CUdevResourceDesc desc;
+  C10_CUDA_DRIVER_CHECK(
+      c10::cuda::DriverAPI::get()->cuDevResourceGenerateDesc_(
+          &desc, result_data, 1));
+
+  // Create green context
+  // CU_GREEN_CTX_DEFAULT_STREAM is required per docs:
+  // https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__GREEN__CONTEXTS.html
+  C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuGreenCtxCreate_(
+      &green_ctx_, desc, device, CU_GREEN_CTX_DEFAULT_STREAM));
+
+  // Convert to regular context
+  C10_CUDA_DRIVER_CHECK(
+      c10::cuda::DriverAPI::get()->cuCtxFromGreenCtx_(&context_, green_ctx_));
+  TORCH_CHECK(context_, "Green ctx conversion to regular ctx failed!");
+#else
+  TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
 #endif
  }

  std::unique_ptr<GreenContext> GreenContext::create(
      uint32_t num_sms,
      std::optional<uint32_t> device_id) {
-#if CUDA_HAS_GREEN_CONTEXT
+#if HAS_CUDA_GREEN_CONTEXT()
    if (!device_id.has_value()) {
      device_id = at::cuda::current_device();
    }
-    return std::make_unique<GreenContext>(device_id.value(), num_sms);
+    return std::unique_ptr<GreenContext>(new GreenContext(device_id.value(), num_sms));
 #else
    TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
 #endif
@ -80,7 +92,7 @@ namespace at::cuda {

  // Implement move operations
  GreenContext::GreenContext(GreenContext&& other) noexcept{
-#if CUDA_HAS_GREEN_CONTEXT
+#if HAS_CUDA_GREEN_CONTEXT()
    device_id_ = std::exchange(other.device_id_, -1);
    green_ctx_ = std::exchange(other.green_ctx_, nullptr);
    context_ = std::exchange(other.context_, nullptr);
@ -91,7 +103,7 @@ namespace at::cuda {
  }

  GreenContext& GreenContext::operator=(GreenContext&& other) noexcept{
-#if CUDA_HAS_GREEN_CONTEXT
+#if HAS_CUDA_GREEN_CONTEXT()
    if (this != &other) {
      // Clean up current resources
      if (green_ctx_) {
@ -120,7 +132,7 @@ namespace at::cuda {
  }

  GreenContext::~GreenContext() noexcept{
-#if CUDA_HAS_GREEN_CONTEXT
+#if HAS_CUDA_GREEN_CONTEXT()
    C10_CUDA_DRIVER_CHECK(
        c10::cuda::DriverAPI::get()->cuGreenCtxDestroy_(green_ctx_));
 #else
@ -128,25 +140,9 @@ namespace at::cuda {
 #endif
  }

-  // Get the underlying CUDA context
-  CUcontext GreenContext::getContext() const {
-#if CUDA_HAS_GREEN_CONTEXT
-    return context_;
-#else
-    TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
-#endif
-  }
-
-  // Get the underlying green context
-#if CUDA_HAS_GREEN_CONTEXT
-  CUgreenCtx GreenContext::getGreenContext() const {
-    return green_ctx_;
-  }
-#endif
-
  // Make this context current
  void GreenContext::setContext() {
-#if CUDA_HAS_GREEN_CONTEXT
+#if HAS_CUDA_GREEN_CONTEXT()
    auto current_stream = c10::cuda::getCurrentCUDAStream();
    parent_stream_ = current_stream.stream();

@ -175,7 +171,7 @@ namespace at::cuda {
  }

  void GreenContext::popContext() {
-#if CUDA_HAS_GREEN_CONTEXT
+#if HAS_CUDA_GREEN_CONTEXT()
    // see above note about stream being hardcoded to the default stream
    at::cuda::CUDAEvent ev;
    ev.record(c10::cuda::getCurrentCUDAStream());
--- a/aten/src/ATen/cuda/CUDAGreenContext.h
+++ b/aten/src/ATen/cuda/CUDAGreenContext.h
@ -1,53 +1,38 @@
 #pragma once
 #include <ATen/cuda/CUDAEvent.h>
-
-#if defined(CUDA_VERSION) && !defined(USE_ROCM) && defined(PYTORCH_C10_DRIVER_API_SUPPORTED)
-#include <c10/cuda/driver_api.h>
 #include <cuda.h>
-#include <memory>
-#include <stdexcept>
-#include <vector>
-#define CUDA_HAS_GREEN_CONTEXT 1
-#else
-#define CUDA_HAS_GREEN_CONTEXT 0
-#endif
+
+// Forward declare green context as opaque ptr
+typedef struct CUgreenCtx_st* CUgreenCtx;

 namespace at::cuda {

 class TORCH_CUDA_CPP_API GreenContext {
 public:
-  GreenContext(uint32_t device_id, uint32_t num_sms);
-
-  static std::unique_ptr<GreenContext> create(uint32_t num_sms, std::optional<uint32_t> device_id);
+  // Green context creation
+  static std::unique_ptr<GreenContext> create(
+      uint32_t num_sms,
+      std::optional<uint32_t> device_id);
+  ~GreenContext() noexcept;

  // Delete copy constructor and assignment
  GreenContext(const GreenContext&) = delete;
  GreenContext& operator=(const GreenContext&) = delete;

-  // Implement move operations
-  GreenContext(GreenContext&& other) noexcept;
-  GreenContext& operator=(GreenContext&& other) noexcept;
-  ~GreenContext() noexcept;
-
-  // Get the underlying CUDA context
-  CUcontext getContext() const;
-
-  // Get the underlying green context
-#if CUDA_HAS_GREEN_CONTEXT
-  CUgreenCtx getGreenContext() const;
-#endif
-
  // Make this context current
  void setContext();

  void popContext();

 private:
-#if CUDA_HAS_GREEN_CONTEXT
+  GreenContext(uint32_t device_id, uint32_t num_sms);
+  // Implement move operations
+  GreenContext(GreenContext&& other) noexcept;
+  GreenContext& operator=(GreenContext&& other) noexcept;
+
  int32_t device_id_ = -1;
  CUgreenCtx green_ctx_ = nullptr;
  CUcontext context_ = nullptr;
  cudaStream_t parent_stream_ = nullptr;
-#endif
 };
 } // namespace at::cuda
--- a/aten/src/ATen/cuda/CUDASparse.h
+++ b/aten/src/ATen/cuda/CUDASparse.h
@ -7,17 +7,6 @@
 #endif


-#if defined(USE_ROCM)
-// hipSparse const API added in v2.4.0
-#if HIPSPARSE_VERSION >= 200400
-#define AT_USE_HIPSPARSE_GENERIC_API() 1
-#else
-#define AT_USE_HIPSPARSE_GENERIC_API() 1
-#endif
-#else // USE_ROCM
-#define AT_USE_HIPSPARSE_GENERIC_API() 0
-#endif // USE_ROCM
-
 // cuSparse Generic API spsv function was added in CUDA 11.3.0
 #if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && (CUSPARSE_VERSION >= 11500)
 #define AT_USE_CUSPARSE_GENERIC_SPSV() 1
--- a/aten/src/ATen/cuda/CUDASparseDescriptors.cpp
+++ b/aten/src/ATen/cuda/CUDASparseDescriptors.cpp
@ -179,7 +179,7 @@ CuSparseSpMatCsrDescriptor::CuSparseSpMatCsrDescriptor(const Tensor& input, int6
          batch_offset * values_batch_stride * values.itemsize(),
      index_type, // data type of row offsets index
      index_type, // data type of col indices
-      CUSPARSE_INDEX_BASE_ZERO, // base index of row offset and col indes
+      CUSPARSE_INDEX_BASE_ZERO, // base index of row offset and col index
      value_type // data type of values
      ));

--- a/aten/src/ATen/cuda/CachingHostAllocator.h
+++ b/aten/src/ATen/cuda/CachingHostAllocator.h
@ -10,7 +10,7 @@ namespace at::cuda {
 //
 // A caching allocator for CUDA host allocations (pinned memory).
 //
-// This provides a drop-in replacement for THCudaHostAllocator, which re-uses
+// This provides a drop-in replacement for THCudaHostAllocator, which reuses
 // freed pinned (page-locked) memory allocations. This avoids device
 // synchronizations due to cudaFreeHost calls.
 //
@ -26,7 +26,7 @@ inline TORCH_CUDA_CPP_API at::HostAllocator* getCachingHostAllocator() {
 }

 // Records an event in the specified stream. The allocation corresponding to the
-// input `ptr`/`ctx` will not be re-used until the event has occurred.
+// input `ptr`/`ctx` will not be reused until the event has occurred.
 C10_DEPRECATED_MESSAGE(
  "at::cuda::CachingHostAllocator_recordEvent(...) is deprecated. Please use at::getHostAllocator(at::kCUDA)->record_event(...) instead.")
 inline TORCH_CUDA_CPP_API bool CachingHostAllocator_recordEvent(
--- a/aten/src/ATen/cuda/Sleep.cu
+++ b/aten/src/ATen/cuda/Sleep.cu
@ -1,6 +1,7 @@
 #include <ATen/cuda/CUDAContextLight.h>
 #include <ATen/cuda/Sleep.h>

+#include <c10/cuda/CUDACachingAllocator.h>
 #include <c10/cuda/CUDAException.h>
 #include <c10/cuda/CUDAStream.h>

@ -24,8 +25,22 @@ __global__ void spin_kernel(int64_t cycles) {
 #endif
  }
 }
+
+thread_local int *flag = nullptr;
+
+__global__ void busy_wait_for_flag_kernel(int *flag) {
+  atomicExch(flag, 1);
+  while (atomicAdd(flag, 0) == 1) {
+    // do nothing
+  }
 }

+__global__ void clear_flag_kernel(int *flag) {
+  atomicExch(flag, 0);
+}
+
+} // anonymous namespace
+
 void sleep(int64_t cycles) {
  dim3 grid(1);
  dim3 block(1);
@ -33,6 +48,26 @@ void sleep(int64_t cycles) {
  C10_CUDA_KERNEL_LAUNCH_CHECK();
 }

+void busy_wait_for_flag() {
+  if (!flag) {
+    flag = (int*)c10::cuda::CUDACachingAllocator::raw_alloc(sizeof(int));
+  }
+  dim3 grid(1);
+  dim3 block(1);
+  busy_wait_for_flag_kernel<<<grid, block, 0, c10::cuda::getCurrentCUDAStream()>>>(flag);
+  C10_CUDA_KERNEL_LAUNCH_CHECK();
+}
+
+void clear_flag() {
+  if (!flag) {
+    flag = (int*)c10::cuda::CUDACachingAllocator::raw_alloc(sizeof(int));
+  }
+  dim3 grid(1);
+  dim3 block(1);
+  clear_flag_kernel<<<grid, block, 0, c10::cuda::getCurrentCUDAStream()>>>(flag);
+  C10_CUDA_KERNEL_LAUNCH_CHECK();
+}
+
 #ifdef USE_ROCM
 __global__ void flush_icache_kernel()
 {
--- a/aten/src/ATen/cuda/Sleep.h
+++ b/aten/src/ATen/cuda/Sleep.h
@ -7,6 +7,11 @@ namespace at::cuda {
 // enqueues a kernel that spins for the specified number of cycles
 TORCH_CUDA_CU_API void sleep(int64_t cycles);

+// enqueues a kernel that spins until a flag is cleared by a
+// corresponding call to clear_flag()
+TORCH_CUDA_CU_API void busy_wait_for_flag();
+TORCH_CUDA_CU_API void clear_flag();
+
 // flushes instruction cache for ROCm; no-op for CUDA
 TORCH_CUDA_CU_API void flush_icache();

--- a/aten/src/ATen/cuda/detail/TensorInfo.cuh
+++ b/aten/src/ATen/cuda/detail/TensorInfo.cuh
@ -93,7 +93,7 @@ struct IndexToOffset {
  }
 };

-// Uses dynamic (runtime) instead of static (compiletime) dims
+// Uses dynamic (runtime) instead of static (compile time) dims
 template <typename T, typename IndexType>
 struct IndexToOffset<T, IndexType, -1> {
  static inline __host__ __device__ IndexType get(
--- a/aten/src/ATen/cuda/jiterator.cu
+++ b/aten/src/ATen/cuda/jiterator.cu
@ -32,7 +32,7 @@ static inline void launch_jitted_vectorized_kernel_dynamic(

  // Different kernels are compiled depending on what we're vectorizing up to (1, 2 or 4 elements)
  //   fn_ptr is set to the appropriate function based on the vec size and GPU used
-  // TODO: Memory use can probably be optimized by re-using kernels across GPUs with
+  // TODO: Memory use can probably be optimized by reusing kernels across GPUs with
  //   the same compute capability

  std::string f_inputs_type_str = at::cuda::jit::typeName(common_dtype);
--- a/aten/src/ATen/cuda/tunable/Tunable.cpp
+++ b/aten/src/ATen/cuda/tunable/Tunable.cpp
@ -580,7 +580,7 @@ std::ofstream& TuningContext::GetUntunedFile(){
      filename.append(device);
    }

-    untuned_file_ = std::ofstream(filename, std::ios::out | std::ios::trunc);
+    untuned_file_ = std::ofstream(filename, std::ios::out | std::ios::app);
  }
  return untuned_file_;
 }
--- a/aten/src/ATen/functorch/LegacyVmapTransforms.h
+++ b/aten/src/ATen/functorch/LegacyVmapTransforms.h
@ -143,7 +143,7 @@ struct TORCH_API VmapPhysicalView {
  // mapping a physical tensor to a new logical tensor (BatchedTensor)
  VmapPhysicalToLogicalMap getPhysicalToLogicalMap() const;

-  // Maps a logical shape to a physical shape by pre-pending the batch
+  // Maps a logical shape to a physical shape by prepending the batch
  // sizes to the logical shape.
  VmapDimVector getPhysicalShape(IntArrayRef logical_shape) const;
  SymDimVector getPhysicalShape(c10::SymIntArrayRef logical_shape) const;
--- a/aten/src/ATen/functorch/TensorWrapper.h
+++ b/aten/src/ATen/functorch/TensorWrapper.h
@ -27,7 +27,7 @@ namespace at::functorch {
 //
 // There are alternative designs we could have chosen (e.g. each grad transform
 // stores a weak map of Tensor -> AutogradMeta); the benefit of the TensorWrapper
-// design is that we can re-use existing VariableType kernels (i.e. Autograd kernels)
+// design is that we can reuse existing VariableType kernels (i.e. Autograd kernels)
 // without much modification. Since a TensorWrapper looks like a regular Tensor,
 // the VariableType kernel can pull out the AutogradMeta struct from where it
 // expects and extend the autograd graph
--- a/aten/src/ATen/native/Convolution.cpp
+++ b/aten/src/ATen/native/Convolution.cpp
@ -410,8 +410,8 @@ struct ConvParams {
      return false;
    }
    static long cudnn_version = detail::getCUDAHooks().versionCuDNN();
-    // broken on cuDNN 9.8
-    if (cudnn_version >= 90800) {
+    // broken on cuDNN 9.8 - 9.14
+    if (cudnn_version >= 90800 && cudnn_version < 91500) {
      if (cudnn_conv_suggest_memory_format(input, weight) == at::MemoryFormat::Contiguous &&
          (input.scalar_type() == at::kBFloat16 || input.scalar_type() == at::kHalf) &&
          weight.dim() == 5) {
@ -689,6 +689,10 @@ static void check_shape_forward(const at::Tensor& input,
             ", but got bias of size ", at::symint::sizes<T>(bias), " instead");

    for (const auto i : c10::irange(2, k)) {
+      // T could be int64_t or SymInt, Specialized numeric_limts<SymInt> in c10/core/SymInt.h
+      TORCH_CHECK(padding[i-2] <= (std::numeric_limits<T>::max() - padding[i-2]),
+                  "Given padding=", padding[i-2], " at dimension ", i-2, " , expected padding to be at most ",
+                  (std::numeric_limits<T>::max() / 2));
      input_shape.push_back(at::symint::size<T>(input, i) + 2 * padding[i-2]);
      // log new kernel size considering dilation
      kernel_shape.push_back(dilation[i-2] * (weight_sizes[i]-1) + 1);
@ -715,6 +719,11 @@ static void check_shape_forward(const at::Tensor& input,
               "Kernel size: (", kernel_ss.str(), "). Kernel size can't be greater than actual input size");
    }
  } else { // transposed
+    for (const auto i : c10::irange(2, k)) {
+      TORCH_CHECK(padding[i-2] <= (std::numeric_limits<T>::max() - padding[i-2]),
+                  "Given padding=", padding[i-2], " at dimension ", i-2, " , expected padding to be at most ",
+                  (std::numeric_limits<T>::max() / 2));
+    }
    TORCH_CHECK(at::symint::size<T>(input, 1) == weight_sizes[0],
             "Given transposed=", transposed, ", weight of size ", weight_sizes,
             ", expected input", at::symint::sizes<T>(input), " to have ", weight_sizes[0],
--- a/aten/src/ATen/native/ConvolutionTBC.cpp
+++ b/aten/src/ATen/native/ConvolutionTBC.cpp
@ -52,8 +52,7 @@ Tensor conv_tbc(const Tensor& self, const Tensor& weight, const Tensor& bias, in
  for (const auto k : c10::irange(kw)) {
    int iShift = std::max(0, static_cast<int>(k - real_pad));
    int oShift = std::max(0, static_cast<int>(real_pad - k));
-    // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-    int t = std::min(ilen + real_pad - k, olen) - oShift;
+    long t = std::min(ilen + real_pad - k, olen) - oShift;
    // Note: gemm assumes column-major matrices
    // input    is l*m (row-major)
    // weight   is m*r (row-major)
--- a/aten/src/ATen/native/IndexingUtils.cpp
+++ b/aten/src/ATen/native/IndexingUtils.cpp
@ -16,8 +16,7 @@ bool canUse32BitIndexMath(const TensorBase& t, int64_t max_elem) {
  auto linearId = elements - 1;

  // NOTE: Assumes all strides are positive, which is true for now
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  for (int i = t.dim() - 1; i >= 0; --i) {
+  for (auto i = t.dim() - 1; i >= 0; --i) {
    auto curDimIndex = linearId % t.sym_size(i);
    auto curDimOffset = curDimIndex * t.sym_stride(i);
    offset += curDimOffset;
--- a/aten/src/ATen/native/QuantizedLinear.cpp
+++ b/aten/src/ATen/native/QuantizedLinear.cpp
@ -68,7 +68,6 @@ Tensor fbgemm_linear_int8_weight_fp32_activation(
  const float* input_ptr = input_contig.const_data_ptr<float>();

  TORCH_CHECK(input.dim() >= 2);
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
  const int64_t M = size_to_dim_(input.dim() - 1, input.sizes());
  const int64_t K = input.size(input.dim() - 1);
  TORCH_CHECK(weight.dim() == 2);
--- a/aten/src/ATen/native/cpu/DistanceOpsKernel.cpp
+++ b/aten/src/ATen/native/cpu/DistanceOpsKernel.cpp
@ -160,10 +160,9 @@ struct Dist {
    // value of k.
    parallel_for(0, combs, internal::GRAIN_SIZE / (16 * m), [p, self_start, self_end, n, m, res_start](int64_t k, int64_t end) {
      const Vec pvec(p);
-      double n2 = n - .5;
+      double n2 = static_cast<double>(n) - .5;
      // The -1 accounts for floating point truncation issues
-      // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-      int64_t i = static_cast<int64_t>((n2 - std::sqrt(n2 * n2 - 2 * k - 1)));
+      int64_t i = static_cast<int64_t>((n2 - std::sqrt(n2 * n2 - 2.0 * static_cast<double>(k) - 1.0)));
      int64_t j = k - n * i + i * (i + 1) / 2 + i + 1;

      const scalar_t * self_i = self_start + i * m;
--- a/aten/src/ATen/native/cpu/UpSampleKernel.cpp
+++ b/aten/src/ATen/native/cpu/UpSampleKernel.cpp
@ -1017,7 +1017,7 @@ struct HelperInterpBase {
      while (aligned_interp_size % sizeof(int32_t) != 0) {
        aligned_interp_size += 1;
      }
-      // assert that we wont go out of bounds
+      // assert that we won't go out of bounds
      TORCH_INTERNAL_ASSERT(aligned_interp_size * sizeof(int16_t) < interp_size * sizeof(double));
    }

--- a/aten/src/ATen/native/cpu/UpSampleKernelAVXAntialias.h
+++ b/aten/src/ATen/native/cpu/UpSampleKernelAVXAntialias.h
@ -655,7 +655,7 @@ void ImagingResampleHorizontalConvolution8u4x(
      // last element
      auto mmk = _mm256_set1_epi32(k[i]);
      // For num_channels == 3 (3 bytes = one pixel) we tolerate to read 4 bytes
-      // lines 0, 1 and 2 wont go out of allocated memory bounds
+      // lines 0, 1 and 2 won't go out of allocated memory bounds
      auto pix = _mm256_inserti128_si256(_mm256_castsi128_si256(
          mm_cvtepu8_epi32(lineIn0_min + stride * i, i32_aligned)),
          mm_cvtepu8_epi32(lineIn1_min + stride * i, i32_aligned), 1);
@ -1312,7 +1312,7 @@ void ImagingResampleVerticalConvolution8u(

    // Here we write 4 bytes to the output even if num_channels < 4, e.g o = {r,g,b,X} for num_channels=3
    // It is OK to write 4th byte (e.g. X) as on the next step we will overwrite it with new data.
-    // We also wont go out of bounds of lineOut memory allocation
+    // We also won't go out of bounds of lineOut memory allocation
    std::memcpy(lineOut + j, (uint8_t *) &o, 4);
  }

--- a/aten/src/ATen/native/cuda/AdaptiveAveragePooling.cu
+++ b/aten/src/ATen/native/cuda/AdaptiveAveragePooling.cu
@ -705,7 +705,7 @@ namespace {
            );
        } while (!done && max_threads);
        if (!done) {
-          TORCH_INTERNAL_ASSERT(false, "Couldn't reduce launch bounds to accomodate sharedMemPerBlock limit");
+          TORCH_INTERNAL_ASSERT(false, "Couldn't reduce launch bounds to accommodate sharedMemPerBlock limit");
        }
        break;
      }
--- a/aten/src/ATen/native/cuda/CUDAJitLoops.cuh
+++ b/aten/src/ATen/native/cuda/CUDAJitLoops.cuh
@ -298,7 +298,7 @@ static void jitted_gpu_kernel_impl(
    at::opmath_type<f_inputs_type> scalar_val,
    const std::tuple<ExtraArgs...>& extra_args) {

-  // TODO: Memory use can probably be optimized by re-using kernels across GPUs with
+  // TODO: Memory use can probably be optimized by reusing kernels across GPUs with
  //   the same compute capability
  static std::mutex jiterator_mutex;
  static std::vector<JittedKernelVariantCache> device_caches(c10::cuda::device_count());
--- a/aten/src/ATen/native/cuda/Dropout.cu
+++ b/aten/src/ATen/native/cuda/Dropout.cu
@ -75,7 +75,7 @@ fused_dropout_kernel_vec(at::cuda::detail::TensorInfo<const scalar_t, IndexType>
    // We'll use this to actually cause vectorized loads later
    LoadT *value = reinterpret_cast<LoadT*>(&src);

-    //curand_uniform_double was pure evil anyway, not doing what it promises, and there's nothing for halfs, so generate float for everything
+    //curand_uniform_double was pure evil anyway, not doing what it promises, and there's nothing for Halfs, so generate float for everything
    // Note: need a new set of random values per 4 elements -- we'll handle VEC elements in this thread, so need ceil(VEC / 4)
    // sets of rand.
    if ((VEC >= 4) || (gridxvec_loop_state == 0)) {
@ -159,7 +159,7 @@ fused_dropout_kernel(cuda::detail::TensorInfo<const scalar_t, IndexType> a,
  for (IndexType linearIndex = idx;
       linearIndex < rounded_size;
       linearIndex += gridDim.x * blockDim.x*UNROLL) {
-//curand_uniform_double was pure evil anyway, not doing what it promises, and there's nothing for halfs, so generate float for everything
+//curand_uniform_double was pure evil anyway, not doing what it promises, and there's nothing for Halfs, so generate float for everything
       float4 rand = curand_uniform4(&state);
       scalar_t src[UNROLL];
       rand.x = rand.x < p;
--- a/aten/src/ATen/native/cuda/EmbeddingBackwardKernel.cu
+++ b/aten/src/ATen/native/cuda/EmbeddingBackwardKernel.cu
@ -24,7 +24,7 @@ namespace at::native {
 namespace {

 /* This code computes the sum of the weights in two-steps:
-  1) Each GPU warp sums `NROWS_PER_THREAD` number of row given by `indeces`
+  1) Each GPU warp sums `NROWS_PER_THREAD` number of row given by `indices`
  2) Each partial-sum from 1) are summed and scatter into `grad_weight`

  Notice, `NROWS_PER_THREAD` impacts the Achieved Occupancy of the
--- a/aten/src/ATen/native/cuda/ForeachBinaryOpScalar.cu
+++ b/aten/src/ATen/native/cuda/ForeachBinaryOpScalar.cu
@ -204,7 +204,7 @@ Scalar scalar_reciprocal(const Scalar& scalar) {
    return Scalar(1. / scalar.toComplexDouble());
  }
  TORCH_INTERNAL_ASSERT(
-      false, "divison with ", scalar.type(), " not supported");
+      false, "division with ", scalar.type(), " not supported");
 }

 void foreach_tensor_div_scalar_kernel_cuda_(
--- a/aten/src/ATen/native/cuda/GridSampler.cu
+++ b/aten/src/ATen/native/cuda/GridSampler.cu
@ -57,7 +57,7 @@ namespace {
      const index_t n = index / (out_H * out_W);
      const index_t grid_offset = n * grid_sN + h * grid_sH + w * grid_sW;

-      // get the corresponding input x, y co-ordinates from grid
+      // get the corresponding input x, y coordinates from grid
      opmath_t x = grid.data[grid_offset];
      opmath_t y = grid.data[grid_offset + grid_sCoor];

@ -193,7 +193,7 @@ namespace {
      const index_t n = index / (out_D * out_H * out_W);
      const index_t grid_offset = n * grid_sN + d * grid_sD + h * grid_sH + w * grid_sW;

-      // get the corresponding input x, y, z co-ordinates from grid
+      // get the corresponding input x, y, z coordinates from grid
      opmath_t x = grid.data[grid_offset];
      opmath_t y = grid.data[grid_offset + grid_sCoor];
      opmath_t z = grid.data[grid_offset + 2 * grid_sCoor];
@ -358,7 +358,7 @@ namespace {
      const index_t n = index / (out_H * out_W);
      const auto grid_offset = n * grid_sN + h * grid_sH + w * grid_sW;

-      // get the corresponding input x, y co-ordinates from grid
+      // get the corresponding input x, y coordinates from grid
      scalar_t x = grid.data[grid_offset];
      scalar_t y = grid.data[grid_offset + grid_sCoor];

@ -572,7 +572,7 @@ namespace {
      const index_t n = index / (out_D * out_H * out_W);
      const auto grid_offset = n * grid_sN + d * grid_sD + h * grid_sH + w * grid_sW;

-      // get the corresponding input x, y, z co-ordinates from grid
+      // get the corresponding input x, y, z coordinates from grid
      scalar_t ix = grid.data[grid_offset];
      scalar_t iy = grid.data[grid_offset + grid_sCoor];
      scalar_t iz = grid.data[grid_offset + 2 * grid_sCoor];
--- a/aten/src/ATen/native/cuda/GroupMM.cu
+++ b/aten/src/ATen/native/cuda/GroupMM.cu
@ -8,7 +8,7 @@
 #include <c10/util/irange.h>


-// Three warninngs in Cutlass included header files
+// Three warnings in Cutlass included header files
 C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wset-but-not-used")
 C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-but-set-parameter")
 C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-but-set-variable")
--- a/aten/src/ATen/native/cuda/GroupedBlas.cpp
+++ b/aten/src/ATen/native/cuda/GroupedBlas.cpp
@ -208,6 +208,62 @@ _f8_f8_bf16_rowwise_grouped_mm(
 #endif
 }

+Tensor&
+_f4_f4_bf16_grouped_mm_fbgemm(
+      const Tensor& mat_a,
+      const Tensor& mat_b,
+      const Tensor& scale_a,
+      const std::optional<Tensor>& global_scale_a,
+      const Tensor& scale_b,
+      const std::optional<Tensor>& global_scale_b,
+      const std::optional<Tensor>& offs,
+      const std::optional<Tensor>& bias,
+      Tensor& out) {
+#if !defined(USE_ROCM) && defined(USE_FBGEMM_GENAI)
+  // Typing checks
+  TORCH_CHECK_VALUE(mat_a.scalar_type() == at::kFloat4_e2m1fn_x2,
+      "mat_a must be Float4_e2n1fn_2, got: ", mat_a.scalar_type());
+  TORCH_CHECK_VALUE(mat_b.scalar_type() == at::kFloat4_e2m1fn_x2,
+      "mat_b must be Float4_e2n1fn_2, got: ", mat_b.scalar_type());
+
+  std::optional<Tensor> combined_global_scale = std::nullopt;
+  if (global_scale_a.has_value() || global_scale_b.has_value()) {
+      // NVFP4
+      TORCH_CHECK_VALUE(global_scale_a.has_value() && global_scale_b.has_value(),
+          "For NVFP4 grouped gemm both of global_scale_{a,b} must have values")
+      TORCH_CHECK_VALUE(scale_a.scalar_type() == at::kFloat8_e4m3fn,
+          "scale_a must be Float8_e4m3fn, got: ", scale_a.scalar_type());
+      TORCH_CHECK_VALUE(scale_b.scalar_type() == at::kFloat8_e4m3fn,
+          "scale_b must be Float8_e4m3fn, got: ", scale_b.scalar_type());
+      TORCH_CHECK_VALUE(global_scale_a.value().scalar_type() == at::kFloat,
+          "global_scale_a must be Float, got: ", global_scale_a.value().scalar_type());
+      TORCH_CHECK_VALUE(global_scale_b.value().scalar_type() == at::kFloat,
+          "global_scale_b must be Float, got: ", global_scale_b.value().scalar_type());
+      combined_global_scale = global_scale_a.value().mul(global_scale_b.value());
+  } else {
+      // MXFP4
+      TORCH_CHECK_VALUE(scale_a.scalar_type() == at::kFloat8_e8m0fnu,
+          "scale_a must be Float8_e8m0fnu, got: ", scale_a.scalar_type());
+      TORCH_CHECK_VALUE(scale_b.scalar_type() == at::kFloat8_e8m0fnu,
+          "scale_b must be Float8_e8m0fnu, got: ", scale_b.scalar_type());
+  }
+
+  auto o = fbgemm_gpu::f4f4bf16_grouped_mm(
+      mat_a,
+      mat_b,
+      scale_a,
+      scale_b,
+      offs.value(),
+      out,
+      combined_global_scale
+  );
+#else
+  TORCH_CHECK_NOT_IMPLEMENTED(false, "nvfp4 grouped gemm is not supported without USE_FBGEMM_GENAI, and only for CUDA")
+#endif
+
+  return out;
+}
+
 void _check_scales_fp8_rowwise(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx, const int scale_multiplier=1) {
  // Checks scales for 2d or 3d target tensors (`mat`).
  if (mat.dim() == 2) {
@ -245,7 +301,15 @@ void _check_scales_fp8_rowwise(const Tensor& mat, const Tensor& scale, const int
  }
 }

-void _check_scales_mxfp8(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx) {
+void _check_scales_blocked(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx) {
+  // if {mx,nv}fp4, will need to modify K later
+  bool is_fp4 = (mat.scalar_type() == kFloat4_e2m1fn_x2);
+  int blocksize = 32;
+  // check for nvfp4 vs. mxfp4 to fix blocksize
+  if (is_fp4 && scale.scalar_type() == kFloat8_e4m3fn) {
+    blocksize = 16;
+  }
+
  // Checks scales for 2d or 3d target tensors (`mat`).
  if (mat.dim() == 2) {
    // For MXFP8, 2d tensors have variable size groups represented as subtensors,
@ -253,17 +317,19 @@ void _check_scales_mxfp8(const Tensor& mat, const Tensor& scale, const int dim,
    // so we can't check the scale sizes without doing a d2h sync to get the group sizes here.
    TORCH_CHECK(
      scale.dim() == mat.dim(),
-      "for mxfp8, scale must have same number of dimensions as parent tensor, but got mat.dim() = ", mat.dim(), " and scale.dim() = ", scale.dim(), " for arg ", arg_idx);
+      "for block-scaled, scale must have same number of dimensions as parent tensor, but got mat.dim() = ", mat.dim(),
+      " and scale.dim() = ", scale.dim(), " for arg ", arg_idx
+    );

-    // LHS mat shape (M, total_K) -> scale shape (rounded_up(M, 128), rounded_up_per_group(K/32, 4))
-    // RHS mat shape (total_K, N) -> scale shape (rounded_up(N, 128), rounded_up_per_group(K/32, 4))
+    // LHS mat shape (M, total_K) -> scale shape (rounded_up(M, 128), rounded_up_per_group(K/blocksize, 4))
+    // RHS mat shape (total_K, N) -> scale shape (rounded_up(N, 128), rounded_up_per_group(K/blocksize, 4))
    //   * weight is transposed prior to the call, scale stays non-transposed.
    bool LHS = arg_idx == 0;
    int scale_dim_to_check = 0;
    int mat_dim_to_check = LHS ? 0 : 1;
    TORCH_CHECK(
        scale.size(scale_dim_to_check) >= mat.size(mat_dim_to_check),
-        "for mxfp8, arg ", arg_idx, " tensor shape (", mat.size(0), ", ", mat.size(1), ") ",
+        "for block-scaled, arg ", arg_idx, " tensor shape (", mat.size(0), ", ", mat.size(1), ") ",
        "must have scale.shape[", scale_dim_to_check, "] >= ", mat.size(mat_dim_to_check), " but got scale.shape=(", scale.size(0), ", ", scale.size(1), ")");
  } else {
    // For MXFP8, 3d tensors have static group sizes (stack of 2d tensors),
@ -273,32 +339,40 @@ void _check_scales_mxfp8(const Tensor& mat, const Tensor& scale, const int dim,
    };

    // TODO: this is for 3d tensor in 2d-3d case specifically.
-    // We'll need to support 3d-3d and 3d-2d cases once mxfp8 grouped gemm supports them.
+    // We'll need to support 3d-3d and 3d-2d cases once mxfp8/nvfp4 grouped gemm supports them.
    int64_t G = mat.size(0);
    int64_t K = mat.size(1);
+    if (is_fp4) {
+      // FP4 packs 2 values into a single 8b word - the "real" K is 2x the
+      // reported K. Reverse that adjustment.
+      const int fp4_elems_per_byte = 2;
+      K *= fp4_elems_per_byte;
+    }
    int64_t N = mat.size(2);
-    int64_t blocked_scale_K = round_up(K/32, 4);
+    int64_t blocked_scale_K = round_up(K/blocksize, 4);
    int64_t blocked_scale_N = round_up(N, 128);

    // fbgemm expects stack of flattened blocked scales for 3d tensor, shape (G, blocked_scale_K * blocked_scale_N).
    TORCH_CHECK(
      scale.dim() == mat.dim() - 1,
-      "for mxfp8 2d-3d grouped GEMM, the 3d tensor of shape (G,K,N) must have a 2d scale of shape (G, blocked_scale_K * blocked_scale_N), but scale is ", scale.dim(), "D for arg ", arg_idx
+      "for block-scaled 2d-3d grouped GEMM, the 3d tensor of shape (G,K,N) must have a 2d scale of shape (G, blocked_scale_K * blocked_scale_N),",
+      "but scale is ", scale.dim(), "D for arg ", arg_idx
    );
    TORCH_CHECK(
      scale.size(0) == G && scale.size(1) == blocked_scale_K * blocked_scale_N,
-      "for mxfp8, the tensor shape (", G, ", ", K, ", ", N, ") must have scale shape (", G, ",", blocked_scale_K, ",", blocked_scale_N, ") for arg ", arg_idx
+      "for block-scaled grouped GEMM, the tensor shape (", G, ", ", K, ", ", N, ") must have scale shape (", G, ",", blocked_scale_K, ",", blocked_scale_N, ")",
+      " for arg ", arg_idx, ", got: ", scale.size(0), ", ", scale.size(1)
    );
  }
 }

 void check_scale(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx, const int scale_multiplier=1) {
  bool using_fp8_rowwise = scale.scalar_type() == kFloat;
-  bool using_mxfp8 = scale.scalar_type() == at::kFloat8_e8m0fnu;
+  bool using_mx = scale.scalar_type() == at::kFloat8_e8m0fnu;
  if (using_fp8_rowwise) {
    _check_scales_fp8_rowwise(mat, scale, dim, arg_idx, scale_multiplier);
-  } else if (using_mxfp8) {
-    _check_scales_mxfp8(mat, scale, dim, arg_idx);
+  } else if (using_mx) {
+    _check_scales_blocked(mat, scale, dim, arg_idx);
  } else {
    TORCH_CHECK(false, "scale must be float32 or float8_e8m0fnu, but got ", scale.dtype());
  }
@ -411,9 +485,11 @@ namespace {

 using acceptance_fn = std::function<bool(c10::ScalarType, std::vector<ScalingType>&, ArrayRef<Tensor>&, c10::ScalarType, std::vector<ScalingType>&, ArrayRef<Tensor>&)>;

-std::array<std::tuple<std::string, acceptance_fn, ScaledGemmImplementation>, 2> scale_grouped_kernel_dispatch = {{
+std::array<std::tuple<std::string, acceptance_fn, ScaledGemmImplementation>, 4> scale_grouped_kernel_dispatch = {{
  { "rowwise_rowwise", scaled_blas::check_rowwise_recipe, ScaledGemmImplementation::ROWWISE_ROWWISE},
-  { "mxfp8_mxfp8", scaled_blas::check_mxfp8_recipe, ScaledGemmImplementation::MXFP8_MXFP8}}};
+  { "mxfp8_mxfp8", scaled_blas::check_mxfp8_recipe, ScaledGemmImplementation::MXFP8_MXFP8},
+  { "mxfp4_mxfp4", scaled_blas::check_mxfp4_recipe, ScaledGemmImplementation::MXFP4_MXFP4},
+  { "nvfp4_nvfp4", scaled_blas::check_nvfp4_recipe, ScaledGemmImplementation::NVFP4_NVFP4}}};

 } // anonymous namespace

@ -449,7 +525,7 @@ _scaled_grouped_mm_cuda_v2(
          "Contraction dimensions (", dim_a, ",", dim_b, ") of mat_a and mat_b must match, got: ", mat_a.size(dim_a), " and ",
          mat_b.size(dim_b));
      // Note: only (-1, -2) is currently supported
-      TORCH_CHECK_VALUE(dim_a == -1 && dim_b == -2, "Curently contraction dims must be (-1, -2) only");
+      TORCH_CHECK_VALUE(dim_a == -1 && dim_b == -2, "Currently contraction dims must be (-1, -2) only");
    } else {
      TORCH_CHECK_VALUE(mat_a.size(-1) == mat_b.size(-2), "contraction dimension of mat_a and mat_b must match");
    }
@ -525,8 +601,9 @@ _scaled_grouped_mm_cuda_v2(
          out);
    }
    case ScaledGemmImplementation::MXFP8_MXFP8: {
-      _check_scales_mxfp8(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
-      _check_scales_mxfp8(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
+      // scale shape checks
+      _check_scales_blocked(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
+      _check_scales_blocked(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
      return _mx8_mx8_bf16_grouped_mm_fbgemm(
          mat_a,
          mat_b,
@ -537,6 +614,36 @@ _scaled_grouped_mm_cuda_v2(
          offs.value(),
          out);
    }
+    case ScaledGemmImplementation::MXFP4_MXFP4: {
+      // scale shape checks
+      _check_scales_blocked(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
+      _check_scales_blocked(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
+      return _f4_f4_bf16_grouped_mm_fbgemm(
+          mat_a,
+          mat_b,
+          scale_a[0], /* block-scale A */
+          std::nullopt, /* global-scale A */
+          scale_b[0], /* block-scale B */
+          std::nullopt, /* global-scale B */
+          offs.value(),
+          std::nullopt, /* bias */
+          out);
+    }
+    case ScaledGemmImplementation::NVFP4_NVFP4: {
+      // scale shape checks
+      _check_scales_blocked(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
+      _check_scales_blocked(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
+      return _f4_f4_bf16_grouped_mm_fbgemm(
+          mat_a,
+          mat_b,
+          scale_a[0], /* block-scale A */
+          scale_a[1], /* global-scale A */
+          scale_b[0], /* block-scale B */
+          scale_b[1], /* global-scale B */
+          offs.value(),
+          std::nullopt, /* bias */
+          out);
+    }
    default:
      TORCH_CHECK_NOT_IMPLEMENTED(false,
          "_scaled_grouped_mm_cuda_v2 is in an inconsistent state - should never reach here");
--- a/aten/src/ATen/native/cuda/IGammaKernel.cu
+++ b/aten/src/ATen/native/cuda/IGammaKernel.cu
@ -377,7 +377,7 @@ __noinline__ __host__ __device__ scalar_t calc_igammac(scalar_t a, scalar_t x) {
   *   result at the boundary
   * - if a is large and a ~ x, then using Uniform Asymptotic Expansions for
   *   Large Parameter (see DLMF 8.12.4 [igam1])
-   * - if x > 1.1 and x < a, using the substraction from the regularized lower
+   * - if x > 1.1 and x < a, using the subtraction from the regularized lower
   *   incomplete gamma
   * - otherwise, calculate the series from [igam2] eq (5)
   */
@ -460,7 +460,7 @@ __noinline__ __host__ __device__ scalar_t calc_igamma(scalar_t a, scalar_t x) {
   *   result at the boundary
   * - if a is large and a ~ x, then using Uniform Asymptotic Expansions for
   *   Large Parameter (see DLMF 8.12.3 [igam1])
-   * - if x > 1 and x > a, using the substraction from the regularized upper
+   * - if x > 1 and x > a, using the subtraction from the regularized upper
   *   incomplete gamma
   * - otherwise, calculate the series from [igam2] eq (4)
   */
--- a/aten/src/ATen/native/cuda/IndexKernel.cu
+++ b/aten/src/ATen/native/cuda/IndexKernel.cu
@ -332,7 +332,7 @@ void cuda_take_put_kernel(
  const auto offset_calc = make_offset_calculator<2>(iter);
  using uindex_t = std::make_unsigned_t<index_t>;

-  // OffsetCalculator needs the sizes and strides reveresed
+  // OffsetCalculator needs the sizes and strides reversed
  const auto indexed_sizes = std::vector<int64_t>(indexed.sizes().rbegin(), indexed.sizes().rend());
  const auto indexed_strides = std::vector<int64_t>(indexed.strides().rbegin(), indexed.strides().rend());
  const auto* indexed_strides_data = indexed_strides.data();
--- a/aten/src/ATen/native/cuda/IndexKernelUtils.cu
+++ b/aten/src/ATen/native/cuda/IndexKernelUtils.cu
@ -13,7 +13,7 @@ __global__ void vectorized_gather_kernel(char * out, char * inp, index_t * idx,
    if (allow_neg_indices) {
        ind = (ind < 0) ? ind + ind_dim_size : ind;
    }
-    CUDA_KERNEL_ASSERT(ind >=0 && ind < ind_dim_size && "vectorized gather kernel index out of bounds");
+    CUDA_KERNEL_ASSERT_VERBOSE(ind >=0 && ind < ind_dim_size && "vectorized gather kernel index out of bounds", "Expected 0 <= index < ind_dim_size(%ld), but got index = %ld", ind_dim_size, ind);
    int32_t off = (blockDim.x * blockIdx.y + threadIdx.x) * Alignment; // off is guaranteed to be within int32 limits
    if (off >= slice_size) return;
    auto vec = at::native::memory::ld_vec<Alignment>(inp + ind * inp_stride + off);
--- a/aten/src/ATen/native/cuda/Indexing.cu
+++ b/aten/src/ATen/native/cuda/Indexing.cu
@ -1611,7 +1611,7 @@ void index_select_out_cuda_impl(

  // SmallIndexKernel is more performant when the number of indices is small, and pre-loading
  // the index reduces memory accesses. When the number of indices is large, we avoid that
-  // and increase parallellism by calling gather_out which is a generalization of index_select
+  // and increase parallelism by calling gather_out which is a generalization of index_select
  if (cuda::detail::canUse32BitIndexMath(out) &&
      cuda::detail::canUse32BitIndexMath(self) &&
      cuda::detail::canUse32BitIndexMath(index) &&
--- a/aten/src/ATen/native/cuda/KernelUtils.cuh
+++ b/aten/src/ATen/native/cuda/KernelUtils.cuh
@ -269,7 +269,7 @@ __device__ __forceinline__ void opportunistic_fastAtomicAdd(

    scalar_t* dst = self_ptr + index;

-    //pack coalseced bf16 and fp16
+    //pack coalesced bf16 and fp16
    if constexpr (std::is_same<scalar_t, c10::BFloat16>::value || std::is_same<scalar_t, c10::Half>::value)
    {
        typedef unsigned short __attribute__((ext_vector_type(2))) vec_short2;
@ -312,7 +312,7 @@ __device__ __forceinline__ void opportunistic_fastAtomicAdd(
        }
    }

-    // not coalsced, so now let try to capture lane-matches...
+    // not coalesced, so now let try to capture lane-matches...

    if (numel > 16 /*<-hueristic threshold*/ * 64 ) {
      // well shucks, unlikely to capture same-dest atomics in a wave.
--- a/aten/src/ATen/native/cuda/LossCTC.cu
+++ b/aten/src/ATen/native/cuda/LossCTC.cu
@ -343,7 +343,7 @@ ctc_loss_backward_log_beta_gpu_kernel(scalar_t* __restrict__ log_beta_data,
  if (input_length == 0)
    return;

-  // "first" row, the beta initialization before eq (10) (t=target_length - differes per batch)
+  // "first" row, the beta initialization before eq (10) (t=target_length - differs per batch)
  for (int64_t block_s = 2*max_target_length - (2*max_target_length % blockDim.x); block_s >= 0; block_s -= blockDim.x) {
    int64_t s = threadIdx.x + block_s;
    scalar_t lb;
--- a/aten/src/ATen/native/cuda/Math.cuh
+++ b/aten/src/ATen/native/cuda/Math.cuh
@ -816,7 +816,7 @@ const auto erfcx_string = jiterator_stringify(
    with the usual checks for overflow etcetera.

    Performance-wise, it seems to be substantially faster than either
-    the SLATEC DERFC function [or an erfcx function derived therefrom]
+    the SLATEC DERFC function [or an erfcx function derived there from]
    or Cody's CALERF function (from netlib.org/specfun), while
    retaining near machine precision in accuracy.
  */
--- a/aten/src/ATen/native/cuda/MemoryAccess.cuh
+++ b/aten/src/ATen/native/cuda/MemoryAccess.cuh
@ -370,7 +370,7 @@ struct vectorized {

 #ifdef USE_ROCM
 // This is similar to vectorized policy above, but this one supports
-// heterogenous input tensor types as templated parameters.
+// heterogeneous input tensor types as templated parameters.
 // Its use should be limited to frequently used heterogeneous data types
 // as each instantiation will generate a separate kernel, leading to code
 // bloating if applied to all combinations supported in PyTorch. Assumption: all
--- a/aten/src/ATen/native/cuda/MultinomialKernel.cu
+++ b/aten/src/ATen/native/cuda/MultinomialKernel.cu
@ -309,7 +309,7 @@ __global__ void sampleMultinomialOnce(
      } else {
        // This should address a rare bug where we don't select a valid index. This likely occurs when
        // due to floating point arithmetic rounding errors, our cumulative sum does not add up to 1, but
-        // and our uniform sample is greater than this value. In this case we likely have unitialized memory
+        // and our uniform sample is greater than this value. In this case we likely have uninitialized memory
        // in dest[curDist]. So basically we will loop through the distribution and pick the largest index
        // where the distribution is non-zero. This is obviously terribly inefficient, but due to the
        // rarity in which this occurs, this should not be an issue.
--- a/aten/src/ATen/native/cuda/Normalization.cuh
+++ b/aten/src/ATen/native/cuda/Normalization.cuh
@ -1654,7 +1654,7 @@ at::Tensor batch_norm_backward_elemt_channels_last_cuda_template(
  const auto stride = input.sizes()[1];
  const auto reduction_size = input.numel() / stride;

-  // Input is guarunteed to be channels-last compatible
+  // Input is guaranteed to be channels-last compatible
  at::Tensor grad_input = at::empty_like(input);

  dim3 block;
@ -1722,7 +1722,7 @@ at::Tensor batch_norm_backward_elemt_channels_last_cuda_template(
  const auto reduction_size = input.numel() / stride;
  auto norm_fct = 1.0 / reduction_size;

-  // Input is guarunteed to be channels-last compatible
+  // Input is guaranteed to be channels-last compatible
  at::Tensor grad_input = at::empty_like(input);

  dim3 block;
--- a/aten/src/ATen/native/cuda/Randperm.cu
+++ b/aten/src/ATen/native/cuda/Randperm.cu
@ -37,7 +37,7 @@ namespace at::native {
 // threshold probability for having non-duplicate keys, then it can be proved that[1]
 // the number of bits required is: ceil(log2(n - (6 n^2 + 1) / (12 log(q))))
 //
-// Then after sort, we lauch a separate kernel that additionally shuffles any islands
+// Then after sort, we launch a separate kernel that additionally shuffles any islands
 // of values whose keys matched. The algorithm of this kernel is as follows:
 // Each thread reads its key and the keys of its neighbors to tell if it's part of an island.
 // For each island, the first thread in the island sees a key match at index i+1 but not index i-1.
--- a/aten/src/ATen/native/cuda/Reduce.cuh
+++ b/aten/src/ATen/native/cuda/Reduce.cuh
@ -1086,12 +1086,12 @@ ReduceConfig setReduceConfig(const TensorIterator& iter){
  // load instructions.
  //
  // Case 1: "vectorize along input"
-  // This case happens when we are reducing along fastest moving dimesion. In such case, threads
+  // This case happens when we are reducing along fastest moving dimension. In such case, threads
  // with the same threadIdx.y works on the same reduction cooperatively and will produce results
  // for the same output. In such case, values in each loaded vector always correspond to the same output.
  //
  // Case 2: "vectorize along output"
-  // This case happens when the fastest moving dimesion is not the dimension of reduction. In such case,
+  // This case happens when the fastest moving dimension is not the dimension of reduction. In such case,
  // threads with different threadIdx.x are independent and will produce results for different outputs.
  // In such case, values in each loaded vector always correspond to different outputs.
  if (fastest_moving_stride == sizeof(scalar_t)) {
--- a/aten/src/ATen/native/cuda/ReflectionPad.cu
+++ b/aten/src/ATen/native/cuda/ReflectionPad.cu
@ -273,7 +273,7 @@ __global__ void reflection_pad2d_backward_det_out_kernel(
        const int64_t dist_cols = ::abs(inp_col - (input_dim_x - 1));

        // we were dist_rows after, now we want to be dist_rows before
-        // we were dist_cols before, now we wnat to be dist_cols after
+        // we were dist_cols before, now we want to be dist_cols after
        const int64_t reflect_tr_out_row = (corner_tr_out_row - dist_rows);
        const int64_t reflect_tr_out_col = (corner_tr_out_col + dist_cols);
        const int64_t reflect_tr_out =
--- a/aten/src/ATen/native/cuda/RowwiseScaledMM.cu
+++ b/aten/src/ATen/native/cuda/RowwiseScaledMM.cu
@ -5,7 +5,7 @@
 #include <ATen/cuda/nvrtc_stub/ATenNVRTC.h>
 #include <c10/macros/Macros.h>

-// Two warninngs in Cutlass included header files
+// Two warnings in Cutlass included header files
 C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wset-but-not-used")
 C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-but-set-parameter")
 C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wmissing-field-initializers")
--- a/aten/src/ATen/native/cuda/ScaledBlas.cpp
+++ b/aten/src/ATen/native/cuda/ScaledBlas.cpp
@ -794,6 +794,24 @@ void _check_deepseek_scale_stride(const Tensor& scale, const Tensor& t, const Sc
  }
 }

+void
+_check_deepseek_support() {
+#ifndef USE_ROCM
+  auto dprops = at::cuda::getCurrentDeviceProperties();
+  if (dprops->major != 9) {
+    // Only on Hopper GPUs
+    TORCH_CHECK_NOT_IMPLEMENTED(
+      dprops->major == 9,
+      "DeepSeek style (1x128, 128x128) scaling only supported in CUDA for SM90")
+  }
+  // Only in cublasLt >= 12.9
+  TORCH_CHECK_NOT_IMPLEMENTED(
+    CUBLAS_VERSION < 120900 || cublasLtGetVersion() < 120900,
+    "DeepSeek style (1x128, 128x128) scaling requires cublasLt >= 12.9"
+  );
+#endif
+}
+
 Tensor&
 _scaled_block1x128_block1x128(
          const Tensor& mat_a, const Tensor& mat_b,
@ -802,8 +820,12 @@ _scaled_block1x128_block1x128(
          const c10::ScalarType out_dtype,
          const bool use_fast_accum,
          Tensor& out) {
+#ifndef USE_ROCM
  // Restrictions:
  // A, B are FP8, scales are fp32, shape K//128
+  // CUDA: Only Hopper GPUs
+  _check_deepseek_support();
+
  TORCH_CHECK_VALUE(isFloat8Type(mat_a.scalar_type()) && isFloat8Type(mat_b.scalar_type()), "mat_a and mat_b must be fp8 types, got: ",
      mat_a.scalar_type(), mat_b.scalar_type());
  TORCH_CHECK_VALUE(scale_a.sizes()[0] == mat_a.sizes()[0] && scale_a.sizes()[1] == mat_a.sizes()[1] / 128 && scale_a.scalar_type() == kFloat,
@ -821,6 +843,12 @@ _scaled_block1x128_block1x128(
  _scaled_gemm(mat_a, mat_b, scale_a, scale_b, scaling_choice_a, scaling_choice_b, bias, use_fast_accum, out);

  return out;
+#else
+  TORCH_CHECK_NOT_IMPLEMENTED(
+    false,
+    "1x128 and 128x128 scaling not available with ROCm"
+  );
+#endif
 }

 Tensor&
@ -831,10 +859,12 @@ _scaled_block128x128_block1x128(
          const c10::ScalarType out_dtype,
          const bool use_fast_accum,
          Tensor& out) {
+#ifndef USE_ROCM
  // Restrictions:
  // A, B are FP8, scales are fp32, shape K//128
-  std::cout << "mat_b: " << mat_b.dim() << ", " << mat_b.sizes() << ", " << mat_b.strides() << std::endl;
-  std::cout << "scale_b: " << scale_b.dim() << ", " << scale_b.sizes() << ", " << scale_b.strides() << std::endl;
+  // CUDA: Only Hopper GPUs
+  _check_deepseek_support();
+
  TORCH_CHECK_VALUE(isFloat8Type(mat_a.scalar_type()) && isFloat8Type(mat_b.scalar_type()), "mat_a and mat_b must be fp8 types, got: ",
      mat_a.scalar_type(), mat_b.scalar_type());
  TORCH_CHECK_VALUE(scale_a.sizes()[0] == ceil_div<int64_t>(mat_a.sizes()[0], 128) && scale_a.sizes()[1] == ceil_div<int64_t>(mat_a.sizes()[1], 128) && scale_a.scalar_type() == kFloat,
@ -852,6 +882,12 @@ _scaled_block128x128_block1x128(
  _scaled_gemm(mat_a, mat_b, scale_a, scale_b, scaling_choice_a, scaling_choice_b, bias, use_fast_accum, out);

  return out;
+#else
+  TORCH_CHECK_NOT_IMPLEMENTED(
+    false,
+    "1x128 and 128x128 scaling not available with ROCm"
+  );
+#endif
 }

 Tensor&
@ -862,8 +898,12 @@ _scaled_block1x128_block128x128(
          const c10::ScalarType out_dtype,
          const bool use_fast_accum,
          Tensor& out) {
+#ifndef USE_ROCM
  // Restrictions:
  // A, B are FP8, scales are fp32, A: shape K//128, B: K//128, N//128
+  // CUDA: Only Hopper GPUs
+  _check_deepseek_support();
+
  TORCH_CHECK_VALUE(isFloat8Type(mat_a.scalar_type()) && isFloat8Type(mat_b.scalar_type()), "mat_a and mat_b must be fp8 types, got: ",
      mat_a.scalar_type(), mat_b.scalar_type());
  TORCH_CHECK_VALUE(scale_a.sizes()[0] == mat_a.sizes()[0] && scale_a.sizes()[1] == mat_a.sizes()[1] / 128 && scale_a.scalar_type() == kFloat,
@ -881,6 +921,12 @@ _scaled_block1x128_block128x128(
  _scaled_gemm(mat_a, mat_b, scale_a, scale_b, scaling_choice_a, scaling_choice_b, bias, use_fast_accum, out);

  return out;
+#else
+  TORCH_CHECK_NOT_IMPLEMENTED(
+    false,
+    "1x128 and 128x128 scaling not available with ROCm"
+  );
+#endif
 }

 Tensor&
--- a/aten/src/ATen/native/cuda/ScaledGroupMM.cu
+++ b/aten/src/ATen/native/cuda/ScaledGroupMM.cu
@ -7,7 +7,7 @@
 #include <c10/macros/Macros.h>
 #include <c10/util/irange.h>

-// Two warninngs in Cutlass included header files
+// Two warnings in Cutlass included header files
 C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wset-but-not-used")
 C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-but-set-parameter")
 C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-but-set-variable")
--- a/aten/src/ATen/native/cuda/ScatterGatherKernel.cu
+++ b/aten/src/ATen/native/cuda/ScatterGatherKernel.cu
@ -160,8 +160,8 @@ struct _cuda_scatter_gather_internal_kernel {
      auto offsets = offset_calc.get(i);

      int64_t idx_dim = *(index_t*)(index_ptr + offsets[2]);
-      CUDA_KERNEL_ASSERT(idx_dim >= 0 && idx_dim < index_size
-        && "scatter gather kernel index out of bounds");
+      CUDA_KERNEL_ASSERT_VERBOSE(idx_dim >= 0 && idx_dim < index_size
+        && "scatter gather kernel index out of bounds", "Expected 0 <= idx_dim < index_size (%ld), but got idx_dim = %ld", index_size, idx_dim);

      f(
        (scalar_t*)(self_ptr + offsets[0]),
@ -406,9 +406,8 @@ struct _cuda_scatter_fill_internal_kernel {
      auto offsets = offset_calc.get(i);

      int64_t idx_dim = *(index_t*)(index_ptr + offsets[1]);
-      CUDA_KERNEL_ASSERT(idx_dim >= 0 && idx_dim < index_size
-        && "index out of bounds"
-      );
+      CUDA_KERNEL_ASSERT_VERBOSE(idx_dim >= 0 && idx_dim < index_size
+        && "index out of bounds", "Expected 0 <= idx_dim < index_size (%ld), but got idx_dim = %ld", index_size, idx_dim);

      f(
        (scalar_t*)(self_ptr + offsets[0]),
--- a/aten/src/ATen/native/cuda/UnaryGeometricTanKernel.cu
+++ b/aten/src/ATen/native/cuda/UnaryGeometricTanKernel.cu
@ -12,14 +12,15 @@

 namespace at::native {

-#if AT_USE_JITERATOR()
+#if 0 && AT_USE_JITERATOR()
 constexpr char tan_name[] = "tan_impl";
 #endif

 void tan_kernel_cuda(TensorIteratorBase& iter) {
  auto common_dtype = iter.common_dtype();
  if (at::isComplexType(common_dtype)) {
-#if AT_USE_JITERATOR()
+    // Disabled due to accuracy issues
+#if 0 && AT_USE_JITERATOR()
    static const auto tan_string = jiterator_stringify(
        template <typename T> T tan_impl(T a) { return std::tan(a); });
    AT_DISPATCH_COMPLEX_TYPES_AND(
--- a/aten/src/ATen/native/cuda/UnaryGeometricTanhKernel.cu
+++ b/aten/src/ATen/native/cuda/UnaryGeometricTanhKernel.cu
@ -12,14 +12,15 @@

 namespace at::native {

-#if AT_USE_JITERATOR()
+#if 0 && AT_USE_JITERATOR()
 constexpr char tanh_name[] = "tanh_impl";
 #endif

 void tanh_kernel_cuda(TensorIteratorBase& iter) {
  auto common_dtype = iter.common_dtype();
  if (at::isComplexType(common_dtype)) {
-#if AT_USE_JITERATOR()
+    // Disabled due to accuracy issues
+#if 0 && AT_USE_JITERATOR()
    static const auto tanh_string = jiterator_stringify(
        template <typename T> T tanh_impl(T a) { return std::tanh(a); });
    AT_DISPATCH_COMPLEX_TYPES_AND(
--- a/aten/src/ATen/native/cuda/group_norm_kernel.cu
+++ b/aten/src/ATen/native/cuda/group_norm_kernel.cu
@ -460,7 +460,7 @@ __global__ void GammaBetaBackwardCUDAKernel2(
    }
  }

-  // Do warp reduce for the 2st 16 cols in the tile.
+  // Do warp reduce for the 2nd 16 cols in the tile.
  sum1 = g_shared[threadIdx.x][threadIdx.y + blockDim.y];
  sum2 = b_shared[threadIdx.x][threadIdx.y + blockDim.y];
  sum1 = cuda_utils::WarpReduceSum<T_ACC>(sum1);
--- a/aten/src/ATen/native/cuda/jit_utils.cpp
+++ b/aten/src/ATen/native/cuda/jit_utils.cpp
@ -1556,19 +1556,19 @@ NvrtcFunction jit_pwise_function(
    ss << "_" << hash_code;
    file_path = ss.str();

-    std::ifstream readin{file_path, std::ios::in | std::ifstream::binary};
-    if (readin.fail()) {
+    std::ifstream read_stream{file_path, std::ios::in | std::ifstream::binary};
+    if (read_stream.fail()) {
      // NOTE: this does not warn because the file might not exist
      // TODO: consider if this should explicitly check for the file's existence or not to throw
      //   an informative warning
-      readin.close();
+      read_stream.close();
    } else {
      // TODO: try passing the "mapped" file directly to cuModuleLoadCall instead of using an intermediate buffer
-      std::vector<char> buffer(std::istreambuf_iterator<char>(readin), {});
+      std::vector<char> buffer(std::istreambuf_iterator<char>(read_stream), {});
      AT_CUDA_DRIVER_CHECK(nvrtc.cuModuleLoadData(&(compiled_kernel_.module), buffer.data()));
      AT_CUDA_DRIVER_CHECK(
        nvrtc.cuModuleGetFunction(&(compiled_kernel_.function), compiled_kernel_.module, name.c_str()));
-      readin.close();
+      read_stream.close();
      return compiled_kernel_;
    }
  }
--- a/aten/src/ATen/native/cuda/layer_norm_kernel.cu
+++ b/aten/src/ATen/native/cuda/layer_norm_kernel.cu
@ -141,7 +141,8 @@ WelfordDataLN cuWelfordOnlineSum(
  if constexpr (!rms_norm){
    U delta = val - curr_sum.mean;
    U new_count = curr_sum.count + 1.f;
-#if defined(USE_ROCM) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
+//Due to low CU count, we run into accuracy issues on gfx90a with `__builtin_amdgcn_rcpf`
+#if defined(USE_ROCM) && !defined(__gfx90a__) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
    U new_mean = curr_sum.mean + delta * __builtin_amdgcn_rcpf(new_count);
 #else
    U new_mean = curr_sum.mean + delta * (1.f/new_count); //proper division is slow, this is less accurate but noticeably faster
@ -163,7 +164,8 @@ WelfordDataLN cuWelfordCombine(
    U count = dataA.count + dataB.count;
    U mean, sigma2;
    if (count > decltype(dataB.count){0}) {
-#if defined(USE_ROCM) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
+//Due to low CU count, we run into accuracy issues on gfx90a with `__builtin_amdgcn_rcpf`
+#if defined(USE_ROCM) && !defined(__gfx90a__) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
      auto coef = __builtin_amdgcn_rcpf(count);
 #else
      auto coef = 1.f/count; //NB we don't use --use_fast_math, but this is emulation, 1./count goes to intrinsic, `* coef` is multiplication, instead of slow fp division
@ -1050,7 +1052,7 @@ void launch_vectorized_layer_norm_kernel(
    C10_CUDA_KERNEL_LAUNCH_CHECK();

 #ifdef USE_ROCM
-    // the blocks.x contains the max grid x dimention without invalid configuration error
+    // the blocks.x contains the max grid x dimension without invalid configuration error
    // Fix invalid configuration https://github.com/pytorch/pytorch/issues/136291
    // Ensure all elements are processed. Prepare for next round
    int64_t remaining = M - blocks.x;
--- a/aten/src/ATen/native/cudnn/MHA.cpp
+++ b/aten/src/ATen/native/cudnn/MHA.cpp
@ -177,7 +177,7 @@ bool use_ragged_in_dense(
    TORCH_WARN_ONCE(
        "TORCH_CUDNN_SDPA_AVOID_RECOMPILE=1 only works with Q, K, V, and output in BSHD memory layout,"
        "e.g., Q, K, V must be allocated with torch.randn((B, S, H, D).transpose(1, 2)."
-        "Falling back to regualr dense case, which may trigger excessive recompilation.");
+        "Falling back to regular dense case, which may trigger excessive recompilation.");
  }
  return all_bshd;
 }
@ -771,7 +771,7 @@ std::unique_ptr<fe::graph::Graph> build_graph_nestedtensor(
  if (attn_bias.has_value()) {
    TORCH_CHECK(
        false,
-        "attn_bias not yet supportd with cuDNN Attention and NestedTensor");
+        "attn_bias not yet supported with cuDNN Attention and NestedTensor");
    scaled_dot_product_flash_attention_options.set_bias(
        mha_graph->tensor(fe::graph::Tensor_attributes()
                              .set_uid(BIAS)
@ -1196,7 +1196,7 @@ std::unique_ptr<fe::graph::Graph> build_graph_backward_nestedtensor(
  if (attn_bias.has_value()) {
    TORCH_CHECK(
        false,
-        "attn_bias not yet supportd with cuDNN Attention and NestedTensor");
+        "attn_bias not yet supported with cuDNN Attention and NestedTensor");
    sdpa_backward_options.set_bias(
        mha_graph->tensor(fe::graph::Tensor_attributes()
                              .set_uid(BIAS)
@ -1864,7 +1864,7 @@ void run_cudnn_SDP_bprop_nestedtensor(
  }
  TORCH_CHECK(
      !attn_bias.has_value(),
-      "attn_bias not yet supportd with cuDNN Attention and NestedTensor");
+      "attn_bias not yet supported with cuDNN Attention and NestedTensor");

  auto workspace_size = mha_graph.get_workspace_size();
  auto workspace_ptr =
--- a/aten/src/ATen/native/hip/ck_bgemm_bfloat16.hip
+++ b/aten/src/ATen/native/hip/ck_bgemm_bfloat16.hip
@ -30,7 +30,7 @@ static const std::unordered_map<
    };


-// This is the heursitic to choose a kernel based on inputs
+// This is the heuristic to choose a kernel based on inputs
 BGEMMKernel_BFloat16 dispatch_bfloat16_bgemm(CUDABLAS_BGEMM_ARGTYPES(at::BFloat16)) {
  // Optional/future use: directly lookup shape tuples to map to instances
  /*
--- a/aten/src/ATen/native/hip/ck_gemm_bfloat16.hip
+++ b/aten/src/ATen/native/hip/ck_gemm_bfloat16.hip
@ -11,7 +11,7 @@ using S = ck::Sequence<Is...>;
 namespace at::native {

 void dispatch_bfloat16_gemm(CUDABLAS_GEMM_ARGTYPES(at::BFloat16)) {
-  // If any of the shapes cant be tiled, we must use padding.
+  // If any of the shapes can't be tiled, we must use padding.
  bool use_padding = ((m % 256 != 0) || (n % 128 != 0) || (k % 64 != 0));
  // Dispatch to best implementation.
  // TODO add more configurations. Optimize.
@ -471,7 +471,7 @@ void dispatch_bfloat16_gemm(CUDABLAS_GEMM_ARGTYPES(at::BFloat16)) {
 }

 void dispatch_bfloat16_gemm_wmma(CUDABLAS_GEMM_ARGTYPES(at::BFloat16)) {
-  // If any of the shapes cant be tiled, we must use padding.
+  // If any of the shapes can't be tiled, we must use padding.
  bool use_padding = ((m % 256 != 0) || (n % 128 != 0) || (k % 64 != 0));
  // Dispatch to best implementation.
  // TODO add more configurations. Optimize.
--- a/aten/src/ATen/native/hip/ck_gemm_float.hip
+++ b/aten/src/ATen/native/hip/ck_gemm_float.hip
@ -11,7 +11,7 @@ using S = ck::Sequence<Is...>;
 namespace at::native {

 void dispatch_float_gemm(CUDABLAS_GEMM_ARGTYPES(float)) {
-  // If any of the shapes cant be tiled, we must use padding.
+  // If any of the shapes can't be tiled, we must use padding.
  bool use_padding = ((m % 256 != 0) || (n % 128 != 0) || (k % 64 != 0));
  // Dispatch to best implementation.
  // TODO add more configurations. Optimize.
--- a/aten/src/ATen/native/hip/ck_gemm_half.hip
+++ b/aten/src/ATen/native/hip/ck_gemm_half.hip
@ -13,7 +13,7 @@ namespace at::native {

 void dispatch_half_gemm(CUDABLAS_GEMM_ARGTYPES(at::Half)) {
 #if 0
-  // If any of the shapes cant be tiled, we must use padding.
+  // If any of the shapes can't be tiled, we must use padding.
  bool use_padding = ((m % 256 != 0) || (n % 128 != 0) || (k % 64 != 0));
  // Dispatch to best implementation.
  // TODO add more configurations. Optimize.
@ -299,7 +299,7 @@ void dispatch_half_gemm(CUDABLAS_GEMM_ARGTYPES(at::Half)) {
 #endif
 }
 void dispatch_half_gemm_wmma(CUDABLAS_GEMM_ARGTYPES(at::Half)) {
-  // If any of the shapes cant be tiled, we must use padding.
+  // If any of the shapes can't be tiled, we must use padding.
  bool use_padding = ((m % 256 != 0) || (n % 128 != 0) || (k % 64 != 0));
  // Dispatch to best implementation.
  // TODO add more configurations. Optimize.
--- a/aten/src/ATen/native/metal/MetalShaders.h
+++ b/aten/src/ATen/native/metal/MetalShaders.h
@ -545,7 +545,7 @@ kernel void reshape(texture2d_array<half, access::read> in_arr[[texture(0), func
    const ushort slices2 = divRoundUp(C2, 4);
    const ushort slices1 = divRoundUp(C1, 4);
    const ushort n2 = gid.z / slices2; //image index
-    const ushort s2 = gid.z - n2 * slices2; // slice offest
+    const ushort s2 = gid.z - n2 * slices2; // slice offset
    half4 value;
    for (int idx = 0; idx < 4; ++idx){
        // we compute the "linear index" of the output element,
--- a/aten/src/ATen/native/metal/ops/MetalNeurons.mm
+++ b/aten/src/ATen/native/metal/ops/MetalNeurons.mm
@ -86,4 +86,4 @@ TORCH_LIBRARY_IMPL(aten, Metal, m) {
  m.impl(TORCH_SELECTIVE_NAME("aten::hardsigmoid_"), TORCH_FN(hardsigmoid_));
 }

-} // namepsace at::native::metal
+} // namespace at::native::metal
--- a/aten/src/ATen/native/mkldnn/xpu/detail/Attr.h
+++ b/aten/src/ATen/native/mkldnn/xpu/detail/Attr.h
@ -34,7 +34,7 @@ namespace at::native::onednn {

 /*
   oneDNN postops usage:
-   Currently, oneDNN supports 5 kinds of post ops. More details can be refered
+   Currently, oneDNN supports 5 kinds of post ops. More details can be referred
 to oneDNN doc.
   https://oneapi-src.github.io/oneDNN/dev_guide_attributes_post_ops.html#doxid-dev-guide-attributes-post-ops-1dev-guide-attributes-post-ops-eltwise

@ -399,7 +399,7 @@ static inline void construct_attr_for_unary(
  } else {
    TORCH_CHECK(
        unary_post_op == "none",
-        "onednn qlinear: unspported unary post op",
+        "onednn qlinear: unsupported unary post op",
        unary_post_op);
  }
 }
--- a/aten/src/ATen/native/mps/OperationUtils.mm
+++ b/aten/src/ATen/native/mps/OperationUtils.mm
@ -856,7 +856,7 @@ id<MTLLibrary> MetalShaderLibrary::getLibrary(const std::initializer_list<std::s
      break;
    }
    default:
-      TORCH_INTERNAL_ASSERT(false, "Unsupported number of paramaters ", nparams);
+      TORCH_INTERNAL_ASSERT(false, "Unsupported number of parameters ", nparams);
  }
  return libMap[key] = lib;
 }
@ -1184,9 +1184,9 @@ void MetalKernelFunction::dispatch(uint64_t length, std::optional<uint64_t> grou
 }

 void MetalKernelFunction::dispatch(c10::ArrayRef<uint64_t> length, c10::OptionalArrayRef<uint64_t> group_size) {
-  TORCH_CHECK(!length.empty() && length.size() < 4, "Dispatch dimentions must be less than 3 and non-empty");
+  TORCH_CHECK(!length.empty() && length.size() < 4, "Dispatch dimensions must be less than 3 and non-empty");
  TORCH_CHECK(!group_size.has_value() || group_size->size() == length.size(),
-              "size and group_size must have same number of dimentions");
+              "size and group_size must have same number of dimensions");
  const auto max_tg_size = getMaxThreadsPerThreadgroup();
  const auto group_size_length = group_size.has_value() ? group_size->size() : 0;
  auto tg_size = MTLSizeMake(group_size_length > 0 ? group_size->at(0) : max_tg_size,
--- a/aten/src/ATen/native/mps/kernels/GridSampler.metal
+++ b/aten/src/ATen/native/mps/kernels/GridSampler.metal
@ -59,7 +59,7 @@ static GridSamplerOffsets find_grid_sampler_offsets(
  return offsets;
 }

-// Mod function which gives postive output when `a` is negative
+// Mod function which gives positive output when `a` is negative
 static int32_t mod(int32_t a, int32_t b) {
  auto r = a % b;
  return r + (r < 0 ? b : 0);
@ -191,9 +191,9 @@ void grid_sampler_single_element(
  int32_t right_indices[3];
  opmath_t<T> scales[3];

-  // For each dimension, find the pair of indices in the cooresponding dimension
+  // For each dimension, find the pair of indices in the corresponding dimension
  // of `input` which surround the grid coordinate in that dimension. We'll do
-  // this by mapping different coordiante spaces onto each other. There are
+  // this by mapping different coordinate spaces onto each other. There are
  // basically three different coordinate spaces to keep in mind:
  //
  //  * aligned grid space
--- a/aten/src/ATen/native/mps/kernels/Indexing.metal
+++ b/aten/src/ATen/native/mps/kernels/Indexing.metal
@ -137,7 +137,7 @@ kernel void index_put_serial(
    constant int64_t* index_strides,
    constant uint4& ndim_nindices_numel,
    uint thread_index [[thread_position_in_grid]]) {
-  (void)thread_index; // Suppress unused vairable varning
+  (void)thread_index; // Suppress unused variable warning
  for (uint idx = 0; idx < ndim_nindices_numel.z; ++idx) {
    index_put_impl(
        output,
--- a/aten/src/ATen/native/mps/kernels/Quantized.metal
+++ b/aten/src/ATen/native/mps/kernels/Quantized.metal
@ -112,7 +112,7 @@ kernel void int4pack_mm(constant T *A [[buffer(0)]],
  constant uchar *B_ptr = B + ((n * K) / k_pack_factor);

  thread float4 result = float4(0.0);
-  // We multipy group of 4 channels with these scales.
+  // We multiply group of 4 channels with these scales.
  // Because corresponding values from weight matrix are effectively left
  // shifted. This is to avoid doing right shift on those values which ends up
  // affecting performance. This is the trick applied in MLX kernels.
--- a/aten/src/ATen/native/mps/kernels/UnaryKernel.metal
+++ b/aten/src/ATen/native/mps/kernels/UnaryKernel.metal
@ -387,7 +387,7 @@ struct log1p_functor {
  }
  template <typename T>
  inline enable_if_t<is_complex_v<T>, T> operator()(const T x) {
-    // TODO: Implement proper log1p algoirthm
+    // TODO: Implement proper log1p algorithm
    auto magnitude = ::precise::sqrt((1.0f + x.x) * (1.0f + x.x) + x.y * x.y);
    auto real = ::precise::log(magnitude);
    auto imag = (x.x == -1 && x.y == 0) ? 0 : ::precise::atan2(x.y, 1.0 + x.x);
--- a/aten/src/ATen/native/mps/kernels/UpSample.metal
+++ b/aten/src/ATen/native/mps/kernels/UpSample.metal
@ -448,7 +448,7 @@ kernel void upsample_trilinear_backward(

 // See Note [ Weights computation for uint8_t and multiplication trick ]
 // Essentially fall back to fixed floating point arithmetic during uint8
-// interpolation, which is not necesserily more accurate (see example below),
+// interpolation, which is not necessarily more accurate (see example below),
 // but matches closes to what CPU can deliver
 // I.e. mid-point 152+249+172+35 is 152, but algorithm yields 153 as horizontal
 // and vertical interpolation is done in separate steps and results are rounded
--- a/aten/src/ATen/native/mps/operations/Indexing.mm
+++ b/aten/src/ATen/native/mps/operations/Indexing.mm
@ -617,6 +617,9 @@ Tensor& index_select_out_mps(const Tensor& self, int64_t dim, const Tensor& inde
  TORCH_CHECK(self.scalar_type() == output.scalar_type(),
              "index_select(): self and output must have the same scalar type");
  TORCH_CHECK(dim == 0 || dim < self.dim(), "index_select(): Indexing dim ", dim, " is out of bounds of tensor");
+  at::assert_no_internal_overlap(output);
+  at::assert_no_overlap(output, self);
+  at::assert_no_overlap(output, index);
  auto output_size = self.sizes().vec();
  if (self.dim() > 0) {
    output_size[dim] = num_indices;
--- a/Show More
+++ b/Show More