Update

[ghstack-poisoned]

.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

.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

.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 -

.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

.ci/docker/ubuntu/Dockerfile

@@ -100,6 +100,8 @@ COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
 COPY ci_commit_pins/timm.txt timm.txt
 COPY ci_commit_pins/torchbench.txt torchbench.txt
+# Only build aoti cpp tests when INDUCTOR_BENCHMARKS is set to True
+ENV BUILD_AOT_INDUCTOR_TEST ${INDUCTOR_BENCHMARKS}
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
 RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
 

.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]

.ci/pytorch/test.sh

@@ -460,28 +460,18 @@ test_inductor_shard() {
     --verbose
 }
 
-test_inductor_aoti() {
-  # docker build uses bdist_wheel which does not work with test_aot_inductor
-  # TODO: need a faster way to build
+test_inductor_aoti_cpp() {
   if [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
     # We need to hipify before building again
     python3 tools/amd_build/build_amd.py
   fi
   if [[ "$BUILD_ENVIRONMENT" == *sm86* ]]; then
-    BUILD_COMMAND=(TORCH_CUDA_ARCH_LIST=8.6 USE_FLASH_ATTENTION=OFF python -m pip install --no-build-isolation -v -e .)
     # TODO: Replace me completely, as one should not use conda libstdc++, nor need special path to TORCH_LIB
     TEST_ENVS=(CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="/opt/conda/envs/py_3.10/lib:${TORCH_LIB_DIR}:${LD_LIBRARY_PATH}")
   else
-    BUILD_COMMAND=(python -m pip install --no-build-isolation -v -e .)
     TEST_ENVS=(CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="${TORCH_LIB_DIR}")
   fi
 
-  # aoti cmake custom command requires `torch` to be installed
-  # initialize the cmake build cache and install torch
-  /usr/bin/env "${BUILD_COMMAND[@]}"
-  # rebuild with the build cache with `BUILD_AOT_INDUCTOR_TEST` enabled
-  /usr/bin/env CMAKE_FRESH=1 BUILD_AOT_INDUCTOR_TEST=1 "${BUILD_COMMAND[@]}"
-
   /usr/bin/env "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference cpp/test_vec_half_AVX2 -dist=loadfile
 }
 
@@ -1776,7 +1766,7 @@ elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
   install_torchvision
   PYTHONPATH=/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
   if [[ "$SHARD_NUMBER" -eq "1" ]]; then
-    test_inductor_aoti
+    test_inductor_aoti_cpp
   fi
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision

.ci/pytorch/windows/internal/install_python.bat

@@ -7,12 +7,9 @@ if "%DESIRED_PYTHON%" == "3.13t" (
     set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.13.0/python-3.13.0-amd64.exe"
     set ADDITIONAL_OPTIONS="Include_freethreaded=1"
     set PYTHON_EXEC="python3.13t"
-) else if "%DESIRED_PYTHON%"=="3.14" (
-    echo Python version is set to 3.14 or 3.14t
-    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0rc1-amd64.exe"
 ) else if "%DESIRED_PYTHON%"=="3.14t" (
     echo Python version is set to 3.14 or 3.14t
-    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0rc1-amd64.exe"
+    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0-amd64.exe"
     set ADDITIONAL_OPTIONS="Include_freethreaded=1"
     set PYTHON_EXEC="python3.14t"
 ) else (

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-69bbe7363897764f9e758d851cd0340147d27f94
+3b0e7a6f192ca2715e7e6cbe5db007aea7165fe2

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-1752fe6809b74921644866275ab80244b96e80bc
+218d2ab791d437309f91e0486eb9fa7f00badc17

.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:
   - '*'

.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

.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'"

.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,

.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 }}

.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

.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

.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

.github/workflows/upload-test-stats.yml

@@ -6,6 +6,7 @@ on:
       - pull
       - trunk
       - periodic
+      - periodic-rocm-mi200
       - periodic-rocm-mi300
       - inductor
       - unstable

.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
 

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)

README.md

@@ -1,4 +1,4 @@
-![PyTorch Logo](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/pytorch-logo-dark.png)
+![PyTorch Logo](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/pytorch-logo-dark.png)
 
 --------------------------------------------------------------------------------
 
@@ -72,7 +72,7 @@ Elaborating Further:
 
 If you use NumPy, then you have used Tensors (a.k.a. ndarray).
 
-![Tensor illustration](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/tensor_illustration.png)
+![Tensor illustration](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/tensor_illustration.png)
 
 PyTorch provides Tensors that can live either on the CPU or the GPU and accelerates the
 computation by a huge amount.
@@ -99,7 +99,7 @@ from several research papers on this topic, as well as current and past work suc
 While this technique is not unique to PyTorch, it's one of the fastest implementations of it to date.
 You get the best of speed and flexibility for your crazy research.
 
-![Dynamic graph](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/dynamic_graph.gif)
+![Dynamic graph](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/dynamic_graph.gif)
 
 ### Python First
 

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}/"
     )
 

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

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__)

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

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()
 {

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();
 

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_;
 }

aten/src/ATen/native/Convolution.cpp

@@ -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],

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)

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;

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);

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;

aten/src/ATen/native/cuda/Blas.cpp

@@ -170,10 +170,14 @@ static bool isInputCompliesAddmmCudaLt(Tensor& result, const Tensor& self, const
   #if defined(CUDA_VERSION) || defined(USE_ROCM)
   const auto scalar_type = mat1.scalar_type();
   return (beta.toComplexDouble() == 1.0
-    // self.dim() == 1 && result.dim() == 2 && self.sizes()[0] == mat2_sizes[1]
-    // is to use lt interface only when self is bias.
-    && self.dim() == 1 && self.sizes()[0] == mat2_sizes[1] && self.is_contiguous()
     && result.dim() == 2 && result.is_contiguous()
+    // Conditions for bias to be fusable
+    && (
+      self.is_contiguous() &&
+      // NOTE: fine to have 1-len dims to the left from the right-most one
+      (self.dim() == 1 || self.squeeze().dim() == 1) &&
+      self.sizes().back() == mat2_sizes[1]
+    )
     && ( // some dtype restrictions
       #ifndef USE_ROCM
       scalar_type == at::ScalarType::Double ||

aten/src/ATen/native/cuda/GroupedBlas.cpp

@@ -208,6 +208,48 @@ _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 Tensor& global_scale_a,
+      const Tensor& scale_b,
+      const 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());
+  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.scalar_type() == at::kFloat,
+      "global_scale_a must be Float, got: ", global_scale_a.scalar_type());
+  TORCH_CHECK_VALUE(global_scale_b.scalar_type() == at::kFloat,
+      "global_scale_b must be Float, got: ", global_scale_b.scalar_type());
+
+  auto o = fbgemm_gpu::f4f4bf16_grouped_mm(
+      mat_a,
+      mat_b,
+      scale_a,
+      scale_b,
+      offs.value(),
+      out,
+      global_scale_a.mul(global_scale_b)
+  );
+#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 +287,15 @@ 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 +303,19 @@ 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 +325,40 @@ 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 +471,10 @@ 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>, 3> 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},
+  { "nvfp4_nvfp4", scaled_blas::check_nvfp4_recipe, ScaledGemmImplementation::NVFP4_NVFP4}}};
 
 } // anonymous namespace
 
@@ -525,8 +586,9 @@ _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 +599,21 @@ _scaled_grouped_mm_cuda_v2(
           offs.value(),
           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");

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(

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(

aten/src/ATen/native/mps/operations/Distributions.mm

@@ -57,6 +57,7 @@ Tensor& random_mps_impl(Tensor& self,
   if (self.numel() == 0) {
     return self;
   }
+  at::assert_no_internal_overlap(self);
   // MPS random is broken for 5D+ tensors, see https://github.com/pytorch/pytorch/issues/147624
   const auto need_reshape = self.ndimension() > 4;
   auto mps_gen = get_generator_or_default<MPSGeneratorImpl>(gen, at::mps::detail::getDefaultMPSGenerator());
@@ -153,8 +154,16 @@ Tensor& random_mps_impl(Tensor& self,
       feeds[meanPlaceholder.getMPSGraphTensor()] = meanPlaceholder.getMPSGraphTensorData();
     }
 
-    Placeholder outputPlaceholder = Placeholder(cachedGraph->resultTensor, self);
+    // Handle non-contiguous output tensors by creating a contiguous temporary
+    const auto needs_gather = needsGather(self);
+    Tensor self_ = needs_gather ? at::empty_like(self, MemoryFormat::Contiguous) : self;
+    Placeholder outputPlaceholder = Placeholder(cachedGraph->resultTensor, self_);
     runMPSGraph(stream, cachedGraph->graph(), feeds, outputPlaceholder);
+
+    // Copy results back to original non-contiguous output
+    if (needs_gather) {
+      self.copy_(self_);
+    }
   }
 
   return self;

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;

aten/src/ATen/native/quantized/cpu/UpSampleBilinear2d.cpp

@@ -73,8 +73,7 @@ void upsample_bilinear2d_out_frame(
   const auto rwidth = area_pixel_compute_scale<float>(
       input_width, output_width, align_corners, scales_w);
 
-  // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
-  float output_scale = output.q_scale() / input.q_scale();
+  float output_scale = static_cast<float>(output.q_scale() / input.q_scale());
 
   const int64_t input_q_zero_point = input.q_zero_point();
   const int64_t output_q_zero_point = output.q_zero_point();

aten/src/ATen/native/quantized/cpu/kernels/QuantizedOpKernels.cpp

@@ -148,7 +148,7 @@ Tensor qcat_nhwc_kernel(
           // Vectorized loop
           if (c + VLEN <= curr_C) {
             auto curr_scale_vec = Vectorized<float>(curr_scale);
-            auto curr_zero_pt_vec = Vectorized<float>((float)curr_zero_pt);
+            auto curr_zero_pt_vec = Vectorized<float>(curr_zero_pt);
             auto scale_neg_zp_premul = curr_scale_vec * curr_zero_pt_vec.neg();
             for (; c + VLEN <= curr_C; c += VLEN) {
               auto inp_vec = Vec::loadu(iptr + c);
@@ -174,7 +174,7 @@ Tensor qcat_nhwc_kernel(
           int64_t elem_size = curr_C - c;
           if ((VLEN == 4 * kVLEN) && elem_size >= kVLEN) {
             auto curr_scale_vec = Vectorized<float>(curr_scale);
-            auto curr_zero_pt_vec = Vectorized<float>((float)curr_zero_pt);
+            auto curr_zero_pt_vec = Vectorized<float>(curr_zero_pt);
             auto scale_neg_zp_premul = curr_scale_vec * curr_zero_pt_vec.neg();
             int64_t vec_num = elem_size / kVLEN;
             std::array<typename scalar_t::underlying, VLEN> buf_in{};
@@ -611,12 +611,10 @@ void qrelu_kernel(const Tensor& qx, Tensor& qy) {
 void leaky_qrelu_out_kernel(Tensor& out, const Tensor& qx,
                                    const Scalar& negval_) {
   int64_t i_zp = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float i_scale = qx.q_scale();
+  float i_scale = static_cast<float>(qx.q_scale());
 
   int64_t o_zp = out.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float o_scale = out.q_scale();
+  float o_scale = static_cast<float>(out.q_scale());
   float o_inv_scale = 1.0f / o_scale;
 
   float negval = negval_.to<float>();
@@ -627,8 +625,8 @@ void leaky_qrelu_out_kernel(Tensor& out, const Tensor& qx,
     Vec zero_vec = Vec(0.0f);
     Vec one_vec = Vec(1.0f);
 
-    Vec i_scale_vec = Vec((float)i_scale);
-    Vec i_zp_vec = Vec((float)i_zp);
+    Vec i_scale_vec = Vec(i_scale);
+    Vec i_zp_vec = Vec(i_zp);
     Vec i_scale_zp_neg_premul_vec = i_scale_vec * i_zp_vec.neg();
 
     Vec negval_vec = Vec(negval);
@@ -738,10 +736,9 @@ void qprelu_out_kernel(Tensor& out,
 
 void qgelu_kernel(const Tensor& qx, Tensor& qy, GeluType approximate) {
   int64_t zero_point = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float scale = qx.q_scale();
+  float scale = static_cast<float>(qx.q_scale());
   auto scale_vec = Vectorized<float>(scale);
-  auto zero_point_vec = Vectorized<float>((float)zero_point);
+  auto zero_point_vec = Vectorized<float>(zero_point);
   auto scale_neg_zp_premul_vec = scale_vec * zero_point_vec.neg();
   int64_t output_zero_point = zero_point;
   float output_scale = scale;
@@ -828,10 +825,9 @@ void qgelu_kernel(const Tensor& qx, Tensor& qy, GeluType approximate) {
 void qsigmoid_kernel(
     const Tensor& qx, Tensor& qy, double output_scale, int64_t output_zero_point ) {
   int64_t zero_point = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float scale = qx.q_scale();
+  float scale = static_cast<float>(qx.q_scale());
   auto scale_vec = Vectorized<float>(scale);
-  auto zero_point_vec = Vectorized<float>((float)zero_point);
+  auto zero_point_vec = Vectorized<float>(zero_point);
 
   AT_DISPATCH_QINT_TYPES(qx.scalar_type(), "qsigmoid", [&]() {
     float inv_output_scale = 1.0 / output_scale;
@@ -870,10 +866,9 @@ void qsigmoid_kernel(
 
 void qhardsigmoid_kernel(const Tensor& qx, Tensor& qy) {
   int64_t zero_point = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float scale = qx.q_scale();
+  float scale = static_cast<float>(qx.q_scale());
   auto scale_vec = Vectorized<float>(scale);
-  auto zero_point_vec = Vectorized<float>((float)zero_point);
+  auto zero_point_vec = Vectorized<float>(zero_point);
   auto scale_neg_zp_premul_vec = scale_vec * zero_point_vec.neg();
 
   AT_DISPATCH_QINT_TYPES(qx.scalar_type(), "qhardsigmoid", [&]() {
@@ -1029,13 +1024,10 @@ void qthreshold_kernel(
 
   // defines input and output scales and zero_points
   int64_t input_zero_point = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float input_scale = qx.q_scale();
+  float input_scale = static_cast<float>(qx.q_scale());
   int64_t output_zero_point = qy.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float output_scale = qy.q_scale();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float inv_output_scale = 1.0 / output_scale;
+  float output_scale = static_cast<float>(qy.q_scale());
+  float inv_output_scale = static_cast<float>(1.0 / output_scale);
 
   AT_DISPATCH_QINT_TYPES(qx.scalar_type(), "qthreshold", [&]() {
     qy = at::_empty_affine_quantized(
@@ -1096,8 +1088,7 @@ void qhardswish_kernel(const Tensor& qx, Tensor& qy) {
 
   const auto o_scale = qy.q_scale();
   const auto o_zero_point = qy.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  const float o_inv_scale = 1.0 / o_scale;
+  const float o_inv_scale = static_cast<float>(1.0 / o_scale);
 
   using fVec = Vectorized<float>;
   fVec i_scale_vec(i_scale);
@@ -1135,10 +1126,9 @@ void qhardswish_kernel(const Tensor& qx, Tensor& qy) {
 
 void qtanh_kernel(const Tensor& qx, Tensor& qy) {
   int64_t zero_point = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float scale = qx.q_scale();
+  float scale = static_cast<float>(qx.q_scale());
   auto scale_vec = Vectorized<float>(scale);
-  auto zero_point_vec = Vectorized<float>((float)zero_point);
+  auto zero_point_vec = Vectorized<float>(zero_point);
   auto scale_neg_zp_premul_vec = scale_vec * zero_point_vec.neg();
 
   AT_DISPATCH_QINT_TYPES(qx.scalar_type(), "qtanh", [&]() {
@@ -1198,16 +1188,13 @@ void qelu_kernel(
   // they are NOT related to the quantization scale term
 
   int64_t i_zp = qx.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float i_scale = qx.q_scale();
+  float i_scale = static_cast<float>(qx.q_scale());
 
   // In a future PR, we can improve on output scale and zero_point
   // selection.
   int64_t o_zp = qy.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float o_scale = qy.q_scale();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float inv_o_scale = 1.0 / o_scale;
+  float o_scale = static_cast<float>(qy.q_scale());
+  float inv_o_scale = static_cast<float>(1.0 / o_scale);
 
   float alpha_float = alpha.to<float>();
   float scale_coef = scale.to<float>();
@@ -1227,7 +1214,7 @@ void qelu_kernel(
     Vec scale_coef_vec = Vec(scale_coef);
     Vec input_scale_coef_vec = Vec(input_scale_coef);
     Vec i_scale_vec = Vec(i_scale);
-    Vec i_zero_point_vec = Vec((float)i_zp);
+    Vec i_zero_point_vec = Vec(i_zp);
     Vec i_scale_neg_zp_premul_vec = i_scale_vec * i_zero_point_vec.neg();
 
     cpu_kernel_vec(
@@ -1326,23 +1313,20 @@ void qadd_scalar_kernel(Tensor& out, const Tensor& self, const Scalar& other) {
 template <bool ReLUFused = false>
 void qadd_kernel(Tensor& out, const Tensor& self, const Tensor& other) {
   int64_t zero_point = out.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float scale = out.q_scale();
+  float scale = static_cast<float>(out.q_scale());
   float inv_scale = 1.0f / scale;
   int64_t self_zero_point = self.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float self_scale = self.q_scale();
+  float self_scale = static_cast<float>(self.q_scale());
   int64_t other_zero_point = other.q_zero_point();
-  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
-  float other_scale = other.q_scale();
+  float other_scale = static_cast<float>(other.q_scale());
 
   // Broadcast out the parameters here to amortize out that cost across
   // loop iterations.
   // TODO: we can optimize dequantization by doing a premultiplication
   // of the zero point by scale and doing FMA on scale*x_q - (scale*zero_point)
-  auto self_zero_point_vec = Vectorized<float>((float)self_zero_point);
+  auto self_zero_point_vec = Vectorized<float>(self_zero_point);
   auto self_scale_vec = Vectorized<float>(self_scale);
-  auto other_zero_point_vec = Vectorized<float>((float)other_zero_point);
+  auto other_zero_point_vec = Vectorized<float>(other_zero_point);
   auto other_scale_vec = Vectorized<float>(other_scale);
 
   auto self_scale_neg_zp_premul_vec = self_scale_vec * self_zero_point_vec.neg();
@@ -2965,7 +2949,7 @@ void quantized_normalize_kernel(
     const bool beta_null = beta_data == nullptr;
     int64_t x_zp = X.q_zero_point();
     float x_scale = X.q_scale();
-    fVec x_zp_vec((float)x_zp);
+    fVec x_zp_vec(x_zp);
     fVec one_vec(1.0f);
     fVec zero_vec(0.0f);
     float x_fake_scale = 1.0f;
@@ -3253,7 +3237,7 @@ void quantized_groupnorm_nhwc_kernel(
     const bool beta_null = beta_data == nullptr;
     int64_t x_zp = X.q_zero_point();
     float x_scale = X.q_scale();
-    fVec x_zp_vec((float)x_zp);
+    fVec x_zp_vec(x_zp);
     fVec one_vec(1.0f);
     fVec zero_vec(0.0f);
     float x_fake_scale = 1.0f;

aten/src/ATen/native/quantized/cpu/qlinear_dynamic.cpp

@@ -414,7 +414,6 @@ at::Tensor& PackedLinearWeightFp16::apply_dynamic_impl(
   TORCH_CHECK(input.size(input.dim() - 1) == packed_weight_fp16.numRows())
   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 N = packed_weight_fp16.numCols();
   std::vector<int64_t> output_sizes = input.sizes().vec();

aten/src/ATen/native/transformers/cuda/flash_attn/flash_api.cpp

@@ -22,6 +22,7 @@
 #else
 #include <ATen/ops/empty.h>
 #include <ATen/ops/empty_like.h>
+#include <ATen/ops/zeros_like.h>
 #include <ATen/ops/reshape.h>
 #include <ATen/ops/scalar_tensor.h>
 #include <ATen/ops/sum.h>
@@ -42,7 +43,6 @@ C10_DIAGNOSTIC_POP()
 #include <static_switch.h>
 #include <ATen/native/transformers/cuda/flash_attn/flash_api.h>
 
-
 #include <c10/util/Exception.h>
 
 namespace FLASH_NAMESPACE {
@@ -417,6 +417,26 @@ mha_fwd(const at::Tensor &q,         // batch_size x seqlen_q x num_heads x head
     const int head_size_og = sizes[3];
     const int seqlen_k = k.size(1);
     const int num_heads_k = k.size(2);
+
+    if (batch_size == 0) {
+        auto opts = q.options();
+        at::Tensor out = at::empty({0, seqlen_q, num_heads, head_size_og}, opts);
+        at::Tensor q_padded = at::empty({0, seqlen_q, num_heads, head_size_og}, opts);
+        at::Tensor k_padded = at::empty({0, seqlen_k, num_heads_k, head_size_og}, opts);
+        at::Tensor v_padded = at::empty({0, seqlen_k, num_heads_k, head_size_og}, opts);
+        at::Tensor softmax_lse = at::empty({0, num_heads, seqlen_q}, opts.dtype(at::kFloat));
+        at::Tensor rng_state = at::empty({2}, at::dtype(c10::kUInt64).device(at::kCUDA));
+        at::Tensor _unused = at::empty({}, at::dtype(c10::kUInt64).device(at::kCUDA));
+        at::Tensor p = at::empty({0}, opts);
+        if (return_softmax) {
+            auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
+            const int seqlen_q_rounded = round_multiple(seqlen_q, 128);
+            const int seqlen_k_rounded = round_multiple(seqlen_k, 128);
+            p = at::empty({0, num_heads, seqlen_q_rounded, seqlen_k_rounded}, opts);
+        }
+        return {std::move(out), std::move(q_padded), std::move(k_padded), std::move(v_padded), std::move(softmax_lse), std::move(rng_state), _unused, std::move(p)};
+    }
+
     TORCH_CHECK(batch_size > 0, "batch size must be positive");
     TORCH_CHECK(head_size_og % 8 == 0, "head_size must be a multiple of 8, this is ensured by padding!");
     TORCH_CHECK(head_size_og <= 256, "FlashAttention forward only supports head dimension at most 256");
@@ -547,7 +567,7 @@ mha_fwd(const at::Tensor &q,         // batch_size x seqlen_q x num_heads x head
         q_padded = q_padded.transpose(1, 2).reshape({batch_size, 1, num_heads_k * seqlen_q, head_size_og});
         softmax_lse = softmax_lse.reshape({batch_size, num_heads_k * seqlen_q, 1});
     }
-    return {out, q_padded, k_padded, v_padded, softmax_lse, rng_state, _unused, p};
+    return {std::move(out), std::move(q_padded), std::move(k_padded), std::move(v_padded), std::move(softmax_lse), std::move(rng_state), std::move(_unused), std::move(p)};
 }
 
 std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>
@@ -852,7 +872,6 @@ mha_bwd(const at::Tensor &dout,  // batch_size x seqlen_q x num_heads, x head_si
     TORCH_CHECK(k.stride(-1) == 1, "Input tensor must have contiguous last dimension");
     TORCH_CHECK(v.stride(-1) == 1, "Input tensor must have contiguous last dimension");
     TORCH_CHECK(out.stride(-1) == 1, "out tensor must have contiguous last dimension");
-    TORCH_CHECK(dout.stride(-1) == 1, "dout tensor must have contiguous last dimension");
 
     const auto sizes = q.sizes();
 
@@ -863,6 +882,20 @@ mha_bwd(const at::Tensor &dout,  // batch_size x seqlen_q x num_heads, x head_si
     const int head_size = sizes[3];
     const int seqlen_k = k.size(1);
     const int num_heads_k = k.size(2);
+
+    if (batch_size == 0) {
+        auto opts = q.options();
+        at::Tensor dq = at::empty_like(q);
+        at::Tensor dk = at::empty_like(k);
+        at::Tensor dv = at::empty_like(v);
+        auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
+        const int seqlen_q_rounded = round_multiple(seqlen_q, 128);
+        at::Tensor softmax_d = at::empty({0, num_heads, seqlen_q_rounded}, opts.dtype(at::kFloat));
+        return {dq, dk, dv, softmax_d};
+    }
+
+    TORCH_CHECK(dout.stride(-1) == 1, "dout tensor must have contiguous last dimension");
+
     TORCH_CHECK(batch_size > 0, "batch size must be positive");
     TORCH_CHECK(head_size % 8 == 0, "head_size should be a multiple of 8");
     TORCH_CHECK(head_size_og % 8 == 0, "head_size_og should be a multiple of 8, this is ensured by padding!");

benchmarks/transformer/config_utils.py

@@ -0,0 +1,157 @@
+"""Configuration utilities for parsing JSON and YAML config files."""
+
+import json
+import re
+
+
+def heads_input_type(s: str) -> tuple[int, int]:
+    """Convert string format 'Hq,Hkv' to tuple (Hq, Hkv)."""
+    try:
+        hq, hkv = map(int, s.split(","))
+        return hq, hkv
+    except Exception as e:
+        raise ValueError("Heads must be Hq,Hkv") from e
+
+
+default_config = {
+    "dynamic": False,
+    "calculate_bwd": False,
+    "dtype": "bfloat16",
+    "b": [2, 8, 16],
+    "nh": ["16,16", "16,2"],
+    "s": [512, 1024, 4096],
+    "d": [64, 128],
+    "mods": ["noop", "causal", "alibi", "sliding_window"],
+    "backend": ["efficient"],
+    "max_autotune": False,
+    "decoding": False,
+    "kv_size": None,
+    "throughput": True,
+    "save_path": None,
+    "output_json_for_dashboard": None,
+    "benchmark_name": "PyTorch operator microbenchmark",
+}
+
+
+def load_config_file(config_path: str) -> dict:
+    """Load configuration from JSON or YAML file.
+
+    Automatically converts 'nh' field from strings to tuples.
+
+    Args:
+        config_path: Path to the configuration file
+
+    Returns:
+        Dictionary containing the configuration
+
+    Raises:
+        FileNotFoundError: If config file doesn't exist
+        ValueError: If config file format is invalid
+    """
+    with open(config_path) as f:
+        config_str = f.read()
+
+    # Try to load as JSON first
+    try:
+        config = json.loads(config_str)
+    except json.JSONDecodeError:
+        # Fall back to YAML parsing
+        config = _parse_simple_yaml(config_str)
+
+    # Apply automatic conversions for 'nh' field
+    if "nh" in config and isinstance(config["nh"], list):
+        config["nh"] = [
+            heads_input_type(h) if isinstance(h, str) else h for h in config["nh"]
+        ]
+
+    return config
+
+
+def _parse_simple_yaml(yaml_str: str) -> dict:
+    """Simple YAML parser for basic configs (without external dependencies).
+
+    Supports:
+    - key: value pairs
+    - booleans (true/false)
+    - null values
+    - integers and floats
+    - strings (quoted and unquoted)
+    - lists in JSON format [item1, item2, ...]
+    - comments (lines starting with # or after #)
+
+    Args:
+        yaml_str: YAML content as string
+
+    Returns:
+        Dictionary containing parsed YAML content
+    """
+    config = {}
+
+    for line in yaml_str.split("\n"):
+        # Remove comments
+        line = line.split("#")[0].strip()
+
+        if not line or ":" not in line:
+            continue
+
+        key, value = line.split(":", 1)
+        key = key.strip()
+        value = value.strip()
+
+        # Parse value based on type
+        if value.lower() == "true":
+            config[key] = True
+        elif value.lower() == "false":
+            config[key] = False
+        elif value.lower() in ("null", "none", ""):
+            config[key] = None
+        elif value.startswith("[") and value.endswith("]"):
+            # Parse list - handle quoted strings properly
+            pattern = r'"([^"]+)"|\'([^\']+)\'|([^,\[\]\s]+)'
+            matches = re.findall(pattern, value[1:-1])  # Remove [ ]
+            parsed_items = []
+            for match in matches:
+                # match is a tuple of (double_quoted, single_quoted, unquoted)
+                item = match[0] or match[1] or match[2]
+                item = item.strip()
+                if item:
+                    try:
+                        parsed_items.append(int(item))
+                    except ValueError:
+                        parsed_items.append(item)
+            config[key] = parsed_items
+        elif value.startswith(('"', "'")):
+            config[key] = value.strip("\"'")
+        else:
+            # Try to parse as number
+            try:
+                config[key] = int(value)
+            except ValueError:
+                try:
+                    config[key] = float(value)
+                except ValueError:
+                    config[key] = value
+
+    return config
+
+
+def print_default_config(output_format: str) -> None:
+    """Print a default configuration template in JSON or YAML format.
+
+    Args:
+        output_format: Either "json" or "yaml"
+    """
+    if output_format == "json":
+        print(json.dumps(default_config, indent=2))
+    else:  # yaml
+        for key, value in default_config.items():
+            if value is None:
+                print(f"{key}: null")
+            elif isinstance(value, bool):
+                print(f"{key}: {str(value).lower()}")
+            elif isinstance(value, str):
+                print(f'{key}: "{value}"')
+            elif isinstance(value, list):
+                print(f"{key}: {json.dumps(value)}")
+            else:
+                print(f"{key}: {value}")

benchmarks/transformer/configs/config_basic.yaml

@@ -0,0 +1,29 @@
+# Basic benchmark configuration for PyTorch transformer benchmarks
+# Usage: python score_mod.py --config config_basic.yaml
+
+# Core parameters
+dynamic: false
+calculate_bwd: true
+dtype: "bfloat16"
+
+# Shape parameters - larger sweep
+b: [1, 2, 4, 8, 16]  # batch sizes
+nh: ["16,16", "16,2", "32,32", "32,4"]  # [query_heads,key_value_heads]
+s: [512, 1024, 2048, 4096, 8192]  # sequence lengths
+d: [64, 128]  # head dimensions (limited to 128 for Flash Attention/cuDNN compatibility)
+
+# All attention types
+mods: ["noop", "causal", "rel", "head_bias", "alibi", "sliding_window", "prefix_lm", "softcap"]
+
+# Multiple backends for comparison (SDPA + Flash Attention) - flex is always included internally
+backend: ["efficient", "math", "cudnn", "fav2"]
+max_autotune: true  # Enable torch.compile with max-autotune for optimal performance
+
+# Decoding and cache settings
+decoding: false
+kv_size: null
+
+# Metrics and output
+throughput: true  # Calculate memory bandwidth & TFLOPS
+save_path: "comprehensive_results.csv"  # Save to CSV
+output_json_for_dashboard: "attn_bench_basic.json"

benchmarks/transformer/score_mod.py

@@ -1,15 +1,19 @@
 import argparse
 import csv
+import gc
 import itertools
+import json
 import random
+import sys
 from collections import defaultdict
 from collections.abc import Callable
 from contextlib import nullcontext
 from dataclasses import asdict, dataclass
-from functools import partial
-from typing import Optional, Union
+from functools import partial, wraps
+from typing import Literal, Optional, Union
 
 import numpy as np
+from config_utils import heads_input_type, load_config_file, print_default_config
 from tabulate import tabulate
 from tqdm import tqdm
 
@@ -33,6 +37,96 @@ torch._dynamo.config.recompile_limit = 1000
 from torch._inductor.runtime.benchmarking import benchmarker
 
 
+def cleanup_memory():
+    """Aggressively free GPU memory"""
+    torch.cuda.empty_cache()
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.synchronize()
+
+
+def safe_backend(backend_name=None, return_dict=False):
+    """Decorator that wraps backend functions with error handling
+
+    Args:
+        backend_name: Name of the backend for error messages
+        return_dict: If True, returns dict of results for all backends (for run_single_experiment)
+                     If False, returns single ExperimentResults (for individual backend functions)
+    """
+
+    def decorator(func):
+        @wraps(func)
+        def wrapper(config, *args, **kwargs):
+            try:
+                return func(config, *args, **kwargs)
+            except torch.OutOfMemoryError:
+                print(
+                    f"[SKIP] OOM for {backend_name or func.__name__} with shape {config.shape}"
+                )
+                cleanup_memory()
+            except RuntimeError as e:
+                error_msg = str(e)
+                if "out of resource" in error_msg or "OutOfMemoryError" in error_msg:
+                    print(
+                        f"[SKIP] Triton OOM for {backend_name or func.__name__} with shape {config.shape}"
+                    )
+                    cleanup_memory()
+                elif "No valid triton configs" in error_msg:
+                    print(
+                        f"[SKIP] No valid Triton config for {backend_name or func.__name__} with shape {config.shape}"
+                    )
+                else:
+                    print(
+                        f"[SKIP] Runtime error for {backend_name or func.__name__} with shape {config.shape}: {str(e)[:100]}"
+                    )
+            except Exception as e:
+                print(
+                    f"[SKIP] Error for {backend_name or func.__name__} with shape {config.shape}: {str(e)[:100]}"
+                )
+
+            # Return appropriate NaN result based on function type
+            if return_dict:
+                # For run_single_experiment: return dict with NaN for all backends
+                nan_result = ExperimentResults(
+                    fwd_time=float("nan"),
+                    bwd_time=float("nan") if config.calculate_bwd_time else None,
+                )
+                results = dict.fromkeys(config.backends, nan_result)
+                results["flex"] = ExperimentResults(
+                    fwd_time=float("nan"),
+                    bwd_time=float("nan") if config.calculate_bwd_time else None,
+                    sparsity=None,
+                )
+                return results
+            else:
+                # For individual backend functions: return single ExperimentResults
+                return ExperimentResults(
+                    fwd_time=float("nan"),
+                    bwd_time=float("nan") if config.calculate_bwd_time else None,
+                )
+
+        return wrapper
+
+    return decorator
+
+
+# Type definitions
+Backend = Literal["math", "efficient", "cudnn", "fav2", "fav3", "fakv", "og-eager"]
+AttentionType = Literal[
+    "noop",
+    "causal",
+    "rel",
+    "head_bias",
+    "alibi",
+    "sliding_window",
+    "document_mask",
+    "prefix_lm",
+    "softcap",
+]
+DtypeString = Literal["bfloat16", "float16", "float32"]
+SpeedupType = Literal["fwd", "bwd"]
+
+
 def benchmark_torch_function_in_microseconds(func: Callable, *args, **kwargs) -> float:
     # warmup
     for _ in range(5):
@@ -48,6 +142,7 @@ class ExperimentConfig:
     calculate_bwd_time: bool
     cal_bandwidth: bool
     backends: list[str]
+    max_autotune: bool
 
     def __post_init__(self):
         assert len(self.shape) == 6, (
@@ -62,6 +157,7 @@ class ExperimentConfig:
         d.pop("cal_bandwidth", None)
         d["shape(B,Hq,M,Hkv,N,D)"] = d.pop("shape")
         d.pop("backends", None)
+        d.pop("max_autotune", False)
         return d
 
 
@@ -209,6 +305,7 @@ def query_key_value_clones(
     return query_ref, key_ref, value_ref
 
 
+@safe_backend("SDPA")
 def run_single_backend_sdpa(
     config: ExperimentConfig,
     query: torch.Tensor,
@@ -223,6 +320,7 @@ def run_single_backend_sdpa(
     backend_context = get_backend_context(backend)
     with backend_context:
         _device = torch.device("cuda")
+
         eager_sdpa = generate_eager_sdpa(
             config.attn_type, config.shape, config.dtype, block_mask, score_mod
         )
@@ -290,6 +388,7 @@ def run_single_backend_sdpa(
             )
 
 
+@safe_backend("FlashAttention")
 def run_single_backend_FA(
     config: ExperimentConfig,
     query: torch.Tensor,
@@ -301,9 +400,9 @@ def run_single_backend_FA(
     mask_kwargs,
     backend: str,
 ) -> ExperimentResults:
-    assert backend in ["fav2", "fav3", "fakv"]
+    assert backend in ["fav3", "fakv"]
     # Generate callable for specific backend.
-    if backend in ["fav2", "fav3"]:
+    if backend in ["fav3"]:
         FA = generate_FA_callable(
             config.attn_type, config.shape, config.dtype, backend, **mask_kwargs
         )
@@ -354,10 +453,10 @@ def run_single_backend_FA(
     )
 
 
+@safe_backend("flex_attention", return_dict=True)
 def run_single_experiment(
     config: ExperimentConfig,
     dynamic=False,
-    max_autotune=False,
 ) -> dict[str, ExperimentResults]:
     device = torch.device("cuda")
     batch_size, q_heads, q_seq_len, kv_heads, kv_seq_len, head_dim = config.shape
@@ -377,7 +476,7 @@ def run_single_experiment(
     block_mask, mask_kwargs = generate_block_mask(config.attn_type, config.shape)
     kernel_options = get_kernel_options(config.attn_type, config.shape)
 
-    if max_autotune:
+    if config.max_autotune:
         compiled_sdpa = torch.compile(
             flex_attention, dynamic=dynamic, mode="max-autotune-no-cudagraphs"
         )
@@ -407,7 +506,7 @@ def run_single_experiment(
 
     results = {}
     for backend in config.backends:
-        if backend in ["fav2", "fav3", "fakv"]:
+        if backend in ["fav3", "fakv"]:
             results[backend] = run_single_backend_FA(
                 config,
                 query,
@@ -419,7 +518,7 @@ def run_single_experiment(
                 mask_kwargs,
                 backend,
             )
-        else:  # sdpa
+        else:  # sdpa (also supports fav2)
             results[backend] = run_single_backend_sdpa(
                 config,
                 query,
@@ -440,7 +539,7 @@ def run_single_experiment(
     sparsity = block_mask.sparsity() / 100.0 if block_mask is not None else 0.0
     sparsity = sparsity if config.attn_type != "document_mask" else 0.5
 
-    results["compiled"] = ExperimentResults(
+    results["flex"] = ExperimentResults(
         fwd_time=forward_compiled_time,
         bwd_time=backward_compile_time if config.calculate_bwd_time else None,
         sparsity=sparsity,
@@ -501,15 +600,15 @@ def calculate_tflops(config: ExperimentConfig, results: ExperimentResults) -> fl
     softmax_flops = M * N * 2  # Not counting online softmax overhead
     o_flops = M * D * N * 2
     # Not counting split k overhead
-    total_flops = B * Hq * (qk_flops + softmax_flops + o_flops) * (1 - results.sparsity)
+    sparsity = results.sparsity if results.sparsity is not None else 0.0
+    total_flops = B * Hq * (qk_flops + softmax_flops + o_flops) * (1 - sparsity)
     return total_flops / results.fwd_time / 1e6  # in TFLOPs/
 
 
 def get_average_speedups(results: list[Experiment], type: str, backend: str):
     # Calculate speedups
     speedups = [
-        calculate_speedup(r.results["compiled"], r.results[backend], type)
-        for r in results
+        calculate_speedup(r.results["flex"], r.results[backend], type) for r in results
     ]
 
     # Find indices of max and min speedups
@@ -537,7 +636,7 @@ def get_average_speedups(results: list[Experiment], type: str, backend: str):
 def print_results(results: list[Experiment], save_path: Optional[str] = None):
     table_data = defaultdict(list)
     for experiment in results:
-        backends = experiment.config.backends + ["compiled"]
+        backends = experiment.config.backends + ["flex"]
         for key, value in experiment.asdict().items():
             if key in backends:
                 if value.fwd_time:
@@ -550,45 +649,43 @@ def print_results(results: list[Experiment], save_path: Optional[str] = None):
     # Calculate speedups
     for backend in results[0].config.backends:
         fwd_speedups = [
-            calculate_speedup(r.results["compiled"], r.results[backend], type="fwd")
+            calculate_speedup(r.results["flex"], r.results[backend], type="fwd")
             for r in results
         ]
-        table_data[f"fwd_{backend}_speedup"] = fwd_speedups
+        table_data[f"fwd_speedup_flex_over_{backend}"] = fwd_speedups
 
     if results[0].config.calculate_bwd_time:
         for backend in results[0].config.backends:
             bwd_speedups = [
-                calculate_speedup(r.results["compiled"], r.results[backend], type="bwd")
+                calculate_speedup(r.results["flex"], r.results[backend], type="bwd")
                 for r in results
             ]
-            table_data[f"bwd_{backend}_speedup"] = bwd_speedups
+            table_data[f"bwd_speedup_flex_over_{backend}"] = bwd_speedups
 
     # Calculate mem + computational throughput
     if results[0].config.cal_bandwidth:
         fwd_bandwidth = [
-            calculate_bandwidth(r.config, r.results["compiled"], type="fwd")
+            calculate_bandwidth(r.config, r.results["flex"], type="fwd")
             for r in results
         ]
         table_data["fwd_mem_bw (TB/s)"] = fwd_bandwidth
-        fwd_tflops = [
-            calculate_tflops(r.config, r.results["compiled"]) for r in results
-        ]
+        fwd_tflops = [calculate_tflops(r.config, r.results["flex"]) for r in results]
         table_data["TFlops/s"] = fwd_tflops
 
     print(tabulate(table_data, headers="keys", tablefmt="github", floatfmt=".3f"))
 
     for backend in results[0].config.backends:
-        if np.isnan(table_data[f"fwd_{backend}_speedup"]).all():
+        if np.isnan(table_data[f"fwd_speedup_flex_over_{backend}"]).all():
             continue
         print("\n")
-        print(f"FWD Speedups vs. {backend}".center(125, "="))
+        print(f"FWD Speedup of Flex over {backend}".center(125, "="))
         print("\n")
         average_data = get_average_speedups(results, type="fwd", backend=backend)
         print(tabulate(average_data, headers="keys", tablefmt="github", floatfmt=".3f"))
 
         if results[0].config.calculate_bwd_time:
             print("\n")
-            print(f"BWD Speedups vs. {backend}".center(125, "="))
+            print(f"BWD Speedup of Flex over {backend}".center(125, "="))
             print("\n")
             average_data = get_average_speedups(results, type="bwd", backend=backend)
             print(
@@ -791,14 +888,14 @@ def get_backend_context(backend: str):
     Returns a context manager for the specified backend.
     Args:
         backend (str): The name of the backend to use.
-                       Valid options are 'fav2', 'cudnn', 'math', 'efficient', 'fav3', 'fakv', 'og-eager'.
+                       Valid options are 'math', 'efficient', 'cudnn', 'fav2', 'fav3', 'fakv', 'og-eager'.
     Returns:
         A context manager for the specified backend.
     Raises:
         ValueError: If an invalid backend is specified.
     """
     backends = {
-        "fav2": nullcontext(),
+        "fav2": sdpa_kernel(SDPBackend.FLASH_ATTENTION),
         "cudnn": sdpa_kernel(SDPBackend.CUDNN_ATTENTION),
         "math": sdpa_kernel(SDPBackend.MATH),
         "efficient": sdpa_kernel(SDPBackend.EFFICIENT_ATTENTION),
@@ -820,15 +917,7 @@ def generate_FA_callable(
 ) -> Callable | None:
     if dtype not in [torch.float16, torch.bfloat16]:
         return None
-    if backend == "fav2":
-        try:
-            from flash_attn import flash_attn_func, flash_attn_varlen_func
-        except ImportError:
-            print(
-                "Flash attention 2 is not installed. Please install it to run fav2 backend. "
-            )
-            raise
-    elif backend == "fav3":
+    if backend == "fav3":
         try:
             from flash_attn.flash_attn_interface import (
                 flash_attn_func,
@@ -1034,6 +1123,7 @@ def generate_experiment_configs(
     kv_cache_size: list[int],
     cal_bandwidth: bool,
     backends: list[str],
+    max_autotune: bool,
 ) -> list[ExperimentConfig]:
     assert not (calculate_bwd and decoding), "Decoding does not support backward"
 
@@ -1077,52 +1167,333 @@ def generate_experiment_configs(
                 calculate_bwd_time=calculate_bwd,
                 cal_bandwidth=cal_bandwidth,
                 backends=backends,
+                max_autotune=max_autotune,
             )
         )
 
     return all_configs
 
 
-def main(args):
+def _output_json_for_dashboard(
+    experiments,
+    output_file,
+    benchmark_name="PyTorch operator microbenchmark",
+):
+    """
+    Write the result into JSON format for PyTorch OSS dashboard.
+    The JSON format is defined at
+    https://github.com/pytorch/pytorch/wiki/How-to-integrate-with-PyTorch-OSS-benchmark-database
+
+    Args:
+        experiments: List of experiment results
+        output_file: Path to output JSON file
+        benchmark_name: Name of the benchmark
+    """
+    if not experiments:
+        return
+
+    import math
+    import platform
+    from dataclasses import asdict, dataclass
+    from typing import Any, Optional
+
+    # Prepare headers and records for JSON output
+    records = []
+    for experiment in experiments:
+        config = experiment.config
+        results_dict = (
+            experiment.results
+        )  # This is a dict: backend -> ExperimentResults
+
+        # Process each backend result
+        for backend, results in results_dict.items():
+            # Skip backends that were not run (NaN results)
+            if math.isnan(results.fwd_time):
+                continue
+
+            # Extract data from experiment
+            test_name = f"{backend}_{config.attn_type}_"
+            input_config = f"shape: {config.shape}, dtype: {config.dtype}"
+
+            # Determine mode based on backward pass
+            mode = "training" if config.calculate_bwd_time else "inference"
+
+            # Extract dtype
+            dtype = (
+                str(config.dtype).split(".")[1]
+                if "." in str(config.dtype)
+                else str(config.dtype)
+            )
+
+            # Determine device
+            device = "cuda"
+
+            # Get device architecture
+            device_arch = (
+                torch.cuda.get_device_name(0)
+                if device == "cuda"
+                else platform.processor()
+                if device == "cpu"
+                else "unknown"
+            )
+
+            # Create dataclasses for JSON structure
+            @dataclass
+            class BenchmarkInfo:
+                name: str
+                mode: Optional[str]
+                dtype: str
+                extra_info: dict[str, Any]
+
+            @dataclass
+            class ModelInfo:
+                name: str
+                type: str
+                origins: list[str]
+                extra_info: dict[str, Any]
+
+            @dataclass
+            class MetricInfo:
+                name: str
+                unit: str
+                benchmark_values: list[float]
+                target_value: Optional[float]
+
+            @dataclass
+            class BenchmarkRecord:
+                benchmark: BenchmarkInfo
+                model: ModelInfo
+                metric: MetricInfo
+
+            # Benchmark extra info
+            benchmark_extra_info = {
+                "input_config": input_config,
+                "device": device,
+                "arch": device_arch,
+                "operator_name": backend,
+                "attn_type": config.attn_type,
+                "shape": str(config.shape),
+                "max_autotune": config.max_autotune,
+            }
+            # Add record for forward latency
+            record_fwd_latency = BenchmarkRecord(
+                benchmark=BenchmarkInfo(
+                    name=benchmark_name,
+                    mode=mode,
+                    dtype=dtype,
+                    extra_info=benchmark_extra_info,
+                ),
+                model=ModelInfo(
+                    name=test_name + str(config.shape),
+                    type="attention-benchmark",
+                    origins=["pytorch"],
+                    extra_info={
+                        "operator_name": backend,
+                        "attn_type": config.attn_type,
+                    },
+                ),
+                metric=MetricInfo(
+                    name="forward latency",
+                    unit="us",
+                    benchmark_values=[results.fwd_time],
+                    target_value=None,
+                ),
+            )
+            records.append(asdict(record_fwd_latency))
+
+            # Add record for forward memory bandwidth (if available)
+            if config.cal_bandwidth:
+                record_fwd_bandwidth = BenchmarkRecord(
+                    benchmark=BenchmarkInfo(
+                        name=benchmark_name,
+                        mode=mode,
+                        dtype=dtype,
+                        extra_info=benchmark_extra_info,
+                    ),
+                    model=ModelInfo(
+                        name=test_name + str(config.shape),
+                        type="attention-benchmark",
+                        origins=["pytorch"],
+                        extra_info={
+                            "operator_name": backend,
+                        },
+                    ),
+                    metric=MetricInfo(
+                        name="memory bandwidth",
+                        unit="TB/s",
+                        benchmark_values=[calculate_bandwidth(config, results, "fwd")],
+                        target_value=None,
+                    ),
+                )
+                records.append(asdict(record_fwd_bandwidth))
+
+            # Add record for forward TFLOPS (if available)
+            if config.cal_bandwidth:
+                record_fwd_tflops = BenchmarkRecord(
+                    benchmark=BenchmarkInfo(
+                        name=benchmark_name,
+                        mode=mode,
+                        dtype=dtype,
+                        extra_info=benchmark_extra_info,
+                    ),
+                    model=ModelInfo(
+                        name=test_name + str(config.shape),
+                        type="attention-benchmark",
+                        origins=["pytorch"],
+                        extra_info={
+                            "operator_name": backend,
+                        },
+                    ),
+                    metric=MetricInfo(
+                        name="tflops",
+                        unit="TFLOPS/s",
+                        benchmark_values=[calculate_tflops(config, results)],
+                        target_value=None,
+                    ),
+                )
+                records.append(asdict(record_fwd_tflops))
+
+            # Add record for backward latency (if available and not NaN)
+            if (
+                config.calculate_bwd_time
+                and results.bwd_time is not None
+                and not math.isnan(results.bwd_time)
+            ):
+                record_bwd_latency = BenchmarkRecord(
+                    benchmark=BenchmarkInfo(
+                        name=benchmark_name,
+                        mode=mode,
+                        dtype=dtype,
+                        extra_info=benchmark_extra_info,
+                    ),
+                    model=ModelInfo(
+                        name=test_name + str(config.shape),
+                        type="attention-benchmark",
+                        origins=["pytorch"],
+                        extra_info={
+                            "operator_name": backend,
+                        },
+                    ),
+                    metric=MetricInfo(
+                        name="backward latency",
+                        unit="us",
+                        benchmark_values=[results.bwd_time],
+                        target_value=None,
+                    ),
+                )
+                records.append(asdict(record_bwd_latency))
+
+    # Write all records to the output file
+    with open(output_file, "w", encoding="utf-8") as f:
+        json.dump(records, f, indent=2)
+
+
+def main(
+    dynamic: bool = False,
+    calculate_bwd: bool = False,
+    dtype: DtypeString = "bfloat16",
+    b: list[int] | None = None,
+    nh: list[str] | None = None,
+    s: list[int] | None = None,
+    d: list[int] | None = None,
+    mods: list[AttentionType] | None = None,
+    backend: list[Backend] | None = None,
+    max_autotune: bool = False,
+    decoding: bool = False,
+    kv_size: Optional[list[int]] = None,
+    throughput: bool = True,
+    save_path: Optional[str] = None,
+    output_json_for_dashboard: Optional[str] = None,
+    benchmark_name: str = "PyTorch operator microbenchmark",
+) -> None:
+    """Run sweep over sizes and score mods for flex attention.
+
+    Usage Examples:
+        # Use a yml config file
+        python score_mod.py --config basic_config.yaml
+
+        # Use a json config file
+        python score_mod.py --config my_config.json
+
+        # Generate a config template
+        python score_mod.py --print-config json > my_config.json # For a json config
+        python score_mod.py --print-config yaml > my_config.yaml # For a yaml config
+
+        # Override config with CLI args
+        python score_mod.py --config my_config.json -dtype float16 --max-autotune
+
+        # Pure CLI usage
+        python score_mod.py -b 4 8 -s 1024 2048 -mods causal alibi --backend efficient
+
+    Args:
+        dynamic: Runs a dynamic shapes version of compiled flex attention
+        calculate_bwd: Calculate backward pass times
+        dtype: Data type for tensors (bfloat16, float16, float32)
+        b: Batch sizes to benchmark
+        nh: Number of query and key/value heads in format "Hq,Hkv"
+        s: Sequence lengths to benchmark
+        d: Head dimensions to benchmark
+        mods: Score modifications: noop, causal, rel, head_bias, alibi, sliding_window, document_mask, prefix_lm, softcap
+        backend: Backends for attention computation: math, efficient, cudnn, fav2, fav3, fakv, og-eager
+        max_autotune: Turn on max-autotune optimization
+        decoding: Benchmark decoding mode (query sequence length = 1)
+        kv_size: Key/value cache size in MiB (ignores batch size if specified)
+        throughput: Calculate kernel memory bandwidth & computational throughput (always True)
+        save_path: Path to save the results CSV file
+        output_json_for_dashboard: Path to save results in JSON format for PyTorch OSS dashboard
+        benchmark_name: Name of the benchmark for dashboard output
+    """
+    # Convert dtype string to torch dtype (if not already converted)
+    import torch
+
+    if isinstance(dtype, str):
+        dtype = getattr(torch, dtype)
+
+    # Always calculate throughput
+    throughput = True
+    print("Backend: ", backend)
     seed = 123
     np.random.seed(seed)
     torch.manual_seed(seed)
     results = []
-    for config in tqdm(
-        generate_experiment_configs(
-            args.calculate_bwd,
-            args.dtype,
-            args.b,
-            args.nh,
-            args.s,
-            args.d,
-            args.mods,
-            args.decoding,
-            args.kv_size,
-            args.throughput,
-            args.backend,
-        )
+    for experiment_count, config in enumerate(
+        tqdm(
+            generate_experiment_configs(
+                calculate_bwd,
+                dtype,
+                b,
+                nh,
+                s,
+                d,
+                mods,
+                decoding,
+                kv_size,
+                throughput,
+                backend,
+                max_autotune,
+            )
+        ),
+        start=1,
     ):
         results.append(
             Experiment(
                 config,
                 run_single_experiment(
                     config,
-                    dynamic=args.dynamic,
-                    max_autotune=args.max_autotune,
+                    dynamic=dynamic,
                 ),
             )
         )
 
-    print_results(results, args.save_path)
+        # Periodic memory cleanup every 50 experiments
+        if experiment_count % 50 == 0:
+            cleanup_memory()
 
+    print_results(results, save_path)
 
-def heads_input_type(s):
-    try:
-        hq, hkv = map(int, s.split(","))
-        return hq, hkv
-    except Exception as e:
-        raise argparse.ArgumentTypeError("Heads must be Hq,Hkv") from e
+    # Output JSON for dashboard if requested
+    if output_json_for_dashboard:
+        _output_json_for_dashboard(results, output_json_for_dashboard, benchmark_name)
 
 
 if __name__ == "__main__":
@@ -1130,6 +1501,12 @@ if __name__ == "__main__":
     parser = argparse.ArgumentParser(
         description="Run sweep over sizes and score mods for flex attention"
     )
+    parser.add_argument(
+        "--config",
+        type=str,
+        help="Path to JSON config file. CLI args override config file values.",
+        default=None,
+    )
     parser.add_argument(
         "--dynamic",
         action="store_true",
@@ -1199,8 +1576,49 @@ Ignores -b batch size and calculate batch size from kv size instead when specifi
         default=["efficient"],
         help="Backend to use for attention computation",
     )
+    parser.add_argument(
+        "--output-json-for-dashboard",
+        type=str,
+        help="Path to save results in JSON format for PyTorch OSS dashboard",
+        default=None,
+    )
+    parser.add_argument(
+        "--benchmark-name",
+        type=str,
+        help="Name of the benchmark for dashboard output",
+        default="PyTorch operator microbenchmark",
+    )
+    parser.add_argument(
+        "--print-config",
+        type=str,
+        choices=["json", "yaml"],
+        help="Print a default config template in JSON or YAML format and exit",
+        default=None,
+    )
     # Parse arguments
     args = parser.parse_args()
-    args.dtype = getattr(torch, args.dtype)
 
-    main(args)
+    # Handle --print-config
+    if args.print_config:
+        print_default_config(args.print_config)
+        sys.exit(0)
+
+    # Load and merge config if provided
+    if args.config:
+        config = load_config_file(args.config)
+
+        # Merge config with CLI args (CLI args take precedence)
+        json_args = argparse.Namespace()
+        json_args.__dict__ = config
+        args = parser.parse_args(namespace=json_args)
+
+    # Convert dtype string to torch dtype (only if it's still a string)
+    if isinstance(args.dtype, str):
+        args.dtype = getattr(torch, args.dtype)
+
+    # Remove config and print_config from args before passing to main
+    args_dict = vars(args)
+    args_dict.pop("config", None)
+    args_dict.pop("print_config", None)
+
+    main(**args_dict)

build_variables.bzl

@@ -482,6 +482,7 @@ inductor_core_resources = [
     "torch/csrc/inductor/aoti_torch/oss_proxy_executor.cpp",
     "torch/csrc/inductor/inductor_ops.cpp",
     "torch/csrc/jit/serialization/pickle.cpp",
+    "torch/csrc/shim_common.cpp",
 ]
 
 libtorch_core_sources = sorted(

c10/core/SymInt.h

@@ -556,3 +556,26 @@ inline SymBool sym_ge(const SymInt& a, const SymInt& b) {
 }
 
 } // namespace c10
+
+#include <limits>
+
+namespace std {
+
+template <>
+class numeric_limits<c10::SymInt> {
+ public:
+  static constexpr bool is_specialized = true;
+
+  static constexpr int64_t max() noexcept {
+    return std::numeric_limits<int64_t>::max();
+  }
+
+  static constexpr int64_t min() noexcept {
+    return std::numeric_limits<int64_t>::min();
+  }
+
+  static constexpr bool is_signed = true;
+  static constexpr bool is_integer = true;
+};
+
+} // namespace std

c10/metal/special_math.h

@@ -1,4 +1,4 @@
-// Implementation of specal math functions for Metal
+// Implementation of special math functions for Metal
 #pragma once
 #include <c10/metal/expm1f.h>
 #include <c10/metal/igamma.h>

c10/mobile/CPUProfilingAllocator.cpp

@@ -34,7 +34,7 @@ struct MemEvent {
 bool overlaps(const MemBlock& a, const MemBlock& b) {
   // two blocks dont overlap if
   // |---a--------|--------------b--------|
-  // strat_a     end_a <= start_b       end_b
+  // start_a     end_a <= start_b       end_b
   return !(
       (a.end_offset <= b.start_offset) || (b.end_offset <= a.start_offset));
 }

c10/util/Bitset.h

@@ -33,7 +33,7 @@ struct bitset final {
   constexpr bitset() noexcept = default;
   constexpr bitset(const bitset&) noexcept = default;
   constexpr bitset(bitset&&) noexcept = default;
-  // there is an issure for gcc 5.3.0 when define default function as constexpr
+  // there is an issue for gcc 5.3.0 when define default function as constexpr
   // see https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68754.
   bitset& operator=(const bitset&) noexcept = default;
   bitset& operator=(bitset&&) noexcept = default;

c10/xpu/XPUCachingAllocator.cpp

@@ -123,6 +123,8 @@ class DeviceCachingAllocator {
   ska::flat_hash_map<xpu::XPUStream, std::deque<std::pair<sycl::event, Block*>>>
       xpu_events;
   DeviceIndex device_index;
+  size_t allowed_memory_maximum = 0;
+  bool set_fraction = false;
 
   size_t try_merge_blocks(Block* dst, Block* src, BlockPool& pool) {
     if (!src || src->allocated || src->event_count > 0 ||
@@ -245,6 +247,12 @@ class DeviceCachingAllocator {
     if (isRetry) {
       stats.num_alloc_retries += 1;
     }
+    if (set_fraction &&
+        stats.reserved_bytes[static_cast<size_t>(StatType::AGGREGATE)].current +
+                size >
+            allowed_memory_maximum) {
+      return false;
+    }
     void* ptr = sycl::aligned_alloc_device(
         kDeviceAlignment,
         size,
@@ -435,6 +443,11 @@ class DeviceCachingAllocator {
         device_free =
             raw_device.get_info<sycl::ext::intel::info::device::free_memory>();
       }
+      std::string allowed_info;
+      if (set_fraction) {
+        allowed_info = format_size(allowed_memory_maximum) + " allowed; ";
+      }
+
       auto allocated_bytes =
           stats.allocated_bytes[static_cast<size_t>(StatType::AGGREGATE)]
               .current;
@@ -459,7 +472,9 @@ class DeviceCachingAllocator {
           format_size(device_total),
           " of which ",
           format_size(device_free),
-          " is free. Of the allocated memory ",
+          " is free. ",
+          allowed_info,
+          "Of the allocated memory ",
           format_size(allocated_bytes),
           " is allocated by PyTorch, and ",
           format_size(reserved_bytes - allocated_bytes),
@@ -538,6 +553,14 @@ class DeviceCachingAllocator {
       stats.requested_bytes[statType].reset_peak();
     }
   }
+
+  void setMemoryFraction(double fraction) {
+    c10::xpu::DeviceProp device_prop;
+    c10::xpu::get_device_properties(&device_prop, device_index);
+    auto device_total = device_prop.global_mem_size;
+    allowed_memory_maximum = static_cast<size_t>(fraction * device_total);
+    set_fraction = true;
+  }
 };
 
 static void local_raw_delete(void* ptr);
@@ -700,6 +723,16 @@ class XPUAllocator : public DeviceAllocator {
     assertValidDevice(device);
     device_allocators[device]->resetAccumulatedStats();
   }
+
+  void setMemoryFraction(double fraction, DeviceIndex device) {
+    assertValidDevice(device);
+    TORCH_CHECK_VALUE(
+        0 < fraction && fraction <= 1,
+        "invalid fraction:",
+        fraction,
+        ". Please set within (0, 1].");
+    device_allocators[device]->setMemoryFraction(fraction);
+  }
 };
 
 static XPUAllocator allocator;
@@ -744,6 +777,10 @@ void recordStream(const DataPtr& dataPtr, XPUStream stream) {
   return allocator.recordStream(dataPtr, stream);
 }
 
+void setMemoryFraction(double fraction, DeviceIndex device) {
+  return allocator.setMemoryFraction(fraction, device);
+}
+
 REGISTER_ALLOCATOR(kXPU, &allocator)
 
 } // namespace c10::xpu::XPUCachingAllocator

c10/xpu/XPUCachingAllocator.h

@@ -25,4 +25,6 @@ C10_XPU_API void raw_delete(void* ptr);
 
 C10_XPU_API void recordStream(const DataPtr& dataPtr, XPUStream stream);
 
+C10_XPU_API void setMemoryFraction(double fraction, DeviceIndex device);
+
 } // namespace c10::xpu::XPUCachingAllocator

caffe2/CMakeLists.txt

@@ -1358,9 +1358,15 @@ if(BUILD_TEST)
     )
   else()
     add_subdirectory(${TORCH_ROOT}/test/cpp/jit ${CMAKE_BINARY_DIR}/test_jit)
+    add_subdirectory(${TORCH_ROOT}/test/cpp/lazy ${CMAKE_BINARY_DIR}/test_lazy)
     # NativeRT is disabled
     # add_subdirectory(${TORCH_ROOT}/test/cpp/nativert ${CMAKE_BINARY_DIR}/test_nativert)
     add_subdirectory(${TORCH_ROOT}/test/inductor ${CMAKE_BINARY_DIR}/test_inductor)
+    add_subdirectory(${TORCH_ROOT}/test/cpp/aoti_abi_check ${CMAKE_BINARY_DIR}/test_aoti_abi_check)
+    if(BUILD_AOT_INDUCTOR_TEST)
+      add_subdirectory(${TORCH_ROOT}/test/cpp/aoti_inference ${CMAKE_BINARY_DIR}/test_aoti_inference)
+    endif()
+
     if(USE_DISTRIBUTED)
       add_subdirectory(${TORCH_ROOT}/test/cpp/c10d ${CMAKE_BINARY_DIR}/test_cpp_c10d)
       if(NOT WIN32)
@@ -1378,16 +1384,6 @@ if(BUILD_TEST)
         ${CMAKE_BINARY_DIR}/test_mobile_nnc
       )
     endif()
-    add_subdirectory(${TORCH_ROOT}/test/cpp/lazy
-                     ${CMAKE_BINARY_DIR}/test_lazy)
-  endif()
-  if(BUILD_AOT_INDUCTOR_TEST)
-    add_subdirectory(
-      ${TORCH_ROOT}/test/cpp/aoti_abi_check
-      ${CMAKE_BINARY_DIR}/test_aoti_abi_check)
-    add_subdirectory(
-      ${TORCH_ROOT}/test/cpp/aoti_inference
-      ${CMAKE_BINARY_DIR}/test_aoti_inference)
   endif()
 endif()
 

caffe2/serialize/crc_alt.h

@@ -38,7 +38,7 @@ uint32_t crc32_combine (uint32_t crcA, uint32_t crcB, size_t lengthB);
 
 /// compute CRC32 (bitwise algorithm)
 uint32_t crc32_bitwise (const void* data, size_t length, uint32_t previousCrc32 = 0);
-/// compute CRC32 (half-byte algoritm)
+/// compute CRC32 (half-byte algorithm)
 uint32_t crc32_halfbyte(const void* data, size_t length, uint32_t previousCrc32 = 0);
 
 #ifdef CRC32_USE_LOOKUP_TABLE_BYTE
@@ -96,7 +96,7 @@ uint32_t crc32_16bytes_prefetch(const void* data, size_t length, uint32_t previo
   #define __BIG_ENDIAN    4321
 #endif
 
-// define endianess and some integer data types
+// define endianness and some integer data types
 #if defined(_MSC_VER) || defined(__MINGW32__)
   // Windows always little endian
   #define __BYTE_ORDER __LITTLE_ENDIAN
@@ -168,7 +168,7 @@ namespace
   /// zlib's CRC32 polynomial
   const uint32_t Polynomial = 0xEDB88320;
 
-  /// swap endianess
+  /// swap endianness
   static inline uint32_t swap(uint32_t x)
   {
   #if defined(__GNUC__) || defined(__clang__)
@@ -229,7 +229,7 @@ uint32_t crc32_bitwise(const void* data, size_t length, uint32_t previousCrc32)
 }
 
 
-/// compute CRC32 (half-byte algoritm)
+/// compute CRC32 (half-byte algorithm)
 uint32_t crc32_halfbyte(const void* data, size_t length, uint32_t previousCrc32)
 {
   uint32_t crc = ~previousCrc32; // same as previousCrc32 ^ 0xFFFFFFFF
@@ -662,7 +662,7 @@ uint32_t crc32_combine(uint32_t crcA, uint32_t crcB, size_t lengthB)
   // - if you append length(B) zeros to A and call it A' (think of it as AAAA000)
   //   and   prepend length(A) zeros to B and call it B' (think of it as 0000BBB)
   //   then exists a C' = A' ^ B'
-  // - remember: if you XOR someting with zero, it remains unchanged: X ^ 0 = X
+  // - remember: if you XOR something with zero, it remains unchanged: X ^ 0 = X
   // - that means C' = A concat B so that crc(A concat B) = crc(C') = crc(A') ^ crc(B')
   // - the trick is to compute crc(A') based on crc(A)
   //                       and crc(B') based on crc(B)

caffe2/serialize/inline_container.h

@@ -76,7 +76,7 @@ typedef struct mz_zip_archive mz_zip_archive;
 // 2) Writing with 1-pass sequential access
 //      -> We must take care not to require updating values that have already
 //         been written. We place the variable-length index at the end and do
-//         not put any indicies into the header to fulfill this constraint.
+//         not put any index into the header to fulfill this constraint.
 
 // The model.json, which contains all the metadata information,
 // should be written as the last file. One reason is that the size of tensor

caffe2/serialize/inline_container_test.cc

@@ -519,7 +519,7 @@ TEST(PyTorchStreamWriterAndReader, SaveAndLoadWithAllocator) {
   std::tie(data_ptr, size) = reader.getRecord("key1", &overrideAllocator);
   EXPECT_EQ(overrideAllocator.getAllocatedBytes(), kBytes1);
   EXPECT_EQ(baseAllocator.getAllocatedBytes(), allocBytes);
-  // allcoate with base allocator
+  // allocate with base allocator
   std::tie(data_ptr, size) = reader.getRecord("key1");
   EXPECT_EQ(overrideAllocator.getAllocatedBytes(), kBytes1);
   EXPECT_EQ(baseAllocator.getAllocatedBytes(), allocBytes + kBytes1);

cmake/public/utils.cmake

@@ -383,7 +383,7 @@ function(torch_compile_options libname)
       -Wno-strict-aliasing
       )
     if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
-      list(APPEND private_compile_options -Wredundant-move)
+      list(APPEND private_compile_options -Wredundant-move -Wno-interference-size)
     endif()
     if(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
       list(APPEND private_compile_options -Wextra-semi -Wmove)

docs/cpp/source/index.rst

@@ -14,7 +14,7 @@ Combining, these building blocks form a research and
 production ready C++ library for tensor computation and dynamic neural
 networks with strong emphasis on GPU acceleration as well as fast CPU
 performance. It is currently in use at Facebook in research and
-production; we are looking forward to welcome more users of the PyTorch C++ API.
+production; we are looking forward to welcoming more users of the PyTorch C++ API.
 
 .. warning::
 

docs/cpp/source/notes/inference_mode.rst

@@ -64,7 +64,7 @@ users should pay additional attention to:
 
   - Both guards affects tensor execution process to skip work not related to inference, but ``InferenceMode``
     also affects tensor creation while ``AutoNonVariableTypeMode`` doesn't. In other words, tensors created
-    inside ``InferenceMode`` are marked as inference tensors so that certain limitation can be applied after
+    inside ``InferenceMode`` are marked as inference tensors so that certain limitations can be applied after
     exiting ``InferenceMode``.
   - Enabled/disabled ``InferenceMode`` states can be nested while ``AutoNonVariableTypeMode`` only allows enabled state.
 

docs/source/checkpoint.md

@@ -17,7 +17,7 @@ restoring the RNG state during each checkpoint.
 The stashing logic saves and restores the RNG state for CPU and another
 device type (infer the device type from Tensor arguments excluding CPU
 tensors by `_infer_device_type`) to the `run_fn`. If there are multiple
-device, device state will only be saved for devices of a single device type,
+devices, device state will only be saved for devices of a single device type,
 and the remaining devices will be ignored. Consequently, if any checkpointed
 functions involve randomness, this may result in incorrect gradients. (Note
 that if CUDA devices are among the devices detected, it will be prioritized;

docs/source/conf.py

@@ -1066,6 +1066,8 @@ coverage_ignore_functions = [
     "set_current_meta",
     "set_grad_fn_seq_nr",
     "set_stack_trace",
+    "set_current_replay_node",
+    "get_current_replay_node",
     # torch.jit.annotations
     "ann_to_type",
     "check_fn",

docs/source/distributed.md

@@ -59,14 +59,14 @@ MPI supports CUDA only if the implementation used to build PyTorch supports it.
 
 ### Backends that come with PyTorch
 
-PyTorch distributed package supports Linux (stable), MacOS (stable), and Windows (prototype).
+PyTorch distributed package supports Linux (stable), macOS (stable), and Windows (prototype).
 By default for Linux, the Gloo and NCCL backends are built and included in PyTorch
 distributed (NCCL only when building with CUDA). MPI is an optional backend that can only be
 included if you build PyTorch from source. (e.g. building PyTorch on a host that has MPI
 installed.)
 
 :::{note}
-As of PyTorch v1.8, Windows supports all collective communications backend but NCCL,
+As of PyTorch v1.8, Windows supports all collective communications backends but NCCL,
 If the `init_method` argument of {func}`init_process_group` points to a file it must adhere
 to the following schema:
 
@@ -394,6 +394,10 @@ an opaque group handle that can be given as a `group` argument to all collective
 .. autofunction:: new_group
 ```
 
+```{eval-rst}
+.. autofunction:: torch.distributed.distributed_c10d.shrink_group
+```
+
 ```{eval-rst}
 .. autofunction:: get_group_rank
 ```

docs/source/distributed.tensor.md

@@ -99,6 +99,12 @@ DTensor supports the following types of {class}`Placement` on each {class}`Devic
   :undoc-members:
 ```
 
+```{eval-rst}
+.. autoclass:: MaskPartial
+  :members:
+  :undoc-members:
+```
+
 ```{eval-rst}
 .. autoclass:: Placement
   :members:

docs/source/mtia.md

@@ -1,6 +1,6 @@
 # torch.mtia
 
-The MTIA backend is implemented out of the tree, only interfaces are be defined here.
+The MTIA backend is implemented out of the tree, only interfaces are defined here.
 
 ```{eval-rst}
 .. automodule:: torch.mtia

docs/source/mtia.memory.md

@@ -1,6 +1,6 @@
 # torch.mtia.memory
 
-The MTIA backend is implemented out of the tree, only interfaces are be defined here.
+The MTIA backend is implemented out of the tree, only interfaces are defined here.
 
 ```{eval-rst}
 .. automodule:: torch.mtia.memory

docs/source/notes/libtorch_stable_abi.md

@@ -122,12 +122,38 @@ The above is relevant in two places:
     }
     ```
 
-2. `aoti_torch_call_dispatcher`
+2. `torch_call_dispatcher`
     This API allows you to call the PyTorch dispatcher from C/C++ code. It has the following signature:
+
     ```cpp
-    aoti_torch_call_dispatcher(const char* opName, const char* overloadName, StableIValue* stack);
+    torch_call_dispatcher(const char* opName, const char* overloadName, StableIValue* stack, uint64_t extension_build_version);
     ```
 
-    `aoti_torch_call_dispatcher` will call the op overload defined by a given `opName`, `overloadName`, and a stack of
-    StableIValues. This call will populate any return values of the op into the stack in their StableIValue form,
-    with `ret0` at index 0, `ret1` at index 1, and so on.
+    `torch_call_dispatcher` will call the op overload defined by a given `opName`, `overloadName`, a stack of
+    StableIValues and the `TORCH_ABI_VERSION` of the user extension. This call will populate any return values of the
+    op into the stack in their StableIValue form, with `ret0` at index 0, `ret1` at index 1, and so on.
+
+    We caution against using this API to call functions that have been registered to the dispatcher by other extensions
+    unless the caller can guarantee that the signature they expect matches that which the custom extension has
+    registered.
+
+### Versioning and Forward/Backward compatibility guarantees
+
+We provide a `TORCH_ABI_VERSION` macro in `torch/headeronly/version.h` of the form
+
+```
+[ byte ][ byte ][ byte ][ byte ][ byte ][ byte ][ byte ][ byte ]
+[MAJ   ][ MIN  ][PATCH ][                 ABI TAG              ]
+```
+
+In the present phase of development, APIs in the C-shim will be versioned based on major.minor.patch release that they are first introduced in, with 2.10 being the first release where this will be enforced. The ABI tag is reserved for future use.
+
+Extensions can select the minimum abi version to be compatible with using:
+
+```
+#define TORCH_TARGET_VERSION (((0ULL + major) << 56) | ((0ULL + minor) << 48))
+```
+
+before including any stable headers or by passing the equivalent `-D` option to the compiler. Otherwise, the default will be the current `TORCH_ABI_VERSION`.
+
+The above ensures that if a user defines `TORCH_TARGET_VERSION` to be 0x0209000000000000 (2.9) and attempts to use a C shim API `foo` that was introduced in version 2.10, a compilation error will be raised. Similarly, the C++ wrapper APIs in `torch/csrc/stable` are compatible with older libtorch binaries up to the TORCH_ABI_VERSION they are exposed in and forward compatible with newer libtorch binaries.

docs/source/xpu.md

@@ -85,6 +85,7 @@
      memory_stats_as_nested_dict
      reset_accumulated_memory_stats
      reset_peak_memory_stats
+     set_per_process_memory_fraction
 ```
 
 ```{eval-rst}

setup.py

@@ -1106,7 +1106,7 @@ class build_ext(setuptools.command.build_ext.build_ext):
                 continue
             self.copy_file(source_lib, target_lib)
             # Delete old rpath and add @loader_lib to the rpath
-            # This should prevent delocate from attempting to package another instance
+            # This should prevent deallocate from attempting to package another instance
             # of OpenMP library in torch wheel as well as loading two libomp.dylib into
             # the address space, as libraries are cached by their unresolved names
             install_name_tool_args = [

test/conftest.py

@@ -238,7 +238,7 @@ def pytest_pycollect_makemodule(module_path, path, parent) -> Module:
 
 @pytest.hookimpl(hookwrapper=True)
 def pytest_report_teststatus(report, config):
-    # Add the test time to the verbose output, unforunately I don't think this
+    # Add the test time to the verbose output, unfortunately I don't think this
     # includes setup or teardown
     pluggy_result = yield
     if not isinstance(report, pytest.TestReport):

test/cpp/aoti_abi_check/CMakeLists.txt

@@ -1,3 +1,8 @@
+# Skip on windows
+if(WIN32)
+  return()
+endif()
+
 set(AOTI_ABI_CHECK_TEST_ROOT ${TORCH_ROOT}/test/cpp/aoti_abi_check)
 
 # Build the cpp gtest binary containing the cpp-only tests.
@@ -30,8 +35,15 @@ target_compile_definitions(test_aoti_abi_check PRIVATE USE_GTEST)
 
 # WARNING: DO NOT LINK torch!!!
 # The purpose is to check if the used aten/c10 headers are written in a header-only way
-target_link_libraries(test_aoti_abi_check PRIVATE gtest_main)
+target_link_libraries(test_aoti_abi_check PRIVATE gtest_main sleef)
 target_include_directories(test_aoti_abi_check PRIVATE ${ATen_CPU_INCLUDE})
+if(NOT USE_SYSTEM_SLEEF)
+  target_include_directories(test_aoti_abi_check PRIVATE ${CMAKE_BINARY_DIR}/include)
+endif()
+
+# Disable unused-variable warnings for variables that are only used to test compilation
+target_compile_options_if_supported(test_aoti_abi_check -Wno-unused-variable)
+target_compile_options_if_supported(test_aoti_abi_check -Wno-unused-but-set-variable)
 
 foreach(test_src ${AOTI_ABI_CHECK_VEC_TEST_SRCS})
   foreach(i RANGE ${NUM_CPU_CAPABILITY_NAMES})
@@ -41,12 +53,17 @@ foreach(test_src ${AOTI_ABI_CHECK_VEC_TEST_SRCS})
     separate_arguments(FLAGS UNIX_COMMAND "${FLAGS}")
     add_executable(${test_name}_${CPU_CAPABILITY} "${test_src}")
 
-    target_link_libraries(${test_name}_${CPU_CAPABILITY} PRIVATE gtest_main)
+    target_link_libraries(${test_name}_${CPU_CAPABILITY} PRIVATE gtest_main sleef)
     target_include_directories(${test_name}_${CPU_CAPABILITY} PRIVATE ${ATen_CPU_INCLUDE})
+    if(NOT USE_SYSTEM_SLEEF)
+      target_include_directories(${test_name}_${CPU_CAPABILITY} PRIVATE ${CMAKE_BINARY_DIR}/include)
+    endif()
 
     # Define CPU_CAPABILITY and CPU_CAPABILITY_XXX macros for conditional compilation
     target_compile_definitions(${test_name}_${CPU_CAPABILITY} PRIVATE CPU_CAPABILITY=${CPU_CAPABILITY} CPU_CAPABILITY_${CPU_CAPABILITY})
     target_compile_options(${test_name}_${CPU_CAPABILITY} PRIVATE ${FLAGS})
+    target_compile_options_if_supported(${test_name}_${CPU_CAPABILITY} -Wno-unused-variable)
+    target_compile_options_if_supported(${test_name}_${CPU_CAPABILITY} -Wno-unused-but-set-variable)
   endforeach()
 endforeach()
 

test/cpp/aoti_abi_check/test_vec.cpp

@@ -2,10 +2,27 @@
 
 #include <ATen/cpu/vec/vec.h>
 
-#include <iostream>
 namespace torch {
 namespace aot_inductor {
 
+template <typename T>
+void ExpectVecEqual(
+    const at::vec::Vectorized<T>& expected,
+    const at::vec::Vectorized<T>& actual) {
+  using Vec = at::vec::Vectorized<T>;
+  // Have to use std::vector for comparison because at::vec::Vectorized doesn't
+  // support operator[] on aarch64
+  std::vector<T> expected_data(Vec::size());
+  std::vector<T> actual_data(Vec::size());
+
+  expected.store(expected_data.data());
+  actual.store(actual_data.data());
+
+  for (int i = 0; i < Vec::size(); i++) {
+    EXPECT_EQ(expected_data[i], actual_data[i]);
+  }
+}
+
 TEST(TestVec, TestAdd) {
   using Vec = at::vec::Vectorized<int>;
   std::vector<int> a(1024, 1);
@@ -16,9 +33,7 @@ TEST(TestVec, TestAdd) {
   std::vector<int> expected(1024, 3);
   Vec expected_vec = Vec::loadu(expected.data());
 
-  for (int i = 0; i < Vec::size(); i++) {
-    EXPECT_EQ(expected_vec[i], actual_vec[i]);
-  }
+  ExpectVecEqual(expected_vec, actual_vec);
 }
 
 TEST(TestVec, TestMax) {
@@ -30,9 +45,7 @@ TEST(TestVec, TestMax) {
   Vec actual_vec = at::vec::maximum(a_vec, b_vec);
   Vec expected_vec = b_vec;
 
-  for (int i = 0; i < Vec::size(); i++) {
-    EXPECT_EQ(expected_vec[i], actual_vec[i]);
-  }
+  ExpectVecEqual(expected_vec, actual_vec);
 }
 
 TEST(TestVec, TestMin) {
@@ -44,9 +57,7 @@ TEST(TestVec, TestMin) {
   Vec actual_vec = at::vec::minimum(a_vec, b_vec);
   Vec expected_vec = a_vec;
 
-  for (int i = 0; i < Vec::size(); i++) {
-    EXPECT_EQ(expected_vec[i], actual_vec[i]);
-  }
+  ExpectVecEqual(expected_vec, actual_vec);
 }
 
 TEST(TestVec, TestConvert) {
@@ -58,9 +69,7 @@ TEST(TestVec, TestConvert) {
   auto actual_vec = at::vec::convert<float>(a_vec);
   auto expected_vec = b_vec;
 
-  for (int i = 0; i < at::vec::Vectorized<int>::size(); i++) {
-    EXPECT_EQ(expected_vec[i], actual_vec[i]);
-  }
+  ExpectVecEqual(expected_vec, actual_vec);
 }
 
 TEST(TestVec, TestClampMin) {
@@ -72,9 +81,7 @@ TEST(TestVec, TestClampMin) {
   Vec actual_vec = at::vec::clamp_min(a_vec, min_vec);
   Vec expected_vec = min_vec;
 
-  for (int i = 0; i < Vec::size(); i++) {
-    EXPECT_EQ(expected_vec[i], actual_vec[i]);
-  }
+  ExpectVecEqual(expected_vec, actual_vec);
 }
 
 } // namespace aot_inductor

test/cpp/aoti_inference/CMakeLists.txt

@@ -1,4 +1,3 @@
-
 set(AOT_INDUCTOR_TEST_ROOT ${TORCH_ROOT}/test/cpp/aoti_inference)
 
 # Build custom TorchScript op for AOTInductor
@@ -8,27 +7,12 @@ set_target_properties(aoti_custom_class PROPERTIES
 if(USE_CUDA)
   target_compile_definitions(aoti_custom_class PRIVATE USE_CUDA)
 elseif(USE_ROCM)
-    target_compile_definitions(aoti_custom_class PRIVATE USE_ROCM)
+  target_compile_definitions(aoti_custom_class PRIVATE USE_ROCM)
 endif()
+
 # Link against LibTorch
 target_link_libraries(aoti_custom_class torch)
 
-# the custom command that generates the TorchScript module
-add_custom_command(
-    OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/script_data.pt
-           ${CMAKE_CURRENT_BINARY_DIR}/script_model_cpu.pt
-           ${CMAKE_CURRENT_BINARY_DIR}/script_model_cuda.pt
-    # This script requires the torch package to be installed.
-    COMMAND python ${AOT_INDUCTOR_TEST_ROOT}/compile_model.py
-    DEPENDS torch torch_python aoti_custom_class ${AOT_INDUCTOR_TEST_ROOT}/compile_model.py
-)
-add_custom_target(aoti_script_model ALL
-    DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/script_data.pt
-    DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/script_model_cpu.pt
-    DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/script_model_cuda.pt
-)
-add_dependencies(aoti_script_model aoti_custom_class)
-
 # Build the cpp gtest binary containing the cpp-only tests.
 set(INDUCTOR_TEST_SRCS
   ${AOT_INDUCTOR_TEST_ROOT}/test.cpp
@@ -37,23 +21,12 @@ set(INDUCTOR_TEST_SRCS
 add_executable(test_aoti_inference
   ${TORCH_ROOT}/test/cpp/common/main.cpp
   ${INDUCTOR_TEST_SRCS}
-  data.pt
-  script_data.pt
-  script_model_cpu.pt
-  script_model_cuda.pt
 )
-add_dependencies(test_aoti_inference aoti_custom_class aoti_script_model)
+add_dependencies(test_aoti_inference aoti_custom_class)
 
 # TODO temporary until we can delete the old gtest polyfills.
 target_compile_definitions(test_aoti_inference PRIVATE USE_GTEST)
 
-# Define a custom command to generate the library
-add_custom_command(
-        OUTPUT data.pt
-        COMMAND python ${AOT_INDUCTOR_TEST_ROOT}/test.py
-        DEPENDS ${AOT_INDUCTOR_TEST_ROOT}/test.py
-)
-
 target_link_libraries(test_aoti_inference PRIVATE
   torch
   gtest_main
@@ -71,6 +44,10 @@ target_compile_definitions(test_aoti_inference PRIVATE
     CMAKE_CURRENT_BINARY_DIR=${CMAKE_CURRENT_BINARY_DIR}
 )
 
+target_compile_options_if_supported(test_aoti_inference -Wno-unused-variable)
+target_compile_options_if_supported(test_aoti_inference -Wno-unused-but-set-variable)
+target_compile_options_if_supported(test_aoti_inference -Wno-unused-function)
+
 if(INSTALL_TEST)
   install(TARGETS test_aoti_inference DESTINATION bin)
   # Install PDB files for MSVC builds

test/cpp/aoti_inference/test.cpp

@@ -2,7 +2,9 @@
 #include <gtest/gtest.h>
 #include <atomic>
 #include <condition_variable>
+#include <cstdlib>
 #include <filesystem>
+#include <fstream>
 #include <functional>
 #include <mutex>
 #include <queue>
@@ -28,6 +30,64 @@
 
 namespace {
 
+// Function to check if test data files exist and are valid
+bool testDataFilesExist() {
+  std::string bindir = STRINGIZE(CMAKE_CURRENT_BINARY_DIR);
+  std::array<std::string, 4> required_files = {
+      "data.pt",
+      "script_data.pt",
+      "script_model_cpu.pt",
+      "script_model_cuda.pt"};
+
+  for (const auto& filename : required_files) {
+    std::string filepath = bindir + "/" + filename;
+    std::ifstream file(filepath);
+    if (!file.good()) {
+      return false;
+    }
+  }
+  return true;
+}
+
+// Function to ensure test data files are generated at runtime
+void ensureTestDataGenerated() {
+  static std::once_flag generated_flag;
+  std::call_once(generated_flag, []() {
+    // Only generate if files don't exist or are placeholders
+    if (testDataFilesExist()) {
+      return;
+    }
+
+    std::string bindir = STRINGIZE(CMAKE_CURRENT_BINARY_DIR);
+
+    // Calculate path to source directory: build/test_aoti_inference -> build ->
+    // pytorch
+    std::string pytorch_root = bindir.substr(0, bindir.find_last_of("/"));
+    pytorch_root = pytorch_root.substr(0, pytorch_root.find_last_of("/"));
+    std::string source_dir = pytorch_root + "/test/cpp/aoti_inference";
+
+    // Generate test data files (data.pt, etc.) by running test.py directly
+    std::string test_script = source_dir + "/test.py";
+    std::string test_data_cmd = "cd " + bindir + " && python " + test_script;
+    std::cout << "Generating test data: " << test_data_cmd << std::endl;
+    int result1 = std::system(test_data_cmd.c_str());
+    if (result1 != 0) {
+      std::cerr << "Warning: Test data generation failed with code " << result1
+                << std::endl;
+    }
+
+    // Generate model files (script_*.pt) by running compile_model.py directly
+    std::string compile_script = source_dir + "/compile_model.py";
+    std::string models_cmd = "cd " + bindir + " && python " + compile_script;
+    std::cout << "Generating model files: " << models_cmd << std::endl;
+    int result2 = std::system(models_cmd.c_str());
+    if (result2 != 0) {
+      std::cerr << "Warning: Model generation failed with code " << result2
+                << std::endl;
+    }
+  });
+}
+
 const std::unordered_map<std::string, at::Tensor> derefTensorConstantMap(
     torch::inductor::TensorConstantMap tensor_constant_map) {
   std::unordered_map<std::string, at::Tensor> ret;
@@ -855,7 +915,6 @@ void test_aoti_free_buffer(bool use_runtime_constant_folding) {
   }
 }
 
-#if defined(USE_CUDA) || defined(USE_ROCM)
 void test_cuda_alloc_test() {
   torch::NoGradGuard no_grad;
 
@@ -895,8 +954,8 @@ void test_cuda_alloc_test() {
       runner->run(data_loader.attr(inputs_attr.c_str()).toTensorList().vec());
   ASSERT_TRUE(torch::allclose(ref_output_tensors[0], actual_output_tensors[0]));
 }
-#endif
 
+#ifdef USE_CUDA
 class ThreadPool {
  private:
   struct Task {
@@ -1037,86 +1096,96 @@ void test_multi_cuda_streams(const std::string& device) {
     ASSERT_TRUE(torch::allclose(ref_output_tensors[0], all_outputs[i][0]));
   }
 }
-#endif
+#endif // USE_CUDA
+#endif // USE_CUDA || USE_ROCM
 } // namespace
 
 namespace torch::aot_inductor {
 
-TEST(AotInductorTest, BasicTestCpu) {
+// Test fixture that ensures test data is generated once for all tests
+class AotInductorTest : public ::testing::Test {
+ public:
+  // This runs once before all tests in this test suite
+  static void SetUpTestSuite() {
+    ensureTestDataGenerated();
+  }
+};
+
+TEST_F(AotInductorTest, BasicTestCpu) {
   test_aoti("cpu", false);
 }
 
-TEST(AotInductorTest, BasicScriptTestCpu) {
+TEST_F(AotInductorTest, BasicScriptTestCpu) {
   test_aoti_script("cpu");
 }
 
-TEST(AotInductorTest, BasicPackageLoaderTestCpu) {
+TEST_F(AotInductorTest, BasicPackageLoaderTestCpu) {
   test_aoti_package_loader("cpu", false);
 }
 
-TEST(AotInductorTest, ExtractConstantsMapCpu) {
+TEST_F(AotInductorTest, ExtractConstantsMapCpu) {
   test_aoti_extract_constants_map("cpu");
 }
 
 #ifdef USE_CUDA
-TEST(AotInductorTest, BasicTestCuda) {
+TEST_F(AotInductorTest, BasicTestCuda) {
   test_aoti("cuda", true);
   test_aoti("cuda", false);
 }
 
-TEST(AotInductorTest, BasicScriptTestCuda) {
+TEST_F(AotInductorTest, BasicScriptTestCuda) {
   test_aoti_script("cuda");
 }
 
-TEST(AotInductorTest, BasicPackageLoaderTestCuda) {
+TEST_F(AotInductorTest, BasicPackageLoaderTestCuda) {
   test_aoti_package_loader("cuda", false);
 }
 
-TEST(AotInductorTest, BasicPackageLoaderTestMultiGpuCuda) {
+TEST_F(AotInductorTest, BasicPackageLoaderTestMultiGpuCuda) {
   test_aoti_package_loader_multi_gpu("cuda", false);
 }
 
-TEST(AotInductorTest, UpdateUserManagedConstantsCuda) {
+TEST_F(AotInductorTest, UpdateUserManagedConstantsCuda) {
   test_aoti_user_managed_buffer();
 }
 
-TEST(AotInductorTest, RuntimeUpdateConstantsCuda) {
+TEST_F(AotInductorTest, RuntimeUpdateConstantsCuda) {
   test_aoti_constants_update("cuda", true);
 }
 
-TEST(AotInductorTest, UpdateConstantsCuda) {
+TEST_F(AotInductorTest, UpdateConstantsCuda) {
   test_aoti_constants_update("cuda", false);
 }
 
-TEST(AotInductorTest, ExtractConstantsMapCuda) {
+TEST_F(AotInductorTest, ExtractConstantsMapCuda) {
   test_aoti_extract_constants_map("cuda");
 }
 
-TEST(AotInductorTest, RuntimeUpdateInactiveConstantsCuda) {
+TEST_F(AotInductorTest, RuntimeUpdateInactiveConstantsCuda) {
   test_aoti_double_buffering("cuda", true);
 }
 
-TEST(AotInductorTest, UpdateInactiveConstantsCuda) {
+TEST_F(AotInductorTest, UpdateInactiveConstantsCuda) {
   test_aoti_double_buffering("cuda", false);
 }
 
-TEST(AotInductorTest, UpdateInactiveConstantsWithTensorConstantsCuda) {
+TEST_F(AotInductorTest, UpdateInactiveConstantsWithTensorConstantsCuda) {
   test_aoti_double_buffering_with_tensor_constants();
 }
 
-TEST(AotInductorTest, FreeInactiveConstantBufferCuda) {
+TEST_F(AotInductorTest, FreeInactiveConstantBufferCuda) {
   test_aoti_free_buffer(false);
 }
 
-TEST(AotInductorTest, FreeInactiveConstantBufferRuntimeConstantFoldingCuda) {
+TEST_F(AotInductorTest, FreeInactiveConstantBufferRuntimeConstantFoldingCuda) {
   test_aoti_free_buffer(true);
 }
 
-TEST(AotInductorTest, MultiStreamTestCuda) {
+TEST_F(AotInductorTest, MultiStreamTestCuda) {
   test_multi_cuda_streams("cuda");
 }
 
-TEST(AotInductorTest, CudaAllocTestCuda) {
+TEST_F(AotInductorTest, CudaAllocTestCuda) {
   test_cuda_alloc_test();
 }
 #endif

test/cpp/api/autograd.cpp

@@ -584,7 +584,7 @@ TEST(CustomAutogradTest, MarkDirty) {
     }
   };
 
-  // Clone here because modifying leafs inplace is not allowed
+  // Clone here because modifying leaves inplace is not allowed
   auto x = torch::randn({5, 5}, torch::requires_grad()).clone();
   auto version_before = x._version();
   auto out = MyFunction::apply(x);

test/cpp/api/parallel.cpp

@@ -264,7 +264,7 @@ TEST_F(ParallelTest, DataParallelNumericalEquivalence_MultiCUDA) {
     input += i;
     input_dp += i;
 
-    // non-prallel training
+    // non-parallel training
     torch::optim::SGD optim(model->parameters(), torch::optim::SGDOptions(0.1));
     auto output = model->forward(input);
     auto loss = torch::mse_loss(output, torch::zeros_like(output));

test/cpp_extensions/open_registration_extension/torch_openreg/README.md

@@ -149,8 +149,8 @@ When `import torch`, installed accelerators (such as `torch_openreg`) will be au
 ### Installation
 
 ```python
-pip3 install --no-build-isolation -e . # for develop
-pip3 install --no-build-isolation . # for install
+python -m pip install --no-build-isolation -e . # for develop
+python -m pip install --no-build-isolation . # for install
 ```
 
 ### Usage Example
@@ -188,7 +188,7 @@ Please refer to [this](https://docs.pytorch.org/docs/main/accelerator/index.html
   - Device-agnostic APIs
   - Memory Management
   - Generator
-  - Distrubuted
+  - Distributed
   - Custom Tensor&Storage
   - ...
 - **Improve Tests**: Add more test cases related to the integration mechanism.

test/cpp_extensions/open_registration_extension/torch_openreg/tests/test_streams.py

@@ -5,6 +5,7 @@ from torch.testing._internal.common_utils import run_tests, skipIfTorchDynamo, T
 
 
 class TestStream(TestCase):
+    @skipIfTorchDynamo()
     def test_stream_create(self):
         stream = torch.Stream(device="openreg")
         self.assertEqual(stream.device_index, torch.openreg.current_device())
@@ -24,6 +25,7 @@ class TestStream(TestCase):
         )
         self.assertEqual(stream, stream1)
 
+    @skipIfTorchDynamo()
     def test_stream_context(self):
         with torch.Stream(device="openreg:1") as stream:
             self.assertEqual(torch.accelerator.current_stream(), stream)
@@ -40,6 +42,7 @@ class TestStream(TestCase):
         current_stream = torch.accelerator.current_stream()
         self.assertEqual(current_stream, stream2)
 
+    @skipIfTorchDynamo()
     def test_stream_synchronize(self):
         stream = torch.Stream(device="openreg:1")
         self.assertEqual(True, stream.query())
@@ -49,12 +52,14 @@ class TestStream(TestCase):
         stream.synchronize()
         self.assertEqual(True, stream.query())
 
+    @skipIfTorchDynamo()
     def test_stream_repr(self):
         stream = torch.Stream(device="openreg:1")
         self.assertTrue(
             "torch.Stream device_type=openreg, device_index=1" in repr(stream)
         )
 
+    @skipIfTorchDynamo()
     def test_stream_wait_stream(self):
         stream_1 = torch.Stream(device="openreg:0")
         stream_2 = torch.Stream(device="openreg:1")

test/distributed/_composable/test_composability/test_2d_composability.py

@@ -218,7 +218,7 @@ class TestFullyShard2DTraining(FSDPTest):
 
         torch.manual_seed(42 + global_mesh.get_local_rank("dp"))
         inp = torch.randint(0, model_args.vocab_size, (2, 16), device=device_type)
-        for iter_idx in range(5):
+        for _ in range(5):
             ref_loss = ref_model(inp).sum()
             loss = model(inp).sum()
             self.assertEqual(ref_loss, loss)
@@ -238,9 +238,7 @@ class TestFullyShard2DTraining(FSDPTest):
             # runs its reduce-scatter
             self.assertIsInstance(model.pos_embeddings.weight.placements[1], Shard)
             self.assertIsInstance(model.pos_embeddings.weight.grad.placements[1], Shard)
-            for ref_param, (param_name, param) in zip(
-                ref_model.parameters(), model.named_parameters()
-            ):
+            for ref_param, param in zip(ref_model.parameters(), model.parameters()):
                 full_grad = param.grad.full_tensor()
                 self.assertEqual(ref_param.grad, full_grad)
 

test/distributed/_composable/test_composability/test_pp_composability.py

@@ -336,7 +336,7 @@ class ComposabilityTest(MultiProcessTestCase):
             for model in partial_models
         ]
 
-        for train_step in range(5):
+        for _train_step in range(5):
             for optimizer in optimizers:
                 optimizer.zero_grad()
             inputs = torch.rand((num_microbatches, dim), device=self.device)
@@ -517,7 +517,7 @@ class ComposabilityTest(MultiProcessTestCase):
             for model in ref_partial_models
         ]
 
-        for train_step in range(5):
+        for _train_step in range(5):
             for optimizer in optimizers:
                 optimizer.zero_grad()
             for ref_optimizer in ref_optimizers:

test/distributed/checkpoint/test_planner.py

@@ -216,7 +216,7 @@ class TestSavePlan(TestCase):
         # Number of plans should remain unchanged
         self.assertEqual(len(all_plans), len(deduped_plans))
 
-        # Numer of items in the deduped plans should be less than the original plans
+        # Number of items in the deduped plans should be less than the original plans
         for new_plan, old_plan in zip(deduped_plans, all_plans):
             self.assertFalse(_compare_save_plans(new_plan, old_plan))
             self.assertTrue(len(new_plan.items) < len(old_plan.items))

test/distributed/checkpoint/test_state_dict.py

@@ -4,7 +4,7 @@ import copy
 import functools
 import sys
 from collections.abc import Callable
-from itertools import chain
+from itertools import chain, product
 from typing import Union
 
 import torch
@@ -708,29 +708,43 @@ class TestStateDict(DTensorTestBase, VerifyStateDictMixin):
     @with_comms
     @skip_if_lt_x_gpu(2)
     def test_flattened_osd(self) -> None:
-        device_mesh = init_device_mesh(device_type, (self.world_size,))
-        model = CompositeParamModel(device=torch.device(device_type))
-        fsdp_model = fully_shard(copy.deepcopy(model), mesh=device_mesh)
-        fsdp_optim = torch.optim.AdamW(fsdp_model.parameters())
-        batch = torch.rand(8, 100, device=device_type)
-        fsdp_model(batch).sum().backward()
-        fsdp_optim.step()
-        fsdp_optim.zero_grad()
-        osd1 = get_optimizer_state_dict(fsdp_model, fsdp_optim)
-        osd2 = get_optimizer_state_dict(
-            fsdp_model,
-            fsdp_optim,
-            options=StateDictOptions(flatten_optimizer_state_dict=True),
-        )
-        fsdp_optim2 = torch.optim.AdamW(fsdp_model.parameters())
-        set_optimizer_state_dict(
-            fsdp_model, optimizers=fsdp_optim2, optim_state_dict=osd2
-        )
-        self.assertEqual(fsdp_optim.state_dict(), fsdp_optim2.state_dict())
-        set_optimizer_state_dict(
-            fsdp_model, optimizers=fsdp_optim2, optim_state_dict=osd1
-        )
-        self.assertEqual(fsdp_optim.state_dict(), fsdp_optim2.state_dict())
+        """
+        Test flattened optimizer state dictionaries with different combinations of
+        flatten_optimizer_state_dict flag for saving and loading.
+
+        This test verifies that:
+        1. We can save optimizer state dict with/without flattening
+        2. We can load optimizer state dict with/without flattening
+        3. The resulting optimizer state is equivalent regardless of flattening options
+        """
+        for flatten_to_save, flatten_to_load in product([True, False], repeat=2):
+            device_mesh = init_device_mesh(device_type, (self.world_size,))
+            model = CompositeParamModel(device=torch.device(device_type))
+            fsdp_model = fully_shard(copy.deepcopy(model), mesh=device_mesh)
+            fsdp_optim = torch.optim.AdamW(fsdp_model.parameters())
+            batch = torch.rand(8, 100, device=device_type)
+            fsdp_model(batch).sum().backward()
+            fsdp_optim.step()
+            fsdp_optim.zero_grad()
+
+            # Get optimizer state dict with/without flattening option
+            osd = get_optimizer_state_dict(
+                fsdp_model,
+                fsdp_optim,
+                options=StateDictOptions(flatten_optimizer_state_dict=flatten_to_save),
+            )
+
+            # Create a new optimizer and load the state from osd
+            fsdp_optim2 = torch.optim.AdamW(fsdp_model.parameters())
+            set_optimizer_state_dict(
+                fsdp_model,
+                optimizers=fsdp_optim2,
+                optim_state_dict=osd,
+                options=StateDictOptions(flatten_optimizer_state_dict=flatten_to_load),
+            )
+
+            # Verify the loaded optimizer state matches the original
+            self.assertEqual(fsdp_optim.state_dict(), fsdp_optim2.state_dict())
 
     def _test_deprecate_partial(self) -> None:
         model = CompositeParamModel(device=torch.device(device_type))

test/distributed/fsdp/test_distributed_checkpoint.py

@@ -31,17 +31,17 @@ if TEST_WITH_DEV_DBG_ASAN:
     sys.exit(0)
 
 
-_DISTRIBUTED_STATE_DICT_IMPLS = (
+_DISTRIBUTED_STATE_DICT_IMPLS = {
     StateDictType.LOCAL_STATE_DICT,
     StateDictType.SHARDED_STATE_DICT,
-)
+}
 
 
 class TestDistributedCheckpoint(FSDPTest):
     @property
     def world_size(self):
-        if torch.cuda.is_available():
-            gpu_cnt = torch.cuda.device_count()
+        if torch.accelerator.is_available():
+            gpu_cnt = torch.accelerator.device_count()
             if gpu_cnt < 2:
                 return gpu_cnt
         return 2
@@ -93,7 +93,9 @@ class TestDistributedCheckpoint(FSDPTest):
         # TODO: add resharding test case.
 
 
-devices = ("cuda", "hpu")
-instantiate_device_type_tests(TestDistributedCheckpoint, globals(), only_for=devices)
+devices = ("cuda", "hpu", "xpu")
+instantiate_device_type_tests(
+    TestDistributedCheckpoint, globals(), only_for=devices, allow_xpu=True
+)
 if __name__ == "__main__":
     run_tests()

test/distributed/fsdp/test_fsdp_apply.py

@@ -36,8 +36,8 @@ device_type = torch.device(get_devtype())
 class TestApply(FSDPTest):
     @property
     def world_size(self):
-        if torch.cuda.is_available():
-            gpu_cnt = torch.cuda.device_count()
+        if torch.accelerator.is_available():
+            gpu_cnt = torch.accelerator.device_count()
             if gpu_cnt < 2:
                 return gpu_cnt
         return 2

test/distributed/logging_utils.py

@@ -0,0 +1,43 @@
+import logging
+import time
+
+
+_start_time = time.time()
+_logger = logging.getLogger(__name__)
+
+
+def _ts():
+    return time.time() - _start_time
+
+
+def configure(level=logging.INFO, force=False):
+    try:
+        logging.basicConfig(
+            level=level,
+            format="%(asctime)s %(name)s %(levelname)s: %(message)s",
+            force=force,
+        )
+    except TypeError:
+        logging.basicConfig(
+            level=level, format="%(asctime)s %(name)s %(levelname)s: %(message)s"
+        )
+
+
+def log_test_info(rank, message):
+    _logger.info("[%7.3fs][Rank %s] %s", _ts(), rank, message)
+
+
+def log_test_success(rank, message):
+    _logger.info("[%7.3fs][Rank %s] ✅ %s", _ts(), rank, message)
+
+
+def log_test_validation(rank, message):
+    _logger.info("[%7.3fs][Rank %s] ✓ %s", _ts(), rank, message)
+
+
+def log_test_warning(rank, message):
+    _logger.warning("[%7.3fs][Rank %s] ⚠️ %s", _ts(), rank, message)
+
+
+def log_test_error(rank, message):
+    _logger.error("[%7.3fs][Rank %s] ✗ %s", _ts(), rank, message)

test/distributed/tensor/test_attention.py

@@ -52,7 +52,9 @@ from torch.testing._internal.common_cuda import (
 from torch.testing._internal.common_distributed import skip_if_lt_x_gpu
 from torch.testing._internal.common_utils import run_tests, skipIfRocm
 from torch.testing._internal.distributed._tensor.common_dtensor import (
+    create_local_tensor_test_class,
     DTensorTestBase,
+    map_local_tensor_for_rank,
     with_comms,
 )
 
@@ -158,7 +160,7 @@ class RingAttentionTest(DTensorTestBase):
             # parameters because when require_grad is True, resize_ is not
             # allowed. But requires_grad of cp_q, cp_k, and cp_v are False
             # now. So we can just use context_parallel() to shard q, k, v.
-            # In reality, context_paralle() should be used to shard the input.
+            # In reality, context_parallel() should be used to shard the input.
             # In reality, context_parallel() should only be used to shard
             # the model inputs (batch).
 
@@ -701,7 +703,7 @@ class CPFlexAttentionTest(DTensorTestBase):
         )
 
         # TODO: change this for-loop to run_subtests
-        # Use a for-loop instead of run_subtests because we need to intialize the mask
+        # Use a for-loop instead of run_subtests because we need to initialize the mask
         # for each subtest. This can be baked into self._test_cp_flex_attention as
         # a str argument denoting mask type.
         for batch_size, max_seq_len, lb_type in itertools.product(
@@ -800,11 +802,47 @@ class TestSharding(DTensorTestBase):
         chunks = freqs_cis.chunk(self.world_size * 2)
         self.assertEqual(
             freqs_cis_shard,
-            torch.cat(
-                [chunks[self.rank], chunks[self.world_size * 2 - self.rank - 1]], dim=0
+            map_local_tensor_for_rank(
+                chunks,
+                self.rank,
+                lambda chunks, rank: torch.cat(
+                    [chunks[rank], chunks[self.world_size * 2 - rank - 1]],
+                    dim=0,
+                ),
             ),
         )
 
 
+RingAttentionTestWithLocalTensor = create_local_tensor_test_class(
+    RingAttentionTest,
+    skipped_tests=[
+        # Need to make attention implementation local tensor friendly, e.g.
+        # rewrite "rank local" logic
+        "test_ring_attention_sdpa",
+    ],
+)
+
+CPFlexAttentionTestWithLocalTensor = create_local_tensor_test_class(
+    CPFlexAttentionTest,
+    skipped_tests=[
+        # Missing support for batched tensors
+        "test_cp_flex_attention_causal_mask",
+        "test_cp_flex_attention_document_mask",
+    ],
+)
+
+TestCPCustomOpsWithLocalTensor = create_local_tensor_test_class(
+    TestCPCustomOps,
+    skipped_tests=[
+        # Missing support for fake tensors
+        "test_flex_cp_custom_op",
+    ],
+)
+
+TestShardingWithLocalTensor = create_local_tensor_test_class(
+    TestSharding,
+)
+
+
 if __name__ == "__main__":
     run_tests()

test/distributed/tensor/test_convolution_ops.py

@@ -16,6 +16,7 @@ from torch.distributed.tensor import (
 from torch.nn import functional as F
 from torch.testing._internal.common_utils import run_tests
 from torch.testing._internal.distributed._tensor.common_dtensor import (
+    create_local_tensor_test_class,
     DTensorTestBase,
     skip_if_lt_x_gpu,
     with_comms,
@@ -232,5 +233,17 @@ class DistConvolutionOpsTest(DTensorTestBase):
         self.assertEqual(out_dt.shape, out.shape)
 
 
+DistConvolutionOpsTestWithLocalTensor = create_local_tensor_test_class(
+    DistConvolutionOpsTest,
+    # Send / recv ops are not supported
+    skipped_tests=[
+        "test_conv1d",
+        "test_conv3d",
+        "test_conv_backward_none_grad_inp",
+        "test_depthwise_convolution",
+        "test_downsampling_convolution",
+    ],
+)
+
 if __name__ == "__main__":
     run_tests()

test/distributed/tensor/test_dtensor_compile.py

@@ -464,25 +464,6 @@ def forward(self, b_parametrizations_buffer_original0, x):
         run(g, 64, 8)
         self.assertEqual(cnt.frame_count, 2)
 
-    def test_dtensor_requires_grad_recompile(self):
-        cnt = torch._dynamo.testing.CompileCounterWithBackend("aot_eager")
-        mesh = DeviceMesh(self.device_type, torch.arange(self.world_size))
-
-        @torch.compile(backend=cnt, fullgraph=True)
-        def f(x):
-            y = x * x
-            return y.to_local()
-
-        full_x = torch.randn(8, 8, requires_grad=False)
-        x = distribute_tensor(full_x, mesh, [Shard(0)])
-        f(x)
-
-        full_x = torch.randn(8, 8, requires_grad=True)
-        x = distribute_tensor(full_x, mesh, [Shard(0)])
-        f(x)
-
-        self.assertEqual(cnt.frame_count, 2)
-
     def test_dtensor_attribute_access_on_intermediate(self):
         mesh = DeviceMesh(self.device_type, torch.arange(self.world_size))
 

test/distributed/tensor/test_dtensor_export.py

@@ -22,7 +22,11 @@ from torch.distributed.tensor.parallel import (
     parallelize_module,
     RowwiseParallel,
 )
-from torch.nn.attention.flex_attention import create_block_mask, flex_attention
+from torch.nn.attention.flex_attention import (
+    BlockMask,
+    create_block_mask,
+    flex_attention,
+)
 from torch.testing._internal.common_utils import (
     instantiate_parametrized_tests,
     parametrize,
@@ -32,6 +36,7 @@ from torch.testing._internal.common_utils import (
 )
 from torch.testing._internal.distributed._tensor.common_dtensor import MLPModule
 from torch.testing._internal.distributed.fake_pg import FakeStore
+from torch.utils._pytree import register_pytree_node
 
 
 class SimpleModel(torch.nn.Module):
@@ -176,6 +181,15 @@ def _count_op(gm, target):
     return sum(1 for node in gm.graph.nodes if node.target == target)
 
 
+register_pytree_node(
+    BlockMask,
+    BlockMask._flatten,
+    BlockMask._unflatten,
+    flatten_with_keys_fn=BlockMask._flatten_with_keys,
+    serialized_type_name="torch.nn.attention.flex_attention.BlockMask",
+)
+
+
 @requires_cuda
 class DTensorExportTest(TestCase):
     def tearDown(self):
@@ -506,6 +520,23 @@ class DTensorExportTest(TestCase):
             2,
         )
 
+    # "Explanation: SourcelessBuilder.create does not know how to wrap <class 'types.UnionType'>"
+    @unittest.expectedFailure
+    def test_union_typed_annotation(self):
+        def fn(leaf: torch.Tensor | DTensor):
+            def nest_fn(leaf: torch.Tensor | DTensor):
+                # def nest_fn(leaf: Union[torch.Tensor, DTensor]):  # this works
+                if isinstance(leaf, DTensor):
+                    leaf = leaf.to_local()
+                return leaf
+
+            return nest_fn(leaf) + 1
+
+        z = torch.randn(16, 16)
+        gm = graph_capture_and_aot_export_joint_with_descriptors(fn, (z,))
+
+        print(gm)
+
 
 instantiate_parametrized_tests(DTensorExportTest)
 

test/distributed/tensor/test_embedding_ops.py

@@ -168,7 +168,7 @@ class TestEmbeddingOp(DTensorTestBase):
         self._run_embedding_op_test(mesh, 0, [6, 7, 6], 13, 22)
         self._run_embedding_op_test(mesh, 0, [34], 15, 14, padding_idx=10)
 
-        from torch.distributed.tensor._ops._embedding_ops import _MaskPartial
+        from torch.distributed.tensor.placement_types import MaskPartial
 
         # test collectives
         embedding_mod = torch.nn.Embedding(10, 20, device=self.device_type)
@@ -176,7 +176,7 @@ class TestEmbeddingOp(DTensorTestBase):
         inp = torch.randint(0, 10, (8, 8), device=self.device_type)
         replicated_inp = DTensor.from_local(inp, mesh, [Replicate()], run_check=False)
         output = sharded_embedding(replicated_inp)
-        self.assertIsInstance(output.placements[0], _MaskPartial)
+        self.assertIsInstance(output.placements[0], MaskPartial)
 
         comm_mode = CommDebugMode()
 
@@ -192,9 +192,9 @@ class TestEmbeddingOp(DTensorTestBase):
         inp = torch.randint(0, 10, (4, 4), device=self.device_type)
         replicated_inp = DTensor.from_local(inp, mesh, [Replicate()], run_check=False)
 
-        from torch.distributed.tensor._ops._embedding_ops import _MaskPartial
+        from torch.distributed.tensor.placement_types import MaskPartial
 
-        # case 1: two embeddings with the same shape, thus sharing the underlying _MaskPartial
+        # case 1: two embeddings with the same shape, thus sharing the underlying MaskPartial
         # and MaskBuffer, because of cache hit from sharding propagation
 
         emb1 = torch.nn.Embedding(10, 23, device=self.device_type)
@@ -206,23 +206,23 @@ class TestEmbeddingOp(DTensorTestBase):
         output2 = sharded_emb2(replicated_inp)
 
         partial_placement1 = output1.placements[0]
-        self.assertIsInstance(partial_placement1, _MaskPartial)
+        self.assertIsInstance(partial_placement1, MaskPartial)
         output1.full_tensor()
 
         partial_placement2 = output2.placements[0]
-        self.assertIsInstance(partial_placement2, _MaskPartial)
+        self.assertIsInstance(partial_placement2, MaskPartial)
         output2.full_tensor()
 
         self.assertTrue(id(partial_placement1), id(partial_placement2))
 
         # case 2: two embeddings with the same logical_dim_size, but different logical_shape
-        # thus they will have different _MaskPartial placements (with no cache hit)
+        # thus they will have different MaskPartial placements (with no cache hit)
 
         emb3 = torch.nn.Embedding(10, 29, device=self.device_type)
         sharded_emb3 = self._apply_sharding(emb3, 0, mesh)
         output3 = sharded_emb3(replicated_inp)
         partial_placement3 = output3.placements[0]
-        self.assertIsInstance(partial_placement3, _MaskPartial)
+        self.assertIsInstance(partial_placement3, MaskPartial)
         output2.full_tensor()
 
         # not equal because of different logical_shape, despite of same logical_dim_size

test/distributed/tensor/test_math_ops.py

@@ -60,9 +60,9 @@ class DistMathOpsTest(DTensorTestBase):
         shard_spec = [Shard(0)]
 
         tensor = torch.randn(12, 8, 8)
-        # TODO: check `all` correctness and test `all` on a bool tensor
-        if op_str in ("any"):
-            # test out a bool tensor for any
+        if op_str in ("any", "all"):
+            # Test bool tensor for any() and all() reduction ops
+            # Previously all() had a bug using sum reduction instead of product
             tensor = tensor < 0
         dtensor = distribute_tensor(tensor, device_mesh, shard_spec)
 

test/distributed/tensor/test_tensor_ops.py

@@ -511,7 +511,7 @@ class DistTensorOpsTest(DTensorTestBase):
         # case 2 input sharding: input sharded, index replicated, output mask partial
         # only works when index has size 1 on the gather dimension and
         # input is sharded on the gather dimension
-        from torch.distributed.tensor._ops._embedding_ops import _MaskPartial
+        from torch.distributed.tensor.placement_types import MaskPartial
 
         gather_dim = 1
         global_input = torch.randn(12, 8, 16)
@@ -522,7 +522,7 @@ class DistTensorOpsTest(DTensorTestBase):
         with comm_mode:
             output_dt = torch.gather(input_dt, gather_dim, index_dt)
             self.assertEqual(comm_mode.get_total_counts(), 0)
-        self.assertIsInstance(output_dt.placements[0], _MaskPartial)
+        self.assertIsInstance(output_dt.placements[0], MaskPartial)
         self.assertEqual(output_dt.full_tensor(), global_output)
 
         # case 3 index sharding: input replicated, index sharded, output sharded

test/distributed/test_aten_comm_compute_reordering.py

@@ -887,6 +887,62 @@ class TestComputeCommReorderingBucketing(TestComputeCommReorderingMultiProc):
             correct = func(a, b, c, d, ranks=ranks)
             self.assertTrue(same(test_out, correct))
 
+    @unittest.skipIf(not HAS_GPU, "Inductor+gpu needs triton and recent GPU arch")
+    @torch._inductor.config.patch(get_bucket_patches())
+    def test_custom_estimation_with_fake_tensor_mode(self):
+        """Test that custom estimation can use FakeTensorMode for analysis."""
+        from torch._subclasses.fake_tensor import FakeTensorMode
+
+        estimation_calls = 0
+
+        def estimate_with_fake_mode(fx_node, compute_multiplier=1.0):
+            with FakeTensorMode():
+                nonlocal estimation_calls
+                estimation_calls += 1
+                assert isinstance(torch.rand([20]), torch._subclasses.FakeTensor)
+
+            return 1.0
+
+        patches = get_bucket_patches()
+        patches["aten_distributed_optimizations.custom_runtime_estimation"] = (
+            estimate_with_fake_mode
+        )
+
+        def func(a, b, *, ranks):
+            # Two independent all_gathers that should be bucketed
+            ag1 = _functional_collectives.all_gather_tensor(a, 0, ranks)
+            ag2 = _functional_collectives.all_gather_tensor(b, 0, ranks)
+
+            # Matmul that can hide the collectives
+            mm1 = torch.matmul(a, a)
+
+            return ag1.sum() + ag2.sum() + mm1.sum()
+
+        with _dynamo_dist_per_rank_init(
+            self.rank,
+            self.world_size,
+            self.backend(device_type),
+            fake_pg=not at_least_x_gpu(2),
+        ):
+            inputs_a = torch.ones(4, 4, dtype=torch.float, device=device_type)
+            inputs_b = torch.ones(4, 4, dtype=torch.float, device=device_type) * 2
+            ranks = list(range(self.world_size))
+
+            func_c = functools.partial(func, ranks=ranks)
+            with torch._inductor.config.patch(patches):
+                compiled = torch.compile(func_c)
+                out, aten_graph_str = run_and_get_aten_graph(
+                    compiled, inputs_a, inputs_b
+                )
+
+            # Verify the custom estimation was called
+            self.assertTrue(
+                estimation_calls > 0, "Custom estimation should have been called"
+            )
+
+            correct = func(inputs_a, inputs_b, ranks=ranks)
+            self.assertTrue(same(out, correct))
+
 
 if __name__ == "__main__":
     from torch._dynamo.test_case import run_tests

test/distributed/test_c10d_nccl.py

@@ -2,6 +2,7 @@
 
 import copy
 import json
+import logging
 import os
 import pickle
 import random
@@ -21,6 +22,7 @@ from unittest import mock, SkipTest
 import torch
 import torch.distributed as c10d
 import torch.distributed._functional_collectives as _functional_collectives
+from torch.distributed.distributed_c10d import SHRINK_ABORT as NCCL_SHRINK_ABORT
 
 
 if not c10d.is_available() or not c10d.is_nccl_available():
@@ -47,12 +49,15 @@ from torch._C._distributed_c10d import ErrorType, OpType, WorkResult
 from torch.nn.parallel import DistributedDataParallel
 from torch.testing._internal.common_cuda import _get_torch_rocm_version, TEST_MULTIGPU
 from torch.testing._internal.common_distributed import (
+    get_required_world_size,
     get_timeout,
     init_multigpu_helper,
     MultiProcessTestCase,
     requires_multicast_support,
     requires_nccl,
+    requires_nccl_shrink,
     requires_nccl_version,
+    requires_world_size,
     skip_if_lt_x_gpu,
     skip_if_rocm_multiprocess,
     sm_is_or_higher_than,
@@ -87,6 +92,17 @@ BFLOAT16_AVAILABLE = torch.cuda.is_available() and (
     torch.version.cuda is not None or torch.version.hip is not None
 )
 
+from logging_utils import (
+    configure as _log_configure,
+    log_test_info,
+    log_test_success,
+    log_test_validation,
+    log_test_warning,
+)
+
+
+_log_configure(level=logging.INFO, force=True)
+
 
 class RendezvousEnvTest(TestCase):
     @retry_on_connect_failures
@@ -317,7 +333,7 @@ class ProcessGroupNCCLGroupTest(MultiProcessTestCase):
 
     @property
     def world_size(self):
-        return 2
+        return get_required_world_size(self, 2)
 
     @property
     def rank_to_GPU(self):
@@ -1255,6 +1271,628 @@ class ProcessGroupNCCLGroupTest(MultiProcessTestCase):
         pg_2 = c10d.new_group([0, 1])
         self.assertEqual(pg_2.group_desc, "undefined")
 
+    @requires_nccl_shrink()
+    @requires_world_size(2)
+    def test_shrink_group_basic(self):
+        """Test basic shrink_group functionality."""
+        self._perform_shrink_test([1], "Basic shrink test")
+
+    @requires_nccl_shrink()
+    @requires_world_size(2)
+    def test_shrink_group_validation(self):
+        """Test input validation in shrink_group."""
+        device, pg = self._setup_shrink_test("validation")
+
+        def _test_invalid_input(ranks, description, expected_exception):
+            """Helper to test invalid inputs."""
+            try:
+                c10d.shrink_group(ranks)
+                self.fail(f"Expected {expected_exception.__name__} for {description}")
+            except expected_exception:
+                log_test_validation(self.rank, f"✓ {description}")
+            except Exception:
+                if expected_exception is Exception:  # Accept any exception
+                    log_test_validation(self.rank, f"✓ {description}")
+                else:
+                    raise
+
+        # Test cases
+        _test_invalid_input([], "Empty exclusion list", ValueError)
+        if self.world_size > 1:
+            _test_invalid_input([0, 0, 1], "Duplicate ranks", Exception)
+        _test_invalid_input([self.world_size + 1], "Out of bounds rank", Exception)
+
+        log_test_success(self.rank, "All validation tests passed")
+        dist.destroy_process_group()
+
+    @requires_nccl_shrink()
+    @requires_world_size(2)
+    def test_shrink_group_backend_properties(self):
+        """Test that backend properties are preserved after shrinking."""
+
+        test_name = "Backend Properties Test"
+        ranks_to_exclude = [0]
+
+        # Reuse _setup_shrink_test for complete setup (device, environment, and process group)
+        device, pg = self._setup_shrink_test("backend_properties")
+
+        # Follow _perform_shrink_test pattern from here
+        log_test_info(self.rank, f"{test_name} (world_size={self.world_size})")
+
+        is_excluded = self.rank in ranks_to_exclude
+        log_test_info(
+            self.rank,
+            f"Excluding ranks: {ranks_to_exclude}, am_excluded: {is_excluded}",
+        )
+
+        # Store original backend property values (not references) before shrinking
+        original_timeout = None
+        original_high_priority = None
+        if not is_excluded:
+            original_backend = pg._get_backend(device)
+            original_timeout = original_backend.options._timeout
+            original_high_priority = original_backend.options.is_high_priority_stream
+            log_test_info(
+                self.rank,
+                f"Storing original backend properties: timeout={original_timeout}, high_priority={original_high_priority}",
+            )
+
+        if is_excluded:
+            log_test_info(
+                self.rank,
+                f"Excluded rank {self.rank} - setup complete, skipping shrink operation",
+            )
+            dist.destroy_process_group()  # hang without it
+            return
+
+        # Only non-excluded ranks proceed with shrink (same as _perform_shrink_test)
+        log_test_info(self.rank, "Non-excluded rank calling shrink_group")
+        shrunk_pg = c10d.shrink_group(ranks_to_exclude)
+
+        # Reuse _validate_shrunk_group helper (same as _perform_shrink_test)
+        expected_size = self.world_size - len(ranks_to_exclude)
+        _ = self._validate_shrunk_group(shrunk_pg, expected_size, test_name)
+
+        # Add custom backend properties validation
+        new_backend = shrunk_pg._get_backend(device)
+        log_test_info(self.rank, "Validating backend properties are preserved")
+
+        new_timeout = new_backend.options._timeout
+        new_high_priority = new_backend.options.is_high_priority_stream
+
+        log_test_info(
+            self.rank,
+            f"Timeout comparison - original: {original_timeout}, new: {new_timeout}",
+        )
+        self.assertEqual(
+            original_timeout, new_timeout, f"{test_name}: timeout not preserved"
+        )
+
+        log_test_info(
+            self.rank,
+            f"High priority stream comparison - original: {original_high_priority}, new: {new_high_priority}",
+        )
+        self.assertEqual(
+            original_high_priority,
+            new_high_priority,
+            f"{test_name}: high_priority_stream not preserved",
+        )
+
+        log_test_validation(
+            self.rank, f"{test_name}: Backend properties preserved successfully"
+        )
+        log_test_success(
+            self.rank, f"{test_name} successful (shrink + backend validation)"
+        )
+
+        # Cleanup (same as _perform_shrink_test)
+        dist.destroy_process_group()
+
+    @requires_nccl_shrink()
+    @requires_world_size(2)
+    def test_shrink_group_multiple_comms(self):
+        """Test shrink_group with multiple communicators and subgroup invalidation."""
+
+        device, pg = self._setup_shrink_test("multiple_comms")
+
+        # Create subgroup [0, 1] and test shrinking it
+        subgroup = c10d.new_group([0, 1])
+        if self.rank <= 1:
+            # Shrink subgroup: exclude rank 1
+            if self.rank == 0:  # Only rank 0 remains
+                shrunk_subgroup = c10d.shrink_group([1], group=subgroup)
+                self.assertEqual(shrunk_subgroup.size(), 1)
+                # Test communication on shrunk subgroup
+                tensor = torch.full((1,), self.rank).cuda(device)
+                c10d.all_reduce(tensor, group=shrunk_subgroup)
+                self.assertEqual(tensor.item(), 0)  # Only rank 0
+                log_test_success(self.rank, "Subgroup shrinking successful")
+
+        dist.barrier()  # Sync before default group test
+
+        # Shrink default group: exclude last rank
+        ranks_to_exclude = [self.world_size - 1]
+        if self.rank not in ranks_to_exclude:
+            shrunk_default = c10d.shrink_group(ranks_to_exclude)
+            expected_size = self.world_size - 1
+            self.assertEqual(shrunk_default.size(), expected_size)
+
+            # Test collective on shrunk default group
+            tensor = torch.full((1,), self.rank).cuda(device)
+            c10d.all_reduce(tensor, group=shrunk_default)
+            expected_sum = sum(
+                range(self.world_size - 1)
+            )  # 0 + 1 + ... + (world_size-2)
+            self.assertEqual(tensor.item(), expected_sum)
+            log_test_success(self.rank, "Default group shrinking successful")
+
+            # Note: After shrinking default group, the old subgroup is invalid
+            # due to global rank reassignment
+
+        dist.destroy_process_group()
+
+    def _test_shrink_group_with_flag(self, shrink_flag, flag_name, rank_to_exclude):
+        """Helper method to test shrink_group with a specific flag."""
+        if self.world_size < 2:
+            log_test_info(self.rank, f"Skipping (needs ≥2 GPUs, got {self.world_size})")
+            return
+        ranks_to_exclude = [rank_to_exclude]
+        log_test_info(self.rank, f"Using {flag_name} flag (value: {shrink_flag})")
+        if flag_name == "NCCL_SHRINK_ABORT":
+            log_test_info(
+                self.rank,
+                "ABORT flag will terminate ongoing operations before shrinking",
+            )
+
+        self._perform_shrink_test(
+            ranks_to_exclude, f"{flag_name} flag test", shrink_flags=shrink_flag
+        )
+
+    @requires_nccl_shrink()
+    @requires_world_size(2)
+    def test_shrink_group_flags(self):
+        """Test shrink_group with different shrink flags."""
+        # Test ABORT flags
+        log_test_info(self.rank, "Testing NCCL_SHRINK_ABORT flag")
+        self._test_shrink_group_with_flag(NCCL_SHRINK_ABORT, "NCCL_SHRINK_ABORT", 1)
+
+    @requires_nccl_shrink()
+    @requires_world_size(2)
+    def test_shrink_group_nccl_config(self):
+        """Verify that passing NCCL config via pg_options influences the shrunk group's backend options."""
+        device, pg = self._setup_shrink_test("config")
+        if self.rank == self.world_size - 1:
+            # excluded rank should not call shrink_group
+            dist.destroy_process_group()
+            return
+
+        # Prepare pg_options with NCCL config overrides
+        # Capture parent's current backend options to ensure we can prove override vs inherit
+        parent_backend = pg._get_backend(torch.device("cuda"))
+        parent_hp = parent_backend.options.is_high_priority_stream
+        parent_blocking = parent_backend.options.config.blocking
+
+        # Choose overrides that differ from the parent (flip where possible)
+        override_hp = not parent_hp
+        if parent_blocking in (0, 1):
+            override_blocking = 1 - parent_blocking
+        else:
+            # If undefined or unexpected, set to 1 which is a concrete value
+            override_blocking = 1
+
+        opts = c10d.ProcessGroupNCCL.Options()
+        opts.is_high_priority_stream = override_hp
+        opts.config.blocking = override_blocking
+
+        shrunk_pg = c10d.shrink_group([self.world_size - 1], pg_options=opts)
+
+        # Validate backend options propagated
+        backend = shrunk_pg._get_backend(torch.device("cuda"))
+        # is_high_priority_stream should exactly match our override and differ from parent
+        self.assertEqual(backend.options.is_high_priority_stream, override_hp)
+        self.assertNotEqual(backend.options.is_high_priority_stream, parent_hp)
+        # config is a struct; check representative field and difference from parent when meaningful
+        self.assertEqual(backend.options.config.blocking, override_blocking)
+        if parent_blocking in (0, 1):
+            self.assertNotEqual(backend.options.config.blocking, parent_blocking)
+
+        dist.destroy_process_group()
+
+    @requires_nccl_shrink()
+    @requires_world_size(2)
+    def test_shrink_group_performance(self):
+        """Test shrink_group performance and regression detection."""
+        import time
+
+        ranks_to_exclude = self._get_default_ranks_to_exclude()
+        is_excluded = self.rank in ranks_to_exclude
+
+        if not ranks_to_exclude:
+            log_test_info(self.rank, "Skipping performance test (world_size=1)")
+            return
+
+        log_test_info(self.rank, f"Performance test with {self.world_size} processes")
+        device, pg = self._setup_shrink_test("performance")
+
+        if not is_excluded:
+            log_test_info(self.rank, "Measuring shrink_group performance")
+            start_time = time.time()
+            shrunk_pg = c10d.shrink_group(ranks_to_exclude)
+            end_time = time.time()
+
+            elapsed_time = end_time - start_time
+            log_test_info(self.rank, f"shrink_group: {elapsed_time:.3f}s")
+
+            # Regression check: should complete within reasonable time
+            self.assertLess(
+                elapsed_time,
+                30.0,
+                f"shrink_group took {elapsed_time:.3f}s, possible regression",
+            )
+
+            # Test collective performance
+            expected_size = self.world_size - len(ranks_to_exclude)
+            self._validate_shrunk_group(shrunk_pg, expected_size, "performance")
+
+            collective_start = time.time()
+            _ = self._test_collective_on_shrunk_group(
+                shrunk_pg, device, ranks_to_exclude, "performance"
+            )
+            collective_time = time.time() - collective_start
+
+            log_test_info(self.rank, f"all_reduce: {collective_time:.3f}s")
+            log_test_success(self.rank, "Performance test passed")
+        else:
+            log_test_info(self.rank, "Excluded rank - waiting")
+
+        dist.destroy_process_group()
+
+    @requires_nccl_shrink()
+    @requires_world_size(4)
+    def test_shrink_group_multiple_exclusions(self):
+        """Test shrink_group with multiple ranks excluded at once."""
+        # Scale exclusions with world size
+        ranks_to_exclude = list(range(2, self.world_size, 2))  # Every other rank from 2
+
+        self._perform_shrink_test(ranks_to_exclude, "Multiple exclusions test")
+
+    @requires_nccl_shrink()
+    @requires_world_size(3)
+    def test_shrink_group_multiple_iterations(self):
+        """Test multiple shrink operations in sequence."""
+        log_test_info(
+            self.rank,
+            f"Starting test_shrink_group_multiple_iterations with world_size={self.world_size}",
+        )
+
+        store = c10d.FileStore(self.file_name, self.world_size)
+        device = torch.device(f"cuda:{self.rank}")
+        _ = self._create_process_group_nccl(store, self.opts(), device_id=device)
+
+        # Track current effective world size throughout shrinking operations
+        current_world_size = self.world_size
+        log_test_info(self.rank, f"Initial world_size: {current_world_size}")
+
+        # First shrinking: exclude the last rank(s)
+        first_exclusion = [self.world_size - 1]
+        if self.world_size >= 6:
+            first_exclusion.append(
+                self.world_size - 2
+            )  # Exclude last two ranks for larger sizes
+
+        log_test_info(self.rank, f"First shrinking: excluding ranks {first_exclusion}")
+
+        if self.rank not in first_exclusion:
+            # Only non-excluded ranks should call shrink_group
+            first_pg = c10d.shrink_group(first_exclusion)
+            self.assertIsNotNone(first_pg)
+            # IMPORTANT: Update world size after first shrinking
+            current_world_size = first_pg.size()
+            expected_first_size = self.world_size - len(first_exclusion)
+            log_test_info(
+                self.rank,
+                f"After first shrinking: world_size {self.world_size} -> {current_world_size}",
+            )
+            self.assertEqual(first_pg.size(), expected_first_size)
+
+            # Second shrinking: exclude another rank from the remaining group
+            # Choose a rank that's in the middle range
+            if current_world_size >= 3:
+                second_exclusion = [
+                    current_world_size - 1
+                ]  # Exclude the new "last" rank
+                log_test_info(
+                    self.rank,
+                    f"Second shrinking from group of size {current_world_size}: excluding ranks {second_exclusion}",
+                )
+
+                if self.rank not in second_exclusion:
+                    # Only non-excluded ranks should call shrink_group for second iteration
+                    second_pg = c10d.shrink_group(second_exclusion, group=first_pg)
+                    self.assertIsNotNone(second_pg)
+                    # IMPORTANT: Update world size after second shrinking
+                    final_world_size = second_pg.size()
+                    expected_final_size = current_world_size - len(second_exclusion)
+                    log_test_info(
+                        self.rank,
+                        f"After second shrinking: world_size {current_world_size} -> {final_world_size}",
+                    )
+                    self.assertEqual(second_pg.size(), expected_final_size)
+
+                    # Test collective on final group
+                    tensor = torch.full((1,), self.rank).cuda(device)
+                    log_test_info(
+                        self.rank,
+                        f"Performing all_reduce on final group (size {final_world_size}) with tensor: {tensor.item()}",
+                    )
+                    c10d.all_reduce(tensor, group=second_pg)
+                    log_test_info(
+                        self.rank,
+                        f"Final all_reduce completed, result: {tensor.item()}",
+                    )
+
+                    # Calculate expected sum of remaining ranks
+                    all_excluded = set(first_exclusion + second_exclusion)
+                    remaining_ranks = [
+                        r for r in range(self.world_size) if r not in all_excluded
+                    ]
+                    expected_sum = sum(remaining_ranks)
+                    log_test_info(
+                        self.rank,
+                        f"Remaining ranks: {remaining_ranks}, expected sum: {expected_sum}, actual: {tensor.item()}",
+                    )
+                    self.assertEqual(tensor.item(), expected_sum)
+                    log_test_info(self.rank, "Final verification passed")
+                else:
+                    log_test_info(
+                        self.rank,
+                        "This rank excluded in second shrinking, not calling shrink_group",
+                    )
+            else:
+                log_test_info(
+                    self.rank, "Skipping second shrinking (remaining group too small)"
+                )
+        else:
+            log_test_info(
+                self.rank,
+                "This rank excluded in first shrinking, not calling shrink_group",
+            )
+
+        log_test_info(self.rank, "Destroying process group")
+        dist.destroy_process_group()
+        log_test_info(self.rank, "test_shrink_group_multiple_iterations completed")
+
+    # Helper methods for optimized shrink group tests
+    def _setup_shrink_test(self, test_suffix, world_size=None, warmup=True):
+        """Common setup for shrink group tests."""
+        os.environ["TORCH_NCCL_USE_COMM_NONBLOCKING"] = "1"
+        world_size = world_size or self.world_size
+        store = c10d.FileStore(self.file_name + f"_{test_suffix}", world_size)
+        device = torch.device(f"cuda:{self.rank}")
+        c10d.init_process_group(
+            "nccl",
+            world_size=world_size,
+            rank=self.rank,
+            store=store,
+            pg_options=self.opts(),
+            device_id=device,
+        )
+        pg = c10d.distributed_c10d._get_default_group()
+
+        if warmup:
+            c10d.all_reduce(torch.ones(1).cuda(device), group=pg)
+
+        return device, pg
+
+    def _validate_shrunk_group(self, shrunk_pg, expected_size, test_name=""):
+        """Validate properties of a shrunk process group."""
+        self.assertIsNotNone(shrunk_pg, f"{test_name}: shrunk_pg should not be None")
+        actual_size = shrunk_pg.size()
+        self.assertEqual(
+            actual_size, expected_size, f"{test_name}: group size mismatch"
+        )
+
+        new_rank = shrunk_pg.rank()
+        self.assertTrue(
+            0 <= new_rank < expected_size, f"{test_name}: invalid new rank {new_rank}"
+        )
+
+        log_test_info(
+            self.rank,
+            f"{test_name}: world_size {self.world_size} -> {actual_size}, rank {self.rank} -> {new_rank}",
+        )
+        return new_rank
+
+    def _test_collective_on_shrunk_group(
+        self, shrunk_pg, device, ranks_to_exclude, test_name=""
+    ):
+        """Test collective communication on shrunk group and verify correctness."""
+        test_tensor = torch.full((1,), self.rank, device=device, dtype=torch.float32)
+        c10d.all_reduce(test_tensor, group=shrunk_pg)
+
+        result = test_tensor.item()
+        expected_sum = sum(
+            r for r in range(self.world_size) if r not in ranks_to_exclude
+        )
+
+        self.assertEqual(
+            result, expected_sum, f"{test_name}: collective result mismatch"
+        )
+        log_test_info(
+            self.rank, f"{test_name}: collective passed ({result} == {expected_sum})"
+        )
+        return result
+
+    def _perform_shrink_test(
+        self, ranks_to_exclude, test_name, shrink_flags=0, with_collective=True
+    ):
+        """Complete shrink test flow: setup, shrink, validate, test collective, cleanup.
+
+        Consistent API: All ranks perform setup to initialize distributed environment.
+        ONLY non-excluded ranks call shrink_group() for both default and non-default groups.
+        Excluded ranks perform setup, then exit without calling shrink_group() or waiting.
+        """
+        log_test_info(self.rank, f"{test_name} (world_size={self.world_size})")
+
+        is_excluded = self.rank in ranks_to_exclude
+        log_test_info(
+            self.rank,
+            f"Excluding ranks: {ranks_to_exclude}, am_excluded: {is_excluded}",
+        )
+
+        # All ranks (including excluded ones) perform setup to initialize distributed environment
+        device, pg = self._setup_shrink_test(test_name.lower().replace(" ", "_"))
+        is_default_group = pg == c10d.distributed_c10d._get_default_group()
+
+        if is_excluded:
+            log_test_info(
+                self.rank,
+                f"Excluded rank {self.rank} - setup complete, skipping shrink operation",
+            )
+            if shrink_flags & NCCL_SHRINK_ABORT:
+                log_test_info(self.rank, f"Using abort for excluded rank {self.rank}")
+                pg._get_backend(torch.device(device)).abort()
+                log_test_info(
+                    self.rank, f"cleanup resources for excluded rank {self.rank}"
+                )
+                dist.destroy_process_group()
+                log_test_info(self.rank, f"Excluded rank {self.rank} - exit")
+            else:
+                log_test_info(
+                    self.rank, f"Using regular destroy for excluded rank {self.rank}"
+                )
+                dist.destroy_process_group()
+            return None
+
+        # Only non-excluded ranks proceed with shrink
+        log_test_info(
+            self.rank,
+            f"Non-excluded rank calling shrink_group (default_group={is_default_group})",
+        )
+        shrunk_pg = c10d.shrink_group(ranks_to_exclude, shrink_flags=shrink_flags)
+        log_test_info(
+            self.rank,
+            f"Non-excluded rank calling shrink_group (default_group={is_default_group}) done",
+        )
+
+        # Non-excluded ranks: validate and test the new group
+        expected_size = self.world_size - len(ranks_to_exclude)
+        _ = self._validate_shrunk_group(shrunk_pg, expected_size, test_name)
+
+        if with_collective:
+            _ = self._test_collective_on_shrunk_group(
+                shrunk_pg, device, ranks_to_exclude, test_name
+            )
+            log_test_success(self.rank, f"{test_name} successful (shrink + collective)")
+        else:
+            log_test_success(self.rank, f"{test_name} successful (shrink only)")
+
+        dist.destroy_process_group()
+        return shrunk_pg
+
+    def _get_default_ranks_to_exclude(self):
+        """Get default ranks to exclude based on world size."""
+        if self.world_size <= 1:
+            return []
+        return [self.world_size - 1]  # Exclude last rank by default
+
+    @requires_nccl_shrink()
+    @requires_world_size(3)
+    def test_shrink_group_vs_abort_reinit_performance(self):
+        """Compare performance of shrink_group vs traditional abort+reinit (simplified for reliability)."""
+        log_test_info(self.rank, "=== TEST 1: abort+reinit ===")
+
+        device, pg1 = self._setup_shrink_test("_perf_reinit")
+        torch.cuda.synchronize(device)
+
+        # Test 1: Traditional abort + reinit
+        start_time = time.perf_counter()
+        dist.destroy_process_group()
+
+        device, new_pg = self._setup_shrink_test("perf_shrink_test1")
+        reinit_time = time.perf_counter() - start_time
+
+        # Test collective with original rank values for fair comparison (non-blocking mode)
+        test_tensor = torch.full((1,), self.rank, device=device, dtype=torch.float32)
+        work = c10d.all_reduce(test_tensor, group=new_pg, async_op=True)
+        work.wait()
+
+        torch.cuda.synchronize(device)
+
+        # Verify correctness
+        expected_sum = sum(r for r in range(self.world_size))
+        self.assertEqual(test_tensor.item(), expected_sum, "Reinit collective failed")
+
+        log_test_info(self.rank, f"abort+reinit: {reinit_time:.4f}s")
+        dist.destroy_process_group(new_pg)
+
+        # Test 2: shrink_group with NCCL_SHRINK_ABORT
+        log_test_info(self.rank, "=== TEST 2: shrink_group ===")
+
+        ranks_to_exclude = [self.world_size - 1]
+        is_excluded = self.rank in ranks_to_exclude
+        log_test_info(
+            self.rank,
+            f"Excluding ranks: {ranks_to_exclude}, am_excluded: {is_excluded}",
+        )
+
+        device, pg1 = self._setup_shrink_test("perf_shrink_test2")  # Unique suffix
+
+        shrink_time = 0
+        if not is_excluded:
+            torch.cuda.synchronize(device)  # Ensure accurate timing
+            start_time = time.perf_counter()
+            shrunk_pg = c10d.shrink_group(
+                ranks_to_exclude, shrink_flags=NCCL_SHRINK_ABORT
+            )
+            c10d.all_reduce(torch.ones(1).cuda(device), group=shrunk_pg)
+            shrink_time = time.perf_counter() - start_time
+
+            # Test collective communication on shrunk group (non-blocking mode)
+            test_tensor = torch.full(
+                (1,), self.rank, device=device, dtype=torch.float32
+            )
+            work = c10d.all_reduce(test_tensor, group=shrunk_pg, async_op=True)
+            work.wait()
+
+            # Verify correctness
+            expected_sum = sum(
+                r for r in range(self.world_size) if r not in ranks_to_exclude
+            )
+            self.assertEqual(
+                test_tensor.item(),
+                expected_sum,
+                "shrink_test: collective result mismatch",
+            )
+
+            torch.cuda.synchronize(device)  # Ensure operations complete
+            log_test_info(self.rank, f"shrink_group: {shrink_time:.4f}s")
+            dist.destroy_process_group()
+        else:
+            log_test_info(self.rank, "Excluded from shrink test - exiting immediately")
+            dist.destroy_process_group()
+            return
+
+        # Performance analysis (only for participating ranks)
+        if shrink_time > 0 and reinit_time > 0:
+            speedup = reinit_time / shrink_time
+            time_saved = reinit_time - shrink_time
+
+            log_test_info(self.rank, "=== PERFORMANCE RESULTS ===")
+            log_test_info(self.rank, f"shrink_group:  {shrink_time:.4f}s")
+            log_test_info(self.rank, f"abort+reinit:  {reinit_time:.4f}s")
+            log_test_info(self.rank, f"time_saved:    {time_saved:+.4f}s")
+            log_test_info(self.rank, f"speedup:       {speedup:.2f}x")
+
+            if speedup > 1.1:
+                log_test_success(self.rank, "shrink_group significantly faster")
+            elif speedup > 0.9:
+                log_test_info(self.rank, "≈ comparable performance")
+            else:
+                log_test_warning(self.rank, "abort+reinit faster")
+
+        log_test_info(self.rank, "Performance test completed")
+
     @requires_nccl()
     @skip_but_pass_in_sandcastle_if(not TEST_MULTIGPU, "NCCL test requires 2+ GPUs")
     def test_deterministic_mode_no_break(self):
@@ -4901,7 +5539,7 @@ class NCCLTraceTest(NCCLTraceTestBase):
             for p2p_op_idx, input_sizes in zip(
                 range(first_op, coalesced_op, 1), op_sizes_per_coalesce
             ):
-                # the indivudal ops inside the coalescing group the individual op metadata,
+                # the individual ops inside the coalescing group the individual op metadata,
                 # but not the timing info coming from the actual coalesced kernel
                 profiling_name = (
                     "nccl:recv 0<-1" if self.rank == 0 else "nccl:send 1->0"

test/distributed/test_nvshmem.py

@@ -398,7 +398,7 @@ class NVSHMEMAll2AllTest(MultiProcContinuousTest):
             nsplits, dtype=torch.int64, device=self.device
         ).copy_(inp_splits)
         # 2 rows: output splits, output offsets
-        # Initiallizing all values to -1 to check if they are updated
+        # Initializing all values to -1 to check if they are updated
         out_splits_offsets = symm_mem.empty(
             (2, nsplits), dtype=torch.int64, device=self.device
         ).fill_(-1)
@@ -503,7 +503,7 @@ class NVSHMEMAll2AllTest(MultiProcContinuousTest):
             (2, nsplits), dtype=torch.int64, device=self.device
         )
         # 2 rows: output splits, output offsets
-        # Initiallizing all values to -1 to check if they are updated
+        # Initializing all values to -1 to check if they are updated
         out_splits_offsets = symm_mem.empty(
             (2, nsplits), dtype=torch.int64, device=self.device
         ).fill_(-1)
@@ -617,7 +617,7 @@ def dispatch_then_combine(device, align: int, group) -> None:
         inp_splits
     )
     # 2 rows: output splits, output offsets
-    # Initiallizing all values to -1 to check if they are updated
+    # Initializing all values to -1 to check if they are updated
     out_splits_offsets = symm_mem.empty(
         (2, nsplits), dtype=torch.int64, device=device
     ).fill_(-1)
@@ -625,7 +625,7 @@ def dispatch_then_combine(device, align: int, group) -> None:
     # Buffers for combine
     combine_out = symm_mem.empty(max_out_numel, dtype=dtype, device=device).fill_(-1)
     # 2 rows: output splits, output offsets
-    # Initiallizing all values to -1 to check if they are updated
+    # Initializing all values to -1 to check if they are updated
     combine_out_splits_offsets = symm_mem.empty(
         (2, nsplits), dtype=torch.int64, device=device
     ).fill_(-1)

test/distributed/test_overlap_bucketing_unit.py

@@ -0,0 +1,413 @@
+# Owner(s): ["module: inductor"]
+import unittest
+
+import torch
+import torch._dynamo
+import torch._dynamo.logging
+import torch._dynamo.test_case
+import torch.distributed as dist
+import torch.fx as fx
+
+# for some reason importing functional collectives after dynamo breaks collectives handling!
+from torch._C import FileCheck
+from torch._inductor.test_case import TestCase as InductorTestCase
+from torch._subclasses.fake_tensor import FakeTensorMode
+from torch.fx.experimental.proxy_tensor import make_fx
+from torch.testing._internal.common_distributed import requires_accelerator_dist_backend
+from torch.testing._internal.common_utils import (
+    instantiate_parametrized_tests,
+    parametrize,
+    run_tests,
+)
+from torch.testing._internal.inductor_utils import HAS_GPU
+from torch.utils._ordered_set import OrderedSet
+
+
+# flake8: noqa: B950
+# Owner(s): ["module: inductor"]
+
+
+aten = torch.ops.aten
+
+from torch.testing._internal.common_fsdp import get_devtype
+
+
+device_type = str(get_devtype())
+
+
+import torch
+import torch._dynamo
+import torch._dynamo.logging
+import torch._dynamo.test_case
+
+
+# for some reason importing functional collectives after dynamo breaks collectives handling!
+
+
+@requires_accelerator_dist_backend(["nccl", "xccl"])
+def build_collective_info(graph, hiding_annotations):
+    """
+    Build CollectiveInfo dict from manual hiding annotations.
+
+    hiding_annotations: dict mapping collective_start -> hiding_compute_node
+    """
+    from torch._inductor.fx_passes.overlap_scheduling import CollectiveInfo
+
+    collective_info = {}
+
+    # Find all collective starts and their corresponding waits
+    start_to_wait = {}
+    for node in graph.nodes:
+        if node.op == "call_function" and "wait_tensor" in str(node.target):
+            wait_input = node.args[0]
+            if isinstance(wait_input, fx.Node):
+                start_to_wait[wait_input] = node
+
+    # Build CollectiveInfo for each collective
+    for start_node, wait_node in start_to_wait.items():
+        hiding_node = hiding_annotations.get(start_node)
+
+        # Estimate size and time
+        size_bytes = 16 * 4  # 4x4 tensor of floats
+        estimated_time_ms = 1.0  # Dummy time
+        exposed_time_ms = 0.0 if hiding_node else 1.0  # Hidden if has hiding_node
+
+        collective_info[start_node] = CollectiveInfo(
+            start_node=start_node,
+            wait_node=wait_node,
+            size_bytes=size_bytes,
+            estimated_time_ms=estimated_time_ms,
+            exposed_time_ms=exposed_time_ms,
+            hiding_node=hiding_node,
+        )
+
+    return collective_info
+
+
+def compute_ancestors(graph):
+    """Compute ancestor sets for all nodes in the graph."""
+    node_ancestors = {}
+
+    for node in graph.nodes:
+        ancestors = OrderedSet()
+        stack = list(node.all_input_nodes)
+        visited = set()
+
+        while stack:
+            current = stack.pop()
+            if current in visited:
+                continue
+            visited.add(current)
+            ancestors.add(current)
+            stack.extend(current.all_input_nodes)
+
+        node_ancestors[node] = ancestors
+
+    return node_ancestors
+
+
+@requires_accelerator_dist_backend()
+@unittest.skipIf(not HAS_GPU, "Inductor+gpu needs triton and recent GPU arch")
+@instantiate_parametrized_tests
+class TestOverlapPreservingBucketing(InductorTestCase):
+    """
+    Unit tests for overlap-preserving bucketing pass.
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        super().setUpClass()
+        from torch.testing._internal.distributed.fake_pg import FakeStore
+
+        store = FakeStore()
+        dist.init_process_group(backend="fake", rank=0, world_size=2, store=store)
+        cls.device = "cuda"
+
+    @classmethod
+    def tearDownClass(cls):
+        super().tearDownClass()
+        dist.destroy_process_group()
+
+    def test_can_bucket_independent_collectives(self):
+        """
+        Test that independent collectives with separate hiding nodes CAN bucket.
+
+        Graph structure:
+        ag1_start -> ag2_start -> mm1 (hides ag1) -> mm2 (hides ag2) -> ag1_wait -> ag2_wait
+        """
+
+        def func(a, b):
+            group_name = "0"
+            group_size = 1
+
+            # Start both collectives
+            ag1 = torch.ops._c10d_functional.all_gather_into_tensor(
+                a, group_size, group_name
+            )
+            ag2 = torch.ops._c10d_functional.all_gather_into_tensor(
+                b, group_size, group_name
+            )
+
+            # Independent compute that can hide both
+            mm1 = torch.mm(a, a)
+            mm2 = torch.mm(b, b)
+
+            # Wait for both
+            ag1_out = torch.ops._c10d_functional.wait_tensor(ag1)
+            ag2_out = torch.ops._c10d_functional.wait_tensor(ag2)
+
+            return ag1_out.sum() + ag2_out.sum() + mm1.sum() + mm2.sum()
+
+        # Use fake mode to trace without executing
+        with FakeTensorMode():
+            a = torch.ones(4, 4, device=self.device)
+            b = torch.ones(4, 4, device=self.device) * 2
+
+            # Trace with make_fx
+            traced = make_fx(func)(a, b)
+
+        # Find nodes using find_nodes
+        ag1, ag2 = traced.graph.find_nodes(
+            op="call_function",
+            target=torch.ops._c10d_functional.all_gather_into_tensor.default,
+        )
+        mm1, mm2 = traced.graph.find_nodes(
+            op="call_function", target=torch.ops.aten.mm.default
+        )
+
+        # Manually annotate hiding relationships
+        hiding_annotations = {
+            ag1: mm1,  # mm1 hides ag1
+            ag2: mm2,  # mm2 hides ag2
+        }
+
+        # Build collective info and ancestors
+        collective_info = build_collective_info(traced.graph, hiding_annotations)
+        node_ancestors = compute_ancestors(traced.graph)
+        scheduled = OrderedSet(traced.graph.nodes)
+
+        # Run bucketing
+        from torch._inductor.fx_passes.overlap_preserving_bucketer import (
+            OverlapPreservingBucketer,
+        )
+
+        bucketer = OverlapPreservingBucketer(
+            traced.graph,
+            collective_info,
+            node_ancestors,
+            scheduled,
+        )
+        bucketer.bucket_collectives()
+
+        # Verify: should have 1 bucketed collective (all_gather_into_tensor_out)
+        graph_str = str(traced.graph)
+        FileCheck().check_count("all_gather_into_tensor_out", 1, exactly=False).run(
+            graph_str
+        )
+
+    def test_cant_bucket_nested_hiding_intervals(self):
+        """
+        Test that nested hiding intervals prevent bucketing.
+
+        Graph structure:
+        ag1_start -> ag2_start -> mm2 (hides ag2) -> ag2_wait -> mm1 (hides ag1) -> ag1_wait
+
+        ag2's hiding interval is nested inside ag1's hiding interval.
+        """
+
+        def func(a, b):
+            group_name = "0"
+            group_size = 1
+
+            # ag1 starts first
+            ag1 = torch.ops._c10d_functional.all_gather_into_tensor(
+                a, group_size, group_name
+            )
+
+            # ag2 starts (inside ag1's interval)
+            ag2 = torch.ops._c10d_functional.all_gather_into_tensor(
+                b, group_size, group_name
+            )
+
+            # mm2 hides ag2
+            mm2 = torch.mm(b[:2, :2], b[:2, :2])
+
+            # ag2 waits (still inside ag1's interval)
+            ag2_out = torch.ops._c10d_functional.wait_tensor(ag2)
+
+            # mm1 uses ag2's result and hides ag1
+            mm1 = torch.mm(a + ag2_out[:4, :4], a)
+
+            # ag1 waits last
+            ag1_out = torch.ops._c10d_functional.wait_tensor(ag1)
+
+            return ag1_out.sum() + ag2_out.sum() + mm1.sum() + mm2.sum()
+
+        # Use fake mode to trace without executing
+        with FakeTensorMode():
+            a = torch.ones(4, 4, device=self.device)
+            b = torch.ones(4, 4, device=self.device) * 2
+
+            # Trace with make_fx
+            traced = make_fx(func)(a, b)
+
+        # Find nodes using find_nodes
+        ag1, ag2 = traced.graph.find_nodes(
+            op="call_function",
+            target=torch.ops._c10d_functional.all_gather_into_tensor.default,
+        )
+        mm_nodes = traced.graph.find_nodes(
+            op="call_function", target=torch.ops.aten.mm.default
+        )
+        # mm2 is the first mm, mm1 is the second (based on graph order)
+        mm2 = mm_nodes[0]
+        mm1 = mm_nodes[1]
+
+        # Manually annotate hiding relationships
+        hiding_annotations = {
+            ag1: mm1,  # mm1 hides ag1
+            ag2: mm2,  # mm2 hides ag2
+        }
+
+        # Build collective info and ancestors
+        collective_info = build_collective_info(traced.graph, hiding_annotations)
+        node_ancestors = compute_ancestors(traced.graph)
+        scheduled = OrderedSet(traced.graph.nodes)
+
+        # Run bucketing
+        from torch._inductor.fx_passes.overlap_preserving_bucketer import (
+            OverlapPreservingBucketer,
+        )
+
+        bucketer = OverlapPreservingBucketer(
+            traced.graph,
+            collective_info,
+            node_ancestors,
+            scheduled,
+        )
+        bucketer.bucket_collectives()
+
+        # Verify: nested hiding intervals should prevent bucketing
+        # Should have 2 separate all_gathers, not 1 bucketed one
+        graph_str = str(traced.graph)
+        FileCheck().check_count("all_gather_into_tensor", 2, exactly=False).run(
+            graph_str
+        )
+
+    @parametrize("final_mm_hidden", (True, False))
+    def test_cant_bucket_ag_with_rs_hiding_interval_between(self, final_mm_hidden):
+        """
+        Test that all_gathers can't bucket when a reduce_scatter's hiding interval is between them.
+
+        Graph structure:
+        ag1_start -> mm1 (hides ag1) -> ag1_wait ->
+        rs_start -> mm2 (hides rs) -> rs_wait ->
+
+        if final_mm_hidden:
+            ag2_start -> mm3 (hides ag2) -> ag2_wait
+
+        if final_mm_hidden:
+            Bucketing ag1 and ag2 would require moving one of them, which would break hiding relationships:
+            - Moving ag2 earlier would break ag2's hiding by mm3
+            - Moving ag1 later would break ag1's hiding by mm1
+            - The rs hiding interval creates an obstacle between them
+
+        otherwise, we can bucket
+        """
+
+        def func(a, b, c):
+            group_name = dist.distributed_c10d._get_default_group().group_name
+            group_size = 1
+
+            # First all_gather
+            ag1 = torch.ops._c10d_functional.all_gather_into_tensor(
+                a, group_size, group_name
+            )
+            mm1 = torch.mm(a, a)  # hides ag1
+            ag1_out = torch.ops._c10d_functional.wait_tensor(ag1)
+
+            # Reduce scatter in between
+            rs = torch.ops._c10d_functional.reduce_scatter_tensor(
+                b, "sum", group_size, group_name
+            )
+            mm2 = torch.mm(b[:4, :4], b[:4, :4])  # hides rs
+            rs_out = torch.ops._c10d_functional.wait_tensor(rs)
+
+            # Second all_gather
+            ag2 = torch.ops._c10d_functional.all_gather_into_tensor(
+                c, group_size, group_name
+            )
+            mm3 = torch.mm(c, c)  # hides ag2
+            ag2_out = torch.ops._c10d_functional.wait_tensor(ag2)
+
+            return ag1_out.sum() + rs_out.sum() + ag2_out.sum(), mm1, mm2, mm3
+
+        # Use fake mode to trace without executing
+        with FakeTensorMode():
+            a = torch.ones(4, 4, device=self.device)
+            b = torch.ones(8, 4, device=self.device)
+            c = torch.ones(4, 4, device=self.device)
+
+            # Trace with make_fx
+            traced = make_fx(func)(a, b, c)
+
+        ag1, ag2 = traced.graph.find_nodes(
+            op="call_function",
+            target=torch.ops._c10d_functional.all_gather_into_tensor.default,
+        )
+        (rs,) = traced.graph.find_nodes(
+            op="call_function",
+            target=torch.ops._c10d_functional.reduce_scatter_tensor.default,
+        )
+        mm1, mm2, mm3 = traced.graph.find_nodes(
+            op="call_function", target=torch.ops.aten.mm.default
+        )
+
+        # Manually annotate hiding relationships
+        hiding_annotations = {
+            ag1: mm1,  # mm1 hides ag1
+            # rs: mm2,   # mm2 hides rs
+            ag2: mm3,
+        }
+        if final_mm_hidden:
+            hiding_annotations[rs] = mm2
+
+        # Build collective info and ancestors
+        collective_info = build_collective_info(traced.graph, hiding_annotations)
+        node_ancestors = compute_ancestors(traced.graph)
+        scheduled = OrderedSet(traced.graph.nodes)
+
+        # Run bucketing logic to find buckets (without applying them, which would require process groups)
+        from torch._inductor.fx_passes.overlap_preserving_bucketer import (
+            OverlapPreservingBucketer,
+        )
+
+        bucketer = OverlapPreservingBucketer(
+            traced.graph,
+            collective_info,
+            node_ancestors,
+            scheduled,
+        )
+
+        bucketer.bucket_collectives()
+
+        graph_str = str(traced.graph)
+
+        # check order of mms preserved
+        FileCheck().check("%mm").check("%mm_1").check("%mm_2").run(graph_str)
+
+        if final_mm_hidden:
+            # Should NOT bucket - 2 separate all_gathers
+            # Count all_gather node names (works even when wrapped in control_deps)
+            FileCheck().check_count("%all_gather_into_tensor", 2, exactly=False).run(
+                graph_str
+            )
+        else:
+            # Should bucket - 1 bucketed all_gather (all_gather_into_tensor_out)
+            FileCheck().check_count(
+                "%all_gather_into_tensor_out", 1, exactly=False
+            ).run(graph_str)
+
+
+if __name__ == "__main__":
+    run_tests()

test/distributed/test_symmetric_memory.py

@@ -274,7 +274,7 @@ class SymmetricMemoryTest(MultiProcContinuousTest):
             self.assertTrue(buf.eq(peer_rank + world.size() // 2).all())
 
 
-# We move AsyncTP tests to a seperate test suite because 1) Async TP ops are not
+# We move AsyncTP tests to a separate test suite because 1) Async TP ops are not
 # the core symmetric memory APIs, they are more like applications, 2)
 # MultiProcContinuousTest will skip all the following tests if a test fails (
 # we should fix this too). We still want to get the test signals for the core
@@ -621,7 +621,7 @@ class AsyncTPTest(MultiProcContinuousTest):
 
 # [READ ME FIRST]
 # The `SymmMemEmptySetDeviceTest` suite parameterizes whether user sets the
-# device before calling symm_mem.emtpy.  Either way should work.
+# device before calling symm_mem.empty.  Either way should work.
 # However, since `set_device` is persistent, we cannot use the
 # `MultiProcContinuousTest` template because the next function will be
 # "contaminated", leading to flaky tests (e.g. hang). Therefore, we use

test/dynamo/cpython/3_13/mathdata/ieee754.txt

@@ -51,7 +51,7 @@ nan
 >>> INF / INF
 nan
 
-However unambigous operations with inf return inf:
+However unambiguous operations with inf return inf:
 >>> INF * INF
 inf
 >>> 1.5 * INF

test/dynamo/cpython/3_13/test_itertools.py

@@ -1711,7 +1711,7 @@ class TestBasicOps(__TestCase):
         t3 = tnew(t1)
         self.assertTrue(list(t1) == list(t2) == list(t3) == list('abc'))
 
-        # test that tee objects are weak referencable
+        # test that tee objects are weak referenceable
         a, b = tee(range(10))
         p = weakref.proxy(a)
         self.assertEqual(getattr(p, '__class__'), type(b))
@@ -2243,7 +2243,7 @@ class TestPurePythonRoughEquivalents(__TestCase):
         t3 = tnew(t1)
         self.assertTrue(list(t1) == list(t2) == list(t3) == list('abc'))
 
-        # test that tee objects are weak referencable
+        # test that tee objects are weak referenceable
         a, b = tee(range(10))
         p = weakref.proxy(a)
         self.assertEqual(getattr(p, '__class__'), type(b))