[dynamo] Revert C++-fying of symbolic shape guards

Moving symbolic shape guards to C++ causes compile time issues. This basically boils down to a tradeoff question. For models that have large amount of dynamic shape guards, this flag will help reduce guard latency. But for most of the models, that have a very few dynamic shape guards, the guard lantecy is anyways small. These models will still see a high compile time hit because of calling gcc during the compile. So a good default value seems to be False. We can write a doc to give guidance on reducing guard latency.
2025-11-06 00:54:56 +08:00 · 2025-10-28 11:13:24 -07:00
1281 changed files with 13355 additions and 37591 deletions
--- a/.bc-linter.yml
+++ b/.bc-linter.yml
@ -13,4 +13,3 @@ exclude:
  - "**/benchmarks/**"
  - "**/test_*.py"
  - "**/*_test.py"
-  - "tools/**"
--- a/.ci/docker/build.sh
+++ b/.ci/docker/build.sh
@ -195,16 +195,13 @@ case "$tag" in
    NINJA_VERSION=1.9.0
    TRITON=yes
    ;;
-  pytorch-linux-jammy-xpu-n-py3 | pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks)
+  pytorch-linux-jammy-xpu-n-py3)
    ANACONDA_PYTHON_VERSION=3.10
    GCC_VERSION=11
    VISION=yes
    XPU_VERSION=2025.2
    NINJA_VERSION=1.9.0
    TRITON=yes
-    if [[ $tag =~ "benchmarks" ]]; then
-      INDUCTOR_BENCHMARKS=yes
-    fi
    ;;
  pytorch-linux-jammy-py3-gcc11-inductor-benchmarks)
    ANACONDA_PYTHON_VERSION=3.10
--- a/.ci/docker/ci_commit_pins/triton.txt
+++ b/.ci/docker/ci_commit_pins/triton.txt
@ -1 +1 @@
-bfeb066872bc1e8b2d2bc0a3b295b99dd77206e7
+7416ffcb92cdbe98d9f97e4e6f95247e46dfc9fd
--- a/.ci/docker/common/install_acl.sh
+++ b/.ci/docker/common/install_acl.sh
@ -3,7 +3,7 @@

 set -eux

-ACL_VERSION=${ACL_VERSION:-"v52.6.0"}
+ACL_VERSION=${ACL_VERSION:-"v25.02"}
 ACL_INSTALL_DIR="/acl"

 # Clone ACL
--- a/.ci/docker/common/install_conda.sh
+++ b/.ci/docker/common/install_conda.sh
@ -49,20 +49,12 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
    export SYSROOT_DEP="sysroot_linux-64=2.17"
  fi

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

-NVSHMEM_VERSION=3.4.5
+NVSHMEM_VERSION=3.3.24

 function install_cuda {
  version=$1
@ -129,7 +129,7 @@ function install_129 {
 }

 function install_128 {
-  CUDNN_VERSION=9.10.2.21
+  CUDNN_VERSION=9.8.0.87
  echo "Installing CUDA 12.8.1 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
  # install CUDA 12.8.1 in the same container
  install_cuda 12.8.1 cuda_12.8.1_570.124.06_linux
--- a/.ci/docker/common/install_rocm.sh
+++ b/.ci/docker/common/install_rocm.sh
@ -40,7 +40,11 @@ EOF

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

    # Add rocm repository
    wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add -
--- a/.ci/docker/common/install_rocm_magma.sh
+++ b/.ci/docker/common/install_rocm_magma.sh
@ -12,8 +12,8 @@ function do_install() {

    rocm_version_nodot=${rocm_version//./}

-    # post merge of https://github.com/icl-utk-edu/magma/pull/65
-    MAGMA_VERSION=c0792ae825fb36872784892ea643dd6f3456bc5f
+    # https://github.com/icl-utk-edu/magma/pull/65
+    MAGMA_VERSION=d6e4117bc88e73f06d26c6c2e14f064e8fc3d1ec
    magma_archive="magma-rocm${rocm_version_nodot}-${MAGMA_VERSION}-1.tar.bz2"

    rocm_dir="/opt/rocm"
--- a/.ci/docker/manywheel/Dockerfile_2_28
+++ b/.ci/docker/manywheel/Dockerfile_2_28
@ -149,7 +149,7 @@ FROM cpu_final as rocm_final
 ARG ROCM_VERSION=6.0
 ARG PYTORCH_ROCM_ARCH
 ENV PYTORCH_ROCM_ARCH ${PYTORCH_ROCM_ARCH}
-ARG DEVTOOLSET_VERSION=13
+ARG DEVTOOLSET_VERSION=11
 ENV LDFLAGS="-Wl,-rpath=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/lib64 -Wl,-rpath=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/lib"
 # Somewhere in ROCm stack, we still use non-existing /opt/rocm/hip path,
 # below workaround helps avoid error
--- a/.ci/docker/manywheel/build.sh
+++ b/.ci/docker/manywheel/build.sh
@ -97,7 +97,7 @@ case ${image} in
    manylinux2_28-builder:xpu)
        TARGET=xpu_final
        GPU_IMAGE=amd64/almalinux:8
-        DOCKER_GPU_BUILD_ARG=" --build-arg DEVTOOLSET_VERSION=13"
+        DOCKER_GPU_BUILD_ARG=" --build-arg DEVTOOLSET_VERSION=11"
        MANY_LINUX_VERSION="2_28"
        ;;
    *)
--- a/.ci/docker/requirements-ci.txt
+++ b/.ci/docker/requirements-ci.txt
@ -138,12 +138,10 @@ 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" and python_version < "3.14"
-numpy==2.3.4; python_version >= "3.14"
+numpy==2.1.2; python_version >= "3.13"

 pandas==2.0.3; 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"
+pandas==2.2.3; python_version >= "3.13"

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

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

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

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

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

-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"
+onnx==1.19.1
 #Description: Required by onnx tests, and mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
@ -366,7 +359,7 @@ pwlf==2.2.1
 #test that import: test_sac_estimator.py

 # To build PyTorch itself
-pyyaml==6.0.3
+pyyaml==6.0.2
 pyzstd
 setuptools==78.1.1
 packaging==23.1
--- a/.ci/docker/triton_version.txt
+++ b/.ci/docker/triton_version.txt
@ -1 +1 @@
-3.5.1
+3.5.0
--- a/.ci/docker/ubuntu-xpu/Dockerfile
+++ b/.ci/docker/ubuntu-xpu/Dockerfile
@ -54,15 +54,12 @@ ENV OPENSSL_DIR /opt/openssl
 RUN rm install_openssl.sh

 ARG INDUCTOR_BENCHMARKS
-ARG ANACONDA_PYTHON_VERSION
-ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
 COPY ci_commit_pins/timm.txt timm.txt
-COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${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
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt

 # Install XPU Dependencies
 ARG XPU_VERSION
--- a/.ci/docker/ubuntu/Dockerfile
+++ b/.ci/docker/ubuntu/Dockerfile
@ -100,8 +100,6 @@ 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

--- a/.ci/lumen_cli/pyproject.toml
+++ b/.ci/lumen_cli/pyproject.toml
@ -6,7 +6,7 @@ dependencies = [
    "GitPython==3.1.45",
    "docker==7.1.0",
    "pytest==7.3.2",
-    "uv==0.9.6"
+    "uv==0.9.5"
 ]

 [tool.setuptools]
--- a/.ci/magma-rocm/Makefile
+++ b/.ci/magma-rocm/Makefile
@ -1,7 +1,7 @@
 SHELL=/usr/bin/env bash

 DOCKER_CMD ?= docker
-DESIRED_ROCM ?= 7.1
+DESIRED_ROCM ?= 7.0
 DESIRED_ROCM_SHORT = $(subst .,,$(DESIRED_ROCM))
 PACKAGE_NAME = magma-rocm
 # inherit this from underlying docker image, do not pass this env var to docker
@ -16,7 +16,6 @@ DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
 	magma-rocm/build_magma.sh

 .PHONY: all
-all: magma-rocm71
 all: magma-rocm70
 all: magma-rocm64

@ -25,11 +24,6 @@ clean:
 	$(RM) -r magma-*
 	$(RM) -r output

-.PHONY: magma-rocm71
-magma-rocm71: DESIRED_ROCM := 7.1
-magma-rocm71:
-	$(DOCKER_RUN)
-
 .PHONY: magma-rocm70
 magma-rocm70: DESIRED_ROCM := 7.0
 magma-rocm70:
--- a/.ci/pytorch/build.sh
+++ b/.ci/pytorch/build.sh
@ -426,7 +426,7 @@ fi
 if [[ "$BUILD_ENVIRONMENT" != *libtorch* && "$BUILD_ENVIRONMENT" != *bazel* ]]; then
  # export test times so that potential sharded tests that'll branch off this build will use consistent data
  # don't do this for libtorch as libtorch is C++ only and thus won't have python tests run on its build
-  PYTHONPATH=. python tools/stats/export_test_times.py
+  python tools/stats/export_test_times.py
 fi
 # don't do this for bazel or s390x or riscv64 as they don't use sccache
 if [[ "$BUILD_ENVIRONMENT" != *s390x* && "$BUILD_ENVIRONMENT" != *riscv64* && "$BUILD_ENVIRONMENT" != *-bazel-* ]]; then
--- a/.ci/pytorch/smoke_test/smoke_test.py
+++ b/.ci/pytorch/smoke_test/smoke_test.py
@ -272,18 +272,6 @@ def smoke_test_cuda(
        torch_cudnn_version = cudnn_to_version_str(torch.backends.cudnn.version())
        print(f"Torch cuDNN version: {torch_cudnn_version}")

-        torch_cudnn_compile_version = torch._C._cudnn.getCompileVersion()
-        print(f"Torch cuDNN compile-time version: {torch_cudnn_compile_version}")
-        torch_cudnn_runtime_version = tuple(
-            [int(x) for x in torch_cudnn_version.split(".")]
-        )
-        if torch_cudnn_runtime_version != torch_cudnn_compile_version:
-            raise RuntimeError(
-                "cuDNN runtime version doesn't match comple version. "
-                f"Loaded: {torch_cudnn_runtime_version} "
-                f"Expected: {torch_cudnn_compile_version}"
-            )
-
        if sys.platform in ["linux", "linux2"]:
            torch_nccl_version = ".".join(str(v) for v in torch.cuda.nccl.version())
            print(f"Torch nccl; version: {torch_nccl_version}")
--- a/.ci/pytorch/test.sh
+++ b/.ci/pytorch/test.sh
@ -460,18 +460,28 @@ test_inductor_shard() {
    --verbose
 }

-test_inductor_aoti_cpp() {
+test_inductor_aoti() {
+  # docker build uses bdist_wheel which does not work with test_aot_inductor
+  # TODO: need a faster way to build
  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
 }

@ -572,8 +582,6 @@ fi

 if [[ "${TEST_CONFIG}" == *cpu* ]]; then
  DYNAMO_BENCHMARK_FLAGS+=(--device cpu)
-elif [[ "${TEST_CONFIG}" == *xpu* ]]; then
-  DYNAMO_BENCHMARK_FLAGS+=(--device xpu)
 else
  DYNAMO_BENCHMARK_FLAGS+=(--device cuda)
 fi
@ -667,8 +675,6 @@ test_perf_for_dashboard() {
    device=cuda_b200
  elif [[ "${TEST_CONFIG}" == *rocm* ]]; then
    device=rocm
-  elif [[ "${TEST_CONFIG}" == *xpu* ]]; then
-    device=xpu
  fi

  for mode in "${modes[@]}"; do
@ -1653,7 +1659,7 @@ test_operator_microbenchmark() {

  cd "${TEST_DIR}"/benchmarks/operator_benchmark

-  for OP_BENCHMARK_TESTS in matmul mm addmm bmm conv; do
+  for OP_BENCHMARK_TESTS in matmul mm addmm bmm; do
    $TASKSET python -m pt.${OP_BENCHMARK_TESTS}_test --tag-filter long \
      --output-json-for-dashboard "${TEST_REPORTS_DIR}/operator_microbenchmark_${OP_BENCHMARK_TESTS}_compile.json" \
      --benchmark-name "PyTorch operator microbenchmark" --use-compile
@ -1761,7 +1767,7 @@ elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
  else
    # Do this after checkout_install_torchbench to ensure we clobber any
    # nightlies that torchbench may pull in
-    if [[ "${TEST_CONFIG}" != *cpu* && "${TEST_CONFIG}" != *xpu* ]]; then
+    if [[ "${TEST_CONFIG}" != *cpu* ]]; then
      install_torchrec_and_fbgemm
    fi
    PYTHONPATH=/torchbench test_dynamo_benchmark torchbench "$id"
@ -1770,7 +1776,7 @@ elif [[ "${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_cpp
+    test_inductor_aoti
  fi
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
  install_torchvision
--- a/.ci/pytorch/windows/internal/install_python.bat
+++ b/.ci/pytorch/windows/internal/install_python.bat
@ -7,9 +7,12 @@ 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-3.14.0-amd64.exe"
+    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0rc1-amd64.exe"
    set ADDITIONAL_OPTIONS="Include_freethreaded=1"
    set PYTHON_EXEC="python3.14t"
 ) else (
--- a/.clang-tidy
+++ b/.clang-tidy
@ -60,11 +60,9 @@ performance-*,
 readability-container-size-empty,
 readability-delete-null-pointer,
 readability-duplicate-include,
-readability-named-parameter,
 readability-misplaced-array-index,
 readability-redundant*,
 readability-simplify-subscript-expr,
-readability-static-definition-in-anonymous-namespace
 readability-string-compare,
 -readability-redundant-access-specifiers,
 -readability-redundant-control-flow,
--- a/.claude/skills/add-uint-support/SKILL.md
+++ b/.claude/skills/add-uint-support/SKILL.md
@ -1,319 +0,0 @@
---
-name: add-uint-support
-description: Add unsigned integer (uint) type support to PyTorch operators by updating AT_DISPATCH macros. Use when adding support for uint16, uint32, uint64 types to operators, kernels, or when user mentions enabling unsigned types, barebones unsigned types, or uint support.
---
-
-# Add Unsigned Integer (uint) Support to Operators
-
-This skill helps add support for unsigned integer types (uint16, uint32, uint64) to PyTorch operators by updating their AT_DISPATCH macros.
-
-## When to use this skill
-
-Use this skill when:
- Adding uint16, uint32, or uint64 support to an operator
- User mentions "unsigned types", "uint support", "barebones unsigned types"
- Enabling support for kUInt16, kUInt32, kUInt64 in kernels
- Working with operator implementations that need expanded type coverage
-
-## Quick reference
-
-**Add unsigned types to existing dispatch:**
-```cpp
-// Before
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
-  kernel<scalar_t>();
-}), AT_EXPAND(AT_ALL_TYPES));
-
-// After (method 1: add unsigned types explicitly)
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
-  kernel<scalar_t>();
-}), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES));
-
-// After (method 2: use V2 integral types if AT_INTEGRAL_TYPES present)
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
-  kernel<scalar_t>();
-}), AT_EXPAND(AT_INTEGRAL_TYPES_V2), AT_EXPAND(AT_FLOATING_TYPES));
-```
-
-## Type group reference
-
-**Unsigned type groups:**
- `AT_BAREBONES_UNSIGNED_TYPES`: kUInt16, kUInt32, kUInt64
- `AT_INTEGRAL_TYPES_V2`: AT_INTEGRAL_TYPES + AT_BAREBONES_UNSIGNED_TYPES
-
-**Relationship:**
-```cpp
-AT_INTEGRAL_TYPES          // kByte, kChar, kInt, kLong, kShort
-AT_BAREBONES_UNSIGNED_TYPES  // kUInt16, kUInt32, kUInt64
-AT_INTEGRAL_TYPES_V2       // INTEGRAL_TYPES + BAREBONES_UNSIGNED_TYPES
-```
-
-## Instructions
-
-### Step 1: Determine if conversion to V2 is needed
-
-Check if the file uses AT_DISPATCH_V2:
-
-**If using old AT_DISPATCH:**
- First convert to AT_DISPATCH_V2 using the at-dispatch-v2 skill
- Then proceed with adding uint support
-
-**If already using AT_DISPATCH_V2:**
- Proceed directly to Step 2
-
-### Step 2: Analyze the current dispatch macro
-
-Identify what type groups are currently in use:
-
-```cpp
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
-  // body
-}), AT_EXPAND(AT_ALL_TYPES), kHalf, kBFloat16);
-    ^^^^^^^^^^^^^^^^^^^^^^^^^
-    Current type coverage
-```
-
-Common patterns:
- `AT_EXPAND(AT_ALL_TYPES)` → includes AT_INTEGRAL_TYPES + AT_FLOATING_TYPES
- `AT_EXPAND(AT_INTEGRAL_TYPES)` → signed integers only
- `AT_EXPAND(AT_FLOATING_TYPES)` → floating point types
-
-### Step 3: Choose the uint addition method
-
-Two approaches:
-
-**Method 1: Add AT_BAREBONES_UNSIGNED_TYPES explicitly**
- Use when: You want to be explicit about adding uint support
- Add `AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES)` to the type list
-
-**Method 2: Substitute AT_INTEGRAL_TYPES with AT_INTEGRAL_TYPES_V2**
- Use when: The dispatch already uses `AT_EXPAND(AT_INTEGRAL_TYPES)`
- More concise: replaces one type group with its superset
- Only applicable if AT_INTEGRAL_TYPES is present
-
-### Step 4: Apply the transformation
-
-**Method 1 example:**
-```cpp
-// Before
-AT_DISPATCH_V2(
-    dtype,
-    "min_values_cuda",
-    AT_WRAP([&]() {
-      kernel_impl<scalar_t>(iter);
-    }),
-    AT_EXPAND(AT_ALL_TYPES),
-    kBFloat16, kHalf, kBool
-);
-
-// After (add unsigned types)
-AT_DISPATCH_V2(
-    dtype,
-    "min_values_cuda",
-    AT_WRAP([&]() {
-      kernel_impl<scalar_t>(iter);
-    }),
-    AT_EXPAND(AT_ALL_TYPES),
-    AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES),
-    kBFloat16, kHalf, kBool
-);
-```
-
-**Method 2 example:**
-```cpp
-// Before
-AT_DISPATCH_V2(
-    dtype,
-    "integral_op",
-    AT_WRAP([&]() {
-      kernel<scalar_t>();
-    }),
-    AT_EXPAND(AT_INTEGRAL_TYPES)
-);
-
-// After (substitute with V2)
-AT_DISPATCH_V2(
-    dtype,
-    "integral_op",
-    AT_WRAP([&]() {
-      kernel<scalar_t>();
-    }),
-    AT_EXPAND(AT_INTEGRAL_TYPES_V2)
-);
-```
-
-### Step 5: Handle AT_ALL_TYPES vs individual type groups
-
-If the dispatch uses `AT_EXPAND(AT_ALL_TYPES)`:
- `AT_ALL_TYPES` = `AT_INTEGRAL_TYPES` + `AT_FLOATING_TYPES`
- To add uint: add `AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES)` to the list
-
-If the dispatch separately lists INTEGRAL and FLOATING:
-```cpp
-// Before
-AT_EXPAND(AT_INTEGRAL_TYPES), AT_EXPAND(AT_FLOATING_TYPES)
-
-// After (Method 2 preferred)
-AT_EXPAND(AT_INTEGRAL_TYPES_V2), AT_EXPAND(AT_FLOATING_TYPES)
-```
-
-### Step 6: Verify all dispatch sites
-
-Check the file for ALL dispatch macros that need uint support:
- Some operators have multiple dispatch sites (CPU, CUDA, different functions)
- Apply the transformation consistently across all sites
- Ensure each gets the same type coverage updates
-
-### Step 7: Validate the changes
-
-Check that:
- [ ] AT_DISPATCH_V2 format is used (not old AT_DISPATCH)
- [ ] Unsigned types are added via one of the two methods
- [ ] All relevant dispatch sites in the file are updated
- [ ] Type groups use `AT_EXPAND()`
- [ ] Arguments are properly formatted and comma-separated
-
-## Common patterns
-
-### Pattern 1: AT_ALL_TYPES + extras
-
-```cpp
-// Before
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
-  kernel<scalar_t>();
-}), AT_EXPAND(AT_ALL_TYPES), kHalf, kBFloat16);
-
-// After
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
-  kernel<scalar_t>();
-}), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kHalf, kBFloat16);
-```
-
-### Pattern 2: Separate INTEGRAL + FLOATING
-
-```cpp
-// Before
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
-  kernel<scalar_t>();
-}), AT_EXPAND(AT_INTEGRAL_TYPES), AT_EXPAND(AT_FLOATING_TYPES));
-
-// After
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
-  kernel<scalar_t>();
-}), AT_EXPAND(AT_INTEGRAL_TYPES_V2), AT_EXPAND(AT_FLOATING_TYPES));
-```
-
-### Pattern 3: Old dispatch needs conversion first
-
-```cpp
-// Before (needs v2 conversion first)
-AT_DISPATCH_ALL_TYPES_AND2(kHalf, kBFloat16, dtype, "op", [&]() {
-  kernel<scalar_t>();
-});
-
-// After v2 conversion
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
-  kernel<scalar_t>();
-}), AT_EXPAND(AT_ALL_TYPES), kHalf, kBFloat16);
-
-// After adding uint support
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
-  kernel<scalar_t>();
-}), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kHalf, kBFloat16);
-```
-
-## Multiple dispatch sites example
-
-For a file with multiple functions:
-
-```cpp
-void min_values_kernel_cuda(TensorIterator& iter) {
-  AT_DISPATCH_V2(iter.dtype(), "min_values_cuda", AT_WRAP([&]() {
-    impl<scalar_t>(iter);
-  }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf);
-  //                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-  //                           Added uint support
-}
-
-void min_launch_kernel(TensorIterator &iter) {
-  AT_DISPATCH_V2(iter.input_dtype(), "min_cuda", AT_WRAP([&]() {
-    gpu_reduce_kernel<scalar_t>(iter);
-  }), AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES), kBFloat16, kHalf);
-  //                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-  //                           Added uint support here too
-}
-```
-
-## Decision tree
-
-Use this decision tree to determine the approach:
-
-```
-Is the file using AT_DISPATCH_V2?
-├─ No → Use at-dispatch-v2 skill first, then continue
-└─ Yes
-   └─ Does it use AT_EXPAND(AT_INTEGRAL_TYPES)?
-      ├─ Yes → Replace with AT_EXPAND(AT_INTEGRAL_TYPES_V2)
-      └─ No → Add AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES) to type list
-```
-
-## Edge cases
-
-### Case 1: Dispatch with only floating types
-
-If the operator only supports floating point types, don't add uint support:
-
-```cpp
-// Leave as-is - floating point only operator
-AT_DISPATCH_V2(dtype, "float_op", AT_WRAP([&]() {
-  kernel<scalar_t>();
-}), AT_EXPAND(AT_FLOATING_TYPES), kHalf);
-```
-
-### Case 2: Complex types present
-
-Unsigned types work alongside complex types:
-
-```cpp
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
-  kernel<scalar_t>();
-}), AT_EXPAND(AT_ALL_TYPES),
-    AT_EXPAND(AT_BAREBONES_UNSIGNED_TYPES),
-    AT_EXPAND(AT_COMPLEX_TYPES),
-    kHalf, kBFloat16);
-```
-
-### Case 3: Already has uint support
-
-Check if uint types are already present:
- If `AT_INTEGRAL_TYPES_V2` is used → already has uint support
- If `AT_BAREBONES_UNSIGNED_TYPES` is already in list → already has uint support
- Skip the file if uint support is already present
-
-## Workflow
-
-When asked to add uint support:
-
-1. Read the target file
-2. Check if using AT_DISPATCH_V2:
-   - If not → use at-dispatch-v2 skill first
-3. Identify all dispatch macro sites
-4. For each dispatch:
-   - Analyze current type groups
-   - Choose method (add BAREBONES_UNSIGNED or upgrade to V2)
-   - Apply transformation with Edit tool
-5. Show the user the changes
-6. Explain what was modified
-
-## Important notes
-
- Always check if v2 conversion is needed first
- Apply changes consistently across all dispatch sites in the file
- Method 2 (AT_INTEGRAL_TYPES_V2) is cleaner when applicable
- Method 1 (explicit AT_BAREBONES_UNSIGNED_TYPES) is more explicit
- Unsigned types are: kUInt16, kUInt32, kUInt64 (not kByte which is uint8)
- Some operators may not semantically support unsigned types - use judgment
-
-## Testing
-
-After adding uint support, the operator should accept uint16, uint32, and uint64 tensors. The user is responsible for functional testing.
--- a/.claude/skills/at-dispatch-v2/SKILL.md
+++ b/.claude/skills/at-dispatch-v2/SKILL.md
@ -1,305 +0,0 @@
---
-name: at-dispatch-v2
-description: Convert PyTorch AT_DISPATCH macros to AT_DISPATCH_V2 format in ATen C++ code. Use when porting AT_DISPATCH_ALL_TYPES_AND*, AT_DISPATCH_FLOATING_TYPES*, or other dispatch macros to the new v2 API. For ATen kernel files, CUDA kernels, and native operator implementations.
---
-
-# AT_DISPATCH to AT_DISPATCH_V2 Converter
-
-This skill helps convert PyTorch's legacy AT_DISPATCH macros to the new AT_DISPATCH_V2 format, as defined in `aten/src/ATen/Dispatch_v2.h`.
-
-## When to use this skill
-
-Use this skill when:
- Converting AT_DISPATCH_* macros to AT_DISPATCH_V2
- Porting ATen kernels to use the new dispatch API
- Working with files in `aten/src/ATen/native/` that use dispatch macros
- User mentions "AT_DISPATCH", "dispatch v2", "Dispatch_v2.h", or macro conversion
-
-## Quick reference
-
-**Old format:**
-```cpp
-AT_DISPATCH_ALL_TYPES_AND3(kBFloat16, kHalf, kBool, dtype, "kernel_name", [&]() {
-  // lambda body
-});
-```
-
-**New format:**
-```cpp
-AT_DISPATCH_V2(dtype, "kernel_name", AT_WRAP([&]() {
-  // lambda body
-}), AT_EXPAND(AT_ALL_TYPES), kBFloat16, kHalf, kBool);
-```
-
-## Key transformations
-
-1. **Reorder arguments**: `scalar_type` and `name` come first, then lambda, then types
-2. **Wrap the lambda**: Use `AT_WRAP(lambda)` to handle internal commas
-3. **Expand type groups**: Use `AT_EXPAND(AT_ALL_TYPES)` instead of implicit expansion
-4. **List individual types**: Add extra types (kHalf, kBFloat16, etc.) after expanded groups
-5. **Add include**: `#include <ATen/Dispatch_v2.h>` near other Dispatch includes
-
-## Instructions
-
-### Step 1: Add the Dispatch_v2.h include
-
-Add the v2 header near the existing `#include <ATen/Dispatch.h>`:
-
-```cpp
-#include <ATen/Dispatch.h>
-#include <ATen/Dispatch_v2.h>
-```
-
-Keep the old Dispatch.h include for now (other code may still need it).
-
-### Step 2: Identify the old dispatch pattern
-
-Common patterns to convert:
-
- `AT_DISPATCH_ALL_TYPES_AND{2,3,4}(type1, type2, ..., scalar_type, name, lambda)`
- `AT_DISPATCH_FLOATING_TYPES_AND{2,3}(type1, type2, ..., scalar_type, name, lambda)`
- `AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND{2,3}(type1, ..., scalar_type, name, lambda)`
- `AT_DISPATCH_FLOATING_AND_COMPLEX_TYPES_AND{2,3}(type1, ..., scalar_type, name, lambda)`
-
-### Step 3: Map the old macro to type groups
-
-Identify which type group macro corresponds to the base types:
-
-| Old macro base | AT_DISPATCH_V2 type group |
-|----------------|---------------------------|
-| `ALL_TYPES` | `AT_EXPAND(AT_ALL_TYPES)` |
-| `FLOATING_TYPES` | `AT_EXPAND(AT_FLOATING_TYPES)` |
-| `INTEGRAL_TYPES` | `AT_EXPAND(AT_INTEGRAL_TYPES)` |
-| `COMPLEX_TYPES` | `AT_EXPAND(AT_COMPLEX_TYPES)` |
-| `ALL_TYPES_AND_COMPLEX` | `AT_EXPAND(AT_ALL_TYPES_AND_COMPLEX)` |
-
-For combined patterns, use multiple `AT_EXPAND()` entries:
-```cpp
-// Old: AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND2(...)
-// New: AT_EXPAND(AT_ALL_TYPES), AT_EXPAND(AT_COMPLEX_TYPES), type1, type2
-```
-
-### Step 4: Extract the individual types
-
-From `AT_DISPATCH_*_AND2(type1, type2, ...)` or `AT_DISPATCH_*_AND3(type1, type2, type3, ...)`, extract the individual types (type1, type2, etc.).
-
-These become the trailing arguments after the type group:
-```cpp
-AT_DISPATCH_V2(..., AT_EXPAND(AT_ALL_TYPES), kBFloat16, kHalf, kBool)
-                                             ^^^^^^^^^^^^^^^^^^^^^^^^
-                                             Individual types from AND3
-```
-
-### Step 5: Transform to AT_DISPATCH_V2
-
-Apply the transformation:
-
-**Pattern:**
-```cpp
-AT_DISPATCH_V2(
-  scalar_type,           // 1st: The dtype expression
-  "name",                // 2nd: The debug string
-  AT_WRAP(lambda),       // 3rd: The lambda wrapped in AT_WRAP
-  type_groups,           // 4th+: Type groups with AT_EXPAND()
-  individual_types       // Last: Individual types
-)
-```
-
-**Example transformation:**
-```cpp
-// BEFORE
-AT_DISPATCH_ALL_TYPES_AND3(
-    kBFloat16, kHalf, kBool,
-    iter.dtype(),
-    "min_values_cuda",
-    [&]() {
-      min_values_kernel_cuda_impl<scalar_t>(iter);
-    }
-);
-
-// AFTER
-AT_DISPATCH_V2(
-    iter.dtype(),
-    "min_values_cuda",
-    AT_WRAP([&]() {
-      min_values_kernel_cuda_impl<scalar_t>(iter);
-    }),
-    AT_EXPAND(AT_ALL_TYPES),
-    kBFloat16, kHalf, kBool
-);
-```
-
-### Step 6: Handle multi-line lambdas
-
-For lambdas with internal commas or complex expressions, AT_WRAP is essential:
-
-```cpp
-AT_DISPATCH_V2(
-    dtype,
-    "complex_kernel",
-    AT_WRAP([&]() {
-      gpu_reduce_kernel<scalar_t, scalar_t>(
-        iter,
-        MinOps<scalar_t>{},
-        thrust::pair<scalar_t, int64_t>(upper_bound(), 0)  // Commas inside!
-      );
-    }),
-    AT_EXPAND(AT_ALL_TYPES)
-);
-```
-
-### Step 7: Verify the conversion
-
-Check that:
- [ ] `AT_WRAP()` wraps the entire lambda
- [ ] Type groups use `AT_EXPAND()`
- [ ] Individual types don't have `AT_EXPAND()` (just `kBFloat16`, not `AT_EXPAND(kBFloat16)`)
- [ ] Argument order is: scalar_type, name, lambda, types
- [ ] Include added: `#include <ATen/Dispatch_v2.h>`
-
-## Type group reference
-
-Available type group macros (use with `AT_EXPAND()`):
-
-```cpp
-AT_INTEGRAL_TYPES      // kByte, kChar, kInt, kLong, kShort
-AT_FLOATING_TYPES      // kDouble, kFloat
-AT_COMPLEX_TYPES       // kComplexDouble, kComplexFloat
-AT_QINT_TYPES         // kQInt8, kQUInt8, kQInt32
-AT_ALL_TYPES          // INTEGRAL_TYPES + FLOATING_TYPES
-AT_ALL_TYPES_AND_COMPLEX  // ALL_TYPES + COMPLEX_TYPES
-AT_INTEGRAL_TYPES_V2  // INTEGRAL_TYPES + unsigned types
-AT_BAREBONES_UNSIGNED_TYPES  // kUInt16, kUInt32, kUInt64
-AT_FLOAT8_TYPES       // Float8 variants
-```
-
-## Common patterns
-
-### Pattern: AT_DISPATCH_ALL_TYPES_AND2
-
-```cpp
-// Before
-AT_DISPATCH_ALL_TYPES_AND2(kHalf, kBFloat16, dtype, "op", [&]() {
-  kernel<scalar_t>(data);
-});
-
-// After
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
-  kernel<scalar_t>(data);
-}), AT_EXPAND(AT_ALL_TYPES), kHalf, kBFloat16);
-```
-
-### Pattern: AT_DISPATCH_FLOATING_TYPES_AND3
-
-```cpp
-// Before
-AT_DISPATCH_FLOATING_TYPES_AND3(kHalf, kBFloat16, kFloat8_e4m3fn,
-    tensor.scalar_type(), "float_op", [&] {
-  process<scalar_t>(tensor);
-});
-
-// After
-AT_DISPATCH_V2(tensor.scalar_type(), "float_op", AT_WRAP([&] {
-  process<scalar_t>(tensor);
-}), AT_EXPAND(AT_FLOATING_TYPES), kHalf, kBFloat16, kFloat8_e4m3fn);
-```
-
-### Pattern: AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND2
-
-```cpp
-// Before
-AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND2(
-    kComplexHalf, kHalf,
-    self.scalar_type(),
-    "complex_op",
-    [&] {
-      result = compute<scalar_t>(self);
-    }
-);
-
-// After
-AT_DISPATCH_V2(
-    self.scalar_type(),
-    "complex_op",
-    AT_WRAP([&] {
-      result = compute<scalar_t>(self);
-    }),
-    AT_EXPAND(AT_ALL_TYPES),
-    AT_EXPAND(AT_COMPLEX_TYPES),
-    kComplexHalf,
-    kHalf
-);
-```
-
-## Edge cases
-
-### Case 1: No extra types (rare)
-
-```cpp
-// Before
-AT_DISPATCH_ALL_TYPES(dtype, "op", [&]() { kernel<scalar_t>(); });
-
-// After
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([&]() {
-  kernel<scalar_t>();
-}), AT_EXPAND(AT_ALL_TYPES));
-```
-
-### Case 2: Many individual types (AND4, AND5, etc.)
-
-```cpp
-// Before
-AT_DISPATCH_FLOATING_TYPES_AND4(kHalf, kBFloat16, kFloat8_e4m3fn, kFloat8_e5m2,
-    dtype, "float8_op", [&]() { kernel<scalar_t>(); });
-
-// After
-AT_DISPATCH_V2(dtype, "float8_op", AT_WRAP([&]() {
-  kernel<scalar_t>();
-}), AT_EXPAND(AT_FLOATING_TYPES), kHalf, kBFloat16, kFloat8_e4m3fn, kFloat8_e5m2);
-```
-
-### Case 3: Lambda with no captures
-
-```cpp
-// Before
-AT_DISPATCH_ALL_TYPES_AND2(kHalf, kBool, dtype, "op", []() {
-  static_kernel<scalar_t>();
-});
-
-// After
-AT_DISPATCH_V2(dtype, "op", AT_WRAP([]() {
-  static_kernel<scalar_t>();
-}), AT_EXPAND(AT_ALL_TYPES), kHalf, kBool);
-```
-
-## Benefits of AT_DISPATCH_V2
-
-1. **No arity in macro name**: Don't need different macros for AND2, AND3, AND4
-2. **Composable type sets**: Mix and match type groups with `AT_EXPAND()`
-3. **Extensible**: Easy to add more types without hitting macro limits
-4. **Clearer**: Type groups are explicit, not implicit in macro name
-
-## Important notes
-
- Keep `#include <ATen/Dispatch.h>` - other code may need it
- The `AT_WRAP()` is mandatory - prevents comma parsing issues in the lambda
- Type groups need `AT_EXPAND()`, individual types don't
- The v2 API is in `aten/src/ATen/Dispatch_v2.h` - refer to it for full docs
- See the header file for the Python script to regenerate the macro implementation
-
-## Workflow
-
-When asked to convert AT_DISPATCH macros:
-
-1. Read the file to identify all AT_DISPATCH uses
-2. Add `#include <ATen/Dispatch_v2.h>` if not present
-3. For each dispatch macro:
-   - Identify the pattern and extract components
-   - Map the base type group
-   - Extract individual types
-   - Construct the AT_DISPATCH_V2 call
-   - Apply with Edit tool
-4. Show the user the complete converted file
-5. Explain what was changed
-
-Do NOT compile or test the code - focus on accurate conversion only.
--- a/.github/actions/diskspace-cleanup/action.yml
+++ b/.github/actions/diskspace-cleanup/action.yml
@ -27,9 +27,7 @@ runs:
            docker system prune -af
            diskspace_new=$(df -H --output=pcent ${docker_root_dir} | sed -n 2p | sed 's/%//' | sed 's/ //')
            if [[ "$diskspace_new" -gt "$diskspace_cutoff" ]] ; then
-                diskspace_cutoff_int=$((diskspace_cutoff + 0))
-                difference=$((100 - diskspace_cutoff_int))
-                echo "Error: Available diskspace is less than $difference percent. Not enough diskspace."
+                echo "Error: Available diskspace is less than $diskspace_cutoff percent. Not enough diskspace."
                echo "$msg"
                exit 1
            else
--- a/.github/actions/pytest-cache-download/action.yml
+++ b/.github/actions/pytest-cache-download/action.yml
@ -38,9 +38,9 @@ runs:
      run: |
        python3 .github/scripts/pytest_cache.py \
          --download \
-          --cache_dir "$GITHUB_WORKSPACE/$CACHE_DIR" \
-          --pr_identifier "$GITHUB_REF" \
-          --job_identifier "$JOB_IDENTIFIER" \
-          --temp_dir "$RUNNER_TEMP" \
-          --repo "$REPO" \
-          --bucket "$BUCKET" \
+          --cache_dir $GITHUB_WORKSPACE/$CACHE_DIR \
+          --pr_identifier $GITHUB_REF \
+          --job_identifier $JOB_IDENTIFIER \
+          --temp_dir $RUNNER_TEMP \
+          --repo $REPO \
+          --bucket $BUCKET \
--- a/.github/actions/pytest-cache-upload/action.yml
+++ b/.github/actions/pytest-cache-upload/action.yml
@ -47,11 +47,11 @@ runs:
      run: |
        python3 .github/scripts/pytest_cache.py \
          --upload \
-          --cache_dir "$GITHUB_WORKSPACE/$CACHE_DIR" \
-          --pr_identifier "$GITHUB_REF" \
-          --job_identifier "$JOB_IDENTIFIER" \
-          --sha "$SHA" \
-          --test_config "$TEST_CONFIG" \
-          --shard "$SHARD" \
-          --repo "$REPO" \
-          --temp_dir "$RUNNER_TEMP" \
+          --cache_dir $GITHUB_WORKSPACE/$CACHE_DIR \
+          --pr_identifier $GITHUB_REF \
+          --job_identifier $JOB_IDENTIFIER \
+          --sha $SHA \
+          --test_config $TEST_CONFIG \
+          --shard $SHARD \
+          --repo $REPO \
+          --temp_dir $RUNNER_TEMP \
--- a/.github/ci_commit_pins/audio.txt
+++ b/.github/ci_commit_pins/audio.txt
@ -1 +1 @@
-3b0e7a6f192ca2715e7e6cbe5db007aea7165fe2
+69bbe7363897764f9e758d851cd0340147d27f94
--- a/.github/ci_commit_pins/vision.txt
+++ b/.github/ci_commit_pins/vision.txt
@ -1 +1 @@
-cfbc5c2f1c798991715a6b06bb3ce46478c4487c
+1752fe6809b74921644866275ab80244b96e80bc
--- a/.github/ci_commit_pins/xla.txt
+++ b/.github/ci_commit_pins/xla.txt
@ -1 +1 @@
-c8b09f5f77d6bf6fb7ed7a9aa83e5d8156b3a5e9
+df6798dfb931ce7c7fe5bed2447cd1092a5981af
--- a/.github/copilot-instructions.md
+++ b/.github/copilot-instructions.md
@ -1,125 +0,0 @@
-# PyTorch Copilot Instructions
-
-This is the PyTorch machine learning framework codebase. These instructions help AI agents navigate and contribute effectively.
-
-## Architecture Overview
-
-### Core Components
-
- **c10/** - Core library (C++-10 compatible) for essential, binary-size-conscious functionality
- **aten/** - ATen tensor library (C++), PyTorch's foundation without autograd
-  - `aten/src/ATen/native/` - Modern operator implementations (CPU/CUDA/MPS/sparse)
-  - `aten/src/ATen/native/native_functions.yaml` - **Critical**: Declarative operator registry
- **torch/** - Python bindings and public API
-  - `torch/csrc/` - C++ Python bindings (hand-written and generated)
-  - `torch/csrc/autograd/` - Reverse-mode automatic differentiation
-  - `torch/csrc/jit/` - TorchScript JIT compiler
- **torchgen/** - Code generation tooling that reads `native_functions.yaml`
- **tools/** - Build scripts, autograd derivatives, code generation
-
-### The Code Generation Workflow
-
-**Most operator changes require editing `native_functions.yaml`**, not direct C++ files. This YAML file:
-1. Declares operator signatures, variants (function/method), and dispatch behavior
-2. Gets processed by `torchgen/` to generate C++/Python bindings
-3. Produces headers in `build/aten/src/ATen/` during compilation
-
-Example entry structure:
-```yaml
- func: my_op(Tensor self, Scalar alpha=1) -> Tensor
-  variants: function, method
-  dispatch:
-    CPU: my_op_cpu
-    CUDA: my_op_cuda
-```
-
-After editing `native_functions.yaml`, implement kernels in `aten/src/ATen/native/` (see `aten/src/ATen/native/README.md`).
-
-## Development Workflows
-
-### Building from Source
-
-**Never run `setup.py` directly** - use pip with editable install:
-```bash
-python -m pip install --no-build-isolation -v -e .
-```
-
-Speed up builds:
- `DEBUG=1` - Debug symbols with `-g -O0`
- `USE_CUDA=0` - Skip CUDA compilation
- `BUILD_TEST=0` - Skip C++ test binaries
- Install `ninja` (`pip install ninja`) for faster builds
- Use `ccache` for incremental compilation caching
-
-Rebuild specific targets: `(cd build && ninja <target>)`
-
-### Testing
-
-**Critical**: DO NOT run entire test suites. Run specific tests only:
-```bash
-python test/test_torch.py TestTorch.test_specific_case
-```
-
-**Test structure**: All tests use `torch.testing._internal.common_utils`:
-```python
-from torch.testing._internal.common_utils import run_tests, TestCase
-
-class TestFeature(TestCase):
-    def test_something(self):
-        # Use self.assertEqual for tensor comparisons
-        pass
-
-if __name__ == "__main__":
-    run_tests()
-```
-
-**For bug fixes**: Create a standalone reproduction script first, verify it fails, then fix and add to appropriate test file.
-
-### Linting
-
-Run linter (not pre-commit): `lintrunner -a` (auto-applies fixes)
-
-## Project-Specific Conventions
-
-### Memory and Storage
- **Storage is never nullptr** (but `StorageImpl.data` may be nullptr for unallocated outputs)
- CUDA device info lives in storage objects
-
-### Python-C++ Integration (`torch/csrc/`)
- Always include `Python.h` **first** to avoid `_XOPEN_SOURCE` redefinition errors
- Use `pybind11::gil_scoped_acquire` before calling Python API or using `THPObjectPtr`
- Wrap entry points with `HANDLE_TH_ERRORS` / `END_HANDLE_TH_ERRORS` for exception conversion
-
-### Dispatch System
- PyTorch uses operator dispatch to route calls to backend-specific kernels
- Prefer `CompositeExplicitAutograd` dispatch when writing device-agnostic compound ops
- See `aten/src/ATen/native/README.md` for dispatch keyword guidance
-
-## Git Workflow (AI Agent Specific)
-
-When preparing PRs from this environment:
-```bash
-git stash -u
-git reset --hard $(cat /tmp/orig_work.txt)  # Reset to LOCAL branch
-git stash pop
-# Resolve conflicts if necessary
-```
-
-## Common Gotchas
-
-1. **Editing generated files** - If it's in `build/`, don't edit it. Edit the source template or `native_functions.yaml`
-2. **NVCC template compilation** - NVCC is stricter about C++ than gcc/clang; code working on Linux may fail Windows CI
-3. **Windows symbol visibility** - Use `TORCH_API` macros for exported symbols (required on Windows, optional on Linux)
-4. **No internet access** - DO NOT attempt to install dependencies during development
-
-## Key Files Reference
-
- `AGENTS.md` - Instructions specific to AI coding agents
- `CONTRIBUTING.md` - Comprehensive human contributor guide
- `GLOSSARY.md` - Terminology (ATen, kernels, operations, JIT, TorchScript)
- `aten/src/ATen/native/README.md` - Operator implementation guide
- `tools/autograd/derivatives.yaml` - Gradient definitions for autograd
-
-## Performance Debugging
-
-Use `TORCH_SHOW_CPP_STACKTRACES=1` for C++ traces in Python errors. For profiling, prefer `py-spy` over manual instrumentation.
--- a/.github/merge_rules.yaml
+++ b/.github/merge_rules.yaml
@ -540,26 +540,6 @@
  - Lint
  - pull

- name: PrivateUse1
-  patterns:
-  - torch/accelerator/**
-  - torch/utils/backend_registration.py
-  - torch/csrc/acc/**
-  - torch/csrc/DeviceAccelerator.*
-  - torch/csrc/profiler/standalone/privateuse1_observer.*
-  - aten/src/ATen/DeviceAccelerator.*
-  - aten/src/ATen/core/GeneratorForPrivateuseone.*
-  - aten/src/ATen/detail/PrivateUse1HooksInterface.*
-  - docs/source/accelerator/**
-  - test/cpp_extensions/open_registration_extension/torch_openreg/**
-  approved_by:
-  - albanD
-  - fffrog
-  mandatory_checks_name:
-  - EasyCLA
-  - Lint
-  - pull
-
 - name: superuser
  patterns:
  - '*'
--- a/.github/pytorch-probot.yml
+++ b/.github/pytorch-probot.yml
@ -19,7 +19,6 @@ ciflow_push_tags:
 - ciflow/inductor-perf-test-nightly-rocm-mi300
 - ciflow/inductor-perf-test-nightly-rocm-mi355
 - ciflow/inductor-perf-test-nightly-x86-zen
- ciflow/inductor-perf-test-nightly-xpu
 - ciflow/inductor-periodic
 - ciflow/inductor-rocm
 - ciflow/linux-aarch64
@ -27,7 +26,6 @@ ciflow_push_tags:
 - ciflow/nightly
 - ciflow/op-benchmark
 - ciflow/periodic
- ciflow/periodic-rocm-mi200
 - ciflow/periodic-rocm-mi300
 - ciflow/pull
 - ciflow/quantization-periodic
--- a/.github/scripts/generate_binary_build_matrix.py
+++ b/.github/scripts/generate_binary_build_matrix.py
@ -11,24 +11,18 @@ architectures:
    * Latest XPU
 """

-import json
 import os
-import re
-from pathlib import Path
 from typing import Optional


-SCRIPT_DIR = Path(__file__).absolute().parent
-REPO_ROOT = SCRIPT_DIR.parent.parent
-
-
+# NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
 CUDA_ARCHES = ["12.6", "12.8", "12.9", "13.0"]
 CUDA_STABLE = "12.8"
 CUDA_ARCHES_FULL_VERSION = {
    "12.6": "12.6.3",
    "12.8": "12.8.1",
    "12.9": "12.9.1",
-    "13.0": "13.0.0",
+    "13.0": "13.0.2",
 }
 CUDA_ARCHES_CUDNN_VERSION = {
    "12.6": "9",
@ -37,7 +31,8 @@ CUDA_ARCHES_CUDNN_VERSION = {
    "13.0": "9",
 }

-ROCM_ARCHES = ["7.0", "7.1"]
+# NOTE: Please also update the ROCm sources in `PIP_SOURCES` in tools/nightly.py when changing this
+ROCM_ARCHES = ["6.4", "7.0"]

 XPU_ARCHES = ["xpu"]

@ -61,7 +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==3.4.5; 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'"
@ -78,7 +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==3.4.5; 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'"
@ -95,7 +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==3.4.5; 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'"
@ -112,7 +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==3.4.5; 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'"
@ -142,48 +137,9 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
 }


-# Used by tools/nightly.py
-PYTORCH_NIGHTLY_PIP_INDEX_URL = "https://download.pytorch.org/whl/nightly"
-NIGHTLY_SOURCE_MATRIX = {
-    "cpu": dict(
-        name="cpu",
-        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/cpu",
-        supported_platforms=["Linux", "macOS", "Windows"],
-        accelerator="cpu",
-    )
-}
-CUDA_NIGHTLY_SOURCE_MATRIX = {
-    f"cuda-{major}.{minor}": dict(
-        name=f"cuda-{major}.{minor}",
-        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/cu{major}{minor}",
-        supported_platforms=["Linux", "Windows"],
-        accelerator="cuda",
-    )
-    for major, minor in (map(int, version.split(".")) for version in CUDA_ARCHES)
-}
-ROCM_NIGHTLY_SOURCE_MATRIX = {
-    f"rocm-{major}.{minor}": dict(
-        name=f"rocm-{major}.{minor}",
-        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/rocm{major}.{minor}",
-        supported_platforms=["Linux"],
-        accelerator="rocm",
-    )
-    for major, minor in (map(int, version.split(".")) for version in ROCM_ARCHES)
-}
-XPU_NIGHTLY_SOURCE_MATRIX = {
-    "xpu": dict(
-        name="xpu",
-        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/xpu",
-        supported_platforms=["Linux"],
-        accelerator="xpu",
-    )
-}
-NIGHTLY_SOURCE_MATRIX.update(CUDA_NIGHTLY_SOURCE_MATRIX)
-NIGHTLY_SOURCE_MATRIX.update(ROCM_NIGHTLY_SOURCE_MATRIX)
-NIGHTLY_SOURCE_MATRIX.update(XPU_NIGHTLY_SOURCE_MATRIX)
-
-
 def get_nccl_wheel_version(arch_version: str) -> str:
+    import re
+
    requirements = map(
        str.strip, re.split("[;|]", PYTORCH_EXTRA_INSTALL_REQUIREMENTS[arch_version])
    )
@ -191,14 +147,17 @@ def get_nccl_wheel_version(arch_version: str) -> str:


 def read_nccl_pin(arch_version: str) -> str:
-    nccl_pin_path = (
-        REPO_ROOT
-        / ".ci"
-        / "docker"
-        / "ci_commit_pins"
-        / f"nccl-cu{arch_version[:2]}.txt"
+    from pathlib import Path
+
+    nccl_pin_path = os.path.join(
+        Path(__file__).absolute().parents[2],
+        ".ci",
+        "docker",
+        "ci_commit_pins",
+        f"nccl-cu{arch_version[:2]}.txt",
    )
-    return nccl_pin_path.read_text().strip()
+    with open(nccl_pin_path) as f:
+        return f.read().strip()


 def validate_nccl_dep_consistency(arch_version: str) -> None:
@ -206,8 +165,7 @@ def validate_nccl_dep_consistency(arch_version: str) -> None:
    wheel_ver = get_nccl_wheel_version(arch_version)
    if not nccl_release_tag.startswith(f"v{wheel_ver}"):
        raise RuntimeError(
-            f"{arch_version} NCCL release tag version {nccl_release_tag} "
-            f"does not correspond to wheel version {wheel_ver}"
+            f"{arch_version} NCCL release tag version {nccl_release_tag} does not correspond to wheel version {wheel_ver}"
        )


@ -454,14 +412,7 @@ def generate_wheels_matrix(
    return ret


-arch_version = ""
-for arch_version in CUDA_ARCHES:
-    validate_nccl_dep_consistency(arch_version)
-del arch_version
-
-
-if __name__ == "__main__":
-    # Used by tools/nightly.py
-    (SCRIPT_DIR / "nightly_source_matrix.json").write_text(
-        json.dumps(NIGHTLY_SOURCE_MATRIX, indent=4) + "\n"
-    )
+validate_nccl_dep_consistency("13.0")
+validate_nccl_dep_consistency("12.9")
+validate_nccl_dep_consistency("12.8")
+validate_nccl_dep_consistency("12.6")
--- a/.github/workflows/_rocm-test.yml
+++ b/.github/workflows/_rocm-test.yml
@ -97,8 +97,8 @@ jobs:
        shell: bash
        run: |
          ngpu=$(rocminfo | grep -c -E 'Name:.*\sgfx')
-          if [[ $ngpu -lt 2 ]]; then #We are temporarily reducing this down to 2 from 4 so that we can run tests on nodes with less gpus.
-            echo "Error: only $ngpu GPU(s) detected, at least 2 GPUs are needed for distributed jobs"
+          if [[ $ngpu -lt 4 ]]; then
+            echo "Error: only $ngpu GPU(s) detected, at least 4 GPUs are needed for distributed jobs"
            exit 1
          fi

--- a/.github/workflows/_xpu-test.yml
+++ b/.github/workflows/_xpu-test.yml
@ -38,10 +38,6 @@ on:
        default: ""
        description: |
          List of tests to include (empty string implies default list)
-      dashboard-tag:
-        required: false
-        type: string
-        default: ""
      disable-monitor:
        description: |
          [Experimental] Disable utilization monitoring for tests.
@ -62,11 +58,6 @@ on:
        required: false
        type: number
        default: 1
-    secrets:
-      HUGGING_FACE_HUB_TOKEN:
-        required: false
-        description: |
-          HF Auth token to avoid rate limits when downloading models or datasets from hub
 permissions:
  id-token: write
  contents: read
@ -205,8 +196,6 @@ jobs:
          PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0' }}
          PYTORCH_TEST_RERUN_DISABLED_TESTS: ${{ matrix.rerun_disabled_tests && '1' || '0' }}
          TESTS_TO_INCLUDE: ${{ inputs.tests-to-include }}
-          DASHBOARD_TAG: ${{ inputs.dashboard-tag }}
-          HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
        timeout-minutes: ${{ fromJson(steps.test-timeout.outputs.timeout) }}
        run: |
          # Fetch aws credential from IMDs
@ -257,8 +246,6 @@ jobs:
            -e PYTORCH_TEST_RERUN_DISABLED_TESTS \
            -e TESTS_TO_INCLUDE \
            -e ZE_AFFINITY_MASK \
-            -e HUGGING_FACE_HUB_TOKEN \
-            -e DASHBOARD_TAG \
            --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
            --ulimit stack=10485760:83886080 \
            --ulimit core=0 \
@ -344,21 +331,5 @@ jobs:
          if-no-files-found: ignore
          path: ./**/core.[1-9]*

-      - name: Authenticate with AWS
-        uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
-        with:
-          role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_upload-benchmark-results
-          # The max duration enforced by the server side
-          role-duration-seconds: 18000
-          aws-region: us-east-1
-
-      - name: Upload the benchmark results
-        uses: pytorch/test-infra/.github/actions/upload-benchmark-results@main
-        with:
-          benchmark-results-dir: test/test-reports
-          dry-run: false
-          schema-version: v3
-          github-token: ${{ secrets.GITHUB_TOKEN }}
-
      - name: Teardown XPU
        uses: ./.github/actions/teardown-xpu
--- a/.github/workflows/build-almalinux-images.yml
+++ b/.github/workflows/build-almalinux-images.yml
@ -36,7 +36,7 @@ jobs:
    runs-on: linux.9xlarge.ephemeral
    strategy:
      matrix:
-        tag: ["cuda12.6", "cuda12.8", "cuda12.9", "cuda13.0", "rocm7.0", "rocm7.1", "cpu"]
+        tag: ["cuda12.6", "cuda12.8", "cuda12.9", "cuda13.0", "rocm6.4", "rocm7.0", "cpu"]
    steps:
      - name: Build docker image
        uses: pytorch/pytorch/.github/actions/binary-docker-build@main
--- a/.github/workflows/build-libtorch-images.yml
+++ b/.github/workflows/build-libtorch-images.yml
@ -52,8 +52,8 @@ jobs:
          { tag: "cuda12.9" },
          { tag: "cuda12.8" },
          { tag: "cuda12.6" },
+          { tag: "rocm6.4"  },
          { tag: "rocm7.0"  },
-          { tag: "rocm7.1"  },
          { tag: "cpu"      },
        ]
    steps:
--- a/.github/workflows/build-magma-rocm-linux.yml
+++ b/.github/workflows/build-magma-rocm-linux.yml
@ -34,7 +34,7 @@ jobs:
      id-token: write
    strategy:
      matrix:
-        rocm_version: ["71", "70"]
+        rocm_version: ["70", "64"]
    steps:
      - name: Checkout PyTorch
        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
--- a/.github/workflows/build-manywheel-images.yml
+++ b/.github/workflows/build-manywheel-images.yml
@ -54,8 +54,8 @@ jobs:
          { name: "manylinuxaarch64-builder",       tag: "cuda12.9",          runner: "linux.arm64.2xlarge.ephemeral" },
          { name: "manylinuxaarch64-builder",       tag: "cuda12.8",          runner: "linux.arm64.2xlarge.ephemeral" },
          { name: "manylinuxaarch64-builder",       tag: "cuda12.6",          runner: "linux.arm64.2xlarge.ephemeral" },
+          { name: "manylinux2_28-builder",          tag: "rocm6.4",           runner: "linux.9xlarge.ephemeral" },
          { name: "manylinux2_28-builder",          tag: "rocm7.0",           runner: "linux.9xlarge.ephemeral" },
-          { name: "manylinux2_28-builder",          tag: "rocm7.1",           runner: "linux.9xlarge.ephemeral" },
          { name: "manylinux2_28-builder",          tag: "cpu",               runner: "linux.9xlarge.ephemeral" },
          { name: "manylinux2_28_aarch64-builder",  tag: "cpu-aarch64",       runner: "linux.arm64.2xlarge.ephemeral" },
          { name: "manylinux2_28-builder",          tag: "xpu",               runner: "linux.9xlarge.ephemeral" },
--- a/.github/workflows/build-triton-wheel.yml
+++ b/.github/workflows/build-triton-wheel.yml
@ -55,7 +55,7 @@ jobs:
        docker-image: ["pytorch/manylinux2_28-builder:cpu"]
        include:
          - device: "rocm"
-            rocm_version: "7.1"
+            rocm_version: "7.0"
            runs_on: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge"
          - device: "cuda"
            rocm_version: ""
@ -159,7 +159,12 @@ jobs:
            WITH_CLANG_LDD="--with-clang-ldd"
          fi

-          docker exec -t "${container_name}" bash -c "${PYTHON_EXECUTABLE} /pytorch/.github/scripts/build_triton_wheel.py --device=$BUILD_DEVICE $RELEASE $WITH_CLANG_LDD"
+          if [[ "${BUILD_DEVICE}" == xpu ]]; then
+            docker exec -t "${container_name}" bash -c "dnf install -y gcc-toolset-13-gcc-c++"
+            docker exec -t "${container_name}" bash -c "source /opt/rh/gcc-toolset-13/enable && ${PYTHON_EXECUTABLE} /pytorch/.github/scripts/build_triton_wheel.py --device=$BUILD_DEVICE $RELEASE"
+          else
+            docker exec -t "${container_name}" bash -c "${PYTHON_EXECUTABLE} /pytorch/.github/scripts/build_triton_wheel.py --device=$BUILD_DEVICE $RELEASE $WITH_CLANG_LDD"
+          fi

          if [[ ("${{ matrix.device }}" == "cuda" || "${{ matrix.device }}" == "xpu") ]]; then
            docker exec -t "${container_name}"  bash -c "auditwheel repair --plat ${PLATFORM} //artifacts/*.whl"
--- a/.github/workflows/docker-builds.yml
+++ b/.github/workflows/docker-builds.yml
@ -57,7 +57,6 @@ 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,
@ -67,7 +66,6 @@ jobs:
          pytorch-linux-jammy-py3.12-halide,
          pytorch-linux-jammy-xpu-n-1-py3,
          pytorch-linux-jammy-xpu-n-py3,
-          pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks,
          pytorch-linux-jammy-py3-clang18-asan,
          pytorch-linux-jammy-py3-clang12-onnx,
          pytorch-linux-jammy-linter,
--- a/.github/workflows/docker-release.yml
+++ b/.github/workflows/docker-release.yml
@ -8,7 +8,6 @@ on:
      - docker.Makefile
      - .github/workflows/docker-release.yml
      - .github/scripts/generate_docker_release_matrix.py
-      - .github/scripts/generate_binary_build_matrix.py
  push:
    branches:
      - nightly
--- a/.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
+++ b/.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
@ -132,7 +132,7 @@ jobs:
      ALPINE_IMAGE: "arm64v8/alpine"
      build_name: manywheel-py3_10-cuda-aarch64-12_6
      build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      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==3.4.5; 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'
+      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'
      timeout-minutes: 420
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
--- a/.github/workflows/generated-linux-binary-libtorch-nightly.yml
+++ b/.github/workflows/generated-linux-binary-libtorch-nightly.yml
@ -384,6 +384,124 @@ jobs:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml

+  libtorch-rocm6_4-shared-with-deps-release-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: rocm6.4
+      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_TYPE: rocm
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm6.4
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      timeout-minutes: 300
+      build_name: libtorch-rocm6_4-shared-with-deps-release
+      build_environment: linux-binary-libtorch
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  libtorch-rocm6_4-shared-with-deps-release-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-rocm6_4-shared-with-deps-release-build
+      - get-label-type
+    runs-on: linux.rocm.gpu.mi250
+    timeout-minutes: 240
+    env:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: rocm6.4
+      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_TYPE: rocm
+      SKIP_ALL_TESTS: 1
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm6.4
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+    permissions:
+      id-token: write
+      contents: read
+    steps:
+      - name: Setup ROCm
+        uses: ./.github/actions/setup-rocm
+      - uses: actions/download-artifact@v4.1.7
+        name: Download Build Artifacts
+        with:
+          name: libtorch-rocm6_4-shared-with-deps-release
+          path: "${{ runner.temp }}/artifacts/"
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      - name: ROCm set GPU_FLAG
+        run: |
+          echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd --device=/dev/dri --group-add video --group-add daemon" >> "${GITHUB_ENV}"
+      - name: configure aws credentials
+        id: aws_creds
+        if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') }}
+        uses: aws-actions/configure-aws-credentials@v4
+        with:
+          role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
+          aws-region: us-east-1
+          role-duration-seconds: 18000
+      - name: Calculate docker image
+        id: calculate-docker-image
+        uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
+        with:
+          docker-registry: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') && '308535385114.dkr.ecr.us-east-1.amazonaws.com' || 'docker.io' }}
+          docker-image-name: libtorch-cxx11-builder
+          custom-tag-prefix: rocm6.4
+          docker-build-dir: .ci/docker
+          working-directory: pytorch
+      - name: Pull Docker image
+        uses: pytorch/test-infra/.github/actions/pull-docker-image@main
+        with:
+          docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
+      - name: Test Pytorch binary
+        uses: ./pytorch/.github/actions/test-pytorch-binary
+        env:
+          DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
+      - name: Teardown ROCm
+        uses: ./.github/actions/teardown-rocm
+  libtorch-rocm6_4-shared-with-deps-release-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-rocm6_4-shared-with-deps-release-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: rocm6.4
+      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_TYPE: rocm
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm6.4
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      build_name: libtorch-rocm6_4-shared-with-deps-release
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
  libtorch-rocm7_0-shared-with-deps-release-build:
    if: ${{ github.repository_owner == 'pytorch' }}
    uses: ./.github/workflows/_binary-build-linux.yml
@ -501,121 +619,3 @@ jobs:
    secrets:
      github-token: ${{ secrets.GITHUB_TOKEN }}
    uses: ./.github/workflows/_binary-upload.yml
-
-  libtorch-rocm7_1-shared-with-deps-release-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: libtorch
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: rocm7.1
-      GPU_ARCH_VERSION: "7.1"
-      GPU_ARCH_TYPE: rocm
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm7.1
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      timeout-minutes: 300
-      build_name: libtorch-rocm7_1-shared-with-deps-release
-      build_environment: linux-binary-libtorch
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  libtorch-rocm7_1-shared-with-deps-release-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - libtorch-rocm7_1-shared-with-deps-release-build
-      - get-label-type
-    runs-on: linux.rocm.gpu.mi250
-    timeout-minutes: 240
-    env:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: libtorch
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: rocm7.1
-      GPU_ARCH_VERSION: "7.1"
-      GPU_ARCH_TYPE: rocm
-      SKIP_ALL_TESTS: 1
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm7.1
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-    permissions:
-      id-token: write
-      contents: read
-    steps:
-      - name: Setup ROCm
-        uses: ./.github/actions/setup-rocm
-      - uses: actions/download-artifact@v4.1.7
-        name: Download Build Artifacts
-        with:
-          name: libtorch-rocm7_1-shared-with-deps-release
-          path: "${{ runner.temp }}/artifacts/"
-      - name: Checkout PyTorch
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-          submodules: recursive
-          path: pytorch
-          show-progress: false
-      - name: Clean PyTorch checkout
-        run: |
-          # Remove any artifacts from the previous checkouts
-          git clean -fxd
-        working-directory: pytorch
-      - name: ROCm set GPU_FLAG
-        run: |
-          echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd --device=/dev/dri --group-add video --group-add daemon" >> "${GITHUB_ENV}"
-      - name: configure aws credentials
-        id: aws_creds
-        if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') }}
-        uses: aws-actions/configure-aws-credentials@v4
-        with:
-          role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-          aws-region: us-east-1
-          role-duration-seconds: 18000
-      - name: Calculate docker image
-        id: calculate-docker-image
-        uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
-        with:
-          docker-registry: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') && '308535385114.dkr.ecr.us-east-1.amazonaws.com' || 'docker.io' }}
-          docker-image-name: libtorch-cxx11-builder
-          custom-tag-prefix: rocm7.1
-          docker-build-dir: .ci/docker
-          working-directory: pytorch
-      - name: Pull Docker image
-        uses: pytorch/test-infra/.github/actions/pull-docker-image@main
-        with:
-          docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
-      - name: Test Pytorch binary
-        uses: ./pytorch/.github/actions/test-pytorch-binary
-        env:
-          DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
-      - name: Teardown ROCm
-        uses: ./.github/actions/teardown-rocm
-  libtorch-rocm7_1-shared-with-deps-release-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: libtorch-rocm7_1-shared-with-deps-release-test
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: libtorch
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: rocm7.1
-      GPU_ARCH_VERSION: "7.1"
-      GPU_ARCH_TYPE: rocm
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm7.1
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-      build_name: libtorch-rocm7_1-shared-with-deps-release
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml
--- a/.github/workflows/generated-linux-binary-manywheel-nightly.yml
+++ b/.github/workflows/generated-linux-binary-manywheel-nightly.yml
--- a/.github/workflows/inductor-perf-test-nightly-xpu.yml
+++ b/.github/workflows/inductor-perf-test-nightly-xpu.yml
@ -1,148 +0,0 @@
-name: inductor-perf-nightly-xpu
-
-on:
-  push:
-    tags:
-      - ciflow/inductor-perf-test-nightly-xpu/*
-  schedule:
-    - cron: 30 17 * * *
-  workflow_dispatch:
-    inputs:
-      training:
-        description: Run training (on by default)?
-        required: false
-        type: boolean
-        default: true
-      inference:
-        description: Run inference (on by default)?
-        required: false
-        type: boolean
-        default: true
-      default:
-        description: Run inductor_default?
-        required: false
-        type: boolean
-        default: false
-      dynamic:
-        description: Run inductor_dynamic_shapes?
-        required: false
-        type: boolean
-        default: false
-      cppwrapper:
-        description: Run inductor_cpp_wrapper?
-        required: false
-        type: boolean
-        default: false
-      cudagraphs:
-        description: Run inductor_cudagraphs?
-        required: false
-        type: boolean
-        default: false
-      freezing_cudagraphs:
-        description: Run inductor_cudagraphs with freezing for inference?
-        required: false
-        type: boolean
-        default: false
-      aotinductor:
-        description: Run aot_inductor for inference?
-        required: false
-        type: boolean
-        default: false
-      maxautotune:
-        description: Run inductor_max_autotune?
-        required: false
-        type: boolean
-        default: false
-      benchmark_configs:
-        description: The list of configs used the benchmark
-        required: false
-        type: string
-        default: inductor_huggingface_perf,inductor_timm_perf,inductor_torchbench_perf,cachebench
-
-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' }}
-  cancel-in-progress: true
-
-permissions: read-all
-
-jobs:
-  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: ${{ github.ref_type }}
-      opt_out_experiments: lf
-
-  xpu-n-py3_10-inductor-benchmark-build:
-    name: xpu-n-py3.10-inductor-benchmark
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-xpu-n-py3.10
-      docker-image-name: ci-image:pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks
-      runner: linux.c7i.12xlarge
-      test-matrix: |
-        { include: [
-          { config: "inductor_huggingface_perf_xpu", shard: 1, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_huggingface_perf_xpu", shard: 2, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_huggingface_perf_xpu", shard: 3, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_huggingface_perf_xpu", shard: 4, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_huggingface_perf_xpu", shard: 5, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 1, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 2, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 3, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 4, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 5, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 6, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 1, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 2, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 3, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 4, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 5, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 6, num_shards: 6, runner: "linux.idc.xpu" },
-        ]}
-    secrets: inherit
-
-  xpu-n-py3_10-inductor-benchmark-test-nightly:
-    permissions:
-      id-token: write
-      contents: read
-    if: github.event_name != 'workflow_dispatch'
-    name: xpu-n-py3.10-inductor-benchmark
-    uses: ./.github/workflows/_xpu-test.yml
-    needs: xpu-n-py3_10-inductor-benchmark-build
-    with:
-      build-environment: linux-jammy-xpu-n-py3.10
-      dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-false-cppwrapper-true-aotinductor-true-freezing_cudagraphs-false-cudagraphs_low_precision-false
-      docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}
-      timeout-minutes: 720
-      # Disable monitor in perf tests for more investigation
-      disable-monitor: true
-      monitor-log-interval: 10
-      monitor-data-collect-interval: 2
-    secrets: inherit
-
-  xpu-n-py3_10-inductor-benchmark-test:
-    permissions:
-      id-token: write
-      contents: read
-    if: github.event_name == 'workflow_dispatch'
-    name: xpu-n-py3.10-inductor-test
-    uses: ./.github/workflows/_xpu-test.yml
-    needs: xpu-n-py3_10-inductor-benchmark-build
-    with:
-      build-environment: linux-jammy-xpu-n-py3.10
-      dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cudagraphs-${{ inputs.cudagraphs }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}-maxautotune-${{ inputs.maxautotune }}-freezing_cudagraphs-${{ inputs.freezing_cudagraphs }}-cudagraphs_low_precision-${{ inputs.cudagraphs }}
-      docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}
-      timeout-minutes: 720
-      disable-monitor: false
-      monitor-log-interval: 15
-      monitor-data-collect-interval: 4
-    secrets: inherit
--- a/.github/workflows/inductor-rocm.yml
+++ b/.github/workflows/inductor-rocm.yml
@ -1,10 +1,9 @@
 name: inductor-rocm

 on:
-  schedule:
-    - cron: 0 * * * *
  push:
    branches:
+      - main
      - release/*
    tags:
      - ciflow/inductor-rocm/*
--- a/.github/workflows/inductor-unittest.yml
+++ b/.github/workflows/inductor-unittest.yml
@ -115,10 +115,10 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      test-matrix: |
        { include: [
-          { config: "inductor_amx", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "inductor_amx", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "inductor_avx2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.avx2" },
-          { config: "inductor_avx2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.avx2" },
+          { config: "inductor_amx", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "inductor_amx", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "inductor_avx2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.10xlarge.avx2" },
+          { config: "inductor_avx2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.10xlarge.avx2" },
        ]}
    secrets: inherit

--- a/.github/workflows/inductor.yml
+++ b/.github/workflows/inductor.yml
@ -84,13 +84,13 @@ jobs:
      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
      test-matrix: |
        { include: [
-          { config: "cpu_inductor_torchbench", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "cpu_inductor_torchbench", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "dynamic_cpu_inductor_huggingface", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "dynamic_cpu_inductor_timm", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "dynamic_cpu_inductor_timm", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "dynamic_cpu_inductor_torchbench", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
-          { config: "dynamic_cpu_inductor_torchbench", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge.amx" },
+          { config: "cpu_inductor_torchbench", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "cpu_inductor_torchbench", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "dynamic_cpu_inductor_huggingface", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "dynamic_cpu_inductor_timm", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "dynamic_cpu_inductor_timm", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "dynamic_cpu_inductor_torchbench", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
+          { config: "dynamic_cpu_inductor_torchbench", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.8xlarge.amx" },
          { config: "inductor_torchbench_cpu_smoketest_perf", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.24xl.spr-metal" },
        ]}
      build-additional-packages: "vision audio torchao"
--- a/.github/workflows/lint.yml
+++ b/.github/workflows/lint.yml
@ -76,12 +76,11 @@ jobs:

  # NOTE: mypy needs its own job because it depends on --all-files, without assessing all files it sometimes
  #       fails to find types when it should
-  # NOTE: We should be able to disable this and consolidate with Pyrefly
-  lintrunner-pyrefly:
+  lintrunner-mypy:
    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
-    name: lintrunner-pyrefly-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
+    name: lintrunner-mypy-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
    needs: [get-label-type, get-changed-files]
-    # Only run if there are changed files relevant to pyrefly
+    # Only run if there are changed files relevant to mypy
    if: |
      github.repository_owner == 'pytorch' && (
        needs.get-changed-files.outputs.changed-files == '*' ||
@ -99,8 +98,8 @@ jobs:
      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
      script: |
        CHANGED_FILES="${{ needs.get-changed-files.outputs.changed-files }}"
-        echo "Running pyrefly"
-        ADDITIONAL_LINTRUNNER_ARGS="--take PYREFLY --all-files" .github/scripts/lintrunner.sh
+        echo "Running mypy"
+        ADDITIONAL_LINTRUNNER_ARGS="--take MYPY,MYPYSTRICT --all-files" .github/scripts/lintrunner.sh

  lintrunner-noclang:
    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
@ -119,9 +118,9 @@ jobs:
        CHANGED_FILES="${{ needs.get-changed-files.outputs.changed-files }}"
        echo "Running all other linters"
        if [ "$CHANGED_FILES" = '*' ]; then
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,PYREFLY --all-files" .github/scripts/lintrunner.sh
+          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY --all-files" .github/scripts/lintrunner.sh
        else
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,PYREFLY ${CHANGED_FILES}" .github/scripts/lintrunner.sh
+          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY ${CHANGED_FILES}" .github/scripts/lintrunner.sh
        fi

  quick-checks:
--- a/.github/workflows/nightly.yml
+++ b/.github/workflows/nightly.yml
@ -41,7 +41,7 @@ jobs:
    uses: ./.github/workflows/_linux-build.yml
    needs: get-label-type
    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge"
      build-environment: linux-jammy-py3.10-gcc11
      docker-image-name: ci-image:pytorch-linux-jammy-py3.10-gcc11
    secrets: inherit
--- a/.github/workflows/periodic-rocm-mi200.yml
+++ b/.github/workflows/periodic-rocm-mi200.yml
@ -1,84 +0,0 @@
-name: periodic-rocm-mi200
-
-on:
-  schedule:
-    # We have several schedules so jobs can check github.event.schedule to activate only for a fraction of the runs.
-    # Also run less frequently on weekends.
-    - cron: 45 0,8,16 * * 1-5
-    - cron: 45 4 * * 0,6
-    - cron: 45 4,12,20 * * 1-5
-    - cron: 45 12 * * 0,6
-    - cron: 29 8 * * *  # about 1:29am PDT, for mem leak check and rerun disabled tests
-  push:
-    tags:
-      - ciflow/periodic/*
-      - ciflow/periodic-rocm-mi200/*
-    branches:
-      - release/*
-  workflow_dispatch:
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}-${{ github.event.schedule }}
-  cancel-in-progress: true
-
-permissions:
-  id-token: write
-  contents: read
-
-jobs:
-  llm-td:
-    if: github.repository_owner == 'pytorch'
-    name: before-test
-    uses: ./.github/workflows/llm_td_retrieval.yml
-    permissions:
-      id-token: write
-      contents: read
-
-  target-determination:
-    name: before-test
-    uses: ./.github/workflows/target_determination.yml
-    needs: llm-td
-    permissions:
-      id-token: write
-      contents: read
-
-  get-label-type:
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    if: (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch'
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-
-  linux-jammy-rocm-py3_10-build:
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
-      test-matrix: |
-        { include: [
-          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
-        ]}
-    secrets: inherit
-
-  linux-jammy-rocm-py3_10-test:
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-jammy-rocm-py3_10-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
-    secrets: inherit
--- a/.github/workflows/periodic.yml
+++ b/.github/workflows/periodic.yml
@ -204,6 +204,37 @@ jobs:
      test-matrix: ${{ needs.linux-jammy-cuda13_0-py3_10-gcc11-build.outputs.test-matrix }}
    secrets: inherit

+  linux-jammy-rocm-py3_10-build:
+    name: linux-jammy-rocm-py3.10
+    uses: ./.github/workflows/_linux-build.yml
+    needs: get-label-type
+    with:
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
+      test-matrix: |
+        { include: [
+          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
+        ]}
+    secrets: inherit
+
+  linux-jammy-rocm-py3_10-test:
+    permissions:
+      id-token: write
+      contents: read
+    name: linux-jammy-rocm-py3.10
+    uses: ./.github/workflows/_rocm-test.yml
+    needs:
+      - linux-jammy-rocm-py3_10-build
+      - target-determination
+    with:
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
+    secrets: inherit
+
  linux-jammy-cuda12_8-py3-gcc11-slow-gradcheck-build:
    name: linux-jammy-cuda12.8-py3-gcc11-slow-gradcheck
    uses: ./.github/workflows/_linux-build.yml
--- a/.github/workflows/pull.yml
+++ b/.github/workflows/pull.yml
@ -66,10 +66,10 @@ jobs:
          { config: "default", shard: 5, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
          { config: "docs_test", shard: 1, num_shards: 1,  runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
          { config: "jit_legacy", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
-          { config: "backwards_compat", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+          { config: "backwards_compat", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
          { config: "distributed", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
          { config: "distributed", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
-          { config: "numpy_2_x", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+          { config: "numpy_2_x", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
        ]}
    secrets: inherit

@ -167,8 +167,8 @@ jobs:
      docker-image-name: ci-image:pytorch-linux-jammy-py3-clang12-onnx
      test-matrix: |
        { include: [
-          { config: "default", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
-          { config: "default", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.c7i.2xlarge" },
+          { config: "default", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
+          { config: "default", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge" },
        ]}
    secrets: inherit

--- a/.github/workflows/rocm.yml
+++ b/.github/workflows/rocm.yml
@ -3,13 +3,13 @@ name: rocm
 on:
  push:
    branches:
+      - main
      - release/*
    tags:
      - ciflow/rocm/*
  workflow_dispatch:
  schedule:
    - cron: 29 8 * * *  # about 1:29am PDT
-    - cron: 0 * * * *

 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' }}
--- a/.github/workflows/trunk.yml
+++ b/.github/workflows/trunk.yml
@ -204,7 +204,6 @@ jobs:
        { include: [
          { config: "default", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
          { config: "default", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "distributed", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.gfx942.4" },
        ]}
    secrets: inherit

@ -222,7 +221,7 @@ jobs:
      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 }}
-      tests-to-include: "test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor distributed/test_c10d_common distributed/test_c10d_nccl"
+      tests-to-include: "test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor"
    secrets: inherit

  inductor-build:
--- a/.github/workflows/upload-test-stats.yml
+++ b/.github/workflows/upload-test-stats.yml
@ -6,7 +6,6 @@ on:
      - pull
      - trunk
      - periodic
-      - periodic-rocm-mi200
      - periodic-rocm-mi300
      - inductor
      - unstable
--- a/.github/workflows/xpu.yml
+++ b/.github/workflows/xpu.yml
@ -59,18 +59,14 @@ jobs:
      runner: linux.c7i.12xlarge
      test-matrix: |
        { include: [
-          { 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" },
+          { 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" },
        ]}
    secrets: inherit

--- a/.gitignore
+++ b/.gitignore
@ -143,7 +143,6 @@ scripts/release_notes/*.json
 sccache-stats*.json
 lint.json
 merge_record.json
-.github/scripts/nightly_source_matrix.json

 # These files get copied over on invoking setup.py
 torchgen/packaged/*
@ -398,4 +397,3 @@ CLAUDE.local.md
 /test_*.py
 /debug_*.py
 CLAUDE_CONTEXT/
-/.claude/settings.local.json
--- a/.lintrunner.toml
+++ b/.lintrunner.toml
@ -121,6 +121,94 @@ command = [
 ]
 is_formatter = true

+[[linter]]
+code = 'MYPY'
+include_patterns = [
+    'setup.py',
+    'functorch/dim/**/*.py',
+    'torch/**/*.py',
+    'torch/**/*.pyi',
+    'caffe2/**/*.py',
+    'caffe2/**/*.pyi',
+    'test/test_bundled_images.py',
+    'test/test_bundled_inputs.py',
+    'test/test_complex.py',
+    'test/test_datapipe.py',
+    'test/test_futures.py',
+    'test/test_numpy_interop.py',
+    'test/test_torch.py',
+    'test/test_type_hints.py',
+    'test/test_type_info.py',
+    'test/test_utils.py',
+]
+exclude_patterns = [
+    '**/fb/**',
+]
+command = [
+    'python3',
+    'tools/linter/adapters/mypy_linter.py',
+    '--config=mypy.ini',
+    '--',
+    '@{{PATHSFILE}}'
+]
+init_command = [
+    'python3',
+    'tools/linter/adapters/pip_init.py',
+    '--dry-run={{DRYRUN}}',
+    'numpy==1.26.4 ; python_version >= "3.10" and python_version <= "3.11"',
+    'numpy==2.1.0 ; python_version >= "3.12"',
+    'expecttest==0.3.0',
+    'mypy==1.16.0',
+    'sympy==1.13.3',
+    'types-requests==2.27.25',
+    'types-pyyaml==6.0.2',
+    'types-tabulate==0.8.8',
+    'types-protobuf==5.29.1.20250403',
+    'types-setuptools==79.0.0.20250422',
+    'types-jinja2==2.11.9',
+    'types-colorama==0.4.6',
+    'filelock==3.18.0',
+    'junitparser==2.1.1',
+    'rich==14.1.0',
+    'pyyaml==6.0.2',
+    'optree==0.13.0',
+    'dataclasses-json==0.6.7',
+    'pandas==2.2.3',
+]
+
+[[linter]]
+code = 'MYPYSTRICT'
+include_patterns = [
+    '.github/**/*.py',
+    'benchmarks/instruction_counts/**/*.py',
+    'tools/**/*.py',
+    'torchgen/**/*.py',
+    'torch/utils/_pytree.py',
+    'torch/utils/_cxx_pytree.py',
+    'torch/utils/benchmark/utils/common.py',
+    'torch/utils/benchmark/utils/timer.py',
+    'torch/utils/benchmark/utils/valgrind_wrapper/**/*.py',
+]
+exclude_patterns = [
+    # (linbinyu) copied from internal repo
+    '**/fb/**',
+    'tools/code_analyzer/gen_operators_yaml.py',
+    'tools/dynamo/verify_dynamo.py',
+    'tools/gen_vulkan_spv.py',
+    'tools/test/gen_operators_yaml_test.py',
+    'tools/test/gen_oplist_test.py',
+    'tools/test/test_selective_build.py',
+    'tools/experimental/torchfuzz/**',
+]
+command = [
+    'python3',
+    'tools/linter/adapters/mypy_linter.py',
+    '--config=mypy-strict.ini',
+    '--code=MYPYSTRICT',
+    '--',
+    '@{{PATHSFILE}}'
+]
+

 [[linter]]
 code = 'PYREFLY'
@ -142,7 +230,6 @@ init_command = [
    'python3',
    'tools/linter/adapters/pip_init.py',
    '--dry-run={{DRYRUN}}',
-    'numpy==1.26.4 ; python_version >= "3.10" and python_version <= "3.11"',
    'numpy==2.1.0 ; python_version >= "3.12"',
    'expecttest==0.3.0',
    'pyrefly==0.36.2',
@ -211,6 +298,7 @@ exclude_patterns = [
    '**/*pb.h',
    '**/*inl.h',
    'aten/src/ATen/cpu/FlushDenormal.cpp',
+    'aten/src/ATen/cpu/Utils.cpp',
    'aten/src/ATen/cpu/vml.h',
    'aten/src/ATen/CPUFixedAllocator.h',
    'aten/src/ATen/Parallel*.h',
@ -229,6 +317,8 @@ exclude_patterns = [
    'c10/util/win32-headers.h',
    'c10/test/**/*.h',
    'third_party/**/*',
+    'torch/csrc/api/include/torch/nn/modules/common.h',
+    'torch/csrc/api/include/torch/linalg.h',
    'torch/csrc/autograd/generated/**',
    'torch/csrc/distributed/**/*.cu',
    'torch/csrc/distributed/c10d/WinSockUtils.hpp',
@ -240,6 +330,7 @@ exclude_patterns = [
    'torch/csrc/utils/generated_serialization_types.h',
    'torch/csrc/utils/pythoncapi_compat.h',
    'torch/csrc/inductor/aoti_runtime/sycl_runtime_wrappers.h',
+    'aten/src/ATen/ExpandBase.h',
 ]
 init_command = [
    'python3',
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -234,17 +234,7 @@ option(USE_COLORIZE_OUTPUT "Colorize output during compilation" ON)
 option(USE_ASAN "Use Address+Undefined Sanitizers" OFF)
 option(USE_LSAN "Use Leak Sanitizer" OFF)
 option(USE_TSAN "Use Thread Sanitizer" OFF)
-
-# Track whether USE_CUDA was explicitly set by the user (before option() is called)
-# If USE_CUDA is already defined in cache, it means user explicitly set it
-if(DEFINED CACHE{USE_CUDA})
-  set(_USE_CUDA_EXPLICITLY_SET TRUE)
-else()
-  set(_USE_CUDA_EXPLICITLY_SET FALSE)
-endif()
-
 option(USE_CUDA "Use CUDA" ON)
-
 option(USE_XPU "Use XPU" ON)
 cmake_dependent_option(
  BUILD_LAZY_CUDA_LINALG "Build cuda linalg ops as separate library" ON
@ -384,7 +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 folder"
+cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin fodler"
                       OFF "USE_CUDA" OFF)
 cmake_dependent_option(USE_KLEIDIAI "Use KleidiAI for the ARM CPU & AARCH64 architecture." ON
                        "CPU_AARCH64" OFF)
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@ -11,6 +11,7 @@ aspects of contributing to PyTorch.
 <!-- toc -->

 - [Developing PyTorch](#developing-pytorch)
+  - [Setup the development environment](#setup-the-development-environment)
  - [Tips and Debugging](#tips-and-debugging)
 - [Nightly Checkout & Pull](#nightly-checkout--pull)
 - [Codebase structure](#codebase-structure)
@ -66,6 +67,23 @@ aspects of contributing to PyTorch.

 Follow the instructions for [installing PyTorch from source](https://github.com/pytorch/pytorch#from-source). If you get stuck when developing PyTorch on your machine, check out the [tips and debugging](#tips-and-debugging) section below for common solutions.

+### Setup the development environment
+
+First, you need to [fork the PyTorch project on GitHub](https://github.com/pytorch/pytorch/fork) and follow the instructions at [Connecting to GitHub with SSH](https://docs.github.com/en/authentication/connecting-to-github-with-ssh) to setup your SSH authentication credentials.
+
+Then clone the PyTorch project and setup the development environment:
+
+```bash
+git clone git@github.com:<USERNAME>/pytorch.git
+cd pytorch
+git remote add upstream git@github.com:pytorch/pytorch.git
+
+make setup-env
+# Or run `make setup-env-cuda` for pre-built CUDA binaries
+# Or run `make setup-env-rocm` for pre-built ROCm binaries
+source venv/bin/activate  # or `. .\venv\Scripts\activate` on Windows
+```
+
 ### Tips and Debugging

 * If you want to have no-op incremental rebuilds (which are fast), see [Make no-op build fast](#make-no-op-build-fast) below.
--- a/README.md
+++ b/README.md
@ -1,4 +1,4 @@
-![PyTorch Logo](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/pytorch-logo-dark.png)
+![PyTorch Logo](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/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/raw/main/docs/source/_static/img/tensor_illustration.png)
+![Tensor illustration](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/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/raw/main/docs/source/_static/img/dynamic_graph.gif)
+![Dynamic graph](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/dynamic_graph.gif)

 ### Python First

--- a/SECURITY.md
+++ b/SECURITY.md
@ -1,7 +1,7 @@
 # Security Policy

 - [**Reporting a Vulnerability**](#reporting-a-vulnerability)
- - [**Using PyTorch Securely**](#using-pytorch-securely)
+ - [**Using Pytorch Securely**](#using-pytorch-securely)
   - [Untrusted models](#untrusted-models)
   - [TorchScript models](#torchscript-models)
   - [Untrusted inputs](#untrusted-inputs)
@ -10,28 +10,28 @@
 - [**CI/CD security principles**](#cicd-security-principles)
 ## Reporting Security Issues

-Beware that none of the topics under [Using PyTorch Securely](#using-pytorch-securely) are considered vulnerabilities of PyTorch.
+Beware that none of the topics under [Using Pytorch Securely](#using-pytorch-securely) are considered vulnerabilities of Pytorch.

 However, if you believe you have found a security vulnerability in PyTorch, we encourage you to let us know right away. We will investigate all legitimate reports and do our best to quickly fix the problem.

 Please report security issues using https://github.com/pytorch/pytorch/security/advisories/new

-All reports submitted through the security advisories mechanism would **either be made public or dismissed by the team within 90 days of the submission**. If advisory has been closed on the grounds that it is not a security issue, please do not hesitate to create an [new issue](https://github.com/pytorch/pytorch/issues/new?template=bug-report.yml) as it is still likely a valid issue within the framework.
+All reports submitted thru the security advisories mechanism would **either be made public or dismissed by the team within 90 days of the submission**. If advisory has been closed on the grounds that it is not a security issue, please do not hesitate to create an [new issue](https://github.com/pytorch/pytorch/issues/new?template=bug-report.yml) as it is still likely a valid issue within the framework.

 Please refer to the following page for our responsible disclosure policy, reward guidelines, and those things that should not be reported:

 https://www.facebook.com/whitehat


-## Using PyTorch Securely
-**PyTorch models are programs**, so treat its security seriously -- running untrusted models is equivalent to running untrusted code. In general we recommend that model weights and the python code for the model are distributed independently. That said, be careful about where you get the python code from and who wrote it (preferentially check for a provenance or checksums, do not run any pip installed package).
+## Using Pytorch Securely
+**Pytorch models are programs**, so treat its security seriously -- running untrusted models is equivalent to running untrusted code. In general we recommend that model weights and the python code for the model are distributed independently. That said, be careful about where you get the python code from and who wrote it (preferentially check for a provenance or checksums, do not run any pip installed package).

 ### Untrusted models
 Be careful when running untrusted models. This classification includes models created by unknown developers or utilizing data obtained from unknown sources[^data-poisoning-sources].

 **Prefer to execute untrusted models within a secure, isolated environment such as a sandbox** (e.g., containers, virtual machines). This helps protect your system from potentially malicious code. You can find further details and instructions in [this page](https://developers.google.com/code-sandboxing).

-**Be mindful of risky model formats**. Give preference to share and load weights with the appropriate format for your use case. [Safetensors](https://huggingface.co/docs/safetensors/en/index) gives the most safety but is the most restricted in what it supports. [`torch.load`](https://pytorch.org/docs/stable/generated/torch.load.html#torch.load) has a significantly larger surface of attack but is more flexible in what it can serialize. See the documentation for more details.
+**Be mindful of risky model formats**. Give preference to share and load weights with the appropriate format for your use case. [safetensors](https://huggingface.co/docs/safetensors/en/index) gives the most safety but is the most restricted in what it supports. [`torch.load`](https://pytorch.org/docs/stable/generated/torch.load.html#torch.load) has a significantly larger surface of attack but is more flexible in what it can serialize. See the documentation for more details.

 Even for more secure serialization formats, unexpected inputs to the downstream system can cause diverse security threats (e.g. denial of service, out of bound reads/writes) and thus we recommend extensive validation of any untrusted inputs.

@ -43,7 +43,7 @@ Important Note: The trustworthiness of a model is not binary. You must always de

 ### TorchScript models

-TorchScript models should be treated the same way as locally executable code from an unknown source. Only run TorchScript models if you trust the provider. Please note, that tools for introspecting TorchScript models (such as `torch.utils.model_dump`) may also execute partial or full code stored in those models, therefore they should be used only if you trust the provider of the binary you are about to load.
+TorchScript models should treated the same way as locally executable code from an unknown source. Only run TorchScript models if you trust the provider. Please note, that tools for introspecting TorchScript models (such as `torch.utils.model_dump`) may also execute partial or full code stored in those models, therefore they should be used only if you trust the provider of the binary you are about to load.

 ### Untrusted inputs during training and prediction

@ -59,9 +59,9 @@ If applicable, prepare your model against bad inputs and prompt injections. Some

 ### Data privacy

-**Take special security measures if you train your models with sensitive data**. Prioritize [sandboxing](https://developers.google.com/code-sandboxing) your models and:
- Do not feed sensitive data to an untrusted model (even if runs in a sandboxed environment)
- If you consider publishing a model that was partially trained with sensitive data, be aware that data can potentially be recovered from the trained weights (especially if the model overfits).
+**Take special security measures if your model if you train models with sensitive data**. Prioritize [sandboxing](https://developers.google.com/code-sandboxing) your models and:
+- Do not feed sensitive data to untrusted model (even if runs in a sandboxed environment)
+- If you consider publishing a model that was partially trained with sensitive data, be aware that data can potentially be recovered from the trained weights (especially if model overfits).

 ### Using distributed features

--- a/aten/src/ATen/CMakeLists.txt
+++ b/aten/src/ATen/CMakeLists.txt
@ -260,7 +260,7 @@ IF(USE_FBGEMM_GENAI)
  if(USE_CUDA)
    # To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
    # If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
-    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*(mx8mx8bf16_grouped|f4f4bf16_grouped|f4f4bf16).*")
+    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*mx8mx8bf16_grouped.*")
    file(GLOB_RECURSE fbgemm_genai_native_cuda_cu
      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/*.cu"
      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/**/*.cu")
@ -291,7 +291,6 @@ IF(USE_FBGEMM_GENAI)

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

--- a/aten/src/ATen/CPUGeneratorImpl.cpp
+++ b/aten/src/ATen/CPUGeneratorImpl.cpp
@ -181,7 +181,7 @@ c10::intrusive_ptr<c10::TensorImpl> CPUGeneratorImpl::get_state() const {
  static const size_t size = sizeof(CPUGeneratorImplState);
  static_assert(std::is_standard_layout_v<CPUGeneratorImplState>, "CPUGeneratorImplState is not a PODType");

-  auto state_tensor = at::detail::empty_cpu({static_cast<int64_t>(size)}, ScalarType::Byte, std::nullopt, std::nullopt, std::nullopt, std::nullopt);
+  auto state_tensor = at::detail::empty_cpu({(int64_t)size}, ScalarType::Byte, std::nullopt, std::nullopt, std::nullopt, std::nullopt);
  auto rng_state = state_tensor.data_ptr();

  // accumulate generator data to be copied into byte tensor
--- a/aten/src/ATen/Context.cpp
+++ b/aten/src/ATen/Context.cpp
@ -23,6 +23,8 @@ C10_DIAGNOSTIC_POP()
 #endif
 namespace at {

+namespace {
+
 /*
  These const variables defined the fp32 precisions for different backend
  We have "generic", "cuda", "mkldnn" backend now and we can choose fp32
@ -39,6 +41,16 @@ namespace at {
                ->rnn
 */

+  C10_ALWAYS_INLINE void warn_deprecated_fp32_precision_api(){
+    TORCH_WARN_ONCE(
+      "Please use the new API settings to control TF32 behavior, such as torch.backends.cudnn.conv.fp32_precision = 'tf32' "
+      "or torch.backends.cuda.matmul.fp32_precision = 'ieee'. Old settings, e.g, torch.backends.cuda.matmul.allow_tf32 = True, "
+      "torch.backends.cudnn.allow_tf32 = True, allowTF32CuDNN() and allowTF32CuBLAS() will be deprecated after Pytorch 2.9. Please see "
+      "https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices"
+    );
+  }
+} // namespace
+
 Float32Backend str2backend(const std::string& name) {
  if (name == "generic")
    return Float32Backend::GENERIC;
@ -194,6 +206,7 @@ bool Context::allowTF32CuDNN(std::optional<Float32Op> op) const {
  } else {
    return float32Precision(Float32Backend::CUDA, op.value()) == Float32Precision::TF32;
  }
+  warn_deprecated_fp32_precision_api();
  return allow_tf32_cudnn;
 }

@ -201,6 +214,7 @@ void Context::setAllowTF32CuDNN(bool b) {
  setFloat32Precision(Float32Backend::CUDA, Float32Op::RNN, b ? Float32Precision::TF32 : Float32Precision::NONE);
  setFloat32Precision(Float32Backend::CUDA, Float32Op::CONV, b ? Float32Precision::TF32 : Float32Precision::NONE);
  allow_tf32_cudnn = b;
+  warn_deprecated_fp32_precision_api();
 }

 void Context::setSDPPriorityOrder(const std::vector<int64_t>& order) {
@ -209,7 +223,7 @@ void Context::setSDPPriorityOrder(const std::vector<int64_t>& order) {
    "setSDPPriority order expected ", sdp_priority_order.size() - 1, " but got ",
    at::num_sdp_backends, " unique backends specified in priority order.");
  for (uint32_t i = 0; i < order.size(); i++) {
-    sdp_priority_order[i] = static_cast<at::SDPBackend>(order[i]);
+    sdp_priority_order[i] = (at::SDPBackend) order[i];
  }
 }

@ -311,6 +325,7 @@ bool Context::allowTF32CuBLAS() const {
      "Current status indicate that you have used mix of the legacy and new APIs to set the TF32 status for cublas matmul. ",
      "We suggest only using the new API to set the TF32 flag. See also: ",
      "https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices");
+  warn_deprecated_fp32_precision_api();
  return allow_tf32_new;
 }

@ -334,6 +349,7 @@ Float32MatmulPrecision Context::float32MatmulPrecision() const {
      "Current status indicate that you have used mix of the legacy and new APIs to set the matmul precision. ",
      "We suggest only using the new API for matmul precision. See also: ",
      "https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices");
+  warn_deprecated_fp32_precision_api();
  return float32_matmul_precision;
 }

@ -361,6 +377,7 @@ Float32Precision Context::float32Precision(Float32Backend backend, Float32Op op)

 void Context::setFloat32MatmulPrecision(const std::string &s) {
  auto match = [this](const std::string & s_) {
+    warn_deprecated_fp32_precision_api();
    // TODO: consider if CuDNN field needs to also be set for potential future CuDNN ops like multi-headed attention
    if (s_ == "highest") {
      float32_matmul_precision = at::Float32MatmulPrecision::HIGHEST;
@ -808,14 +825,6 @@ void Context::setDisplayVmapFallbackWarnings(bool enabled) {
  display_vmap_fallback_warnings_ = enabled;
 }

-bool Context::warnOnAccumulateGradStreamMismatch() const {
-  return warn_on_accumulate_grad_stream_mismatch_;
-}
-
-void Context::setWarnOnAccumulateGradStreamMismatch(bool enabled) {
-  warn_on_accumulate_grad_stream_mismatch_ = enabled;
-}
-
 bool Context::isDefaultMobileCPUAllocatorSet() {
  return prev_allocator_ptr_ != nullptr;
 }
--- a/aten/src/ATen/Context.h
+++ b/aten/src/ATen/Context.h
@ -404,9 +404,6 @@ class TORCH_API Context {
  void setDisplayVmapFallbackWarnings(bool enabled);
  bool areVmapFallbackWarningsEnabled() const;

-  void setWarnOnAccumulateGradStreamMismatch(bool enabled);
-  bool warnOnAccumulateGradStreamMismatch() const;
-
  bool isDefaultMobileCPUAllocatorSet();
  void setDefaultMobileCPUAllocator();
  void unsetDefaultMobileCPUAllocator();
@ -497,7 +494,6 @@ class TORCH_API Context {
  bool release_original_weights = false;
 #endif
  bool display_vmap_fallback_warnings_ = false;
-  bool warn_on_accumulate_grad_stream_mismatch_ = true;
  std::atomic<at::QEngine> quantized_engine = at::QEngine::NoQEngine;
  bool enable_sparse_tensor_invariant_checks = false;
  bool allow_fp16_reduction_cpu = false;
--- a/aten/src/ATen/Dispatch.h
+++ b/aten/src/ATen/Dispatch.h
@ -197,7 +197,6 @@ inline at::ScalarType scalar_type(at::ScalarType s) {
    /* don't use TYPE again in case it is an expensive or side-effect op */ \
    at::ScalarType _st = ::detail::scalar_type(the_type);                   \
    RECORD_KERNEL_FUNCTION_DTYPE(at_dispatch_name, _st);                    \
-    C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum")             \
    switch (_st) {                                                          \
      __VA_ARGS__                                                           \
      default:                                                              \
@ -209,7 +208,6 @@ inline at::ScalarType scalar_type(at::ScalarType s) {
            toString(_st),                                                  \
            "'");                                                           \
    }                                                                       \
-    C10_DIAGNOSTIC_POP()                                                    \
  }()

 #define AT_DISPATCH_CASE_FLOATING_TYPES(...)            \
--- a/aten/src/ATen/MapAllocator.cpp
+++ b/aten/src/ATen/MapAllocator.cpp
@ -252,13 +252,13 @@ MapAllocator::MapAllocator(WithFd /*unused*/, std::string_view filename, int fd,
    if (!(flags_ & ALLOCATOR_MAPPED_FROMFD)) {
      if (flags_ & ALLOCATOR_MAPPED_SHARED) {
        // NOLINTNEXTLINE(bugprone-assignment-in-if-condition)
-        if ((fd = open(filename_.c_str(), flags, static_cast<mode_t>(0600))) == -1) {
+        if ((fd = open(filename_.c_str(), flags, (mode_t)0600)) == -1) {
          TORCH_CHECK(false, "unable to open file <", filename_, "> in read-write mode: ", c10::utils::str_error(errno), " (", errno, ")");
        }
      } else if (flags_ & ALLOCATOR_MAPPED_SHAREDMEM) {
 #ifdef HAVE_SHM_OPEN
        // NOLINTNEXTLINE(bugprone-assignment-in-if-condition)
-        if((fd = shm_open(filename_.c_str(), flags, static_cast<mode_t>(0600))) == -1) {
+        if((fd = shm_open(filename_.c_str(), flags, (mode_t)0600)) == -1) {
          TORCH_CHECK(false, "unable to open shared memory object <", filename_, "> in read-write mode: ", c10::utils::str_error(errno), " (", errno, ")");
        }
 #else
@ -503,7 +503,7 @@ RefcountedMapAllocator::RefcountedMapAllocator(WithFd /*unused*/, const char *fi

 void RefcountedMapAllocator::initializeAlloc() {
  TORCH_CHECK(base_ptr_, "base_ptr_ is null");
-  MapInfo *map_info = static_cast<MapInfo*>(base_ptr_);
+  MapInfo *map_info = (MapInfo*)base_ptr_;

 #ifdef _WIN32
  ReleaseContext* r_ctx = new ReleaseContext;
@ -539,7 +539,7 @@ void RefcountedMapAllocator::close() {
  }
 #else /* _WIN32 */

-  MapInfo *info = static_cast<MapInfo*>(data);
+  MapInfo *info = (MapInfo*)(data);
  if (--info->refcount == 0) {
 #ifdef HAVE_SHM_UNLINK
    if (shm_unlink(filename_.c_str()) == -1) {
--- a/aten/src/ATen/TensorIterator.cpp
+++ b/aten/src/ATen/TensorIterator.cpp
@ -862,7 +862,7 @@ void TensorIteratorBase::narrow(int dim, int64_t start, int64_t size) {
  shape_[dim] = size;
  view_offsets_[dim] += start;
  for (auto& op : operands_) {
-    op.data = (static_cast<char*>(op.data)) + op.stride_bytes[dim] * start;
+    op.data = ((char*)op.data) + op.stride_bytes[dim] * start;
  }
  if (size == 1 && !is_reduction_) {
    coalesce_dimensions();
@ -873,7 +873,7 @@ void TensorIteratorBase::select_all_keeping_dim(int start_dim, IntArrayRef indic
  TORCH_INTERNAL_ASSERT(start_dim <= ndim());
  for (const auto i : c10::irange(start_dim, ndim())) {
    for (auto& op : operands_) {
-      op.data = (static_cast<char*>(op.data)) + op.stride_bytes[i] * indices[i - start_dim];
+      op.data = ((char*)op.data) + op.stride_bytes[i] * indices[i - start_dim];
    }
    shape_[i] = 1;
  }
--- a/aten/src/ATen/TensorIteratorInternal.h
+++ b/aten/src/ATen/TensorIteratorInternal.h
@ -41,7 +41,7 @@ inline void serial_for_each(
    IntArrayRef strides,
    char** base_ptrs,
    size_t ntensors,
-    TensorIteratorBase::loop2d_t loop,
+    typename TensorIteratorBase::loop2d_t loop,
    Range range) {
  const auto ndim = shape.size();
  TORCH_INTERNAL_ASSERT_DEBUG_ONLY(
--- a/aten/src/ATen/VmapModeRegistrations.cpp
+++ b/aten/src/ATen/VmapModeRegistrations.cpp
@ -72,16 +72,10 @@ TORCH_LIBRARY_IMPL(aten, VmapMode, m) {
  m.impl("random_", unsupportedRandomOp_<Tensor&, std::optional<Generator>>);

  m.impl("rand_like", unsupportedRandomOp<const Tensor&, TENSOROPTIONS, std::optional<MemoryFormat>>);
-  m.impl("rand_like.generator", unsupportedRandomOp<const Tensor&, std::optional<Generator>, TENSOROPTIONS, std::optional<MemoryFormat>>);
  m.impl("randn_like", unsupportedRandomOp<const Tensor&, TENSOROPTIONS, std::optional<MemoryFormat>>);
-  m.impl("randn_like.generator", unsupportedRandomOp<const Tensor&, std::optional<Generator>, TENSOROPTIONS, std::optional<MemoryFormat>>);

  m.impl("randint_like", unsupportedRandomOp<const Tensor&, int64_t, TENSOROPTIONS, std::optional<MemoryFormat>>);
-  m.impl("randint_like.Tensor", unsupportedRandomOp<const Tensor&, const Tensor&, TENSOROPTIONS, std::optional<MemoryFormat>>);
  m.impl("randint_like.low_dtype", unsupportedRandomOp<const Tensor&, int64_t, int64_t, TENSOROPTIONS, std::optional<MemoryFormat>>);
-  m.impl("randint_like.generator", unsupportedRandomOp<const Tensor&, int64_t, std::optional<Generator>, TENSOROPTIONS, std::optional<MemoryFormat>>);
-  m.impl("randint_like.Tensor_generator", unsupportedRandomOp<const Tensor&, const Tensor&, std::optional<Generator>, TENSOROPTIONS, std::optional<MemoryFormat>>);
-  m.impl("randint_like.low_generator_dtype", unsupportedRandomOp<const Tensor&, int64_t, int64_t, std::optional<Generator>, TENSOROPTIONS, std::optional<MemoryFormat>>);

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

  alignas(hardware_destructive_interference_size) std::mutex events_mutex_;
  std::deque<std::pair<E, B*>> events_; // event queue paired with block
--- a/aten/src/ATen/core/IListRef.h
+++ b/aten/src/ATen/core/IListRef.h
@ -190,14 +190,12 @@ class IListRef;
 * it to a function (e.g. `ImplT::<dispatch-function>(this_)`).
 */
 #define TORCH_ILISTREF_UNWRAP(TAG, BODY)                         \
-  C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum")    \
  switch (TAG) {                                                 \
    TORCH_ILISTREF_FORALL_TAGS(TORCH_ILISTREF_UNWRAP_CASE, BODY) \
    break;                                                       \
    default:                                                     \
      TORCH_INTERNAL_ASSERT(false, "invalid IListRef tag.");     \
-  } \
-  C10_DIAGNOSTIC_POP()
+  }

 enum class IListRefTag {
 #define DEFINE_TAG(tag, ...) tag,
--- a/aten/src/ATen/core/TransformationHelper.h
+++ b/aten/src/ATen/core/TransformationHelper.h
@ -56,7 +56,7 @@ C10_HOST_DEVICE inline T uniform_int_full_range(V val) {
 * in this overloaded version
 */
 template <typename T, typename V>
-C10_HOST_DEVICE inline std::enable_if_t<!std::is_floating_point_v<T>, T>uniform_int(V val) {
+C10_HOST_DEVICE inline std::enable_if_t<!(std::is_floating_point_v<T>), T>uniform_int(V val) {
  if constexpr (std::is_same_v<T, bool>) {
    return static_cast<bool>(val & 1);
  } else if constexpr (std::is_same_v<T, int64_t>) {
--- a/aten/src/ATen/core/boxing/KernelFunction_impl.h
+++ b/aten/src/ATen/core/boxing/KernelFunction_impl.h
@ -114,25 +114,25 @@ inline typename remove_symint<T>::type unpackSymInt(T x) {
 }

 template <>
-inline remove_symint<c10::SymInt>::type unpackSymInt(c10::SymInt x) {
+inline typename remove_symint<c10::SymInt>::type unpackSymInt(c10::SymInt x) {
  return x.guard_int(__FILE__, __LINE__);
 }

 template <>
-inline remove_symint<c10::SymIntArrayRef>::type unpackSymInt(
+inline typename remove_symint<c10::SymIntArrayRef>::type unpackSymInt(
    c10::SymIntArrayRef x) {
  return C10_AS_INTARRAYREF_SLOW(x);
 }

 template <>
-inline remove_symint<std::optional<c10::SymInt>>::type unpackSymInt(
+inline typename remove_symint<std::optional<c10::SymInt>>::type unpackSymInt(
    std::optional<c10::SymInt> x) {
  return x.has_value() ? std::make_optional(x->guard_int(__FILE__, __LINE__))
                       : std::nullopt;
 }

 template <>
-inline remove_symint<at::OptionalSymIntArrayRef>::type unpackSymInt(
+inline typename remove_symint<at::OptionalSymIntArrayRef>::type unpackSymInt(
    at::OptionalSymIntArrayRef x) {
  return x.has_value() ? std::make_optional(C10_AS_INTARRAYREF_SLOW(*x))
                       : std::nullopt;
--- a/aten/src/ATen/core/boxing/impl/make_boxed_from_unboxed_functor.h
+++ b/aten/src/ATen/core/boxing/impl/make_boxed_from_unboxed_functor.h
@ -631,8 +631,8 @@ call_functor_with_args_from_stack_(
    Stack* stack,
    std::index_sequence<ivalue_arg_indices...> /*unused*/,
    guts::typelist::typelist<ArgTypes...>* /*unused*/) {
-  (void)stack; // when sizeof...(ivalue_arg_indices) == 0, this argument would
-               // be unused and we have to silence the compiler warning.
+  (void)(stack); // when sizeof...(ivalue_arg_indices) == 0, this argument would
+                 // be unused and we have to silence the compiler warning.

  // We're explicitly filtering out DispatchKeySet from the argument list.
  // Some kernels take a DispatchKeySet as their first argument in order to
--- a/aten/src/ATen/core/enum_type.h
+++ b/aten/src/ATen/core/enum_type.h
@ -18,7 +18,6 @@ struct TORCH_API EnumType : public NamedType {
      TypePtr value,
      std::vector<EnumNameValue> enum_names_values,
      std::weak_ptr<::torch::jit::CompilationUnit> cu) {
-    C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum")
    switch (value->kind()) {
      case TypeKind::IntType:
      case TypeKind::FloatType:
@ -35,7 +34,6 @@ struct TORCH_API EnumType : public NamedType {
            value->str(),
            "', only int, float and string are supported");
    }
-    C10_DIAGNOSTIC_POP()
  }

  std::string str() const override {
--- a/aten/src/ATen/core/ivalue.cpp
+++ b/aten/src/ATen/core/ivalue.cpp
@ -601,8 +601,8 @@ std::ostream& IValue::repr(
      double d = v.toDouble();
      int c = std::fpclassify(d);
      if ((c == FP_NORMAL || c == FP_ZERO ) && std::abs(d) < 1e10) {
-        int64_t i = static_cast<int64_t>(d);
-        if (static_cast<double>(i) == d) {
+        int64_t i = int64_t(d);
+        if (double(i) == d) {
          // -0.0 (signed zero) needs to be parsed as -0.
          if (i == 0 && std::signbit(d)) {
            return out << "-" << i << ".";
@ -799,8 +799,8 @@ std::ostream& operator<<(std::ostream & out, const IValue & v) {
      double d = v.toDouble();
      int c = std::fpclassify(d);
      if (c == FP_NORMAL || c == FP_ZERO) {
-        int64_t i = static_cast<int64_t>(d);
-        if (static_cast<double>(i) == d) {
+        int64_t i = int64_t(d);
+        if (double(i) == d) {
          return out << i << ".";
        }
      }
--- a/aten/src/ATen/core/jit_type.h
+++ b/aten/src/ATen/core/jit_type.h
@ -41,7 +41,7 @@ void standardizeVectorForUnion(std::vector<TypePtr>* to_flatten);
 inline bool is_contiguous_strides(
    const IntArrayRef sizes,
    const IntArrayRef strides) {
-  size_t n_dim = sizes.size();
+  int n_dim = static_cast<int>(sizes.size());
  if (n_dim == 0) {
    return true;
  }
@ -50,7 +50,7 @@ inline bool is_contiguous_strides(
    return false;
  }

-  for (int i = static_cast<int>(n_dim) - 2; i >= 0; i--) {
+  for (int i = n_dim - 2; i >= 0; i--) {
    if (strides[i] != strides[i + 1] * sizes[i + 1]) {
      return false;
    }
@ -922,7 +922,6 @@ struct TORCH_API DictType : public SharedType {
    if (auto dyn = key->castRaw<DynamicType>()) {
      kind = dyn->dynamicKind();
    }
-    C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum")
    switch (kind) {
      case TypeKind::AnyType:
      case TypeKind::IntType:
@ -939,7 +938,6 @@ struct TORCH_API DictType : public SharedType {
            key->str(),
            "', only int, float, complex, Tensor, device and string keys are supported");
    }
-    C10_DIAGNOSTIC_POP()
  }

  // aligned with the format in FunctionSchema
@ -2373,7 +2371,7 @@ private:
 };

 template<>
-inline detail::CastReturnType<NamedType>::type Type::cast<NamedType>() {
+inline typename detail::CastReturnType<NamedType>::type Type::cast<NamedType>() {
  if (kind() == TypeKind::TupleType || kind() == TypeKind::FunctionType ||
      kind() == TypeKind::ClassType || kind() == TypeKind::InterfaceType) {
    return std::static_pointer_cast<NamedType>(static_cast<NamedType *>(this)->shared_from_this());
@ -2382,7 +2380,7 @@ inline detail::CastReturnType<NamedType>::type Type::cast<NamedType>() {
 }

 template<>
-inline detail::CastConstReturnType<NamedType>::type Type::cast<NamedType>() const {
+inline typename detail::CastConstReturnType<NamedType>::type Type::cast<NamedType>() const {
  if (kind() == TypeKind::TupleType || kind() == TypeKind::FunctionType ||
      kind() == TypeKind::ClassType || kind() == TypeKind::InterfaceType) {
    return std::static_pointer_cast<const NamedType>(static_cast<const NamedType *>(this)->shared_from_this());
--- a/aten/src/ATen/cpu/vec/vec128/vec128_bfloat16_neon.h
+++ b/aten/src/ATen/cpu/vec/vec128/vec128_bfloat16_neon.h
@ -19,13 +19,6 @@ inline namespace CPU_CAPABILITY {
 #error "Big endian is not supported."
 #endif

-// GCC does not properly optimize bf16 operators
-#if defined(__ARM_FEATURE_BF16) && (__clang_major__ >= 19)
-#define BF16_ARITHMETIC_SUPPORTED() 1
-#else
-#define BF16_ARITHMETIC_SUPPORTED() 0
-#endif
-
 // Unlike the float16_t family of types, bfloat16_t is not available
 // when we're not targeting bfloat16 hardware support on some
 // platforms (but not Mac, so we have to be careful not to shadow the
@ -359,35 +352,18 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
        other, &Vectorized<float>::name);                        \
  }

+  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(abs)
  Vectorized frac() const;
  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(trunc)
  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(sqrt)

 #ifdef __ARM_FEATURE_BF16
-  // Flip sign bit
  Vectorized<c10::BFloat16> neg() const {
-    return vreinterpretq_bf16_s16(vreinterpretq_s16_bf16(values) ^ (-32768));
+    return -values;
  }
-  // Fast reciprocal is fine because we are truncating results
  Vectorized<c10::BFloat16> reciprocal() const {
-    auto x = vcvtq_low_f32_bf16(values);
-    auto y = vcvtq_high_f32_bf16(values);
-    x = vrecpeq_f32(x);
-    y = vrecpeq_f32(y);
-    return vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(x), y);
+    return 1.0f / values;
  }
-  // Clearing the sign bit
-  Vectorized<c10::BFloat16> abs() const {
-    return vreinterpretq_bf16_u16(vreinterpretq_u16_bf16(values) & 0x7FFF);
-  }
-#else
-  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(abs)
-  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(neg)
-  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(reciprocal)
-#endif
-
-// These functions are optimized on clang-21+
-#if BF16_ARITHMETIC_SUPPORTED() && (__clang_major__ >= 21)
  Vectorized<c10::BFloat16> operator==(
      const Vectorized<c10::BFloat16>& other) const {
    return values == other.values;
@ -418,6 +394,8 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
    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!=)
  DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator<)
@ -473,7 +451,7 @@ template <>
 Vectorized<c10::BFloat16> inline operator+(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
  bfloat16x8_t x = a;
  bfloat16x8_t y = b;
  return x + y;
@ -486,7 +464,7 @@ template <>
 Vectorized<c10::BFloat16> inline operator-(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
  bfloat16x8_t x = a;
  bfloat16x8_t y = b;
  return x - y;
@ -499,7 +477,7 @@ template <>
 Vectorized<c10::BFloat16> inline operator*(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
  bfloat16x8_t x = a;
  bfloat16x8_t y = b;
  return x * y;
@ -512,7 +490,7 @@ template <>
 Vectorized<c10::BFloat16> inline operator/(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
  bfloat16x8_t x = a;
  bfloat16x8_t y = b;
  return x / y;
@ -629,7 +607,7 @@ Vectorized<c10::BFloat16> inline fmadd(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b,
    const Vectorized<c10::BFloat16>& c) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
  bfloat16x8_t x = a;
  bfloat16x8_t y = b;
  bfloat16x8_t z = c;
@ -649,7 +627,7 @@ Vectorized<c10::BFloat16> inline fnmadd(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b,
    const Vectorized<c10::BFloat16>& c) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
  bfloat16x8_t x = a;
  bfloat16x8_t y = b;
  bfloat16x8_t z = c;
@ -665,7 +643,7 @@ Vectorized<c10::BFloat16> inline fmsub(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b,
    const Vectorized<c10::BFloat16>& c) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
  bfloat16x8_t x = a;
  bfloat16x8_t y = b;
  bfloat16x8_t z = c;
@ -681,7 +659,7 @@ Vectorized<c10::BFloat16> inline fnmsub(
    const Vectorized<c10::BFloat16>& a,
    const Vectorized<c10::BFloat16>& b,
    const Vectorized<c10::BFloat16>& c) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
  bfloat16x8_t x = a;
  bfloat16x8_t y = b;
  bfloat16x8_t z = c;
--- a/aten/src/ATen/cpu/vec/vec128/vec128_convert.h
+++ b/aten/src/ATen/cpu/vec/vec128/vec128_convert.h
@ -6,9 +6,9 @@ namespace at::vec {
 inline namespace CPU_CAPABILITY {
 #if (defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256))

-// Enable auto-vectorization for clang-17+
+// Enable auto-vectorization for GCC-13+ and clang-17+
 // GCC-12 has a bug: gcc.gnu.org/bugzilla/show_bug.cgi?id=117001
-#if defined(__clang__) && (__clang_major__ >= 17)
+#if __GNUC__ > 12 || (defined(__clang__) && (__clang_major__ >= 17))

 template <typename from_type, typename to_type>
 inline void convertImpl(
@ -21,46 +21,12 @@ 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)
@ -68,7 +34,6 @@ 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)
@ -76,7 +41,6 @@ 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)
@ -84,7 +48,6 @@ 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)
@ -92,7 +55,6 @@ 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)
@ -100,7 +62,6 @@ 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)
@ -108,7 +69,6 @@ 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)
@ -116,14 +76,6 @@ 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)                            \
@ -155,111 +107,25 @@ 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
-
-template <typename to_type>
-inline void convertFromBf16Impl(
-    const c10::BFloat16* __restrict src,
-    to_type* __restrict dst,
-    int64_t n) {
-  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++) {
-    uint32_t tmp = static_cast<uint32_t>(srcPtr[i]) << 16;
-    float tmpF;
-    __builtin_memcpy(&tmpF, &tmp, sizeof(float));
-    dst[i] = static_cast<to_type>(tmpF);
-  }
-}
-#define CONVERT_FROM_BF16_TEMPLATE(to_type)                                \
-  template <>                                                              \
-  inline void convert(const c10::BFloat16* src, to_type* dst, int64_t n) { \
-    return convertFromBf16Impl<to_type>(src, dst, n);                      \
-  }
-
-CONVERT_FROM_BF16_TEMPLATE(uint8_t)
-CONVERT_FROM_BF16_TEMPLATE(int8_t)
-CONVERT_FROM_BF16_TEMPLATE(int16_t)
-CONVERT_FROM_BF16_TEMPLATE(int32_t)
-CONVERT_FROM_BF16_TEMPLATE(int64_t)
-CONVERT_FROM_BF16_TEMPLATE(float)
-CONVERT_FROM_BF16_TEMPLATE(double)
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-CONVERT_FROM_BF16_TEMPLATE(float16_t)
+#ifdef __ARM_FEATURE_BF16
+CONVERT_TEMPLATE(bfloat16_t, uint8_t)
+CONVERT_TEMPLATE(bfloat16_t, int8_t)
+CONVERT_TEMPLATE(bfloat16_t, int16_t)
+CONVERT_TEMPLATE(bfloat16_t, int32_t)
+CONVERT_TEMPLATE(bfloat16_t, int64_t)
+CONVERT_TEMPLATE(bfloat16_t, bfloat16_t)
+CONVERT_TEMPLATE(bfloat16_t, float)
+CONVERT_TEMPLATE(bfloat16_t, double)
+CONVERT_TEMPLATE(uint8_t, bfloat16_t)
+CONVERT_TEMPLATE(int8_t, bfloat16_t)
+CONVERT_TEMPLATE(int16_t, bfloat16_t)
+CONVERT_TEMPLATE(int32_t, bfloat16_t)
+CONVERT_TEMPLATE(int64_t, bfloat16_t)
+CONVERT_TEMPLATE(float, bfloat16_t)
+CONVERT_TEMPLATE(double, bfloat16_t)
 #endif

-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

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

    return vfmaq_f32(scale, poly, scale);
  }
-  Vectorized<float> exp_u20() const {
-    return vexpq_f32_u20();
-  }
  Vectorized<float> fexp_u20() const {
    return exp_u20();
  }
@ -637,7 +634,7 @@ inline Vectorized<float> Vectorized<float>::erf() const {
  // - exp(- x * x)
  auto pow_2 = (*this) * (*this);
  auto neg_pow_2 = pow_2 ^ neg_zero_vec;
-  auto tmp4 = neg_pow_2.vexpq_f32_u20();
+  auto tmp4 = neg_pow_2.exp();
  auto tmp5 = tmp4 ^ neg_zero_vec;
  // erf(x) = sign(x) * (1 - r * t * exp(- x * x))
  auto tmp6 = t * tmp5;
--- a/aten/src/ATen/cpu/vec/vec256/vec256_qint.h
+++ b/aten/src/ATen/cpu/vec/vec256/vec256_qint.h
@ -514,7 +514,7 @@ struct Vectorized<c10::qint8> : public Vectorizedqi {

  using float_vec_return_type = std::array<Vectorized<float>, kFloatNumVecs>;
  using int_vec_return_type = std::array<Vectorized<c10::qint32>, kIntNumVecs>;
-  using value_type = c10::qint8::underlying;
+  using value_type = typename c10::qint8::underlying;

 public:
  using Vectorizedqi::Vectorizedqi;
@ -727,7 +727,7 @@ struct Vectorized<c10::quint8> : public Vectorizedqi {

  using float_vec_return_type = std::array<Vectorized<float>, kFloatNumVecs>;
  using int_vec_return_type = std::array<Vectorized<c10::qint32>, kIntNumVecs>;
-  using value_type = c10::quint8::underlying;
+  using value_type = typename c10::quint8::underlying;

 public:
  using Vectorizedqi::Vectorizedqi;
--- a/aten/src/ATen/cpu/vec/vec512/vec512_qint.h
+++ b/aten/src/ATen/cpu/vec/vec512/vec512_qint.h
@ -567,7 +567,7 @@ struct Vectorized<c10::qint8> : public Vectorizedqi {

  using float_vec_return_type = std::array<Vectorized<float>, 4>;
  using int_vec_return_type = std::array<Vectorized<c10::qint32>, 4>;
-  using value_type = c10::qint8::underlying;
+  using value_type = typename c10::qint8::underlying;

 public:
  using Vectorizedqi::Vectorizedqi;
@ -804,7 +804,7 @@ struct Vectorized<c10::quint8> : public Vectorizedqi {

  using float_vec_return_type = std::array<Vectorized<float>, 4>;
  using int_vec_return_type = std::array<Vectorized<c10::qint32>, 4>;
-  using value_type = c10::quint8::underlying;
+  using value_type = typename c10::quint8::underlying;

 public:
  using Vectorizedqi::Vectorizedqi;
--- a/aten/src/ATen/cpu/vec/vec_base.h
+++ b/aten/src/ATen/cpu/vec/vec_base.h
@ -672,7 +672,7 @@ struct Vectorized {
    return map(std::sqrt);
  }
  Vectorized<T> reciprocal() const {
-    return map([](T x) { return (T)1 / x; });
+    return map([](T x) { return (T)(1) / x; });
  }
  Vectorized<T> rsqrt() const {
    return map([](T x) { return (T)1 / std::sqrt(x); });
--- a/aten/src/ATen/cpu/vml.h
+++ b/aten/src/ATen/cpu/vml.h
@ -46,7 +46,7 @@ inline void vrsqrt(scalar_t* out, scalar_t* in, int64_t size) {
  parallel_for(0, size, 2048, [out, in](int64_t begin, int64_t end) {
    map(
        [](const Vectorized<scalar_t>& x) {
-          return Vectorized<scalar_t>((scalar_t)1) / x.sqrt();
+          return Vectorized<scalar_t>((scalar_t)(1)) / x.sqrt();
        },
        out + begin,
        in + begin,
--- a/aten/src/ATen/cuda/CUDAGeneratorImpl.cpp
+++ b/aten/src/ATen/cuda/CUDAGeneratorImpl.cpp
@ -194,8 +194,8 @@ void CUDAGeneratorState::unregister_graph(cuda::CUDAGraph* graph) {
 void CUDAGeneratorState::capture_prologue() {
  capturing_ = true;
  offset_intragraph_ = 0;
-  seed_extragraph_.fill_(static_cast<int64_t>(seed_));
-  offset_extragraph_.fill_(0);
+  seed_extragraph_.fill_(int64_t(seed_));
+  offset_extragraph_.fill_(int64_t(0));
 }

 /**
@ -216,8 +216,8 @@ void CUDAGeneratorState::replay_prologue(uint64_t wholegraph_increment) {
  at::cuda::assertNotCapturing(
      "Cannot prepare for replay during capturing stage.");
  if (wholegraph_increment) {
-      seed_extragraph_.fill_(static_cast<int64_t>(seed_));
-      offset_extragraph_.fill_(static_cast<int64_t>(philox_offset_per_thread_));
+      seed_extragraph_.fill_(int64_t(seed_));
+      offset_extragraph_.fill_(int64_t(philox_offset_per_thread_));
      // Applies the total increment achieved during previous captures to update the
      // offset.
      increase(wholegraph_increment);
@ -329,7 +329,7 @@ c10::intrusive_ptr<c10::TensorImpl> CUDAGeneratorImpl::get_state() const {
  constexpr size_t offset_size = sizeof(int64_t);
  constexpr size_t total_size = seed_size + offset_size;

-  auto state_tensor = at::detail::empty_cpu({static_cast<int64_t>(total_size)}, ScalarType::Byte, std::nullopt, std::nullopt, std::nullopt, std::nullopt);
+  auto state_tensor = at::detail::empty_cpu({(int64_t)total_size}, ScalarType::Byte, std::nullopt, std::nullopt, std::nullopt, std::nullopt);
  auto rng_state = state_tensor.data_ptr<uint8_t>();
  auto current_seed = this->current_seed();
  auto offset = static_cast<int64_t>(this->philox_offset_per_thread()); // Note that old THCGeneratorState had offset as std::atomic<int64_t>
--- a/aten/src/ATen/cuda/CUDAGreenContext.cpp
+++ b/aten/src/ATen/cuda/CUDAGreenContext.cpp
@ -1,90 +1,78 @@
 #include <ATen/cuda/CUDAGreenContext.h>

-#if defined(CUDA_VERSION) && (CUDA_VERSION >= 12030) && !defined(USE_ROCM) && defined(PYTORCH_C10_DRIVER_API_SUPPORTED)
-#include <c10/cuda/driver_api.h>
-#include <stdexcept>
-#include <vector>
-#define HAS_CUDA_GREEN_CONTEXT() 1
-#else
-#define HAS_CUDA_GREEN_CONTEXT() 0
-// Suppress unsued private field warnings as this class is not supposed to be called
-C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-private-field")
-#endif
-
 namespace at::cuda {
+  GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
+#if CUDA_HAS_GREEN_CONTEXT
+    int driver_version;
+    C10_CUDA_CHECK(cudaDriverGetVersion(&driver_version));
+    TORCH_CHECK(
+        driver_version >= 12080, "cuda driver too old to use green context!");
+    CUcontext pctx = nullptr;
+    C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuCtxGetCurrent_(&pctx));
+    if (C10_UNLIKELY(!pctx)) {
+      TORCH_WARN(
+          "Attempted to create a green context but"
+          " there was no primary context! Creating a primary context...");

-GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
-#if HAS_CUDA_GREEN_CONTEXT()
-  int driver_version;
-  C10_CUDA_CHECK(cudaDriverGetVersion(&driver_version));
-  TORCH_CHECK(
-      driver_version >= 12080, "cuda driver too old to use green context!");
-  CUcontext pctx = nullptr;
-  C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuCtxGetCurrent_(&pctx));
-  if (C10_UNLIKELY(!pctx)) {
-    TORCH_WARN(
-        "Attempted to create a green context but"
-        " there was no primary context! Creating a primary context...");
+      cudaFree(0);
+    }

-    cudaFree(0);
-  }
+    CUdevice device;
+    device_id_ = device_id;
+    C10_CUDA_DRIVER_CHECK(
+        c10::cuda::DriverAPI::get()->cuDeviceGet_(&device, device_id));

-   CUdevice device;
-  device_id_ = device_id;
-  C10_CUDA_DRIVER_CHECK(
-      c10::cuda::DriverAPI::get()->cuDeviceGet_(&device, device_id));
+    // Get device resources
+    CUdevResource device_resource;
+    C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuDeviceGetDevResource_(
+        device, &device_resource, CU_DEV_RESOURCE_TYPE_SM));

-  // Get device resources
-  CUdevResource device_resource;
-  C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuDeviceGetDevResource_(
-      device, &device_resource, CU_DEV_RESOURCE_TYPE_SM));
+    // Split resources
+    std::vector<CUdevResource> result(1);
+    auto result_data = result.data();
+    unsigned int nb_groups = 1;
+    CUdevResource remaining;

-  // Split resources
-  std::vector<CUdevResource> result(1);
-  auto result_data = result.data();
-  unsigned int nb_groups = 1;
-  CUdevResource remaining;
+    C10_CUDA_DRIVER_CHECK(
+        c10::cuda::DriverAPI::get()->cuDevSmResourceSplitByCount_(
+            result_data,
+            &nb_groups,
+            &device_resource,
+            &remaining,
+            0, // default flags
+            num_sms));

-  C10_CUDA_DRIVER_CHECK(
-      c10::cuda::DriverAPI::get()->cuDevSmResourceSplitByCount_(
-          result_data,
-          &nb_groups,
-          &device_resource,
-          &remaining,
-          0, // default flags
-          num_sms));
+    TORCH_CHECK(nb_groups == 1, "Failed to create single resource group");

-  TORCH_CHECK(nb_groups == 1, "Failed to create single resource group");
+    // Generate resource descriptor
+    CUdevResourceDesc desc;
+    C10_CUDA_DRIVER_CHECK(
+        c10::cuda::DriverAPI::get()->cuDevResourceGenerateDesc_(
+            &desc, result_data, 1));

-  // Generate resource descriptor
-  CUdevResourceDesc desc;
-  C10_CUDA_DRIVER_CHECK(
-      c10::cuda::DriverAPI::get()->cuDevResourceGenerateDesc_(
-          &desc, result_data, 1));
+    // Create green context
+    // CU_GREEN_CTX_DEFAULT_STREAM is required per docs:
+    // https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__GREEN__CONTEXTS.html
+    C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuGreenCtxCreate_(
+        &green_ctx_, desc, device, CU_GREEN_CTX_DEFAULT_STREAM));

-  // Create green context
-  // CU_GREEN_CTX_DEFAULT_STREAM is required per docs:
-  // https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__GREEN__CONTEXTS.html
-  C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuGreenCtxCreate_(
-      &green_ctx_, desc, device, CU_GREEN_CTX_DEFAULT_STREAM));
-
-  // Convert to regular context
-  C10_CUDA_DRIVER_CHECK(
-      c10::cuda::DriverAPI::get()->cuCtxFromGreenCtx_(&context_, green_ctx_));
-  TORCH_CHECK(context_, "Green ctx conversion to regular ctx failed!");
+    // Convert to regular context
+    C10_CUDA_DRIVER_CHECK(
+        c10::cuda::DriverAPI::get()->cuCtxFromGreenCtx_(&context_, green_ctx_));
+    TORCH_CHECK(context_, "Green ctx conversion to regular ctx failed!");
 #else
-  TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
+    TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
 #endif
  }

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

  // Implement move operations
  GreenContext::GreenContext(GreenContext&& other) noexcept{
-#if HAS_CUDA_GREEN_CONTEXT()
+#if CUDA_HAS_GREEN_CONTEXT
    device_id_ = std::exchange(other.device_id_, -1);
    green_ctx_ = std::exchange(other.green_ctx_, nullptr);
    context_ = std::exchange(other.context_, nullptr);
@ -103,7 +91,7 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
  }

  GreenContext& GreenContext::operator=(GreenContext&& other) noexcept{
-#if HAS_CUDA_GREEN_CONTEXT()
+#if CUDA_HAS_GREEN_CONTEXT
    if (this != &other) {
      // Clean up current resources
      if (green_ctx_) {
@ -132,7 +120,7 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
  }

  GreenContext::~GreenContext() noexcept{
-#if HAS_CUDA_GREEN_CONTEXT()
+#if CUDA_HAS_GREEN_CONTEXT
    C10_CUDA_DRIVER_CHECK(
        c10::cuda::DriverAPI::get()->cuGreenCtxDestroy_(green_ctx_));
 #else
@ -140,9 +128,25 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
 #endif
  }

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

@ -171,7 +175,7 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
  }

  void GreenContext::popContext() {
-#if HAS_CUDA_GREEN_CONTEXT()
+#if CUDA_HAS_GREEN_CONTEXT
    // see above note about stream being hardcoded to the default stream
    at::cuda::CUDAEvent ev;
    ev.record(c10::cuda::getCurrentCUDAStream());
--- a/aten/src/ATen/cuda/CUDAGreenContext.h
+++ b/aten/src/ATen/cuda/CUDAGreenContext.h
@ -1,38 +1,53 @@
 #pragma once
 #include <ATen/cuda/CUDAEvent.h>
-#include <cuda.h>

-// Forward declare green context as opaque ptr
-typedef struct CUgreenCtx_st* CUgreenCtx;
+#if defined(CUDA_VERSION) && !defined(USE_ROCM) && defined(PYTORCH_C10_DRIVER_API_SUPPORTED)
+#include <c10/cuda/driver_api.h>
+#include <cuda.h>
+#include <memory>
+#include <stdexcept>
+#include <vector>
+#define CUDA_HAS_GREEN_CONTEXT 1
+#else
+#define CUDA_HAS_GREEN_CONTEXT 0
+#endif

 namespace at::cuda {

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

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

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

  void popContext();

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


+#if defined(USE_ROCM)
+// hipSparse const API added in v2.4.0
+#if HIPSPARSE_VERSION >= 200400
+#define AT_USE_HIPSPARSE_GENERIC_API() 1
+#else
+#define AT_USE_HIPSPARSE_GENERIC_API() 1
+#endif
+#else // USE_ROCM
+#define AT_USE_HIPSPARSE_GENERIC_API() 0
+#endif // USE_ROCM
+
 // cuSparse Generic API spsv function was added in CUDA 11.3.0
 #if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && (CUSPARSE_VERSION >= 11500)
 #define AT_USE_CUSPARSE_GENERIC_SPSV() 1
--- a/aten/src/ATen/cuda/CublasHandlePool.cpp
+++ b/aten/src/ATen/cuda/CublasHandlePool.cpp
@ -155,8 +155,8 @@ size_t parseChosenWorkspaceSize() {
    while (next != end) {
      std::smatch match = *next;
      TORCH_CHECK(match.size() == 3, "Expected CUBLAS_WORKSPACE_SPACE_CONFIG match of size 3 (Format :SIZE:COUNT)");
-      size_t curr_size = std::stoull(match.str(1));
-      size_t count = std::stoull(match.str(2));
+      size_t curr_size = (size_t) std::stoi(match.str(1));
+      size_t count = (size_t) std::stoi(match.str(2));
      total_size += curr_size * 1024 * count;
      next++;
    }
--- a/aten/src/ATen/cuda/NumericLimits.cuh
+++ b/aten/src/ATen/cuda/NumericLimits.cuh
@ -55,14 +55,6 @@ struct numeric_limits<int8_t> {
  static inline __host__ __device__ int8_t upper_bound() { return INT8_MAX; }
 };

-template <>
-struct numeric_limits<uint16_t> {
-  static inline __host__ __device__ uint16_t lowest() { return 0; }
-  static inline __host__ __device__ uint16_t max() { return UINT16_MAX; }
-  static inline __host__ __device__ uint16_t lower_bound() { return 0; }
-  static inline __host__ __device__ uint16_t upper_bound() { return UINT16_MAX; }
-};
-
 template <>
 struct numeric_limits<int16_t> {
  static inline __host__ __device__ int16_t lowest() { return INT16_MIN; }
@ -71,14 +63,6 @@ struct numeric_limits<int16_t> {
  static inline __host__ __device__ int16_t upper_bound() { return INT16_MAX; }
 };

-template <>
-struct numeric_limits<uint32_t> {
-  static inline __host__ __device__ uint32_t lowest() { return 0; }
-  static inline __host__ __device__ uint32_t max() { return UINT32_MAX; }
-  static inline __host__ __device__ uint32_t lower_bound() { return 0; }
-  static inline __host__ __device__ uint32_t upper_bound() { return UINT32_MAX; }
-};
-
 template <>
 struct numeric_limits<int32_t> {
  static inline __host__ __device__ int32_t lowest() { return INT32_MIN; }
@ -87,21 +71,6 @@ struct numeric_limits<int32_t> {
  static inline __host__ __device__ int32_t upper_bound() { return INT32_MAX; }
 };

-template <>
-struct numeric_limits<uint64_t> {
-#ifdef _MSC_VER
-  static inline __host__ __device__ uint64_t lowest() { return 0; }
-  static inline __host__ __device__ uint64_t max() { return _UI64_MAX; }
-  static inline __host__ __device__ uint64_t lower_bound() { return 0; }
-  static inline __host__ __device__ uint64_t upper_bound() { return _UI64_MAX; }
-#else
-  static inline __host__ __device__ uint64_t lowest() { return 0; }
-  static inline __host__ __device__ uint64_t max() { return UINT64_MAX; }
-  static inline __host__ __device__ uint64_t lower_bound() { return 0; }
-  static inline __host__ __device__ uint64_t upper_bound() { return UINT64_MAX; }
-#endif
-};
-
 template <>
 struct numeric_limits<int64_t> {
 #ifdef _MSC_VER
--- a/aten/src/ATen/cuda/Sleep.cu
+++ b/aten/src/ATen/cuda/Sleep.cu
@ -1,7 +1,6 @@
 #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>

@ -25,22 +24,8 @@ __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);
@ -48,26 +33,6 @@ void sleep(int64_t cycles) {
  C10_CUDA_KERNEL_LAUNCH_CHECK();
 }

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

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

--- a/aten/src/ATen/cuda/detail/BLASConstants.cu
+++ b/aten/src/ATen/cuda/detail/BLASConstants.cu
@ -2,6 +2,8 @@
 #include <ATen/Tensor.h>
 #include <ATen/cuda/Exceptions.h>

+#include <mutex>
+
 namespace at {
 namespace cuda {
 namespace detail {
@ -10,36 +12,39 @@ __device__ __constant__ float cublas_one_device;
 __device__ __constant__ float cublas_zero_device;

 float *get_cublas_device_one() {
-  static float *ptr = nullptr;
-  static auto init_flag = [&]() {
+  static c10::once_flag init_flag;
+
+  c10::call_once(init_flag, []() {
    const float one = 1.f;
    AT_CUDA_CHECK(cudaMemcpyToSymbol(cublas_one_device, &one, sizeof(float)));
-    AT_CUDA_CHECK(cudaGetSymbolAddress(reinterpret_cast<void**>(&ptr), cublas_one_device));
-    return true;
-  }();
+  });

+  float *ptr;
+  AT_CUDA_CHECK(cudaGetSymbolAddress(reinterpret_cast<void**>(&ptr), cublas_one_device));
  return ptr;
 }

 float *get_cublas_device_zero() {
-  static float *ptr = nullptr;
-  static auto init_flag = [&]() {
+  static c10::once_flag init_flag;
+
+  c10::call_once(init_flag, []() {
    const float zero = 0.f;
    AT_CUDA_CHECK(cudaMemcpyToSymbol(cublas_zero_device, &zero, sizeof(float)));
-    AT_CUDA_CHECK(cudaGetSymbolAddress(reinterpret_cast<void**>(&ptr), cublas_zero_device));
-    return true;
-  }();
+  });

+  float *ptr;
+  AT_CUDA_CHECK(cudaGetSymbolAddress(reinterpret_cast<void**>(&ptr), cublas_zero_device));
  return ptr;
 }

 float *get_user_alpha_ptr() {
  static float *alpha_ptr;

-  static bool init_flag [[maybe_unused]] = []() {
+  static c10::once_flag init_flag;
+
+  c10::call_once(init_flag, []() {
    AT_CUDA_CHECK(cudaMalloc(&alpha_ptr, sizeof(float)));
-    return true;
-  }();
+  });

  return alpha_ptr;
 }
--- a/aten/src/ATen/cuda/tunable/Tunable.cpp
+++ b/aten/src/ATen/cuda/tunable/Tunable.cpp
@ -580,7 +580,7 @@ std::ofstream& TuningContext::GetUntunedFile(){
      filename.append(device);
    }

-    untuned_file_ = std::ofstream(filename, std::ios::out | std::ios::app);
+    untuned_file_ = std::ofstream(filename, std::ios::out | std::ios::trunc);
  }
  return untuned_file_;
 }
--- a/aten/src/ATen/cudnn/Descriptors.cpp
+++ b/aten/src/ATen/cudnn/Descriptors.cpp
@ -3,7 +3,6 @@
 #include <ATen/ATen.h>
 #include <c10/util/irange.h>

-#include <array>
 #include <iostream>
 #include <sstream>

@ -137,9 +136,9 @@ void FilterDescriptor::set(const at::Tensor &t, const at::MemoryFormat memory_fo
    "Weight strides: ", t.strides(), "\n",
    "cuDNN suggested memory_format: ", memory_format);

-  std::array<int, CUDNN_DIM_MAX> size;
+  int size[CUDNN_DIM_MAX];
  for (const auto i : c10::irange(dim)) {
-    size[i] = static_cast<int>(t.size(i));
+    size[i] = (int) t.size(i);
  }
  for (const auto i : c10::irange(dim, pad)) {
    size[i] = 1;
@ -157,7 +156,7 @@ void FilterDescriptor::set(const at::Tensor &t, const at::MemoryFormat memory_fo
    default:
      TORCH_INTERNAL_ASSERT(false, "unsupported memory_format for cuDNN filters");
  }
-  set(getDataType(t), static_cast<int>(dim), size.data(), filter_format);
+  set(getDataType(t), static_cast<int>(dim), size, filter_format);
 }

 std::string cudnnMemoryFormatToString(cudnnTensorFormat_t tformat) {
--- a/aten/src/ATen/detail/MTIAHooksInterface.h
+++ b/aten/src/ATen/detail/MTIAHooksInterface.h
@ -1,6 +1,5 @@
 #pragma once

-#include <c10/core/CachingDeviceAllocator.h>
 #include <c10/core/Device.h>
 #include <c10/util/Exception.h>

@ -9,8 +8,8 @@

 #include <c10/core/Allocator.h>

-#include <ATen/detail/AcceleratorHooksInterface.h>
 #include <c10/util/python_stub.h>
+#include <ATen/detail/AcceleratorHooksInterface.h>

 #include <string>
 namespace at {
@ -26,7 +25,8 @@ constexpr const char* MTIA_HELP =
 struct TORCH_API MTIAHooksInterface : AcceleratorHooksInterface {
 // this fails the implementation if MTIAHooks functions are called, but
 // MTIA backend is not present.
-#define FAIL_MTIAHOOKS_FUNC(func) TORCH_CHECK(false, "Cannot execute ", func, "() without MTIA backend.");
+#define FAIL_MTIAHOOKS_FUNC(func) \
+  TORCH_CHECK(false, "Cannot execute ", func, "() without MTIA backend.");

  ~MTIAHooksInterface() override = default;

@ -91,7 +91,7 @@ struct TORCH_API MTIAHooksInterface : AcceleratorHooksInterface {
    return c10::Stream::unpack3(-1, 0, c10::DeviceType::MTIA);
  }

-  virtual void setCurrentStream(const c10::Stream& /*stream*/) const {
+  virtual void setCurrentStream(const c10::Stream& /*stream*/ ) const {
    FAIL_MTIAHOOKS_FUNC(__func__);
  }

@ -123,9 +123,11 @@ struct TORCH_API MTIAHooksInterface : AcceleratorHooksInterface {
    FAIL_MTIAHOOKS_FUNC(__func__);
  }

-  virtual void recordMemoryHistory(const std::optional<std::string>& /*enabled*/,
-                                   const std::string& /*stacks*/,
-                                   size_t /*max_entries*/) const {
+
+  virtual void recordMemoryHistory(
+    const std::optional<std::string>& /*enabled*/,
+    const std::string& /*stacks*/,
+    size_t /*max_entries*/) const {
    FAIL_MTIAHOOKS_FUNC(__func__);
  }

@ -149,46 +151,13 @@ struct TORCH_API MTIAHooksInterface : AcceleratorHooksInterface {
  }

  virtual bool isAvailable() const override;
-
-  /* MTIAGraph related APIs */
-  virtual int64_t mtiagraphCreate(bool keep_graph = false) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-    return -1;
-  }
-
-  virtual void mtiagraphDestroy(int64_t handle) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-  }
-
-  virtual void mtiagraphCaptureBegin(int64_t handle, MempoolId_t pool) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-  }
-
-  virtual void mtiagraphCaptureEnd(int64_t handle) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-  }
-
-  virtual void mtiagraphInstantiate(int64_t handle) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-  }
-
-  virtual void mtiagraphReplay(int64_t handle) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-  }
-
-  virtual void mtiagraphReset(int64_t handle) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-  }
-
-  virtual MempoolId_t mtiagraphPool(int64_t handle) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-  }
 };

 struct TORCH_API MTIAHooksArgs {};

 TORCH_DECLARE_REGISTRY(MTIAHooksRegistry, MTIAHooksInterface, MTIAHooksArgs);
-#define REGISTER_MTIA_HOOKS(clsname) C10_REGISTER_CLASS(MTIAHooksRegistry, clsname, clsname)
+#define REGISTER_MTIA_HOOKS(clsname) \
+  C10_REGISTER_CLASS(MTIAHooksRegistry, clsname, clsname)

 namespace detail {
 TORCH_API const MTIAHooksInterface& getMTIAHooks();
--- a/Show More
+++ b/Show More
 @ -1 +1 @@
 .5.1
 .5.0