[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.
2025-11-05 00:14:54 +08:00 · 2025-10-28 11:13:24 -07:00
996 changed files with 10770 additions and 26612 deletions
--- 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/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
--- 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/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/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:
-	$(DOCKER_RUN)
-
 .PHONY: magma-rocm70
 magma-rocm70: DESIRED_ROCM := 7.0
 magma-rocm70:
--- a/.ci/magma-rocm/build_magma.sh
+++ b/.ci/magma-rocm/build_magma.sh
@ -6,8 +6,8 @@ set -eou pipefail
 # The script expects DESIRED_CUDA and PACKAGE_NAME to be set
 ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"

-# 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

 # Folders for the build
 PACKAGE_FILES=${ROOT_DIR}/magma-rocm/package_files # metadata
@ -20,7 +20,7 @@ mkdir -p ${PACKAGE_DIR} ${PACKAGE_OUTPUT}/linux-64 ${PACKAGE_BUILD} ${PACKAGE_RE

 # Fetch magma sources and verify checksum
 pushd ${PACKAGE_DIR}
-git clone https://github.com/icl-utk-edu/magma
+git clone https://github.com/jeffdaily/magma
 pushd magma
 git checkout ${MAGMA_VERSION}
 popd
--- a/.ci/pytorch/build.sh
+++ b/.ci/pytorch/build.sh
@ -426,7 +426,7 @@ fi
 if [[ "$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/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 "${modes[@]}"; do
@ -1761,7 +1767,7 @@ elif [[ "${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/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/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
@ -11,17 +11,11 @@ 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 = {
@ -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/_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 \
--- 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/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/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/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/CMakeLists.txt
+++ b/CMakeLists.txt
@ -374,7 +374,7 @@ cmake_dependent_option(
  "Build the lazy Torchscript backend, not compatible with mobile builds" ON
  "NOT INTERN_BUILD_MOBILE" OFF)
 cmake_dependent_option(BUILD_FUNCTORCH "Build Functorch" ON "BUILD_PYTHON" OFF)
-cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin 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/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).*")
+    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/Context.cpp
+++ b/aten/src/ATen/Context.cpp
@ -825,14 +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/MapAllocator.cpp
+++ b/aten/src/ATen/MapAllocator.cpp
@ -539,7 +539,7 @@ void RefcountedMapAllocator::close() {
  }
 #else /* _WIN32 */

-  MapInfo *info = (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/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/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/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/jit_type.h
+++ b/aten/src/ATen/core/jit_type.h
@ -2371,7 +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());
@ -2380,7 +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)                            \
@ -155,41 +107,6 @@ CONVERT_TO_FP16_TEMPLATE(int32_t)
 CONVERT_TO_FP16_TEMPLATE(int64_t)
 CONVERT_TO_FP16_TEMPLATE(float)
 CONVERT_TO_FP16_TEMPLATE(double)
-
-inline void convertBoolToFp16Impl(
-    const bool* __restrict src,
-    at::Half* __restrict dst,
-    int64_t n) {
-  const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
-  float16_t* dstPtr = reinterpret_cast<float16_t*>(dst);
-  uint64_t len = static_cast<uint64_t>(n);
-  for (uint64_t i = 0; i < len; i++) {
-    dstPtr[i] = srcPtr[i] != 0 ? 1.0 : 0;
-  }
-}
-
-template <>
-inline void convert(const bool* src, at::Half* dst, int64_t n) {
-  return convertBoolToFp16Impl(src, dst, n);
-}
-
-inline void convertFp16ToBoolImpl(
-    const at::Half* __restrict src,
-    bool* __restrict dst,
-    int64_t n) {
-  const float16_t* srcPtr = reinterpret_cast<const float16_t*>(src);
-  uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
-  uint64_t len = static_cast<uint64_t>(n);
-  for (uint64_t i = 0; i < len; i++) {
-    dstPtr[i] = srcPtr[i] != 0.0 ? 1 : 0;
-  }
-}
-
-template <>
-inline void convert(const at::Half* src, bool* dst, int64_t n) {
-  return convertFp16ToBoolImpl(src, dst, n);
-}
-
 #endif
 #ifdef __ARM_FEATURE_BF16
 CONVERT_TEMPLATE(bfloat16_t, uint8_t)
@ -207,44 +124,6 @@ CONVERT_TEMPLATE(int32_t, bfloat16_t)
 CONVERT_TEMPLATE(int64_t, bfloat16_t)
 CONVERT_TEMPLATE(float, bfloat16_t)
 CONVERT_TEMPLATE(double, bfloat16_t)
-
-inline void convertBoolToBfloat16Impl(
-    const bool* __restrict src,
-    c10::BFloat16* __restrict dst,
-    int64_t n) {
-  const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
-  uint16_t* dstPtr = reinterpret_cast<uint16_t*>(dst);
-  uint64_t len = static_cast<uint64_t>(n);
-  constexpr uint16_t kBf16One = 0x3f80; // 1.0 in bfloat16
-  for (uint64_t i = 0; i < len; i++) {
-    dstPtr[i] = srcPtr[i] != 0 ? kBf16One : 0;
-  }
-}
-
-template <>
-inline void convert(const bool* src, c10::BFloat16* dst, int64_t n) {
-  return convertBoolToBfloat16Impl(src, dst, n);
-}
-
-inline void convertBfloat16ToBoolImpl(
-    const c10::BFloat16* __restrict src,
-    bool* __restrict dst,
-    int64_t n) {
-  uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
-  const uint16_t* srcPtr = reinterpret_cast<const uint16_t*>(src);
-  uint64_t len = static_cast<uint64_t>(n);
-  for (uint64_t i = 0; i < len; i++) {
-    // Check if all non-sign bits are 0
-    bool isBf16Zero = (srcPtr[i] & 0x7fff) == 0;
-    dstPtr[i] = isBf16Zero ? 0 : 1;
-  }
-}
-
-template <>
-inline void convert(const c10::BFloat16* src, bool* dst, int64_t n) {
-  return convertBfloat16ToBoolImpl(src, dst, n);
-}
-
 #endif

 #endif
--- a/aten/src/ATen/cpu/vec/vec128/vec128_float_neon.h
+++ b/aten/src/ATen/cpu/vec/vec128/vec128_float_neon.h
@ -309,7 +309,7 @@ class Vectorized<float> {
  DEFINE_SLEEF_COMPATIBLE_UNARY_ELEMENTWISE_FUNC(expm1)
  // Implementation copied from Arm Optimized Routine
  // https://github.com/ARM-software/optimized-routines/blob/master/math/aarch64/advsimd/expf.c
-  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/CUDAGreenContext.cpp
+++ b/aten/src/ATen/cuda/CUDAGreenContext.cpp
@ -1,90 +1,78 @@
 #include <ATen/cuda/CUDAGreenContext.h>

-#if defined(CUDA_VERSION) && !defined(USE_ROCM) && defined(PYTORCH_C10_DRIVER_API_SUPPORTED)
-#include <c10/cuda/driver_api.h>
-#include <stdexcept>
-#include <vector>
-#define HAS_CUDA_GREEN_CONTEXT() 1
-#else
-#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/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/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>

@ -152,36 +151,6 @@ 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 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 {};
--- a/aten/src/ATen/functorch/BatchRulesViews.cpp
+++ b/aten/src/ATen/functorch/BatchRulesViews.cpp
@ -534,20 +534,20 @@ Tensor trace_decomp(const Tensor& tensor) {
 std::tuple<Tensor, std::optional<int64_t>> tril_batch_rule(
    const Tensor& self,
    std::optional<int64_t> self_bdim,
-    c10::SymInt diagonal = 0) {
+    int64_t diagonal = 0) {
  TORCH_CHECK(self.dim() >= 2, "tril: The input tensor must have at least 2 dimensions.");
  auto self_ = moveBatchDimToFront(self, self_bdim);
-  auto result = at::tril_symint(self_, std::move(diagonal));
+  auto result = at::tril(self_, diagonal);
  return std::make_tuple(std::move(result), 0);
 }

 std::tuple<Tensor, std::optional<int64_t>> triu_batch_rule(
    const Tensor& self,
    std::optional<int64_t> self_bdim,
-    c10::SymInt diagonal = 0) {
+    int64_t diagonal = 0) {
  TORCH_CHECK(self.dim() >= 2, "triu: The input tensor must have at least 2 dimensions.");
  auto self_ = moveBatchDimToFront(self, self_bdim);
-  auto result = at::triu_symint(self_, std::move(diagonal));
+  auto result = at::triu(self_, diagonal);
  return std::make_tuple(std::move(result), 0);
 }

--- a/aten/src/ATen/mps/MPSDevice.mm
+++ b/aten/src/ATen/mps/MPSDevice.mm
@ -1,5 +1,7 @@
 //  Copyright © 2022 Apple Inc.

+#include <c10/util/CallOnce.h>
+
 #include <ATen/mps/IndexKernels.h>
 #include <ATen/mps/MPSAllocatorInterface.h>
 #include <ATen/mps/MPSDevice.h>
@ -8,6 +10,9 @@

 namespace at::mps {

+static std::unique_ptr<MPSDevice> mps_device;
+static c10::once_flag mpsdev_init;
+
 static inline MTLLanguageVersion getMetalLanguageVersion(const id<MTLDevice>& device) {
  // MPS Advanced Indexing needs at least Metal 2.0 (support for Argument Buffers and function constants)
  // host_name attribute needs at least Metal 2.2 and ulong needs Metal 2.3 (supported on MacOS 11+
@ -16,8 +21,8 @@ static inline MTLLanguageVersion getMetalLanguageVersion(const id<MTLDevice>& de
 }

 MPSDevice* MPSDevice::getInstance() {
-  static MPSDevice mps_device;
-  return &mps_device;
+  c10::call_once(mpsdev_init, [] { mps_device = std::unique_ptr<MPSDevice>(new MPSDevice()); });
+  return mps_device.get();
 }

 MPSDevice::~MPSDevice() {
--- a/aten/src/ATen/native/AveragePool2d.cpp
+++ b/aten/src/ATen/native/AveragePool2d.cpp
@ -25,19 +25,18 @@ TORCH_PRECOMPUTE_META_FUNC(avg_pool2d)
  // #20866, #22032: Guarantee this for the official C++ API?
  TORCH_CHECK(kernel_size.size() == 1 || kernel_size.size() == 2,
    "avg_pool2d: kernel_size must either be a single int, or a tuple of two ints");
-  const int kH = safe_downcast<int, int64_t>(kernel_size[0]);
-  const int kW = kernel_size.size() == 1 ? kH : safe_downcast<int, int64_t>(kernel_size[1]);
+  const int64_t kH = kernel_size[0];
+  const int64_t kW = kernel_size.size() == 1 ? kH : kernel_size[1];

  TORCH_CHECK(stride.empty() || stride.size() == 1 || stride.size() == 2,
    "avg_pool2d: stride must either be omitted, a single int, or a tuple of two ints");
-  const int dH = stride.empty() ? kH : safe_downcast<int, int64_t>(stride[0]);
-  const int dW = stride.empty() ? kW :
-                 stride.size() == 1 ? dH : safe_downcast<int, int64_t>(stride[1]);
+  const int64_t dH = stride.empty() ? kH : stride[0];
+  const int64_t dW = stride.empty() ? kW : stride.size() == 1 ? dH : stride[1];

  TORCH_CHECK(padding.size() == 1 || padding.size() == 2,
    "avg_pool2d: padding must either be a single int, or a tuple of two ints");
-  const int padH = safe_downcast<int, int64_t>(padding[0]);
-  const int padW = padding.size() == 1 ? padH : safe_downcast<int, int64_t>(padding[1]);
+  const int64_t padH = padding[0];
+  const int64_t padW = padding.size() == 1 ? padH : padding[1];

  TORCH_CHECK(!divisor_override.has_value() || divisor_override.value() != 0,
    "divisor must be not zero");
--- a/aten/src/ATen/native/Convolution.cpp
+++ b/aten/src/ATen/native/Convolution.cpp
@ -410,8 +410,8 @@ struct ConvParams {
      return false;
    }
    static long cudnn_version = detail::getCUDAHooks().versionCuDNN();
-    // broken on cuDNN 9.8 - 9.14
-    if (cudnn_version >= 90800 && cudnn_version < 91500) {
+    // broken on cuDNN 9.8
+    if (cudnn_version >= 90800) {
      if (cudnn_conv_suggest_memory_format(input, weight) == at::MemoryFormat::Contiguous &&
          (input.scalar_type() == at::kBFloat16 || input.scalar_type() == at::kHalf) &&
          weight.dim() == 5) {
@ -689,10 +689,6 @@ static void check_shape_forward(const at::Tensor& input,
             ", but got bias of size ", at::symint::sizes<T>(bias), " instead");

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

  // NOTE: Assumes all strides are positive, which is true for now
-  for (auto i = t.dim() - 1; i >= 0; --i) {
+  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+  for (int i = t.dim() - 1; i >= 0; --i) {
    auto curDimIndex = linearId % t.sym_size(i);
    auto curDimOffset = curDimIndex * t.sym_stride(i);
    offset += curDimOffset;
--- a/aten/src/ATen/native/Integration.cpp
+++ b/aten/src/ATen/native/Integration.cpp
@ -40,7 +40,7 @@ Tensor do_trapezoid(const Tensor& y, const Tensor& dx, int64_t dim) {
 // When dx is constant, the above formula simplifies
 // to dx * [(\sum_{i=1}^n y_i) - (y_1 + y_n)/2]
 Tensor do_trapezoid(const Tensor& y, double dx, int64_t dim) {
-    return (y.sum(dim) - (y.select(dim, 0) + y.select(dim, -1)) * 0.5) * dx;
+    return (y.sum(dim) - (y.select(dim, 0) + y.select(dim, -1)) * (0.5)) * dx;
 }

 Tensor zeros_like_except(const Tensor& y, int64_t dim) {
--- a/aten/src/ATen/native/Linear.cpp
+++ b/aten/src/ATen/native/Linear.cpp
@ -201,7 +201,7 @@ static Tensor sumproduct_pair(const Tensor& left_, const Tensor& right_, IntArra
  out_size.reserve(out_num_dim);
  for (auto& d : lro) out_size.push_back(left.sym_size(d));
  for (auto& d : lo) out_size.push_back(left.sym_size(d));
-  for (auto& d : sum_dims_) { out_size.emplace_back(1); (void)d; }; // avoid warning about not using d
+  for (auto& d : sum_dims_) { out_size.emplace_back(1); (void)(d); }; // avoid warning about not using d
  for (auto& d : ro) out_size.push_back(right.sym_size(d));

  std::vector<int64_t> lpermutation(lro);
@ -640,7 +640,7 @@ Tensor einsum(std::string_view equation, TensorList operands, at::OptionalIntArr
    }
  }

-  return std::move(ops[0]);
+  return ops[0];
 }

 // _trilinear computes a trilinear einstein sum with an unrolled dimension
--- a/aten/src/ATen/native/Math.h
+++ b/aten/src/ATen/native/Math.h
@ -192,7 +192,7 @@ Date:  February 1996
  x = x - (std::erf(x) - y) / ((static_cast<T>(2.0)/static_cast<T>(std::sqrt(c10::pi<double>)))*std::exp(-x*x));
  x = x - (std::erf(x) - y) / ((static_cast<T>(2.0)/static_cast<T>(std::sqrt(c10::pi<double>)))*std::exp(-x*x));

-  return x;
+  return(x);
 }

 #undef CENTRAL_RANGE
@ -3819,7 +3819,7 @@ inline C10_HOST_DEVICE T shifted_chebyshev_polynomial_v_forward(T x, int64_t n)

    if ((n > 6) && (std::abs(x + x - T(1.0)) < T(1.0))) {
        if (std::sin(std::acos(x + x - T(1.0)) / T(2.0)) != T(1.0)) {
-            return std::cos((n + T(0.5)) * std::acos(x + x - T(1.0))) / std::cos(std::acos(x + x - T(1.0)) / T(2.0));
+            return std::cos(((n) + T(0.5)) * std::acos(x + x - T(1.0))) / std::cos(std::acos(x + x - T(1.0)) / T(2.0));
        }

        if (n % 2 == 0) {
--- a/aten/src/ATen/native/QuantizedLinear.cpp
+++ b/aten/src/ATen/native/QuantizedLinear.cpp
@ -68,6 +68,7 @@ Tensor fbgemm_linear_int8_weight_fp32_activation(
  const float* input_ptr = input_contig.const_data_ptr<float>();

  TORCH_CHECK(input.dim() >= 2);
+  // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
  const int64_t M = size_to_dim_(input.dim() - 1, input.sizes());
  const int64_t K = input.size(input.dim() - 1);
  TORCH_CHECK(weight.dim() == 2);
--- a/aten/src/ATen/native/RowwisePrune.cpp
+++ b/aten/src/ATen/native/RowwisePrune.cpp
@ -25,7 +25,7 @@ std::tuple<Tensor, Tensor> _rowwise_prune_helper(
  auto mask_contig = mask.contiguous();
  auto mask_data = mask_contig.data_ptr<bool>();
  for (const auto i : c10::irange(mask.numel())) {
-    num_non_masked_rows += ((mask_data[i] == true) ? 1 : 0);
+    num_non_masked_rows += (((mask_data[i] == true)) ? 1 : 0);
  }
  int num_cols = weights.size(1);
  auto pruned_2d_tensor = at::empty({num_non_masked_rows, num_cols},
--- a/aten/src/ATen/native/ao_sparse/quantized/cpu/qlinear_dynamic.cpp
+++ b/aten/src/ATen/native/ao_sparse/quantized/cpu/qlinear_dynamic.cpp
@ -40,7 +40,7 @@ at::Tensor PackedLinearWeightQnnp::apply_dynamic_impl<false>(
      "quantized_sparse_linear(): Input tensor rank should be >= 2");

  const auto rows_input = c10::multiply_integers(input.sizes().begin(), input.sizes().end() - 1);
-  const auto cols_input = input.size(input.dim() - 1);
+  const auto cols_input = static_cast<int64_t>(input.size(input.dim() - 1));
  TORCH_CHECK(
      cols_input == input_channels_,
      "quantized_sparse_linear: Input tensor's last and weight tensor's"
--- a/aten/src/ATen/native/ao_sparse/quantized/cpu/qlinear_unpack.cpp
+++ b/aten/src/ATen/native/ao_sparse/quantized/cpu/qlinear_unpack.cpp
@ -65,8 +65,8 @@ LinearPackedSerializationType PackedLinearWeight::unpack() {
 #ifdef USE_PYTORCH_QNNPACK

 LinearPackedSerializationType PackedLinearWeightQnnp::unpack() {
-  const int64_t N = output_channels_;
-  const int64_t K = input_channels_;
+  const int64_t N = static_cast<int64_t>(output_channels_);
+  const int64_t K = static_cast<int64_t>(input_channels_);

  float* w_scales_ptr = w_scales_.data_ptr<float>();

--- a/aten/src/ATen/native/cpu/AtomicAddFloat.h
+++ b/aten/src/ATen/native/cpu/AtomicAddFloat.h
@ -17,7 +17,7 @@ static inline void cpu_atomic_add_float(float* dst, float fvalue)
  } uf32_t;

  uf32_t new_value, old_value;
-  std::atomic<unsigned>* dst_intV = (std::atomic<unsigned>*)dst;
+  std::atomic<unsigned>* dst_intV = (std::atomic<unsigned>*)(dst);

  old_value.floatV = *dst;
  new_value.floatV = old_value.floatV + fvalue;
--- a/aten/src/ATen/native/cpu/BinaryOpsKernel.cpp
+++ b/aten/src/ATen/native/cpu/BinaryOpsKernel.cpp
@ -851,7 +851,7 @@ void sigmoid_backward_kernel(TensorIteratorBase& iter) {
          });
    });
  } else if (iter.dtype() == kBFloat16) {
-    auto one_vec = Vectorized<float>((float)1);
+    auto one_vec = Vectorized<float>((float)(1));
    cpu_kernel_vec(
        iter,
        [=](BFloat16 a, BFloat16 b) -> BFloat16 {
--- a/aten/src/ATen/native/cpu/DistanceOpsKernel.cpp
+++ b/aten/src/ATen/native/cpu/DistanceOpsKernel.cpp
@ -160,9 +160,10 @@ struct Dist {
    // value of k.
    parallel_for(0, combs, internal::GRAIN_SIZE / (16 * m), [p, self_start, self_end, n, m, res_start](int64_t k, int64_t end) {
      const Vec pvec(p);
-      double n2 = static_cast<double>(n) - .5;
+      double n2 = n - .5;
      // The -1 accounts for floating point truncation issues
-      int64_t i = static_cast<int64_t>((n2 - std::sqrt(n2 * n2 - 2.0 * static_cast<double>(k) - 1.0)));
+      // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
+      int64_t i = static_cast<int64_t>((n2 - std::sqrt(n2 * n2 - 2 * k - 1)));
      int64_t j = k - n * i + i * (i + 1) / 2 + i + 1;

      const scalar_t * self_i = self_start + i * m;
--- a/aten/src/ATen/native/cpu/PointwiseOpsKernel.cpp
+++ b/aten/src/ATen/native/cpu/PointwiseOpsKernel.cpp
@ -139,7 +139,7 @@ void smooth_l1_backward_cpu_kernel(TensorIterator& iter, const Scalar& norm, dou
      }
    );
  } else {
-    AT_DISPATCH_ALL_TYPES_AND(kHalf, dtype, "smooth_l1_backward_cpu_out", [&] {
+    AT_DISPATCH_ALL_TYPES(dtype, "smooth_l1_backward_cpu_out", [&] {
    auto norm_val = norm.to<scalar_t>();
    scalar_t beta_val(beta);
    auto norm_val_vec = Vectorized<scalar_t>(norm_val);
--- a/aten/src/ATen/native/cpu/Unfold2d.cpp
+++ b/aten/src/ATen/native/cpu/Unfold2d.cpp
@ -298,7 +298,7 @@ void unfolded2d_copy(
                      memcpy(
                          dst + (size_t)y * output_width + x,
                          src + (size_t)iy * input_width + ix,
-                          sizeof(scalar_t) * 1);
+                          sizeof(scalar_t) * (1));
                  }
                }
              }
@ -317,7 +317,7 @@ void unfolded2d_copy(
                  memcpy(
                      dst + (size_t)y * output_width + x,
                      src + (size_t)iy * input_width + ix + x * dW,
-                      sizeof(scalar_t) * 1);
+                      sizeof(scalar_t) * (1));
              }
            }
          }
--- a/aten/src/ATen/native/cpu/UpSampleKernelAVXAntialias.h
+++ b/aten/src/ATen/native/cpu/UpSampleKernelAVXAntialias.h
@ -342,7 +342,7 @@ void upsample_avx_bilinear_bicubic_uint8(

  if (need_horizontal) {
    int interp_dim = 3;
-    auto stride = skip_unpacking ? num_channels : 4;
+    auto stride = (skip_unpacking) ? num_channels : 4;
    std::tie(horiz_indices_weights, ksize_horiz, horiz_weights_precision) =
        F::compute_index_ranges_int16_weights(
            /*input_size=*/xin,
@ -358,7 +358,7 @@ void upsample_avx_bilinear_bicubic_uint8(

  if (need_vertical) {
    int interp_dim = 2;
-    auto stride = skip_unpacking ? num_channels * xout : 4 * xout;
+    auto stride = (skip_unpacking) ? num_channels * xout : 4 * xout;
    std::tie(vert_indices_weights, ksize_vert, vert_weights_precision) =
        F::compute_index_ranges_int16_weights(
            /*input_size=*/yin,
@ -377,17 +377,17 @@ void upsample_avx_bilinear_bicubic_uint8(
  // horizontal-only or vertical-only interpolation, and if the tensor doesn't
  // need repacking
  if (need_horizontal && (need_vertical || !skip_packing)) {
-    auto c = skip_unpacking ? num_channels : 4;
+    auto c = (skip_unpacking) ? num_channels : 4;
    buffer_horiz = at::empty({c, yin, xout}, input.options());
  }
  if (need_vertical && !skip_packing) {
-    auto c = skip_unpacking ? num_channels : 4;
+    auto c = (skip_unpacking) ? num_channels : 4;
    buffer_vert = at::empty({c, yout, xout}, input.options());
  }

  for (const auto i : c10::irange(batch_size)) {

-    at::Tensor unpacked_input = skip_unpacking ? input[i] : unpack_rgb(input[i]);
+    at::Tensor unpacked_input = (skip_unpacking) ? input[i] : unpack_rgb(input[i]);
    at::Tensor unpacked_output;

    if (need_horizontal) {
@ -411,7 +411,7 @@ void upsample_avx_bilinear_bicubic_uint8(
      unpacked_output = unpacked_input = unpacked_output_temp;
    }
    if (need_vertical) {
-      unpacked_output = skip_packing ? output[i] : buffer_vert;
+      unpacked_output = (skip_packing) ? output[i] : buffer_vert;

      ImagingResampleVertical(
          unpacked_output,
@ -502,7 +502,7 @@ void ImagingResampleHorizontalConvolution8u4x(
  // RGBA: b4_delta = b4_delta_soft = 3
  // RGB : b4_delta = 5
  // RGB : b4_delta_soft = 4
-  const auto b4_delta = (stride == 4) ? 3 : (is_last_line ? 5 : 4);
+  const auto b4_delta = (stride == 4) ? 3 : ((is_last_line) ? 5 : 4);

  // In block 2 (2 means we process 2 weights values together), we read input data
  // with _mm_loadl_epi64, i.e. 8 bytes, per one line:
@ -515,7 +515,7 @@ void ImagingResampleHorizontalConvolution8u4x(
  // RGBA: b2_delta = b2_delta_soft = 1
  // RGB : b2_delta = 2
  // RGB : b2_delta_soft = 1
-  const auto b2_delta = (stride == 4) ? 1 : (is_last_line ? 2 : 1);
+  const auto b2_delta = (stride == 4) ? 1 : ((is_last_line) ? 2 : 1);

  const auto max_out_x_strided = out_xsize * stride;
  const auto max_in_x_strided = in_xsize * stride;
@ -819,7 +819,7 @@ void ImagingResampleHorizontalConvolution8u(
  // RGBA: b8_delta = b8_delta_soft = 7
  // RGB : b8_delta = 10
  // RGB : b8_delta_soft = 9
-  const auto b8_delta = (stride == 4) ? 7 : (is_last_line ? 10 : 9);
+  const auto b8_delta = (stride == 4) ? 7 : ((is_last_line) ? 10 : 9);

  // In block 4 (4 means we process 4 weight values together), we read
  // 16 bytes of input data.
@ -832,7 +832,7 @@ void ImagingResampleHorizontalConvolution8u(
  // RGBA: b4_delta = b4_delta_soft = 3
  // RGB : b4_delta = 5
  // RGB : b4_delta_soft = 4
-  const auto b4_delta = (stride == 4) ? 3 : (is_last_line ? 5 : 4);
+  const auto b4_delta = (stride == 4) ? 3 : ((is_last_line) ? 5 : 4);

  // In block 2 (2 means we process 2 weight values together), we read
  // 8 bytes of input data.
@ -845,7 +845,7 @@ void ImagingResampleHorizontalConvolution8u(
  // RGBA: b2_delta = b2_delta_soft = 1
  // RGB : b2_delta = 2
  // RGB : b2_delta_soft = 1
-  const auto b2_delta = (stride == 4) ? 1 : (is_last_line ? 2 : 1);
+  const auto b2_delta = (stride == 4) ? 1 : ((is_last_line) ? 2 : 1);

  const auto max_out_x_strided = out_xsize * stride;
  const auto max_in_x_strided = in_xsize * stride;
--- a/aten/src/ATen/native/cpu/int4mm_kernel.cpp
+++ b/aten/src/ATen/native/cpu/int4mm_kernel.cpp
@ -644,8 +644,8 @@ void weight_to_int4pack_kernel(
            int32_t val2 = src[(d + 32) * K + k];
            int32_t val3 = src[(d + 48) * K + k];

-            uint8_t packed02 = ((uint8_t)val2 << 4) | ((uint8_t)val0);
-            uint8_t packed13 = ((uint8_t)val3 << 4) | ((uint8_t)val1);
+            uint8_t packed02 = (((uint8_t)(val2) << 4)) | ((uint8_t)(val0));
+            uint8_t packed13 = (((uint8_t)(val3) << 4)) | ((uint8_t)(val1));

            dst[k * 32 + d] = packed02;
            dst[k * 32 + 16 + d] = packed13;
@ -656,7 +656,7 @@ void weight_to_int4pack_kernel(
            int32_t val0 = src[n * K + k];
            int32_t val1 = src[n * K + K + k];

-            uint8_t packed = ((uint8_t)val1 << 4) | ((uint8_t)val0);
+            uint8_t packed = (((uint8_t)(val1) << 4)) | ((uint8_t)(val0));
            dst[k * nb_size / 2 + n / 2] = packed;
          }
        }
@ -667,7 +667,7 @@ void weight_to_int4pack_kernel(
            int32_t val0 = src[(d + 0) * K + k];
            int32_t val1 = src[(d + 16) * K + k];

-            uint8_t packed01 = ((uint8_t)val1 << 4) | ((uint8_t)val0);
+            uint8_t packed01 = (((uint8_t)(val1) << 4)) | ((uint8_t)(val0));
            dst[k * 16 + d] = packed01;
          }
        } else {
@ -676,7 +676,7 @@ void weight_to_int4pack_kernel(
            int32_t val0 = src[n * K + k];
            int32_t val1 = src[n * K + K + k];

-            uint8_t packed = ((uint8_t)val1 << 4) | ((uint8_t)val0);
+            uint8_t packed = (((uint8_t)(val1) << 4)) | ((uint8_t)(val0));
            dst[k * nb_size / 2 + n / 2] = packed;
          }
        }
@ -685,7 +685,7 @@ void weight_to_int4pack_kernel(
          int32_t val0 = src[n * K + k];
          int32_t val1 = src[n * K + K + k];

-          uint8_t packed = ((uint8_t)val1 << 4) | ((uint8_t)val0);
+          uint8_t packed = (((uint8_t)(val1) << 4)) | ((uint8_t)(val0));
          dst[k * nb_size / 2 + n / 2] = packed;
        }
 #endif
@ -872,16 +872,16 @@ void ref_dyn_quant_matmul_4bit_channelwise_kernel(
          for (size_t k_idx = 0; k_idx < k; ++k_idx) {
            const float src0_0 = src_ptr[k_idx];

-            max0 = std::max(src0_0, max0);
-            min0 = std::min(src0_0, min0);
+            max0 = (std::max)(src0_0, max0);
+            min0 = (std::min)(src0_0, min0);
          }

          // Maximum/minimum int8 values
          const float qmin = (float)INT8_MIN;
          const float qmax = (float)INT8_MAX;

-          const float rmin0 = std::min(0.0f, min0);
-          const float rmax0 = std::max(0.0f, max0);
+          const float rmin0 = (std::min)(0.0f, min0);
+          const float rmax0 = (std::max)(0.0f, max0);

          const float scale0 =
              rmin0 == rmax0 ? 1.f : (qmax - qmin) / (rmax0 - rmin0);
@ -900,8 +900,8 @@ void ref_dyn_quant_matmul_4bit_channelwise_kernel(
              ? qmin - descaled_min0
              : qmax - descaled_max0;

-          zero_point0 = std::max(zero_point0, qmin);
-          zero_point0 = std::min(zero_point0, qmax);
+          zero_point0 = (std::max)(zero_point0, qmin);
+          zero_point0 = (std::min)(zero_point0, qmax);

          // Round to nearest integer
          const int32_t nudged_zero_point0 = lrintf(zero_point0);
@ -909,9 +909,9 @@ void ref_dyn_quant_matmul_4bit_channelwise_kernel(
          int8_t* dst_ptr = lhs_qa8dx + m_idx * dst_stride;

          // LHS offset at the beginning of the row
-          *((float*)dst_ptr) = recip_scale0;
+          *((float*)(dst_ptr)) = recip_scale0;
          dst_ptr += sizeof(float);
-          *((int32_t*)dst_ptr) = -nudged_zero_point0;
+          *((int32_t*)(dst_ptr)) = -nudged_zero_point0;
          dst_ptr += sizeof(int32_t);

          // Quantize the channels
@ -922,8 +922,8 @@ void ref_dyn_quant_matmul_4bit_channelwise_kernel(
            int32_t v0_s32 = (int32_t)(std::round(src0_0 * scale0));

            v0_s32 = v0_s32 + nudged_zero_point0;
-            v0_s32 = std::max(v0_s32, static_cast<int32_t>(INT8_MIN));
-            v0_s32 = std::min(v0_s32, static_cast<int32_t>(INT8_MAX));
+            v0_s32 = (std::max)(v0_s32, static_cast<int32_t>(INT8_MIN));
+            v0_s32 = (std::min)(v0_s32, static_cast<int32_t>(INT8_MAX));
            dst_ptr[0] = (int8_t)v0_s32;
            dst_ptr += sizeof(int8_t);
          }
@ -988,8 +988,8 @@ void ref_dyn_quant_matmul_4bit_channelwise_kernel(
      main_acc = main_acc * lhs_scale;

      // Clamp (min-max) operation
-      main_acc = std::max(main_acc, scalar_min);
-      main_acc = std::min(main_acc, scalar_max);
+      main_acc = (std::max)(main_acc, scalar_min);
+      main_acc = (std::min)(main_acc, scalar_max);

      dst_f32[0] = main_acc;
      dst_f32 += 1;
@ -1024,15 +1024,15 @@ void ref_dyn_quant_matmul_4bit_groupwise_kernel(

      for (size_t k_idx = 0; k_idx < k; ++k_idx) {
        const float src0_0 = src_ptr[k_idx];
-        max0 = std::max(src0_0, max0);
-        min0 = std::min(src0_0, min0);
+        max0 = (std::max)(src0_0, max0);
+        min0 = (std::min)(src0_0, min0);
      }

      const float qmin = (float)INT8_MIN;
      const float qmax = (float)INT8_MAX;

-      const float rmin0 = std::min(0.0f, min0);
-      const float rmax0 = std::max(0.0f, max0);
+      const float rmin0 = (std::min)(0.0f, min0);
+      const float rmax0 = (std::max)(0.0f, max0);
      const float scale0 =
          (rmin0 == rmax0) ? 1.f : (qmax - qmin) / (rmax0 - rmin0);
      const float recip_scale0 = scale0 ? 1.0f / scale0 : 0.0f;
@ -1044,22 +1044,22 @@ void ref_dyn_quant_matmul_4bit_groupwise_kernel(
          ? qmin - descaled_min0
          : qmax - descaled_max0;

-      zero_point0 = std::max(zero_point0, qmin);
-      zero_point0 = std::min(zero_point0, qmax);
+      zero_point0 = (std::max)(zero_point0, qmin);
+      zero_point0 = (std::min)(zero_point0, qmax);
      const int32_t nudged_zero_point0 = lrintf(zero_point0);

      int8_t* dst_ptr = lhs_qa8dx + row_idx * dst_stride;

-      *((float*)dst_ptr) = recip_scale0;
+      *((float*)(dst_ptr)) = recip_scale0;
      dst_ptr += sizeof(float);
-      *((int32_t*)dst_ptr) = -nudged_zero_point0;
+      *((int32_t*)(dst_ptr)) = -nudged_zero_point0;
      dst_ptr += sizeof(int32_t);

      for (size_t k_idx = 0; k_idx < k; ++k_idx) {
        const float src0_0 = src_ptr[k_idx];
        int32_t v0_s32 = (int32_t)(std::round(src0_0 * scale0));
-        v0_s32 = std::max(
-            std::min(
+        v0_s32 = (std::max)(
+            (std::min)(
                v0_s32 + nudged_zero_point0, static_cast<int32_t>(INT8_MAX)),
            static_cast<int32_t>(INT8_MIN));
        dst_ptr[0] = (int8_t)v0_s32;
@ -1118,8 +1118,8 @@ void ref_dyn_quant_matmul_4bit_groupwise_kernel(
      }

      main_acc = main_acc * lhs_scale;
-      main_acc = std::max(main_acc, scalar_min);
-      main_acc = std::min(main_acc, scalar_max);
+      main_acc = (std::max)(main_acc, scalar_min);
+      main_acc = (std::min)(main_acc, scalar_max);

      dst_f32[0] = main_acc;
      dst_f32 += 1;
--- a/aten/src/ATen/native/cuda/Blas.cpp
+++ b/aten/src/ATen/native/cuda/Blas.cpp
@ -4,6 +4,7 @@
 #include <c10/util/SmallVector.h>
 #include <c10/core/Scalar.h>
 #include <c10/core/ScalarType.h>
+#include <c10/util/Exception.h>
 #define TORCH_ASSERT_ONLY_METHOD_OPERATORS
 #include <ATen/core/Tensor.h>
 #include <ATen/core/NamedTensor.h>
@ -169,14 +170,10 @@ static bool isInputCompliesAddmmCudaLt(Tensor& result, const Tensor& self, const
  #if defined(CUDA_VERSION) || defined(USE_ROCM)
  const auto scalar_type = mat1.scalar_type();
  return (beta.toComplexDouble() == 1.0
+    // self.dim() == 1 && result.dim() == 2 && self.sizes()[0] == mat2_sizes[1]
+    // is to use lt interface only when self is bias.
+    && self.dim() == 1 && self.sizes()[0] == mat2_sizes[1] && self.is_contiguous()
    && result.dim() == 2 && result.is_contiguous()
-    // Conditions for bias to be fusable
-    && (
-      self.is_contiguous() &&
-      // NOTE: fine to have 1-len dims to the left from the right-most one
-      (self.dim() == 1 || self.squeeze().dim() == 1) &&
-      self.sizes().back() == mat2_sizes[1]
-    )
    && ( // some dtype restrictions
      #ifndef USE_ROCM
      scalar_type == at::ScalarType::Double ||
--- a/aten/src/ATen/native/cuda/GroupedBlas.cpp
+++ b/aten/src/ATen/native/cuda/GroupedBlas.cpp
@ -208,62 +208,6 @@ _f8_f8_bf16_rowwise_grouped_mm(
 #endif
 }

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

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

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

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

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

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

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

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

 } // anonymous namespace

@ -601,9 +525,8 @@ _scaled_grouped_mm_cuda_v2(
          out);
    }
    case ScaledGemmImplementation::MXFP8_MXFP8: {
-      // scale shape checks
-      _check_scales_blocked(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
-      _check_scales_blocked(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
+      _check_scales_mxfp8(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
+      _check_scales_mxfp8(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
      return _mx8_mx8_bf16_grouped_mm_fbgemm(
          mat_a,
          mat_b,
@ -614,36 +537,6 @@ _scaled_grouped_mm_cuda_v2(
          offs.value(),
          out);
    }
-    case ScaledGemmImplementation::MXFP4_MXFP4: {
-      // scale shape checks
-      _check_scales_blocked(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
-      _check_scales_blocked(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
-      return _f4_f4_bf16_grouped_mm_fbgemm(
-          mat_a,
-          mat_b,
-          scale_a[0], /* block-scale A */
-          std::nullopt, /* global-scale A */
-          scale_b[0], /* block-scale B */
-          std::nullopt, /* global-scale B */
-          offs.value(),
-          std::nullopt, /* bias */
-          out);
-    }
-    case ScaledGemmImplementation::NVFP4_NVFP4: {
-      // scale shape checks
-      _check_scales_blocked(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
-      _check_scales_blocked(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
-      return _f4_f4_bf16_grouped_mm_fbgemm(
-          mat_a,
-          mat_b,
-          scale_a[0], /* block-scale A */
-          scale_a[1], /* global-scale A */
-          scale_b[0], /* block-scale B */
-          scale_b[1], /* global-scale B */
-          offs.value(),
-          std::nullopt, /* bias */
-          out);
-    }
    default:
      TORCH_CHECK_NOT_IMPLEMENTED(false,
          "_scaled_grouped_mm_cuda_v2 is in an inconsistent state - should never reach here");
--- a/aten/src/ATen/native/cuda/IndexKernelUtils.cu
+++ b/aten/src/ATen/native/cuda/IndexKernelUtils.cu
@ -13,7 +13,7 @@ __global__ void vectorized_gather_kernel(char * out, char * inp, index_t * idx,
    if (allow_neg_indices) {
        ind = (ind < 0) ? ind + ind_dim_size : ind;
    }
-    CUDA_KERNEL_ASSERT_VERBOSE(ind >=0 && ind < ind_dim_size && "vectorized gather kernel index out of bounds", "Expected 0 <= index < ind_dim_size(%ld), but got index = %ld", ind_dim_size, ind);
+    CUDA_KERNEL_ASSERT(ind >=0 && ind < ind_dim_size && "vectorized gather kernel index out of bounds");
    int32_t off = (blockDim.x * blockIdx.y + threadIdx.x) * Alignment; // off is guaranteed to be within int32 limits
    if (off >= slice_size) return;
    auto vec = at::native::memory::ld_vec<Alignment>(inp + ind * inp_stride + off);
--- a/aten/src/ATen/native/cuda/ScaledBlas.cpp
+++ b/aten/src/ATen/native/cuda/ScaledBlas.cpp
@ -794,24 +794,6 @@ void _check_deepseek_scale_stride(const Tensor& scale, const Tensor& t, const Sc
  }
 }

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

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

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

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

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

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

 Tensor&
--- a/aten/src/ATen/native/cuda/ScatterGatherKernel.cu
+++ b/aten/src/ATen/native/cuda/ScatterGatherKernel.cu
@ -160,8 +160,8 @@ struct _cuda_scatter_gather_internal_kernel {
      auto offsets = offset_calc.get(i);

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

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

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

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

 namespace at::native {

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

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

 namespace at::native {

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

 void tanh_kernel_cuda(TensorIteratorBase& iter) {
  auto common_dtype = iter.common_dtype();
  if (at::isComplexType(common_dtype)) {
-    // Disabled due to accuracy issues
-#if 0 && AT_USE_JITERATOR()
+#if AT_USE_JITERATOR()
    static const auto tanh_string = jiterator_stringify(
        template <typename T> T tanh_impl(T a) { return std::tanh(a); });
    AT_DISPATCH_COMPLEX_TYPES_AND(
--- a/aten/src/ATen/native/cuda/layer_norm_kernel.cu
+++ b/aten/src/ATen/native/cuda/layer_norm_kernel.cu
@ -141,8 +141,7 @@ WelfordDataLN cuWelfordOnlineSum(
  if constexpr (!rms_norm){
    U delta = val - curr_sum.mean;
    U new_count = curr_sum.count + 1.f;
-//Due to low CU count, we run into accuracy issues on gfx90a with `__builtin_amdgcn_rcpf`
-#if defined(USE_ROCM) && !defined(__gfx90a__) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
+#if defined(USE_ROCM) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
    U new_mean = curr_sum.mean + delta * __builtin_amdgcn_rcpf(new_count);
 #else
    U new_mean = curr_sum.mean + delta * (1.f/new_count); //proper division is slow, this is less accurate but noticeably faster
@ -164,8 +163,7 @@ WelfordDataLN cuWelfordCombine(
    U count = dataA.count + dataB.count;
    U mean, sigma2;
    if (count > decltype(dataB.count){0}) {
-//Due to low CU count, we run into accuracy issues on gfx90a with `__builtin_amdgcn_rcpf`
-#if defined(USE_ROCM) && !defined(__gfx90a__) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
+#if defined(USE_ROCM) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
      auto coef = __builtin_amdgcn_rcpf(count);
 #else
      auto coef = 1.f/count; //NB we don't use --use_fast_math, but this is emulation, 1./count goes to intrinsic, `* coef` is multiplication, instead of slow fp division
--- a/aten/src/ATen/native/mkl/SpectralOps.cpp
+++ b/aten/src/ATen/native/mkl/SpectralOps.cpp
@ -330,6 +330,7 @@ Tensor _fft_c2c_mkl(const Tensor& self, IntArrayRef dim, int64_t normalization,
 }

 #elif AT_MKL_ENABLED()
+#include <ATen/Dispatch.h>

 #include <algorithm>
 #include <numeric>
--- a/aten/src/ATen/native/mkldnn/xpu/detail/WoQMatmul.cpp
+++ b/aten/src/ATen/native/mkldnn/xpu/detail/WoQMatmul.cpp
@ -41,7 +41,7 @@ void woq_matmul_int4_impl(
      dst_usr_dims;
  dnnl::memory::dims m1_usr_strides, m2_usr_strides, scale_usr_strides,
      zp_usr_strides, dst_usr_strides;
-  int compressed_k = k / 8;
+  int compressed_k = (int)(k / 8);
  int num_groups = (int)(k / group_size);
  m1_usr_dims = {m, k};
  m1_usr_strides = {m1.stride(0), m1.stride(1)};
--- a/aten/src/ATen/native/mps/kernels/BinaryKernel.metal
+++ b/aten/src/ATen/native/mps/kernels/BinaryKernel.metal
@ -86,28 +86,6 @@ struct zeta_functor {
  }
 };

-struct logaddexp_functor {
-  template <typename T, enable_if_t<is_floating_point_v<T>, bool> = true>
-  inline T operator()(const T a, const T b) {
-    return c10::metal::logaddexp(a, b);
-  }
-  template <typename T, enable_if_t<is_integral_v<T>, bool> = true>
-  inline float operator()(const T a, const T b) {
-    return c10::metal::logaddexp(float(a), float(b));
-  }
-};
-
-struct logaddexp2_functor {
-  template <typename T, enable_if_t<is_floating_point_v<T>, bool> = true>
-  inline T operator()(const T a, const T b) {
-    return c10::metal::logaddexp2(a, b);
-  }
-  template <typename T, enable_if_t<is_integral_v<T>, bool> = true>
-  inline float operator()(const T a, const T b) {
-    return c10::metal::logaddexp2(float(a), float(b));
-  }
-};
-
 struct xlog1py_functor {
  template <typename T, enable_if_t<is_floating_point_v<T>, bool> = true>
  inline T operator()(const T a, const T b) {
@ -399,10 +377,6 @@ REGISTER_FLOAT_BINARY_OP(fmin);
 REGISTER_FLOAT_BINARY_OP(nextafter);
 REGISTER_FLOAT_BINARY_OP(zeta);
 REGISTER_INT2FLOAT_BINARY_OP(zeta);
-REGISTER_FLOAT_BINARY_OP(logaddexp);
-REGISTER_INT2FLOAT_BINARY_OP(logaddexp);
-REGISTER_FLOAT_BINARY_OP(logaddexp2);
-REGISTER_INT2FLOAT_BINARY_OP(logaddexp2);
 REGISTER_FLOAT_BINARY_OP(xlog1py);
 REGISTER_INT2FLOAT_BINARY_OP(xlog1py);
 REGISTER_FLOAT_BINARY_OP(chebyshev_polynomial_t);
@ -489,8 +463,6 @@ REGISTER_BINARY_OP(add, float2, float2);
 REGISTER_BINARY_OP(add, half2, half2);
 REGISTER_BINARY_OP(sub, float2, float2);
 REGISTER_BINARY_OP(sub, half2, half2);
-REGISTER_BINARY_OP(logaddexp, float2, float2);
-REGISTER_BINARY_OP(logaddexp, half2, half2);
 REGISTER_BINARY_ALPHA_OP(add_alpha, float2, float2, float2);
 REGISTER_BINARY_ALPHA_OP(add_alpha, half2, half2, half2);
 REGISTER_BINARY_ALPHA_OP(sub_alpha, float2, float2, float2);
--- a/Show More
+++ b/Show More