Add sequoia ops

Add macos15 check
Remove MPSGraph usage from cat_mps_out
2025-10-24 07:27:32 +08:00 · 2025-05-07 13:40:12 -07:00 · 2025-05-07 13:33:50 -07:00 · 2025-05-07 13:30:53 -07:00 · 2025-05-05 09:36:25 -07:00
5890 changed files with 81993 additions and 219782 deletions
--- a/.ci/aarch64_linux/aarch64_ci_build.sh
+++ b/.ci/aarch64_linux/aarch64_ci_build.sh
@ -3,7 +3,9 @@ set -eux -o pipefail

 GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}

-if [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
+if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
+    export TORCH_CUDA_ARCH_LIST="9.0"
+elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
    export TORCH_CUDA_ARCH_LIST="9.0;10.0;12.0"
 fi

@ -25,7 +27,6 @@ if [ "$DESIRED_CUDA" = "cpu" ]; then
    USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
 else
    echo "BASE_CUDA_VERSION is set to: $DESIRED_CUDA"
-    export USE_SYSTEM_NCCL=1
    #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
    USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
 fi
--- a/.ci/aarch64_linux/aarch64_wheel_ci_build.py
+++ b/.ci/aarch64_linux/aarch64_wheel_ci_build.py
@ -31,47 +31,33 @@ def build_ArmComputeLibrary() -> None:
        "build=native",
    ]
    acl_install_dir = "/acl"
-    acl_checkout_dir = os.getenv("ACL_SOURCE_DIR", "ComputeLibrary")
-    if os.path.isdir(acl_install_dir):
-        shutil.rmtree(acl_install_dir)
-    if not os.path.isdir(acl_checkout_dir) or not len(os.listdir(acl_checkout_dir)):
-        check_call(
-            [
-                "git",
-                "clone",
-                "https://github.com/ARM-software/ComputeLibrary.git",
-                "-b",
-                "v25.02",
-                "--depth",
-                "1",
-                "--shallow-submodules",
-            ]
-        )
+    acl_checkout_dir = "ComputeLibrary"
+    os.makedirs(acl_install_dir)
+    check_call(
+        [
+            "git",
+            "clone",
+            "https://github.com/ARM-software/ComputeLibrary.git",
+            "-b",
+            "v25.02",
+            "--depth",
+            "1",
+            "--shallow-submodules",
+        ]
+    )

    check_call(
-        ["scons", "Werror=1", f"-j{os.cpu_count()}"] + acl_build_flags,
+        ["scons", "Werror=1", "-j8", f"build_dir=/{acl_install_dir}/build"]
+        + acl_build_flags,
        cwd=acl_checkout_dir,
    )
-    for d in ["arm_compute", "include", "utils", "support", "src", "build"]:
+    for d in ["arm_compute", "include", "utils", "support", "src"]:
        shutil.copytree(f"{acl_checkout_dir}/{d}", f"{acl_install_dir}/{d}")


-def replace_tag(filename) -> None:
-    with open(filename) as f:
-        lines = f.readlines()
-    for i, line in enumerate(lines):
-        if line.startswith("Tag:"):
-            lines[i] = line.replace("-linux_", "-manylinux_2_28_")
-            print(f"Updated tag from {line} to {lines[i]}")
-            break
-
-    with open(filename, "w") as f:
-        f.writelines(lines)
-
-
-def package_cuda_wheel(wheel_path, desired_cuda) -> None:
+def update_wheel(wheel_path, desired_cuda) -> None:
    """
-    Package the cuda wheel libraries
+    Update the cuda wheel libraries
    """
    folder = os.path.dirname(wheel_path)
    wheelname = os.path.basename(wheel_path)
@ -79,7 +65,6 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
    os.system(f"unzip {wheel_path} -d {folder}/tmp")
    libs_to_copy = [
        "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12",
-        "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so",
        "/usr/local/cuda/lib64/libcudnn.so.9",
        "/usr/local/cuda/lib64/libcublas.so.12",
        "/usr/local/cuda/lib64/libcublasLt.so.12",
@ -89,7 +74,7 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
        "/usr/local/cuda/lib64/libcusparseLt.so.0",
        "/usr/local/cuda/lib64/libcusolver.so.11",
        "/usr/local/cuda/lib64/libcurand.so.10",
-        "/usr/local/cuda/lib64/libnccl.so.2",
+        "/usr/local/cuda/lib64/libnvToolsExt.so.1",
        "/usr/local/cuda/lib64/libnvJitLink.so.12",
        "/usr/local/cuda/lib64/libnvrtc.so.12",
        "/usr/local/cuda/lib64/libcudnn_adv.so.9",
@ -103,19 +88,30 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
        "/usr/lib64/libgfortran.so.5",
        "/acl/build/libarm_compute.so",
        "/acl/build/libarm_compute_graph.so",
-        "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
-        "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
-        "/usr/local/lib/libnvpl_lapack_core.so.0",
-        "/usr/local/lib/libnvpl_blas_core.so.0",
    ]
-
-    if "129" in desired_cuda:
+    if enable_cuda:
        libs_to_copy += [
-            "/usr/local/cuda/lib64/libnvrtc-builtins.so.12.9",
-            "/usr/local/cuda/lib64/libcufile.so.0",
-            "/usr/local/cuda/lib64/libcufile_rdma.so.1",
+            "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
+            "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
+            "/usr/local/lib/libnvpl_lapack_core.so.0",
+            "/usr/local/lib/libnvpl_blas_core.so.0",
+        ]
+        if "126" in desired_cuda:
+            libs_to_copy += [
+                "/usr/local/cuda/lib64/libnvrtc-builtins.so.12.6",
+                "/usr/local/cuda/lib64/libcufile.so.0",
+                "/usr/local/cuda/lib64/libcufile_rdma.so.1",
+            ]
+        elif "128" in desired_cuda:
+            libs_to_copy += [
+                "/usr/local/cuda/lib64/libnvrtc-builtins.so.12.8",
+                "/usr/local/cuda/lib64/libcufile.so.0",
+                "/usr/local/cuda/lib64/libcufile_rdma.so.1",
+            ]
+    else:
+        libs_to_copy += [
+            "/opt/OpenBLAS/lib/libopenblas.so.0",
        ]
-
    # Copy libraries to unzipped_folder/a/lib
    for lib_path in libs_to_copy:
        lib_name = os.path.basename(lib_path)
@ -124,13 +120,6 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
            f"cd {folder}/tmp/torch/lib/; "
            f"patchelf --set-rpath '$ORIGIN' --force-rpath {folder}/tmp/torch/lib/{lib_name}"
        )
-
-    # Make sure the wheel is tagged with manylinux_2_28
-    for f in os.scandir(f"{folder}/tmp/"):
-        if f.is_dir() and f.name.endswith(".dist-info"):
-            replace_tag(f"{f.path}/WHEEL")
-            break
-
    os.mkdir(f"{folder}/cuda_wheel")
    os.system(f"cd {folder}/tmp/; zip -r {folder}/cuda_wheel/{wheelname} *")
    shutil.move(
@ -205,10 +194,8 @@ if __name__ == "__main__":
    ).decode()

    print("Building PyTorch wheel")
-    build_vars = "CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000 "
-    # MAX_JOB=5 is not required for CPU backend (see commit 465d98b)
-    if enable_cuda:
-        build_vars = "MAX_JOBS=5 " + build_vars
+    build_vars = "MAX_JOBS=5 CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000 "
+    os.system("cd /pytorch; python setup.py clean")

    override_package_version = os.getenv("OVERRIDE_PACKAGE_VERSION")
    desired_cuda = os.getenv("DESIRED_CUDA")
@ -255,6 +242,6 @@ if __name__ == "__main__":
        print("Updating Cuda Dependency")
        filename = os.listdir("/pytorch/dist/")
        wheel_path = f"/pytorch/dist/{filename[0]}"
-        package_cuda_wheel(wheel_path, desired_cuda)
+        update_wheel(wheel_path, desired_cuda)
    pytorch_wheel_name = complete_wheel("/pytorch/")
    print(f"Build Complete. Created {pytorch_wheel_name}..")
--- a/.ci/caffe2/README.md
+++ b/.ci/caffe2/README.md
@ -10,3 +10,5 @@ example: `py2-cuda9.0-cudnn7-ubuntu16.04`. The Docker images that are
 built on Jenkins and are used in triggered builds already have this
 environment variable set in their manifest. Also see
 `./docker/jenkins/*/Dockerfile` and search for `BUILD_ENVIRONMENT`.
+
+Our Jenkins installation is located at https://ci.pytorch.org/jenkins/.
--- a/.ci/docker/almalinux/Dockerfile
+++ b/.ci/docker/almalinux/Dockerfile
@ -1,7 +1,7 @@
-ARG CUDA_VERSION=12.6
+ARG CUDA_VERSION=12.4
 ARG BASE_TARGET=cuda${CUDA_VERSION}
 ARG ROCM_IMAGE=rocm/dev-almalinux-8:6.3-complete
-FROM amd64/almalinux:8.10-20250519 as base
+FROM amd64/almalinux:8 as base

 ENV LC_ALL en_US.UTF-8
 ENV LANG en_US.UTF-8
@ -11,8 +11,6 @@ ARG DEVTOOLSET_VERSION=11

 RUN yum -y update
 RUN yum -y install epel-release
-# install glibc-langpack-en make sure en_US.UTF-8 locale is available
-RUN yum -y install glibc-langpack-en
 RUN yum install -y sudo wget curl perl util-linux xz bzip2 git patch which perl zlib-devel openssl-devel yum-utils autoconf automake make gcc-toolset-${DEVTOOLSET_VERSION}-toolchain
 # Just add everything as a safe.directory for git since these will be used in multiple places with git
 RUN git config --global --add safe.directory '*'
@ -52,6 +50,10 @@ ENV CUDA_VERSION=${CUDA_VERSION}
 # Make things in our path by default
 ENV PATH=/usr/local/cuda-${CUDA_VERSION}/bin:$PATH

+FROM cuda as cuda11.8
+RUN bash ./install_cuda.sh 11.8
+ENV DESIRED_CUDA=11.8
+
 FROM cuda as cuda12.6
 RUN bash ./install_cuda.sh 12.6
 ENV DESIRED_CUDA=12.6
@ -60,10 +62,6 @@ FROM cuda as cuda12.8
 RUN bash ./install_cuda.sh 12.8
 ENV DESIRED_CUDA=12.8

-FROM cuda as cuda12.9
-RUN bash ./install_cuda.sh 12.9
-ENV DESIRED_CUDA=12.9
-
 FROM ${ROCM_IMAGE} as rocm
 ENV PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
 ADD ./common/install_mkl.sh install_mkl.sh
@ -78,8 +76,7 @@ RUN bash ./install_mnist.sh
 FROM base as all_cuda
 COPY --from=cuda11.8  /usr/local/cuda-11.8 /usr/local/cuda-11.8
 COPY --from=cuda12.6  /usr/local/cuda-12.6 /usr/local/cuda-12.6
-COPY --from=cuda12.8  /usr/local/cuda-12.8 /usr/local/cuda-12.8
-COPY --from=cuda12.9  /usr/local/cuda-12.9 /usr/local/cuda-12.9
+COPY --from=cuda12.4  /usr/local/cuda-12.8 /usr/local/cuda-12.8

 # Final step
 FROM ${BASE_TARGET} as final
--- a/.ci/docker/build.sh
+++ b/.ci/docker/build.sh
@ -50,21 +50,30 @@ if [[ "$image" == *xla* ]]; then
  exit 0
 fi

-if [[ "$image" == *-jammy* ]]; then
+if [[ "$image" == *-focal* ]]; then
+  UBUNTU_VERSION=20.04
+elif [[ "$image" == *-jammy* ]]; then
  UBUNTU_VERSION=22.04
 elif [[ "$image" == *ubuntu* ]]; then
  extract_version_from_image_name ubuntu UBUNTU_VERSION
+elif [[ "$image" == *centos* ]]; then
+  extract_version_from_image_name centos CENTOS_VERSION
 fi

 if [ -n "${UBUNTU_VERSION}" ]; then
  OS="ubuntu"
+elif [ -n "${CENTOS_VERSION}" ]; then
+  OS="centos"
 else
  echo "Unable to derive operating system base..."
  exit 1
 fi

 DOCKERFILE="${OS}/Dockerfile"
-if [[ "$image" == *rocm* ]]; then
+# When using ubuntu - 22.04, start from Ubuntu docker image, instead of nvidia/cuda docker image.
+if [[ "$image" == *cuda* && "$UBUNTU_VERSION" != "22.04" ]]; then
+  DOCKERFILE="${OS}-cuda/Dockerfile"
+elif [[ "$image" == *rocm* ]]; then
  DOCKERFILE="${OS}-rocm/Dockerfile"
 elif [[ "$image" == *xpu* ]]; then
  DOCKERFILE="${OS}-xpu/Dockerfile"
@ -76,6 +85,9 @@ elif [[ "$image" == *linter* ]]; then
  DOCKERFILE="linter/Dockerfile"
 fi

+# CMake 3.18 is needed to support CUDA17 language variant
+CMAKE_VERSION=3.18.5
+
 _UCX_COMMIT=7bb2722ff2187a0cad557ae4a6afa090569f83fb
 _UCC_COMMIT=20eae37090a4ce1b32bcce6144ccad0b49943e0b
 if [[ "$image" == *rocm* ]]; then
@ -83,14 +95,12 @@ if [[ "$image" == *rocm* ]]; then
  _UCC_COMMIT=0c0fc21559835044ab107199e334f7157d6a0d3d
 fi

-tag=$(echo $image | awk -F':' '{print $2}')
-
 # It's annoying to rename jobs every time you want to rewrite a
 # configuration, so we hardcode everything here rather than do it
 # from scratch
-case "$tag" in
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11)
-    CUDA_VERSION=12.8.1
+case "$image" in
+  pytorch-linux-focal-cuda12.6-cudnn9-py3-gcc11)
+    CUDA_VERSION=12.6.3
    CUDNN_VERSION=9
    ANACONDA_PYTHON_VERSION=3.10
    GCC_VERSION=11
@ -98,10 +108,11 @@ case "$tag" in
    KATEX=yes
    UCX_COMMIT=${_UCX_COMMIT}
    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
    TRITON=yes
    ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.8.1
+  pytorch-linux-focal-cuda12.4-cudnn9-py3-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.4.1
    CUDNN_VERSION=9
    ANACONDA_PYTHON_VERSION=3.10
    GCC_VERSION=9
@ -109,11 +120,12 @@ case "$tag" in
    KATEX=yes
    UCX_COMMIT=${_UCX_COMMIT}
    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
    TRITON=yes
    INDUCTOR_BENCHMARKS=yes
    ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.8.1
+  pytorch-linux-focal-cuda12.4-cudnn9-py3.12-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.4.1
    CUDNN_VERSION=9
    ANACONDA_PYTHON_VERSION=3.12
    GCC_VERSION=9
@ -121,11 +133,12 @@ case "$tag" in
    KATEX=yes
    UCX_COMMIT=${_UCX_COMMIT}
    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
    TRITON=yes
    INDUCTOR_BENCHMARKS=yes
    ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.8.1
+  pytorch-linux-focal-cuda12.4-cudnn9-py3.13-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.4.1
    CUDNN_VERSION=9
    ANACONDA_PYTHON_VERSION=3.13
    GCC_VERSION=9
@ -133,10 +146,11 @@ case "$tag" in
    KATEX=yes
    UCX_COMMIT=${_UCX_COMMIT}
    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
    TRITON=yes
    INDUCTOR_BENCHMARKS=yes
    ;;
-  pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9)
+  pytorch-linux-focal-cuda12.6-cudnn9-py3-gcc9)
    CUDA_VERSION=12.6.3
    CUDNN_VERSION=9
    ANACONDA_PYTHON_VERSION=3.10
@ -145,10 +159,11 @@ case "$tag" in
    KATEX=yes
    UCX_COMMIT=${_UCX_COMMIT}
    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
    TRITON=yes
    ;;
-  pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.6
+  pytorch-linux-focal-cuda12.6-cudnn9-py3-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6.3
    CUDNN_VERSION=9
    ANACONDA_PYTHON_VERSION=3.10
    GCC_VERSION=9
@ -156,11 +171,12 @@ case "$tag" in
    KATEX=yes
    UCX_COMMIT=${_UCX_COMMIT}
    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
    TRITON=yes
    INDUCTOR_BENCHMARKS=yes
    ;;
-  pytorch-linux-jammy-cuda12.6-cudnn9-py3.12-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.6
+  pytorch-linux-focal-cuda12.6-cudnn9-py3.12-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6.3
    CUDNN_VERSION=9
    ANACONDA_PYTHON_VERSION=3.12
    GCC_VERSION=9
@ -168,11 +184,12 @@ case "$tag" in
    KATEX=yes
    UCX_COMMIT=${_UCX_COMMIT}
    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
    TRITON=yes
    INDUCTOR_BENCHMARKS=yes
    ;;
-  pytorch-linux-jammy-cuda12.6-cudnn9-py3.13-gcc9-inductor-benchmarks)
-    CUDA_VERSION=12.6
+  pytorch-linux-focal-cuda12.6-cudnn9-py3.13-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6.3
    CUDNN_VERSION=9
    ANACONDA_PYTHON_VERSION=3.13
    GCC_VERSION=9
@ -180,11 +197,12 @@ case "$tag" in
    KATEX=yes
    UCX_COMMIT=${_UCX_COMMIT}
    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
    TRITON=yes
    INDUCTOR_BENCHMARKS=yes
    ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9)
-    CUDA_VERSION=12.8.1
+  pytorch-linux-focal-cuda11.8-cudnn9-py3-gcc9)
+    CUDA_VERSION=11.8.0
    CUDNN_VERSION=9
    ANACONDA_PYTHON_VERSION=3.10
    GCC_VERSION=9
@ -192,55 +210,71 @@ case "$tag" in
    KATEX=yes
    UCX_COMMIT=${_UCX_COMMIT}
    UCC_COMMIT=${_UCC_COMMIT}
+    CONDA_CMAKE=yes
    TRITON=yes
    ;;
-  pytorch-linux-jammy-py3-clang12-onnx)
+  pytorch-linux-focal-py3-clang10-onnx)
    ANACONDA_PYTHON_VERSION=3.9
-    CLANG_VERSION=12
+    CLANG_VERSION=10
    VISION=yes
+    CONDA_CMAKE=yes
    ONNX=yes
    ;;
-  pytorch-linux-jammy-py3.9-clang12)
+  pytorch-linux-focal-py3.9-clang10)
    ANACONDA_PYTHON_VERSION=3.9
-    CLANG_VERSION=12
+    CLANG_VERSION=10
    VISION=yes
+    CONDA_CMAKE=yes
    TRITON=yes
    ;;
-  pytorch-linux-jammy-py3.11-clang12)
+  pytorch-linux-focal-py3.11-clang10)
    ANACONDA_PYTHON_VERSION=3.11
-    CLANG_VERSION=12
+    CLANG_VERSION=10
    VISION=yes
+    CONDA_CMAKE=yes
    TRITON=yes
    ;;
-  pytorch-linux-jammy-py3.9-gcc9)
+  pytorch-linux-focal-py3.9-gcc9)
    ANACONDA_PYTHON_VERSION=3.9
    GCC_VERSION=9
    VISION=yes
+    CONDA_CMAKE=yes
    TRITON=yes
    ;;
-  pytorch-linux-jammy-rocm-n-1-py3)
+  pytorch-linux-focal-rocm-n-1-py3)
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=11
+    VISION=yes
+    ROCM_VERSION=6.2.4
+    NINJA_VERSION=1.9.0
+    CONDA_CMAKE=yes
+    TRITON=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-focal-rocm-n-py3)
    ANACONDA_PYTHON_VERSION=3.10
    GCC_VERSION=11
    VISION=yes
    ROCM_VERSION=6.3
    NINJA_VERSION=1.9.0
+    CONDA_CMAKE=yes
    TRITON=yes
    KATEX=yes
    UCX_COMMIT=${_UCX_COMMIT}
    UCC_COMMIT=${_UCC_COMMIT}
    INDUCTOR_BENCHMARKS=yes
    ;;
-  pytorch-linux-jammy-rocm-n-py3)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-xpu-2024.0-py3)
+    ANACONDA_PYTHON_VERSION=3.9
    GCC_VERSION=11
    VISION=yes
-    ROCM_VERSION=6.4
+    XPU_VERSION=0.5
    NINJA_VERSION=1.9.0
+    CONDA_CMAKE=yes
    TRITON=yes
-    KATEX=yes
-    UCX_COMMIT=${_UCX_COMMIT}
-    UCC_COMMIT=${_UCC_COMMIT}
-    INDUCTOR_BENCHMARKS=yes
    ;;
  pytorch-linux-jammy-xpu-2025.0-py3)
    ANACONDA_PYTHON_VERSION=3.9
@ -248,14 +282,7 @@ case "$tag" in
    VISION=yes
    XPU_VERSION=2025.0
    NINJA_VERSION=1.9.0
-    TRITON=yes
-    ;;
-  pytorch-linux-jammy-xpu-2025.1-py3)
-    ANACONDA_PYTHON_VERSION=3.9
-    GCC_VERSION=11
-    VISION=yes
-    XPU_VERSION=2025.1
-    NINJA_VERSION=1.9.0
+    CONDA_CMAKE=yes
    TRITON=yes
    ;;
    pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
@ -263,21 +290,36 @@ case "$tag" in
    GCC_VERSION=11
    VISION=yes
    KATEX=yes
+    CONDA_CMAKE=yes
    TRITON=yes
    DOCS=yes
    INDUCTOR_BENCHMARKS=yes
    ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12)
+  pytorch-linux-jammy-cuda11.8-cudnn9-py3.9-clang12)
    ANACONDA_PYTHON_VERSION=3.9
-    CUDA_VERSION=12.8.1
+    CUDA_VERSION=11.8
    CUDNN_VERSION=9
    CLANG_VERSION=12
    VISION=yes
    TRITON=yes
    ;;
+  pytorch-linux-jammy-py3-clang12-asan)
+    ANACONDA_PYTHON_VERSION=3.9
+    CLANG_VERSION=12
+    VISION=yes
+    CONDA_CMAKE=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-py3-clang15-asan)
+    ANACONDA_PYTHON_VERSION=3.10
+    CLANG_VERSION=15
+    CONDA_CMAKE=yes
+    VISION=yes
+    ;;
  pytorch-linux-jammy-py3-clang18-asan)
    ANACONDA_PYTHON_VERSION=3.10
    CLANG_VERSION=18
+    CONDA_CMAKE=yes
    VISION=yes
    ;;
  pytorch-linux-jammy-py3.9-gcc11)
@ -285,6 +327,7 @@ case "$tag" in
    GCC_VERSION=11
    VISION=yes
    KATEX=yes
+    CONDA_CMAKE=yes
    TRITON=yes
    DOCS=yes
    UNINSTALL_DILL=yes
@ -292,12 +335,14 @@ case "$tag" in
  pytorch-linux-jammy-py3-clang12-executorch)
    ANACONDA_PYTHON_VERSION=3.10
    CLANG_VERSION=12
+    CONDA_CMAKE=yes
    EXECUTORCH=yes
    ;;
  pytorch-linux-jammy-py3.12-halide)
    CUDA_VERSION=12.6
    ANACONDA_PYTHON_VERSION=3.12
    GCC_VERSION=11
+    CONDA_CMAKE=yes
    HALIDE=yes
    TRITON=yes
    ;;
@ -305,23 +350,27 @@ case "$tag" in
    CUDA_VERSION=12.6
    ANACONDA_PYTHON_VERSION=3.12
    GCC_VERSION=11
+    CONDA_CMAKE=yes
    TRITON_CPU=yes
    ;;
-  pytorch-linux-jammy-linter)
+  pytorch-linux-focal-linter)
    # TODO: Use 3.9 here because of this issue https://github.com/python/mypy/issues/13627.
    # We will need to update mypy version eventually, but that's for another day. The task
    # would be to upgrade mypy to 1.0.0 with Python 3.11
    PYTHON_VERSION=3.9
+    PIP_CMAKE=yes
    ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter)
+  pytorch-linux-jammy-cuda11.8-cudnn9-py3.9-linter)
    PYTHON_VERSION=3.9
-    CUDA_VERSION=12.8.1
+    CUDA_VERSION=11.8
+    PIP_CMAKE=yes
    ;;
  pytorch-linux-jammy-aarch64-py3.10-gcc11)
    ANACONDA_PYTHON_VERSION=3.10
    GCC_VERSION=11
    ACL=yes
    VISION=yes
+    CONDA_CMAKE=yes
    # snadampal: skipping llvm src build install because the current version
    # from pytorch/llvm:9.0.1 is x86 specific
    SKIP_LLVM_SRC_BUILD_INSTALL=yes
@ -331,6 +380,7 @@ case "$tag" in
    GCC_VERSION=11
    ACL=yes
    VISION=yes
+    CONDA_CMAKE=yes
    # snadampal: skipping llvm src build install because the current version
    # from pytorch/llvm:9.0.1 is x86 specific
    SKIP_LLVM_SRC_BUILD_INSTALL=yes
@ -353,7 +403,8 @@ case "$tag" in
      TRITON=yes
      # To ensure that any ROCm config will build using conda cmake
      # and thus have LAPACK/MKL enabled
-      fi
+      CONDA_CMAKE=yes
+    fi
    if [[ "$image" == *centos7* ]]; then
      NINJA_VERSION=1.10.2
    fi
@ -369,11 +420,22 @@ case "$tag" in
    if [[ "$image" == *glibc* ]]; then
      extract_version_from_image_name glibc GLIBC_VERSION
    fi
+    if [[ "$image" == *cmake* ]]; then
+      extract_version_from_image_name cmake CMAKE_VERSION
+    fi
  ;;
 esac

 tmp_tag=$(basename "$(mktemp -u)" | tr '[:upper:]' '[:lower:]')

+#when using cudnn version 8 install it separately from cuda
+if [[ "$image" == *cuda*  && ${OS} == "ubuntu" ]]; then
+  IMAGE_NAME="nvidia/cuda:${CUDA_VERSION}-cudnn${CUDNN_VERSION}-devel-ubuntu${UBUNTU_VERSION}"
+  if [[ ${CUDNN_VERSION} == 9 ]]; then
+    IMAGE_NAME="nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}"
+  fi
+fi
+
 no_cache_flag=""
 progress_flag=""
 # Do not use cache and progress=plain when in CI
@ -390,6 +452,7 @@ docker build \
       --build-arg "LLVMDEV=${LLVMDEV:-}" \
       --build-arg "VISION=${VISION:-}" \
       --build-arg "UBUNTU_VERSION=${UBUNTU_VERSION}" \
+       --build-arg "CENTOS_VERSION=${CENTOS_VERSION}" \
       --build-arg "DEVTOOLSET_VERSION=${DEVTOOLSET_VERSION}" \
       --build-arg "GLIBC_VERSION=${GLIBC_VERSION}" \
       --build-arg "CLANG_VERSION=${CLANG_VERSION}" \
@ -400,6 +463,7 @@ docker build \
       --build-arg "CUDNN_VERSION=${CUDNN_VERSION}" \
       --build-arg "TENSORRT_VERSION=${TENSORRT_VERSION}" \
       --build-arg "GRADLE_VERSION=${GRADLE_VERSION}" \
+       --build-arg "CMAKE_VERSION=${CMAKE_VERSION:-}" \
       --build-arg "NINJA_VERSION=${NINJA_VERSION:-}" \
       --build-arg "KATEX=${KATEX:-}" \
       --build-arg "ROCM_VERSION=${ROCM_VERSION:-}" \
@ -407,6 +471,8 @@ docker build \
       --build-arg "IMAGE_NAME=${IMAGE_NAME}" \
       --build-arg "UCX_COMMIT=${UCX_COMMIT}" \
       --build-arg "UCC_COMMIT=${UCC_COMMIT}" \
+       --build-arg "CONDA_CMAKE=${CONDA_CMAKE}" \
+       --build-arg "PIP_CMAKE=${PIP_CMAKE}" \
       --build-arg "TRITON=${TRITON}" \
       --build-arg "TRITON_CPU=${TRITON_CPU}" \
       --build-arg "ONNX=${ONNX}" \
@ -415,7 +481,6 @@ docker build \
       --build-arg "EXECUTORCH=${EXECUTORCH}" \
       --build-arg "HALIDE=${HALIDE}" \
       --build-arg "XPU_VERSION=${XPU_VERSION}" \
-       --build-arg "UNINSTALL_DILL=${UNINSTALL_DILL}" \
       --build-arg "ACL=${ACL:-}" \
       --build-arg "SKIP_SCCACHE_INSTALL=${SKIP_SCCACHE_INSTALL:-}" \
       --build-arg "SKIP_LLVM_SRC_BUILD_INSTALL=${SKIP_LLVM_SRC_BUILD_INSTALL:-}" \
@ -492,12 +557,3 @@ elif [ "$HAS_TRITON" = "yes" ]; then
  echo "expecting triton to not be installed, but it is"
  exit 1
 fi
-
-# Sanity check cmake version.  Executorch reinstalls cmake and I'm not sure if
-# they support 4.0.0 yet, so exclude them from this check.
-CMAKE_VERSION=$(drun cmake --version)
-if [[ "$EXECUTORCH" != *yes* && "$CMAKE_VERSION" != *4.* ]]; then
-  echo "CMake version is not 4.0.0:"
-  drun cmake --version
-  exit 1
-fi
--- a/.ci/docker/centos-rocm/Dockerfile
+++ b/.ci/docker/centos-rocm/Dockerfile
@ -17,8 +17,9 @@ RUN bash ./install_base.sh && rm install_base.sh
 # Update CentOS git version
 RUN yum -y remove git
 RUN yum -y remove git-*
-RUN yum -y install https://packages.endpointdev.com/rhel/7/os/x86_64/endpoint-repo-1.9-1.x86_64.rpm && \
-    sed -i 's/packages.endpoint/packages.endpointdev/' /etc/yum.repos.d/endpoint.repo
+RUN yum -y install https://packages.endpoint.com/rhel/7/os/x86_64/endpoint-repo-1.9-1.x86_64.rpm || \
+    (yum -y install https://packages.endpointdev.com/rhel/7/os/x86_64/endpoint-repo-1.9-1.x86_64.rpm && \
+    sed -i "s/packages.endpoint/packages.endpointdev/" /etc/yum.repos.d/endpoint.repo)
 RUN yum install -y git

 # Install devtoolset
@ -39,7 +40,7 @@ RUN bash ./install_user.sh && rm install_user.sh

 # Install conda and other packages (e.g., numpy, pytest)
 ARG ANACONDA_PYTHON_VERSION
-ARG BUILD_ENVIRONMENT
+ARG CONDA_CMAKE
 ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
 ENV PATH /opt/conda/envs/py_$ANACONDA_PYTHON_VERSION/bin:/opt/conda/bin:$PATH
 COPY requirements-ci.txt /opt/conda/requirements-ci.txt
@ -74,6 +75,12 @@ ENV MAGMA_HOME /opt/rocm/magma
 ENV LANG en_US.utf8
 ENV LC_ALL en_US.utf8

+# (optional) Install non-default CMake version
+ARG CMAKE_VERSION
+COPY ./common/install_cmake.sh install_cmake.sh
+RUN if [ -n "${CMAKE_VERSION}" ]; then bash ./install_cmake.sh; fi
+RUN rm install_cmake.sh
+
 # (optional) Install non-default Ninja version
 ARG NINJA_VERSION
 COPY ./common/install_ninja.sh install_ninja.sh
--- a/.ci/docker/ci_commit_pins/executorch.txt
+++ b/.ci/docker/ci_commit_pins/executorch.txt
@ -1 +1 @@
-56392aa978594cc155fa8af48cd949f5b5f1823a
+a325c56864c5b25630c1ebda03c8f16a96a02a19
--- a/.ci/docker/ci_commit_pins/nccl-cu12.txt
+++ b/.ci/docker/ci_commit_pins/nccl-cu12.txt
@ -1 +1 @@
-v2.27.3-1
+v2.26.2-1
--- a/.ci/docker/ci_commit_pins/triton-xpu.txt
+++ b/.ci/docker/ci_commit_pins/triton-xpu.txt
@ -1 +1 @@
-ae324eeac8e102a2b40370e341460f3791353398
+0bcc8265e677e5321606a3311bf71470f14456a8
--- a/.ci/docker/ci_commit_pins/triton.txt
+++ b/.ci/docker/ci_commit_pins/triton.txt
@ -1 +1 @@
-c8757738a7418249896224430ce84888e8ecdd79
+96316ce50fade7e209553aba4898cd9b82aab83b
--- a/.ci/docker/common/install_base.sh
+++ b/.ci/docker/common/install_base.sh
@ -30,6 +30,18 @@ install_ubuntu() {
    maybe_libomp_dev=""
  fi

+  # HACK: UCC testing relies on libnccl library from NVIDIA repo, and version 2.16 crashes
+  # See https://github.com/pytorch/pytorch/pull/105260#issuecomment-1673399729
+  # TODO: Eliminate this hack, we should not relay on apt-get installation
+  # See https://github.com/pytorch/pytorch/issues/144768
+  if [[ "$UBUNTU_VERSION" == "20.04"* && "$CUDA_VERSION" == "11.8"* ]]; then
+    maybe_libnccl_dev="libnccl2=2.15.5-1+cuda11.8 libnccl-dev=2.15.5-1+cuda11.8 --allow-downgrades --allow-change-held-packages"
+  elif [[ "$UBUNTU_VERSION" == "20.04"* && "$CUDA_VERSION" == "12.4"* ]]; then
+    maybe_libnccl_dev="libnccl2=2.26.2-1+cuda12.4 libnccl-dev=2.26.2-1+cuda12.4 --allow-downgrades --allow-change-held-packages"
+  else
+    maybe_libnccl_dev=""
+  fi
+
  # Install common dependencies
  apt-get update
  # TODO: Some of these may not be necessary
@ -58,6 +70,7 @@ install_ubuntu() {
    libasound2-dev \
    libsndfile-dev \
    ${maybe_libomp_dev} \
+    ${maybe_libnccl_dev} \
    software-properties-common \
    wget \
    sudo \
--- a/.ci/docker/common/install_cache.sh
+++ b/.ci/docker/common/install_cache.sh
@ -9,7 +9,7 @@ install_ubuntu() {
  # Instead use lib and headers from OpenSSL1.1 installed in `install_openssl.sh``
  apt-get install -y cargo
  echo "Checking out sccache repo"
-  git clone https://github.com/mozilla/sccache -b v0.10.0
+  git clone https://github.com/mozilla/sccache -b v0.9.1
  cd sccache
  echo "Building sccache"
  cargo build --release
--- a/.ci/docker/common/install_cmake.sh
+++ b/.ci/docker/common/install_cmake.sh
@ -0,0 +1,31 @@
+#!/bin/bash
+
+set -ex
+
+[ -n "$CMAKE_VERSION" ]
+
+# Remove system cmake install so it won't get used instead
+ID=$(grep -oP '(?<=^ID=).+' /etc/os-release | tr -d '"')
+case "$ID" in
+  ubuntu)
+    apt-get remove cmake -y
+    ;;
+  centos)
+    yum remove cmake -y
+    ;;
+  *)
+    echo "Unable to determine OS..."
+    exit 1
+    ;;
+esac
+
+# Turn 3.6.3 into v3.6
+path=$(echo "${CMAKE_VERSION}" | sed -e 's/\([0-9].[0-9]\+\).*/v\1/')
+file="cmake-${CMAKE_VERSION}-Linux-x86_64.tar.gz"
+
+# Download and install specific CMake version in /usr/local
+pushd /tmp
+curl -Os --retry 3 "https://cmake.org/files/${path}/${file}"
+tar -C /usr/local --strip-components 1 --no-same-owner -zxf cmake-*.tar.gz
+rm -f cmake-*.tar.gz
+popd
--- a/.ci/docker/common/install_conda.sh
+++ b/.ci/docker/common/install_conda.sh
@ -6,8 +6,8 @@ set -ex
 if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
  BASE_URL="https://repo.anaconda.com/miniconda"
  CONDA_FILE="Miniconda3-latest-Linux-x86_64.sh"
-  if [[ $(uname -m) == "aarch64" ]] || [[ "$BUILD_ENVIRONMENT" == *xpu* ]] || [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
-    BASE_URL="https://github.com/conda-forge/miniforge/releases/latest/download"  # @lint-ignore
+  if [[ $(uname -m) == "aarch64" ]] || [[ "$BUILD_ENVIRONMENT" == *xpu* ]]; then
+    BASE_URL="https://github.com/conda-forge/miniforge/releases/latest/download"
    CONDA_FILE="Miniforge3-Linux-$(uname -m).sh"
  fi

@ -64,11 +64,6 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
  # which is provided in libstdcxx 12 and up.
  conda_install libstdcxx-ng=12.3.0 --update-deps -c conda-forge

-  # Miniforge installer doesn't install sqlite by default
-  if [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
-    conda_install sqlite
-  fi
-
  # Install PyTorch conda deps, as per https://github.com/pytorch/pytorch README
  if [[ $(uname -m) == "aarch64" ]]; then
    conda_install "openblas==0.3.29=*openmp*"
@ -80,6 +75,14 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
  # and libpython-static for torch deploy
  conda_install llvmdev=8.0.0 "libpython-static=${ANACONDA_PYTHON_VERSION}"

+  # Use conda cmake in some cases. Conda cmake will be newer than our supported
+  # min version (3.5 for xenial and 3.10 for bionic), so we only do it in those
+  # following builds that we know should use conda. Specifically, Ubuntu bionic
+  # and focal cannot find conda mkl with stock cmake, so we need a cmake from conda
+  if [ -n "${CONDA_CMAKE}" ]; then
+    pip_install cmake
+  fi
+
  # Magma package names are concatenation of CUDA major and minor ignoring revision
  # I.e. magma-cuda102 package corresponds to CUDA_VERSION=10.2 and CUDA_VERSION=10.2.89
  # Magma is installed from a tarball in the ossci-linux bucket into the conda env
--- a/.ci/docker/common/install_cpython.sh
+++ b/.ci/docker/common/install_cpython.sh
@ -3,10 +3,11 @@
 set -uex -o pipefail

 PYTHON_DOWNLOAD_URL=https://www.python.org/ftp/python
+PYTHON_DOWNLOAD_GITHUB_BRANCH=https://github.com/python/cpython/archive/refs/heads
 GET_PIP_URL=https://bootstrap.pypa.io/get-pip.py

 # Python versions to be installed in /opt/$VERSION_NO
-CPYTHON_VERSIONS=${CPYTHON_VERSIONS:-"3.9.0 3.10.1 3.11.0 3.12.0 3.13.0 3.13.0t 3.14.0 3.14.0t"}
+CPYTHON_VERSIONS=${CPYTHON_VERSIONS:-"3.9.0 3.10.1 3.11.0 3.12.0 3.13.0 3.13.0t"}

 function check_var {
    if [ -z "$1" ]; then
@ -23,8 +24,9 @@ function do_cpython_build {
    tar -xzf Python-$py_ver.tgz

    local additional_flags=""
-    if [[ "$py_ver" == *"t" ]]; then
+    if [ "$py_ver" == "3.13.0t" ]; then
        additional_flags=" --disable-gil"
+        mv cpython-3.13/ cpython-3.13t/
    fi

    pushd $py_folder
@ -74,20 +76,24 @@ function do_cpython_build {
 function build_cpython {
    local py_ver=$1
    check_var $py_ver
-    local py_suffix=$py_ver
-    local py_folder=$py_ver
+    check_var $PYTHON_DOWNLOAD_URL
+    local py_ver_folder=$py_ver

-    # Special handling for nogil
-    if [[ "${py_ver}" == *"t" ]]; then
-        py_suffix=${py_ver::-1}
-        py_folder=$py_suffix
+    if [ "$py_ver" = "3.13.0t" ]; then
+        PY_VER_SHORT="3.13"
+        PYT_VER_SHORT="3.13t"
+        check_var $PYTHON_DOWNLOAD_GITHUB_BRANCH
+        wget $PYTHON_DOWNLOAD_GITHUB_BRANCH/$PY_VER_SHORT.tar.gz -O Python-$py_ver.tgz
+        do_cpython_build $py_ver cpython-$PYT_VER_SHORT
+    elif [ "$py_ver" = "3.13.0" ]; then
+        PY_VER_SHORT="3.13"
+        check_var $PYTHON_DOWNLOAD_GITHUB_BRANCH
+        wget $PYTHON_DOWNLOAD_GITHUB_BRANCH/$PY_VER_SHORT.tar.gz -O Python-$py_ver.tgz
+        do_cpython_build $py_ver cpython-$PY_VER_SHORT
+    else
+        wget -q $PYTHON_DOWNLOAD_URL/$py_ver_folder/Python-$py_ver.tgz
+        do_cpython_build $py_ver Python-$py_ver
    fi
-    # Only b3 is available now
-    if [ "$py_suffix" == "3.14.0" ]; then
-        py_suffix="3.14.0b3"
-    fi
-    wget -q $PYTHON_DOWNLOAD_URL/$py_folder/Python-$py_suffix.tgz -O Python-$py_ver.tgz
-    do_cpython_build $py_ver Python-$py_suffix

    rm -f Python-$py_ver.tgz
 }
--- a/.ci/docker/common/install_cuda.sh
+++ b/.ci/docker/common/install_cuda.sh
@ -10,8 +10,6 @@ else
  arch_path='sbsa'
 fi

-NVSHMEM_VERSION=3.3.9
-
 function install_cuda {
  version=$1
  runfile=$2
@ -42,51 +40,40 @@ function install_cudnn {
  rm -rf tmp_cudnn
 }

-function install_nvshmem {
-  cuda_major_version=$1      # e.g. "12"
-  nvshmem_version=$2         # e.g. "3.3.9"
+function install_118 {
+    CUDNN_VERSION=9.1.0.70
+    echo "Installing CUDA 11.8 and cuDNN ${CUDNN_VERSION} and NCCL and cuSparseLt-0.4.0"
+    install_cuda 11.8.0 cuda_11.8.0_520.61.05_linux

-  case "${arch_path}" in
-    sbsa)
-      dl_arch="aarch64"
-      ;;
-    x86_64)
-      dl_arch="x64"
-      ;;
-    *)
-      dl_arch="${arch}"
-      ;;
-  esac
+    install_cudnn 11 $CUDNN_VERSION

-  tmpdir="tmp_nvshmem"
-  mkdir -p "${tmpdir}" && cd "${tmpdir}"
+    CUDA_VERSION=11.8 bash install_nccl.sh

-  # nvSHMEM license: https://docs.nvidia.com/nvshmem/api/sla.html
-  filename="libnvshmem_cuda${cuda_major_version}-linux-${arch_path}-${nvshmem_version}"
-  url="https://developer.download.nvidia.com/compute/redist/nvshmem/${nvshmem_version}/builds/cuda${cuda_major_version}/txz/agnostic/${dl_arch}/${filename}.tar.gz"
+    CUDA_VERSION=11.8 bash install_cusparselt.sh

-  # download, unpack, install
-  wget -q "${url}"
-  tar xf "${filename}.tar.gz"
-  cp -a "libnvshmem/include/"* /usr/local/include/
-  cp -a "libnvshmem/lib/"*     /usr/local/lib/
-
-  # cleanup
-  cd ..
-  rm -rf "${tmpdir}"
-
-  echo "nvSHMEM ${nvshmem_version} for CUDA ${cuda_major_version} (${arch_path}) installed."
+    ldconfig
 }

-
-function install_126 {
-  CUDNN_VERSION=9.10.2.21
-  echo "Installing CUDA 12.6.3 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
-  install_cuda 12.6.3 cuda_12.6.3_560.35.05_linux
+function install_124 {
+  CUDNN_VERSION=9.1.0.70
+  echo "Installing CUDA 12.4.1 and cuDNN ${CUDNN_VERSION} and NCCL and cuSparseLt-0.6.2"
+  install_cuda 12.4.1 cuda_12.4.1_550.54.15_linux

  install_cudnn 12 $CUDNN_VERSION

-  install_nvshmem 12 $NVSHMEM_VERSION
+  CUDA_VERSION=12.4 bash install_nccl.sh
+
+  CUDA_VERSION=12.4 bash install_cusparselt.sh
+
+  ldconfig
+}
+
+function install_126 {
+  CUDNN_VERSION=9.5.1.17
+  echo "Installing CUDA 12.6.3 and cuDNN ${CUDNN_VERSION} and NCCL and cuSparseLt-0.6.3"
+  install_cuda 12.6.3 cuda_12.6.3_560.35.05_linux
+
+  install_cudnn 12 $CUDNN_VERSION

  CUDA_VERSION=12.6 bash install_nccl.sh

@ -95,22 +82,69 @@ function install_126 {
  ldconfig
 }

-function install_129 {
-  CUDNN_VERSION=9.10.2.21
-  echo "Installing CUDA 12.9.1 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
-  # install CUDA 12.9.1 in the same container
-  install_cuda 12.9.1 cuda_12.9.1_575.57.08_linux
+function prune_118 {
+    echo "Pruning CUDA 11.8 and cuDNN"
+    #####################################################################################
+    # CUDA 11.8 prune static libs
+    #####################################################################################
+    export NVPRUNE="/usr/local/cuda-11.8/bin/nvprune"
+    export CUDA_LIB_DIR="/usr/local/cuda-11.8/lib64"

-  # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
-  install_cudnn 12 $CUDNN_VERSION
+    export GENCODE="-gencode arch=compute_35,code=sm_35 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90"
+    export GENCODE_CUDNN="-gencode arch=compute_35,code=sm_35 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90"

-  install_nvshmem 12 $NVSHMEM_VERSION
+    if [[ -n "$OVERRIDE_GENCODE" ]]; then
+        export GENCODE=$OVERRIDE_GENCODE
+    fi

-  CUDA_VERSION=12.9 bash install_nccl.sh
+    # all CUDA libs except CuDNN and CuBLAS (cudnn and cublas need arch 3.7 included)
+    ls $CUDA_LIB_DIR/ | grep "\.a" | grep -v "culibos" | grep -v "cudart" | grep -v "cudnn" | grep -v "cublas" | grep -v "metis"  \
+      | xargs -I {} bash -c \
+                "echo {} && $NVPRUNE $GENCODE $CUDA_LIB_DIR/{} -o $CUDA_LIB_DIR/{}"

-  CUDA_VERSION=12.9 bash install_cusparselt.sh
+    # prune CuDNN and CuBLAS
+    $NVPRUNE $GENCODE_CUDNN $CUDA_LIB_DIR/libcublas_static.a -o $CUDA_LIB_DIR/libcublas_static.a
+    $NVPRUNE $GENCODE_CUDNN $CUDA_LIB_DIR/libcublasLt_static.a -o $CUDA_LIB_DIR/libcublasLt_static.a

-  ldconfig
+    #####################################################################################
+    # CUDA 11.8 prune visual tools
+    #####################################################################################
+    export CUDA_BASE="/usr/local/cuda-11.8/"
+    rm -rf $CUDA_BASE/libnvvp $CUDA_BASE/nsightee_plugins $CUDA_BASE/nsight-compute-2022.3.0 $CUDA_BASE/nsight-systems-2022.4.2/
+}
+
+function prune_124 {
+  echo "Pruning CUDA 12.4"
+  #####################################################################################
+  # CUDA 12.4 prune static libs
+  #####################################################################################
+  export NVPRUNE="/usr/local/cuda-12.4/bin/nvprune"
+  export CUDA_LIB_DIR="/usr/local/cuda-12.4/lib64"
+
+  export GENCODE="-gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90"
+  export GENCODE_CUDNN="-gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90"
+
+  if [[ -n "$OVERRIDE_GENCODE" ]]; then
+      export GENCODE=$OVERRIDE_GENCODE
+  fi
+  if [[ -n "$OVERRIDE_GENCODE_CUDNN" ]]; then
+      export GENCODE_CUDNN=$OVERRIDE_GENCODE_CUDNN
+  fi
+
+  # all CUDA libs except CuDNN and CuBLAS
+  ls $CUDA_LIB_DIR/ | grep "\.a" | grep -v "culibos" | grep -v "cudart" | grep -v "cudnn" | grep -v "cublas" | grep -v "metis"  \
+      | xargs -I {} bash -c \
+                "echo {} && $NVPRUNE $GENCODE $CUDA_LIB_DIR/{} -o $CUDA_LIB_DIR/{}"
+
+  # prune CuDNN and CuBLAS
+  $NVPRUNE $GENCODE_CUDNN $CUDA_LIB_DIR/libcublas_static.a -o $CUDA_LIB_DIR/libcublas_static.a
+  $NVPRUNE $GENCODE_CUDNN $CUDA_LIB_DIR/libcublasLt_static.a -o $CUDA_LIB_DIR/libcublasLt_static.a
+
+  #####################################################################################
+  # CUDA 12.4 prune visual tools
+  #####################################################################################
+  export CUDA_BASE="/usr/local/cuda-12.4/"
+  rm -rf $CUDA_BASE/libnvvp $CUDA_BASE/nsightee_plugins $CUDA_BASE/nsight-compute-2024.1.0 $CUDA_BASE/nsight-systems-2023.4.4/
 }

 function prune_126 {
@ -149,15 +183,13 @@ function prune_126 {

 function install_128 {
  CUDNN_VERSION=9.8.0.87
-  echo "Installing CUDA 12.8.1 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
-  # install CUDA 12.8.1 in the same container
-  install_cuda 12.8.1 cuda_12.8.1_570.124.06_linux
+  echo "Installing CUDA 12.8.0 and cuDNN ${CUDNN_VERSION} and NCCL and cuSparseLt-0.6.3"
+  # install CUDA 12.8.0 in the same container
+  install_cuda 12.8.0 cuda_12.8.0_570.86.10_linux

  # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
  install_cudnn 12 $CUDNN_VERSION

-  install_nvshmem 12 $NVSHMEM_VERSION
-
  CUDA_VERSION=12.8 bash install_nccl.sh

  CUDA_VERSION=12.8 bash install_cusparselt.sh
@ -169,11 +201,13 @@ function install_128 {
 while test $# -gt 0
 do
    case "$1" in
-    12.6|12.6.*) install_126; prune_126
+    11.8) install_118; prune_118
        ;;
-    12.8|12.8.*) install_128;
+    12.4) install_124; prune_124
        ;;
-    12.9|12.9.*) install_129;
+    12.6) install_126; prune_126
+        ;;
+    12.8) install_128;
        ;;
    *) echo "bad argument $1"; exit 1
        ;;
--- a/.ci/docker/common/install_cudnn.sh
+++ b/.ci/docker/common/install_cudnn.sh
@ -4,10 +4,12 @@ if [[ -n "${CUDNN_VERSION}" ]]; then
    # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
    mkdir tmp_cudnn
    pushd tmp_cudnn
-    if [[ ${CUDA_VERSION:0:4} == "12.9" || ${CUDA_VERSION:0:4} == "12.8" ]]; then
-        CUDNN_NAME="cudnn-linux-x86_64-9.10.2.21_cuda12-archive"
+    if [[ ${CUDA_VERSION:0:4} == "12.8" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.8.0.87_cuda12-archive"
    elif [[ ${CUDA_VERSION:0:4} == "12.6" ]]; then
-        CUDNN_NAME="cudnn-linux-x86_64-9.10.2.21_cuda12-archive"
+        CUDNN_NAME="cudnn-linux-x86_64-9.5.1.17_cuda12-archive"
+    elif [[ ${CUDA_VERSION:0:2} == "12" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.1.0.70_cuda12-archive"
    elif [[ ${CUDA_VERSION:0:2} == "11" ]]; then
        CUDNN_NAME="cudnn-linux-x86_64-9.1.0.70_cuda11-archive"
    else
--- a/.ci/docker/common/install_cusparselt.sh
+++ b/.ci/docker/common/install_cusparselt.sh
@ -5,14 +5,25 @@ set -ex
 # cuSPARSELt license: https://docs.nvidia.com/cuda/cusparselt/license.html
 mkdir tmp_cusparselt && cd tmp_cusparselt

-if [[ ${CUDA_VERSION:0:4} =~ ^12\.[5-9]$ ]]; then
+if [[ ${CUDA_VERSION:0:4} =~ ^12\.[5-8]$ ]]; then
    arch_path='sbsa'
    export TARGETARCH=${TARGETARCH:-$(uname -m)}
    if [ ${TARGETARCH} = 'amd64' ] || [ "${TARGETARCH}" = 'x86_64' ]; then
        arch_path='x86_64'
    fi
-    CUSPARSELT_NAME="libcusparse_lt-linux-${arch_path}-0.7.1.0-archive"
+    CUSPARSELT_NAME="libcusparse_lt-linux-${arch_path}-0.6.3.2-archive"
    curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-${arch_path}/${CUSPARSELT_NAME}.tar.xz
+elif [[ ${CUDA_VERSION:0:4} == "12.4" ]]; then
+    arch_path='sbsa'
+    export TARGETARCH=${TARGETARCH:-$(uname -m)}
+    if [ ${TARGETARCH} = 'amd64' ] || [ "${TARGETARCH}" = 'x86_64' ]; then
+        arch_path='x86_64'
+    fi
+    CUSPARSELT_NAME="libcusparse_lt-linux-${arch_path}-0.6.2.3-archive"
+    curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-${arch_path}/${CUSPARSELT_NAME}.tar.xz
+elif [[ ${CUDA_VERSION:0:4} == "11.8" ]]; then
+    CUSPARSELT_NAME="libcusparse_lt-linux-x86_64-0.4.0.7-archive"
+    curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-x86_64/${CUSPARSELT_NAME}.tar.xz
 else
    echo "Not sure which libcusparselt version to install for this ${CUDA_VERSION}"
 fi
--- a/.ci/docker/common/install_executorch.sh
+++ b/.ci/docker/common/install_executorch.sh
@ -13,7 +13,7 @@ clone_executorch() {
  # and fetch the target commit
  pushd executorch
  git checkout "${EXECUTORCH_PINNED_COMMIT}"
-  git submodule update --init --recursive
+  git submodule update --init
  popd

  chown -R jenkins executorch
--- a/.ci/docker/common/install_inductor_benchmark_deps.sh
+++ b/.ci/docker/common/install_inductor_benchmark_deps.sh
@ -16,7 +16,7 @@ function install_timm() {

  pip_install "git+https://github.com/huggingface/pytorch-image-models@${commit}"
  # Clean up
-  conda_run pip uninstall -y torch torchvision triton
+  conda_run pip uninstall -y cmake torch torchvision triton
 }

 # Pango is needed for weasyprint which is needed for doctr
--- a/.ci/docker/common/install_onnx.sh
+++ b/.ci/docker/common/install_onnx.sh
@ -8,6 +8,16 @@ retry () {
    "$@" || (sleep 10 && "$@") || (sleep 20 && "$@") || (sleep 40 && "$@")
 }

+# A bunch of custom pip dependencies for ONNX
+pip_install \
+  beartype==0.15.0 \
+  filelock==3.9.0 \
+  flatbuffers==2.0 \
+  mock==5.0.1 \
+  ninja==1.10.2 \
+  networkx==2.5 \
+  numpy==1.24.2
+
 # ONNXRuntime should be installed before installing
 # onnx-weekly. Otherwise, onnx-weekly could be
 # overwritten by onnx.
@ -19,8 +29,12 @@ pip_install \
  transformers==4.36.2

 pip_install coloredlogs packaging
+
 pip_install onnxruntime==1.18.1
-pip_install onnxscript==0.3.1
+pip_install onnx==1.17.0
+pip_install onnxscript==0.2.2 --no-deps
+# required by onnxscript
+pip_install ml_dtypes

 # Cache the transformers model to be used later by ONNX tests. We need to run the transformers
 # package to download the model. By default, the model is cached at ~/.cache/huggingface/hub/
--- a/.ci/docker/common/install_openblas.sh
+++ b/.ci/docker/common/install_openblas.sh
@ -4,7 +4,8 @@
 set -ex

 cd /
-git clone https://github.com/OpenMathLib/OpenBLAS.git -b "${OPENBLAS_VERSION:-v0.3.29}" --depth 1 --shallow-submodules
+git clone https://github.com/OpenMathLib/OpenBLAS.git -b v0.3.29 --depth 1 --shallow-submodules
+

 OPENBLAS_BUILD_FLAGS="
 NUM_THREADS=128
--- a/.ci/docker/common/install_python.sh
+++ b/.ci/docker/common/install_python.sh
@ -13,3 +13,6 @@ source /var/lib/jenkins/ci_env/bin/activate

 python -mpip install --upgrade pip
 python -mpip install -r /opt/requirements-ci.txt
+if [ -n "${PIP_CMAKE}" ]; then
+  python -mpip install cmake==3.31.6
+fi
--- a/.ci/docker/common/install_rocm.sh
+++ b/.ci/docker/common/install_rocm.sh
@ -26,11 +26,6 @@ Pin: release o=repo.radeon.com
 Pin-Priority: 600
 EOF

-    # we want the patch version of 6.4 instead
-    if [[ $(ver $ROCM_VERSION) -eq $(ver 6.4) ]]; then
-        ROCM_VERSION="${ROCM_VERSION}.1"
-    fi
-
    # Add amdgpu repository
    UBUNTU_VERSION_NAME=`cat /etc/os-release | grep UBUNTU_CODENAME | awk -F= '{print $2}'`
    echo "deb [arch=amd64] https://repo.radeon.com/amdgpu/${ROCM_VERSION}/ubuntu ${UBUNTU_VERSION_NAME} main" > /etc/apt/sources.list.d/amdgpu.list
@ -71,29 +66,17 @@ EOF
    done

    # ROCm 6.3 had a regression where initializing static code objects had significant overhead
-    # ROCm 6.4 did not yet fix the regression, also HIP branch names are different
-    if [[ $(ver $ROCM_VERSION) -ge $(ver 6.3) ]] && [[ $(ver $ROCM_VERSION) -lt $(ver 7.0) ]]; then
-        if [[ $(ver $ROCM_VERSION) -eq $(ver 6.4.1) ]]; then
-            HIP_BRANCH=release/rocm-rel-6.4
-            VER_STR=6.4
-            VER_PATCH=.1
-        elif [[ $(ver $ROCM_VERSION) -eq $(ver 6.4) ]]; then
-            HIP_BRANCH=release/rocm-rel-6.4
-            VER_STR=6.4
-        elif [[ $(ver $ROCM_VERSION) -eq $(ver 6.3) ]]; then
-            HIP_BRANCH=rocm-6.3.x
-            VER_STR=6.3
-        fi
+    if [[ $(ver $ROCM_VERSION) -eq $(ver 6.3) ]]; then
        # clr build needs CppHeaderParser but can only find it using conda's python
        /opt/conda/bin/python -m pip install CppHeaderParser
-        git clone https://github.com/ROCm/HIP -b $HIP_BRANCH
+        git clone https://github.com/ROCm/HIP -b rocm-6.3.x
        HIP_COMMON_DIR=$(readlink -f HIP)
-        git clone https://github.com/jeffdaily/clr -b release/rocm-rel-${VER_STR}${VER_PATCH}-statco-hotfix
+        git clone https://github.com/jeffdaily/clr -b release/rocm-rel-6.3-statco-hotfix
        mkdir -p clr/build
        pushd clr/build
        cmake .. -DCLR_BUILD_HIP=ON -DHIP_COMMON_DIR=$HIP_COMMON_DIR
        make -j
-        cp hipamd/lib/libamdhip64.so.${VER_STR}.* /opt/rocm/lib/libamdhip64.so.${VER_STR}.*
+        cp hipamd/lib/libamdhip64.so.6.3.* /opt/rocm/lib/libamdhip64.so.6.3.*
        popd
        rm -rf HIP clr
    fi
--- a/.ci/docker/common/install_rocm_magma.sh
+++ b/.ci/docker/common/install_rocm_magma.sh
@ -5,12 +5,7 @@ set -eou pipefail

 function do_install() {
    rocm_version=$1
-    if [[ ${rocm_version} =~ ^[0-9]+\.[0-9]+\.[0-9]+$ ]]; then
-        # chop off any patch version
-        rocm_version="${rocm_version%.*}"
-    fi
-
-    rocm_version_nodot=${rocm_version//./}
+    rocm_version_nodot=${1//./}

    # Version 2.7.2 + ROCm related updates
    MAGMA_VERSION=a1625ff4d9bc362906bd01f805dbbe12612953f6
--- a/.ci/docker/common/install_triton.sh
+++ b/.ci/docker/common/install_triton.sh
@ -33,9 +33,11 @@ if [ -n "${UBUNTU_VERSION}" ];then
    apt-get install -y gpg-agent
 fi

-# Keep the current cmake and numpy version here, so we can reinstall them later
-CMAKE_VERSION=$(get_pip_version cmake)
-NUMPY_VERSION=$(get_pip_version numpy)
+if [ -n "${CONDA_CMAKE}" ]; then
+  # Keep the current cmake and numpy version here, so we can reinstall them later
+  CMAKE_VERSION=$(get_pip_version cmake)
+  NUMPY_VERSION=$(get_pip_version numpy)
+fi

 if [ -z "${MAX_JOBS}" ]; then
    export MAX_JOBS=$(nproc)
@ -51,12 +53,7 @@ as_jenkins git clone --recursive ${TRITON_REPO} triton
 cd triton
 as_jenkins git checkout ${TRITON_PINNED_COMMIT}
 as_jenkins git submodule update --init --recursive
-
-# Old versions of python have setup.py in ./python; newer versions have it in ./
-if [ ! -f setup.py ]; then
-  cd python
-fi
-
+cd python
 pip_install pybind11==2.13.6

 # TODO: remove patch setup.py once we have a proper fix for https://github.com/triton-lang/triton/issues/4527
@ -82,26 +79,21 @@ cp dist/*.whl /opt/triton
 # Install the wheel for docker builds that don't use multi stage
 pip_install dist/*.whl

-# TODO: This is to make sure that the same cmake and numpy version from install conda
-# script is used. Without this step, the newer cmake version (3.25.2) downloaded by
-# triton build step via pip will fail to detect conda MKL. Once that issue is fixed,
-# this can be removed.
-#
-# The correct numpy version also needs to be set here because conda claims that it
-# causes inconsistent environment.  Without this, conda will attempt to install the
-# latest numpy version, which fails ASAN tests with the following import error: Numba
-# needs NumPy 1.20 or less.
-# Note that we install numpy with pip as conda might not have the version we want
-if [ -n "${CMAKE_VERSION}" ]; then
-  pip_install "cmake==${CMAKE_VERSION}"
-fi
-if [ -n "${NUMPY_VERSION}" ]; then
-  pip_install "numpy==${NUMPY_VERSION}"
-fi
-
-# IMPORTANT: helion needs to be installed without dependencies.
-# It depends on torch and triton. We don't want to install
-# triton and torch from production on Docker CI images
-if [[ "$ANACONDA_PYTHON_VERSION" != 3.9* ]]; then
-  pip_install helion --no-deps
+if [ -n "${CONDA_CMAKE}" ]; then
+  # TODO: This is to make sure that the same cmake and numpy version from install conda
+  # script is used. Without this step, the newer cmake version (3.25.2) downloaded by
+  # triton build step via pip will fail to detect conda MKL. Once that issue is fixed,
+  # this can be removed.
+  #
+  # The correct numpy version also needs to be set here because conda claims that it
+  # causes inconsistent environment.  Without this, conda will attempt to install the
+  # latest numpy version, which fails ASAN tests with the following import error: Numba
+  # needs NumPy 1.20 or less.
+  # Note that we install numpy with pip as conda might not have the version we want
+  if [ -n "${CMAKE_VERSION}" ]; then
+    pip_install "cmake==${CMAKE_VERSION}"
+  fi
+  if [ -n "${NUMPY_VERSION}" ]; then
+    pip_install "numpy==${NUMPY_VERSION}"
+  fi
 fi
--- a/.ci/docker/common/install_xpu.sh
+++ b/.ci/docker/common/install_xpu.sh
@ -26,7 +26,7 @@ function install_ubuntu() {
    wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB \
        | gpg --dearmor > /usr/share/keyrings/oneapi-archive-keyring.gpg.gpg
    echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg.gpg] \
-        https://apt.repos.intel.com/oneapi all main" \
+        https://apt.repos.intel.com/${XPU_REPO_NAME} all main" \
        | tee /etc/apt/sources.list.d/oneAPI.list

    # Update the packages list and repository index
@ -74,7 +74,7 @@ function install_rhel() {
    tee > /etc/yum.repos.d/oneAPI.repo << EOF
 [oneAPI]
 name=Intel for Pytorch GPU dev repository
-baseurl=https://yum.repos.intel.com/oneapi
+baseurl=https://yum.repos.intel.com/${XPU_REPO_NAME}
 enabled=1
 gpgcheck=1
 repo_gpgcheck=1
@ -118,7 +118,7 @@ function install_sles() {
        https://repositories.intel.com/gpu/sles/${VERSION_SP}${XPU_DRIVER_VERSION}/unified/intel-gpu-${VERSION_SP}.repo
    rpm --import https://repositories.intel.com/gpu/intel-graphics.key
    # To add the online network network package repository for the Intel Support Packages
-    zypper addrepo https://yum.repos.intel.com/oneapi oneAPI
+    zypper addrepo https://yum.repos.intel.com/${XPU_REPO_NAME} oneAPI
    rpm --import https://yum.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB

    # The xpu-smi packages
@ -141,10 +141,10 @@ if [[ "${XPU_DRIVER_TYPE,,}" == "rolling" ]]; then
    XPU_DRIVER_VERSION=""
 fi

-# Default use Intel® oneAPI Deep Learning Essentials 2025.0
-if [[ "$XPU_VERSION" == "2025.1" ]]; then
-    XPU_PACKAGES="intel-deep-learning-essentials-2025.1"
-else
+XPU_REPO_NAME="intel-for-pytorch-gpu-dev"
+XPU_PACKAGES="intel-for-pytorch-gpu-dev-0.5 intel-pti-dev-0.9"
+if [[ "$XPU_VERSION" == "2025.0" ]]; then
+    XPU_REPO_NAME="oneapi"
    XPU_PACKAGES="intel-deep-learning-essentials-2025.0"
 fi

--- a/.ci/docker/libtorch/Dockerfile
+++ b/.ci/docker/libtorch/Dockerfile
@ -54,6 +54,16 @@ COPY ./ci_commit_pins/nccl-cu* /ci_commit_pins/
 COPY ./common/install_cusparselt.sh install_cusparselt.sh
 ENV CUDA_HOME /usr/local/cuda

+FROM cuda as cuda11.8
+RUN bash ./install_cuda.sh 11.8
+RUN bash ./install_magma.sh 11.8
+RUN ln -sf /usr/local/cuda-11.8 /usr/local/cuda
+
+FROM cuda as cuda12.4
+RUN bash ./install_cuda.sh 12.4
+RUN bash ./install_magma.sh 12.4
+RUN ln -sf /usr/local/cuda-12.4 /usr/local/cuda
+
 FROM cuda as cuda12.6
 RUN bash ./install_cuda.sh 12.6
 RUN bash ./install_magma.sh 12.6
@ -64,11 +74,6 @@ RUN bash ./install_cuda.sh 12.8
 RUN bash ./install_magma.sh 12.8
 RUN ln -sf /usr/local/cuda-12.8 /usr/local/cuda

-FROM cuda as cuda12.9
-RUN bash ./install_cuda.sh 12.9
-RUN bash ./install_magma.sh 12.9
-RUN ln -sf /usr/local/cuda-12.9 /usr/local/cuda
-
 FROM cpu as rocm
 ARG ROCM_VERSION
 ARG PYTORCH_ROCM_ARCH
--- a/.ci/docker/libtorch/build.sh
+++ b/.ci/docker/libtorch/build.sh
@ -39,10 +39,6 @@ case ${DOCKER_TAG_PREFIX} in
        DOCKER_GPU_BUILD_ARG=""
        ;;
    rocm*)
-        # we want the patch version of 6.4 instead
-        if [[ $(ver $GPU_ARCH_VERSION) -eq $(ver 6.4) ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.1"
-        fi
        BASE_TARGET=rocm
        GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete
        PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
--- a/.ci/docker/linter/Dockerfile
+++ b/.ci/docker/linter/Dockerfile
@ -16,6 +16,7 @@ RUN bash ./install_user.sh && rm install_user.sh

 # Install conda and other packages (e.g., numpy, pytest)
 ARG PYTHON_VERSION
+ARG PIP_CMAKE
 ENV PATH /var/lib/jenkins/ci_env/bin:$PATH
 ENV VIRTUAL_ENV /var/lib/jenkins/ci_env
 COPY requirements-ci.txt /opt/requirements-ci.txt
--- a/.ci/docker/manywheel/Dockerfile_2_28
+++ b/.ci/docker/manywheel/Dockerfile_2_28
@ -7,8 +7,8 @@ ENV LC_ALL en_US.UTF-8
 ENV LANG en_US.UTF-8
 ENV LANGUAGE en_US.UTF-8

-ARG DEVTOOLSET_VERSION=13
-RUN yum install -y sudo wget curl perl util-linux xz bzip2 git patch which perl zlib-devel yum-utils gcc-toolset-${DEVTOOLSET_VERSION}-gcc gcc-toolset-${DEVTOOLSET_VERSION}-gcc-c++ gcc-toolset-${DEVTOOLSET_VERSION}-gcc-gfortran gcc-toolset-${DEVTOOLSET_VERSION}-gdb
+ARG DEVTOOLSET_VERSION=11
+RUN yum install -y sudo wget curl perl util-linux xz bzip2 git patch which perl zlib-devel yum-utils gcc-toolset-${DEVTOOLSET_VERSION}-toolchain
 ENV PATH=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/bin:$PATH
 ENV LD_LIBRARY_PATH=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/lib64:/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/lib:$LD_LIBRARY_PATH

@ -26,14 +26,14 @@ ADD ./common/install_openssl.sh install_openssl.sh
 RUN bash ./install_openssl.sh && rm install_openssl.sh


-# remove unnecessary python versions
+# remove unncessary python versions
 RUN rm -rf /opt/python/cp26-cp26m /opt/_internal/cpython-2.6.9-ucs2
 RUN rm -rf /opt/python/cp26-cp26mu /opt/_internal/cpython-2.6.9-ucs4
 RUN rm -rf /opt/python/cp33-cp33m /opt/_internal/cpython-3.3.6
 RUN rm -rf /opt/python/cp34-cp34m /opt/_internal/cpython-3.4.6

 FROM base as cuda
-ARG BASE_CUDA_VERSION=12.6
+ARG BASE_CUDA_VERSION=11.8
 # Install CUDA
 ADD ./common/install_cuda.sh install_cuda.sh
 COPY ./common/install_nccl.sh install_nccl.sh
@ -47,7 +47,7 @@ ADD ./common/install_mkl.sh install_mkl.sh
 RUN bash ./install_mkl.sh && rm install_mkl.sh

 FROM base as magma
-ARG BASE_CUDA_VERSION=12.6
+ARG BASE_CUDA_VERSION=10.2
 # Install magma
 ADD ./common/install_magma.sh install_magma.sh
 RUN bash ./install_magma.sh ${BASE_CUDA_VERSION} && rm install_magma.sh
@ -64,7 +64,7 @@ ADD ./common/install_libpng.sh install_libpng.sh
 RUN bash ./install_libpng.sh && rm install_libpng.sh

 FROM ${GPU_IMAGE} as common
-ARG DEVTOOLSET_VERSION=13
+ARG DEVTOOLSET_VERSION=11
 ENV LC_ALL en_US.UTF-8
 ENV LANG en_US.UTF-8
 ENV LANGUAGE en_US.UTF-8
@ -87,12 +87,13 @@ RUN yum install -y \
        wget \
        which \
        xz \
-        glibc-langpack-en \
-        gcc-toolset-${DEVTOOLSET_VERSION}-gcc \
-        gcc-toolset-${DEVTOOLSET_VERSION}-gcc-c++ \
-        gcc-toolset-${DEVTOOLSET_VERSION}-gcc-gfortran \
-        gcc-toolset-${DEVTOOLSET_VERSION}-gdb
+        gcc-toolset-${DEVTOOLSET_VERSION}-toolchain \
+        glibc-langpack-en
+RUN yum install -y \
+    https://repo.ius.io/ius-release-el7.rpm \
+    https://ossci-linux.s3.amazonaws.com/epel-release-7-14.noarch.rpm

+RUN yum swap -y git git236-core
 # git236+ would refuse to run git commands in repos owned by other users
 # Which causes version check to fail, as pytorch repo is bind-mounted into the image
 # Override this behaviour by treating every folder as safe
@ -103,7 +104,6 @@ ENV SSL_CERT_FILE=/opt/_internal/certs.pem
 # Install LLVM version
 COPY --from=openssl            /opt/openssl                          /opt/openssl
 COPY --from=base               /opt/python                           /opt/python
-COPY --from=base               /usr/local/lib/                       /usr/local/lib/
 COPY --from=base               /opt/_internal                        /opt/_internal
 COPY --from=base               /usr/local/bin/auditwheel             /usr/local/bin/auditwheel
 COPY --from=intel              /opt/intel                            /opt/intel
@ -117,8 +117,8 @@ COPY --from=libpng             /usr/local/lib/pkgconfig              /usr/local/
 COPY --from=jni                /usr/local/include/jni.h              /usr/local/include/jni.h

 FROM common as cpu_final
-ARG BASE_CUDA_VERSION=12.6
-ARG DEVTOOLSET_VERSION=13
+ARG BASE_CUDA_VERSION=11.8
+ARG DEVTOOLSET_VERSION=11
 # Install Anaconda
 ADD ./common/install_conda_docker.sh install_conda.sh
 RUN bash ./install_conda.sh && rm install_conda.sh
@ -157,11 +157,8 @@ ENV ROCM_PATH /opt/rocm
 # and avoid 3.21.0 cmake+ninja issues with ninja inserting "-Wl,--no-as-needed" in LINK_FLAGS for static linker
 RUN python3 -m pip install --upgrade pip && \
    python3 -mpip install cmake==3.28.4
-# replace the libdrm in /opt/amdgpu with custom amdgpu.ids lookup path
 ADD ./common/install_rocm_drm.sh install_rocm_drm.sh
 RUN bash ./install_rocm_drm.sh && rm install_rocm_drm.sh
-# ROCm 6.4 rocm-smi depends on system drm.h header
-RUN yum install -y libdrm-devel
 ENV MKLROOT /opt/intel
 ADD ./common/install_rocm_magma.sh install_rocm_magma.sh
 RUN bash ./install_rocm_magma.sh ${ROCM_VERSION} && rm install_rocm_magma.sh
@ -175,6 +172,6 @@ ENV XPU_DRIVER_TYPE ROLLING
 RUN python3 -m pip install --upgrade pip && \
    python3 -mpip install cmake==3.28.4
 ADD ./common/install_xpu.sh install_xpu.sh
-ENV XPU_VERSION 2025.1
+ENV XPU_VERSION 2025.0
 RUN bash ./install_xpu.sh && rm install_xpu.sh
 RUN pushd /opt/_internal && tar -xJf static-libs-for-embedding-only.tar.xz && popd
--- a/.ci/docker/manywheel/Dockerfile_2_28_aarch64
+++ b/.ci/docker/manywheel/Dockerfile_2_28_aarch64
@ -1,8 +1,9 @@
 FROM quay.io/pypa/manylinux_2_28_aarch64 as base

-ARG GCCTOOLSET_VERSION=13
+# Graviton needs GCC 10 or above for the build. GCC12 is the default version in almalinux-8.
+ARG GCCTOOLSET_VERSION=11

-# Language variables
+# Language variabes
 ENV LC_ALL=en_US.UTF-8
 ENV LANG=en_US.UTF-8
 ENV LANGUAGE=en_US.UTF-8
@ -35,10 +36,7 @@ RUN yum install -y \
  yasm \
  zstd \
  sudo \
-  gcc-toolset-${GCCTOOLSET_VERSION}-gcc \
-  gcc-toolset-${GCCTOOLSET_VERSION}-gcc-c++ \
-  gcc-toolset-${GCCTOOLSET_VERSION}-gcc-gfortran \
-  gcc-toolset-${GCCTOOLSET_VERSION}-gdb
+  gcc-toolset-${GCCTOOLSET_VERSION}-toolchain

 # (optional) Install non-default Ninja version
 ARG NINJA_VERSION
@ -58,13 +56,12 @@ RUN git config --global --add safe.directory "*"

 FROM base as openblas
 # Install openblas
-ARG OPENBLAS_VERSION
 ADD ./common/install_openblas.sh install_openblas.sh
 RUN bash ./install_openblas.sh && rm install_openblas.sh

 FROM base as final

-# remove unnecessary python versions
+# remove unncessary python versions
 RUN rm -rf /opt/python/cp26-cp26m /opt/_internal/cpython-2.6.9-ucs2
 RUN rm -rf /opt/python/cp26-cp26mu /opt/_internal/cpython-2.6.9-ucs4
 RUN rm -rf /opt/python/cp33-cp33m /opt/_internal/cpython-3.3.6
--- a/.ci/docker/manywheel/Dockerfile_cuda_aarch64
+++ b/.ci/docker/manywheel/Dockerfile_cuda_aarch64
@ -1,7 +1,7 @@
 FROM quay.io/pypa/manylinux_2_28_aarch64 as base

 # Cuda ARM build needs gcc 11
-ARG DEVTOOLSET_VERSION=13
+ARG DEVTOOLSET_VERSION=11

 # Language variables
 ENV LC_ALL=en_US.UTF-8
@ -34,10 +34,7 @@ RUN yum install -y \
  zstd \
  libgomp \
  sudo \
-  gcc-toolset-${DEVTOOLSET_VERSION}-gcc \
-  gcc-toolset-${DEVTOOLSET_VERSION}-gcc-c++ \
-  gcc-toolset-${DEVTOOLSET_VERSION}-gcc-gfortran \
-  gcc-toolset-${DEVTOOLSET_VERSION}-gdb
+  gcc-toolset-${DEVTOOLSET_VERSION}-toolchain

 # Ensure the expected devtoolset is used
 ENV PATH=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/bin:$PATH
@ -60,7 +57,7 @@ RUN bash ./install_openssl.sh && rm install_openssl.sh
 ENV SSL_CERT_FILE=/opt/_internal/certs.pem

 FROM openssl as final
-# remove unnecessary python versions
+# remove unncessary python versions
 RUN rm -rf /opt/python/cp26-cp26m /opt/_internal/cpython-2.6.9-ucs2
 RUN rm -rf /opt/python/cp26-cp26mu /opt/_internal/cpython-2.6.9-ucs4
 RUN rm -rf /opt/python/cp33-cp33m /opt/_internal/cpython-3.3.6
--- a/.ci/docker/manywheel/Dockerfile_s390x
+++ b/.ci/docker/manywheel/Dockerfile_s390x
@ -5,9 +5,7 @@ ENV LC_ALL=C.UTF-8
 ENV LANG=C.UTF-8
 ENV LANGUAGE=C.UTF-8

-# there is a bugfix in gcc >= 14 for precompiled headers and s390x vectorization interaction.
-# with earlier gcc versions test/inductor/test_cpu_cpp_wrapper.py will fail.
-ARG DEVTOOLSET_VERSION=14
+ARG DEVTOOLSET_VERSION=13
 # Installed needed OS packages. This is to support all
 # the binary builds (torch, vision, audio, text, data)
 RUN yum -y install epel-release
@ -60,8 +58,7 @@ RUN yum install -y \
  libxslt-devel \
  libxml2-devel \
  openssl-devel \
-  valgrind \
-  ninja-build
+  valgrind

 ENV PATH=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/bin:$PATH
 ENV LD_LIBRARY_PATH=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/lib64:/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/lib:$LD_LIBRARY_PATH
@ -106,6 +103,9 @@ CMD ["/bin/bash"]
 # install test dependencies:
 # - grpcio requires system openssl, bundled crypto fails to build
 RUN dnf install -y \
+  protobuf-devel \
+  protobuf-c-devel \
+  protobuf-lite-devel \
  hdf5-devel \
  python3-h5py \
  git
@ -120,22 +120,15 @@ RUN python3 -mpip install cmake==3.28.0
 # so just build it from upstream repository.
 # h5py is dependency of onnxruntime_training.
 # h5py==3.11.0 builds with hdf5-devel 1.10.5 from repository.
-# h5py 3.11.0 doesn't build with numpy >= 2.3.0.
 # install newest flatbuffers version first:
 # for some reason old version is getting pulled in otherwise.
 # packaging package is required for onnxruntime wheel build.
 RUN pip3 install flatbuffers && \
-  pip3 install cython 'pkgconfig>=1.5.5' 'setuptools>=77' 'numpy<2.3.0' && \
-  pip3 install --no-build-isolation h5py==3.11.0 && \
+  pip3 install h5py==3.11.0 && \
  pip3 install packaging && \
  git clone https://github.com/microsoft/onnxruntime && \
  cd onnxruntime && git checkout v1.21.0 && \
  git submodule update --init --recursive && \
-  wget https://github.com/microsoft/onnxruntime/commit/f57db79743c4d1a3553aa05cf95bcd10966030e6.patch && \
-  patch -p1 < f57db79743c4d1a3553aa05cf95bcd10966030e6.patch && \
-  ./build.sh --config Release --parallel 0 --enable_pybind \
-  --build_wheel --enable_training --enable_training_apis \
-  --enable_training_ops --skip_tests --allow_running_as_root \
-  --compile_no_warning_as_error && \
+  ./build.sh --config Release --parallel 0 --enable_pybind --build_wheel --enable_training --enable_training_apis --enable_training_ops --skip_tests --allow_running_as_root && \
  pip3 install ./build/Linux/Release/dist/onnxruntime_training-*.whl && \
  cd .. && /bin/rm -rf ./onnxruntime
--- a/.ci/docker/manywheel/build.sh
+++ b/.ci/docker/manywheel/build.sh
@ -27,21 +27,19 @@ fi

 MANY_LINUX_VERSION=${MANY_LINUX_VERSION:-}
 DOCKERFILE_SUFFIX=${DOCKERFILE_SUFFIX:-}
-OPENBLAS_VERSION=${OPENBLAS_VERSION:-}

 case ${image} in
    manylinux2_28-builder:cpu)
        TARGET=cpu_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"
        ;;
    manylinux2_28_aarch64-builder:cpu-aarch64)
        TARGET=final
        GPU_IMAGE=arm64v8/almalinux:8
-        DOCKER_GPU_BUILD_ARG=" --build-arg DEVTOOLSET_VERSION=13 --build-arg NINJA_VERSION=1.12.1"
+        DOCKER_GPU_BUILD_ARG=" --build-arg DEVTOOLSET_VERSION=11 --build-arg NINJA_VERSION=1.12.1"
        MANY_LINUX_VERSION="2_28_aarch64"
-        OPENBLAS_VERSION="v0.3.29"
        ;;
    manylinuxcxx11-abi-builder:cpu-cxx11-abi)
        TARGET=final
@ -55,30 +53,20 @@ case ${image} in
        DOCKER_GPU_BUILD_ARG=""
        MANY_LINUX_VERSION="s390x"
        ;;
-    manylinux2_28-builder:cuda11*)
+    manylinux2_28-builder:cuda*)
        TARGET=cuda_final
        GPU_IMAGE=amd64/almalinux:8
        DOCKER_GPU_BUILD_ARG="--build-arg BASE_CUDA_VERSION=${GPU_ARCH_VERSION} --build-arg DEVTOOLSET_VERSION=11"
        MANY_LINUX_VERSION="2_28"
        ;;
-    manylinux2_28-builder:cuda12*)
-        TARGET=cuda_final
-        GPU_IMAGE=amd64/almalinux:8
-        DOCKER_GPU_BUILD_ARG="--build-arg BASE_CUDA_VERSION=${GPU_ARCH_VERSION} --build-arg DEVTOOLSET_VERSION=13"
-        MANY_LINUX_VERSION="2_28"
-        ;;
    manylinuxaarch64-builder:cuda*)
        TARGET=cuda_final
        GPU_IMAGE=amd64/almalinux:8
-        DOCKER_GPU_BUILD_ARG="--build-arg BASE_CUDA_VERSION=${GPU_ARCH_VERSION} --build-arg DEVTOOLSET_VERSION=13"
+        DOCKER_GPU_BUILD_ARG="--build-arg BASE_CUDA_VERSION=${GPU_ARCH_VERSION} --build-arg DEVTOOLSET_VERSION=11"
        MANY_LINUX_VERSION="aarch64"
        DOCKERFILE_SUFFIX="_cuda_aarch64"
        ;;
    manylinux2_28-builder:rocm*)
-        # we want the patch version of 6.4 instead
-        if [[ $(ver $GPU_ARCH_VERSION) -eq $(ver 6.4) ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.1"
-        fi
        TARGET=rocm_final
        MANY_LINUX_VERSION="2_28"
        DEVTOOLSET_VERSION="11"
@ -115,7 +103,6 @@ tmp_tag=$(basename "$(mktemp -u)" | tr '[:upper:]' '[:lower:]')
 DOCKER_BUILDKIT=1 docker build  \
    ${DOCKER_GPU_BUILD_ARG} \
    --build-arg "GPU_IMAGE=${GPU_IMAGE}" \
-    --build-arg "OPENBLAS_VERSION=${OPENBLAS_VERSION}" \
    --target "${TARGET}" \
    -t "${tmp_tag}" \
    $@ \
--- a/.ci/docker/manywheel/build_scripts/build.sh
+++ b/.ci/docker/manywheel/build_scripts/build.sh
@ -97,7 +97,7 @@ find /opt/_internal -type f -print0 \
    | xargs -0 -n1 strip --strip-unneeded 2>/dev/null || true
 # We do not need the Python test suites, or indeed the precompiled .pyc and
 # .pyo files. Partially cribbed from:
-#    https://github.com/docker-library/python/blob/master/3.4/slim/Dockerfile  # @lint-ignore
+#    https://github.com/docker-library/python/blob/master/3.4/slim/Dockerfile
 find /opt/_internal \
     \( -type d -a -name test -o -name tests \) \
  -o \( -type f -a -name '*.pyc' -o -name '*.pyo' \) \
--- a/.ci/docker/manywheel/build_scripts/build_utils.sh
+++ b/.ci/docker/manywheel/build_scripts/build_utils.sh
@ -2,7 +2,7 @@
 # Helper utilities for build
 # Script used only in CD pipeline

-OPENSSL_DOWNLOAD_URL=https://www.openssl.org/source/old/1.1.1/  # @lint-ignore
+OPENSSL_DOWNLOAD_URL=https://www.openssl.org/source/old/1.1.1/
 CURL_DOWNLOAD_URL=https://curl.se/download

 AUTOCONF_DOWNLOAD_URL=https://ftp.gnu.org/gnu/autoconf
--- a/.ci/docker/requirements-ci.txt
+++ b/.ci/docker/requirements-ci.txt
@ -41,11 +41,14 @@ fbscribelogger==0.1.7
 #Pinned versions: 0.1.6
 #test that import:

-flatbuffers==24.12.23
+flatbuffers==2.0 ; platform_machine != "s390x"
 #Description: cross platform serialization library
-#Pinned versions: 24.12.23
+#Pinned versions: 2.0
 #test that import:

+flatbuffers ; platform_machine == "s390x"
+#Description: cross platform serialization library; Newer version is required on s390x for new python version
+
 hypothesis==5.35.1
 # Pin hypothesis to avoid flakiness: https://github.com/pytorch/pytorch/issues/31136
 #Description: advanced library for generating parametrized tests
@ -90,10 +93,10 @@ librosa>=0.6.2 ; python_version < "3.11"
 #Pinned versions:
 #test that import:

-mypy==1.16.0
+mypy==1.14.0
 # Pin MyPy version because new errors are likely to appear with each release
 #Description: linter
-#Pinned versions: 1.16.0
+#Pinned versions: 1.14.0
 #test that import: test_typing.py, test_type_hints.py

 networkx==2.8.8
@ -163,10 +166,10 @@ pillow==11.0.0
 #Pinned versions: 10.3.0
 #test that import:

-protobuf==5.29.4
-#Description:  Google's data interchange format
-#Pinned versions: 5.29.4
-#test that import: test_tensorboard.py, test/onnx/*
+protobuf==3.20.2
+#Description:  Google’s data interchange format
+#Pinned versions: 3.20.1
+#test that import: test_tensorboard.py

 psutil
 #Description: information on running processes and system utilization
@ -334,12 +337,12 @@ sympy==1.13.3
 #Pinned versions:
 #test that import:

-onnx==1.18.0
-#Description: Required by onnx tests, and mypy and test_public_bindings.py when checking torch.onnx._internal
+onnx==1.17.0
+#Description: Required by mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:

-onnxscript==0.3.1
+onnxscript==0.2.2
 #Description: Required by mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
@ -376,13 +379,3 @@ dataclasses_json==0.6.7
 #Description: required for data pipeline and scripts under tools/stats
 #Pinned versions: 0.6.7
 #test that import:
-
-cmake==4.0.0
-#Description: required for building
-
-tlparse==0.3.30
-#Description: required for log parsing
-
-cuda-bindings>=12.0,<13.0 ; platform_machine != "s390x"
-#Description: required for testing CUDAGraph::raw_cuda_graph(). See https://nvidia.github.io/cuda-python/cuda-bindings/latest/support.html for how this version was chosen. Note "Any fix in the latest bindings would be backported to the prior major version" means that only the newest version of cuda-bindings will get fixes. Depending on the latest version of 12.x is okay because all 12.y versions will be supported via "CUDA minor version compatibility". Pytorch builds against 13.z versions of cuda toolkit work with 12.x versions of cuda-bindings as well because newer drivers work with old toolkits.
-#test that import: test_cuda.py
--- a/.ci/docker/requirements-docs.txt
+++ b/.ci/docker/requirements-docs.txt
@ -15,14 +15,9 @@ sphinxext-opengraph==0.9.1
 #Description: This is used to generate PyTorch docs
 #Pinned versions: 0.9.1

-sphinx_sitemap==2.6.0
-#Description: This is used to generate sitemap for PyTorch docs
-#Pinned versions: 2.6.0
-
-matplotlib==3.5.3 ; python_version < "3.13"
-matplotlib==3.6.3 ; python_version >= "3.13"
+matplotlib==3.5.3
 #Description: This is used to generate PyTorch docs
-#Pinned versions: 3.6.3 if python > 3.12. Otherwise 3.5.3.
+#Pinned versions: 3.5.3

 tensorboard==2.13.0 ; python_version < "3.13"
 tensorboard==2.18.0 ; python_version >= "3.13"
--- a/.ci/docker/triton_version.txt
+++ b/.ci/docker/triton_version.txt
@ -1 +1 @@
-3.3.1
+3.3.0
--- a/.ci/docker/triton_xpu_version.txt
+++ b/.ci/docker/triton_xpu_version.txt
@ -1 +0,0 @@
-3.4.0
--- a/.ci/docker/ubuntu-cuda/Dockerfile
+++ b/.ci/docker/ubuntu-cuda/Dockerfile
@ -0,0 +1,177 @@
+ARG UBUNTU_VERSION
+ARG CUDA_VERSION
+ARG IMAGE_NAME
+
+FROM ${IMAGE_NAME} as base
+
+ARG UBUNTU_VERSION
+ARG CUDA_VERSION
+
+ENV DEBIAN_FRONTEND noninteractive
+
+# Install common dependencies (so that this step can be cached separately)
+COPY ./common/install_base.sh install_base.sh
+RUN bash ./install_base.sh && rm install_base.sh
+
+# Install user
+COPY ./common/install_user.sh install_user.sh
+RUN bash ./install_user.sh && rm install_user.sh
+
+# Install katex
+ARG KATEX
+COPY ./common/install_docs_reqs.sh install_docs_reqs.sh
+RUN bash ./install_docs_reqs.sh && rm install_docs_reqs.sh
+
+# Install conda and other packages (e.g., numpy, pytest)
+ARG ANACONDA_PYTHON_VERSION
+ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
+ENV PATH /opt/conda/envs/py_$ANACONDA_PYTHON_VERSION/bin:/opt/conda/bin:$PATH
+ARG CONDA_CMAKE
+COPY requirements-ci.txt /opt/conda/requirements-ci.txt
+COPY ./common/install_conda.sh install_conda.sh
+COPY ./common/common_utils.sh common_utils.sh
+COPY ./common/install_magma_conda.sh install_magma_conda.sh
+RUN bash ./install_conda.sh && rm install_conda.sh install_magma_conda.sh common_utils.sh /opt/conda/requirements-ci.txt
+
+# Install gcc
+ARG GCC_VERSION
+COPY ./common/install_gcc.sh install_gcc.sh
+RUN bash ./install_gcc.sh && rm install_gcc.sh
+
+# Install clang
+ARG CLANG_VERSION
+COPY ./common/install_clang.sh install_clang.sh
+RUN bash ./install_clang.sh && rm install_clang.sh
+
+# (optional) Install vision packages like OpenCV
+ARG VISION
+COPY ./common/install_vision.sh ./common/cache_vision_models.sh ./common/common_utils.sh ./
+RUN if [ -n "${VISION}" ]; then bash ./install_vision.sh; fi
+RUN rm install_vision.sh cache_vision_models.sh common_utils.sh
+ENV INSTALLED_VISION ${VISION}
+
+# (optional) Install UCC
+ARG UCX_COMMIT
+ARG UCC_COMMIT
+ENV UCX_COMMIT $UCX_COMMIT
+ENV UCC_COMMIT $UCC_COMMIT
+ENV UCX_HOME /usr
+ENV UCC_HOME /usr
+ADD ./common/install_ucc.sh install_ucc.sh
+RUN if [ -n "${UCX_COMMIT}" ] && [ -n "${UCC_COMMIT}" ]; then bash ./install_ucc.sh; fi
+RUN rm install_ucc.sh
+
+COPY ./common/install_openssl.sh install_openssl.sh
+ENV OPENSSL_ROOT_DIR /opt/openssl
+RUN bash ./install_openssl.sh
+ENV OPENSSL_DIR /opt/openssl
+
+ARG INDUCTOR_BENCHMARKS
+ARG ANACONDA_PYTHON_VERSION
+ENV ANACONDA_PYTHON_VERSION=$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.txt huggingface.txt
+COPY ci_commit_pins/timm.txt timm.txt
+RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
+
+# (optional) Install non-default CMake version
+ARG CMAKE_VERSION
+COPY ./common/install_cmake.sh install_cmake.sh
+RUN if [ -n "${CMAKE_VERSION}" ]; then bash ./install_cmake.sh; fi
+RUN rm install_cmake.sh
+
+ARG TRITON
+
+FROM base as triton-builder
+# Install triton, this needs to be done before sccache because the latter will
+# try to reach out to S3, which docker build runners don't have access
+COPY ./common/install_triton.sh install_triton.sh
+COPY ./common/common_utils.sh common_utils.sh
+COPY ci_commit_pins/triton.txt triton.txt
+COPY triton_version.txt triton_version.txt
+RUN bash ./install_triton.sh
+
+FROM base as final
+COPY --from=triton-builder /opt/triton /opt/triton
+RUN if [ -n "${TRITON}" ]; then pip install /opt/triton/*.whl; chown -R jenkins:jenkins /opt/conda; fi
+RUN rm -rf /opt/triton
+
+ARG HALIDE
+# Build and install halide
+COPY ./common/install_halide.sh install_halide.sh
+COPY ./common/common_utils.sh common_utils.sh
+COPY ci_commit_pins/halide.txt halide.txt
+RUN if [ -n "${HALIDE}" ]; then bash ./install_halide.sh; fi
+RUN rm install_halide.sh common_utils.sh halide.txt
+
+# Install ccache/sccache (do this last, so we get priority in PATH)
+COPY ./common/install_cache.sh install_cache.sh
+ENV PATH /opt/cache/bin:$PATH
+# See https://github.com/pytorch/pytorch/issues/82174
+# TODO(sdym@fb.com):
+# check if this is needed after full off Xenial migration
+ENV CARGO_NET_GIT_FETCH_WITH_CLI true
+RUN bash ./install_cache.sh && rm install_cache.sh
+ENV CMAKE_CUDA_COMPILER_LAUNCHER=/opt/cache/bin/sccache
+
+# Add jni.h for java host build
+COPY ./common/install_jni.sh install_jni.sh
+COPY ./java/jni.h jni.h
+RUN bash ./install_jni.sh && rm install_jni.sh
+
+# Install Open MPI for CUDA
+COPY ./common/install_openmpi.sh install_openmpi.sh
+RUN if [ -n "${CUDA_VERSION}" ]; then bash install_openmpi.sh; fi
+RUN rm install_openmpi.sh
+
+# Include BUILD_ENVIRONMENT environment variable in image
+ARG BUILD_ENVIRONMENT
+ENV BUILD_ENVIRONMENT ${BUILD_ENVIRONMENT}
+
+# AWS specific CUDA build guidance
+ENV TORCH_CUDA_ARCH_LIST Maxwell
+ENV TORCH_NVCC_FLAGS "-Xfatbin -compress-all"
+ENV CUDA_PATH /usr/local/cuda
+
+# Install LLVM dev version (Defined in the pytorch/builder github repository)
+COPY --from=pytorch/llvm:9.0.1 /opt/llvm /opt/llvm
+
+# Install CUDNN
+ARG CUDNN_VERSION
+ARG CUDA_VERSION
+COPY ./common/install_cudnn.sh install_cudnn.sh
+RUN if [ -n "${CUDNN_VERSION}" ]; then bash install_cudnn.sh; fi
+RUN rm install_cudnn.sh
+
+# Install CUSPARSELT
+ARG CUDA_VERSION
+COPY ./common/install_cusparselt.sh install_cusparselt.sh
+RUN bash install_cusparselt.sh
+RUN rm install_cusparselt.sh
+
+# Install NCCL
+ARG CUDA_VERSION
+COPY ./common/install_nccl.sh install_nccl.sh
+COPY ./ci_commit_pins/nccl-cu* /ci_commit_pins/
+RUN bash install_nccl.sh
+RUN rm install_nccl.sh /ci_commit_pins/nccl-cu*
+ENV USE_SYSTEM_NCCL=1
+ENV NCCL_INCLUDE_DIR="/usr/local/cuda/include/"
+ENV NCCL_LIB_DIR="/usr/local/cuda/lib64/"
+
+# Install CUDSS
+ARG CUDA_VERSION
+COPY ./common/install_cudss.sh install_cudss.sh
+RUN bash install_cudss.sh
+RUN rm install_cudss.sh
+
+# Delete /usr/local/cuda-11.X/cuda-11.X symlinks
+RUN if [ -h /usr/local/cuda-11.6/cuda-11.6 ]; then rm /usr/local/cuda-11.6/cuda-11.6; fi
+RUN if [ -h /usr/local/cuda-11.7/cuda-11.7 ]; then rm /usr/local/cuda-11.7/cuda-11.7; fi
+RUN if [ -h /usr/local/cuda-12.1/cuda-12.1 ]; then rm /usr/local/cuda-12.1/cuda-12.1; fi
+RUN if [ -h /usr/local/cuda-12.4/cuda-12.4 ]; then rm /usr/local/cuda-12.4/cuda-12.4; fi
+
+USER jenkins
+CMD ["bash"]
--- a/.ci/docker/ubuntu-rocm/Dockerfile
+++ b/.ci/docker/ubuntu-rocm/Dockerfile
@ -25,9 +25,9 @@ RUN bash ./install_docs_reqs.sh && rm install_docs_reqs.sh

 # Install conda and other packages (e.g., numpy, pytest)
 ARG ANACONDA_PYTHON_VERSION
-ARG BUILD_ENVIRONMENT
 ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
 ENV PATH /opt/conda/envs/py_$ANACONDA_PYTHON_VERSION/bin:/opt/conda/bin:$PATH
+ARG CONDA_CMAKE
 COPY requirements-ci.txt /opt/conda/requirements-ci.txt
 COPY ./common/install_conda.sh install_conda.sh
 COPY ./common/common_utils.sh common_utils.sh
@ -101,6 +101,12 @@ COPY ci_commit_pins/timm.txt timm.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
 RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt

+# (optional) Install non-default CMake version
+ARG CMAKE_VERSION
+COPY ./common/install_cmake.sh install_cmake.sh
+RUN if [ -n "${CMAKE_VERSION}" ]; then bash ./install_cmake.sh; fi
+RUN rm install_cmake.sh
+
 # (optional) Install non-default Ninja version
 ARG NINJA_VERSION
 COPY ./common/install_ninja.sh install_ninja.sh
--- a/.ci/docker/ubuntu-xpu/Dockerfile
+++ b/.ci/docker/ubuntu-xpu/Dockerfile
@ -28,6 +28,7 @@ RUN bash ./install_docs_reqs.sh && rm install_docs_reqs.sh

 # Install conda and other packages (e.g., numpy, pytest)
 ARG ANACONDA_PYTHON_VERSION
+ARG CONDA_CMAKE
 ARG DOCS
 ARG BUILD_ENVIRONMENT
 ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
@ -72,7 +73,7 @@ ARG TRITON
 COPY ./common/install_triton.sh install_triton.sh
 COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/triton-xpu.txt triton-xpu.txt
-COPY triton_xpu_version.txt triton_version.txt
+COPY triton_version.txt triton_version.txt
 RUN if [ -n "${TRITON}" ]; then bash ./install_triton.sh; fi
 RUN rm install_triton.sh common_utils.sh triton-xpu.txt triton_version.txt

@ -83,6 +84,12 @@ RUN if [ -n "${VISION}" ]; then bash ./install_vision.sh; fi
 RUN rm install_vision.sh cache_vision_models.sh common_utils.sh
 ENV INSTALLED_VISION ${VISION}

+# (optional) Install non-default CMake version
+ARG CMAKE_VERSION
+COPY ./common/install_cmake.sh install_cmake.sh
+RUN if [ -n "${CMAKE_VERSION}" ]; then bash ./install_cmake.sh; fi
+RUN rm install_cmake.sh
+
 # (optional) Install non-default Ninja version
 ARG NINJA_VERSION
 COPY ./common/install_ninja.sh install_ninja.sh
--- a/.ci/docker/ubuntu/Dockerfile
+++ b/.ci/docker/ubuntu/Dockerfile
@ -28,6 +28,7 @@ RUN bash ./install_docs_reqs.sh && rm install_docs_reqs.sh

 # Install conda and other packages (e.g., numpy, pytest)
 ARG ANACONDA_PYTHON_VERSION
+ARG CONDA_CMAKE
 ARG DOCS
 ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
 ENV PATH /opt/conda/envs/py_$ANACONDA_PYTHON_VERSION/bin:/opt/conda/bin:$PATH
@ -81,6 +82,12 @@ RUN if [ -n "${VISION}" ]; then bash ./install_vision.sh; fi
 RUN rm install_vision.sh cache_vision_models.sh common_utils.sh
 ENV INSTALLED_VISION ${VISION}

+# (optional) Install non-default CMake version
+ARG CMAKE_VERSION
+COPY ./common/install_cmake.sh install_cmake.sh
+RUN if [ -n "${CMAKE_VERSION}" ]; then bash ./install_cmake.sh; fi
+RUN rm install_cmake.sh
+
 # (optional) Install non-default Ninja version
 ARG NINJA_VERSION
 COPY ./common/install_ninja.sh install_ninja.sh
--- a/.ci/magma/Makefile
+++ b/.ci/magma/Makefile
@ -1,7 +1,7 @@
 SHELL=/usr/bin/env bash

 DOCKER_CMD ?= docker
-DESIRED_CUDA ?= 12.8
+DESIRED_CUDA ?= 11.8
 DESIRED_CUDA_SHORT = $(subst .,,$(DESIRED_CUDA))
 PACKAGE_NAME = magma-cuda
 CUDA_ARCH_LIST ?= -gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90
@ -16,21 +16,15 @@ DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
 	magma/build_magma.sh

 .PHONY: all
-all: magma-cuda129
 all: magma-cuda128
 all: magma-cuda126
+all: magma-cuda118

 .PHONY:
 clean:
 	$(RM) -r magma-*
 	$(RM) -r output

-.PHONY: magma-cuda129
-magma-cuda129: DESIRED_CUDA := 12.9
-magma-cuda129: CUDA_ARCH_LIST += -gencode arch=compute_100,code=sm_100 -gencode arch=compute_120,code=sm_120
-magma-cuda129:
-	$(DOCKER_RUN)
-
 .PHONY: magma-cuda128
 magma-cuda128: DESIRED_CUDA := 12.8
 magma-cuda128: CUDA_ARCH_LIST += -gencode arch=compute_100,code=sm_100 -gencode arch=compute_120,code=sm_120
@ -41,3 +35,9 @@ magma-cuda128:
 magma-cuda126: DESIRED_CUDA := 12.6
 magma-cuda126:
 	$(DOCKER_RUN)
+
+.PHONY: magma-cuda118
+magma-cuda118: DESIRED_CUDA := 11.8
+magma-cuda118: CUDA_ARCH_LIST += -gencode arch=compute_37,code=sm_37
+magma-cuda118:
+	$(DOCKER_RUN)
--- a/.ci/manywheel/build_common.sh
+++ b/.ci/manywheel/build_common.sh
@ -18,10 +18,12 @@ retry () {
    $*  || (sleep 1 && $*) || (sleep 2 && $*) || (sleep 4 && $*) || (sleep 8 && $*)
 }

-PLATFORM=""
+PLATFORM="manylinux2014_x86_64"
 # TODO move this into the Docker images
 OS_NAME=$(awk -F= '/^NAME/{print $2}' /etc/os-release)
-if [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
+if [[ "$OS_NAME" == *"CentOS Linux"* ]]; then
+    retry yum install -q -y zip openssl
+elif [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
    retry yum install -q -y zip openssl
    PLATFORM="manylinux_2_28_x86_64"
 elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
@ -31,11 +33,9 @@ elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
    # Comment out nvidia repositories to prevent them from getting apt-get updated, see https://github.com/pytorch/pytorch/issues/74968
    # shellcheck disable=SC2046
    sed -i 's/.*nvidia.*/# &/' $(find /etc/apt/ -type f -name "*.list")
+
    retry apt-get update
    retry apt-get -y install zip openssl
-else
-    echo "Unknown OS: '$OS_NAME'"
-    exit 1
 fi

 # We use the package name to test the package by passing this to 'pip install'
@ -79,6 +79,8 @@ if [[ -e /opt/openssl ]]; then
    export CMAKE_INCLUDE_PATH="/opt/openssl/include":$CMAKE_INCLUDE_PATH
 fi

+
+
 mkdir -p /tmp/$WHEELHOUSE_DIR

 export PATCHELF_BIN=/usr/local/bin/patchelf
@ -97,7 +99,6 @@ if [[ -z "$PYTORCH_ROOT" ]]; then
    exit 1
 fi
 pushd "$PYTORCH_ROOT"
-retry pip install -q cmake
 python setup.py clean
 retry pip install -qr requirements.txt
 case ${DESIRED_PYTHON} in
@ -151,7 +152,7 @@ if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
    BUILD_LIBTORCH_WHL=0 BUILD_PYTHON_ONLY=1 \
    BUILD_LIBTORCH_CPU_WITH_DEBUG=$BUILD_DEBUG_INFO \
    USE_NCCL=${USE_NCCL} USE_RCCL=${USE_RCCL} USE_KINETO=${USE_KINETO} \
-    CMAKE_FRESH=1 python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR
+    python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR --cmake
    echo "Finished setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
 else
    time CMAKE_ARGS=${CMAKE_ARGS[@]} \
@ -320,8 +321,8 @@ for pkg in /$WHEELHOUSE_DIR/torch_no_python*.whl /$WHEELHOUSE_DIR/torch*linux*.w
            # ROCm workaround for roctracer dlopens
            if [[ "$DESIRED_CUDA" == *"rocm"* ]]; then
                patchedpath=$(fname_without_so_number $destpath)
-            # Keep the so number for XPU dependencies and libgomp.so.1 to avoid twice load
-            elif [[ "$DESIRED_CUDA" == *"xpu"* || "$filename" == "libgomp.so.1" ]]; then
+            # Keep the so number for XPU dependencies
+            elif [[ "$DESIRED_CUDA" == *"xpu"* ]]; then
                patchedpath=$destpath
            else
                patchedpath=$(fname_with_sha256 $destpath)
--- a/.ci/manywheel/build_cuda.sh
+++ b/.ci/manywheel/build_cuda.sh
@ -15,9 +15,6 @@ export INSTALL_TEST=0 # dont install test binaries into site-packages
 export USE_CUPTI_SO=0
 export USE_CUSPARSELT=${USE_CUSPARSELT:-1} # Enable if not disabled by libtorch build
 export USE_CUFILE=${USE_CUFILE:-1}
-export USE_SYSTEM_NCCL=1
-export NCCL_INCLUDE_DIR="/usr/local/cuda/include/"
-export NCCL_LIB_DIR="/usr/local/cuda/lib64/"

 # Keep an array of cmake variables to add to
 if [[ -z "$CMAKE_ARGS" ]]; then
@ -39,8 +36,10 @@ if [[ -n "$DESIRED_CUDA" ]]; then
    if [[ ${DESIRED_CUDA} =~ ^[0-9]+\.[0-9]+$ ]]; then
        CUDA_VERSION=${DESIRED_CUDA}
    else
-        # cu126, cu128 etc...
-        if [[ ${#DESIRED_CUDA} -eq 5 ]]; then
+        # cu90, cu92, cu100, cu101
+        if [[ ${#DESIRED_CUDA} -eq 4 ]]; then
+            CUDA_VERSION="${DESIRED_CUDA:2:1}.${DESIRED_CUDA:3:1}"
+        elif [[ ${#DESIRED_CUDA} -eq 5 ]]; then
            CUDA_VERSION="${DESIRED_CUDA:2:2}.${DESIRED_CUDA:4:1}"
        fi
    fi
@ -51,22 +50,24 @@ else
 fi

 cuda_version_nodot=$(echo $CUDA_VERSION | tr -d '.')
-EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")

+TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;7.5;8.0;8.6"
 case ${CUDA_VERSION} in
-    #removing sm_50-sm_70 as these architectures are deprecated in CUDA 12.8/9 and will be removed in future releases
    12.8)
-        TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;9.0;10.0;12.0"
-        ;;
-    12.9)
-        TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;9.0;10.0;12.0+PTX"
-        # WAR to resolve the ld error in libtorch build with CUDA 12.9
-        if [[ "$PACKAGE_TYPE" == "libtorch" ]]; then
-            TORCH_CUDA_ARCH_LIST="7.5;8.0;9.0;10.0;12.0+PTX"
-        fi
+        TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;9.0;10.0;12.0+PTX" #removing sm_50-sm_70 as these architectures are deprecated in CUDA 12.8 and will be removed in future releases
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
        ;;
    12.6)
-        TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;7.5;8.0;8.6;9.0"
+        TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0"
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
+        ;;
+    12.4)
+        TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};9.0"
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
+        ;;
+    11.8)
+        TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};3.7;9.0"
+        EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
        ;;
    *)
        echo "unknown cuda version $CUDA_VERSION"
@ -90,15 +91,14 @@ fi
 mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR" || true

 OS_NAME=$(awk -F= '/^NAME/{print $2}' /etc/os-release)
-if [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
+if [[ "$OS_NAME" == *"CentOS Linux"* ]]; then
+    LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
+elif [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
    LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
 elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
    LIBGOMP_PATH="/usr/lib64/libgomp.so.1"
 elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
    LIBGOMP_PATH="/usr/lib/x86_64-linux-gnu/libgomp.so.1"
-else
-    echo "Unknown OS: '$OS_NAME'"
-    exit 1
 fi

 DEPS_LIST=(
@ -108,12 +108,31 @@ DEPS_SONAME=(
    "libgomp.so.1"
 )

+# CUDA 11.8 have to ship the libcusparseLt.so.0 with the binary
+# since nvidia-cusparselt-cu11 is not available in PYPI
+if [[ $USE_CUSPARSELT == "1" && $CUDA_VERSION == "11.8" ]]; then
+        DEPS_SONAME+=(
+            "libcusparseLt.so.0"
+        )
+        DEPS_LIST+=(
+            "/usr/local/cuda/lib64/libcusparseLt.so.0"
+        )
+fi

-# CUDA_VERSION 12.6, 12.8, 12.9
+
+# Turn USE_CUFILE off for CUDA 11.8, 12.4 since nvidia-cufile-cu11 and 1.9.0.20 are
+# not available in PYPI
+if [[ $CUDA_VERSION == "11.8" || $CUDA_VERSION == "12.4" ]]; then
+    export USE_CUFILE=0
+fi
+
+
+# CUDA_VERSION 12.4, 12.6, 12.8
 if [[ $CUDA_VERSION == 12* ]]; then
    export USE_STATIC_CUDNN=0
    # Try parallelizing nvcc as well
    export TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
+
    if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
        echo "Bundling with cudnn and cublas."
        DEPS_LIST+=(
@ -129,12 +148,9 @@ if [[ $CUDA_VERSION == 12* ]]; then
            "/usr/local/cuda/lib64/libcublasLt.so.12"
            "/usr/local/cuda/lib64/libcusparseLt.so.0"
            "/usr/local/cuda/lib64/libcudart.so.12"
+            "/usr/local/cuda/lib64/libnvToolsExt.so.1"
            "/usr/local/cuda/lib64/libnvrtc.so.12"
            "/usr/local/cuda/lib64/libnvrtc-builtins.so"
-            "/usr/local/cuda/lib64/libcufile.so.0"
-            "/usr/local/cuda/lib64/libcufile_rdma.so.1"
-            "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12"
-            "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so"
        )
        DEPS_SONAME+=(
            "libcudnn_adv.so.9"
@ -149,17 +165,19 @@ if [[ $CUDA_VERSION == 12* ]]; then
            "libcublasLt.so.12"
            "libcusparseLt.so.0"
            "libcudart.so.12"
+            "libnvToolsExt.so.1"
            "libnvrtc.so.12"
            "libnvrtc-builtins.so"
-            "libcufile.so.0"
-            "libcufile_rdma.so.1"
-            "libcupti.so.12"
-            "libnvperf_host.so"
        )
-        # Add libnvToolsExt only if CUDA version is not 12.9
-        if [[ $CUDA_VERSION != 12.9* ]]; then
-            DEPS_LIST+=("/usr/local/cuda/lib64/libnvToolsExt.so.1")
-            DEPS_SONAME+=("libnvToolsExt.so.1")
+        if [[ $USE_CUFILE == 1 ]]; then
+            DEPS_LIST+=(
+                "/usr/local/cuda/lib64/libcufile.so.0"
+                "/usr/local/cuda/lib64/libcufile_rdma.so.1"
+            )
+            DEPS_SONAME+=(
+                "libcufile.so.0"
+                "libcufile_rdma.so.1"
+            )
        fi
    else
        echo "Using nvidia libs from pypi."
@ -173,21 +191,94 @@ if [[ $CUDA_VERSION == 12* ]]; then
            '$ORIGIN/../../nvidia/curand/lib'
            '$ORIGIN/../../nvidia/cusolver/lib'
            '$ORIGIN/../../nvidia/cusparse/lib'
-            '$ORIGIN/../../nvidia/cusparselt/lib'
            '$ORIGIN/../../cusparselt/lib'
            '$ORIGIN/../../nvidia/nccl/lib'
-            '$ORIGIN/../../nvidia/nvshmem/lib'
            '$ORIGIN/../../nvidia/nvtx/lib'
-            '$ORIGIN/../../nvidia/cufile/lib'
+        )
+        if [[ $USE_CUFILE == 1 ]]; then
+            CUDA_RPATHS+=(
+                '$ORIGIN/../../nvidia/cufile/lib'
+            )
+        fi
+        CUDA_RPATHS=$(IFS=: ; echo "${CUDA_RPATHS[*]}")
+        export C_SO_RPATH=$CUDA_RPATHS':$ORIGIN:$ORIGIN/lib'
+        export LIB_SO_RPATH=$CUDA_RPATHS':$ORIGIN'
+        export FORCE_RPATH="--force-rpath"
+        export USE_STATIC_NCCL=0
+        export USE_SYSTEM_NCCL=1
+        export ATEN_STATIC_CUDA=0
+        export USE_CUDA_STATIC_LINK=0
+        export USE_CUPTI_SO=1
+        export NCCL_INCLUDE_DIR="/usr/local/cuda/include/"
+        export NCCL_LIB_DIR="/usr/local/cuda/lib64/"
+    fi
+elif [[ $CUDA_VERSION == "11.8" ]]; then
+    export USE_STATIC_CUDNN=0
+    # Try parallelizing nvcc as well
+    export TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
+    # Bundle ptxas into the wheel, see https://github.com/pytorch/pytorch/pull/119750
+    export BUILD_BUNDLE_PTXAS=1
+
+    if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
+        echo "Bundling with cudnn and cublas."
+        DEPS_LIST+=(
+            "/usr/local/cuda/lib64/libcudnn_adv.so.9"
+            "/usr/local/cuda/lib64/libcudnn_cnn.so.9"
+            "/usr/local/cuda/lib64/libcudnn_graph.so.9"
+            "/usr/local/cuda/lib64/libcudnn_ops.so.9"
+            "/usr/local/cuda/lib64/libcudnn_engines_runtime_compiled.so.9"
+            "/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9"
+            "/usr/local/cuda/lib64/libcudnn_heuristic.so.9"
+            "/usr/local/cuda/lib64/libcudnn.so.9"
+            "/usr/local/cuda/lib64/libcublas.so.11"
+            "/usr/local/cuda/lib64/libcublasLt.so.11"
+            "/usr/local/cuda/lib64/libcudart.so.11.0"
+            "/usr/local/cuda/lib64/libnvToolsExt.so.1"
+            "/usr/local/cuda/lib64/libnvrtc.so.11.2"    # this is not a mistake, it links to more specific cuda version
+            "/usr/local/cuda/lib64/libnvrtc-builtins.so.11.8"
+        )
+        DEPS_SONAME+=(
+            "libcudnn_adv.so.9"
+            "libcudnn_cnn.so.9"
+            "libcudnn_graph.so.9"
+            "libcudnn_ops.so.9"
+            "libcudnn_engines_runtime_compiled.so.9"
+            "libcudnn_engines_precompiled.so.9"
+            "libcudnn_heuristic.so.9"
+            "libcudnn.so.9"
+            "libcublas.so.11"
+            "libcublasLt.so.11"
+            "libcudart.so.11.0"
+            "libnvToolsExt.so.1"
+            "libnvrtc.so.11.2"
+            "libnvrtc-builtins.so.11.8"
+        )
+    else
+        echo "Using nvidia libs from pypi."
+        CUDA_RPATHS=(
+            '$ORIGIN/../../nvidia/cublas/lib'
+            '$ORIGIN/../../nvidia/cuda_cupti/lib'
+            '$ORIGIN/../../nvidia/cuda_nvrtc/lib'
+            '$ORIGIN/../../nvidia/cuda_runtime/lib'
+            '$ORIGIN/../../nvidia/cudnn/lib'
+            '$ORIGIN/../../nvidia/cufft/lib'
+            '$ORIGIN/../../nvidia/curand/lib'
+            '$ORIGIN/../../nvidia/cusolver/lib'
+            '$ORIGIN/../../nvidia/cusparse/lib'
+            '$ORIGIN/../../nvidia/nccl/lib'
+            '$ORIGIN/../../nvidia/nvtx/lib'
        )
        CUDA_RPATHS=$(IFS=: ; echo "${CUDA_RPATHS[*]}")
        export C_SO_RPATH=$CUDA_RPATHS':$ORIGIN:$ORIGIN/lib'
        export LIB_SO_RPATH=$CUDA_RPATHS':$ORIGIN'
        export FORCE_RPATH="--force-rpath"
        export USE_STATIC_NCCL=0
+        export USE_SYSTEM_NCCL=1
        export ATEN_STATIC_CUDA=0
        export USE_CUDA_STATIC_LINK=0
        export USE_CUPTI_SO=1
+        export NCCL_INCLUDE_DIR="/usr/local/cuda/include/"
+        export NCCL_LIB_DIR="/usr/local/cuda/lib64/"
    fi
 else
    echo "Unknown cuda version $CUDA_VERSION"
--- a/.ci/manywheel/build_libtorch.sh
+++ b/.ci/manywheel/build_libtorch.sh
@ -22,7 +22,9 @@ retry () {

 # TODO move this into the Docker images
 OS_NAME=`awk -F= '/^NAME/{print $2}' /etc/os-release`
-if [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
+if [[ "$OS_NAME" == *"CentOS Linux"* ]]; then
+    retry yum install -q -y zip openssl
+elif [[ "$OS_NAME" == *"AlmaLinux"* ]]; then
    retry yum install -q -y zip openssl
 elif [[ "$OS_NAME" == *"Red Hat Enterprise Linux"* ]]; then
    retry dnf install -q -y zip openssl
@ -33,9 +35,6 @@ elif [[ "$OS_NAME" == *"Ubuntu"* ]]; then
    sed -i 's/.*nvidia.*/# &/' $(find /etc/apt/ -type f -name "*.list")
    retry apt-get update
    retry apt-get -y install zip openssl
-else
-    echo "Unknown OS: '$OS_NAME'"
-    exit 1
 fi

 # Version: setup.py uses $PYTORCH_BUILD_VERSION.post$PYTORCH_BUILD_NUMBER if
@ -92,7 +91,6 @@ if [[ -z "$PYTORCH_ROOT" ]]; then
    exit 1
 fi
 pushd "$PYTORCH_ROOT"
-retry pip install -q cmake
 python setup.py clean
 retry pip install -qr requirements.txt
 retry pip install -q numpy==2.0.1
--- a/.ci/manywheel/build_rocm.sh
+++ b/.ci/manywheel/build_rocm.sh
@ -95,7 +95,6 @@ ROCM_SO_FILES=(
    "libroctracer64.so"
    "libroctx64.so"
    "libhipblaslt.so"
-    "libhipsparselt.so"
    "libhiprtc.so"
 )

@ -187,28 +186,20 @@ do
    OS_SO_FILES[${#OS_SO_FILES[@]}]=$file_name # Append lib to array
 done

-ARCH=$(echo $PYTORCH_ROCM_ARCH | sed 's/;/|/g') # Replace ; separated arch list to bar for grep
-
 # rocBLAS library files
 ROCBLAS_LIB_SRC=$ROCM_HOME/lib/rocblas/library
 ROCBLAS_LIB_DST=lib/rocblas/library
-ROCBLAS_ARCH_SPECIFIC_FILES=$(ls $ROCBLAS_LIB_SRC | grep -E $ARCH)
-ROCBLAS_OTHER_FILES=$(ls $ROCBLAS_LIB_SRC | grep -v gfx)
-ROCBLAS_LIB_FILES=($ROCBLAS_ARCH_SPECIFIC_FILES $OTHER_FILES)
+ARCH=$(echo $PYTORCH_ROCM_ARCH | sed 's/;/|/g') # Replace ; seperated arch list to bar for grep
+ARCH_SPECIFIC_FILES=$(ls $ROCBLAS_LIB_SRC | grep -E $ARCH)
+OTHER_FILES=$(ls $ROCBLAS_LIB_SRC | grep -v gfx)
+ROCBLAS_LIB_FILES=($ARCH_SPECIFIC_FILES $OTHER_FILES)

 # hipblaslt library files
 HIPBLASLT_LIB_SRC=$ROCM_HOME/lib/hipblaslt/library
 HIPBLASLT_LIB_DST=lib/hipblaslt/library
-HIPBLASLT_ARCH_SPECIFIC_FILES=$(ls $HIPBLASLT_LIB_SRC | grep -E $ARCH)
-HIPBLASLT_OTHER_FILES=$(ls $HIPBLASLT_LIB_SRC | grep -v gfx)
-HIPBLASLT_LIB_FILES=($HIPBLASLT_ARCH_SPECIFIC_FILES $HIPBLASLT_OTHER_FILES)
-
-# hipsparselt library files
-HIPSPARSELT_LIB_SRC=$ROCM_HOME/lib/hipsparselt/library
-HIPSPARSELT_LIB_DST=lib/hipsparselt/library
-HIPSPARSELT_ARCH_SPECIFIC_FILES=$(ls $HIPSPARSELT_LIB_SRC | grep -E $ARCH)
-#HIPSPARSELT_OTHER_FILES=$(ls $HIPSPARSELT_LIB_SRC | grep -v gfx)
-HIPSPARSELT_LIB_FILES=($HIPSPARSELT_ARCH_SPECIFIC_FILES $HIPSPARSELT_OTHER_FILES)
+ARCH_SPECIFIC_FILES=$(ls $HIPBLASLT_LIB_SRC | grep -E $ARCH)
+OTHER_FILES=$(ls $HIPBLASLT_LIB_SRC | grep -v gfx)
+HIPBLASLT_LIB_FILES=($ARCH_SPECIFIC_FILES $OTHER_FILES)

 # ROCm library files
 ROCM_SO_PATHS=()
@ -243,14 +234,12 @@ DEPS_SONAME=(
 DEPS_AUX_SRCLIST=(
    "${ROCBLAS_LIB_FILES[@]/#/$ROCBLAS_LIB_SRC/}"
    "${HIPBLASLT_LIB_FILES[@]/#/$HIPBLASLT_LIB_SRC/}"
-    "${HIPSPARSELT_LIB_FILES[@]/#/$HIPSPARSELT_LIB_SRC/}"
    "/opt/amdgpu/share/libdrm/amdgpu.ids"
 )

 DEPS_AUX_DSTLIST=(
    "${ROCBLAS_LIB_FILES[@]/#/$ROCBLAS_LIB_DST/}"
    "${HIPBLASLT_LIB_FILES[@]/#/$HIPBLASLT_LIB_DST/}"
-    "${HIPSPARSELT_LIB_FILES[@]/#/$HIPSPARSELT_LIB_DST/}"
    "share/libdrm/amdgpu.ids"
 )

--- a/.ci/manywheel/build_xpu.sh
+++ b/.ci/manywheel/build_xpu.sh
@ -20,11 +20,7 @@ fi
 source /opt/intel/oneapi/compiler/latest/env/vars.sh
 source /opt/intel/oneapi/pti/latest/env/vars.sh
 source /opt/intel/oneapi/umf/latest/env/vars.sh
-source /opt/intel/oneapi/ccl/latest/env/vars.sh
-source /opt/intel/oneapi/mpi/latest/env/vars.sh
 export USE_STATIC_MKL=1
-export USE_ONEMKL=1
-export USE_XCCL=1

 WHEELHOUSE_DIR="wheelhousexpu"
 LIBTORCH_HOUSE_DIR="libtorch_housexpu"
--- a/.ci/onnx/README.md
+++ b/.ci/onnx/README.md
@ -10,3 +10,5 @@ example: `py2-cuda9.0-cudnn7-ubuntu16.04`. The Docker images that are
 built on Jenkins and are used in triggered builds already have this
 environment variable set in their manifest. Also see
 `./docker/jenkins/*/Dockerfile` and search for `BUILD_ENVIRONMENT`.
+
+Our Jenkins installation is located at https://ci.pytorch.org/jenkins/.
--- a/.ci/pytorch/build.sh
+++ b/.ci/pytorch/build.sh
@ -27,12 +27,6 @@ cmake --version
 echo "Environment variables:"
 env

-# The sccache wrapped version of nvcc gets put in /opt/cache/lib in docker since
-# there are some issues if it is always wrapped, so we need to add it to PATH
-# during CI builds.
-# https://github.com/pytorch/pytorch/blob/0b6c0898e6c352c8ea93daec854e704b41485375/.ci/docker/common/install_cache.sh#L97
-export PATH="/opt/cache/lib:$PATH"
-
 if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
  # Use jemalloc during compilation to mitigate https://github.com/pytorch/pytorch/issues/116289
  export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libjemalloc.so.2
@ -58,6 +52,12 @@ fi
 export USE_LLVM=/opt/llvm
 export LLVM_DIR=/opt/llvm/lib/cmake/llvm

+if [[ "$BUILD_ENVIRONMENT" == *executorch* ]]; then
+  # To build test_edge_op_registration
+  export BUILD_EXECUTORCH=ON
+  export USE_CUDA=0
+fi
+
 if ! which conda; then
  # In ROCm CIs, we are doing cross compilation on build machines with
  # intel cpu and later run tests on machines with amd cpu.
@ -171,12 +171,6 @@ fi
 if [[ "$BUILD_ENVIRONMENT" == *xpu* ]]; then
  # shellcheck disable=SC1091
  source /opt/intel/oneapi/compiler/latest/env/vars.sh
-  # shellcheck disable=SC1091
-  source /opt/intel/oneapi/ccl/latest/env/vars.sh
-  # shellcheck disable=SC1091
-  source /opt/intel/oneapi/mpi/latest/env/vars.sh
-  # Enable XCCL build
-  export USE_XCCL=1
  # XPU kineto feature dependencies are not fully ready, disable kineto build as temp WA
  export USE_KINETO=0
  export TORCH_XPU_ARCH_LIST=pvc
@ -198,8 +192,10 @@ fi

 # We only build FlashAttention files for CUDA 8.0+, and they require large amounts of
 # memory to build and will OOM
-if [[ "$BUILD_ENVIRONMENT" == *cuda* ]] && [[ 1 -eq $(echo "${TORCH_CUDA_ARCH_LIST} >= 8.0" | bc) ]]; then
-  export BUILD_CUSTOM_STEP="ninja -C build flash_attention -j 2"
+if [[ "$BUILD_ENVIRONMENT" == *cuda* ]] && [[ 1 -eq $(echo "${TORCH_CUDA_ARCH_LIST} >= 8.0" | bc) ]] && [ -z "$MAX_JOBS_OVERRIDE" ]; then
+  echo "WARNING: FlashAttention files require large amounts of memory to build and will OOM"
+  echo "Setting MAX_JOBS=(nproc-2)/3 to reduce memory usage"
+  export MAX_JOBS="$(( $(nproc --ignore=2) / 3 ))"
 fi

 if [[ "${BUILD_ENVIRONMENT}" == *clang* ]]; then
@ -255,7 +251,6 @@ if [[ "$BUILD_ENVIRONMENT" == *-bazel-* ]]; then
  set -e -o pipefail

  get_bazel
-  python3 tools/optional_submodules.py checkout_eigen

  # Leave 1 CPU free and use only up to 80% of memory to reduce the change of crashing
  # the runner
@ -393,8 +388,10 @@ else
    # This is an attempt to mitigate flaky libtorch build OOM error. By default, the build parallelization
    # is set to be the number of CPU minus 2. So, let's try a more conservative value here. A 4xlarge has
    # 16 CPUs
-    MAX_JOBS=$(nproc --ignore=4)
-    export MAX_JOBS
+    if [ -z "$MAX_JOBS_OVERRIDE" ]; then
+      MAX_JOBS=$(nproc --ignore=4)
+      export MAX_JOBS
+    fi

    # NB: Install outside of source directory (at the same level as the root
    # pytorch folder) so that it doesn't get cleaned away prior to docker push.
--- a/.ci/pytorch/check_binary.sh
+++ b/.ci/pytorch/check_binary.sh
@ -302,22 +302,19 @@ except RuntimeError as e:
 fi

 ###############################################################################
-# Check for C++ ABI compatibility to GCC-11 - GCC 13
+# Check for C++ ABI compatibility to GCC-11
 ###############################################################################
 if [[ "$(uname)" == 'Linux' &&  "$PACKAGE_TYPE" == 'manywheel' ]]; then
  pushd /tmp
-  # Per https://gcc.gnu.org/onlinedocs/gcc/C_002b_002b-Dialect-Options.html
-  # gcc-11 is ABI16, gcc-13 is ABI18, gcc-14 is ABI19
-  # gcc 11 - CUDA 11.8, xpu, rocm
-  # gcc 13 - CUDA 12.6, 12.8 and cpu
-  # Please see issue for reference: https://github.com/pytorch/pytorch/issues/152426
-  if [[ "$(uname -m)" == "s390x" ]]; then
-    cxx_abi="19"
-  elif [[ "$DESIRED_CUDA" != 'xpu' && "$DESIRED_CUDA" != 'rocm'* ]]; then
-    cxx_abi="18"
-  else
-    cxx_abi="16"
+  # Per https://gcc.gnu.org/onlinedocs/gcc/C_002b_002b-Dialect-Options.html gcc-11 is ABI16
+  # Though manylinux_2.28 should have been build with gcc-14, per
+  # https://github.com/pypa/manylinux?tab=readme-ov-file#manylinux_2_28-almalinux-8-based
+  # On s390x gcc 14 is used because it contains fix for interaction
+  # between precompiled headers and vectorization builtins.
+  # This fix is not available in earlier gcc versions.
+  # gcc-14 uses ABI19.
+  if [[ "$(uname -m)" != "s390x" ]]; then
+    python -c "import torch; exit(0 if torch._C._PYBIND11_BUILD_ABI == '_cxxabi1016' else 1)"
  fi
-  python -c "import torch; exit(0 if torch._C._PYBIND11_BUILD_ABI == '_cxxabi10${cxx_abi}' else 1)"
  popd
 fi
--- a/.ci/pytorch/common-build.sh
+++ b/.ci/pytorch/common-build.sh
@ -13,13 +13,6 @@ if [[ "$BUILD_ENVIRONMENT" != *win-* ]]; then
    fi

    if which sccache > /dev/null; then
-        # Clear SCCACHE_BUCKET and SCCACHE_REGION if they are empty, otherwise
-        # sccache will complain about invalid bucket configuration
-        if [[ -z "${SCCACHE_BUCKET:-}" ]]; then
-          unset SCCACHE_BUCKET
-          unset SCCACHE_REGION
-        fi
-
        # Save sccache logs to file
        sccache --stop-server > /dev/null  2>&1 || true
        rm -f ~/sccache_error.log || true
--- a/.ci/pytorch/common.sh
+++ b/.ci/pytorch/common.sh
@ -15,6 +15,6 @@ if [[ "${BUILD_ENVIRONMENT}" == *rocm* ]]; then
  export PYTORCH_TEST_WITH_ROCM=1
 fi

-# TODO: Reenable libtorch testing for MacOS, see https://github.com/pytorch/pytorch/issues/62598
+# TODO: Renable libtorch testing for MacOS, see https://github.com/pytorch/pytorch/issues/62598
 # shellcheck disable=SC2034
 BUILD_TEST_LIBTORCH=0
--- a/.ci/pytorch/common_utils.sh
+++ b/.ci/pytorch/common_utils.sh
@ -159,6 +159,11 @@ function install_torchvision() {
  fi
 }

+function install_tlparse() {
+  pip_install --user "tlparse==0.3.30"
+  PATH="$(python -m site --user-base)/bin:$PATH"
+}
+
 function install_torchrec_and_fbgemm() {
  local torchrec_commit
  torchrec_commit=$(get_pinned_commit torchrec)
--- a/.ci/pytorch/install_cache_xla.sh
+++ b/.ci/pytorch/install_cache_xla.sh
@ -1,50 +1,31 @@
 #!/bin/bash

 # Script for installing sccache on the xla build job, which uses xla's docker
-# image, which has sccache installed but doesn't write the stubs.  This is
-# mostly copied from .ci/docker/install_cache.sh.  Changes are: removing checks
-# that will always return the same thing, ex checks for for rocm, CUDA, changing
-# the path where sccache is installed, not changing /etc/environment, and not
-# installing/downloading sccache as it is already in the docker image.
+# image and doesn't have sccache installed on it.  This is mostly copied from
+# .ci/docker/install_cache.sh.  Changes are: removing checks that will always
+# return the same thing, ex checks for for rocm, CUDA, and changing the path
+# where sccache is installed, and not changing /etc/environment.

 set -ex -o pipefail

+install_binary() {
+  echo "Downloading sccache binary from S3 repo"
+  curl --retry 3 https://s3.amazonaws.com/ossci-linux/sccache -o /tmp/cache/bin/sccache
+}
+
 mkdir -p /tmp/cache/bin
+mkdir -p /tmp/cache/lib
 export PATH="/tmp/cache/bin:$PATH"

+install_binary
+chmod a+x /tmp/cache/bin/sccache
+
 function write_sccache_stub() {
  # Unset LD_PRELOAD for ps because of asan + ps issues
  # https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90589
-  if [ "$1" == "gcc" ]; then
-    # Do not call sccache recursively when dumping preprocessor argument
-    # For some reason it's very important for the first cached nvcc invocation
-    cat >"/tmp/cache/bin/$1" <<EOF
-#!/bin/sh
-
-# sccache does not support -E flag, so we need to call the original compiler directly in order to avoid calling this wrapper recursively
-for arg in "\$@"; do
-  if [ "\$arg" = "-E" ]; then
-    exec $(which "$1") "\$@"
-  fi
-done
-
-if [ \$(env -u LD_PRELOAD ps -p \$PPID -o comm=) != sccache ]; then
-  exec sccache $(which "$1") "\$@"
-else
-  exec $(which "$1") "\$@"
-fi
-EOF
-  else
-    cat >"/tmp/cache/bin/$1" <<EOF
-#!/bin/sh
-
-if [ \$(env -u LD_PRELOAD ps -p \$PPID -o comm=) != sccache ]; then
-  exec sccache $(which "$1") "\$@"
-else
-  exec $(which "$1") "\$@"
-fi
-EOF
-  fi
+  # shellcheck disable=SC2086
+  # shellcheck disable=SC2059
+  printf "#!/bin/sh\nif [ \$(env -u LD_PRELOAD ps -p \$PPID -o comm=) != sccache ]; then\n  exec sccache $(which $1) \"\$@\"\nelse\n  exec $(which $1) \"\$@\"\nfi" > "/tmp/cache/bin/$1"
  chmod a+x "/tmp/cache/bin/$1"
 }

--- a/.ci/pytorch/macos-build.sh
+++ b/.ci/pytorch/macos-build.sh
@ -40,7 +40,7 @@ if [[ ${BUILD_ENVIRONMENT} == *"distributed"* ]]; then
 else
  # Explicitly set USE_DISTRIBUTED=0 to align with the default build config on mac. This also serves as the sole CI config that tests
  # that building with USE_DISTRIBUTED=0 works at all. See https://github.com/pytorch/pytorch/issues/86448
-  USE_DISTRIBUTED=0 USE_OPENMP=1 MACOSX_DEPLOYMENT_TARGET=11.0 WERROR=1 BUILD_TEST=OFF USE_PYTORCH_METAL=1 python setup.py bdist_wheel --plat-name macosx_11_0_arm64
+  USE_DISTRIBUTED=0 USE_OPENMP=1 MACOSX_DEPLOYMENT_TARGET=11.0 WERROR=1 BUILD_TEST=OFF USE_PYTORCH_METAL=1 python setup.py bdist_wheel
 fi
 if which sccache > /dev/null; then
  print_sccache_stats
--- a/.ci/pytorch/macos-common.sh
+++ b/.ci/pytorch/macos-common.sh
@ -20,4 +20,14 @@ print_cmake_info() {
  CONDA_INSTALLATION_DIR=$(dirname "$CMAKE_EXEC")
  # Print all libraries under cmake rpath for debugging
  ls -la "$CONDA_INSTALLATION_DIR/../lib"
+
+  export CMAKE_EXEC
+  # Explicitly add conda env lib folder to cmake rpath to address the flaky issue
+  # where cmake dependencies couldn't be found. This seems to point to how conda
+  # links $CMAKE_EXEC to its package cache when cloning a new environment
+  install_name_tool -add_rpath @executable_path/../lib "${CMAKE_EXEC}" || true
+  # Adding the rpath will invalidate cmake signature, so signing it again here
+  # to trust the executable. EXC_BAD_ACCESS (SIGKILL (Code Signature Invalid))
+  # with an exit code 137 otherwise
+  codesign -f -s - "${CMAKE_EXEC}" || true
 }
--- a/.ci/pytorch/macos-test.sh
+++ b/.ci/pytorch/macos-test.sh
@ -5,6 +5,11 @@ set -x
 # shellcheck source=./macos-common.sh
 source "$(dirname "${BASH_SOURCE[0]}")/macos-common.sh"

+if [[ -n "$CONDA_ENV" ]]; then
+  # Use binaries under conda environment
+  export PATH="$CONDA_ENV/bin":$PATH
+fi
+
 # Test that OpenMP is enabled
 pushd test
 if [[ ! $(python -c "import torch; print(int(torch.backends.openmp.is_available()))") == "1" ]]; then
@ -37,16 +42,6 @@ test_python_all() {
  assert_git_not_dirty
 }

-test_python_mps() {
-  setup_test_python
-
-  time python test/run_test.py --verbose --mps
-  MTL_CAPTURE_ENABLED=1 ${CONDA_RUN} python3 test/test_mps.py --verbose -k test_metal_capture
-
-  assert_git_not_dirty
-}
-
-
 test_python_shard() {
  if [[ -z "$NUM_TEST_SHARDS" ]]; then
    echo "NUM_TEST_SHARDS must be defined to run a Python test shard"
@ -160,7 +155,6 @@ test_jit_hooks() {
 torchbench_setup_macos() {
  git clone --recursive https://github.com/pytorch/vision torchvision
  git clone --recursive https://github.com/pytorch/audio torchaudio
-  brew install jpeg-turbo libpng

  pushd torchvision
  git fetch
@ -175,8 +169,7 @@ torchbench_setup_macos() {
  git checkout "$(cat ../.github/ci_commit_pins/audio.txt)"
  git submodule update --init --recursive
  python setup.py clean
-  #TODO: Remove me, when figure out how to make TorchAudio find brew installed openmp
-  USE_OPENMP=0 python setup.py develop
+  python setup.py develop
  popd

  # Shellcheck doesn't like it when you pass no arguments to a function that can take args. See https://www.shellcheck.net/wiki/SC2120
@ -184,8 +177,9 @@ torchbench_setup_macos() {
  checkout_install_torchbench
 }

-pip_benchmark_deps() {
-  python -mpip install --no-input astunparse requests cython scikit-learn
+conda_benchmark_deps() {
+  conda install -y astunparse numpy scipy ninja pyyaml setuptools cmake typing-extensions requests protobuf numba cython scikit-learn
+  conda install -y -c conda-forge librosa
 }


@ -193,7 +187,7 @@ test_torchbench_perf() {
  print_cmake_info

  echo "Launching torchbench setup"
-  pip_benchmark_deps
+  conda_benchmark_deps
  torchbench_setup_macos

  TEST_REPORTS_DIR=$(pwd)/test/test-reports
@ -220,7 +214,7 @@ test_torchbench_smoketest() {
  print_cmake_info

  echo "Launching torchbench setup"
-  pip_benchmark_deps
+  conda_benchmark_deps
  # shellcheck disable=SC2119,SC2120
  torchbench_setup_macos

@ -228,52 +222,37 @@ test_torchbench_smoketest() {
  mkdir -p "$TEST_REPORTS_DIR"

  local device=mps
-  local dtypes=(undefined float16 bfloat16 notset)
-  local dtype=${dtypes[$1]}
-  local models=(hf_T5 llama BERT_pytorch dcgan hf_GPT2 yolov3 resnet152 sam sam_fast pytorch_unet stable_diffusion_text_encoder speech_transformer Super_SloMo doctr_det_predictor doctr_reco_predictor timm_resnet timm_vovnet vgg16)
+  local models=(hf_T5 llama BERT_pytorch dcgan hf_GPT2 yolov3 resnet152 sam pytorch_unet stable_diffusion_text_encoder moco speech_transformer)

  for backend in eager inductor; do

-    echo "Launching torchbench inference performance run for backend ${backend} and dtype ${dtype}"
-    local dtype_arg="--${dtype}"
-    if [ "$dtype" == notset ]; then
-        dtype_arg="--float32"
-    fi
-    touch "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_inference_${device}_performance.csv"
-    for model in "${models[@]}"; do
-      PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
-        --performance --only "$model" --backend "$backend" --inference --devices "$device" "$dtype_arg" \
-        --output "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_inference_${device}_performance.csv" || true
-      if [ "$backend" == "inductor" ]; then
-        PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
-          --accuracy --only "$model" --backend "$backend" --inference --devices "$device" "$dtype_arg" \
-          --output "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_inference_${device}_accuracy.csv" || true
-      fi
-    done
-    if [ "$backend" == "inductor" ]; then
-      PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/huggingface.py \
-        --performance --backend "$backend" --inference --devices "$device" "$dtype_arg" \
-        --output "$TEST_REPORTS_DIR/inductor_${backend}_huggingface_${dtype}_inference_${device}_performance.csv" || true
-      PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/huggingface.py \
-        --accuracy --backend "$backend" --inference --devices "$device" "$dtype_arg" \
-        --output "$TEST_REPORTS_DIR/inductor_${backend}_huggingface_${dtype}_inference_${device}_accuracy.csv" || true
-    fi
-
-    if [ "$dtype" == notset ]; then
-      for dtype_ in notset amp; do
-        echo "Launching torchbench training performance run for backend ${backend} and dtype ${dtype_}"
-        touch "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype_}_training_${device}_performance.csv"
-        local dtype_arg="--${dtype_}"
-        if [ "$dtype_" == notset ]; then
+    for dtype in notset float16 bfloat16; do
+      echo "Launching torchbench inference performance run for backend ${backend} and dtype ${dtype}"
+      local dtype_arg="--${dtype}"
+      if [ "$dtype" == notset ]; then
          dtype_arg="--float32"
-        fi
-        for model in "${models[@]}"; do
-          PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
-            --performance --only "$model" --backend "$backend" --training --devices "$device" "$dtype_arg" \
-            --output "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype_}_training_${device}_performance.csv" || true
-        done
+      fi
+      touch "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_inference_${device}_performance.csv"
+      for model in "${models[@]}"; do
+        PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
+          --performance --only "$model" --backend "$backend" --inference --devices "$device" "$dtype_arg" \
+          --output "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_inference_${device}_performance.csv" || true
      done
-    fi
+    done
+
+    for dtype in notset amp; do
+      echo "Launching torchbench training performance run for backend ${backend} and dtype ${dtype}"
+      touch "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_training_${device}_performance.csv"
+      local dtype_arg="--${dtype}"
+      if [ "$dtype" == notset ]; then
+          dtype_arg="--float32"
+      fi
+      for model in "${models[@]}"; do
+        PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
+          --performance --only "$model" --backend "$backend" --training --devices "$device" "$dtype_arg" \
+          --output "$TEST_REPORTS_DIR/inductor_${backend}_torchbench_${dtype}_training_${device}_performance.csv" || true
+      done
+    done

  done

@ -284,7 +263,7 @@ test_hf_perf() {
  print_cmake_info
  TEST_REPORTS_DIR=$(pwd)/test/test-reports
  mkdir -p "$TEST_REPORTS_DIR"
-  pip_benchmark_deps
+  conda_benchmark_deps
  torchbench_setup_macos

  echo "Launching HuggingFace training perf run"
@ -300,7 +279,7 @@ test_timm_perf() {
  print_cmake_info
  TEST_REPORTS_DIR=$(pwd)/test/test-reports
  mkdir -p "$TEST_REPORTS_DIR"
-  pip_benchmark_deps
+  conda_benchmark_deps
  torchbench_setup_macos

  echo "Launching timm training perf run"
@ -312,6 +291,8 @@ test_timm_perf() {
  echo "timm benchmark on mps device completed"
 }

+install_tlparse
+
 if [[ $TEST_CONFIG == *"perf_all"* ]]; then
  test_torchbench_perf
  test_hf_perf
@ -323,9 +304,7 @@ elif [[ $TEST_CONFIG == *"perf_hf"* ]]; then
 elif [[ $TEST_CONFIG == *"perf_timm"* ]]; then
  test_timm_perf
 elif [[ $TEST_CONFIG == *"perf_smoketest"* ]]; then
-  test_torchbench_smoketest "${SHARD_NUMBER}"
-elif [[ $TEST_CONFIG == *"mps"* ]]; then
-  test_python_mps
+  test_torchbench_smoketest
 elif [[ $NUM_TEST_SHARDS -gt 1 ]]; then
  test_python_shard "${SHARD_NUMBER}"
  if [[ "${SHARD_NUMBER}" == 1 ]]; then
--- a/.ci/pytorch/run_tests.sh
+++ b/.ci/pytorch/run_tests.sh
@ -76,7 +76,7 @@ fi
 # Environment initialization
 if [[ "$(uname)" == Darwin ]]; then
    # Install the testing dependencies
-    retry pip install -q future hypothesis ${NUMPY_PACKAGE} ${PROTOBUF_PACKAGE} pytest setuptools six typing_extensions pyyaml
+    retry conda install -yq future hypothesis ${NUMPY_PACKAGE} ${PROTOBUF_PACKAGE} pytest setuptools six typing_extensions pyyaml
 else
    retry pip install -qr requirements.txt || true
    retry pip install -q hypothesis protobuf pytest setuptools || true
@ -91,6 +91,7 @@ fi

 echo "Testing with:"
 pip freeze
+conda list || true

 ##############################################################################
 # Smoke tests
--- a/.ci/pytorch/smoke_test/check_binary_symbols.py
+++ b/.ci/pytorch/smoke_test/check_binary_symbols.py
@ -93,7 +93,7 @@ def check_lib_symbols_for_abi_correctness(lib: str) -> None:
            f"Found pre-cxx11 symbols, but there shouldn't be any, see: {pre_cxx11_symbols[:100]}"
        )
    if num_cxx11_symbols < 100:
-        raise RuntimeError("Didn't find enough cxx11 symbols")
+        raise RuntimeError("Didn't find enought cxx11 symbols")


 def main() -> None:
--- a/.ci/pytorch/smoke_test/check_gomp.py
+++ b/.ci/pytorch/smoke_test/check_gomp.py
@ -1,77 +0,0 @@
-import ctypes
-import os
-import sys
-from pathlib import Path
-
-
-def get_gomp_thread():
-    """
-    Retrieves the maximum number of OpenMP threads after loading the `libgomp.so.1` library
-    and the `libtorch_cpu.so` library. It then queries the
-    maximum number of threads available for OpenMP parallel regions using the
-    `omp_get_max_threads` function.
-
-    Returns:
-        int: The maximum number of OpenMP threads available.
-
-    Notes:
-        - The function assumes the default path for `libgomp.so.1` on AlmaLinux OS.
-        - The path to `libtorch_cpu.so` is constructed based on the Python executable's
-          installation directory.
-        - This function is specific to environments where PyTorch and OpenMP are used
-          together and may require adjustments for other setups.
-    """
-    python_path = Path(sys.executable).resolve()
-    python_prefix = (
-        python_path.parent.parent
-    )  # Typically goes to the Python installation root
-
-    # Get the additional ABI flags (if any); it may be an empty string.
-    abiflags = getattr(sys, "abiflags", "")
-
-    # Construct the Python directory name correctly (e.g., "python3.13t").
-    python_version = (
-        f"python{sys.version_info.major}.{sys.version_info.minor}{abiflags}"
-    )
-
-    libtorch_cpu_path = (
-        python_prefix
-        / "lib"
-        / python_version
-        / "site-packages"
-        / "torch"
-        / "lib"
-        / "libtorch_cpu.so"
-    )
-
-    # use the default gomp path of AlmaLinux OS
-    libgomp_path = "/usr/lib64/libgomp.so.1"
-    # if it does not exist, try Ubuntu path
-    if not os.path.exists(libgomp_path):
-        libgomp_path = f"/usr/lib/{os.uname().machine}-linux-gnu/libgomp.so.1"
-
-    os.environ["GOMP_CPU_AFFINITY"] = "0-3"
-
-    libgomp = ctypes.CDLL(libgomp_path)
-    libgomp = ctypes.CDLL(libtorch_cpu_path)
-
-    libgomp.omp_get_max_threads.restype = ctypes.c_int
-    libgomp.omp_get_max_threads.argtypes = []
-
-    omp_max_threads = libgomp.omp_get_max_threads()
-    return omp_max_threads
-
-
-def main():
-    omp_max_threads = get_gomp_thread()
-    print(
-        f"omp_max_threads after loading libgomp.so and libtorch_cpu.so: {omp_max_threads}"
-    )
-    if omp_max_threads == 1:
-        raise RuntimeError(
-            "omp_max_threads is 1. Check whether libgomp.so is loaded twice."
-        )
-
-
-if __name__ == "__main__":
-    main()
--- a/.ci/pytorch/smoke_test/smoke_test.py
+++ b/.ci/pytorch/smoke_test/smoke_test.py
@ -276,7 +276,7 @@ def smoke_test_cuda(
            torch_nccl_version = ".".join(str(v) for v in torch.cuda.nccl.version())
            print(f"Torch nccl; version: {torch_nccl_version}")

-        # Pypi dependencies are installed on linux only and nccl is available only on Linux.
+        # Pypi dependencies are installed on linux ony and nccl is availbale only on Linux.
        if pypi_pkg_check == "enabled" and sys.platform in ["linux", "linux2"]:
            compare_pypi_to_torch_versions(
                "cudnn", find_pypi_package_version("nvidia-cudnn"), torch_cudnn_version
--- a/.ci/pytorch/test.sh
+++ b/.ci/pytorch/test.sh
@ -11,8 +11,6 @@ export TERM=vt100

 # shellcheck source=./common.sh
 source "$(dirname "${BASH_SOURCE[0]}")/common.sh"
-# shellcheck source=./common-build.sh
-source "$(dirname "${BASH_SOURCE[0]}")/common-build.sh"

 # Do not change workspace permissions for ROCm and s390x CI jobs
 # as it can leave workspace with bad permissions for cancelled jobs
@ -193,12 +191,8 @@ if [[ "$BUILD_ENVIRONMENT" == *xpu* ]]; then
    # shellcheck disable=SC1091
    source /opt/intel/oneapi/umf/latest/env/vars.sh
  fi
-  # shellcheck disable=SC1091
-  source /opt/intel/oneapi/ccl/latest/env/vars.sh
-  # shellcheck disable=SC1091
-  source /opt/intel/oneapi/mpi/latest/env/vars.sh
  # Check XPU status before testing
-  timeout 30 xpu-smi discovery || true
+  xpu-smi discovery
 fi

 if [[ "$BUILD_ENVIRONMENT" != *-bazel-* ]] ; then
@ -214,6 +208,8 @@ if [[ "$BUILD_ENVIRONMENT" == *aarch64* ]]; then
  export VALGRIND=OFF
 fi

+install_tlparse
+
 # DANGER WILL ROBINSON.  The LD_PRELOAD here could cause you problems
 # if you're not careful.  Check this if you made some changes and the
 # ASAN test is not working
@ -226,7 +222,7 @@ if [[ "$BUILD_ENVIRONMENT" == *asan* ]]; then
    export PYTORCH_TEST_WITH_ASAN=1
    export PYTORCH_TEST_WITH_UBSAN=1
    # TODO: Figure out how to avoid hard-coding these paths
-    export ASAN_SYMBOLIZER_PATH=/usr/lib/llvm-18/bin/llvm-symbolizer
+    export ASAN_SYMBOLIZER_PATH=/usr/lib/llvm-15/bin/llvm-symbolizer
    export TORCH_USE_RTLD_GLOBAL=1
    # NB: We load libtorch.so with RTLD_GLOBAL for UBSAN, unlike our
    # default behavior.
@ -318,29 +314,6 @@ test_python() {
  assert_git_not_dirty
 }

-test_python_smoke() {
-  # Smoke tests for H100
-  time python test/run_test.py --include test_matmul_cuda inductor/test_fp8 inductor/test_max_autotune $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  assert_git_not_dirty
-}
-
-test_h100_distributed() {
-  # Distributed tests at H100
-  time python test/run_test.py --include distributed/_composable/test_composability/test_pp_composability.py  $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  # This test requires multicast support
-  time python test/run_test.py --include distributed/_composable/fsdp/test_fully_shard_comm.py -k TestFullyShardAllocFromPG $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  assert_git_not_dirty
-}
-
-test_h100_symm_mem() {
-  # symmetric memory test
-  time python test/run_test.py --include distributed/test_symmetric_memory.py  $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  time python test/run_test.py --include distributed/test_nvshmem.py $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  time python test/run_test.py --include distributed/test_nvshmem_triton.py $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  time python test/run_test.py --include distributed/test_nccl.py $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  assert_git_not_dirty
-}
-
 test_lazy_tensor_meta_reference_disabled() {
  export TORCH_DISABLE_FUNCTIONALIZATION_META_REFERENCE=1
  echo "Testing lazy tensor operations without meta reference"
@ -355,7 +328,6 @@ test_dynamo_wrapped_shard() {
    exit 1
  fi
  python tools/dynamo/verify_dynamo.py
-  python tools/dynamo/gb_id_mapping.py verify
  # PLEASE DO NOT ADD ADDITIONAL EXCLUDES HERE.
  # Instead, use @skipIfTorchDynamo on your tests.
  time python test/run_test.py --dynamo \
@ -370,17 +342,6 @@ test_dynamo_wrapped_shard() {
  assert_git_not_dirty
 }

-test_einops() {
-  pip install einops==0.6.1
-  time python test/run_test.py --einops --verbose --upload-artifacts-while-running
-  pip install einops==0.7.0
-  time python test/run_test.py --einops --verbose --upload-artifacts-while-running
-  pip install einops==0.8.1
-  time python test/run_test.py --einops --verbose --upload-artifacts-while-running
-  assert_git_not_dirty
-}
-
-
 test_inductor_distributed() {
  # Smuggle a few multi-gpu tests here so that we don't have to request another large node
  echo "Testing multi_gpu tests in test_torchinductor"
@ -437,22 +398,8 @@ test_inductor_aoti() {
    # We need to hipify before building again
    python3 tools/amd_build/build_amd.py
  fi
-  if [[ "$BUILD_ENVIRONMENT" == *sm86* ]]; then
-    BUILD_COMMAND=(TORCH_CUDA_ARCH_LIST=8.6 USE_FLASH_ATTENTION=OFF python setup.py develop)
-    # TODO: Replace me completely, as one should not use conda libstdc++, nor need special path to TORCH_LIB
-    TEST_ENVS=(CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="/opt/conda/envs/py_3.10/lib:${TORCH_LIB_DIR}:${LD_LIBRARY_PATH}")
-  else
-    BUILD_COMMAND=(python setup.py develop)
-    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 "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference -dist=loadfile
+  BUILD_AOT_INDUCTOR_TEST=1 python setup.py develop
+  CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="${TORCH_LIB_DIR}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference
 }

 test_inductor_cpp_wrapper_shard() {
@ -465,26 +412,46 @@ test_inductor_cpp_wrapper_shard() {
  TEST_REPORTS_DIR=$(pwd)/test/test-reports
  mkdir -p "$TEST_REPORTS_DIR"

+  if [[ "$1" -eq "2" ]]; then
+    # For now, manually put the opinfo tests in shard 2, and all other tests in
+    # shard 1.  Test specific things triggering past bugs, for now.
+    python test/run_test.py \
+      --include inductor/test_torchinductor_opinfo \
+      -k 'linalg or to_sparse' \
+      --verbose
+    exit
+  fi
+
  # Run certain inductor unit tests with cpp wrapper. In the end state, we
  # should be able to run all the inductor unit tests with cpp_wrapper.
-  #
-  # TODO: I'm pretty sure that "TestInductorOpInfoCPU" is not a valid filter,
-  # but change that in another PR to more accurately monitor the increased CI
-  # usage.
-  python test/run_test.py \
-    --include inductor/test_torchinductor_opinfo \
-    -k 'linalg or to_sparse or TestInductorOpInfoCPU' \
-    --shard "$1" "$NUM_TEST_SHARDS" \
-    --verbose
  python test/run_test.py \
    --include inductor/test_torchinductor inductor/test_max_autotune inductor/test_cpu_repro \
-    --shard "$1" "$NUM_TEST_SHARDS" \
-    --verbose
-  python test/run_test.py --inductor \
-    --include test_torch \
-    -k 'take' \
-    --shard "$1" "$NUM_TEST_SHARDS" \
    --verbose
+  python test/run_test.py --inductor --include test_torch -k 'take' --verbose
+
+  # Run inductor benchmark tests with cpp wrapper.
+  # Skip benchmark tests if it's in rerun-disabled-mode.
+  if [[ "${PYTORCH_TEST_RERUN_DISABLED_TESTS}" == "1" ]]; then
+    echo "skip dynamo benchmark tests for rerun-disabled-test"
+  else
+    echo "run dynamo benchmark tests with cpp wrapper"
+    python benchmarks/dynamo/timm_models.py --device cuda --accuracy --amp \
+    --training --inductor --disable-cudagraphs --only vit_base_patch16_224 \
+    --output "$TEST_REPORTS_DIR/inductor_cpp_wrapper_training.csv"
+    python benchmarks/dynamo/check_accuracy.py \
+      --actual "$TEST_REPORTS_DIR/inductor_cpp_wrapper_training.csv" \
+      --expected "benchmarks/dynamo/ci_expected_accuracy/${MAYBE_ROCM}inductor_timm_training.csv"
+
+    python benchmarks/dynamo/torchbench.py --device cuda --accuracy \
+      --bfloat16 --inference --inductor --only hf_T5 --output "$TEST_REPORTS_DIR/inductor_cpp_wrapper_inference.csv"
+    python benchmarks/dynamo/torchbench.py --device cuda --accuracy \
+      --bfloat16 --inference --inductor --only llama --output "$TEST_REPORTS_DIR/inductor_cpp_wrapper_inference.csv"
+    python benchmarks/dynamo/torchbench.py --device cuda --accuracy \
+      --bfloat16 --inference --inductor --only moco --output "$TEST_REPORTS_DIR/inductor_cpp_wrapper_inference.csv"
+    python benchmarks/dynamo/check_accuracy.py \
+      --actual "$TEST_REPORTS_DIR/inductor_cpp_wrapper_inference.csv" \
+      --expected "benchmarks/dynamo/ci_expected_accuracy/${MAYBE_ROCM}inductor_torchbench_inference.csv"
+  fi
 }

 # "Global" flags for inductor benchmarking controlled by TEST_CONFIG
@ -605,9 +572,7 @@ test_perf_for_dashboard() {

  local device=cuda
  if [[ "${TEST_CONFIG}" == *cpu* ]]; then
-    if [[ "${TEST_CONFIG}" == *cpu_x86_zen* ]]; then
-      device=cpu_x86_zen
-    elif [[ "${TEST_CONFIG}" == *cpu_x86* ]]; then
+    if [[ "${TEST_CONFIG}" == *cpu_x86* ]]; then
      device=cpu_x86
    elif [[ "${TEST_CONFIG}" == *cpu_aarch64* ]]; then
      device=cpu_aarch64
@ -623,11 +588,7 @@ test_perf_for_dashboard() {

  for mode in "${modes[@]}"; do
    if [[ "$mode" == "inference" ]]; then
-      if [[ "$device" == "cpu_x86" ]]; then
-        dtype=amp
-      else
-        dtype=bfloat16
-      fi
+      dtype=bfloat16
    elif [[ "$mode" == "training" ]]; then
      dtype=amp
    fi
@ -639,10 +600,6 @@ test_perf_for_dashboard() {
        target_flag+=( --no-translation-validation)
      fi

-      if [[ "$DASHBOARD_TAG" == *freezing-true* ]]; then
-        target_flag+=( --freezing)
-      fi
-
      if [[ "$DASHBOARD_TAG" == *default-true* ]]; then
        $TASKSET python "benchmarks/dynamo/$suite.py" \
            "${target_flag[@]}" --"$mode" --"$dtype" --backend "$backend" --disable-cudagraphs "$@" \
@ -845,7 +802,16 @@ test_inductor_torchbench_smoketest_perf() {
  done
 }

+test_inductor_get_core_number() {
+  if [[ "${TEST_CONFIG}" == *aarch64* ]]; then
+    echo "$(($(lscpu | grep 'Cluster(s):' | awk '{print $2}') * $(lscpu | grep 'Core(s) per cluster:' | awk '{print $4}')))"
+  else
+    echo "$(($(lscpu | grep 'Socket(s):' | awk '{print $2}') * $(lscpu | grep 'Core(s) per socket:' | awk '{print $4}')))"
+  fi
+}
+
 test_inductor_set_cpu_affinity(){
+  #set jemalloc
  JEMALLOC_LIB="$(find /usr/lib -name libjemalloc.so.2)"
  export LD_PRELOAD="$JEMALLOC_LIB":"$LD_PRELOAD"
  export MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:-1,muzzy_decay_ms:-1"
@ -857,23 +823,14 @@ test_inductor_set_cpu_affinity(){
    export KMP_AFFINITY=granularity=fine,compact,1,0
    export KMP_BLOCKTIME=1
  fi
-
-  # Use nproc here instead of lscpu because it takes into account cgroups slice
-  cpus=$(nproc)
-  thread_per_core=$(lscpu | grep 'Thread(s) per core:' | awk '{print $4}')
-  cores=$((cpus / thread_per_core))
-
-  # Set number of cores to 16 on aarch64 for performance runs
+  cores=$(test_inductor_get_core_number)
+  # Set number of cores to 16 on Aarch64 for performance runs.
  if [[ "${TEST_CONFIG}" == *aarch64* && $cores -gt 16 ]]; then
    cores=16
  fi
  export OMP_NUM_THREADS=$cores
-
-  # Handle cgroups slice start and end CPU
-  start_cpu=$(python -c 'import os; print(min(os.sched_getaffinity(0)))')
-  # Leaving one physical CPU for other tasks
-  end_cpu=$(($(python -c 'import os; print(max(os.sched_getaffinity(0)))') - thread_per_core))
-  export TASKSET="taskset -c $start_cpu-$end_cpu"
+  end_core=$((cores-1))
+  export TASKSET="taskset -c 0-$end_core"
 }

 test_inductor_torchbench_cpu_smoketest_perf(){
@ -1156,12 +1113,6 @@ test_custom_backend() {

 test_custom_script_ops() {
  echo "Testing custom script operators"
-
-  if [[ "$BUILD_ENVIRONMENT" == *s390x* ]]; then
-    echo "Skipping custom script operators until it's fixed"
-    return 0
-  fi
-
  CUSTOM_OP_BUILD="${CUSTOM_TEST_ARTIFACT_BUILD_DIR}/custom-op-build"
  pushd test/custom_operator
  cp -a "$CUSTOM_OP_BUILD" build
@ -1525,6 +1476,8 @@ test_executorch() {
  export PYTHON_EXECUTABLE=python
  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON"

+  # For llama3
+  bash examples/models/llama3_2_vision/install_requirements.sh
  # NB: We need to rebuild ExecuTorch runner here because it depends on PyTorch
  # from the PR
  bash .ci/scripts/setup-linux.sh --build-tool cmake
@ -1551,7 +1504,7 @@ test_executorch() {
 test_linux_aarch64() {
  python test/run_test.py --include test_modules test_mkldnn test_mkldnn_fusion test_openmp test_torch test_dynamic_shapes \
        test_transformers test_multiprocessing test_numpy_interop test_autograd test_binary_ufuncs test_complex test_spectral_ops \
-        test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops test_cpp_extensions_open_device_registration \
+        test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops \
        --shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose

  # Dynamo tests
@ -1585,8 +1538,7 @@ test_operator_benchmark() {

  cd "${TEST_DIR}"/benchmarks/operator_benchmark
  $TASKSET python -m benchmark_all_test --device "$1" --tag-filter "$2" \
-      --output-csv "${TEST_REPORTS_DIR}/operator_benchmark_eager_float32_cpu.csv" \
-      --output-json-for-dashboard "${TEST_REPORTS_DIR}/operator_benchmark_eager_float32_cpu.json" \
+      --output-dir "${TEST_REPORTS_DIR}/operator_benchmark_eager_float32_cpu.csv"

  pip_install pandas
  python check_perf_csv.py \
@ -1671,7 +1623,7 @@ elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
    install_torchaudio cuda
  fi
  install_torchvision
-  TORCH_CUDA_ARCH_LIST="8.0;8.6" install_torchao
+  TORCH_CUDA_ARCH_LIST="8.0;8.6" pip_install git+https://github.com/pytorch/ao.git
  id=$((SHARD_NUMBER-1))
  # https://github.com/opencv/opencv-python/issues/885
  pip_install opencv-python==4.8.0.74
@ -1696,11 +1648,11 @@ elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
    PYTHONPATH=$(pwd)/torchbench test_dynamo_benchmark torchbench "$id"
  fi
 elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
+  install_torchaudio cuda
  install_torchvision
+  checkout_install_torchbench hf_T5 llama moco
  PYTHONPATH=$(pwd)/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
-  if [[ "$SHARD_NUMBER" -eq "1" ]]; then
-    test_inductor_aoti
-  fi
+  test_inductor_aoti
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
  install_torchvision
  test_inductor_shard "${SHARD_NUMBER}"
@ -1709,8 +1661,6 @@ elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
      test_inductor_distributed
    fi
  fi
-elif [[ "${TEST_CONFIG}" == *einops* ]]; then
-  test_einops
 elif [[ "${TEST_CONFIG}" == *dynamo_wrapped* ]]; then
  install_torchvision
  test_dynamo_wrapped_shard "${SHARD_NUMBER}"
@ -1756,12 +1706,6 @@ elif [[ "${BUILD_ENVIRONMENT}" == *xpu* ]]; then
  test_python
  test_aten
  test_xpu_bin
-elif [[ "${TEST_CONFIG}" == smoke ]]; then
-  test_python_smoke
-elif [[ "${TEST_CONFIG}" == h100_distributed ]]; then
-  test_h100_distributed
-elif [[ "${TEST_CONFIG}" == "h100-symm-mem" ]]; then
-  test_h100_symm_mem
 else
  install_torchvision
  install_monkeytype
--- a/.ci/pytorch/win-build.sh
+++ b/.ci/pytorch/win-build.sh
@ -31,7 +31,7 @@ PYLONG_API_CHECK=$?
 if [[ $PYLONG_API_CHECK == 0 ]]; then
  echo "Usage of PyLong_{From,As}{Unsigned}Long API may lead to overflow errors on Windows"
  echo "because \`sizeof(long) == 4\` and \`sizeof(unsigned long) == 4\`."
-  echo "Please include \"torch/csrc/utils/python_numbers.h\" and use the corresponding APIs instead."
+  echo "Please include \"torch/csrc/utils/python_numbers.h\" and use the correspoding APIs instead."
  echo "PyLong_FromLong -> THPUtils_packInt32 / THPUtils_packInt64"
  echo "PyLong_AsLong -> THPUtils_unpackInt (32-bit) / THPUtils_unpackLong (64-bit)"
  echo "PyLong_FromUnsignedLong -> THPUtils_packUInt32 / THPUtils_packUInt64"
--- a/.ci/pytorch/win-test-helpers/build_pytorch.bat
+++ b/.ci/pytorch/win-test-helpers/build_pytorch.bat
@ -10,7 +10,7 @@ set PATH=C:\Program Files\CMake\bin;C:\Program Files\7-Zip;C:\ProgramData\chocol
 :: able to see what our cl.exe commands are (since you can actually
 :: just copy-paste them into a local Windows setup to just rebuild a
 :: single file.)
-:: log sizes are too long, but leaving this here in case someone wants to use it locally
+:: log sizes are too long, but leaving this here incase someone wants to use it locally
 :: set CMAKE_VERBOSE_MAKEFILE=1


@ -37,11 +37,6 @@ call %INSTALLER_DIR%\activate_miniconda3.bat
 if errorlevel 1 goto fail
 if not errorlevel 0 goto fail

-:: Update CMake
-call choco upgrade -y cmake --no-progress --installargs 'ADD_CMAKE_TO_PATH=System' --apply-install-arguments-to-dependencies --version=3.27.9
-if errorlevel 1 goto fail
-if not errorlevel 0 goto fail
-
 call pip install mkl-include==2021.4.0 mkl-devel==2021.4.0
 if errorlevel 1 goto fail
 if not errorlevel 0 goto fail
@ -93,7 +88,7 @@ set PATH=%CUDA_PATH%\bin;%CUDA_PATH%\libnvvp;%PATH%
 :cuda_build_end

 set DISTUTILS_USE_SDK=1
-set PATH=%TMP_DIR_WIN%\bin;C:\Program Files\CMake\bin;%PATH%
+set PATH=%TMP_DIR_WIN%\bin;%PATH%

 :: The latest Windows CUDA test is running on AWS G5 runner with A10G GPU
 if "%TORCH_CUDA_ARCH_LIST%" == "" set TORCH_CUDA_ARCH_LIST=8.6
--- a/.ci/pytorch/win-test-helpers/installation-helpers/install_magma.bat
+++ b/.ci/pytorch/win-test-helpers/installation-helpers/install_magma.bat
@ -24,7 +24,7 @@ if "%CUDA_SUFFIX%" == "" (

 if "%REBUILD%"=="" (
  if "%BUILD_ENVIRONMENT%"=="" (
-    curl --retry 3 --retry-all-errors -k https://s3.amazonaws.com/ossci-windows/magma_2.5.4_%CUDA_SUFFIX%_%BUILD_TYPE%.7z --output %TMP_DIR_WIN%\magma_2.5.4_%CUDA_SUFFIX%_%BUILD_TYPE%.7z & REM @lint-ignore
+    curl --retry 3 --retry-all-errors -k https://s3.amazonaws.com/ossci-windows/magma_2.5.4_%CUDA_SUFFIX%_%BUILD_TYPE%.7z --output %TMP_DIR_WIN%\magma_2.5.4_%CUDA_SUFFIX%_%BUILD_TYPE%.7z
  ) else (
    aws s3 cp s3://ossci-windows/magma_2.5.4_%CUDA_SUFFIX%_%BUILD_TYPE%.7z %TMP_DIR_WIN%\magma_2.5.4_%CUDA_SUFFIX%_%BUILD_TYPE%.7z --quiet
  )
--- a/.ci/pytorch/win-test-helpers/run_python_nn_smoketests.py
+++ b/.ci/pytorch/win-test-helpers/run_python_nn_smoketests.py
@ -52,7 +52,7 @@ if __name__ == "__main__":
            if os.path.exists(debugger):
                command_args = [debugger, "-o", "-c", "~*g; q"] + command_args
                command_string = " ".join(command_args)
-                print("Rerunning with traceback enabled")
+                print("Reruning with traceback enabled")
                print("Command:", command_string)
                subprocess.run(command_args, check=False)
            sys.exit(e.returncode)
--- a/.ci/pytorch/win-test.sh
+++ b/.ci/pytorch/win-test.sh
@ -38,7 +38,7 @@ if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
 fi

 # TODO: Move both of them to Windows AMI
-python -m pip install pytest-rerunfailures==10.3 pytest-cpp==2.3.0 tensorboard==2.13.0 protobuf==5.29.4 pytest-subtests==0.13.1
+python -m pip install pytest-rerunfailures==10.3 pytest-cpp==2.3.0 tensorboard==2.13.0 pytest-subtests==0.13.1

 # Install Z3 optional dependency for Windows builds.
 python -m pip install z3-solver==4.12.2.0
@ -52,9 +52,6 @@ python -m pip install parameterized==0.8.1
 # Install pulp for testing ilps under torch\distributed\_tools
 python -m pip install pulp==2.9.0

-# Install expecttest to merge https://github.com/pytorch/pytorch/pull/155308
-python -m pip install expecttest==0.3.0
-
 run_tests() {
    # Run nvidia-smi if available
    for path in '/c/Program Files/NVIDIA Corporation/NVSMI/nvidia-smi.exe' /c/Windows/System32/nvidia-smi.exe; do
--- a/.ci/pytorch/windows/arm64/bootstrap_libuv.bat
+++ b/.ci/pytorch/windows/arm64/bootstrap_libuv.bat
@ -7,7 +7,7 @@ if not exist "%DOWNLOADS_DIR%" mkdir %DOWNLOADS_DIR%
 if not exist "%DEPENDENCIES_DIR%" mkdir %DEPENDENCIES_DIR%

 :: activate visual studio
-call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64
+call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
 where cl.exe

 cd %DEPENDENCIES_DIR%
--- a/.ci/pytorch/windows/arm64/bootstrap_openblas.bat
+++ b/.ci/pytorch/windows/arm64/bootstrap_openblas.bat
@ -7,7 +7,7 @@ if not exist "%DOWNLOADS_DIR%" mkdir %DOWNLOADS_DIR%
 if not exist "%DEPENDENCIES_DIR%" mkdir %DEPENDENCIES_DIR%

 :: activate visual studio
-call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64
+call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
 where cl.exe

 :: Clone OpenBLAS
--- a/.ci/pytorch/windows/arm64/bootstrap_tests.bat
+++ b/.ci/pytorch/windows/arm64/bootstrap_tests.bat
@ -2,7 +2,7 @@
 cd %PYTORCH_ROOT%

 :: activate visual studio
-call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64
+call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
 where cl.exe

 :: create virtual environment
--- a/.ci/pytorch/windows/arm64/build_libtorch.bat
+++ b/.ci/pytorch/windows/arm64/build_libtorch.bat
@ -21,7 +21,7 @@ if %ENABLE_APL% == 1 (
 )

 :: activate visual studio
-call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64
+call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
 where cl.exe

 :: change to source directory
--- a/.ci/pytorch/windows/arm64/build_pytorch.bat
+++ b/.ci/pytorch/windows/arm64/build_pytorch.bat
@ -21,7 +21,7 @@ if %ENABLE_APL% == 1 (
 )

 :: activate visual studio
-call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64
+call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64
 where cl.exe

 :: change to source directory
--- a/.ci/pytorch/windows/arm64/smoke_test.bat
+++ b/.ci/pytorch/windows/arm64/smoke_test.bat
@ -33,7 +33,7 @@ pushd tmp
 set VC_VERSION_LOWER=14
 set VC_VERSION_UPPER=36

-call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64
+call "%DEPENDENCIES_DIR%\VSBuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64

 set install_root=%CD%
 set INCLUDE=%INCLUDE%;%install_root%\include;%install_root%\include\torch\csrc\api\include
--- a/.ci/pytorch/windows/cuda118.bat
+++ b/.ci/pytorch/windows/cuda118.bat
@ -0,0 +1,59 @@
+@echo off
+
+set MODULE_NAME=pytorch
+
+IF NOT EXIST "setup.py" IF NOT EXIST "%MODULE_NAME%" (
+    call internal\clone.bat
+    cd %~dp0
+) ELSE (
+    call internal\clean.bat
+)
+IF ERRORLEVEL 1 goto :eof
+
+call internal\check_deps.bat
+IF ERRORLEVEL 1 goto :eof
+
+REM Check for optional components
+
+set USE_CUDA=
+set CMAKE_GENERATOR=Visual Studio 15 2017 Win64
+
+IF "%NVTOOLSEXT_PATH%"=="" (
+    IF EXIST "C:\Program Files\NVIDIA Corporation\NvToolsExt\lib\x64\nvToolsExt64_1.lib"  (
+        set NVTOOLSEXT_PATH=C:\Program Files\NVIDIA Corporation\NvToolsExt
+    ) ELSE (
+        echo NVTX ^(Visual Studio Extension ^for CUDA^) ^not installed, failing
+        exit /b 1
+    )
+)
+
+IF "%CUDA_PATH_V118%"=="" (
+    IF EXIST "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin\nvcc.exe" (
+        set "CUDA_PATH_V118=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8"
+    ) ELSE (
+        echo CUDA 11.8 not found, failing
+        exit /b 1
+    )
+)
+
+IF "%BUILD_VISION%" == "" (
+    set TORCH_CUDA_ARCH_LIST=3.7+PTX;5.0;6.0;6.1;7.0;7.5;8.0;8.6;9.0
+    set TORCH_NVCC_FLAGS=-Xfatbin -compress-all
+) ELSE (
+    set NVCC_FLAGS=-D__CUDA_NO_HALF_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_90,code=compute_90
+)
+
+set "CUDA_PATH=%CUDA_PATH_V118%"
+set "PATH=%CUDA_PATH_V118%\bin;%PATH%"
+
+:optcheck
+
+call internal\check_opts.bat
+IF ERRORLEVEL 1 goto :eof
+
+if exist "%NIGHTLIES_PYTORCH_ROOT%" cd %NIGHTLIES_PYTORCH_ROOT%\..
+call  %~dp0\internal\copy.bat
+IF ERRORLEVEL 1 goto :eof
+
+call  %~dp0\internal\setup.bat
+IF ERRORLEVEL 1 goto :eof
--- a/.ci/pytorch/windows/cuda124.bat
+++ b/.ci/pytorch/windows/cuda124.bat
@ -27,24 +27,24 @@ IF "%NVTOOLSEXT_PATH%"=="" (
    )
 )

-IF "%CUDA_PATH_V129%"=="" (
-    IF EXIST "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\bin\nvcc.exe" (
-        set "CUDA_PATH_V129=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9"
+IF "%CUDA_PATH_V124%"=="" (
+    IF EXIST "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin\nvcc.exe" (
+        set "CUDA_PATH_V124=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
    ) ELSE (
-        echo CUDA 12.9 not found, failing
+        echo CUDA 12.4 not found, failing
        exit /b 1
    )
 )

 IF "%BUILD_VISION%" == "" (
-    set TORCH_CUDA_ARCH_LIST=7.5;8.0;8.6;9.0;10.0;12.0
+    set TORCH_CUDA_ARCH_LIST=6.1;7.0;7.5;8.0;8.6;9.0
    set TORCH_NVCC_FLAGS=-Xfatbin -compress-all
 ) ELSE (
-    set NVCC_FLAGS=-D__CUDA_NO_HALF_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_90,code=compute_90 -gencode=arch=compute_100,code=compute_100 -gencode=arch=compute_120,code=compute_120
+    set NVCC_FLAGS=-D__CUDA_NO_HALF_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_90,code=compute_90
 )

-set "CUDA_PATH=%CUDA_PATH_V129%"
-set "PATH=%CUDA_PATH_V129%\bin;%PATH%"
+set "CUDA_PATH=%CUDA_PATH_V124%"
+set "PATH=%CUDA_PATH_V124%\bin;%PATH%"

 :optcheck

--- a/.ci/pytorch/windows/internal/7z_install.bat
+++ b/.ci/pytorch/windows/internal/7z_install.bat
@ -1,6 +1,6 @@
@echo off

-curl -k -L "https://sourceforge.net/projects/sevenzip/files/7-Zip/18.05/7z1805-x64.exe/download" -o 7z1805-x64.exe
+curl -k https://www.7-zip.org/a/7z1805-x64.exe -O
 if errorlevel 1 exit /b 1

 start /wait 7z1805-x64.exe /S
--- a/.ci/pytorch/windows/internal/check_deps.bat
+++ b/.ci/pytorch/windows/internal/check_deps.bat
@ -65,7 +65,7 @@ for /F "usebackq delims=" %%i in (`python -c "import sys; print('{0[0]}{0[1]}'.f
 if  %PYVER% LSS 35 (
    echo Warning: PyTorch for Python 2 under Windows is experimental.
    echo Python x64 3.5 or up is recommended to compile PyTorch on Windows
-    echo Maybe you can create a virtual environment if you have conda installed:
+    echo Maybe you can create a virual environment if you have conda installed:
    echo ^> conda create -n test python=3.6 pyyaml numpy
    echo ^> activate test
 )
--- a/.ci/pytorch/windows/internal/clone.bat
+++ b/.ci/pytorch/windows/internal/clone.bat
@ -8,7 +8,7 @@ goto submodule

 :clone_pytorch

-git clone https://github.com/%PYTORCH_REPO%/%MODULE_NAME% & REM @lint-ignore
+git clone https://github.com/%PYTORCH_REPO%/%MODULE_NAME%

 cd %MODULE_NAME%

--- a/.ci/pytorch/windows/internal/copy.bat
+++ b/.ci/pytorch/windows/internal/copy.bat
@ -8,7 +8,6 @@ copy "%CUDA_PATH%\bin\cusolver*64_*.dll*" pytorch\torch\lib
 copy "%CUDA_PATH%\bin\cudnn*64_*.dll*" pytorch\torch\lib
 copy "%CUDA_PATH%\bin\nvrtc*64_*.dll*" pytorch\torch\lib
 copy "%CUDA_PATH%\extras\CUPTI\lib64\cupti64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\extras\CUPTI\lib64\nvperf_host*.dll*" pytorch\torch\lib

 copy "C:\Program Files\NVIDIA Corporation\NvToolsExt\bin\x64\nvToolsExt64_1.dll*" pytorch\torch\lib
 copy "%PYTHON_LIB_PATH%\libiomp*5md.dll" pytorch\torch\lib
--- a/.ci/pytorch/windows/internal/cuda_install.bat
+++ b/.ci/pytorch/windows/internal/cuda_install.bat
@ -23,20 +23,73 @@ set CUDNN_LIB_FOLDER="lib\x64"
 :: Skip all of this if we already have cuda installed
 if exist "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v%CUDA_VERSION_STR%\bin\nvcc.exe" goto set_cuda_env_vars

+if %CUDA_VER% EQU 118 goto cuda118
+if %CUDA_VER% EQU 124 goto cuda124
 if %CUDA_VER% EQU 126 goto cuda126
 if %CUDA_VER% EQU 128 goto cuda128
-if %CUDA_VER% EQU 129 goto cuda129

 echo CUDA %CUDA_VERSION_STR% is not supported
 exit /b 1

+:cuda118
+
+set CUDA_INSTALL_EXE=cuda_11.8.0_522.06_windows.exe
+if not exist "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" (
+    curl -k -L "https://ossci-windows.s3.amazonaws.com/%CUDA_INSTALL_EXE%" --output "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
+    if errorlevel 1 exit /b 1
+    set "CUDA_SETUP_FILE=%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
+    set "ARGS=cuda_profiler_api_11.8 thrust_11.8 nvcc_11.8 cuobjdump_11.8 nvprune_11.8 nvprof_11.8 cupti_11.8 cublas_11.8 cublas_dev_11.8 cudart_11.8 cufft_11.8 cufft_dev_11.8 curand_11.8 curand_dev_11.8 cusolver_11.8 cusolver_dev_11.8 cusparse_11.8 cusparse_dev_11.8 npp_11.8 npp_dev_11.8 nvrtc_11.8 nvrtc_dev_11.8 nvml_dev_11.8 nvtx_11.8"
+)
+
+set CUDNN_FOLDER=cudnn-windows-x86_64-9.5.0.50_cuda11-archive
+set CUDNN_LIB_FOLDER="lib"
+set "CUDNN_INSTALL_ZIP=%CUDNN_FOLDER%.zip"
+if not exist "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" (
+    curl -k -L "http://s3.amazonaws.com/ossci-windows/%CUDNN_INSTALL_ZIP%" --output "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
+    if errorlevel 1 exit /b 1
+    set "CUDNN_SETUP_FILE=%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
+)
+
+@REM cuDNN 8.3+ required zlib to be installed on the path
+echo Installing ZLIB dlls
+curl -k -L "http://s3.amazonaws.com/ossci-windows/zlib123dllx64.zip" --output "%SRC_DIR%\temp_build\zlib123dllx64.zip"
+7z x "%SRC_DIR%\temp_build\zlib123dllx64.zip" -o"%SRC_DIR%\temp_build\zlib"
+xcopy /Y "%SRC_DIR%\temp_build\zlib\dll_x64\*.dll" "C:\Windows\System32"
+
+goto cuda_common
+
+:cuda124
+
+set CUDA_INSTALL_EXE=cuda_12.4.0_551.61_windows.exe
+if not exist "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" (
+    curl -k -L "https://ossci-windows.s3.amazonaws.com/%CUDA_INSTALL_EXE%" --output "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
+    if errorlevel 1 exit /b 1
+    set "CUDA_SETUP_FILE=%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
+    set "ARGS=cuda_profiler_api_12.4 thrust_12.4 nvcc_12.4 cuobjdump_12.4 nvprune_12.4 nvprof_12.4 cupti_12.4 cublas_12.4 cublas_dev_12.4 cudart_12.4 cufft_12.4 cufft_dev_12.4 curand_12.4 curand_dev_12.4 cusolver_12.4 cusolver_dev_12.4 cusparse_12.4 cusparse_dev_12.4 npp_12.4 npp_dev_12.4 nvrtc_12.4 nvrtc_dev_12.4 nvml_dev_12.4 nvjitlink_12.4 nvtx_12.4"
+)
+
+set CUDNN_FOLDER=cudnn-windows-x86_64-9.5.0.50_cuda12-archive
+set CUDNN_LIB_FOLDER="lib"
+set "CUDNN_INSTALL_ZIP=%CUDNN_FOLDER%.zip"
+if not exist "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" (
+    curl -k -L "http://s3.amazonaws.com/ossci-windows/%CUDNN_INSTALL_ZIP%" --output "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
+    if errorlevel 1 exit /b 1
+    set "CUDNN_SETUP_FILE=%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
+)
+
+@REM cuDNN 8.3+ required zlib to be installed on the path
+echo Installing ZLIB dlls
+curl -k -L "http://s3.amazonaws.com/ossci-windows/zlib123dllx64.zip" --output "%SRC_DIR%\temp_build\zlib123dllx64.zip"
+7z x "%SRC_DIR%\temp_build\zlib123dllx64.zip" -o"%SRC_DIR%\temp_build\zlib"
+xcopy /Y "%SRC_DIR%\temp_build\zlib\dll_x64\*.dll" "C:\Windows\System32"
+
 goto cuda_common

 :cuda126

 set CUDA_INSTALL_EXE=cuda_12.6.2_560.94_windows.exe
 if not exist "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" (
-    curl -k -L "https://ossci-windows.s3.amazonaws.com/%CUDA_INSTALL_EXE%" --output "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" & REM @lint-ignore
+    curl -k -L "https://ossci-windows.s3.amazonaws.com/%CUDA_INSTALL_EXE%" --output "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
    if errorlevel 1 exit /b 1
    set "CUDA_SETUP_FILE=%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
    set "ARGS=cuda_profiler_api_12.6 thrust_12.6 nvcc_12.6 cuobjdump_12.6 nvprune_12.6 nvprof_12.6 cupti_12.6 cublas_12.6 cublas_dev_12.6 cudart_12.6 cufft_12.6 cufft_dev_12.6 curand_12.6 curand_dev_12.6 cusolver_12.6 cusolver_dev_12.6 cusparse_12.6 cusparse_dev_12.6 npp_12.6 npp_dev_12.6 nvrtc_12.6 nvrtc_dev_12.6 nvml_dev_12.6 nvjitlink_12.6 nvtx_12.6"
@ -46,7 +99,7 @@ set CUDNN_FOLDER=cudnn-windows-x86_64-9.5.0.50_cuda12-archive
 set CUDNN_LIB_FOLDER="lib"
 set "CUDNN_INSTALL_ZIP=%CUDNN_FOLDER%.zip"
 if not exist "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" (
-    curl -k -L "http://s3.amazonaws.com/ossci-windows/%CUDNN_INSTALL_ZIP%" --output "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" & REM @lint-ignore
+    curl -k -L "http://s3.amazonaws.com/ossci-windows/%CUDNN_INSTALL_ZIP%" --output "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
    if errorlevel 1 exit /b 1
    set "CUDNN_SETUP_FILE=%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
 )
@ -63,7 +116,7 @@ goto cuda_common

 set CUDA_INSTALL_EXE=cuda_12.8.0_571.96_windows.exe
 if not exist "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" (
-    curl -k -L "https://ossci-windows.s3.amazonaws.com/%CUDA_INSTALL_EXE%" --output "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" & REM @lint-ignore
+    curl -k -L "https://ossci-windows.s3.amazonaws.com/%CUDA_INSTALL_EXE%" --output "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
    if errorlevel 1 exit /b 1
    set "CUDA_SETUP_FILE=%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
    set "ARGS=cuda_profiler_api_12.8 thrust_12.8 nvcc_12.8 cuobjdump_12.8 nvprune_12.8 nvprof_12.8 cupti_12.8 cublas_12.8 cublas_dev_12.8 cudart_12.8 cufft_12.8 cufft_dev_12.8 curand_12.8 curand_dev_12.8 cusolver_12.8 cusolver_dev_12.8 cusparse_12.8 cusparse_dev_12.8 npp_12.8 npp_dev_12.8 nvrtc_12.8 nvrtc_dev_12.8 nvml_dev_12.8 nvjitlink_12.8 nvtx_12.8"
@ -73,34 +126,7 @@ set CUDNN_FOLDER=cudnn-windows-x86_64-9.7.0.66_cuda12-archive
 set CUDNN_LIB_FOLDER="lib"
 set "CUDNN_INSTALL_ZIP=%CUDNN_FOLDER%.zip"
 if not exist "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" (
-    curl -k -L "http://s3.amazonaws.com/ossci-windows/%CUDNN_INSTALL_ZIP%" --output "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" & REM @lint-ignore
-    if errorlevel 1 exit /b 1
-    set "CUDNN_SETUP_FILE=%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
-)
-
-@REM cuDNN 8.3+ required zlib to be installed on the path
-echo Installing ZLIB dlls
-curl -k -L "http://s3.amazonaws.com/ossci-windows/zlib123dllx64.zip" --output "%SRC_DIR%\temp_build\zlib123dllx64.zip"
-7z x "%SRC_DIR%\temp_build\zlib123dllx64.zip" -o"%SRC_DIR%\temp_build\zlib"
-xcopy /Y "%SRC_DIR%\temp_build\zlib\dll_x64\*.dll" "C:\Windows\System32"
-
-goto cuda_common
-
-:cuda129
-
-set CUDA_INSTALL_EXE=cuda_12.9.1_576.57_windows.exe
-if not exist "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" (
-    curl -k -L "https://ossci-windows.s3.amazonaws.com/%CUDA_INSTALL_EXE%" --output "%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%" & REM @lint-ignore
-    if errorlevel 1 exit /b 1
-    set "CUDA_SETUP_FILE=%SRC_DIR%\temp_build\%CUDA_INSTALL_EXE%"
-    set "ARGS=cuda_profiler_api_12.9 thrust_12.9 nvcc_12.9 cuobjdump_12.9 nvprune_12.9 nvprof_12.9 cupti_12.9 cublas_12.9 cublas_dev_12.9 cudart_12.9 cufft_12.9 cufft_dev_12.9 curand_12.9 curand_dev_12.9 cusolver_12.9 cusolver_dev_12.9 cusparse_12.9 cusparse_dev_12.9 npp_12.9 npp_dev_12.9 nvrtc_12.9 nvrtc_dev_12.9 nvml_dev_12.9 nvjitlink_12.9 nvtx_12.9"
-)
-
-set CUDNN_FOLDER=cudnn-windows-x86_64-9.10.2.21_cuda12-archive
-set CUDNN_LIB_FOLDER="lib"
-set "CUDNN_INSTALL_ZIP=%CUDNN_FOLDER%.zip"
-if not exist "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" (
-    curl -k -L "http://s3.amazonaws.com/ossci-windows/%CUDNN_INSTALL_ZIP%" --output "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%" & REM @lint-ignore
+    curl -k -L "http://s3.amazonaws.com/ossci-windows/%CUDNN_INSTALL_ZIP%" --output "%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
    if errorlevel 1 exit /b 1
    set "CUDNN_SETUP_FILE=%SRC_DIR%\temp_build\%CUDNN_INSTALL_ZIP%"
 )
--- a/.ci/pytorch/windows/internal/driver_update.bat
+++ b/.ci/pytorch/windows/internal/driver_update.bat
@ -1,5 +1,5 @@
 set WIN_DRIVER_VN=528.89
-set "DRIVER_DOWNLOAD_LINK=https://ossci-windows.s3.amazonaws.com/%WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe" & REM @lint-ignore
+set "DRIVER_DOWNLOAD_LINK=https://ossci-windows.s3.amazonaws.com/%WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe"
 curl --retry 3 -kL %DRIVER_DOWNLOAD_LINK% --output %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe
 if errorlevel 1 exit /b 1

--- a/.ci/pytorch/windows/internal/install_python.bat
+++ b/.ci/pytorch/windows/internal/install_python.bat
@ -18,5 +18,3 @@ start /wait "" python-amd64.exe /quiet InstallAllUsers=1 PrependPath=0 Include_t
 if errorlevel 1 exit /b 1

 set "PATH=%CD%\Python\Scripts;%CD%\Python;%PATH%"
-%PYTHON_EXEC% -m pip install --upgrade pip setuptools packaging wheel
-if errorlevel 1 exit /b 1
--- a/.ci/pytorch/windows/internal/static_lib_test.bat
+++ b/.ci/pytorch/windows/internal/static_lib_test.bat
@ -1,132 +0,0 @@
-set SRC_DIR=%~dp0
-
-pushd %SRC_DIR%\..
-
-if "%CUDA_VERSION%" == "cpu" call internal\driver_update.bat
-if errorlevel 1 exit /b 1
-
-call internal\cuda_install.bat
-set LIB=%CUDA_PATH%\lib\x64;%LIB%
-if errorlevel 1 exit /b 1
-set "ORIG_PATH=%PATH%"
-
-setlocal EnableDelayedExpansion
-set NVIDIA_GPU_EXISTS=0
-for /F "delims=" %%i in ('wmic path win32_VideoController get name') do (
-    set GPUS=%%i
-    if not "x!GPUS:NVIDIA=!" == "x!GPUS!" (
-        SET NVIDIA_GPU_EXISTS=1
-        goto gpu_check_end
-    )
-)
-:gpu_check_end
-endlocal & set NVIDIA_GPU_EXISTS=%NVIDIA_GPU_EXISTS%
-
-:: Download MAGMA Files on CUDA builds
-set MAGMA_VERSION=2.5.4
-set CUDA_PREFIX=cuda%CUDA_VERSION%
-if "%CUDA_VERSION%" == "92" set MAGMA_VERSION=2.5.2
-if "%CUDA_VERSION%" == "100" set MAGMA_VERSION=2.5.2
-
-if "%DEBUG%" == "1" (
-    set BUILD_TYPE=debug
-) else (
-    set BUILD_TYPE=release
-)
-
-if not "%CUDA_VERSION%" == "cpu" (
-    rmdir /s /q magma_%CUDA_PREFIX%_%BUILD_TYPE%
-    del magma_%CUDA_PREFIX%_%BUILD_TYPE%.7z
-    curl -k https://s3.amazonaws.com/ossci-windows/magma_%MAGMA_VERSION%_%CUDA_PREFIX%_%BUILD_TYPE%.7z -o magma_%CUDA_PREFIX%_%BUILD_TYPE%.7z & REM @lint-ignore
-    7z x -aoa magma_%CUDA_PREFIX%_%BUILD_TYPE%.7z -omagma_%CUDA_PREFIX%_%BUILD_TYPE%
-    set LIB=%CD%\magma_%CUDA_PREFIX%_%BUILD_TYPE%\lib;%LIB%
-)
-
-echo "install conda package"
-
-:: Install Miniconda3
-set "CONDA_HOME=%CD%\conda"
-set "tmp_conda=%CONDA_HOME%"
-set "miniconda_exe=%CD%\miniconda.exe"
-
-rmdir /s /q conda
-del miniconda.exe
-curl -k https://repo.anaconda.com/miniconda/Miniconda3-latest-Windows-x86_64.exe -o "%miniconda_exe%"
-start /wait "" "%miniconda_exe%" /S /InstallationType=JustMe /RegisterPython=0 /AddToPath=0 /D=%tmp_conda%
-if ERRORLEVEL 1 exit /b 1
-
-set "PATH=%CONDA_HOME%;%CONDA_HOME%\scripts;%CONDA_HOME%\Library\bin;%PATH%"
-
-conda create -qyn testenv python=%DESIRED_PYTHON%
-if errorlevel 1 exit /b 1
-
-call %CONDA_HOME%\condabin\activate.bat testenv
-if errorlevel 1 exit /b 1
-
-call conda install  -y -q -c conda-forge libuv=1.39
-call conda install -y -q intel-openmp
-
-echo "install and test libtorch"
-pip install cmake
-echo "installing cmake"
-
-if "%VC_YEAR%" == "2019" powershell internal\vs2019_install.ps1
-if "%VC_YEAR%" == "2022" powershell internal\vs2022_install.ps1
-
-if ERRORLEVEL 1 exit /b 1
-
-for /F "delims=" %%i in ('where /R "%PYTORCH_FINAL_PACKAGE_DIR:/=\%" *-latest.zip') do 7z x "%%i" -otmp
-if ERRORLEVEL 1 exit /b 1
-
-
-pushd tmp\libtorch
-
-set VC_VERSION_LOWER=17
-set VC_VERSION_UPPER=18
-IF "%VC_YEAR%" == "2019" (
-    set VC_VERSION_LOWER=16
-    set VC_VERSION_UPPER=17
-)
-
-for /f "usebackq tokens=*" %%i in (`"%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe" -legacy -products * -version [%VC_VERSION_LOWER%^,%VC_VERSION_UPPER%^) -property installationPath`) do (
-    if exist "%%i" if exist "%%i\VC\Auxiliary\Build\vcvarsall.bat" (
-        set "VS15INSTALLDIR=%%i"
-        set "VS15VCVARSALL=%%i\VC\Auxiliary\Build\vcvarsall.bat"
-        goto vswhere
-    )
-)
-
-:vswhere
-IF "%VS15VCVARSALL%"=="" (
-    echo Visual Studio %VC_YEAR% C++ BuildTools is required to compile PyTorch test on Windows
-    exit /b 1
-)
-call "%VS15VCVARSALL%" x64
-
-set install_root=%CD%
-set INCLUDE=%INCLUDE%;%install_root%\include;%install_root%\include\torch\csrc\api\include
-set LIB=%LIB%;%install_root%\lib\x64
-set PATH=%PATH%;%install_root%\lib
-
-
-cd %PYTORCH_ROOT%\.ci\pytorch\test_example_code\
-mkdir build
-cd build
-
-cmake -DCMAKE_PREFIX_PATH=%install_root% ..
-
-if ERRORLEVEL 1 exit /b 1
-
-cmake --build . --config Release
-
-.\Release\simple-torch-test.exe
-if ERRORLEVEL 1 exit /b 1
-
-popd
-
-echo Cleaning temp files
-rd /s /q "tmp" || ver > nul
-
-:end
-set "PATH=%ORIG_PATH%"
-popd
--- a/.ci/pytorch/windows/internal/vc_install_helper.bat
+++ b/.ci/pytorch/windows/internal/vc_install_helper.bat
@ -3,6 +3,7 @@ if "%VC_YEAR%" == "2022" powershell windows/internal/vs2022_install.ps1
 set VC_VERSION_LOWER=17
 set VC_VERSION_UPPER=18

+
 for /f "usebackq tokens=*" %%i in (`"%ProgramFiles(x86)%\Microsoft Visual Studio\Installer\vswhere.exe"  -products Microsoft.VisualStudio.Product.BuildTools -version [%VC_VERSION_LOWER%^,%VC_VERSION_UPPER%^) -property installationPath`) do (
    if exist "%%i" if exist "%%i\VC\Auxiliary\Build\vcvarsall.bat" (
        set "VS15INSTALLDIR=%%i"
--- a/.ci/pytorch/windows/internal/xpu_install.bat
+++ b/.ci/pytorch/windows/internal/xpu_install.bat
@ -10,23 +10,53 @@ if not "%CUDA_VERSION%" == "xpu" (
 set SRC_DIR=%NIGHTLIES_PYTORCH_ROOT%
 if not exist "%SRC_DIR%\temp_build" mkdir "%SRC_DIR%\temp_build"

+set XPU_INSTALL_MODE=%~1
+if "%XPU_INSTALL_MODE%"=="" goto xpu_bundle_install_start
+if "%XPU_INSTALL_MODE%"=="bundle" goto xpu_bundle_install_start
+if "%XPU_INSTALL_MODE%"=="driver" goto xpu_driver_install_start
+if "%XPU_INSTALL_MODE%"=="all" goto xpu_driver_install_start
+
+:arg_error
+
+echo Illegal XPU installation mode. The value can be "bundle"/"driver"/"all"
+echo If keep the value as space, will use default "bundle" mode
+exit /b 1
+
+:xpu_driver_install_start
+:: TODO Need more testing for driver installation
+set XPU_DRIVER_LINK=https://downloadmirror.intel.com/830975/gfx_win_101.5972.exe
+curl -o xpu_driver.exe --retry 3 --retry-all-errors -k %XPU_DRIVER_LINK%
+echo "XPU Driver installing..."
+start /wait "Intel XPU Driver Installer" "xpu_driver.exe"
+if errorlevel 1 exit /b 1
+del xpu_driver.exe
+if "%XPU_INSTALL_MODE%"=="driver" goto xpu_install_end
+
 :xpu_bundle_install_start

 set XPU_BUNDLE_PARENT_DIR=C:\Program Files (x86)\Intel\oneAPI
-set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/9d6d6c17-ca2d-4735-9331-99447e4a1280/intel-deep-learning-essentials-2025.0.1.28_offline.exe
-set XPU_BUNDLE_PRODUCT_NAME=intel.oneapi.win.deep-learning-essentials.product
-set XPU_BUNDLE_VERSION=2025.0.1+20
+set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/9d1a91e2-e8b8-40a5-8c7f-5db768a6a60c/w_intel-for-pytorch-gpu-dev_p_0.5.3.37_offline.exe
+set XPU_BUNDLE_PRODUCT_NAME=intel.oneapi.win.intel-for-pytorch-gpu-dev.product
+set XPU_BUNDLE_VERSION=0.5.3+31
 set XPU_BUNDLE_INSTALLED=0
 set XPU_BUNDLE_UNINSTALL=0
-set XPU_EXTRA_URL=NULL
-set XPU_EXTRA_PRODUCT_NAME=intel.oneapi.win.compiler.product
-set XPU_EXTRA_VERSION=2025.0.1+1226
+set XPU_EXTRA_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/9d1a91e2-e8b8-40a5-8c7f-5db768a6a60c/w_intel-pti-dev_p_0.9.0.37_offline.exe
+set XPU_EXTRA_PRODUCT_NAME=intel.oneapi.win.intel-pti-dev.product
+set XPU_EXTRA_VERSION=0.9.0+36
 set XPU_EXTRA_INSTALLED=0
 set XPU_EXTRA_UNINSTALL=0

-if not [%XPU_VERSION%]==[] if [%XPU_VERSION%]==[2025.1] (
-    set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/75d4eb97-914a-4a95-852c-7b9733d80f74/intel-deep-learning-essentials-2025.1.3.8_offline.exe
-    set XPU_BUNDLE_VERSION=2025.1.3+5
+if not [%XPU_VERSION%]==[] if [%XPU_VERSION%]==[2025.0] (
+    set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/9d6d6c17-ca2d-4735-9331-99447e4a1280/intel-deep-learning-essentials-2025.0.1.28_offline.exe
+    set XPU_BUNDLE_PRODUCT_NAME=intel.oneapi.win.deep-learning-essentials.product
+    set XPU_BUNDLE_VERSION=2025.0.1+20
+    set XPU_BUNDLE_INSTALLED=0
+    set XPU_BUNDLE_UNINSTALL=0
+    set XPU_EXTRA_URL=NULL
+    set XPU_EXTRA_PRODUCT_NAME=intel.oneapi.win.compiler.product
+    set XPU_EXTRA_VERSION=2025.0.1+1226
+    set XPU_EXTRA_INSTALLED=0
+    set XPU_EXTRA_UNINSTALL=0
 )

 :: Check if XPU bundle is target version or already installed
--- a/.ci/pytorch/windows/xpu.bat
+++ b/.ci/pytorch/windows/xpu.bat
@ -26,7 +26,6 @@ set VS2022INSTALLDIR=%VS15INSTALLDIR%
 set XPU_BUNDLE_ROOT=%ProgramFiles(x86)%\Intel\oneAPI
 call "%XPU_BUNDLE_ROOT%\compiler\latest\env\vars.bat"
 call "%XPU_BUNDLE_ROOT%\ocloc\latest\env\vars.bat"
-set USE_ONEMKL=1
 IF ERRORLEVEL 1 goto :eof

 if exist "%NIGHTLIES_PYTORCH_ROOT%" cd %NIGHTLIES_PYTORCH_ROOT%\..
--- a/.ci/wheel/build_wheel.sh
+++ b/.ci/wheel/build_wheel.sh
@ -206,7 +206,7 @@ if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
    BUILD_LIBTORCH_WHL=1 BUILD_PYTHON_ONLY=0 python setup.py bdist_wheel -d "$whl_tmp_dir"
    echo "Finished setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
    echo "Calling setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
-    BUILD_LIBTORCH_WHL=0 BUILD_PYTHON_ONLY=1 CMAKE_FRESH=1 python setup.py bdist_wheel -d "$whl_tmp_dir"
+    BUILD_PYTHON_ONLY=1 BUILD_LIBTORCH_WHL=0 python setup.py bdist_wheel -d "$whl_tmp_dir" --cmake
    echo "Finished setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
 else
    python setup.py bdist_wheel -d "$whl_tmp_dir"
--- a/.circleci/scripts/binary_linux_test.sh
+++ b/.circleci/scripts/binary_linux_test.sh
@ -101,11 +101,6 @@ if [[ "\$GPU_ARCH_TYPE" != *s390x* && "\$GPU_ARCH_TYPE" != *xpu* && "\$GPU_ARCH_
  else
    python /pytorch/.ci/pytorch/smoke_test/smoke_test.py --package=torchonly --torch-compile-check disabled $extra_parameters
  fi
-
-  if [[ "\$GPU_ARCH_TYPE" != *cpu-aarch64* ]]; then
-    # https://github.com/pytorch/pytorch/issues/149422
-    python /pytorch/.ci/pytorch/smoke_test/check_gomp.py
-  fi
 fi

 # Clean temp files
--- a/.circleci/scripts/binary_populate_env.sh
+++ b/.circleci/scripts/binary_populate_env.sh
@ -75,8 +75,8 @@ TRITON_VERSION=$(cat $PYTORCH_ROOT/.ci/docker/triton_version.txt)
 # Here PYTORCH_EXTRA_INSTALL_REQUIREMENTS is already set for the all the wheel builds hence append TRITON_CONSTRAINT
 TRITON_CONSTRAINT="platform_system == 'Linux' and platform_machine == 'x86_64'"

-# CUDA 12.9 builds have triton for Linux and Linux aarch64 binaries.
-if [[ "$DESIRED_CUDA" == "cu129" ]]; then
+# CUDA 12.8 builds have triton for Linux and Linux aarch64 binaries.
+if [[ "$DESIRED_CUDA" == cu128 ]]; then
  TRITON_CONSTRAINT="platform_system == 'Linux'"
 fi

@ -105,7 +105,6 @@ fi

 # Set triton via PYTORCH_EXTRA_INSTALL_REQUIREMENTS for triton xpu package
 if [[ "$PACKAGE_TYPE" =~ .*wheel.* && -n "$PYTORCH_BUILD_VERSION" && "$PYTORCH_BUILD_VERSION" =~ .*xpu.* ]]; then
-    TRITON_VERSION=$(cat $PYTORCH_ROOT/.ci/docker/triton_xpu_version.txt)
    TRITON_REQUIREMENT="pytorch-triton-xpu==${TRITON_VERSION}"
    if [[ -n "$PYTORCH_BUILD_VERSION" && "$PYTORCH_BUILD_VERSION" =~ .*dev.* ]]; then
        TRITON_SHORTHASH=$(cut -c1-8 $PYTORCH_ROOT/.ci/docker/ci_commit_pins/triton-xpu.txt)
--- a/.circleci/scripts/binary_windows_build.sh
+++ b/.circleci/scripts/binary_windows_build.sh
@ -13,9 +13,8 @@ if [[ "$OS" != "windows-arm64" ]]; then
 fi

 if [[ "$DESIRED_CUDA" == 'xpu' ]]; then
-    export VC_YEAR=2022
    export USE_SCCACHE=0
-    export XPU_VERSION=2025.1
+    export XPU_VERSION=2025.0
    export XPU_ENABLE_KINETO=1
 fi

--- a/.circleci/scripts/binary_windows_test.sh
+++ b/.circleci/scripts/binary_windows_test.sh
@ -7,8 +7,7 @@ export CUDA_VERSION="${DESIRED_CUDA/cu/}"
 export VC_YEAR=2022

 if [[ "$DESIRED_CUDA" == 'xpu' ]]; then
-    export VC_YEAR=2022
-    export XPU_VERSION=2025.1
+    export XPU_VERSION=2025.0
 fi

 pushd "$PYTORCH_ROOT/.ci/pytorch/"
--- a/.circleci/scripts/trigger_azure_pipeline.py
+++ b/.circleci/scripts/trigger_azure_pipeline.py
@ -0,0 +1,157 @@
+# Documentation: https://docs.microsoft.com/en-us/rest/api/azure/devops/build/?view=azure-devops-rest-6.0
+
+import json
+import os
+import re
+import sys
+import time
+
+import requests
+
+
+AZURE_PIPELINE_BASE_URL = "https://aiinfra.visualstudio.com/PyTorch/"
+AZURE_DEVOPS_PAT_BASE64 = os.environ.get("AZURE_DEVOPS_PAT_BASE64_SECRET", "")
+PIPELINE_ID = "911"
+PROJECT_ID = "0628bce4-2d33-499e-bac5-530e12db160f"
+TARGET_BRANCH = os.environ.get("CIRCLE_BRANCH", "main")
+TARGET_COMMIT = os.environ.get("CIRCLE_SHA1", "")
+
+build_base_url = AZURE_PIPELINE_BASE_URL + "_apis/build/builds?api-version=6.0"
+
+s = requests.Session()
+s.headers.update({"Authorization": "Basic " + AZURE_DEVOPS_PAT_BASE64})
+
+
+def submit_build(pipeline_id, project_id, source_branch, source_version):
+    print("Submitting build for branch: " + source_branch)
+    print("Commit SHA1: ", source_version)
+
+    run_build_raw = s.post(
+        build_base_url,
+        json={
+            "definition": {"id": pipeline_id},
+            "project": {"id": project_id},
+            "sourceBranch": source_branch,
+            "sourceVersion": source_version,
+        },
+    )
+
+    try:
+        run_build_json = run_build_raw.json()
+    except json.decoder.JSONDecodeError as e:
+        print(e)
+        print(
+            "Failed to parse the response. Check if the Azure DevOps PAT is incorrect or expired."
+        )
+        sys.exit(-1)
+
+    build_id = run_build_json["id"]
+
+    print("Submitted bulid: " + str(build_id))
+    print("Bulid URL: " + run_build_json["url"])
+    return build_id
+
+
+def get_build(_id):
+    get_build_url = (
+        AZURE_PIPELINE_BASE_URL + f"/_apis/build/builds/{_id}?api-version=6.0"
+    )
+    get_build_raw = s.get(get_build_url)
+    return get_build_raw.json()
+
+
+def get_build_logs(_id):
+    get_build_logs_url = (
+        AZURE_PIPELINE_BASE_URL + f"/_apis/build/builds/{_id}/logs?api-version=6.0"
+    )
+    get_build_logs_raw = s.get(get_build_logs_url)
+    return get_build_logs_raw.json()
+
+
+def get_log_content(url):
+    resp = s.get(url)
+    return resp.text
+
+
+def wait_for_build(_id):
+    build_detail = get_build(_id)
+    build_status = build_detail["status"]
+
+    while build_status == "notStarted":
+        print("Waiting for run to start: " + str(_id))
+        sys.stdout.flush()
+        try:
+            build_detail = get_build(_id)
+            build_status = build_detail["status"]
+        except Exception as e:
+            print("Error getting build")
+            print(e)
+
+        time.sleep(30)
+
+    print("Bulid started: ", str(_id))
+
+    handled_logs = set()
+    while build_status == "inProgress":
+        try:
+            print("Waiting for log: " + str(_id))
+            logs = get_build_logs(_id)
+        except Exception as e:
+            print("Error fetching logs")
+            print(e)
+            time.sleep(30)
+            continue
+
+        for log in logs["value"]:
+            log_id = log["id"]
+            if log_id in handled_logs:
+                continue
+            handled_logs.add(log_id)
+            print("Fetching log: \n" + log["url"])
+            try:
+                log_content = get_log_content(log["url"])
+                print(log_content)
+            except Exception as e:
+                print("Error getting log content")
+                print(e)
+            sys.stdout.flush()
+        build_detail = get_build(_id)
+        build_status = build_detail["status"]
+        time.sleep(30)
+
+    build_result = build_detail["result"]
+
+    print("Bulid status: " + build_status)
+    print("Bulid result: " + build_result)
+
+    return build_status, build_result
+
+
+if __name__ == "__main__":
+    # Convert the branch name for Azure DevOps
+    match = re.search(r"pull/(\d+)", TARGET_BRANCH)
+    if match is not None:
+        pr_num = match.group(1)
+        SOURCE_BRANCH = f"refs/pull/{pr_num}/head"
+    else:
+        SOURCE_BRANCH = f"refs/heads/{TARGET_BRANCH}"
+
+    MAX_RETRY = 2
+    retry = MAX_RETRY
+
+    while retry > 0:
+        build_id = submit_build(PIPELINE_ID, PROJECT_ID, SOURCE_BRANCH, TARGET_COMMIT)
+        build_status, build_result = wait_for_build(build_id)
+
+        if build_result != "succeeded":
+            retry = retry - 1
+            if retry > 0:
+                print("Retrying... remaining attempt: " + str(retry))
+                # Wait a bit before retrying
+                time.sleep((MAX_RETRY - retry) * 120)
+                continue
+            else:
+                print("No more chance to retry. Giving up.")
+                sys.exit(-1)
+        else:
+            break
--- a/.devcontainer/Dockerfile
+++ b/.devcontainer/Dockerfile
@ -1,51 +1,34 @@
-FROM mcr.microsoft.com/vscode/devcontainers/base:ubuntu-22.04
+FROM mcr.microsoft.com/vscode/devcontainers/miniconda:0-3

-# Tools needed for development
-RUN apt-get -y update && \
-    apt-get install -y \
-    build-essential \
-    cmake \
-    ninja-build \
-    git \
-    python3 \
-    python3-pip \
-    python3-dev \
-    python3-venv \
-    libopenblas-dev
+#  I am suprised this is needed
+RUN conda init
+
+# Copy environment.yml (if found) to a temp location so we update the environment. Also
+# copy "noop.txt" so the COPY instruction does not fail if no environment.yml exists.
+COPY .devcontainer/cuda/environment.yml .devcontainer/noop.txt /tmp/conda-tmp/
+RUN if [ -f "/tmp/conda-tmp/environment.yml" ]; then umask 0002 && /opt/conda/bin/conda env update -n base -f /tmp/conda-tmp/environment.yml; fi \
+    && sudo rm -rf /tmp/conda-tmp

 # Tools needed for llvm
-RUN apt-get install --no-install-recommends -y lsb-release wget software-properties-common gnupg && \
-    sudo apt-get clean -y
-
-# Create Python virtual environment
-# RUN python3 -m venv /opt/venv
-# ENV PATH="/opt/venv/bin:$PATH"
-RUN pip3 install --upgrade pip
+RUN sudo apt-get -y update
+RUN sudo apt install -y lsb-release wget software-properties-common gnupg

 # Install CLANG if version is specified
 ARG CLANG_VERSION
 RUN if [ -n "$CLANG_VERSION" ]; then \
-    wget https://apt.llvm.org/llvm.sh; \
+    sudo wget https://apt.llvm.org/llvm.sh; \
    chmod +x llvm.sh; \
-    ./llvm.sh "${CLANG_VERSION}"; \
+    sudo ./llvm.sh "${CLANG_VERSION}"; \
    echo 'export CC=clang' >> ~/.bashrc; \
    echo 'export CXX=clang++' >> ~/.bashrc; \
-    apt-get install --no-install-recommends -y clang libomp-dev && \
-    apt-get clean -y; \
+    sudo apt update; \
+    sudo apt install -y clang; \
+    sudo apt install -y libomp-dev; \
    fi


-# Install CUDA if version is specified
+# Install cuda if version is specified
 ARG CUDA_VERSION
 RUN if [ -n "$CUDA_VERSION" ]; then \
-    CUDA_REPO_VERSION=$(echo ${CUDA_VERSION} | sed 's/\./\-/g'); \
-    wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.0-1_all.deb && \
-    dpkg -i cuda-keyring_1.0-1_all.deb && \
-    apt-get install --no-install-recommends -y cuda-toolkit-${CUDA_VERSION} && \
-    apt-get clean -y; \
+       conda install -y cuda -c "nvidia/label/cuda-${CUDA_VERSION}"; \
    fi
-
-# Set PATH for CUDA
-ENV PATH="/usr/local/cuda/bin:${PATH}"
-ENV LD_LIBRARY_PATH="/usr/local/cuda/lib64:${LD_LIBRARY_PATH}"
-ENV PIP_BREAK_SYSTEM_PACKAGES=1
--- a/.devcontainer/cpu/devcontainer.json
+++ b/.devcontainer/cpu/devcontainer.json
@ -3,7 +3,7 @@
 {
  "name": "PyTorch - CPU",
  "build": {
-    "context": "./",
+    "context": "../..",
    "dockerfile": "../Dockerfile",
    "args": {
      "USERNAME": "vscode",
@ -11,12 +11,6 @@
      "CLANG_VERSION": ""
    }
  },
-  // Mount the full repo only after the container starts
-  "workspaceMount": "source=${localWorkspaceFolder},target=/workspace/pytorch,type=bind,consistency=cached",
-  "workspaceFolder": "/workspace/pytorch",
-  "containerEnv": {
-    "PIP_USER": "0" // <‑‑ disable implicit --user
-  },

  // Features to add to the dev container. More info: https://containers.dev/features.
  "features": {
--- a/.devcontainer/cpu/environment.yml
+++ b/.devcontainer/cpu/environment.yml
@ -0,0 +1,6 @@
+# This environment is specific to Debian
+name: PyTorch
+dependencies:
+  - cmake
+  - ninja
+  - libopenblas
--- a/.devcontainer/cuda/devcontainer.json
+++ b/.devcontainer/cuda/devcontainer.json
@ -3,22 +3,16 @@
 {
  "name": "PyTorch - CUDA",
  "build": {
-    "context": "./",
+    "context": "../..",
    "dockerfile": "../Dockerfile",
    "args": {
      "USERNAME": "vscode",
      "BUILDKIT_INLINE_CACHE": "0",
-      "CUDA_VERSION": "12.8.0",
+      "CUDA_VERSION": "11.8.0",
      "CLANG_VERSION": ""
    }
  },
-  "runArgs": ["--runtime", "nvidia", "--gpus", "all"],
-  // Mount the full repo only after the container starts
-  "workspaceMount": "source=${localWorkspaceFolder},target=/workspace/pytorch,type=bind,consistency=cached",
-  "workspaceFolder": "/workspace/pytorch",
-  "containerEnv": {
-    "PIP_USER": "0" // <‑‑ disable implicit --user
-  },
+  "runArgs": ["--gpus", "all"],
 // Use 'forwardPorts' to make a list of ports inside the container available locally.
  // "forwardPorts": [],

--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Joona Havukainen	4402750a68	Add sequoia ops	2025-05-07 13:40:12 -07:00
Joona Havukainen	21fb7457a9	Add macos15 check	2025-05-07 13:33:50 -07:00
Joona Havukainen	a29349b633	Remove MPSGraph usage from cat_mps_out	2025-05-07 13:30:53 -07:00
Joona Havukainen	51dd0770e4	Synchronize mps backend in the timer	2025-05-05 09:36:25 -07:00
 @ -1 +1 @@
 .3.1
 .3.0