Update (base update)

[ghstack-poisoned]

.bc-linter.yml

@@ -1,15 +0,0 @@
-version: 1
-paths:
-include:
-  - "**/*.py"
-exclude:
-  - ".*"
-  - ".*/**"
-  - "**/.*/**"
-  - "**/.*"
-  - "**/_*/**"
-  - "**/_*.py"
-  - "**/test/**"
-  - "**/benchmarks/**"
-  - "**/test_*.py"
-  - "**/*_test.py"

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -3,18 +3,8 @@ set -eux -o pipefail
 
 GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}
 
-# Set CUDA architecture lists to match x86 build_cuda.sh
-if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0"
-elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
+if [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
-elif [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0+PTX"
-fi
-
-# Compress the fatbin with -compress-mode=size for CUDA 13
-if [[ "$DESIRED_CUDA" == *"13"* ]]; then
-    export TORCH_NVCC_FLAGS="-compress-mode=size"
 fi
 
 SCRIPTPATH="$( cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P )"
@@ -28,22 +18,14 @@ cd /
 # on the mounted pytorch repo
 git config --global --add safe.directory /pytorch
 pip install -r /pytorch/requirements.txt
-pip install auditwheel==6.2.0 wheel
+pip install auditwheel==6.2.0
 if [ "$DESIRED_CUDA" = "cpu" ]; then
     echo "BASE_CUDA_VERSION is not set. Building cpu wheel."
-    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
+    #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
 else
     echo "BASE_CUDA_VERSION is set to: $DESIRED_CUDA"
     export USE_SYSTEM_NCCL=1
-
-    # Check if we should use NVIDIA libs from PyPI (similar to x86 build_cuda.sh logic)
-    if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
-        echo "Bundling CUDA libraries with wheel for aarch64."
-    else
-        echo "Using nvidia libs from pypi for aarch64."
-        echo "Updated PYTORCH_EXTRA_INSTALL_REQUIREMENTS for aarch64: $PYTORCH_EXTRA_INSTALL_REQUIREMENTS"
-        export USE_NVIDIA_PYPI_LIBS=1
-    fi
-
-    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
+    #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

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -13,6 +13,49 @@ def list_dir(path: str) -> list[str]:
     return check_output(["ls", "-1", path]).decode().split("\n")
 
 
+def build_ArmComputeLibrary() -> None:
+    """
+    Using ArmComputeLibrary for aarch64 PyTorch
+    """
+    print("Building Arm Compute Library")
+    acl_build_flags = [
+        "debug=0",
+        "neon=1",
+        "opencl=0",
+        "os=linux",
+        "openmp=1",
+        "cppthreads=0",
+        "arch=armv8a",
+        "multi_isa=1",
+        "fixed_format_kernels=1",
+        "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",
+            ]
+        )
+
+    check_call(
+        ["scons", "Werror=1", f"-j{os.cpu_count()}"] + acl_build_flags,
+        cwd=acl_checkout_dir,
+    )
+    for d in ["arm_compute", "include", "utils", "support", "src", "build"]:
+        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()
@@ -26,186 +69,61 @@ def replace_tag(filename) -> None:
         f.writelines(lines)
 
 
-def patch_library_rpath(
-    folder: str,
-    lib_name: str,
-    use_nvidia_pypi_libs: bool = False,
-    desired_cuda: str = "",
-) -> None:
-    """Apply patchelf to set RPATH for a library in torch/lib"""
-    lib_path = f"{folder}/tmp/torch/lib/{lib_name}"
-
-    if use_nvidia_pypi_libs:
-        # For PyPI NVIDIA libraries, construct CUDA RPATH
-        cuda_rpaths = [
-            "$ORIGIN/../../nvidia/cudnn/lib",
-            "$ORIGIN/../../nvidia/nvshmem/lib",
-            "$ORIGIN/../../nvidia/nccl/lib",
-            "$ORIGIN/../../nvidia/cusparselt/lib",
-        ]
-
-        if "130" in desired_cuda:
-            cuda_rpaths.append("$ORIGIN/../../nvidia/cu13/lib")
-        else:
-            cuda_rpaths.extend(
-                [
-                    "$ORIGIN/../../nvidia/cublas/lib",
-                    "$ORIGIN/../../nvidia/cuda_cupti/lib",
-                    "$ORIGIN/../../nvidia/cuda_nvrtc/lib",
-                    "$ORIGIN/../../nvidia/cuda_runtime/lib",
-                    "$ORIGIN/../../nvidia/cufft/lib",
-                    "$ORIGIN/../../nvidia/curand/lib",
-                    "$ORIGIN/../../nvidia/cusolver/lib",
-                    "$ORIGIN/../../nvidia/cusparse/lib",
-                    "$ORIGIN/../../nvidia/nvtx/lib",
-                    "$ORIGIN/../../nvidia/cufile/lib",
-                ]
-            )
-
-        # Add $ORIGIN for local torch libs
-        rpath = ":".join(cuda_rpaths) + ":$ORIGIN"
-    else:
-        # For bundled libraries, just use $ORIGIN
-        rpath = "$ORIGIN"
-
-    if os.path.exists(lib_path):
-        os.system(
-            f"cd {folder}/tmp/torch/lib/; "
-            f"patchelf --set-rpath '{rpath}' --force-rpath {lib_name}"
-        )
-
-
-def copy_and_patch_library(
-    src_path: str,
-    folder: str,
-    use_nvidia_pypi_libs: bool = False,
-    desired_cuda: str = "",
-) -> None:
-    """Copy a library to torch/lib and patch its RPATH"""
-    if os.path.exists(src_path):
-        lib_name = os.path.basename(src_path)
-        shutil.copy2(src_path, f"{folder}/tmp/torch/lib/{lib_name}")
-        patch_library_rpath(folder, lib_name, use_nvidia_pypi_libs, desired_cuda)
-
-
 def package_cuda_wheel(wheel_path, desired_cuda) -> None:
     """
     Package the cuda wheel libraries
     """
     folder = os.path.dirname(wheel_path)
+    wheelname = os.path.basename(wheel_path)
     os.mkdir(f"{folder}/tmp")
     os.system(f"unzip {wheel_path} -d {folder}/tmp")
-    # Delete original wheel since it will be repackaged
-    os.system(f"rm {wheel_path}")
+    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",
+        "/usr/local/cuda/lib64/libcudart.so.12",
+        "/usr/local/cuda/lib64/libcufft.so.11",
+        "/usr/local/cuda/lib64/libcusparse.so.12",
+        "/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/libnvJitLink.so.12",
+        "/usr/local/cuda/lib64/libnvrtc.so.12",
+        "/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",
+        "/lib64/libgomp.so.1",
+        "/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",
+    ]
 
-    # Check if we should use PyPI NVIDIA libraries or bundle system libraries
-    use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
-
-    if use_nvidia_pypi_libs:
-        print("Using nvidia libs from pypi - skipping CUDA library bundling")
-        # For PyPI approach, we don't bundle CUDA libraries - they come from PyPI packages
-        # We only need to bundle non-NVIDIA libraries
-        minimal_libs_to_copy = [
-            "/lib64/libgomp.so.1",
-            "/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",
-        ]
-
-        # Copy minimal libraries to unzipped_folder/torch/lib
-        for lib_path in minimal_libs_to_copy:
-            copy_and_patch_library(lib_path, folder, use_nvidia_pypi_libs, desired_cuda)
-
-        # Patch torch libraries used for searching libraries
-        torch_libs_to_patch = [
-            "libtorch.so",
-            "libtorch_cpu.so",
-            "libtorch_cuda.so",
-            "libtorch_cuda_linalg.so",
-            "libtorch_global_deps.so",
-            "libtorch_python.so",
-            "libtorch_nvshmem.so",
-            "libc10.so",
-            "libc10_cuda.so",
-            "libcaffe2_nvrtc.so",
-            "libshm.so",
-        ]
-        for lib_name in torch_libs_to_patch:
-            patch_library_rpath(folder, lib_name, use_nvidia_pypi_libs, desired_cuda)
-    else:
-        print("Bundling CUDA libraries with wheel")
-        # Original logic for bundling system CUDA libraries
-        # Common libraries for all CUDA versions
-        common_libs = [
-            # Non-NVIDIA system libraries
-            "/lib64/libgomp.so.1",
-            "/usr/lib64/libgfortran.so.5",
-            "/acl/build/libarm_compute.so",
-            "/acl/build/libarm_compute_graph.so",
-            # Common CUDA libraries (same for all versions)
-            "/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",
-            "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so",
-            "/usr/local/cuda/lib64/libcudnn.so.9",
-            "/usr/local/cuda/lib64/libcusparseLt.so.0",
-            "/usr/local/cuda/lib64/libcurand.so.10",
-            "/usr/local/cuda/lib64/libnccl.so.2",
-            "/usr/local/cuda/lib64/libnvshmem_host.so.3",
-            "/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",
+    if "129" in desired_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/cuda/lib64/libcusparse.so.12",
         ]
 
-        # CUDA version-specific libraries
-        if "13" in desired_cuda:
-            minor_version = desired_cuda[-1]
-            version_specific_libs = [
-                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13",
-                "/usr/local/cuda/lib64/libcublas.so.13",
-                "/usr/local/cuda/lib64/libcublasLt.so.13",
-                "/usr/local/cuda/lib64/libcudart.so.13",
-                "/usr/local/cuda/lib64/libcufft.so.12",
-                "/usr/local/cuda/lib64/libcusolver.so.12",
-                "/usr/local/cuda/lib64/libnvJitLink.so.13",
-                "/usr/local/cuda/lib64/libnvrtc.so.13",
-                f"/usr/local/cuda/lib64/libnvrtc-builtins.so.13.{minor_version}",
-            ]
-        elif "12" in desired_cuda:
-            # Get the last character for libnvrtc-builtins version (e.g., "129" -> "9")
-            minor_version = desired_cuda[-1]
-            version_specific_libs = [
-                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12",
-                "/usr/local/cuda/lib64/libcublas.so.12",
-                "/usr/local/cuda/lib64/libcublasLt.so.12",
-                "/usr/local/cuda/lib64/libcudart.so.12",
-                "/usr/local/cuda/lib64/libcufft.so.11",
-                "/usr/local/cuda/lib64/libcusolver.so.11",
-                "/usr/local/cuda/lib64/libnvJitLink.so.12",
-                "/usr/local/cuda/lib64/libnvrtc.so.12",
-                f"/usr/local/cuda/lib64/libnvrtc-builtins.so.12.{minor_version}",
-            ]
-        else:
-            raise ValueError(f"Unsupported CUDA version: {desired_cuda}.")
-
-        # Combine all libraries
-        libs_to_copy = common_libs + version_specific_libs
-
-        # Copy libraries to unzipped_folder/torch/lib
-        for lib_path in libs_to_copy:
-            copy_and_patch_library(lib_path, folder, use_nvidia_pypi_libs, desired_cuda)
+    # Copy libraries to unzipped_folder/a/lib
+    for lib_path in libs_to_copy:
+        lib_name = os.path.basename(lib_path)
+        shutil.copy2(lib_path, f"{folder}/tmp/torch/lib/{lib_name}")
+        os.system(
+            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/"):
@@ -213,8 +131,14 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
             replace_tag(f"{f.path}/WHEEL")
             break
 
-    os.system(f"wheel pack {folder}/tmp/ -d {folder}")
-    os.system(f"rm -rf {folder}/tmp/")
+    os.mkdir(f"{folder}/cuda_wheel")
+    os.system(f"cd {folder}/tmp/; zip -r {folder}/cuda_wheel/{wheelname} *")
+    shutil.move(
+        f"{folder}/cuda_wheel/{wheelname}",
+        f"{folder}/{wheelname}",
+        copy_function=shutil.copy2,
+    )
+    os.system(f"rm -rf {folder}/tmp/ {folder}/cuda_wheel/")
 
 
 def complete_wheel(folder: str) -> str:
@@ -237,7 +161,14 @@ def complete_wheel(folder: str) -> str:
             f"/{folder}/dist/{repaired_wheel_name}",
         )
     else:
-        repaired_wheel_name = list_dir(f"/{folder}/dist")[0]
+        repaired_wheel_name = wheel_name.replace(
+            "linux_aarch64", "manylinux_2_28_aarch64"
+        )
+        print(f"Renaming {wheel_name} wheel to {repaired_wheel_name}")
+        os.rename(
+            f"/{folder}/dist/{wheel_name}",
+            f"/{folder}/dist/{repaired_wheel_name}",
+        )
 
     print(f"Copying {repaired_wheel_name} to artifacts")
     shutil.copy2(
@@ -274,20 +205,12 @@ if __name__ == "__main__":
     ).decode()
 
     print("Building PyTorch wheel")
-    build_vars = ""
+    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 "
-
-        # Handle PyPI NVIDIA libraries vs bundled libraries
-        use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
-        if use_nvidia_pypi_libs:
-            print("Configuring build for PyPI NVIDIA libraries")
-            # Configure for dynamic linking (matching x86 logic)
-            build_vars += "ATEN_STATIC_CUDA=0 USE_CUDA_STATIC_LINK=0 USE_CUPTI_SO=1 "
-        else:
-            print("Configuring build for bundled NVIDIA libraries")
-            # Keep existing static linking approach - already configured above
+        # nvshmem is broken for aarch64 see https://github.com/pytorch/pytorch/issues/160425
+        build_vars += "USE_NVSHMEM=OFF "
 
     override_package_version = os.getenv("OVERRIDE_PACKAGE_VERSION")
     desired_cuda = os.getenv("DESIRED_CUDA")
@@ -313,13 +236,19 @@ if __name__ == "__main__":
         build_vars += f"BUILD_TEST=0 PYTORCH_BUILD_VERSION={branch[1 : branch.find('-')]} PYTORCH_BUILD_NUMBER=1 "
 
     if enable_mkldnn:
+        build_ArmComputeLibrary()
         print("build pytorch with mkldnn+acl backend")
-        build_vars += "USE_MKLDNN=ON USE_MKLDNN_ACL=ON "
-        build_vars += "ACL_ROOT_DIR=/acl "
+        build_vars += (
+            "USE_MKLDNN=ON USE_MKLDNN_ACL=ON "
+            "ACL_ROOT_DIR=/acl "
+            "LD_LIBRARY_PATH=/pytorch/build/lib:/acl/build:$LD_LIBRARY_PATH "
+            "ACL_INCLUDE_DIR=/acl/build "
+            "ACL_LIBRARY=/acl/build "
+        )
         if enable_cuda:
             build_vars += "BLAS=NVPL "
         else:
-            build_vars += "BLAS=OpenBLAS OpenBLAS_HOME=/opt/OpenBLAS "
+            build_vars += "BLAS=OpenBLAS OpenBLAS_HOME=/OpenBLAS "
     else:
         print("build pytorch without mkldnn backend")
 

.ci/aarch64_linux/build_aarch64_wheel.py

@@ -299,6 +299,40 @@ def install_condaforge_python(host: RemoteHost, python_version="3.8") -> None:
         )
 
 
+def build_OpenBLAS(host: RemoteHost, git_clone_flags: str = "") -> None:
+    print("Building OpenBLAS")
+    host.run_cmd(
+        f"git clone https://github.com/xianyi/OpenBLAS -b v0.3.28 {git_clone_flags}"
+    )
+    make_flags = "NUM_THREADS=64 USE_OPENMP=1 NO_SHARED=1 DYNAMIC_ARCH=1 TARGET=ARMV8"
+    host.run_cmd(
+        f"pushd OpenBLAS && make {make_flags} -j8 && sudo make {make_flags} install && popd && rm -rf OpenBLAS"
+    )
+
+
+def build_ArmComputeLibrary(host: RemoteHost, git_clone_flags: str = "") -> None:
+    print("Building Arm Compute Library")
+    acl_build_flags = " ".join(
+        [
+            "debug=0",
+            "neon=1",
+            "opencl=0",
+            "os=linux",
+            "openmp=1",
+            "cppthreads=0",
+            "arch=armv8a",
+            "multi_isa=1",
+            "fixed_format_kernels=1",
+            "build=native",
+        ]
+    )
+    host.run_cmd(
+        f"git clone https://github.com/ARM-software/ComputeLibrary.git -b v25.02 {git_clone_flags}"
+    )
+
+    host.run_cmd(f"cd ComputeLibrary && scons Werror=1 -j8 {acl_build_flags}")
+
+
 def embed_libgomp(host: RemoteHost, use_conda, wheel_name) -> None:
     host.run_cmd("pip3 install auditwheel")
     host.run_cmd(
@@ -666,6 +700,7 @@ def start_build(
     configure_system(
         host, compiler=compiler, use_conda=use_conda, python_version=python_version
     )
+    build_OpenBLAS(host, git_clone_flags)
 
     if host.using_docker():
         print("Move libgfortant.a into a standard location")
@@ -688,8 +723,6 @@ def start_build(
         f"git clone --recurse-submodules -b {branch} https://github.com/pytorch/pytorch {git_clone_flags}"
     )
 
-    host.run_cmd("pytorch/.ci/docker/common/install_openblas.sh")
-
     print("Building PyTorch wheel")
     build_opts = ""
     if pytorch_build_number is not None:
@@ -710,18 +743,15 @@ def start_build(
     if host.using_docker():
         build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
     if enable_mkldnn:
-        host.run_cmd("pytorch/.ci/docker/common/install_acl.sh")
+        build_ArmComputeLibrary(host, git_clone_flags)
         print("build pytorch with mkldnn+acl backend")
         build_vars += " USE_MKLDNN=ON USE_MKLDNN_ACL=ON"
-        build_vars += " BLAS=OpenBLAS"
-        build_vars += " OpenBLAS_HOME=/opt/OpenBLAS"
-        build_vars += " ACL_ROOT_DIR=/acl"
         host.run_cmd(
-            f"cd $HOME/pytorch && {build_vars} python3 setup.py bdist_wheel{build_opts}"
+            f"cd $HOME/pytorch && export ACL_ROOT_DIR=$HOME/ComputeLibrary && {build_vars} python3 setup.py bdist_wheel{build_opts}"
         )
         print("Repair the wheel")
         pytorch_wheel_name = host.list_dir("pytorch/dist")[0]
-        ld_library_path = "/acl/build:$HOME/pytorch/build/lib"
+        ld_library_path = "$HOME/acl/build:$HOME/pytorch/build/lib"
         host.run_cmd(
             f"export LD_LIBRARY_PATH={ld_library_path} && auditwheel repair $HOME/pytorch/dist/{pytorch_wheel_name}"
         )
@@ -877,7 +907,7 @@ def terminate_instances(instance_type: str) -> None:
 def parse_arguments():
     from argparse import ArgumentParser
 
-    parser = ArgumentParser("Build and test AARCH64 wheels using EC2")
+    parser = ArgumentParser("Builid and test AARCH64 wheels using EC2")
     parser.add_argument("--key-name", type=str)
     parser.add_argument("--debug", action="store_true")
     parser.add_argument("--build-only", action="store_true")

.ci/docker/README.md

@@ -120,8 +120,8 @@ If your new Docker image needs a library installed from a specific pinned commit
    If you're introducing a new argument to the Docker build, make sure to add it in the Docker build step in `.ci/docker/build.sh`:
    ```bash
    docker build \
-     ....
-     --build-arg "NEW_ARG_1=${NEW_ARG_1}"
+      ....
+      --build-arg "NEW_ARG_1=${NEW_ARG_1}"
    ```
 
 3. **Update Dockerfile logic**:

.ci/docker/almalinux/Dockerfile

@@ -64,10 +64,6 @@ FROM cuda as cuda12.9
 RUN bash ./install_cuda.sh 12.9
 ENV DESIRED_CUDA=12.9
 
-FROM cuda as cuda13.0
-RUN bash ./install_cuda.sh 13.0
-ENV DESIRED_CUDA=13.0
-
 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
@@ -80,10 +76,10 @@ ADD ./common/install_mnist.sh install_mnist.sh
 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=cuda13.0  /usr/local/cuda-13.0 /usr/local/cuda-13.0
 
 # Final step
 FROM ${BASE_TARGET} as final

.ci/docker/build.sh

@@ -81,8 +81,8 @@ elif [[ "$image" == *riscv* ]]; then
   DOCKERFILE="ubuntu-cross-riscv/Dockerfile"
 fi
 
-_UCX_COMMIT=7836b165abdbe468a2f607e7254011c07d788152
-_UCC_COMMIT=430e241bf5d38cbc73fc7a6b89155397232e3f96
+_UCX_COMMIT=7bb2722ff2187a0cad557ae4a6afa090569f83fb
+_UCC_COMMIT=20eae37090a4ce1b32bcce6144ccad0b49943e0b
 if [[ "$image" == *rocm* ]]; then
   _UCX_COMMIT=cc312eaa4655c0cc5c2bcd796db938f90563bcf6
   _UCC_COMMIT=0c0fc21559835044ab107199e334f7157d6a0d3d
@@ -114,19 +114,31 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
     ;;
-  pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11)
-    CUDA_VERSION=13.0.0
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.8.1
     ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=11
+    GCC_VERSION=9
     VISION=yes
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks)
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.8.1
-    ANACONDA_PYTHON_VERSION=3.10
+    ANACONDA_PYTHON_VERSION=3.12
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.8.1
+    ANACONDA_PYTHON_VERSION=3.13
     GCC_VERSION=9
     VISION=yes
     KATEX=yes
@@ -156,13 +168,13 @@ case "$tag" in
     TRITON=yes
     ;;
   pytorch-linux-jammy-py3-clang12-onnx)
-    ANACONDA_PYTHON_VERSION=3.10
+    ANACONDA_PYTHON_VERSION=3.9
     CLANG_VERSION=12
     VISION=yes
     ONNX=yes
     ;;
-  pytorch-linux-jammy-py3.10-clang12)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-py3.9-clang12)
+    ANACONDA_PYTHON_VERSION=3.9
     CLANG_VERSION=12
     VISION=yes
     TRITON=yes
@@ -197,24 +209,23 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950"
     ;;
-  pytorch-linux-jammy-xpu-n-1-py3)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-xpu-2025.0-py3)
+    ANACONDA_PYTHON_VERSION=3.9
+    GCC_VERSION=11
+    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
     TRITON=yes
     ;;
-  pytorch-linux-jammy-xpu-n-py3)
-    ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=11
-    VISION=yes
-    XPU_VERSION=2025.2
-    NINJA_VERSION=1.9.0
-    TRITON=yes
-    ;;
-  pytorch-linux-jammy-py3-gcc11-inductor-benchmarks)
-    # TODO (huydhn): Upgrade this to Python >= 3.10
+  pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
     ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
     VISION=yes
@@ -223,8 +234,8 @@ case "$tag" in
     DOCS=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-clang12)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12)
+    ANACONDA_PYTHON_VERSION=3.9
     CUDA_VERSION=12.8.1
     CLANG_VERSION=12
     VISION=yes
@@ -235,8 +246,8 @@ case "$tag" in
     CLANG_VERSION=18
     VISION=yes
     ;;
-  pytorch-linux-jammy-py3.10-gcc11)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-py3.9-gcc11)
+    ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
     VISION=yes
     KATEX=yes
@@ -277,6 +288,7 @@ case "$tag" in
     GCC_VERSION=11
     ACL=yes
     VISION=yes
+    CONDA_CMAKE=yes
     OPENBLAS=yes
     # snadampal: skipping llvm src build install because the current version
     # from pytorch/llvm:9.0.1 is x86 specific
@@ -287,6 +299,7 @@ case "$tag" in
     GCC_VERSION=11
     ACL=yes
     VISION=yes
+    CONDA_CMAKE=yes
     OPENBLAS=yes
     # snadampal: skipping llvm src build install because the current version
     # from pytorch/llvm:9.0.1 is x86 specific

.ci/docker/centos-rocm/Dockerfile

@@ -56,13 +56,9 @@ ENV INSTALLED_VISION ${VISION}
 
 # Install rocm
 ARG ROCM_VERSION
-RUN mkdir ci_commit_pins
-COPY ./common/common_utils.sh common_utils.sh
-COPY ./ci_commit_pins/rocm-composable-kernel.txt ci_commit_pins/rocm-composable-kernel.txt
 COPY ./common/install_rocm.sh install_rocm.sh
 RUN bash ./install_rocm.sh
-RUN rm install_rocm.sh common_utils.sh
-RUN rm -r ci_commit_pins
+RUN rm install_rocm.sh
 COPY ./common/install_rocm_magma.sh install_rocm_magma.sh
 RUN bash ./install_rocm_magma.sh ${ROCM_VERSION}
 RUN rm install_rocm_magma.sh

.ci/docker/ci_commit_pins/huggingface-requirements.txt

@@ -1,2 +0,0 @@
-transformers==4.54.0
-soxr==0.5.0

.ci/docker/ci_commit_pins/huggingface.txt

@@ -0,0 +1 @@
+v4.54.0

.ci/docker/ci_commit_pins/nccl-cu13.txt

@@ -1 +0,0 @@
-v2.27.7-1

.ci/docker/ci_commit_pins/rocm-composable-kernel.txt

@@ -1 +0,0 @@
-7fe50dc3da2069d6645d9deb8c017a876472a977

.ci/docker/ci_commit_pins/torchbench.txt

@@ -1 +1 @@
-74a23feff57432129df84d8099e622773cf77925
+e03a63be43e33596f7f0a43b0f530353785e4a59

.ci/docker/ci_commit_pins/triton-xpu.txt

@@ -1 +1 @@
-1b0418a9a454b2b93ab8d71f40e59d2297157fae
+ae324eeac8e102a2b40370e341460f3791353398

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-fccfc522864cf8bc172abe0cd58ae5581e2d44b9
+f7888497a1eb9e98d4c07537f0d0bcfe180d1363

.ci/docker/common/install_acl.sh

@@ -1,27 +1,16 @@
-#!/bin/bash
-# Script used only in CD pipeline
+set -euo pipefail
 
-set -eux
-
-ACL_VERSION=${ACL_VERSION:-"v25.02"}
-ACL_INSTALL_DIR="/acl"
+readonly version=v25.02
+readonly src_host=https://github.com/ARM-software
+readonly src_repo=ComputeLibrary
 
 # Clone ACL
-git clone https://github.com/ARM-software/ComputeLibrary.git -b "${ACL_VERSION}" --depth 1 --shallow-submodules
+[[ ! -d ${src_repo} ]] && git clone ${src_host}/${src_repo}.git
+cd ${src_repo}
+
+git checkout $version
 
-ACL_CHECKOUT_DIR="ComputeLibrary"
 # Build with scons
-pushd $ACL_CHECKOUT_DIR
 scons -j8  Werror=0 debug=0 neon=1 opencl=0 embed_kernels=0 \
   os=linux arch=armv8a build=native multi_isa=1 \
   fixed_format_kernels=1 openmp=1 cppthreads=0
-popd
-
-# Install ACL
-sudo mkdir -p ${ACL_INSTALL_DIR}
-for d in arm_compute include utils support src build
-do
-  sudo cp -r ${ACL_CHECKOUT_DIR}/${d} ${ACL_INSTALL_DIR}/${d}
-done
-
-rm -rf $ACL_CHECKOUT_DIR

.ci/docker/common/install_cpython.sh

@@ -83,9 +83,9 @@ function build_cpython {
         py_suffix=${py_ver::-1}
         py_folder=$py_suffix
     fi
-    # Update to rc2 due to https://github.com/python/cpython/commit/c72699086fe4
+    # Only b3 is available now
     if [ "$py_suffix" == "3.14.0" ]; then
-        py_suffix="3.14.0rc2"
+        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

.ci/docker/common/install_cuda.sh

@@ -10,7 +10,7 @@ else
   arch_path='sbsa'
 fi
 
-NVSHMEM_VERSION=3.3.24
+NVSHMEM_VERSION=3.3.9
 
 function install_cuda {
   version=$1
@@ -62,16 +62,14 @@ function install_nvshmem {
   mkdir -p "${tmpdir}" && cd "${tmpdir}"
 
   # nvSHMEM license: https://docs.nvidia.com/nvshmem/api/sla.html
-  # This pattern is a lie as it is not consistent across versions, for 3.3.9 it was cuda_ver-arch-nvshhem-ver
-  filename="libnvshmem-linux-${arch_path}-${nvshmem_version}_cuda${cuda_major_version}-archive"
-  suffix=".tar.xz"
-  url="https://developer.download.nvidia.com/compute/nvshmem/redist/libnvshmem/linux-${arch_path}/${filename}${suffix}"
+  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"
 
   # download, unpack, install
   wget -q "${url}"
-  tar xf "${filename}${suffix}"
-  cp -a "${filename}/include/"* /usr/local/cuda/include/
-  cp -a "${filename}/lib/"*     /usr/local/cuda/lib64/
+  tar xf "${filename}.tar.gz"
+  cp -a "libnvshmem/include/"* /usr/local/cuda/include/
+  cp -a "libnvshmem/lib/"*     /usr/local/cuda/lib64/
 
   # cleanup
   cd ..
@@ -128,6 +126,74 @@ function install_129 {
   ldconfig
 }
 
+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 $CUDA_BASE/nsight-systems-2023.4.4/
+}
+
+function prune_126 {
+  echo "Pruning CUDA 12.6"
+  #####################################################################################
+  # CUDA 12.6 prune static libs
+  #####################################################################################
+  export NVPRUNE="/usr/local/cuda-12.6/bin/nvprune"
+  export CUDA_LIB_DIR="/usr/local/cuda-12.6/lib64"
+
+  export GENCODE="-gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90"
+  export GENCODE_CUDNN="-gencode arch=compute_50,code=sm_50 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90"
+
+  if [[ -n "$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 12.6 prune visual tools
+  #####################################################################################
+  export CUDA_BASE="/usr/local/cuda-12.6/"
+  rm -rf $CUDA_BASE/libnvvp $CUDA_BASE/nsightee_plugins $CUDA_BASE/nsight-compute-2024.3.2 $CUDA_BASE/nsight-systems-2024.5.1/
+}
+
 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"
@@ -146,38 +212,18 @@ function install_128 {
   ldconfig
 }
 
-function install_130 {
-  CUDNN_VERSION=9.13.0.50
-  echo "Installing CUDA 13.0 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
-  # install CUDA 13.0 in the same container
-  install_cuda 13.0.0 cuda_13.0.0_580.65.06_linux
-
-  # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
-  install_cudnn 13 $CUDNN_VERSION
-
-  install_nvshmem 13 $NVSHMEM_VERSION
-
-  CUDA_VERSION=13.0 bash install_nccl.sh
-
-  CUDA_VERSION=13.0 bash install_cusparselt.sh
-
-  ldconfig
-}
-
 # idiomatic parameter and option handling in sh
 while test $# -gt 0
 do
     case "$1" in
-    12.4) install_124;
+    12.4) install_124; prune_124
         ;;
-    12.6|12.6.*) install_126;
+    12.6|12.6.*) install_126; prune_126
         ;;
     12.8|12.8.*) install_128;
         ;;
     12.9|12.9.*) install_129;
         ;;
-    13.0|13.0.*) install_130;
-        ;;
     *) echo "bad argument $1"; exit 1
         ;;
     esac

.ci/docker/common/install_cusparselt.sh

@@ -5,15 +5,7 @@ set -ex
 # cuSPARSELt license: https://docs.nvidia.com/cuda/cusparselt/license.html
 mkdir tmp_cusparselt && cd tmp_cusparselt
 
-if [[ ${CUDA_VERSION:0:4} =~ "13" ]]; then
-    arch_path='sbsa'
-    export TARGETARCH=${TARGETARCH:-$(uname -m)}
-    if [ ${TARGETARCH} = 'amd64' ] || [ "${TARGETARCH}" = 'x86_64' ]; then
-        arch_path='x86_64'
-    fi
-    CUSPARSELT_NAME="libcusparse_lt-linux-${arch_path}-0.8.0.4_cuda13-archive"
-    curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cusparselt/redist/libcusparse_lt/linux-${arch_path}/${CUSPARSELT_NAME}.tar.xz
-elif [[ ${CUDA_VERSION:0:4} =~ ^12\.[5-9]$ ]]; then
+if [[ ${CUDA_VERSION:0:4} =~ ^12\.[5-9]$ ]]; then
     arch_path='sbsa'
     export TARGETARCH=${TARGETARCH:-$(uname -m)}
     if [ ${TARGETARCH} = 'amd64' ] || [ "${TARGETARCH}" = 'x86_64' ]; then

.ci/docker/common/install_inductor_benchmark_deps.sh

@@ -5,7 +5,9 @@ set -ex
 source "$(dirname "${BASH_SOURCE[0]}")/common_utils.sh"
 
 function install_huggingface() {
-  pip_install -r huggingface-requirements.txt
+  local version
+  commit=$(get_pinned_commit huggingface)
+  pip_install "git+https://github.com/huggingface/transformers@${commit}"
 }
 
 function install_timm() {
@@ -24,6 +26,9 @@ function install_torchbench() {
 
   python install.py --continue_on_fail
 
+  # soxr comes from https://github.com/huggingface/transformers/pull/39429
+  pip install transformers==4.54.0 soxr==0.5.0
+
   echo "Print all dependencies after TorchBench is installed"
   python -mpip freeze
   popd

.ci/docker/common/install_nccl.sh

@@ -7,8 +7,6 @@ if [[ ${CUDA_VERSION:0:2} == "11" ]]; then
   NCCL_VERSION=$(cat ci_commit_pins/nccl-cu11.txt)
 elif [[ ${CUDA_VERSION:0:2} == "12" ]]; then
   NCCL_VERSION=$(cat ci_commit_pins/nccl-cu12.txt)
-elif [[ ${CUDA_VERSION:0:2} == "13" ]]; then
-  NCCL_VERSION=$(cat ci_commit_pins/nccl-cu13.txt)
 else
   echo "Unexpected CUDA_VERSION ${CUDA_VERSION}"
   exit 1

.ci/docker/common/install_onnx.sh

@@ -19,8 +19,8 @@ pip_install \
   transformers==4.36.2
 
 pip_install coloredlogs packaging
-pip_install onnxruntime==1.22.1
-pip_install onnxscript==0.4.0
+pip_install onnxruntime==1.18.1
+pip_install onnxscript==0.3.1
 
 # 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/

.ci/docker/common/install_openblas.sh

@@ -3,10 +3,8 @@
 
 set -ex
 
-OPENBLAS_VERSION=${OPENBLAS_VERSION:-"v0.3.30"}
-
-# Clone OpenBLAS
-git clone https://github.com/OpenMathLib/OpenBLAS.git -b "${OPENBLAS_VERSION}" --depth 1 --shallow-submodules
+cd /
+git clone https://github.com/OpenMathLib/OpenBLAS.git -b "${OPENBLAS_VERSION:-v0.3.30}" --depth 1 --shallow-submodules
 
 OPENBLAS_CHECKOUT_DIR="OpenBLAS"
 OPENBLAS_BUILD_FLAGS="
@@ -19,7 +17,5 @@ CFLAGS=-O3
 BUILD_BFLOAT16=1
 "
 
-make -j8 ${OPENBLAS_BUILD_FLAGS} -C $OPENBLAS_CHECKOUT_DIR
-sudo make install -C $OPENBLAS_CHECKOUT_DIR
-
-rm -rf $OPENBLAS_CHECKOUT_DIR
+make -j8 ${OPENBLAS_BUILD_FLAGS} -C ${OPENBLAS_CHECKOUT_DIR}
+make -j8 ${OPENBLAS_BUILD_FLAGS} install -C ${OPENBLAS_CHECKOUT_DIR}

.ci/docker/common/install_rocm.sh

@@ -2,11 +2,6 @@
 
 set -ex
 
-# for pip_install function
-source "$(dirname "${BASH_SOURCE[0]}")/common_utils.sh"
-
-ROCM_COMPOSABLE_KERNEL_VERSION="$(cat $(dirname $0)/../ci_commit_pins/rocm-composable-kernel.txt)"
-
 ver() {
     printf "%3d%03d%03d%03d" $(echo "$1" | tr '.' ' ');
 }
@@ -118,8 +113,6 @@ EOF
         rm -rf HIP clr
     fi
 
-    pip_install "git+https://github.com/rocm/composable_kernel@$ROCM_COMPOSABLE_KERNEL_VERSION"
-
     # Cleanup
     apt-get autoclean && apt-get clean
     rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
@@ -183,8 +176,6 @@ install_centos() {
       sqlite3 $kdb "PRAGMA journal_mode=off; PRAGMA VACUUM;"
   done
 
-  pip_install "git+https://github.com/rocm/composable_kernel@$ROCM_COMPOSABLE_KERNEL_VERSION"
-
   # Cleanup
   yum clean all
   rm -rf /var/cache/yum

.ci/docker/common/install_triton.sh

@@ -57,7 +57,7 @@ if [ ! -f setup.py ]; then
   cd python
 fi
 
-pip_install pybind11==3.0.1
+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
 as_jenkins sed -i -e 's/https:\/\/tritonlang.blob.core.windows.net\/llvm-builds/https:\/\/oaitriton.blob.core.windows.net\/public\/llvm-builds/g' setup.py

.ci/docker/common/install_ucc.sh

@@ -44,12 +44,8 @@ function install_ucc() {
 
   ./autogen.sh
 
-  if [[ -n "$CUDA_VERSION"  && $CUDA_VERSION == 13* ]]; then
-    NVCC_GENCODE="-gencode=arch=compute_86,code=compute_86"
-  else
-    # We only run distributed tests on Tesla M60 and A10G
-    NVCC_GENCODE="-gencode=arch=compute_52,code=sm_52 -gencode=arch=compute_86,code=compute_86"
-  fi
+  # We only run distributed tests on Tesla M60 and A10G
+  NVCC_GENCODE="-gencode=arch=compute_52,code=sm_52 -gencode=arch=compute_86,code=compute_86"
 
   if [[ -n "$ROCM_VERSION" ]]; then
     if [[ -n "$PYTORCH_ROCM_ARCH" ]]; then

.ci/docker/common/install_xpu.sh

@@ -65,14 +65,10 @@ function install_ubuntu() {
 
 function install_rhel() {
     . /etc/os-release
-    if [[ "${ID}" == "rhel" ]]; then
-        if [[ ! " 8.8 8.9 9.0 9.2 9.3 " =~ " ${VERSION_ID} " ]]; then
-            echo "RHEL version ${VERSION_ID} not supported"
-            exit
-        fi
-    elif [[ "${ID}" == "almalinux" ]]; then
-        # Workaround for almalinux8 which used by quay.io/pypa/manylinux_2_28_x86_64
-        VERSION_ID="8.8"
+
+    if [[ ! " 8.8 8.10 9.0 9.2 9.3 " =~ " ${VERSION_ID} " ]]; then
+        echo "RHEL version ${VERSION_ID} not supported"
+        exit
     fi
 
     dnf install -y 'dnf-command(config-manager)'
@@ -150,11 +146,11 @@ if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
     XPU_DRIVER_VERSION="/lts/2350"
 fi
 
-# Default use Intel® oneAPI Deep Learning Essentials 2025.1
-if [[ "$XPU_VERSION" == "2025.2" ]]; then
-    XPU_PACKAGES="intel-deep-learning-essentials-2025.2"
-else
+# Default use Intel® oneAPI Deep Learning Essentials 2025.0
+if [[ "$XPU_VERSION" == "2025.1" ]]; then
     XPU_PACKAGES="intel-deep-learning-essentials-2025.1"
+else
+    XPU_PACKAGES="intel-deep-learning-essentials-2025.0"
 fi
 
 # The installation depends on the base OS

.ci/docker/libtorch/Dockerfile

@@ -69,19 +69,6 @@ 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 cuda as cuda13.0
-RUN bash ./install_cuda.sh 13.0
-RUN bash ./install_magma.sh 13.0
-RUN ln -sf /usr/local/cuda-13.0 /usr/local/cuda
-
-# Install libibverbs for libtorch and copy to CUDA directory
-RUN apt-get update -y && \
-    apt-get install -y libibverbs-dev librdmacm-dev && \
-    cp /usr/lib/x86_64-linux-gnu/libmlx5.so* /usr/local/cuda/lib64/ && \
-    cp /usr/lib/x86_64-linux-gnu/librdmacm.so* /usr/local/cuda/lib64/ && \
-    cp /usr/lib/x86_64-linux-gnu/libibverbs.so* /usr/local/cuda/lib64/ && \
-    cp /usr/lib/x86_64-linux-gnu/libnl* /usr/local/cuda/lib64/
-
 FROM cpu as rocm
 ARG ROCM_VERSION
 ARG PYTORCH_ROCM_ARCH

.ci/docker/manywheel/Dockerfile_2_28

@@ -175,6 +175,6 @@ 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 2025.2
+ENV XPU_VERSION 2025.1
 RUN bash ./install_xpu.sh && rm install_xpu.sh
 RUN pushd /opt/_internal && tar -xJf static-libs-for-embedding-only.tar.xz && popd

.ci/docker/manywheel/Dockerfile_2_28_aarch64

@@ -62,13 +62,6 @@ ARG OPENBLAS_VERSION
 ADD ./common/install_openblas.sh install_openblas.sh
 RUN bash ./install_openblas.sh && rm install_openblas.sh
 
-# Install Arm Compute Library
-FROM base as arm_compute
-# use python3.9 to install scons
-RUN python3.9 -m pip install scons==4.7.0
-RUN ln -sf /opt/python/cp39-cp39/bin/scons /usr/local/bin
-COPY ./common/install_acl.sh install_acl.sh
-RUN bash ./install_acl.sh && rm install_acl.sh
 FROM base as final
 
 # remove unnecessary python versions
@@ -77,5 +70,4 @@ 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
 COPY --from=openblas     /opt/OpenBLAS/  /opt/OpenBLAS/
-COPY --from=arm_compute /acl /acl
-ENV LD_LIBRARY_PATH=/opt/OpenBLAS/lib:/acl/build/:$LD_LIBRARY_PATH
+ENV LD_LIBRARY_PATH=/opt/OpenBLAS/lib:$LD_LIBRARY_PATH

.ci/docker/manywheel/build.sh

@@ -28,7 +28,6 @@ fi
 MANY_LINUX_VERSION=${MANY_LINUX_VERSION:-}
 DOCKERFILE_SUFFIX=${DOCKERFILE_SUFFIX:-}
 OPENBLAS_VERSION=${OPENBLAS_VERSION:-}
-ACL_VERSION=${ACL_VERSION:-}
 
 case ${image} in
     manylinux2_28-builder:cpu)
@@ -42,6 +41,7 @@ case ${image} in
         GPU_IMAGE=arm64v8/almalinux:8
         DOCKER_GPU_BUILD_ARG=" --build-arg DEVTOOLSET_VERSION=13 --build-arg NINJA_VERSION=1.12.1"
         MANY_LINUX_VERSION="2_28_aarch64"
+        OPENBLAS_VERSION="v0.3.30"
         ;;
     manylinuxcxx11-abi-builder:cpu-cxx11-abi)
         TARGET=final
@@ -67,12 +67,6 @@ case ${image} in
         DOCKER_GPU_BUILD_ARG="--build-arg BASE_CUDA_VERSION=${GPU_ARCH_VERSION} --build-arg DEVTOOLSET_VERSION=13"
         MANY_LINUX_VERSION="2_28"
         ;;
-    manylinux2_28-builder:cuda13*)
-        TARGET=cuda_final
-        GPU_IMAGE=amd64/almalinux:8
-        DOCKER_GPU_BUILD_ARG="--build-arg BASE_CUDA_VERSION=${GPU_ARCH_VERSION} --build-arg DEVTOOLSET_VERSION=13"
-        MANY_LINUX_VERSION="2_28"
-        ;;
     manylinuxaarch64-builder:cuda*)
         TARGET=cuda_final
         GPU_IMAGE=amd64/almalinux:8
@@ -121,8 +115,7 @@ tmp_tag=$(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=${OPENBLAS_VERSION:-}" \
-    --build-arg "ACL_VERSION=${ACL_VERSION:-}" \
+    --build-arg "OPENBLAS_VERSION=${OPENBLAS_VERSION}" \
     --target "${TARGET}" \
     -t "${tmp_tag}" \
     $@ \

.ci/docker/requirements-ci.txt

@@ -263,6 +263,11 @@ scipy==1.14.1 ; python_version >= "3.12"
 #Pinned versions:
 #test that import:
 
+tb-nightly==2.13.0a20230426
+#Description: TensorBoard
+#Pinned versions:
+#test that import:
+
 # needed by torchgen utils
 typing-extensions>=4.10.0
 #Description: type hints for python
@@ -339,7 +344,7 @@ onnx==1.18.0
 #Pinned versions:
 #test that import:
 
-onnxscript==0.4.0
+onnxscript==0.3.1
 #Description: Required by mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
@@ -379,7 +384,7 @@ dataclasses_json==0.6.7
 cmake==4.0.0
 #Description: required for building
 
-tlparse==0.4.0
+tlparse==0.3.30
 #Description: required for log parsing
 
 cuda-bindings>=12.0,<13.0 ; platform_machine != "s390x"

.ci/docker/requirements-docs.txt

@@ -1,7 +1,7 @@
 sphinx==5.3.0
 #Description: This is used to generate PyTorch docs
 #Pinned versions: 5.3.0
--e git+https://github.com/pytorch/pytorch_sphinx_theme.git@1657ad2fc1acdc98aa719eebecbb0128a7c13ce4#egg=pytorch_sphinx_theme2
+-e git+https://github.com/pytorch/pytorch_sphinx_theme.git@722b7e6f9ca512fcc526ad07d62b3d28c50bb6cd#egg=pytorch_sphinx_theme2
 
 # TODO: sphinxcontrib.katex 0.9.0 adds a local KaTeX server to speed up pre-rendering
 # but it doesn't seem to work and hangs around idly. The initial thought that it is probably

.ci/docker/triton_version.txt

@@ -1 +1 @@
-3.5.0
+3.4.0

.ci/docker/triton_xpu_version.txt

@@ -1 +1 @@
-3.5.0
+3.4.0

.ci/docker/ubuntu-rocm/Dockerfile

@@ -52,13 +52,9 @@ ENV INSTALLED_VISION ${VISION}
 
 # Install rocm
 ARG ROCM_VERSION
-RUN mkdir ci_commit_pins
-COPY ./common/common_utils.sh common_utils.sh
-COPY ./ci_commit_pins/rocm-composable-kernel.txt ci_commit_pins/rocm-composable-kernel.txt
 COPY ./common/install_rocm.sh install_rocm.sh
 RUN bash ./install_rocm.sh
-RUN rm install_rocm.sh common_utils.sh
-RUN rm -r ci_commit_pins
+RUN rm install_rocm.sh
 COPY ./common/install_rocm_magma.sh install_rocm_magma.sh
 RUN bash ./install_rocm_magma.sh ${ROCM_VERSION}
 RUN rm install_rocm_magma.sh
@@ -100,11 +96,11 @@ ARG ANACONDA_PYTHON_VERSION
 ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
-COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
+COPY ci_commit_pins/huggingface.txt huggingface.txt
 COPY ci_commit_pins/timm.txt timm.txt
 COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt torchbench.txt
 
 # (optional) Install non-default Ninja version
 ARG NINJA_VERSION

.ci/docker/ubuntu-xpu/Dockerfile

@@ -56,10 +56,10 @@ RUN rm install_openssl.sh
 ARG INDUCTOR_BENCHMARKS
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
-COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
+COPY ci_commit_pins/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-requirements.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
 
 # Install XPU Dependencies
 ARG XPU_VERSION

.ci/docker/ubuntu/Dockerfile

@@ -66,7 +66,6 @@ ENV NCCL_LIB_DIR="/usr/local/cuda/lib64/"
 # (optional) Install UCC
 ARG UCX_COMMIT
 ARG UCC_COMMIT
-ARG CUDA_VERSION
 ENV UCX_COMMIT $UCX_COMMIT
 ENV UCC_COMMIT $UCC_COMMIT
 ENV UCX_HOME /usr
@@ -97,11 +96,11 @@ RUN rm install_openssl.sh
 ARG INDUCTOR_BENCHMARKS
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
-COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
+COPY ci_commit_pins/huggingface.txt huggingface.txt
 COPY ci_commit_pins/timm.txt timm.txt
 COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt torchbench.txt
 
 ARG TRITON
 ARG TRITON_CPU
@@ -182,6 +181,7 @@ COPY --from=pytorch/llvm:9.0.1 /opt/llvm /opt/llvm
 RUN if [ -n "${SKIP_LLVM_SRC_BUILD_INSTALL}" ]; then set -eu; rm -rf /opt/llvm; fi
 
 # AWS specific CUDA build guidance
+ENV TORCH_CUDA_ARCH_LIST Maxwell
 ENV TORCH_NVCC_FLAGS "-Xfatbin -compress-all"
 ENV CUDA_PATH /usr/local/cuda
 

.ci/libtorch/build.sh

@@ -7,4 +7,4 @@ set -ex
 
 SCRIPTPATH="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 
-USE_NVSHMEM=0 USE_CUSPARSELT=0 BUILD_PYTHONLESS=1 DESIRED_PYTHON="3.9" ${SCRIPTPATH}/../manywheel/build.sh
+USE_CUSPARSELT=0 BUILD_PYTHONLESS=1 DESIRED_PYTHON="3.9" ${SCRIPTPATH}/../manywheel/build.sh

.ci/lumen_cli/cli/build_cli/register_build.py

@@ -2,7 +2,7 @@ import argparse
 import logging
 
 from cli.lib.common.cli_helper import register_targets, RichHelp, TargetSpec
-from cli.lib.core.vllm.vllm_build import VllmBuildRunner
+from cli.lib.core.vllm import VllmBuildRunner
 
 
 logger = logging.getLogger(__name__)

.ci/lumen_cli/cli/lib/common/gh_summary.py

@@ -1,143 +0,0 @@
-from __future__ import annotations
-
-import logging
-import os
-import textwrap
-from pathlib import Path
-from typing import TYPE_CHECKING
-
-from cli.lib.common.utils import get_wheels
-from jinja2 import Template
-
-
-if TYPE_CHECKING:
-    from collections.abc import Iterable, Mapping
-
-
-logger = logging.getLogger(__name__)
-
-_TPL_CONTENT = Template(
-    textwrap.dedent("""\
-    ## {{ title }}
-
-    ```{{ lang }}
-    {{ content }}
-    ```
-""")
-)
-
-_TPL_LIST_ITEMS = Template(
-    textwrap.dedent("""\
-    ## {{ title }}
-    {% for it in items %}
-    - {{ it.pkg }}: {{ it.relpath }}
-    {% else %}
-    _(no item found)_
-    {% endfor %}
-    """)
-)
-
-_TPL_TABLE = Template(
-    textwrap.dedent("""\
-    {%- if rows %}
-    | {{ cols | join(' | ') }} |
-    |{%- for _ in cols %} --- |{%- endfor %}
-    {%- for r in rows %}
-    | {%- for c in cols %} {{ r.get(c, "") }} |{%- endfor %}
-    {%- endfor %}
-    {%- else %}
-    _(no data)_
-    {%- endif %}
-""")
-)
-
-
-def gh_summary_path() -> Path | None:
-    """Return the Path to the GitHub step summary file, or None if not set."""
-    p = os.environ.get("GITHUB_STEP_SUMMARY")
-    return Path(p) if p else None
-
-
-def write_gh_step_summary(md: str, *, append_content: bool = True) -> bool:
-    """
-    Write Markdown content to the GitHub Step Summary file if GITHUB_STEP_SUMMARY is set.
-    append_content: default true, if True, append to the end of the file, else overwrite the whole file
-
-    Returns:
-        True if written successfully (in GitHub Actions environment),
-        False if skipped (e.g., running locally where the variable is not set).
-    """
-    sp = gh_summary_path()
-    if not sp:
-        logger.info("[gh-summary] GITHUB_STEP_SUMMARY not set, skipping write.")
-        return False
-
-    md_clean = textwrap.dedent(md).strip() + "\n"
-
-    mode = "a" if append_content else "w"
-    with sp.open(mode, encoding="utf-8") as f:
-        f.write(md_clean)
-    return True
-
-
-def md_heading(text: str, level: int = 2) -> str:
-    """Generate a Markdown heading string with the given level (1-6)."""
-    return f"{'#' * max(1, min(level, 6))} {text}\n"
-
-
-def md_details(summary: str, content: str) -> str:
-    """Generate a collapsible <details> block with a summary and inner content."""
-    return f"<details>\n<summary>{summary}</summary>\n\n{content}\n\n</details>\n"
-
-
-def summarize_content_from_file(
-    output_dir: Path,
-    freeze_file: str,
-    title: str = "Content from file",
-    code_lang: str = "",  # e.g. "text" or "ini"
-) -> bool:
-    f = Path(output_dir) / freeze_file
-    if not f.exists():
-        return False
-    content = f.read_text(encoding="utf-8").strip()
-    md = render_content(content, title=title, lang=code_lang)
-    return write_gh_step_summary(md)
-
-
-def summarize_wheels(path: Path, title: str = "Wheels", max_depth: int = 3):
-    items = get_wheels(path, max_depth=max_depth)
-    if not items:
-        return False
-    md = render_list(items, title=title)
-    return write_gh_step_summary(md)
-
-
-def md_kv_table(rows: Iterable[Mapping[str, str | int | float]]) -> str:
-    """
-    Render a list of dicts as a Markdown table using Jinja template.
-    """
-    rows = list(rows)
-    cols = list({k for r in rows for k in r.keys()})
-    md = _TPL_TABLE.render(cols=cols, rows=rows).strip() + "\n"
-    return md
-
-
-def render_list(
-    items: Iterable[str],
-    *,
-    title: str = "List",
-) -> str:
-    tpl = _TPL_LIST_ITEMS
-    md = tpl.render(title=title, items=items)
-    return md
-
-
-def render_content(
-    content: str,
-    *,
-    title: str = "Content",
-    lang: str = "text",
-) -> str:
-    tpl = _TPL_CONTENT
-    md = tpl.render(title=title, content=content, lang=lang)
-    return md

.ci/lumen_cli/cli/lib/common/git_helper.py

@@ -45,7 +45,7 @@ def clone_external_repo(target: str, repo: str, dst: str = "", update_submodules
 
         # Checkout pinned commit
         commit = get_post_build_pinned_commit(target)
-        logger.info("Checking out pinned %s commit %s", target, commit)
+        logger.info("Checking out pinned commit %s", commit)
         r.git.checkout(commit)
 
         # Update submodules if requested
@@ -55,7 +55,7 @@ def clone_external_repo(target: str, repo: str, dst: str = "", update_submodules
                 sm.update(init=True, recursive=True, progress=PrintProgress())
 
         logger.info("Successfully cloned %s", target)
-        return r, commit
+        return r
 
     except GitCommandError as e:
         logger.error("Git operation failed: %s", e)

.ci/lumen_cli/cli/lib/common/pip_helper.py

@@ -1,71 +0,0 @@
-import glob
-import logging
-import shlex
-import shutil
-import sys
-from collections.abc import Iterable
-from importlib.metadata import PackageNotFoundError, version  # noqa: UP035
-from typing import Optional, Union
-
-from cli.lib.common.utils import run_command
-
-
-logger = logging.getLogger(__name__)
-
-
-def pip_install_packages(
-    packages: Iterable[str] = (),
-    env=None,
-    *,
-    requirements: Optional[str] = None,
-    constraints: Optional[str] = None,
-    prefer_uv: bool = False,
-) -> None:
-    use_uv = prefer_uv and shutil.which("uv") is not None
-    base = (
-        [sys.executable, "-m", "uv", "pip", "install"]
-        if use_uv
-        else [sys.executable, "-m", "pip", "install"]
-    )
-    cmd = base[:]
-    if requirements:
-        cmd += ["-r", requirements]
-    if constraints:
-        cmd += ["-c", constraints]
-    cmd += list(packages)
-    logger.info("pip installing packages: %s", " ".join(map(shlex.quote, cmd)))
-    run_command(" ".join(map(shlex.quote, cmd)), env=env)
-
-
-def pip_install_first_match(pattern: str, extras: Optional[str] = None, pref_uv=False):
-    wheel = first_matching_pkg(pattern)
-    target = f"{wheel}[{extras}]" if extras else wheel
-    logger.info("Installing %s...", target)
-    pip_install_packages([target], prefer_uv=pref_uv)
-
-
-def run_python(args: Union[str, list[str]], env=None):
-    """
-    Run the python in the current environment.
-    """
-    if isinstance(args, str):
-        args = shlex.split(args)
-    cmd = [sys.executable] + args
-    run_command(" ".join(map(shlex.quote, cmd)), env=env)
-
-
-def pkg_exists(name: str) -> bool:
-    try:
-        pkg_version = version(name)
-        logger.info("%s already exist with version: %s", name, pkg_version)
-        return True
-    except PackageNotFoundError:
-        logger.info("%s is not installed", name)
-        return False
-
-
-def first_matching_pkg(pattern: str) -> str:
-    matches = sorted(glob.glob(pattern))
-    if not matches:
-        raise FileNotFoundError(f"No wheel matching: {pattern}")
-    return matches[0]

.ci/lumen_cli/cli/lib/common/utils.py

@@ -7,8 +7,6 @@ import os
 import shlex
 import subprocess
 import sys
-from contextlib import contextmanager
-from pathlib import Path
 from typing import Optional
 
 
@@ -79,61 +77,3 @@ def str2bool(value: Optional[str]) -> bool:
     if value in false_value_set:
         return False
     raise ValueError(f"Invalid string value for boolean conversion: {value}")
-
-
-@contextmanager
-def temp_environ(updates: dict[str, str]):
-    """
-    Temporarily set environment variables and restore them after the block.
-    Args:
-        updates: Dict of environment variables to set.
-    """
-    missing = object()
-    old: dict[str, str | object] = {k: os.environ.get(k, missing) for k in updates}
-    try:
-        os.environ.update(updates)
-        yield
-    finally:
-        for k, v in old.items():
-            if v is missing:
-                os.environ.pop(k, None)
-            else:
-                os.environ[k] = v  # type: ignore[arg-type]
-
-
-@contextmanager
-def working_directory(path: str):
-    """
-    Temporarily change the working directory inside a context.
-    """
-    if not path:
-        # No-op context
-        yield
-        return
-    prev_cwd = os.getcwd()
-    try:
-        os.chdir(path)
-        yield
-    finally:
-        os.chdir(prev_cwd)
-
-
-def get_wheels(
-    output_dir: Path,
-    max_depth: Optional[int] = None,
-) -> list[str]:
-    """Return a list of wheels found in the given output directory."""
-    root = Path(output_dir)
-    if not root.exists():
-        return []
-    items = []
-    for dirpath, _, filenames in os.walk(root):
-        depth = Path(dirpath).relative_to(root).parts
-        if max_depth is not None and len(depth) > max_depth:
-            continue
-        for fname in sorted(filenames):
-            if fname.endswith(".whl"):
-                pkg = fname.split("-")[0]
-                relpath = str((Path(dirpath) / fname).relative_to(root))
-                items.append({"pkg": pkg, "relpath": relpath})
-    return items

.ci/lumen_cli/cli/lib/core/vllm.py

@@ -13,11 +13,7 @@ from cli.lib.common.envs_helper import (
     env_str_field,
     with_params_help,
 )
-from cli.lib.common.gh_summary import (
-    gh_summary_path,
-    summarize_content_from_file,
-    summarize_wheels,
-)
+from cli.lib.common.git_helper import clone_external_repo
 from cli.lib.common.path_helper import (
     copy,
     ensure_dir_exists,
@@ -26,7 +22,6 @@ from cli.lib.common.path_helper import (
     is_path_exist,
 )
 from cli.lib.common.utils import run_command
-from cli.lib.core.vllm.lib import clone_vllm, summarize_build_info
 
 
 logger = logging.getLogger(__name__)
@@ -47,7 +42,7 @@ class VllmBuildParameters:
     """
 
     # USE_TORCH_WHEEL: when true, use local Torch wheels; requires TORCH_WHEELS_PATH.
-    # Otherwise docker build pull torch nightly during build
+    #  Otherwise docker build pull torch nightly during build
     # TORCH_WHEELS_PATH: directory containing local torch wheels when use_torch_whl is True
     use_torch_whl: bool = env_bool_field("USE_TORCH_WHEEL", True)
     torch_whls_path: Path = env_path_field("TORCH_WHEELS_PATH", "./dist")
@@ -67,7 +62,7 @@ class VllmBuildParameters:
     )
 
     # OUTPUT_DIR: where docker buildx (local exporter) will write artifacts
-    output_dir: Path = env_path_field("OUTPUT_DIR", "external/vllm")
+    output_dir: Path = env_path_field("OUTPUT_DIR", "shared")
 
     # --- Build args ----------------------------------------------------------
     target_stage: str = env_str_field("TARGET_STAGE", "export-wheels")
@@ -157,44 +152,18 @@ class VllmBuildRunner(BaseRunner):
         3. run docker build
         """
         inputs = VllmBuildParameters()
-        logger.info("Running vllm build with inputs: %s", inputs)
-        vllm_commit = clone_vllm()
+        clone_vllm()
 
         self.cp_dockerfile_if_exist(inputs)
+
         # cp torch wheels from root direct to vllm workspace if exist
         self.cp_torch_whls_if_exist(inputs)
 
-        # make sure the output dir to store the build artifacts exist
-        ensure_dir_exists(Path(inputs.output_dir))
+        ensure_dir_exists(inputs.output_dir)
 
         cmd = self._generate_docker_build_cmd(inputs)
         logger.info("Running docker build: \n %s", cmd)
-
-        try:
-            run_command(cmd, cwd="vllm", env=os.environ.copy())
-        finally:
-            self.genearte_vllm_build_summary(vllm_commit, inputs)
-
-    def genearte_vllm_build_summary(
-        self, vllm_commit: str, inputs: VllmBuildParameters
-    ):
-        if not gh_summary_path():
-            return logger.info("Skipping, not detect GH Summary env var....")
-        logger.info("Generate GH Summary ...")
-        # summarize vllm build info
-        summarize_build_info(vllm_commit)
-
-        # summarize vllm build artifacts
-        vllm_artifact_dir = inputs.output_dir / "wheels"
-        summarize_content_from_file(
-            vllm_artifact_dir,
-            "build_summary.txt",
-            title="Vllm build env pip package summary",
-        )
-        summarize_wheels(
-            inputs.torch_whls_path, max_depth=3, title="Torch Wheels Artifacts"
-        )
-        summarize_wheels(vllm_artifact_dir, max_depth=3, title="Vllm Wheels Artifacts")
+        run_command(cmd, cwd="vllm", env=os.environ.copy())
 
     def cp_torch_whls_if_exist(self, inputs: VllmBuildParameters) -> str:
         if not inputs.use_torch_whl:
@@ -283,3 +252,12 @@ class VllmBuildRunner(BaseRunner):
                 --progress=plain .
         """
         ).strip()
+
+
+def clone_vllm():
+    clone_external_repo(
+        target="vllm",
+        repo="https://github.com/vllm-project/vllm.git",
+        dst="vllm",
+        update_submodules=True,
+    )

.ci/lumen_cli/cli/lib/core/vllm/lib.py

@@ -1,296 +0,0 @@
-import logging
-import os
-import textwrap
-from typing import Any
-
-from cli.lib.common.gh_summary import write_gh_step_summary
-from cli.lib.common.git_helper import clone_external_repo
-from cli.lib.common.pip_helper import pip_install_packages
-from cli.lib.common.utils import run_command, temp_environ, working_directory
-from jinja2 import Template
-
-
-logger = logging.getLogger(__name__)
-
-_TPL_VLLM_INFO = Template(
-    textwrap.dedent("""\
-    ##  Vllm against Pytorch CI Test Summary
-    **Vllm Commit**: [{{ vllm_commit }}](https://github.com/vllm-project/vllm/commit/{{ vllm_commit }})
-    {%- if torch_sha %}
-    **Pytorch Commit**: [{{ torch_sha }}](https://github.com/pytorch/pytorch/commit/{{ torch_sha }})
-    {%- endif %}
-""")
-)
-
-
-def sample_vllm_test_library():
-    """
-    Simple sample to unblock the vllm ci development, which is mimic to
-    https://github.com/vllm-project/vllm/blob/main/.buildkite/test-pipeline.yaml
-    see run_test_plan for more details
-    """
-    # TODO(elainewy): Read from yaml file to handle the env and tests for vllm
-    return {
-        "vllm_basic_correctness_test": {
-            "title": "Basic Correctness Test",
-            "id": "vllm_basic_correctness_test",
-            "env_vars": {
-                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
-            },
-            "steps": [
-                "pytest -v -s basic_correctness/test_cumem.py",
-                "pytest -v -s basic_correctness/test_basic_correctness.py",
-                "pytest -v -s basic_correctness/test_cpu_offload.py",
-                "VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py",
-            ],
-        },
-        "vllm_basic_models_test": {
-            "title": "Basic models test",
-            "id": "vllm_basic_models_test",
-            "steps": [
-                "pytest -v -s models/test_transformers.py",
-                "pytest -v -s models/test_registry.py",
-                "pytest -v -s models/test_utils.py",
-                "pytest -v -s models/test_vision.py",
-                "pytest -v -s models/test_initialization.py",
-            ],
-        },
-        "vllm_entrypoints_test": {
-            "title": "Entrypoints Test ",
-            "id": "vllm_entrypoints_test",
-            "env_vars": {
-                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
-            },
-            "steps": [
-                " ".join(
-                    [
-                        "pytest",
-                        "-v",
-                        "-s",
-                        "entrypoints/llm",
-                        "--ignore=entrypoints/llm/test_lazy_outlines.py",
-                        "--ignore=entrypoints/llm/test_generate.py",
-                        "--ignore=entrypoints/llm/test_generate_multiple_loras.py",
-                        "--ignore=entrypoints/llm/test_collective_rpc.py",
-                    ]
-                ),
-                "pytest -v -s entrypoints/llm/test_lazy_outlines.py",
-                "pytest -v -s entrypoints/llm/test_generate.py ",
-                "VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode",
-            ],
-        },
-        "vllm_regression_test": {
-            "title": "Regression Test",
-            "id": "vllm_regression_test",
-            "package_install": ["modelscope"],
-            "steps": [
-                "pytest -v -s test_regression.py",
-            ],
-        },
-        "vllm_lora_tp_test_distributed": {
-            "title": "LoRA TP Test (Distributed)",
-            "id": "vllm_lora_tp_test_distributed",
-            "env_vars": {
-                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
-            },
-            "num_gpus": 4,
-            "steps": [
-                "pytest -v -s -x lora/test_chatglm3_tp.py",
-                "pytest -v -s -x lora/test_llama_tp.py",
-                "pytest -v -s -x lora/test_llm_with_multi_loras.py",
-            ],
-        },
-        "vllm_distributed_test_28_failure_test": {
-            "title": "Distributed Tests (2 GPUs) pytorch 2.8 release failure",
-            "id": "vllm_distributed_test_28_failure_test",
-            "env_vars": {
-                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
-            },
-            "num_gpus": 4,
-            "steps": [
-                "pytest -v -s distributed/test_sequence_parallel.py",
-            ],
-        },
-        "vllm_lora_28_failure_test": {
-            "title": "LoRA pytorch 2.8 failure test",
-            "id": "vllm_lora_28_failure_test",
-            "steps": ["pytest -v lora/test_quant_model.py"],
-        },
-        "vllm_multi_model_processor_test": {
-            "title": "Multi-Modal Processor Test",
-            "id": "vllm_multi_model_processor_test",
-            "package_install": ["git+https://github.com/TIGER-AI-Lab/Mantis.git"],
-            "steps": [
-                "pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py",
-            ],
-        },
-        "vllm_multi_model_test_28_failure_test": {
-            "title": "Multi-Model Test (Failed 2.8 release)",
-            "id": "vllm_multi_model_test_28_failure_test",
-            "package_install": ["git+https://github.com/TIGER-AI-Lab/Mantis.git"],
-            "steps": [
-                "pytest -v -s models/multimodal/generation/test_voxtral.py",
-                "pytest -v -s models/multimodal/pooling",
-            ],
-        },
-        "vllm_pytorch_compilation_unit_tests": {
-            "title": "PyTorch Compilation Unit Tests",
-            "id": "vllm_pytorch_compilation_unit_tests",
-            "steps": [
-                "pytest -v -s compile/test_pass_manager.py",
-                "pytest -v -s compile/test_fusion.py",
-                "pytest -v -s compile/test_fusion_attn.py",
-                "pytest -v -s compile/test_silu_mul_quant_fusion.py",
-                "pytest -v -s compile/test_sequence_parallelism.py",
-                "pytest -v -s compile/test_async_tp.py",
-                "pytest -v -s compile/test_fusion_all_reduce.py",
-                "pytest -v -s compile/test_decorator.py",
-            ],
-        },
-        "vllm_languagde_model_test_extended_generation_28_failure_test": {
-            "title": "Language Models Test (Extended Generation) 2.8 release failure",
-            "id": "vllm_languagde_model_test_extended_generation_28_failure_test",
-            "package_install": [
-                "--no-build-isolation",
-                "git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8",
-            ],
-            "steps": [
-                "pytest -v -s models/language/generation/test_mistral.py",
-            ],
-        },
-        "vllm_distributed_test_2_gpu_28_failure_test": {
-            "title": "Distributed Tests (2 GPUs) pytorch 2.8 release failure",
-            "id": "vllm_distributed_test_2_gpu_28_failure_test",
-            "env_vars": {
-                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
-            },
-            "num_gpus": 4,
-            "steps": [
-                "pytest -v -s distributed/test_sequence_parallel.py",
-            ],
-        },
-        # TODO(elainewy):need to add g6 with 4 gpus to run this test
-        "vllm_lora_test": {
-            "title": "LoRA Test %N",
-            "id": "lora_test",
-            "parallelism": 4,
-            "steps": [
-                "echo '[checking] list sharded lora tests:'",
-                " ".join(
-                    [
-                        "pytest -q --collect-only lora",
-                        "--shard-id=$$BUILDKITE_PARALLEL_JOB",
-                        "--num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT",
-                        "--ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py",
-                    ]
-                ),
-                "echo '[checking] Done. list lora tests'",
-                " ".join(
-                    [
-                        "pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB",
-                        "--num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT",
-                        "--ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py",
-                    ]
-                ),
-            ],
-        },
-    }
-
-
-def check_parallelism(tests: Any, title: str, shard_id: int = 0, num_shards: int = 0):
-    """
-    a method to check if the test plan is parallelism or not.
-    """
-    parallelism = int(tests.get("parallelism", "0"))
-    is_parallel = parallelism and parallelism > 1
-
-    if not is_parallel:
-        return False
-
-    if shard_id > num_shards:
-        raise RuntimeError(
-            f"Test {title} expects {num_shards} shards, but invalid {shard_id} is provided"
-        )
-
-    if num_shards != parallelism:
-        raise RuntimeError(
-            f"Test {title} expects {parallelism} shards, but invalid {num_shards} is provided"
-        )
-
-    return True
-
-
-def run_test_plan(
-    test_plan: str,
-    test_target: str,
-    tests_map: dict[str, Any],
-    shard_id: int = 0,
-    num_shards: int = 0,
-):
-    """
-    a method to run list of tests based on the test plan.
-    """
-    logger.info("run %s tests.....", test_target)
-    if test_plan not in tests_map:
-        raise RuntimeError(
-            f"test {test_plan} not found, please add it to test plan pool"
-        )
-    tests = tests_map[test_plan]
-    pkgs = tests.get("package_install", [])
-    title = tests.get("title", "unknown test")
-
-    is_parallel = check_parallelism(tests, title, shard_id, num_shards)
-    if is_parallel:
-        title = title.replace("%N", f"{shard_id}/{num_shards}")
-
-    logger.info("Running tests: %s", title)
-    if pkgs:
-        logger.info("Installing packages: %s", pkgs)
-        pip_install_packages(packages=pkgs, prefer_uv=True)
-    with (
-        working_directory(tests.get("working_directory", "tests")),
-        temp_environ(tests.get("env_vars", {})),
-    ):
-        failures = []
-        for step in tests["steps"]:
-            logger.info("Running step: %s", step)
-            if is_parallel:
-                step = replace_buildkite_placeholders(step, shard_id, num_shards)
-                logger.info("Running parallel step: %s", step)
-            code = run_command(cmd=step, check=False, use_shell=True)
-            if code != 0:
-                failures.append(step)
-            logger.info("Finish running step: %s", step)
-        if failures:
-            logger.error("Failed tests: %s", failures)
-            raise RuntimeError(f"{len(failures)} pytest runs failed: {failures}")
-        logger.info("Done. All tests passed")
-
-
-def clone_vllm(dst: str = "vllm"):
-    _, commit = clone_external_repo(
-        target="vllm",
-        repo="https://github.com/vllm-project/vllm.git",
-        dst=dst,
-        update_submodules=True,
-    )
-    return commit
-
-
-def replace_buildkite_placeholders(step: str, shard_id: int, num_shards: int) -> str:
-    mapping = {
-        "$$BUILDKITE_PARALLEL_JOB_COUNT": str(num_shards),
-        "$$BUILDKITE_PARALLEL_JOB": str(shard_id),
-    }
-    for k in sorted(mapping, key=len, reverse=True):
-        step = step.replace(k, mapping[k])
-    return step
-
-
-def summarize_build_info(vllm_commit: str) -> bool:
-    torch_sha = os.getenv("GITHUB_SHA")
-    md = (
-        _TPL_VLLM_INFO.render(vllm_commit=vllm_commit, torch_sha=torch_sha).strip()
-        + "\n"
-    )
-    return write_gh_step_summary(md)

.ci/lumen_cli/cli/lib/core/vllm/vllm_test.py

@@ -1,269 +0,0 @@
-import logging
-import os
-import re
-import subprocess
-import sys
-from collections.abc import Iterable
-from dataclasses import dataclass
-from enum import Enum
-from pathlib import Path
-from typing import Any
-
-from cli.lib.common.cli_helper import BaseRunner
-from cli.lib.common.envs_helper import env_path_field, env_str_field, get_env
-from cli.lib.common.path_helper import copy, remove_dir
-from cli.lib.common.pip_helper import (
-    pip_install_first_match,
-    pip_install_packages,
-    pkg_exists,
-    run_python,
-)
-from cli.lib.common.utils import run_command, working_directory
-from cli.lib.core.vllm.lib import clone_vllm, run_test_plan, sample_vllm_test_library
-
-
-logger = logging.getLogger(__name__)
-
-
-@dataclass
-class VllmTestParameters:
-    """
-    Parameters defining the vllm external test input
-
-    !!!DO NOT ADD SECRETS IN THIS CLASS!!!
-    you can put environment variable name in VllmTestParameters if it's not the same as the secret one
-    fetch secrests directly from env variables during runtime
-    """
-
-    torch_whls_path: Path = env_path_field("WHEELS_PATH", "./dist")
-
-    vllm_whls_path: Path = env_path_field(
-        "VLLM_WHEELS_PATH", "./dist/external/vllm/wheels"
-    )
-
-    torch_cuda_arch_list: str = env_str_field("TORCH_CUDA_ARCH_LIST", "8.9")
-
-    def __post_init__(self):
-        if not self.torch_whls_path.exists():
-            raise ValueError("missing torch_whls_path")
-        if not self.vllm_whls_path.exists():
-            raise ValueError("missing vllm_whls_path")
-
-
-class TestInpuType(Enum):
-    TEST_PLAN = "test_plan"
-    UNKNOWN = "unknown"
-
-
-class VllmTestRunner(BaseRunner):
-    def __init__(self, args: Any):
-        self.work_directory = "vllm"
-        self.test_plan = ""
-        self.test_type = TestInpuType.UNKNOWN
-
-        self.shard_id = args.shard_id
-        self.num_shards = args.num_shards
-
-        if args.test_plan:
-            self.test_plan = args.test_plan
-            self.test_type = TestInpuType.TEST_PLAN
-
-        # Matches the structeur in the artifacts.zip from torcb build
-        self.TORCH_WHL_PATH_REGEX = "torch*.whl"
-        self.TORCH_WHL_EXTRA = "opt-einsum"
-        self.TORCH_ADDITIONAL_WHLS_REGEX = [
-            "vision/torchvision*.whl",
-            "audio/torchaudio*.whl",
-        ]
-
-        # Match the structure of the artifacts.zip from vllm external build
-        self.VLLM_TEST_WHLS_REGEX = [
-            "xformers/*.whl",
-            "vllm/vllm*.whl",
-            "flashinfer-python/flashinfer*.whl",
-        ]
-
-    def prepare(self):
-        """
-        prepare test environment for vllm. This includes clone vllm repo, install all wheels, test dependencies and set env
-        """
-        params = VllmTestParameters()
-        logger.info("Display VllmTestParameters %s", params)
-        self._set_envs(params)
-
-        clone_vllm(dst=self.work_directory)
-        with working_directory(self.work_directory):
-            remove_dir(Path("vllm"))
-            self._install_wheels(params)
-            self._install_dependencies()
-        # verify the torches are not overridden by test dependencies
-        check_versions()
-
-    def run(self):
-        """
-        main function to run vllm test
-        """
-        self.prepare()
-        try:
-            with working_directory(self.work_directory):
-                if self.test_type == TestInpuType.TEST_PLAN:
-                    if self.num_shards > 1:
-                        run_test_plan(
-                            self.test_plan,
-                            "vllm",
-                            sample_vllm_test_library(),
-                            self.shard_id,
-                            self.num_shards,
-                        )
-                    else:
-                        run_test_plan(
-                            self.test_plan, "vllm", sample_vllm_test_library()
-                        )
-                else:
-                    raise ValueError(f"Unknown test type {self.test_type}")
-        finally:
-            # double check the torches are not overridden by other packages
-            check_versions()
-
-    def _install_wheels(self, params: VllmTestParameters):
-        logger.info("Running vllm test with inputs: %s", params)
-        if not pkg_exists("torch"):
-            # install torch from local whls if it's not installed yet.
-            torch_p = f"{str(params.torch_whls_path)}/{self.TORCH_WHL_PATH_REGEX}"
-            pip_install_first_match(torch_p, self.TORCH_WHL_EXTRA)
-
-        torch_whls_path = [
-            f"{str(params.torch_whls_path)}/{whl_path}"
-            for whl_path in self.TORCH_ADDITIONAL_WHLS_REGEX
-        ]
-        for torch_whl in torch_whls_path:
-            pip_install_first_match(torch_whl)
-        logger.info("Done. Installed torch and other torch-related wheels ")
-
-        logger.info("Installing vllm wheels")
-        vllm_whls_path = [
-            f"{str(params.vllm_whls_path)}/{whl_path}"
-            for whl_path in self.VLLM_TEST_WHLS_REGEX
-        ]
-        for vllm_whl in vllm_whls_path:
-            pip_install_first_match(vllm_whl)
-        logger.info("Done. Installed vllm wheels")
-
-    def _install_test_dependencies(self):
-        """
-        This method replaces torch dependencies with local torch wheel info in
-        requirements/test.in file from vllm repo. then generates the test.txt
-        in runtime
-        """
-        logger.info("generate test.txt from requirements/test.in with local torch whls")
-        preprocess_test_in()
-        copy("requirements/test.txt", "snapshot_constraint.txt")
-
-        run_command(
-            f"{sys.executable} -m uv pip compile requirements/test.in "
-            "-o test.txt "
-            "--index-strategy unsafe-best-match "
-            "--constraint snapshot_constraint.txt "
-            "--torch-backend cu128"
-        )
-        pip_install_packages(requirements="test.txt", prefer_uv=True)
-        logger.info("Done. installed requirements for test dependencies")
-
-    def _install_dependencies(self):
-        pip_install_packages(packages=["-e", "tests/vllm_test_utils"], prefer_uv=True)
-        pip_install_packages(packages=["hf_transfer"], prefer_uv=True)
-        os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
-
-        # using script from vllm repo to remove all torch packages from requirements txt
-        run_python("use_existing_torch.py")
-
-        # install common packages
-        for requirements in ["requirements/common.txt", "requirements/build.txt"]:
-            pip_install_packages(
-                requirements=requirements,
-                prefer_uv=True,
-            )
-        # install test packages
-        self._install_test_dependencies()
-
-    def _set_envs(self, inputs: VllmTestParameters):
-        os.environ["TORCH_CUDA_ARCH_LIST"] = inputs.torch_cuda_arch_list
-        if not validate_cuda(get_env("TORCH_CUDA_ARCH_LIST")):
-            logger.warning(
-                "Missing supported TORCH_CUDA_ARCH_LIST. "
-                "Currently support TORCH_CUDA_ARCH_LIST env var "
-                "with supported arch [8.0, 8.9, 9.0]"
-            )
-
-        os.environ["HF_TOKEN"] = os.getenv("VLLM_TEST_HUGGING_FACE_TOKEN", "")
-        if not get_env("HF_TOKEN"):
-            raise ValueError(
-                "missing required HF_TOKEN, please set VLLM_TEST_HUGGING_FACE_TOKEN env var"
-            )
-        if not get_env("TORCH_CUDA_ARCH_LIST"):
-            raise ValueError(
-                "missing required TORCH_CUDA_ARCH_LIST, please set TORCH_CUDA_ARCH_LIST env var"
-            )
-
-
-def preprocess_test_in(
-    target_file: str = "requirements/test.in", additional_packages: Iterable[str] = ()
-):
-    """
-    This modifies the target_file file in place in vllm work directory.
-    It removes torch and unwanted packages in target_file and replace with local torch whls
-    package  with format "$WHEEL_PACKAGE_NAME @ file://<LOCAL_PATH>"
-    """
-    additional_package_to_move = list(additional_packages or ())
-    pkgs_to_remove = [
-        "torch",
-        "torchvision",
-        "torchaudio",
-        "xformers",
-        "mamba_ssm",
-    ] + additional_package_to_move
-    # Read current requirements
-    target_path = Path(target_file)
-    lines = target_path.read_text().splitlines()
-
-    pkgs_to_add = []
-
-    # Remove lines starting with the package names (==, @, >=) — case-insensitive
-    pattern = re.compile(rf"^({'|'.join(pkgs_to_remove)})\s*(==|@|>=)", re.IGNORECASE)
-    kept_lines = [line for line in lines if not pattern.match(line)]
-
-    # Get local installed torch/vision/audio from pip freeze
-    # This is hacky, but it works
-    pip_freeze = subprocess.check_output(["pip", "freeze"], text=True)
-    header_lines = [
-        line
-        for line in pip_freeze.splitlines()
-        if re.match(
-            r"^(torch|torchvision|torchaudio)\s*@\s*file://", line, re.IGNORECASE
-        )
-    ]
-
-    # Write back: header_lines + blank + kept_lines
-    out_lines = header_lines + [""] + kept_lines
-    if pkgs_to_add:
-        out_lines += [""] + pkgs_to_add
-
-    out = "\n".join(out_lines) + "\n"
-    target_path.write_text(out)
-    logger.info("[INFO] Updated %s", target_file)
-
-
-def validate_cuda(value: str) -> bool:
-    VALID_VALUES = {"8.0", "8.9", "9.0"}
-    return all(v in VALID_VALUES for v in value.split())
-
-
-def check_versions():
-    """
-    check installed packages version
-    """
-    logger.info("Double check installed packages")
-    patterns = ["torch", "xformers", "torchvision", "torchaudio", "vllm"]
-    for pkg in patterns:
-        pkg_exists(pkg)
-    logger.info("Done. checked installed packages")

.ci/lumen_cli/cli/run.py

@@ -5,7 +5,6 @@ import logging
 
 from cli.build_cli.register_build import register_build_commands
 from cli.lib.common.logger import setup_logging
-from cli.test_cli.register_test import register_test_commands
 
 
 logger = logging.getLogger(__name__)
@@ -21,7 +20,6 @@ def main():
 
     # registers second-level subcommands
     register_build_commands(subparsers)
-    register_test_commands(subparsers)
 
     # parse args after all options are registered
     args = parser.parse_args()

.ci/lumen_cli/cli/test_cli/register_test.py

@@ -1,62 +0,0 @@
-import argparse
-import logging
-
-from cli.lib.common.cli_helper import register_targets, RichHelp, TargetSpec
-from cli.lib.core.vllm.vllm_test import VllmTestRunner
-
-
-logger = logging.getLogger(__name__)
-
-# Maps targets to their argparse configuration and runner
-# it adds new target to path python -m cli.run build external {target} with buildrunner
-_TARGETS: dict[str, TargetSpec] = {
-    "vllm": {
-        "runner": VllmTestRunner,
-        "help": "test vLLM with pytorch main",
-    }
-    # add yours ...
-}
-
-
-def common_args(parser: argparse.ArgumentParser) -> None:
-    """
-    Add common CLI arguments to the given parser.
-    """
-    parser.add_argument(
-        "--shard-id",
-        type=int,
-        default=1,
-        help="a shard id to run, e.g. '0,1,2,3'",
-    )
-    parser.add_argument(
-        "--num-shards",
-        type=int,
-        default=1,
-        help="a number of shards to run, e.g. '4'",
-    )
-    group = parser.add_mutually_exclusive_group(required=True)
-    group.add_argument(
-        "-tp",
-        "--test-plan",
-        type=str,
-        help="a pre-defined test plan to run, e.g. 'basic_correctness_test'",
-    )
-
-
-def register_test_commands(subparsers: argparse._SubParsersAction) -> None:
-    build_parser = subparsers.add_parser(
-        "test",
-        help="test related commands",
-        formatter_class=RichHelp,
-    )
-    build_subparsers = build_parser.add_subparsers(dest="test_command", required=True)
-    overview = "\n".join(
-        f"  {name:12} {spec.get('help', '')}" for name, spec in _TARGETS.items()
-    )
-    external_parser = build_subparsers.add_parser(
-        "external",
-        help="Test external targets",
-        description="Test third-party targets.\n\nAvailable targets:\n" + overview,
-        formatter_class=RichHelp,
-    )
-    register_targets(external_parser, _TARGETS, common_args=common_args)

.ci/lumen_cli/pyproject.toml

@@ -6,7 +6,6 @@ dependencies = [
     "GitPython==3.1.45",
     "docker==7.1.0",
     "pytest==7.3.2",
-    "uv==0.8.6"
 ]
 
 [tool.setuptools]

.ci/lumen_cli/tests/test_run_plan.py

@@ -1,185 +0,0 @@
-# tests/test_run_test_plan.py
-import importlib
-from contextlib import nullcontext
-from types import SimpleNamespace
-from unittest.mock import MagicMock
-
-import pytest
-
-
-MOD = "cli.lib.core.vllm.lib"
-
-# We import inside tests so the MOD override above applies everywhere
-run_test_plan_import_path = f"{MOD}.run_test_plan"
-
-
-def _get_cmd(c):
-    # Support both kwargs and positional args
-    return c.kwargs.get("cmd", c.args[0] if c.args else None)
-
-
-def _get_check(c):
-    if "check" in c.kwargs:
-        return c.kwargs["check"]
-    # If positional, assume second arg is 'check' when present; default False
-    return c.args[1] if len(c.args) > 1 else False
-
-
-@pytest.fixture
-def patch_module(monkeypatch):
-    """
-    Patch helpers ('pip_install_packages', 'temp_environ', 'working_directory',
-    'run_command', 'logger') inside the target module and expose them.
-    """
-    module = importlib.import_module(MOD)
-
-    # Create fakes/mocks
-    pip_install_packages = MagicMock(name="pip_install_packages")
-    run_command = MagicMock(name="run_command", return_value=0)
-
-    # temp_environ / working_directory: record calls but act as context managers
-    temp_calls: list[dict] = []
-    workdir_calls: list[str] = []
-
-    def fake_working_directory(path: str):
-        workdir_calls.append(path)
-        return nullcontext()
-
-    def fake_temp_env(map: dict[str, str]):
-        temp_calls.append(map)
-        return nullcontext()
-
-    logger = SimpleNamespace(
-        info=MagicMock(name="logger.info"),
-        error=MagicMock(name="logger.error"),
-    )
-
-    # Apply patches (raise if attribute doesn't exist)
-    monkeypatch.setattr(
-        module, "pip_install_packages", pip_install_packages, raising=True
-    )
-    monkeypatch.setattr(module, "run_command", run_command, raising=True)
-    monkeypatch.setattr(
-        module, "working_directory", fake_working_directory, raising=True
-    )
-    monkeypatch.setattr(module, "temp_environ", fake_temp_env, raising=True)
-    monkeypatch.setattr(module, "logger", logger, raising=True)
-
-    return SimpleNamespace(
-        module=module,
-        run_test_plan=module.run_test_plan,  # expose to avoid getattr("constant") (Ruff B009)
-        pip_install_packages=pip_install_packages,
-        run_command=run_command,
-        temp_calls=temp_calls,
-        workdir_calls=workdir_calls,
-        logger=logger,
-    )
-
-
-def test_success_runs_all_steps_and_uses_env_and_workdir(monkeypatch, patch_module):
-    run_test_plan = patch_module.run_test_plan
-
-    tests_map = {
-        "basic": {
-            "title": "Basic suite",
-            "package_install": [],
-            "working_directory": "tests",
-            "env_vars": {"GLOBAL_FLAG": "1"},
-            "steps": [
-                "export A=x && pytest -q",
-                "export B=y && pytest -q tests/unit",
-            ],
-        }
-    }
-
-    # One exit code per step (export + two pytest)
-    patch_module.run_command.side_effect = [0, 0, 0]
-
-    run_test_plan("basic", "cpu", tests_map)
-
-    calls = patch_module.run_command.call_args_list
-    cmds = [_get_cmd(c) for c in calls]
-    checks = [_get_check(c) for c in calls]
-
-    assert cmds == [
-        "export A=x && pytest -q",
-        "export B=y && pytest -q tests/unit",
-    ]
-    assert all(chk is False for chk in checks)
-
-    assert patch_module.workdir_calls == ["tests"]
-    assert patch_module.temp_calls == [{"GLOBAL_FLAG": "1"}]
-
-
-def test_installs_packages_when_present(monkeypatch, patch_module):
-    run_test_plan = patch_module.module.run_test_plan
-
-    tests_map = {
-        "with_pkgs": {
-            "title": "Needs deps",
-            "package_install": ["timm==1.0.0", "flash-attn"],
-            "steps": ["pytest -q"],
-        }
-    }
-
-    patch_module.run_command.return_value = 0
-
-    run_test_plan("with_pkgs", "gpu", tests_map)
-
-    patch_module.pip_install_packages.assert_called_once_with(
-        packages=["timm==1.0.0", "flash-attn"],
-        prefer_uv=True,
-    )
-
-
-def test_raises_on_missing_plan(patch_module):
-    run_test_plan = patch_module.module.run_test_plan
-    with pytest.raises(RuntimeError) as ei:
-        run_test_plan("nope", "cpu", tests_map={})
-
-    assert "test nope not found" in str(ei.value)
-
-
-def test_aggregates_failures_and_raises(monkeypatch, patch_module):
-    run_test_plan = patch_module.module.run_test_plan
-
-    tests_map = {
-        "mix": {
-            "title": "Some pass some fail",
-            "steps": [
-                "pytest test_a.py",  # 0 → pass
-                "pytest test_b.py",  # 1 → fail
-                "pytest test_c.py",  # 2 → fail
-            ],
-        }
-    }
-
-    # Simulate pass, fail, fail
-    patch_module.run_command.side_effect = [0, 1, 2]
-
-    with pytest.raises(RuntimeError) as ei:
-        run_test_plan("mix", "cpu", tests_map)
-
-    msg = str(ei.value)
-    assert "2 pytest runs failed" in msg
-    # Ensure logger captured failed tests list
-    patch_module.logger.error.assert_called_once()
-    # And we attempted all three commands
-    assert patch_module.run_command.call_count == 3
-
-
-def test_custom_working_directory_used(patch_module):
-    run_test_plan = patch_module.module.run_test_plan
-
-    tests_map = {
-        "customwd": {
-            "title": "Custom wd",
-            "working_directory": "examples/ci",
-            "steps": ["pytest -q"],
-        }
-    }
-
-    patch_module.run_command.return_value = 0
-    run_test_plan("customwd", "cpu", tests_map)
-
-    assert patch_module.workdir_calls == ["examples/ci"]

.ci/lumen_cli/tests/test_utils.py

@@ -1,143 +0,0 @@
-import os
-import tempfile
-import unittest
-from pathlib import Path
-
-from cli.lib.common.utils import temp_environ, working_directory  # <-- replace import
-
-
-class EnvIsolatedTestCase(unittest.TestCase):
-    """Base class that snapshots os.environ and CWD for isolation."""
-
-    def setUp(self):
-        import os
-        import tempfile
-
-        self._env_backup = dict(os.environ)
-
-        # Snapshot/repair CWD if it's gone
-        try:
-            self._cwd_backup = os.getcwd()
-        except FileNotFoundError:
-            # If CWD no longer exists, switch to a safe place and record that
-            self._cwd_backup = tempfile.gettempdir()
-            os.chdir(self._cwd_backup)
-
-        # Create a temporary directory for the test to run in
-        self._temp_dir = tempfile.mkdtemp()
-        os.chdir(self._temp_dir)
-
-    def tearDown(self):
-        import os
-        import shutil
-        import tempfile
-
-        # Restore cwd first (before cleaning up temp dir)
-        try:
-            os.chdir(self._cwd_backup)
-        except OSError:
-            os.chdir(tempfile.gettempdir())
-
-        # Clean up temporary directory
-        try:
-            shutil.rmtree(self._temp_dir, ignore_errors=True)
-        except Exception:
-            pass  # Ignore cleanup errors
-
-        # Restore env
-        to_del = set(os.environ.keys()) - set(self._env_backup.keys())
-        for k in to_del:
-            os.environ.pop(k, None)
-        for k, v in self._env_backup.items():
-            os.environ[k] = v
-
-
-class TestTempEnviron(EnvIsolatedTestCase):
-    def test_sets_and_restores_new_var(self):
-        var = "TEST_TMP_ENV_NEW"
-        self.assertNotIn(var, os.environ)
-
-        with temp_environ({var: "123"}):
-            self.assertEqual(os.environ[var], "123")
-
-        self.assertNotIn(var, os.environ)  # removed after exit
-
-    def test_overwrites_and_restores_existing_var(self):
-        var = "TEST_TMP_ENV_OVERWRITE"
-        os.environ[var] = "orig"
-
-        with temp_environ({var: "override"}):
-            self.assertEqual(os.environ[var], "override")
-
-        self.assertEqual(os.environ[var], "orig")  # restored
-
-    def test_multiple_vars_and_missing_cleanup(self):
-        v1, v2 = "TEST_ENV_V1", "TEST_ENV_V2"
-        os.environ.pop(v1, None)
-        os.environ[v2] = "keep"
-
-        with temp_environ({v1: "a", v2: "b"}):
-            self.assertEqual(os.environ[v1], "a")
-            self.assertEqual(os.environ[v2], "b")
-
-        self.assertNotIn(v1, os.environ)  # newly-added -> removed
-        self.assertEqual(os.environ[v2], "keep")  # pre-existing -> restored
-
-    def test_restores_even_on_exception(self):
-        var = "TEST_TMP_ENV_EXCEPTION"
-        self.assertNotIn(var, os.environ)
-
-        with self.assertRaises(RuntimeError):
-            with temp_environ({var: "x"}):
-                self.assertEqual(os.environ[var], "x")
-                raise RuntimeError("boom")
-
-        self.assertNotIn(var, os.environ)  # removed after exception
-
-
-class TestWorkingDirectory(EnvIsolatedTestCase):
-    def test_changes_and_restores(self):
-        start = Path.cwd()
-        with tempfile.TemporaryDirectory() as td:
-            target = Path(td) / "wd"
-            target.mkdir()
-
-            with working_directory(str(target)):
-                self.assertEqual(Path.cwd().resolve(), target.resolve())
-
-        self.assertEqual(Path.cwd(), start)
-
-    def test_noop_when_empty_path(self):
-        start = Path.cwd()
-        with working_directory(""):
-            self.assertEqual(Path.cwd(), start)
-        self.assertEqual(Path.cwd(), start)
-
-    def test_restores_on_exception(self):
-        start = Path.cwd()
-
-        with tempfile.TemporaryDirectory() as td:
-            target = Path(td) / "wd_exc"
-            target.mkdir()
-
-            with self.assertRaises(ValueError):
-                with working_directory(str(target)):
-                    # Normalize both sides to handle /var -> /private/var
-                    self.assertEqual(Path.cwd().resolve(), target.resolve())
-                    raise ValueError("boom")
-
-        self.assertEqual(Path.cwd().resolve(), start.resolve())
-
-    def test_raises_for_missing_dir(self):
-        start = Path.cwd()
-        with tempfile.TemporaryDirectory() as td:
-            missing = Path(td) / "does_not_exist"
-            with self.assertRaises(FileNotFoundError):
-                # os.chdir should raise before yielding
-                with working_directory(str(missing)):
-                    pass
-        self.assertEqual(Path.cwd(), start)
-
-
-if __name__ == "__main__":
-    unittest.main(verbosity=2)

.ci/lumen_cli/tests/test_vllm.py

@@ -4,15 +4,12 @@ import unittest
 from pathlib import Path
 from unittest.mock import MagicMock, patch
 
-import cli.lib.core.vllm.vllm_build as vllm_build
-
-
-_VLLM_BUILD_MODULE = "cli.lib.core.vllm.vllm_build"
+import cli.lib.core.vllm as vllm
 
 
 class TestVllmBuildParameters(unittest.TestCase):
-    @patch(f"{_VLLM_BUILD_MODULE}.local_image_exists", return_value=True)
-    @patch(f"{_VLLM_BUILD_MODULE}.is_path_exist", return_value=True)
+    @patch("cli.lib.core.vllm.local_image_exists", return_value=True)
+    @patch("cli.lib.core.vllm.is_path_exist", return_value=True)
     @patch(
         "cli.lib.common.envs_helper.env_path_optional",
         side_effect=lambda name, default=None, resolve=True: {
@@ -37,13 +34,13 @@ class TestVllmBuildParameters(unittest.TestCase):
     def test_params_success_normalizes_and_validates(
         self, mock_env_path, mock_is_path, mock_local_img
     ):
-        params = vllm_build.VllmBuildParameters()
+        params = vllm.VllmBuildParameters()
         self.assertEqual(params.torch_whls_path, Path("/abs/dist"))
         self.assertEqual(params.dockerfile_path, Path("/abs/vllm/Dockerfile"))
         self.assertEqual(params.output_dir, Path("/abs/shared"))
         self.assertEqual(params.base_image, "my/image:tag")
 
-    @patch(f"{_VLLM_BUILD_MODULE}.is_path_exist", return_value=False)
+    @patch("cli.lib.core.vllm.is_path_exist", return_value=False)
     @patch.dict(
         os.environ, {"USE_TORCH_WHEEL": "1", "TORCH_WHEELS_PATH": "dist"}, clear=True
     )
@@ -51,14 +48,14 @@ class TestVllmBuildParameters(unittest.TestCase):
         with tempfile.TemporaryDirectory() as td:
             os.chdir(td)
             with self.assertRaises(ValueError) as cm:
-                vllm_build.VllmBuildParameters(
+                vllm.VllmBuildParameters(
                     use_local_base_image=False,
                     use_local_dockerfile=False,
                 )
         err = cm.exception
         self.assertIn("TORCH_WHEELS_PATH", str(err))
 
-    @patch(f"{_VLLM_BUILD_MODULE}.local_image_exists", return_value=False)
+    @patch("cli.lib.core.vllm.local_image_exists", return_value=False)
     @patch.dict(
         os.environ, {"USE_LOCAL_BASE_IMAGE": "1", "BASE_IMAGE": "img:tag"}, clear=True
     )
@@ -66,14 +63,14 @@ class TestVllmBuildParameters(unittest.TestCase):
         with tempfile.TemporaryDirectory() as td:
             os.chdir(td)
             with self.assertRaises(ValueError) as cm:
-                vllm_build.VllmBuildParameters(
+                vllm.VllmBuildParameters(
                     use_torch_whl=False,
                     use_local_dockerfile=False,
                 )
         err = cm.exception
         self.assertIn("BASE_IMAGE", str(err))
 
-    @patch(f"{_VLLM_BUILD_MODULE}.is_path_exist", return_value=False)
+    @patch("cli.lib.core.vllm.is_path_exist", return_value=False)
     @patch.dict(
         os.environ,
         {"USE_LOCAL_DOCKERFILE": "1", "DOCKERFILE_PATH": "Dockerfile"},
@@ -83,14 +80,14 @@ class TestVllmBuildParameters(unittest.TestCase):
         with tempfile.TemporaryDirectory() as td:
             os.chdir(td)
             with self.assertRaises(ValueError) as cm:
-                vllm_build.VllmBuildParameters(
+                vllm.VllmBuildParameters(
                     use_torch_whl=False,
                     use_local_base_image=False,
                 )
         err = cm.exception
         self.assertIn("DOCKERFILE_PATH", str(err))
 
-    @patch(f"{_VLLM_BUILD_MODULE}.is_path_exist", return_value=False)
+    @patch("cli.lib.core.vllm.is_path_exist", return_value=False)
     @patch.dict(
         os.environ,
         {"OUTPUT_DIR": ""},
@@ -98,13 +95,14 @@ class TestVllmBuildParameters(unittest.TestCase):
     )
     def test_params_missing_output_dir(self, _is_path):
         with self.assertRaises(FileNotFoundError):
-            vllm_build.VllmBuildParameters()
+            vllm.VllmBuildParameters()
 
 
 class TestBuildCmdAndRun(unittest.TestCase):
-    @patch(f"{_VLLM_BUILD_MODULE}.local_image_exists", return_value=True)
+    @patch("cli.lib.core.vllm.local_image_exists", return_value=True)
     def test_generate_docker_build_cmd_includes_bits(self, _exists):
-        runner = vllm_build.VllmBuildRunner()
+        runner = vllm.VllmBuildRunner()
+        # Craft inputs that simulate a prepared build
         inputs = MagicMock()
         inputs.output_dir = Path("/abs/out")
         inputs.use_local_base_image = True
@@ -120,7 +118,7 @@ class TestBuildCmdAndRun(unittest.TestCase):
         inputs.tag_name = "vllm-wheels"
 
         cmd = runner._generate_docker_build_cmd(inputs)
-        squashed = " ".join(cmd.split())
+        squashed = " ".join(cmd.split())  # normalize whitespace for matching
 
         self.assertIn("--output type=local,dest=/abs/out", squashed)
         self.assertIn("-f docker/Dockerfile.nightly_torch", squashed)
@@ -138,17 +136,18 @@ class TestBuildCmdAndRun(unittest.TestCase):
         self.assertIn("--target export-wheels", squashed)
         self.assertIn("-t vllm-wheels", squashed)
 
-    @patch(f"{_VLLM_BUILD_MODULE}.run_command")
-    @patch(f"{_VLLM_BUILD_MODULE}.ensure_dir_exists")
-    @patch(f"{_VLLM_BUILD_MODULE}.clone_vllm")
+    @patch("cli.lib.core.vllm.run_command")
+    @patch("cli.lib.core.vllm.ensure_dir_exists")
+    @patch("cli.lib.core.vllm.clone_vllm")
     @patch.object(
-        vllm_build.VllmBuildRunner,
+        vllm.VllmBuildRunner,
         "_generate_docker_build_cmd",
         return_value="docker buildx ...",
     )
     @patch.dict(
         os.environ,
         {
+            # Make __post_init__ validations pass cheaply
             "USE_TORCH_WHEEL": "0",
             "USE_LOCAL_BASE_IMAGE": "0",
             "USE_LOCAL_DOCKERFILE": "0",
@@ -159,18 +158,24 @@ class TestBuildCmdAndRun(unittest.TestCase):
     def test_run_calls_clone_prepare_and_build(
         self, mock_gen, mock_clone, mock_ensure, mock_run
     ):
+        # Stub parameters instance so we avoid FS/Docker accesses in run()
         params = MagicMock()
         params.output_dir = Path("shared")
         params.use_local_dockerfile = False
         params.use_torch_whl = False
 
-        with patch(f"{_VLLM_BUILD_MODULE}.VllmBuildParameters", return_value=params):
-            runner = vllm_build.VllmBuildRunner()
+        with patch("cli.lib.core.vllm.VllmBuildParameters", return_value=params):
+            runner = vllm.VllmBuildRunner()
             runner.run()
 
         mock_clone.assert_called_once()
         mock_ensure.assert_called_once_with(Path("shared"))
         mock_gen.assert_called_once_with(params)
         mock_run.assert_called_once()
+        # ensure we run in vllm workdir
         _, kwargs = mock_run.call_args
         assert kwargs.get("cwd") == "vllm"
+
+
+if __name__ == "__main__":
+    unittest.main()

.ci/magma/Makefile

@@ -16,7 +16,6 @@ DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
 	magma/build_magma.sh
 
 .PHONY: all
-all: magma-cuda130
 all: magma-cuda129
 all: magma-cuda128
 all: magma-cuda126
@@ -26,12 +25,6 @@ clean:
 	$(RM) -r magma-*
 	$(RM) -r output
 
-.PHONY: magma-cuda130
-magma-cuda130: DESIRED_CUDA := 13.0
-magma-cuda130: CUDA_ARCH_LIST := -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_90,code=sm_90 -gencode arch=compute_100,code=sm_100 -gencode arch=compute_120,code=sm_120
-magma-cuda130:
-	$(DOCKER_RUN)
-
 .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

.ci/magma/build_magma.sh

@@ -28,7 +28,6 @@ pushd ${PACKAGE_DIR}/magma-${MAGMA_VERSION}
 patch < ${PACKAGE_FILES}/CMake.patch
 patch < ${PACKAGE_FILES}/cmakelists.patch
 patch -p0 < ${PACKAGE_FILES}/thread_queue.patch
-patch -p1 < ${PACKAGE_FILES}/cuda13.patch
 patch -p1 < ${PACKAGE_FILES}/getrf_shfl.patch
 patch -p1 < ${PACKAGE_FILES}/getrf_nbparam.patch
 # The build.sh script expects to be executed from the sources root folder
@@ -38,7 +37,6 @@ popd
 # Package recipe, license and tarball
 # Folder and package name are backward compatible for the build workflow
 cp ${PACKAGE_FILES}/build.sh ${PACKAGE_RECIPE}/build.sh
-cp ${PACKAGE_FILES}/cuda13.patch ${PACKAGE_RECIPE}/cuda13.patch
 cp ${PACKAGE_FILES}/thread_queue.patch ${PACKAGE_RECIPE}/thread_queue.patch
 cp ${PACKAGE_FILES}/cmakelists.patch ${PACKAGE_RECIPE}/cmakelists.patch
 cp ${PACKAGE_FILES}/getrf_shfl.patch ${PACKAGE_RECIPE}/getrf_shfl.patch

.ci/magma/package_files/cuda13.patch

@@ -1,26 +0,0 @@
-diff --git a/interface_cuda/interface.cpp b/interface_cuda/interface.cpp
-index 73fed1b20..e77519bfe 100644
---- a/interface_cuda/interface.cpp
-+++ b/interface_cuda/interface.cpp
-@@ -438,14 +438,20 @@ magma_print_environment()
-         cudaDeviceProp prop;
-         err = cudaGetDeviceProperties( &prop, dev );
-         check_error( err );
-+        #ifdef MAGMA_HAVE_CUDA
-+#if CUDA_VERSION < 13000
-         printf( "%% device %d: %s, %.1f MHz clock, %.1f MiB memory, capability %d.%d\n",
-                 dev,
-                 prop.name,
-                 prop.clockRate / 1000.,
-+#else
-+        printf( "%% device %d: %s, ??? MHz clock, %.1f MiB memory, capability %d.%d\n",
-+                dev,
-+                prop.name,
-+#endif
-                 prop.totalGlobalMem / (1024.*1024.),
-                 prop.major,
-                 prop.minor );
--        #ifdef MAGMA_HAVE_CUDA
-         int arch = prop.major*100 + prop.minor*10;
-         if ( arch < MAGMA_CUDA_ARCH_MIN ) {
-             printf("\n"

.ci/manywheel/build_cuda.sh

@@ -66,9 +66,6 @@ case ${CUDA_VERSION} in
             TORCH_CUDA_ARCH_LIST="7.5;8.0;9.0;10.0;12.0+PTX"
         fi
         ;;
-    13.0)
-        TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6;9.0;10.0;12.0+PTX"
-        ;;
     12.6)
         TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;7.5;8.0;8.6;9.0"
         ;;
@@ -113,18 +110,13 @@ DEPS_SONAME=(
 )
 
 
-# CUDA_VERSION 12.*, 13.*
-if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
+# CUDA_VERSION 12.6, 12.8, 12.9
+if [[ $CUDA_VERSION == 12* ]]; then
     export USE_STATIC_CUDNN=0
     # Try parallelizing nvcc as well
-    TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
-    # Compress the fatbin with -compress-mode=size for CUDA 13
-    if [[ $CUDA_VERSION == 13* ]]; then
-        export TORCH_NVCC_FLAGS="$TORCH_NVCC_FLAGS -compress-mode=size"
-    fi
+    export TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
     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"
@@ -134,11 +126,16 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
             "/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.12"
+            "/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/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/lib64/libnvshmem_host.so.3"
+            "/usr/local/cuda/lib64/libnvshem_host.so.3"
+            "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12"
             "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so"
         )
         DEPS_SONAME+=(
@@ -150,83 +147,42 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
             "libcudnn_engines_precompiled.so.9"
             "libcudnn_heuristic.so.9"
             "libcudnn.so.9"
+            "libcublas.so.12"
+            "libcublasLt.so.12"
             "libcusparseLt.so.0"
+            "libcudart.so.12"
+            "libnvrtc.so.12"
             "libnvrtc-builtins.so"
             "libnvshmem_host.so.3"
             "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 == 13* ]]; then
-            DEPS_LIST+=(
-                "/usr/local/cuda/lib64/libcublas.so.13"
-                "/usr/local/cuda/lib64/libcublasLt.so.13"
-                "/usr/local/cuda/lib64/libcudart.so.13"
-                "/usr/local/cuda/lib64/libnvrtc.so.13"
-                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13"
-                "/usr/local/cuda/lib64/libibverbs.so.1"
-                "/usr/local/cuda/lib64/librdmacm.so.1"
-                "/usr/local/cuda/lib64/libmlx5.so.1"
-                "/usr/local/cuda/lib64/libnl-3.so.200"
-                "/usr/local/cuda/lib64/libnl-route-3.so.200")
-            DEPS_SONAME+=(
-                "libcublas.so.13"
-                "libcublasLt.so.13"
-                "libcudart.so.13"
-                "libnvrtc.so.13"
-                "libcupti.so.13"
-                "libibverbs.so.1"
-                "librdmacm.so.1"
-                "libmlx5.so.1"
-                "libnl-3.so.200"
-                "libnl-route-3.so.200")
-            export USE_CUPTI_SO=1
-            export ATEN_STATIC_CUDA=0
-            export USE_CUDA_STATIC_LINK=0
-            export USE_CUFILE=0
-        else
-            DEPS_LIST+=(
-                "/usr/local/cuda/lib64/libnvToolsExt.so.1"
-                "/usr/local/cuda/lib64/libcublas.so.12"
-                "/usr/local/cuda/lib64/libcublasLt.so.12"
-                "/usr/local/cuda/lib64/libcudart.so.12"
-                "/usr/local/cuda/lib64/libnvrtc.so.12"
-                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12")
-            DEPS_SONAME+=(
-                "libnvToolsExt.so.1"
-                "libcublas.so.12"
-                "libcublasLt.so.12"
-                "libcudart.so.12"
-                "libnvrtc.so.12"
-                "libcupti.so.12")
+        if [[ $CUDA_VERSION != 12.9* ]]; then
+            DEPS_LIST+=("/usr/local/cuda/lib64/libnvToolsExt.so.1")
+            DEPS_SONAME+=("libnvToolsExt.so.1")
         fi
     else
         echo "Using nvidia libs from pypi."
         CUDA_RPATHS=(
+            '$ORIGIN/../../nvidia/cublas/lib'
+            '$ORIGIN/../../nvidia/cuda_cupti/lib'
+            '$ORIGIN/../../nvidia/cuda_nvrtc/lib'
+            '$ORIGIN/../../nvidia/cuda_runtime/lib'
             '$ORIGIN/../../nvidia/cudnn/lib'
-            '$ORIGIN/../../nvidia/nvshmem/lib'
-            '$ORIGIN/../../nvidia/nccl/lib'
+            '$ORIGIN/../../nvidia/cufft/lib'
+            '$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 [[ $CUDA_VERSION == 13* ]]; then
-            CUDA_RPATHS+=('$ORIGIN/../../nvidia/cu13/lib')
-        else
-            CUDA_RPATHS+=(
-                '$ORIGIN/../../nvidia/cublas/lib'
-                '$ORIGIN/../../nvidia/cuda_cupti/lib'
-                '$ORIGIN/../../nvidia/cuda_nvrtc/lib'
-                '$ORIGIN/../../nvidia/cuda_runtime/lib'
-                '$ORIGIN/../../nvidia/cufft/lib'
-                '$ORIGIN/../../nvidia/curand/lib'
-                '$ORIGIN/../../nvidia/cusolver/lib'
-                '$ORIGIN/../../nvidia/cusparse/lib'
-                '$ORIGIN/../../cusparselt/lib'
-                '$ORIGIN/../../nvidia/nvtx/lib'
-                '$ORIGIN/../../nvidia/cufile/lib'
-            )
-        fi
-
         CUDA_RPATHS=$(IFS=: ; echo "${CUDA_RPATHS[*]}")
         export C_SO_RPATH=$CUDA_RPATHS':$ORIGIN:$ORIGIN/lib'
         export LIB_SO_RPATH=$CUDA_RPATHS':$ORIGIN'

.ci/manywheel/build_xpu.sh

@@ -25,7 +25,6 @@ source /opt/intel/oneapi/mpi/latest/env/vars.sh
 export USE_STATIC_MKL=1
 export USE_ONEMKL=1
 export USE_XCCL=1
-export USE_MPI=0
 
 WHEELHOUSE_DIR="wheelhousexpu"
 LIBTORCH_HOUSE_DIR="libtorch_housexpu"

.ci/pytorch/build.sh

@@ -89,7 +89,7 @@ fi
 if [[ "$BUILD_ENVIRONMENT" == *aarch64* ]]; then
   export USE_MKLDNN=1
   export USE_MKLDNN_ACL=1
-  export ACL_ROOT_DIR=/acl
+  export ACL_ROOT_DIR=/ComputeLibrary
 fi
 
 if [[ "$BUILD_ENVIRONMENT" == *riscv64* ]]; then
@@ -173,7 +173,6 @@ if [[ "$BUILD_ENVIRONMENT" == *xpu* ]]; then
   source /opt/intel/oneapi/mpi/latest/env/vars.sh
   # Enable XCCL build
   export USE_XCCL=1
-  export USE_MPI=0
   # XPU kineto feature dependencies are not fully ready, disable kineto build as temp WA
   export USE_KINETO=0
   export TORCH_XPU_ARCH_LIST=pvc
@@ -195,16 +194,8 @@ fi
 
 # We only build FlashAttention files for CUDA 8.0+, and they require large amounts of
 # memory to build and will OOM
-
 if [[ "$BUILD_ENVIRONMENT" == *cuda* ]] && echo "${TORCH_CUDA_ARCH_LIST}" | tr ' ' '\n' | sed 's/$/>= 8.0/' | bc | grep -q 1; then
-  J=2  # default to 2 jobs
-  case "$RUNNER" in
-    linux.12xlarge.memory|linux.24xlarge.memory)
-      J=24
-      ;;
-  esac
-  echo "Building FlashAttention with job limit $J"
-  export BUILD_CUSTOM_STEP="ninja -C build flash_attention -j ${J}"
+  export BUILD_CUSTOM_STEP="ninja -C build flash_attention -j 2"
 fi
 
 if [[ "${BUILD_ENVIRONMENT}" == *clang* ]]; then

.ci/pytorch/check_binary.sh

@@ -300,3 +300,24 @@ except RuntimeError as e:
     exit 1
   fi
 fi
+
+###############################################################################
+# Check for C++ ABI compatibility to GCC-11 - GCC 13
+###############################################################################
+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' && "$DESIRED_CUDA" != 'rocm'* ]]; then
+    cxx_abi="18"
+  else
+    cxx_abi="16"
+  fi
+  python -c "import torch; exit(0 if torch._C._PYBIND11_BUILD_ABI == '_cxxabi10${cxx_abi}' else 1)"
+  popd
+fi

.ci/pytorch/common_utils.sh

@@ -149,19 +149,6 @@ function get_pinned_commit() {
   cat .github/ci_commit_pins/"${1}".txt
 }
 
-function detect_cuda_arch() {
-  if [[ "${BUILD_ENVIRONMENT}" == *cuda* ]]; then
-    if command -v nvidia-smi; then
-      TORCH_CUDA_ARCH_LIST=$(nvidia-smi --query-gpu=compute_cap --format=csv | tail -n 1)
-    elif [[ "${TEST_CONFIG}" == *nogpu* ]]; then
-      # There won't be nvidia-smi in nogpu tests, so just set TORCH_CUDA_ARCH_LIST to the default
-      # minimum supported value here
-      TORCH_CUDA_ARCH_LIST=8.0
-    fi
-    export TORCH_CUDA_ARCH_LIST
-  fi
-}
-
 function install_torchaudio() {
   local commit
   commit=$(get_pinned_commit audio)

.ci/pytorch/macos-test.sh

@@ -174,12 +174,13 @@ checkout_install_torchbench() {
     # to install and test other models
     python install.py --continue_on_fail
   fi
-  popd
 
-  pip install -r .ci/docker/ci_commit_pins/huggingface-requirements.txt
+  # soxr comes from https://github.com/huggingface/transformers/pull/39429
+  pip install transformers==4.54.0 soxr==0.5.0
 
   echo "Print all dependencies after TorchBench is installed"
   python -mpip freeze
+  popd
 }
 
 torchbench_setup_macos() {
@@ -192,7 +193,7 @@ torchbench_setup_macos() {
   git checkout "$(cat ../.github/ci_commit_pins/vision.txt)"
   git submodule update --init --recursive
   python setup.py clean
-  python -m pip install -e . -v --no-build-isolation
+  python setup.py develop
   popd
 
   pushd torchaudio
@@ -201,7 +202,7 @@ torchbench_setup_macos() {
   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 -m pip install -e . -v --no-build-isolation
+  USE_OPENMP=0 python setup.py develop
   popd
 
   checkout_install_torchbench
@@ -299,47 +300,6 @@ test_torchbench_smoketest() {
     fi
 
   done
-  echo "Pytorch benchmark on mps device completed"
-}
-
-test_aoti_torchbench_smoketest() {
-  print_cmake_info
-
-  echo "Launching AOTInductor torchbench setup"
-  pip_benchmark_deps
-  # shellcheck disable=SC2119,SC2120
-  torchbench_setup_macos
-
-  TEST_REPORTS_DIR=$(pwd)/test/test-reports
-  mkdir -p "$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)
-
-  echo "Launching torchbench inference performance run for AOT Inductor and dtype ${dtype}"
-  local dtype_arg="--${dtype}"
-  if [ "$dtype" == notset ]; then
-      dtype_arg="--float32"
-  fi
-  touch "$TEST_REPORTS_DIR/aot_inductor_torchbench_${dtype}_inference_${device}_performance.csv"
-  for model in "${models[@]}"; do
-    PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
-      --performance --only "$model" --export-aot-inductor --inference --devices "$device" "$dtype_arg" \
-      --output "$TEST_REPORTS_DIR/aot_inductor_torchbench_${dtype}_inference_${device}_performance.csv" || true
-    PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
-      --accuracy --only "$model" --export-aot-inductor --inference --devices "$device" "$dtype_arg" \
-      --output "$TEST_REPORTS_DIR/aot_inductor_torchbench_${dtype}_inference_${device}_accuracy.csv" || true
-  done
-
-  echo "Launching HuggingFace inference performance run for AOT Inductor and dtype ${dtype}"
-  PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/huggingface.py \
-    --performance --export-aot-inductor --inference --devices "$device" "$dtype_arg" \
-    --output "$TEST_REPORTS_DIR/aot_inductor_huggingface_${dtype}_inference_${device}_performance.csv" || true
-  PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/huggingface.py \
-    --accuracy --export-aot-inductor --inference --devices "$device" "$dtype_arg" \
-    --output "$TEST_REPORTS_DIR/aot_inductor_huggingface_${dtype}_inference_${device}_accuracy.csv" || true
 
   echo "Pytorch benchmark on mps device completed"
 }
@@ -388,8 +348,6 @@ elif [[ $TEST_CONFIG == *"perf_timm"* ]]; then
   test_timm_perf
 elif [[ $TEST_CONFIG == *"perf_smoketest"* ]]; then
   test_torchbench_smoketest "${SHARD_NUMBER}"
-elif [[ $TEST_CONFIG == *"aot_inductor_perf_smoketest"* ]]; then
-  test_aoti_torchbench_smoketest "${SHARD_NUMBER}"
 elif [[ $TEST_CONFIG == *"mps"* ]]; then
   test_python_mps
 elif [[ $NUM_TEST_SHARDS -gt 1 ]]; then

.ci/pytorch/multigpu-test.sh

@@ -45,7 +45,6 @@ if [[ "${SHARD_NUMBER:-2}" == "2" ]]; then
     # DTensor tests
     time python test/run_test.py --verbose -i distributed/tensor/test_random_ops
     time python test/run_test.py --verbose -i distributed/tensor/test_dtensor_compile
-    time python test/run_test.py --verbose -i distributed/tensor/test_utils.py
 
     # DeviceMesh test
     time python test/run_test.py --verbose -i distributed/test_device_mesh

.ci/pytorch/test.sh

@@ -91,7 +91,6 @@ if [[ "$BUILD_ENVIRONMENT" == *clang9* || "$BUILD_ENVIRONMENT" == *xpu* ]]; then
   export VALGRIND=OFF
 fi
 
-detect_cuda_arch
 
 if [[ "$BUILD_ENVIRONMENT" == *s390x* ]]; then
   # There are additional warnings on s390x, maybe due to newer gcc.
@@ -496,14 +495,6 @@ test_inductor_cpp_wrapper_shard() {
     -k 'take' \
     --shard "$1" "$NUM_TEST_SHARDS" \
     --verbose
-
-  if [[ "${BUILD_ENVIRONMENT}" == *xpu* ]]; then
-    python test/run_test.py \
-      --include inductor/test_mkldnn_pattern_matcher \
-      -k 'xpu' \
-      --shard "$1" "$NUM_TEST_SHARDS" \
-      --verbose
-  fi
 }
 
 # "Global" flags for inductor benchmarking controlled by TEST_CONFIG
@@ -778,6 +769,11 @@ test_single_dynamo_benchmark() {
 }
 
 test_inductor_micro_benchmark() {
+  # torchao requires cuda 8.0 or above for bfloat16 support
+  if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
+    export TORCH_CUDA_ARCH_LIST="8.0;8.6"
+  fi
+  install_torchao
   TEST_REPORTS_DIR=$(pwd)/test/test-reports
   if [[ "${TEST_CONFIG}" == *cpu* ]]; then
     test_inductor_set_cpu_affinity
@@ -1633,10 +1629,6 @@ elif [[ "${TEST_CONFIG}" == *xla* ]]; then
   install_torchvision
   build_xla
   test_xla
-elif [[ "$TEST_CONFIG" == *vllm* ]]; then
-    echo "vLLM CI uses TORCH_CUDA_ARCH_LIST: $TORCH_CUDA_ARCH_LIST"
-    (cd .ci/lumen_cli && python -m pip install -e .)
-    python -m cli.run test external vllm --test-plan "$TEST_CONFIG" --shard-id "$SHARD_NUMBER" --num-shards "$NUM_TEST_SHARDS"
 elif [[ "${TEST_CONFIG}" == *executorch* ]]; then
   test_executorch
 elif [[ "$TEST_CONFIG" == 'jit_legacy' ]]; then
@@ -1659,50 +1651,37 @@ elif [[ "${TEST_CONFIG}" == *operator_benchmark* ]]; then
     elif [[ "${TEST_CONFIG}" == *all* ]]; then
       TEST_MODE="all"
     fi
+
     test_operator_benchmark cpu ${TEST_MODE}
+
   fi
 elif [[ "${TEST_CONFIG}" == *inductor_distributed* ]]; then
-  if [[ "$BUILD_ENVIRONMENT" != *rocm* ]]; then
-    install_torchao
-  fi
   test_inductor_distributed
 elif [[ "${TEST_CONFIG}" == *inductor-halide* ]]; then
   test_inductor_halide
 elif [[ "${TEST_CONFIG}" == *inductor-triton-cpu* ]]; then
   test_inductor_triton_cpu
 elif [[ "${TEST_CONFIG}" == *inductor-micro-benchmark* ]]; then
-  install_torchao
   test_inductor_micro_benchmark
 elif [[ "${TEST_CONFIG}" == *huggingface* ]]; then
   install_torchvision
-  if [[ "$BUILD_ENVIRONMENT" != *rocm* ]]; then
-    install_torchao
-  fi
   id=$((SHARD_NUMBER-1))
   test_dynamo_benchmark huggingface "$id"
 elif [[ "${TEST_CONFIG}" == *timm* ]]; then
   install_torchvision
-  if [[ "$BUILD_ENVIRONMENT" != *rocm* ]]; then
-    install_torchao
-  fi
   id=$((SHARD_NUMBER-1))
   test_dynamo_benchmark timm_models "$id"
 elif [[ "${TEST_CONFIG}" == cachebench ]]; then
   install_torchaudio
   install_torchvision
-  install_torchao
   PYTHONPATH=/torchbench test_cachebench
 elif [[ "${TEST_CONFIG}" == verify_cachebench ]]; then
   install_torchaudio
   install_torchvision
-  install_torchao
   PYTHONPATH=/torchbench test_verify_cachebench
 elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
   install_torchaudio
   install_torchvision
-  if [[ "$BUILD_ENVIRONMENT" != *rocm* ]]; then
-    install_torchao
-  fi
   id=$((SHARD_NUMBER-1))
   # https://github.com/opencv/opencv-python/issues/885
   pip_install opencv-python==4.8.0.74
@@ -1722,18 +1701,12 @@ elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
   fi
 elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
   install_torchvision
-  if [[ "$BUILD_ENVIRONMENT" != *rocm* ]]; then
-    install_torchao
-  fi
   PYTHONPATH=/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
   if [[ "$SHARD_NUMBER" -eq "1" ]]; then
     test_inductor_aoti
   fi
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision
-  if [[ "$BUILD_ENVIRONMENT" != *rocm* ]]; then
-    install_torchao
-  fi
   test_inductor_shard "${SHARD_NUMBER}"
   if [[ "${SHARD_NUMBER}" == 1 ]]; then
     if [[ "${BUILD_ENVIRONMENT}" != linux-jammy-py3.9-gcc11-build ]]; then

.ci/pytorch/win-test.sh

@@ -44,7 +44,7 @@ python -m pip install pytest-rerunfailures==10.3 pytest-cpp==2.3.0 tensorboard==
 python -m pip install z3-solver==4.15.1.0
 
 # Install tlparse for test\dynamo\test_structured_trace.py UTs.
-python -m pip install tlparse==0.4.0
+python -m pip install tlparse==0.3.30
 
 # Install parameterized
 python -m pip install parameterized==0.8.1

.ci/pytorch/windows/cuda130.bat

@@ -1,59 +0,0 @@
-@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_V130%"=="" (
-    IF EXIST "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.0\bin\nvcc.exe" (
-        set "CUDA_PATH_V130=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.0"
-    ) ELSE (
-        echo CUDA 13.0 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_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 "CUDA_PATH=%CUDA_PATH_V130%"
-set "PATH=%CUDA_PATH_V130%\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

.ci/pytorch/windows/internal/copy.bat

@@ -1,20 +1,12 @@
-
-if %CUDA_VERSION% geq 130 (
-    set "dll_path=bin\x64"
-) else (
-    set "dll_path=bin"
-)
-
-copy "%CUDA_PATH%\%dll_path%\cusparse*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\cublas*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\cudart*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\curand*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\cufft*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\cusolver*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\nvrtc*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\nvJitLink_*.dll*"  pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cusparse*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cublas*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cudart*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\curand*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cufft*64_*.dll*" pytorch\torch\lib
+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
 
@@ -28,3 +20,8 @@ copy "%libuv_ROOT%\bin\uv.dll" pytorch\torch\lib
 if exist "C:\Windows\System32\zlibwapi.dll" (
     copy "C:\Windows\System32\zlibwapi.dll"  pytorch\torch\lib
 )
+
+::copy nvJitLink dll is requires for cuda 12+
+if exist "%CUDA_PATH%\bin\nvJitLink_*.dll*" (
+    copy "%CUDA_PATH%\bin\nvJitLink_*.dll*"  pytorch\torch\lib
+)

.ci/pytorch/windows/internal/cuda_install.bat

@@ -26,7 +26,6 @@ if exist "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v%CUDA_VERSION_STR%
 if %CUDA_VER% EQU 126 goto cuda126
 if %CUDA_VER% EQU 128 goto cuda128
 if %CUDA_VER% EQU 129 goto cuda129
-if %CUDA_VER% EQU 130 goto cuda130
 
 echo CUDA %CUDA_VERSION_STR% is not supported
 exit /b 1
@@ -114,33 +113,6 @@ xcopy /Y "%SRC_DIR%\temp_build\zlib\dll_x64\*.dll" "C:\Windows\System32"
 
 goto cuda_common
 
-:cuda130
-
-set CUDA_INSTALL_EXE=cuda_13.0.0_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="
-)
-
-set CUDNN_FOLDER=cudnn-windows-x86_64-9.12.0.46_cuda13-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
-
 :cuda_common
 :: NOTE: We only install CUDA if we don't have it installed already.
 :: With GHA runners these should be pre-installed as part of our AMI process

.ci/pytorch/windows/internal/driver_update.bat

@@ -1,9 +1,9 @@
-set WIN_DRIVER_VN=580.88
-set "DRIVER_DOWNLOAD_LINK=https://ossci-windows.s3.amazonaws.com/%WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe" & REM @lint-ignore
-curl --retry 3 -kL %DRIVER_DOWNLOAD_LINK% --output %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe
+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
+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
 
-start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe -s -noreboot
+start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe -s -noreboot
 if errorlevel 1 exit /b 1
 
-del %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe || ver > NUL
+del %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe || ver > NUL

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

@@ -1,22 +1,12 @@
 set ADDITIONAL_OPTIONS=""
 set PYTHON_EXEC="python"
-
-
 if "%DESIRED_PYTHON%" == "3.13t" (
     echo Python version is set to 3.13t
     set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.13.0/python-3.13.0-amd64.exe"
     set ADDITIONAL_OPTIONS="Include_freethreaded=1"
     set PYTHON_EXEC="python3.13t"
-) else if "%DESIRED_PYTHON%"=="3.14" (
-    echo Python version is set to 3.14 or 3.14t
-    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0rc1-amd64.exe"
-) else if "%DESIRED_PYTHON%"=="3.14t" (
-    echo Python version is set to 3.14 or 3.14t
-    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0rc1-amd64.exe"
-    set ADDITIONAL_OPTIONS="Include_freethreaded=1"
-    set PYTHON_EXEC="python3.14t"
 ) else (
-    echo Python version is set to %DESIRED_PYTHON%
+    echo DESIRED_PYTHON not defined, Python version is set to %DESIRED_PYTHON%
     set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/%DESIRED_PYTHON%.0/python-%DESIRED_PYTHON%.0-amd64.exe" %= @lint-ignore =%
 )
 

.ci/pytorch/windows/internal/xpu_install.bat

@@ -13,9 +13,9 @@ if not exist "%SRC_DIR%\temp_build" mkdir "%SRC_DIR%\temp_build"
 :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/75d4eb97-914a-4a95-852c-7b9733d80f74/intel-deep-learning-essentials-2025.1.3.8_offline.exe
+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=2025.1.3+5
+set XPU_BUNDLE_VERSION=2025.0.1+20
 set XPU_BUNDLE_INSTALLED=0
 set XPU_BUNDLE_UNINSTALL=0
 set XPU_EXTRA_URL=NULL
@@ -24,9 +24,9 @@ set XPU_EXTRA_VERSION=2025.0.1+1226
 set XPU_EXTRA_INSTALLED=0
 set XPU_EXTRA_UNINSTALL=0
 
-if not [%XPU_VERSION%]==[] if [%XPU_VERSION%]==[2025.2] (
-    set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/24751ead-ddc5-4479-b9e6-f9fe2ff8b9f2/intel-deep-learning-essentials-2025.2.1.25_offline.exe
-    set XPU_BUNDLE_VERSION=2025.2.1+20
+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
 )
 
 :: Check if XPU bundle is target version or already installed
@@ -90,3 +90,14 @@ if errorlevel 1 exit /b 1
 del xpu_extra.exe
 
 :xpu_install_end
+
+if not "%XPU_ENABLE_KINETO%"=="1" goto install_end
+:: Install Level Zero SDK
+set XPU_EXTRA_LZ_URL=https://github.com/oneapi-src/level-zero/releases/download/v1.14.0/level-zero-sdk_1.14.0.zip
+curl -k -L %XPU_EXTRA_LZ_URL% --output "%SRC_DIR%\temp_build\level_zero_sdk.zip"
+echo "Installing level zero SDK..."
+7z x "%SRC_DIR%\temp_build\level_zero_sdk.zip" -o"%SRC_DIR%\temp_build\level_zero"
+set "INCLUDE=%SRC_DIR%\temp_build\level_zero\include;%INCLUDE%"
+del "%SRC_DIR%\temp_build\level_zero_sdk.zip"
+
+:install_end

.ci/pytorch/windows/setup_build.bat

@@ -7,8 +7,6 @@ call "internal\install_python.bat"
 
 %PYTHON_EXEC% --version
 set "PATH=%CD%\Python\Lib\site-packages\cmake\data\bin;%CD%\Python\Scripts;%CD%\Python;%PATH%"
-if "%DESIRED_PYTHON%" == "3.14t" %PYTHON_EXEC% -m pip install numpy==2.3.2 cmake
-if "%DESIRED_PYTHON%" == "3.14" %PYTHON_EXEC% -m pip install numpy==2.3.2 cmake
 if "%DESIRED_PYTHON%" == "3.13t" %PYTHON_EXEC% -m pip install numpy==2.2.1 cmake
 if "%DESIRED_PYTHON%" == "3.13" %PYTHON_EXEC% -m pip install numpy==2.1.2 cmake
 if "%DESIRED_PYTHON%" == "3.12" %PYTHON_EXEC% -m pip install numpy==2.0.2 cmake

.ci/wheel/build_wheel.sh

@@ -124,31 +124,20 @@ popd
 
 export TH_BINARY_BUILD=1
 export INSTALL_TEST=0 # dont install test binaries into site-packages
-export MACOSX_DEPLOYMENT_TARGET=11.0
+export MACOSX_DEPLOYMENT_TARGET=10.15
 export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
 
+SETUPTOOLS_PINNED_VERSION="==70.1.0"
+PYYAML_PINNED_VERSION="=5.3"
 EXTRA_CONDA_INSTALL_FLAGS=""
 CONDA_ENV_CREATE_FLAGS=""
 RENAME_WHEEL=true
 case $desired_python in
-    3.14t)
-        echo "Using 3.14 deps"
-        NUMPY_PINNED_VERSION="==2.1.0"
-        CONDA_ENV_CREATE_FLAGS="python-freethreading"
-        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-        desired_python="3.14.0rc1"
-        RENAME_WHEEL=false
-        ;;
-    3.14)
-        echo "Using 3.14t deps"
-        NUMPY_PINNED_VERSION="==2.1.0"
-        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-        desired_python="3.14.0rc1"
-        RENAME_WHEEL=false
-        ;;
     3.13t)
         echo "Using 3.13 deps"
-        NUMPY_PINNED_VERSION="==2.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=6.0.1"
+        NUMPY_PINNED_VERSION="=2.1.0"
         CONDA_ENV_CREATE_FLAGS="python-freethreading"
         EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
         desired_python="3.13"
@@ -156,23 +145,37 @@ case $desired_python in
         ;;
     3.13)
         echo "Using 3.13 deps"
-        NUMPY_PINNED_VERSION="==2.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=6.0.1"
+        NUMPY_PINNED_VERSION="=2.1.0"
         ;;
     3.12)
         echo "Using 3.12 deps"
-        NUMPY_PINNED_VERSION="==2.0.2"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=6.0.1"
+        NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     3.11)
         echo "Using 3.11 deps"
-        NUMPY_PINNED_VERSION="==2.0.2"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=5.3"
+        NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     3.10)
         echo "Using 3.10 deps"
-        NUMPY_PINNED_VERSION="==2.0.2"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=5.3"
+        NUMPY_PINNED_VERSION="=2.0.2"
+        ;;
+    3.9)
+        echo "Using 3.9 deps"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=5.3"
+        NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     *)
-        echo "Unsupported version $desired_python"
-        exit 1
+        echo "Using default deps"
+        NUMPY_PINNED_VERSION="=1.11.3"
         ;;
 esac
 
@@ -181,11 +184,8 @@ tmp_env_name="wheel_py$python_nodot"
 conda create ${EXTRA_CONDA_INSTALL_FLAGS} -yn "$tmp_env_name" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS}
 source activate "$tmp_env_name"
 
-PINNED_PACKAGES=(
-    "numpy${NUMPY_PINNED_VERSION}"
-)
-retry pip install "${PINNED_PACKAGES[@]}" -r "${pytorch_rootdir}/requirements-build.txt"
-pip install requests ninja typing-extensions
+retry pip install -r "${pytorch_rootdir}/requirements-build.txt"
+pip install "numpy=${NUMPY_PINNED_VERSION}"  "pyyaml${PYYAML_PINNED_VERSION}" requests ninja "setuptools${SETUPTOOLS_PINNED_VERSION}" typing-extensions
 retry pip install -r "${pytorch_rootdir}/requirements.txt" || true
 retry brew install libomp
 
@@ -199,7 +199,7 @@ export BUILD_TEST=OFF
 pushd "$pytorch_rootdir"
 echo "Calling setup.py bdist_wheel at $(date)"
 
-python setup.py bdist_wheel -d "$whl_tmp_dir" --plat-name ${mac_version}
+python setup.py bdist_wheel -d "$whl_tmp_dir"
 
 echo "Finished setup.py bdist_wheel at $(date)"
 

.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 12.9/13.0 builds have triton for Linux and Linux aarch64 binaries.
-if [[ "$DESIRED_CUDA" == "cu129" ]] || [[ "$DESIRED_CUDA" == "cu130" ]]; then
+# CUDA 12.9 builds have triton for Linux and Linux aarch64 binaries.
+if [[ "$DESIRED_CUDA" == "cu129" ]]; then
   TRITON_CONSTRAINT="platform_system == 'Linux'"
 fi
 

.circleci/scripts/binary_upload.sh

@@ -51,12 +51,16 @@ s3_upload() {
     s3_upload_dir="${s3_root_dir}/${UPLOAD_SUBFOLDER}/"
   fi
   (
+    cache_control_flag=""
+    if [[ "${UPLOAD_CHANNEL}" = "test" ]]; then
+      cache_control_flag="--cache-control='no-cache,no-store,must-revalidate'"
+    fi
     for pkg in ${PKG_DIR}/*.${extension}; do
       (
         set -x
         shm_id=$(sha256sum "${pkg}" | awk '{print $1}')
         ${AWS_S3_CP} --no-progress --acl public-read "${pkg}" "${s3_upload_dir}" \
-          --metadata "checksum-sha256=${shm_id}"
+          --metadata "checksum-sha256=${shm_id}" ${cache_control_flag}
       )
     done
   )

.circleci/scripts/binary_windows_build.sh

@@ -15,7 +15,8 @@ fi
 if [[ "$DESIRED_CUDA" == 'xpu' ]]; then
     export VC_YEAR=2022
     export USE_SCCACHE=0
-    export XPU_VERSION=2025.2
+    export XPU_VERSION=2025.1
+    export XPU_ENABLE_KINETO=1
 fi
 
 echo "Free space on filesystem before build:"

.circleci/scripts/binary_windows_test.sh

@@ -8,7 +8,7 @@ export VC_YEAR=2022
 
 if [[ "$DESIRED_CUDA" == 'xpu' ]]; then
     export VC_YEAR=2022
-    export XPU_VERSION=2025.2
+    export XPU_VERSION=2025.1
 fi
 
 pushd "$PYTORCH_ROOT/.ci/pytorch/"

.flake8

@@ -48,7 +48,6 @@ per-file-ignores =
     torch/__init__.py: F401,TOR901
     torch/_custom_op/impl.py: TOR901
     torch/_export/serde/upgrade.py: TOR901
-    torch/_functorch/predispatch.py: TOR901
     torch/_functorch/vmap.py: TOR901
     torch/_inductor/test_operators.py: TOR901
     torch/_library/abstract_impl.py: TOR901

.github/actionlint.yaml

@@ -12,9 +12,7 @@ self-hosted-runner:
     - linux.9xlarge.ephemeral
     - am2.linux.9xlarge.ephemeral
     - linux.12xlarge
-    - linux.12xlarge.memory
     - linux.24xlarge
-    - linux.24xlarge.memory
     - linux.24xlarge.ephemeral
     - linux.24xlarge.amd
     - linux.arm64.2xlarge

.github/actions/build-external-packages/action.yml

@@ -1,86 +0,0 @@
-# .github/workflows/build-external.yml
-name: Build External packages
-
-description: build external packages for PyTorch
-
-inputs:
-  cuda-version:
-    description: CUDA version to use
-    type: string
-    required: true
-    default: '12.8.1'
-  cuda-arch-list:
-    description: TORCH_CUDA_ARCH_LIST (e.g., "8.0;8.9;9.0")
-    type: string
-    required: true
-    default: ""
-  docker-image:
-    description: Base image to use
-    type: string
-    required: true
-  build-targets:
-    description: Build targets
-    type: string
-    required: true
-  torch-wheel-dir:
-    description: Directory to built torch wheel
-    type: string
-    required: false
-    default: dist
-  output-dir:
-    description: Directory to store build artifact
-    default: external
-    type: string
-    required: false
-
-outputs:
-  build_time:
-    description: "Total build time in seconds"
-    value: ${{ steps.build-external.outputs.build_time }}
-  output_dir:
-    description: "Directory where build artifact is stored"
-    value: ${{ steps.build-external.outputs.output_dir }}
-
-runs:
-  using: composite
-  steps:
-    - name: Build external packages in sequence
-      id: build-external
-      env:
-        SCCACHE_BUCKET: ossci-compiler-cache-circleci-v2
-        SCCACHE_REGION: us-east-1
-        CUDA_VERSION: ${{ inputs.cuda-version }}
-        TORCH_CUDA_ARCH_LIST: ${{ inputs.cuda-arch-list }}
-        BASE_IMAGE: ${{ inputs.docker-image }}
-        BUILD_TARGETS: ${{ inputs.build-targets }}
-        PARENT_OUTPUT_DIR: ${{ inputs.output-dir }}
-        TORCH_WHEELS_PATH: ${{ inputs.torch-wheel-dir }}
-      shell: bash
-      run: |
-        set -euo pipefail
-        python3 --version
-        docker images
-        START_TIME=$(date +%s)
-        (
-          cd .ci/lumen_cli
-          python3 -m pip install -e .
-        )
-        MAX_JOBS="$(nproc --ignore=6)"
-        export MAX_JOBS
-
-        # Split the comma-separated list and build each target
-        IFS=',' read -ra TARGETS <<< "$BUILD_TARGETS"
-        for target in "${TARGETS[@]}"; do
-          OUTPUT_DIR="$PARENT_OUTPUT_DIR/$target"
-          export OUTPUT_DIR
-          echo "Building external package: $target in directory $OUTPUT_DIR"
-          python3 -m cli.run build external "$target"
-        done
-
-        END_TIME=$(date +%s)
-        {
-          echo "build_time=$((END_TIME - START_TIME))"
-          if [ -d "$PARENT_OUTPUT_DIR" ]; then
-            echo "output_dir=$PARENT_OUTPUT_DIR"
-          fi
-        } >> "$GITHUB_OUTPUT"

.github/actions/checkout-pytorch/action.yml

@@ -57,21 +57,6 @@ runs:
         submodules: ${{ inputs.submodules }}
         show-progress: false
 
-    - name: Clean submodules post checkout
-      id: clean-submodules
-      if: ${{ steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' }}
-      shell: bash
-      env:
-        NO_SUDO: ${{ inputs.no-sudo }}
-      run: |
-        cd "${GITHUB_WORKSPACE}"
-        # Clean stale submodule dirs
-        if [ -z "${NO_SUDO}" ]; then
-          sudo git submodule foreach --recursive git clean -ffdx
-        else
-          git submodule foreach --recursive git clean -ffdx
-        fi
-
     - name: Clean workspace (try again)
       if: ${{ steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' &&
         (steps.first-clean.outcome != 'success' || steps.first-checkout-attempt.outcome != 'success') }}

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-fa5142928ee157aa65137c4ecff2fe9b1a9e0648
+bdb88e1d66f272cad72156c90ac8428ca61a601c

.github/ci_commit_pins/fbgemm_rocm.txt

@@ -1 +1 @@
-08ae0af1395c8d8471f4025deb6af9aef90b342f
+7f1de94a4c2d14f59ad4ca84538c36084ea6b2c8

.github/ci_commit_pins/torchao.txt

@@ -1 +1 @@
-f32431e593d0e9db86c502d3872dd67ee40a005f
+51c87b6ead6b7e098ada95d6a7609ee873b854cf

.github/ci_commit_pins/vllm.txt

@@ -1 +1 @@
-cc99baf14dacc2497d0c5ed84e076ef2c37f6a4d
+0ca2393b47e72c4424a49aa3b32c7c5d0e378a72

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-6c5478ff7c3d50dd1e3047d72ec5909bea474073
+095faec1e7b6cc47220181e74ae9cde2605f9b00

.github/ci_configs/vllm/Dockerfile.tmp_vllm

@@ -12,46 +12,54 @@ ARG BUILD_BASE_IMAGE=torch-nightly-base
 # by default, it uses devel-ubuntu22.04 official image.
 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
 
-# The logic is copied from https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile
-ARG GET_PIP_URL="https://bootstrap.pypa.io/get-pip.py"
 
-
-#################### TORCH NIGHTLY BASE IMAGE ####################
+#################### TORCH NIGHTLY  BASE IMAGE ####################
 # A base image for building vLLM with devel ubuntu 22.04, this is mainly used to build vllm in vllm builtkite ci
-FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 as torch-nightly-base
+From nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 as torch-nightly-base
+ARG CUDA_VERSION=12.8.1
+ARG PYTHON_VERSION=3.12
+ARG TARGETPLATFORM
+ENV DEBIAN_FRONTEND=noninteractive
 
-ARG CUDA_VERSION
-ARG PYTHON_VERSION
-ARG GET_PIP_URL
+RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
+    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
 
-# Install Python and other dependencies
-RUN apt-get update -y \
-    && apt-get install -y ccache software-properties-common git curl wget sudo vim \
-    && add-apt-repository -y ppa:deadsnakes/ppa \
-    && apt-get update -y \
-    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-    && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION} \
-    && python3 --version && python3 -m pip --version
+# Install Python and other dependencies if it does not existed
+RUN if ! command -v python3 >/dev/null || ! python3 --version | grep -q "${PYTHON_VERSION}"; then \
+      echo "Installing Python ${PYTHON_VERSION}..." && \
+      echo 'tzdata tzdata/Areas select America' | debconf-set-selections && \
+      echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections && \
+      apt-get update -y && \
+      apt-get install -y ccache software-properties-common git curl sudo && \
+      for i in 1 2 3; do \
+        add-apt-repository -y ppa:deadsnakes/ppa && break || \
+        { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
+      done && \
+      apt-get update -y && \
+      apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv && \
+      update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 && \
+      update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} && \
+      ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config && \
+      curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}; \
+   else \
+      echo "Python ${PYTHON_VERSION} already present, skipping setup."; \
+   fi \
+   && python3 --version && python3 -m pip --version
 
 # Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
 # as it was causing spam when compiling the CUTLASS kernels
 # Ensure gcc >= 10 to avoid CUTLASS issues (bug 92519)
 RUN current_gcc_version=$(gcc -dumpversion | cut -f1 -d.) && \
-    if command -v apt-get >/dev/null; then \
-        if [ "$current_gcc_version" -lt 10 ]; then \
-            echo "GCC version is $current_gcc_version, installing gcc-10..."; \
-            apt-get update \
-            && apt-get install -y gcc-10 g++-10 \
-            && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 \
-            && update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100; \
-        else \
-            echo "GCC version is $current_gcc_version, no need to install gcc-10."; \
-        fi \
-    fi \
-    && gcc --version && g++ --version
+    if [ "$current_gcc_version" -lt 10 ]; then \
+      echo "GCC version is $current_gcc_version, installing gcc-10..."; \
+      apt-get update && \
+      apt-get install -y gcc-10 g++-10 && \
+      update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 && \
+      update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100; \
+    else \
+      echo "GCC version is $current_gcc_version, no need to install gcc-10."; \
+    fi && \
+    gcc --version && g++ --version
 
 # install uv for faster pip installs
 RUN --mount=type=cache,target=/root/.cache/uv \
@@ -59,8 +67,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
-# Use copy mode to avoid hardlink failures with Docker cache mounts
-ENV UV_LINK_MODE=copy
 
 #################### TORCH NIGHTLY  BASE IMAGE ####################
 
@@ -71,21 +77,6 @@ ENV UV_LINK_MODE=copy
 FROM ${BUILD_BASE_IMAGE} AS base
 USER root
 
-ARG CUDA_VERSION
-ARG PYTHON_VERSION
-
-# TODO (huydhn): Only work with PyTorch manylinux builder
-ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
-
-# Install some system dependencies and double check python version
-RUN if command -v apt-get >/dev/null; then \
-        apt-get update -y \
-        && apt-get install -y ccache software-properties-common git curl wget sudo vim; \
-    else \
-        dnf install -y git curl wget sudo vim; \
-    fi \
-    && python3 --version && python3 -m pip --version
-
 # Workaround for https://github.com/openai/triton/issues/2507 and
 # https://github.com/pytorch/pytorch/issues/107960 -- hopefully
 # this won't be needed for future versions of this docker image
@@ -99,8 +90,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     fi
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
-# Use copy mode to avoid hardlink failures with Docker cache mounts
-ENV UV_LINK_MODE=copy
 
 WORKDIR /workspace
 
@@ -123,17 +112,18 @@ ARG PINNED_TORCH_VERSION
 RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
     --mount=type=cache,target=/root/.cache/uv \
     if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
-        echo "[INFO] Installing torch wheels to build vllm"; \
         torch_whl=$(find /dist -maxdepth 1 -name 'torch-*.whl' -print -quit); \
-        vision_whl=$(find /dist -name 'torchvision*.whl' | head -n1 | xargs); \
-        audio_whl=$(find /dist -name 'torchaudio*.whl' | head -n1 | xargs); \
-        uv pip install --system "${torch_whl}[opt-einsum]" "${vision_whl}" "${audio_whl}" /dist/*.whl; \
+        vision_whl=$(find /dist/vision -name 'torchvision*.whl' | head -n1 | xargs); \
+        audio_whl=$(find /dist/audio -name 'torchaudio*.whl' | head -n1 | xargs); \
+        uv pip install --system "${torch_whl}[opt-einsum]"; \
+        uv pip install --system "${vision_whl}"; \
+        uv pip install --system "${audio_whl}"; \
     elif [ -n "$PINNED_TORCH_VERSION" ]; then \
-        echo "[INFO] Installing pinned torch nightly version to build vllm: $PINNED_TORCH_VERSION"; \
-        uv pip install --system "$PINNED_TORCH_VERSION" --index-url https://download.pytorch.org/whl/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
+        echo "[INFO] Installing pinned torch nightly version: $PINNED_TORCH_VERSION"; \
+        uv pip install --system "$PINNED_TORCH_VERSION" --index-url https://download.pytorch.org/whl/nightly/cu128; \
     else \
-        echo "[INFO] Installing torch nightly with latest one to build vllm"; \
-        uv pip install --system torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
+        echo "[INFO] Installing torch nightly with latest one"; \
+        uv pip install --system torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128; \
     fi
 
 # Install numba 0.61.2 for cuda environment
@@ -142,25 +132,19 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 # Install common dependencies from vllm common.txt
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/common.txt
+uv pip install --system -r requirements/common.txt
+
 
 # Must put before installing xformers, so it can install the correct version of xfomrers.
-ARG xformers_cuda_arch_list='7.5;8.0+PTX;9.0a'
-ENV TORCH_CUDA_ARCH_LIST=${xformers_cuda_arch_list}
-
+ARG torch_cuda_arch_list='8.0;8.6;8.9;9.0'
+ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
 
-RUN echo ${TORCH_CUDA_ARCH_LIST}
-RUN echo ${MAX_JOBS}
-RUN pip freeze | grep -E 'ninja'
-
 # Build xformers with cuda and torch nightly/wheel
 # following official xformers guidance: https://github.com/facebookresearch/xformers#build
-# sha for https://github.com/facebookresearch/xformers/tree/v0.0.32.post2
-ARG XFORMERS_COMMIT=5d4b92a5e5a9c6c6d4878283f47d82e17995b468
+ARG XFORMERS_COMMIT=f2de641ef670510cadab099ce6954031f52f191c
 ENV CCACHE_DIR=/root/.cache/ccache
-
 RUN --mount=type=cache,target=/root/.cache/ccache \
     --mount=type=cache,target=/root/.cache/uv \
     echo 'git clone xformers...' \
@@ -173,15 +157,14 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
     && python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose \
     && cd .. \
     && rm -rf xformers
-
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system xformers-dist/*.whl --verbose
 
 # Build can take a long time, and the torch nightly version fetched from url can be different in next docker stage.
 # track the nightly torch version used in the build, when we set up runtime environment we can make sure the version is the same
 RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio' > torch_build_versions.txt
+RUN cat  torch_build_versions.txt
 
-RUN cat torch_build_versions.txt
 RUN pip freeze | grep -E 'torch|xformers|torchvision|torchaudio'
 
 #################### BASE BUILD IMAGE ####################
@@ -192,6 +175,9 @@ RUN pip freeze | grep -E 'torch|xformers|torchvision|torchaudio'
 FROM base AS build
 ARG TARGETPLATFORM
 
+ENV UV_HTTP_TIMEOUT=500
+ENV UV_INDEX_STRATEGY="unsafe-best-match"
+
 COPY . .
 
 RUN python3 use_existing_torch.py
@@ -206,7 +192,7 @@ RUN --mount=type=bind,source=.git,target=.git \
 # Max jobs used by Ninja to build extensions
 ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
-ARG nvcc_threads=4
+ARG nvcc_threads=2
 ENV NVCC_THREADS=$nvcc_threads
 ARG torch_cuda_arch_list='8.0;8.6;8.9;9.0'
 ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
@@ -230,14 +216,11 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         && export SCCACHE_S3_NO_CREDENTIALS=${SCCACHE_S3_NO_CREDENTIALS} \
         && export SCCACHE_IDLE_TIMEOUT=0 \
         && export CMAKE_BUILD_TYPE=Release \
-        && export VLLM_DOCKER_BUILD_CONTEXT=1 \
         && sccache --show-stats \
         && python3 setup.py bdist_wheel --dist-dir=vllm-dist --py-limited-api=cp38 \
         && sccache --show-stats; \
     fi
 
-ARG vllm_target_device="cuda"
-ENV VLLM_TARGET_DEVICE=${vllm_target_device}
 ENV CCACHE_DIR=/root/.cache/ccache
 RUN --mount=type=cache,target=/root/.cache/ccache \
     --mount=type=cache,target=/root/.cache/uv \
@@ -246,13 +229,12 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
         # Clean any existing CMake artifacts
         rm -rf .deps && \
         mkdir -p .deps && \
-        export VLLM_DOCKER_BUILD_CONTEXT=1 && \
         python3 setup.py bdist_wheel --dist-dir=vllm-dist --py-limited-api=cp38; \
     fi
 
-RUN echo "[INFO] Listing current directory:" && \
+RUN echo "[DEBUG] Listing  current directory:" && \
     ls -al && \
-    echo "[INFO] Showing torch_build_versions.txt content:" && \
+    echo "[DEBUG] Showing torch_build_versions.txt content:" && \
     cat torch_build_versions.txt
 
 #################### WHEEL BUILD IMAGE ####################
@@ -262,40 +244,42 @@ RUN echo "[INFO] Listing current directory:" && \
 # Setup clean environment for vLLM for test and api server using ubuntu22.04 with AOT flashinfer
 FROM ${FINAL_BASE_IMAGE} AS vllm-base
 USER root
-
-ARG CUDA_VERSION
-ARG PYTHON_VERSION
-ARG GET_PIP_URL
-
-# TODO (huydhn): Only work with PyTorch manylinux builder
-ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
-
 # prepare for environment starts
 WORKDIR /workspace
 
-# Install Python and other dependencies
-RUN if command -v apt-get >/dev/null; then \
-        apt-get update -y \
-        && apt-get install -y ccache software-properties-common git curl wget sudo vim \
-        && add-apt-repository -y ppa:deadsnakes/ppa \
-        && apt-get update -y \
-        && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-        && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-        && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-        && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-        && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION}; \
-    else \
-        dnf install -y git curl wget sudo vim; \
-    fi \
-    && python3 --version && python3 -m pip --version
+RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
+    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
+
+# Install Python and other dependencies if it does not existed
+RUN if ! command -v python3 >/dev/null || ! python3 --version | grep -q "${PYTHON_VERSION}"; then \
+      echo "Installing Python ${PYTHON_VERSION}..." && \
+      echo 'tzdata tzdata/Areas select America' | debconf-set-selections && \
+      echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections && \
+      apt-get update -y && \
+      apt-get install -y ccache software-properties-common git curl sudo && \
+      for i in 1 2 3; do \
+        add-apt-repository -y ppa:deadsnakes/ppa && break || \
+        { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
+      done && \
+      apt-get update -y && \
+      apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv && \
+      update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 && \
+      update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} && \
+      ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config && \
+      curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}; \
+   else \
+      echo "Python ${PYTHON_VERSION} already present, skipping setup."; \
+   fi \
+   && python3 --version && python3 -m pip --version
+
 
 # Get the torch versions, and whls used in previous stagtes for consistency
 COPY --from=base /workspace/torch_build_versions.txt ./torch_build_versions.txt
 COPY --from=base /workspace/xformers-dist /wheels/xformers
 COPY --from=build /workspace/vllm-dist /wheels/vllm
-RUN echo "[INFO] Listing current directory before torch install step:" && \
+RUN echo "[DEBUG] Listing current directory before torch install step:" && \
     ls -al && \
-    echo "[INFO] Showing torch_build_versions.txt content:" && \
+    echo "[DEBUG] Showing torch_build_versions.txt content:" && \
     cat torch_build_versions.txt
 
 # Workaround for https://github.com/openai/triton/issues/2507 and
@@ -304,6 +288,7 @@ RUN echo "[INFO] Listing current directory before torch install step:" && \
 # or future versions of triton.
 RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 
+
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
     if ! python3 -m uv --version > /dev/null 2>&1; then \
@@ -311,8 +296,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     fi
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
-# Use copy mode to avoid hardlink failures with Docker cache mounts
-ENV UV_LINK_MODE=copy
 
 # Default mount file as placeholder, this just avoid the mount error
 ARG TORCH_WHEELS_PATH="./requirements"
@@ -323,13 +306,15 @@ RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
     --mount=type=cache,target=/root/.cache/uv \
     if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
         torch_whl=$(find /dist -maxdepth 1 -name 'torch-*.whl' -print -quit); \
-        vision_whl=$(find /dist -name 'torchvision*.whl' | head -n1 | xargs); \
-        audio_whl=$(find /dist -name 'torchaudio*.whl' | head -n1 | xargs); \
-        echo "[INFO] Use wheels to build : '${torch_whl}' '${audio_whl}' '${vision_whl}'"; \
-        uv pip install --system "${torch_whl}[opt-einsum]" "${vision_whl}" "${audio_whl}" /dist/*.whl; \
+        vision_whl=$(find /dist/vision -name 'torchvision*.whl' | head -n1 | xargs); \
+        audio_whl=$(find /dist/audio -name 'torchaudio*.whl' | head -n1 | xargs); \
+        echo "Found: '${torch_whl}' '${audio_whl}' '${vision_whl}'"; \
+        uv pip install --system "${torch_whl}[opt-einsum]"; \
+        uv pip install --system "${vision_whl}"; \
+        uv pip install --system "${audio_whl}"; \
     else \
         echo "[INFO] Installing torch versions from torch_build_versions.txt"; \
-        uv pip install --system $(cat torch_build_versions.txt | xargs) --index-url https://download.pytorch.org/whl/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
+        uv pip install --system $(cat torch_build_versions.txt | xargs) --index-url https://download.pytorch.org/whl/nightly/cu128; \
     fi
 
 # Install the vllm wheel from previous stage
@@ -340,8 +325,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system /wheels/xformers/*.whl --verbose
 
+
 # Build flashinfer from source.
-ARG torch_cuda_arch_list='8.0;8.9;9.0a;10.0a;12.0'
+ARG torch_cuda_arch_list='8.0;8.9;9.0a'
 # install package for build flashinfer
 # see issue: https://github.com/flashinfer-ai/flashinfer/issues/738
 
@@ -352,7 +338,7 @@ ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 # Build flashinfer for torch nightly from source around 10 mins
 ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
 # Keep this in sync with https://github.com/vllm-project/vllm/blob/main/requirements/cuda.txt
-ARG FLASHINFER_GIT_REF="v0.2.14.post1"
+ARG FLASHINFER_GIT_REF="v0.2.9rc2"
 RUN --mount=type=cache,target=/root/.cache/uv \
     git clone --depth 1 --recursive --shallow-submodules \
         --branch ${FLASHINFER_GIT_REF} \
@@ -370,7 +356,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 # Logging to confirm the torch versions
 RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'
-RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio\|^xformers\|^vllm\|^flashinfer' > build_summary.txt
 ################### VLLM INSTALLED IMAGE ####################
 
 
@@ -379,8 +364,6 @@ FROM vllm-base as test
 
 ENV UV_HTTP_TIMEOUT=500
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
-# Use copy mode to avoid hardlink failures with Docker cache mounts
-ENV UV_LINK_MODE=copy
 
 COPY tests/ tests/
 COPY examples examples
@@ -409,6 +392,11 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/nightly_torch_test.txt
 
+# Workaround for #17068
+# pinned commit for v2.2.4
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system --no-build-isolation "git+https://github.com/state-spaces/mamba@95d8aba8a8c75aedcaa6143713b11e745e7cd0d9#egg=mamba-ssm"
+
 # Logging to confirm the torch versions
 RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'
 
@@ -423,5 +411,4 @@ FROM scratch as export-wheels
 # Just copy the wheels we prepared in previous stages
 COPY --from=base /workspace/xformers-dist /wheels/xformers
 COPY --from=build /workspace/vllm-dist /wheels/vllm
-COPY --from=vllm-base /workspace/build_summary.txt /wheels/build_summary.txt
 COPY --from=vllm-base /workspace/wheels/flashinfer /wheels/flashinfer-python

.github/dependabot.yml

@@ -1,24 +0,0 @@
-version: 2
-updates:
-  # Update to the latest transformers version with dependabot
-  - package-ecosystem: "pip"
-    directory: "/.ci/docker/ci_commit_pins"
-    schedule:
-      interval: "daily"
-    target-branch: "main"
-    allow:
-      - dependency-name: "transformers"
-    ignore:
-      - dependency-name: "*"
-        update-types: ["version-update:semver-patch"]
-    commit-message:
-      prefix: "[Dependabot] Update"
-      include: "scope"
-    labels:
-      - "dependencies"
-      - "open source"
-      - "python"
-      - "topic: not user facing"
-      - "module: ci"
-      - "module: inductor"
-      - "ciflow/inductor"

.github/pytorch-probot.yml

@@ -27,7 +27,6 @@ ciflow_push_tags:
 - ciflow/trunk
 - ciflow/unstable
 - ciflow/xpu
-- ciflow/vllm
 - ciflow/torchbench
 - ciflow/op-benchmark
 - ciflow/pull

.github/requirements/pip-requirements-macOS.txt

@@ -28,7 +28,7 @@ pyyaml==6.0.2
 scipy==1.12.0
 setuptools==72.1.0
 sympy==1.13.3
-tlparse==0.4.0
+tlparse==0.3.30
 tensorboard==2.13.0
 typing-extensions==4.12.2
 unittest-xml-reporting<=3.2.0,>=2.0.0

.github/scripts/build_triton_wheel.py

@@ -84,7 +84,6 @@ def build_triton(
                 ["git", "checkout", f"release/{ver}.{rev}.x"], cwd=triton_basedir
             )
         else:
-            check_call(["git", "fetch", "origin", commit_hash], cwd=triton_basedir)
             check_call(["git", "checkout", commit_hash], cwd=triton_basedir)
 
         # change built wheel name and version

.github/scripts/generate_binary_build_matrix.py

@@ -16,17 +16,17 @@ from typing import Optional
 
 
 # NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
-CUDA_ARCHES = ["12.6", "12.8", "13.0"]
+CUDA_ARCHES = ["12.6", "12.8", "12.9"]
 CUDA_STABLE = "12.8"
 CUDA_ARCHES_FULL_VERSION = {
     "12.6": "12.6.3",
     "12.8": "12.8.1",
-    "13.0": "13.0.0",
+    "12.9": "12.9.1",
 }
 CUDA_ARCHES_CUDNN_VERSION = {
     "12.6": "9",
     "12.8": "9",
-    "13.0": "9",
+    "12.9": "9",
 }
 
 # NOTE: Please also update the ROCm sources in `PIP_SOURCES` in tools/nightly.py when changing this
@@ -38,82 +38,82 @@ CPU_AARCH64_ARCH = ["cpu-aarch64"]
 
 CPU_S390X_ARCH = ["cpu-s390x"]
 
-CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "13.0-aarch64"]
+CUDA_AARCH64_ARCHES = ["12.9-aarch64"]
 
 
 PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
     "12.6": (
-        "nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
-        "nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | "
-        "nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | "
-        "nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | "
-        "nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | "
-        "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
-        "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | "
-        "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | "
-        "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'"
+        "nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
     "12.8": (
-        "nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
-        "nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | "
-        "nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | "
-        "nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | "
-        "nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | "
-        "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
-        "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | "
-        "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
-        "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
+        "nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
-    "13.0": (
-        "nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | "
-        "nvidia-cublas==13.0.0.19; platform_system == 'Linux' | "
-        "nvidia-cufft==12.0.0.15; platform_system == 'Linux' | "
-        "nvidia-curand==10.4.0.35; platform_system == 'Linux' | "
-        "nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | "
-        "nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | "
-        "nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | "
-        "nvidia-nvtx==13.0.39; platform_system == 'Linux' | "
-        "nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | "
-        "nvidia-cufile==1.15.0.42; platform_system == 'Linux'"
+    "12.9": (
+        "nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
     "xpu": (
-        "intel-cmplr-lib-rt==2025.2.1 | "
-        "intel-cmplr-lib-ur==2025.2.1 | "
-        "intel-cmplr-lic-rt==2025.2.1 | "
-        "intel-sycl-rt==2025.2.1 | "
-        "oneccl-devel==2021.16.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "oneccl==2021.16.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "impi-rt==2021.16.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "onemkl-sycl-blas==2025.2.0 | "
-        "onemkl-sycl-dft==2025.2.0 | "
-        "onemkl-sycl-lapack==2025.2.0 | "
-        "onemkl-sycl-rng==2025.2.0 | "
-        "onemkl-sycl-sparse==2025.2.0 | "
-        "dpcpp-cpp-rt==2025.2.1 | "
-        "intel-opencl-rt==2025.2.1 | "
-        "mkl==2025.2.0 | "
-        "intel-openmp==2025.2.1 | "
-        "tbb==2022.2.0 | "
-        "tcmlib==1.4.0 | "
-        "umf==0.11.0 | "
-        "intel-pti==0.13.1"
+        "intel-cmplr-lib-rt==2025.1.1 | "
+        "intel-cmplr-lib-ur==2025.1.1 | "
+        "intel-cmplr-lic-rt==2025.1.1 | "
+        "intel-sycl-rt==2025.1.1 | "
+        "oneccl-devel==2021.15.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "oneccl==2021.15.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "impi-rt==2021.15.0; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "onemkl-sycl-blas==2025.1.0 | "
+        "onemkl-sycl-dft==2025.1.0 | "
+        "onemkl-sycl-lapack==2025.1.0 | "
+        "onemkl-sycl-rng==2025.1.0 | "
+        "onemkl-sycl-sparse==2025.1.0 | "
+        "dpcpp-cpp-rt==2025.1.1 | "
+        "intel-opencl-rt==2025.1.1 | "
+        "mkl==2025.1.0 | "
+        "intel-openmp==2025.1.1 | "
+        "tbb==2022.1.0 | "
+        "tcmlib==1.3.0 | "
+        "umf==0.10.0 | "
+        "intel-pti==0.12.3"
     ),
 }
 
@@ -124,7 +124,9 @@ def get_nccl_wheel_version(arch_version: str) -> str:
     requirements = map(
         str.strip, re.split("[;|]", PYTORCH_EXTRA_INSTALL_REQUIREMENTS[arch_version])
     )
-    return next(x for x in requirements if x.startswith("nvidia-nccl")).split("==")[1]
+    return next(x for x in requirements if x.startswith("nvidia-nccl-cu")).split("==")[
+        1
+    ]
 
 
 def read_nccl_pin(arch_version: str) -> str:
@@ -191,7 +193,7 @@ LIBTORCH_CONTAINER_IMAGES: dict[str, str] = {
     "cpu": "libtorch-cxx11-builder:cpu",
 }
 
-FULL_PYTHON_VERSIONS = ["3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t"]
+FULL_PYTHON_VERSIONS = ["3.9", "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t"]
 
 
 def translate_desired_cuda(gpu_arch_type: str, gpu_arch_version: str) -> str:
@@ -309,20 +311,19 @@ def generate_wheels_matrix(
                 else arch_version
             )
 
-            # TODO: Enable python 3.14 for rest
-            if os not in [
-                "linux",
-                "linux-aarch64",
-                "linux-s390x",
-                "macos-arm64",
-                "windows",
-            ] and (python_version == "3.14" or python_version == "3.14t"):
+            # TODO: Enable python 3.13t on cpu-s390x
+            if gpu_arch_type == "cpu-s390x" and python_version == "3.13t":
+                continue
+            # TODO: Enable python 3.14 on non linux OSes
+            if os != "linux" and (
+                python_version == "3.14" or python_version == "3.14t"
+            ):
                 continue
 
             # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
 
             if (
-                arch_version in ["13.0", "12.8", "12.6"]
+                arch_version in ["12.9", "12.8", "12.6"]
                 and os == "linux"
                 or arch_version in CUDA_AARCH64_ARCHES
             ):
@@ -355,6 +356,29 @@ def generate_wheels_matrix(
                         ),  # include special case for aarch64 build, remove the -aarch64 postfix
                     }
                 )
+                # Special build building to use on Colab. Python 3.11 for 12.6 CUDA
+                if python_version == "3.11" and arch_version == CUDA_STABLE:
+                    ret.append(
+                        {
+                            "python_version": python_version,
+                            "gpu_arch_type": gpu_arch_type,
+                            "gpu_arch_version": gpu_arch_version,
+                            "desired_cuda": translate_desired_cuda(
+                                gpu_arch_type, gpu_arch_version
+                            ),
+                            "container_image": WHEEL_CONTAINER_IMAGES[
+                                arch_version
+                            ].split(":")[0],
+                            "container_image_tag_prefix": WHEEL_CONTAINER_IMAGES[
+                                arch_version
+                            ].split(":")[1],
+                            "package_type": package_type,
+                            "pytorch_extra_install_requirements": "",
+                            "build_name": f"{package_type}-py{python_version}-{gpu_arch_type}{gpu_arch_version}-full".replace(  # noqa: B950
+                                ".", "_"
+                            ),
+                        }
+                    )
             else:
                 ret.append(
                     {
@@ -385,6 +409,6 @@ def generate_wheels_matrix(
     return ret
 
 
-validate_nccl_dep_consistency("13.0")
+validate_nccl_dep_consistency("12.9")
 validate_nccl_dep_consistency("12.8")
 validate_nccl_dep_consistency("12.6")

.github/scripts/prepare_vllm_wheels.sh

@@ -1,91 +0,0 @@
-#!/usr/bin/env bash
-
-set -eux
-
-torch_version=$(unzip -p torch-* '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
-nightly=$(echo ${torch_version} | cut -d'.' -f4)
-
-# Copied from .ci/manywheel/build_common.sh
-make_wheel_record() {
-  fpath=$1
-  if echo $fpath | grep RECORD >/dev/null 2>&1; then
-    echo "$fpath,,"
-  else
-    fhash=$(openssl dgst -sha256 -binary $fpath | openssl base64 | sed -e 's/+/-/g' | sed -e 's/\//_/g' | sed -e 's/=//g')
-    fsize=$(ls -nl $fpath | awk '{print $5}')
-    echo "$fpath,sha256=$fhash,$fsize"
-  fi
-}
-
-change_wheel_version() {
-  local package=$1
-  local wheel=$2
-  local f_version=$3
-  local t_version=$4
-
-  # Extract the wheel
-  ${PYTHON_EXECUTABLE} -mwheel unpack $wheel
-
-  mv "${package}-${f_version}" "${package}-${t_version}"
-  # Change the version from f_version to t_version in the dist-info dir
-  pushd "${package}-${t_version}"
-  mv "${package}-${f_version}.dist-info" "${package}-${t_version}.dist-info"
-
-  pushd "${package}-${t_version}.dist-info"
-  sed -i "s/${package}-${f_version}.dist-info/${package}-${t_version}.dist-info/g" RECORD
-
-  # Update the version in METADATA and its SHA256 hash
-  sed -i "s/Version: ${f_version}/Version: ${t_version}/g" METADATA
-  # then add PyTorch nightly dependency of vLLM
-  if [[ "${package}" == vllm ]] || [[ "${package}" == xformers ]]; then
-    sed -i "/License-File/a\Requires-Dist: torch==${torch_version}" METADATA
-  fi
-  sed -i '/METADATA,sha256/d' RECORD
-  popd
-
-  make_wheel_record "${package}-${t_version}.dist-info/METADATA" >> "${package}-${t_version}.dist-info/RECORD"
-  popd
-
-  # Repack the wheel
-  ${PYTHON_EXECUTABLE} -mwheel pack "${package}-${t_version}"
-
-  # Clean up
-  rm -rf "${package}-${t_version}"
-}
-
-repackage_wheel() {
-  local package=$1
-  pushd $package
-
-  local orig_wheel=$(find . -name *${package//-/_}*)
-  local orig_version=$(unzip -p $orig_wheel '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
-
-  local version=""
-  if [[ "${package}" == vllm ]]; then
-    # Copied from vllm/.buildkite/scripts/upload-wheels.sh
-    version=1.0.0
-  else
-    version=$(echo $orig_version | tr '.+' '.' | cut -d'.' -f1-3)
-  fi
-  local nightly_version=$version.$nightly
-
-  # Use nightly version
-  change_wheel_version ${package//-/_} $orig_wheel $orig_version $nightly_version
-  # Clean up
-  rm "${orig_wheel}"
-
-  auditwheel repair --plat $PLATFORM *.whl \
-    --exclude libc10* --exclude libtorch* --exclude libcu* --exclude libnv*
-  local repair_wheel=$(find wheelhouse -name *${PLATFORM}*)
-  local repair_wheel=$(basename ${repair_wheel})
-  popd
-
-  cp ${package}/wheelhouse/${repair_wheel} .
-  rm -rf $package
-}
-
-pushd externals/vllm/wheels
-for package in xformers flashinfer-python vllm; do
-  repackage_wheel $package
-done
-popd

.github/scripts/test_trymerge.py

@@ -27,7 +27,6 @@ from trymerge import (
     get_drci_classifications,
     gh_get_team_members,
     GitHubPR,
-    iter_issue_timeline_until_comment,
     JobCheckState,
     main as trymerge_main,
     MandatoryChecksMissingError,
@@ -35,8 +34,6 @@ from trymerge import (
     RE_GHSTACK_DESC,
     read_merge_rules,
     remove_job_name_suffix,
-    sha_from_committed_event,
-    sha_from_force_push_after,
     validate_revert,
 )
 
@@ -127,7 +124,7 @@ def mock_parse_args(revert: bool = False, force: bool = False) -> Any:
             self.force = force
             self.pr_num = 76123
             self.dry_run = True
-            self.comment_id = 12345  # Set to non-zero value
+            self.comment_id = 0
             self.reason = "this is for testing"
             self.ignore_current = False
             self.check_mergeability = False
@@ -155,9 +152,9 @@ def mock_revert(
 def mock_merge(
     pr: GitHubPR,
     repo: GitRepo,
-    comment_id: int,
     dry_run: bool = False,
     skip_mandatory_checks: bool = False,
+    comment_id: Optional[int] = None,
     timeout_minutes: int = 400,
     stale_pr_days: int = 3,
     ignore_current: bool = False,
@@ -473,9 +470,9 @@ class TestTryMerge(TestCase):
         mock_merge.assert_called_once_with(
             mock.ANY,
             mock.ANY,
-            comment_id=mock.ANY,
             dry_run=mock.ANY,
             skip_mandatory_checks=True,
+            comment_id=mock.ANY,
             ignore_current=False,
         )
 
@@ -488,9 +485,9 @@ class TestTryMerge(TestCase):
         mock_merge.assert_called_once_with(
             mock.ANY,
             mock.ANY,
-            comment_id=mock.ANY,
             dry_run=mock.ANY,
             skip_mandatory_checks=False,
+            comment_id=mock.ANY,
             ignore_current=False,
         )
 
@@ -1141,176 +1138,5 @@ Pull Request resolved: https://github.com/pytorch/pytorch/pull/154394"""
         )
 
 
-@mock.patch("trymerge.gh_graphql", side_effect=mocked_gh_graphql)
-@mock.patch("trymerge.gh_fetch_merge_base", return_value="")
-@mock.patch(
-    "trymerge.get_drci_classifications", side_effect=mocked_drci_classifications
-)
-class TestTimelineFunctions(TestCase):
-    """Tests for the new timeline-related functions"""
-
-    def test_sha_from_committed_event(self, *args: Any) -> None:
-        """Test extracting SHA from committed event"""
-        # Based on actual GitHub API format - committed events have "sha" at top level
-        event = {
-            "event": "committed",
-            "sha": "fb21ce932ded6670c918804a0d9151b773770a7c",
-        }
-        self.assertEqual(
-            sha_from_committed_event(event), "fb21ce932ded6670c918804a0d9151b773770a7c"
-        )
-
-        # Test with missing SHA
-        event_no_sha = {"event": "committed"}
-        self.assertIsNone(sha_from_committed_event(event_no_sha))
-
-    def test_sha_from_force_push_after(self, *args: Any) -> None:
-        """Test extracting SHA from force push event"""
-        # NOTE: The current function doesn't handle the actual GitHub API format
-        # Real force push events have "commit_id" at top level, but this function
-        # looks for "after", "after_commit", "after_sha", or "head_sha" fields
-
-        # Test with the legacy format the current function handles
-        event_legacy = {
-            "event": "head_ref_force_pushed",
-            "after": {"sha": "ef22bcbc54bb0f787e1e4ffd3d83df18fc407f5e"},
-        }
-        self.assertEqual(
-            sha_from_force_push_after(event_legacy),
-            "ef22bcbc54bb0f787e1e4ffd3d83df18fc407f5e",
-        )
-
-        # Test with current GitHub API format (should return None with current implementation)
-        event_real_api = {
-            "event": "head_ref_force_pushed",
-            "commit_id": "ef22bcbc54bb0f787e1e4ffd3d83df18fc407f5e",
-        }
-        self.assertEqual(
-            sha_from_force_push_after(event_real_api),
-            "ef22bcbc54bb0f787e1e4ffd3d83df18fc407f5e",
-        )  # Current function doesn't handle commit_id
-
-        # Test with missing SHA
-        event_no_sha = {"event": "head_ref_force_pushed"}
-        self.assertIsNone(sha_from_force_push_after(event_no_sha))
-
-    @mock.patch("trymerge.gh_fetch_json_list")
-    def test_iter_issue_timeline_until_comment(
-        self, mock_gh_fetch_json_list: Any, *args: Any
-    ) -> None:
-        """Test timeline iteration until target comment"""
-        # Mock timeline data based on actual GitHub API format
-        timeline_data = [
-            {"event": "commented", "id": 100, "body": "first comment"},
-            {"event": "committed", "sha": "fb21ce932ded6670c918804a0d9151b773770a7c"},
-            {"event": "commented", "id": 200, "body": "target comment"},
-            {"event": "commented", "id": 300, "body": "after target"},
-        ]
-        mock_gh_fetch_json_list.return_value = timeline_data
-
-        # Test iteration stops at target comment
-        events = list(iter_issue_timeline_until_comment("pytorch", "pytorch", 123, 200))
-        self.assertEqual(len(events), 3)  # Should stop at target comment
-        self.assertEqual(events[0]["event"], "commented")
-        self.assertEqual(events[0]["id"], 100)
-        self.assertEqual(events[1]["event"], "committed")
-        self.assertEqual(events[1]["sha"], "fb21ce932ded6670c918804a0d9151b773770a7c")
-        self.assertEqual(events[2]["event"], "commented")
-        self.assertEqual(events[2]["id"], 200)
-
-    @mock.patch("trymerge.gh_fetch_json_list")
-    def test_iter_issue_timeline_until_comment_not_found(
-        self, mock_gh_fetch_json_list: Any, *args: Any
-    ) -> None:
-        """Test timeline iteration when target comment is not found"""
-        # Mock empty timeline
-        mock_gh_fetch_json_list.return_value = []
-
-        events = list(iter_issue_timeline_until_comment("pytorch", "pytorch", 123, 999))
-        self.assertEqual(len(events), 0)
-
-    @mock.patch("trymerge.iter_issue_timeline_until_comment")
-    def test_get_commit_sha_at_comment_commit_after_comment(
-        self, mock_iter_timeline: Any, *args: Any
-    ) -> None:
-        """Test get_commit_sha_at_comment returns correct SHA after comment"""
-        mock_iter_timeline.return_value = [
-            {"event": "committed", "sha": "commit1"},
-            {"event": "committed", "sha": "commit2"},
-            {"event": "commented", "id": 100},
-            {"event": "head_ref_force_pushed", "after": {"sha": "commit3"}},
-        ]
-        pr = GitHubPR("pytorch", "pytorch", 77700)
-        sha = pr.get_commit_sha_at_comment(100)
-        self.assertEqual(sha, "commit2")
-
-    @mock.patch("trymerge.iter_issue_timeline_until_comment")
-    def test_get_commit_sha_at_comment_force_push_before_comment(
-        self, mock_iter_timeline: Any, *args: Any
-    ) -> None:
-        mock_iter_timeline.return_value = [
-            {"event": "committed", "sha": "commit1"},
-            {"event": "committed", "sha": "commit2"},
-            {"event": "head_ref_force_pushed", "commit_id": "commit3"},
-            {"event": "commented", "id": 100},
-        ]
-        pr = GitHubPR("pytorch", "pytorch", 77700)
-        sha = pr.get_commit_sha_at_comment(100)
-        self.assertEqual(sha, "commit3")
-
-    @mock.patch("trymerge.iter_issue_timeline_until_comment")
-    def test_get_commit_sha_at_comment_force_push_before_comment_legacy_mode(
-        self, mock_iter_timeline: Any, *args: Any
-    ) -> None:
-        mock_iter_timeline.return_value = [
-            {"event": "committed", "sha": "commit1"},
-            {"event": "committed", "sha": "commit2"},
-            {"event": "head_ref_force_pushed", "after": {"sha": "commit3"}},
-            {"event": "commented", "id": 100},
-        ]
-        pr = GitHubPR("pytorch", "pytorch", 77700)
-        sha = pr.get_commit_sha_at_comment(100)
-        self.assertEqual(sha, "commit3")
-
-    @mock.patch("trymerge.iter_issue_timeline_until_comment")
-    def test_get_commit_sha_at_comment_multiple_comments(
-        self, mock_iter_timeline: Any, *args: Any
-    ) -> None:
-        mock_iter_timeline.return_value = [
-            {"event": "committed", "sha": "commit1"},
-            {"event": "commented", "id": 100},
-            {"event": "committed", "sha": "commit2"},
-            {"event": "commented", "id": 200},
-            {"event": "head_ref_force_pushed", "after": {"sha": "commit3"}},
-            {"event": "commented", "id": 300},
-        ]
-        pr = GitHubPR("pytorch", "pytorch", 77700)
-        sha = pr.get_commit_sha_at_comment(200)
-        self.assertEqual(sha, "commit2")
-        sha = pr.get_commit_sha_at_comment(300)
-        self.assertEqual(sha, "commit3")
-
-    @mock.patch("trymerge.iter_issue_timeline_until_comment")
-    def test_get_commit_sha_at_comment_no_events(
-        self, mock_iter_timeline: Any, *args: Any
-    ) -> None:
-        mock_iter_timeline.return_value = [
-            {"event": "commented", "id": 100},
-            {"event": "labeled", "label": {"name": "test"}},
-        ]
-        pr = GitHubPR("pytorch", "pytorch", 77700)
-        sha = pr.get_commit_sha_at_comment(100)
-        self.assertIsNone(sha)
-
-    @mock.patch("trymerge.iter_issue_timeline_until_comment")
-    def test_get_commit_sha_at_comment_exception(
-        self, mock_iter_timeline: Any, *args: Any
-    ) -> None:
-        mock_iter_timeline.side_effect = Exception("API error")
-        pr = GitHubPR("pytorch", "pytorch", 77700)
-        sha = pr.get_commit_sha_at_comment(100)
-        self.assertIsNone(sha)
-
-
 if __name__ == "__main__":
     main()

.github/scripts/trymerge.py

@@ -450,63 +450,6 @@ HAS_NO_CONNECTED_DIFF_TITLE = (
 IGNORABLE_FAILED_CHECKS_THESHOLD = 10
 
 
-def iter_issue_timeline_until_comment(
-    org: str, repo: str, issue_number: int, target_comment_id: int, max_pages: int = 200
-) -> Any:
-    """
-    Yield timeline entries in order until (and including) the entry whose id == target_comment_id
-    for a 'commented' event. Stops once the target comment is encountered.
-    """
-    page = 1
-
-    while page <= max_pages:
-        url = (
-            f"https://api.github.com/repos/{org}/{repo}/issues/{issue_number}/timeline"
-        )
-        params = {"per_page": 100, "page": page}
-
-        batch = gh_fetch_json_list(url, params)
-
-        if not batch:
-            return
-        for ev in batch:
-            # The target is the issue comment row with event == "commented" and id == issue_comment_id
-            if ev.get("event") == "commented" and ev.get("id") == target_comment_id:
-                yield ev  # nothing in the timeline after this matters, so stop early
-                return
-            yield ev
-        if len(batch) < 100:
-            return
-        page += 1
-
-    # If we got here without finding the comment, then we either hit a bug or some github PR
-    # has a _really_ long timeline.
-    # The max # of pages found on any pytorch/pytorch PR at the time of this change was 41
-    raise RuntimeError(
-        f"Could not find a merge commit in the first {max_pages} pages of the timeline at url {url}."
-        f"This is most likely a bug, please report it to the @pytorch/pytorch-dev-infra team."
-    )
-
-
-def sha_from_committed_event(ev: dict[str, Any]) -> Optional[str]:
-    """Extract SHA from committed event in timeline"""
-    return ev.get("sha")
-
-
-def sha_from_force_push_after(ev: dict[str, Any]) -> Optional[str]:
-    """Extract SHA from force push event in timeline"""
-    # The current GitHub API format
-    commit_id = ev.get("commit_id")
-    if commit_id:
-        return str(commit_id)
-
-    # Legacy format
-    after = ev.get("after") or ev.get("after_commit") or {}
-    if isinstance(after, dict):
-        return after.get("sha") or after.get("oid")
-    return ev.get("after_sha") or ev.get("head_sha")
-
-
 def gh_get_pr_info(org: str, proj: str, pr_no: int) -> Any:
     rc = gh_graphql(GH_GET_PR_INFO_QUERY, name=proj, owner=org, number=pr_no)
     return rc["data"]["repository"]["pullRequest"]
@@ -794,24 +737,16 @@ class GitHubPR:
     def last_commit(self) -> Any:
         return self.info["commits"]["nodes"][-1]["commit"]
 
-    def last_commit_sha(self, default: Optional[str] = None) -> str:
-        # for commits, the oid is the sha
-
-        if default is None:
-            return str(self.last_commit()["oid"])
-
-        return str(self.last_commit().get("oid", default))
-
     def get_merge_base(self) -> str:
         if self.merge_base:
             return self.merge_base
 
-        last_commit_sha = self.last_commit_sha()
+        last_commit_oid = self.last_commit()["oid"]
         # NB: We could use self.base_ref() here for regular PR, however, that doesn't
         # work for ghstack where the base is the custom branch, i.e. gh/USER/ID/base,
         # so let's just use main instead
         self.merge_base = gh_fetch_merge_base(
-            self.org, self.project, last_commit_sha, self.default_branch()
+            self.org, self.project, last_commit_oid, self.default_branch()
         )
 
         # Fallback to baseRefOid if the API call fails, i.e. rate limit. Note that baseRefOid
@@ -900,44 +835,6 @@ class GitHubPR:
     def get_commit_count(self) -> int:
         return int(self.info["commits_with_authors"]["totalCount"])
 
-    def get_commit_sha_at_comment(self, comment_id: int) -> Optional[str]:
-        """
-        Get the PR head commit SHA that was present when a specific comment was posted.
-        This ensures we only merge the state of the PR at the time the merge command was issued,
-        not any subsequent commits that may have been pushed after.
-
-        Returns None if no head-changing events found before the comment or if the comment was not found.
-        """
-        head = None
-
-        try:
-            for event in iter_issue_timeline_until_comment(
-                self.org, self.project, self.pr_num, comment_id
-            ):
-                etype = event.get("event")
-                if etype == "committed":
-                    sha = sha_from_committed_event(event)
-                    if sha:
-                        head = sha
-                        print(f"Timeline: Found commit event for SHA {sha}")
-                elif etype == "head_ref_force_pushed":
-                    sha = sha_from_force_push_after(event)
-                    if sha:
-                        head = sha
-                        print(f"Timeline: Found force push event for SHA {sha}")
-                elif etype == "commented":
-                    if event.get("id") == comment_id:
-                        print(f"Timeline: Found final comment with sha {sha}")
-                        return head
-        except Exception as e:
-            print(
-                f"Warning: Failed to reconstruct timeline for comment {comment_id}: {e}"
-            )
-            return None
-
-        print(f"Did not find comment with id {comment_id} in the PR timeline")
-        return None
-
     def get_pr_creator_login(self) -> str:
         return cast(str, self.info["author"]["login"])
 
@@ -1254,7 +1151,7 @@ class GitHubPR:
         *,
         skip_mandatory_checks: bool = False,
         dry_run: bool = False,
-        comment_id: int,
+        comment_id: Optional[int] = None,
         ignore_current_checks: Optional[list[str]] = None,
     ) -> None:
         # Raises exception if matching rule is not found
@@ -1270,7 +1167,7 @@ class GitHubPR:
             skip_internal_checks=can_skip_internal_checks(self, comment_id),
             ignore_current_checks=ignore_current_checks,
         )
-        additional_merged_prs = self.merge_changes_locally(
+        additional_merged_prs = self.merge_changes(
             repo, skip_mandatory_checks, comment_id
         )
 
@@ -1299,7 +1196,7 @@ class GitHubPR:
                 broken_trunk_checks=ignorable_checks.get("BROKEN_TRUNK", []),
                 flaky_checks=ignorable_checks.get("FLAKY", []),
                 unstable_checks=ignorable_checks.get("UNSTABLE", []),
-                last_commit_sha=self.last_commit_sha(default=""),
+                last_commit_sha=self.last_commit().get("oid", ""),
                 merge_base_sha=self.get_merge_base(),
                 merge_commit_sha=merge_commit_sha,
                 is_failed=False,
@@ -1320,7 +1217,7 @@ class GitHubPR:
             dry_run=dry_run,
         )
 
-    def merge_changes_locally(
+    def merge_changes(
         self,
         repo: GitRepo,
         skip_mandatory_checks: bool = False,
@@ -1329,15 +1226,27 @@ class GitHubPR:
         skip_all_rule_checks: bool = False,
     ) -> list["GitHubPR"]:
         """
-        :param skip_all_rule_checks: If true, skips all rule checks on ghstack PRs, useful for dry-running merge locally
+        :param skip_all_rule_checks: If true, skips all rule checks, useful for dry-running merge locally
         """
         branch_to_merge_into = self.default_branch() if branch is None else branch
         if repo.current_branch() != branch_to_merge_into:
             repo.checkout(branch_to_merge_into)
+        if not self.is_ghstack_pr():
+            msg = self.gen_commit_message()
+            pr_branch_name = f"__pull-request-{self.pr_num}__init__"
+            repo.fetch(self.last_commit()["oid"], pr_branch_name)
+            repo._run_git("merge", "--squash", pr_branch_name)
+            repo._run_git("commit", f'--author="{self.get_author()}"', "-m", msg)
 
-        # It's okay to skip the commit SHA check for ghstack PRs since
-        # authoring requires write access to the repo.
-        if self.is_ghstack_pr():
+            # Did the PR change since we started the merge?
+            pulled_sha = repo.show_ref(pr_branch_name)
+            latest_pr_status = GitHubPR(self.org, self.project, self.pr_num)
+            if pulled_sha != latest_pr_status.last_commit()["oid"]:
+                raise RuntimeError(
+                    "PR has been updated since CI checks last passed. Please rerun the merge command."
+                )
+            return []
+        else:
             return self.merge_ghstack_into(
                 repo,
                 skip_mandatory_checks,
@@ -1345,48 +1254,6 @@ class GitHubPR:
                 skip_all_rule_checks=skip_all_rule_checks,
             )
 
-        msg = self.gen_commit_message()
-        pr_branch_name = f"__pull-request-{self.pr_num}__init__"
-
-        # Determine which commit SHA to merge
-        commit_to_merge = None
-        if not comment_id:
-            raise ValueError("Must provide --comment-id when merging regular PRs")
-
-        # Get the commit SHA that was present when the comment was made
-        commit_to_merge = self.get_commit_sha_at_comment(comment_id)
-        if not commit_to_merge:
-            raise RuntimeError(
-                f"Could not find commit that was pushed before comment {comment_id}"
-            )
-
-        # Validate that this commit is the latest commit on the PR
-        latest_commit = self.last_commit_sha()
-        if commit_to_merge != latest_commit:
-            raise RuntimeError(
-                f"Commit {commit_to_merge} was HEAD when comment {comment_id} was posted "
-                f"but now the latest commit on the PR is {latest_commit}. "
-                f"Please re-issue the merge command to merge the latest commit."
-            )
-
-        print(f"Merging commit {commit_to_merge} locally")
-
-        repo.fetch(commit_to_merge, pr_branch_name)
-        repo._run_git("merge", "--squash", pr_branch_name)
-        repo._run_git("commit", f'--author="{self.get_author()}"', "-m", msg)
-
-        # Did the PR change since we started the merge?
-        pulled_sha = repo.show_ref(pr_branch_name)
-        latest_pr_status = GitHubPR(self.org, self.project, self.pr_num)
-        if (
-            pulled_sha != latest_pr_status.last_commit_sha()
-            or pulled_sha != commit_to_merge
-        ):
-            raise RuntimeError(
-                "PR has been updated since CI checks last passed. Please rerun the merge command."
-            )
-        return []
-
 
 class MergeRuleFailedError(RuntimeError):
     def __init__(self, message: str, rule: Optional["MergeRule"] = None) -> None:
@@ -1591,7 +1458,7 @@ def find_matching_merge_rule(
             pending_checks = []
             failed_checks = []
 
-        hud_link = f"https://hud.pytorch.org/{pr.org}/{pr.project}/commit/{pr.last_commit_sha()}"
+        hud_link = f"https://hud.pytorch.org/{pr.org}/{pr.project}/commit/{pr.last_commit()['oid']}"
         if len(failed_checks) > 0:
             if reject_reason_score < 30000:
                 reject_reason_score = 30000
@@ -2289,14 +2156,14 @@ def categorize_checks(
 def merge(
     pr: GitHubPR,
     repo: GitRepo,
-    comment_id: int,
     dry_run: bool = False,
     skip_mandatory_checks: bool = False,
+    comment_id: Optional[int] = None,
     timeout_minutes: int = 400,
     stale_pr_days: int = 3,
     ignore_current: bool = False,
 ) -> None:
-    initial_commit_sha = pr.last_commit_sha()
+    initial_commit_sha = pr.last_commit()["oid"]
     pr_link = f"https://github.com/{pr.org}/{pr.project}/pull/{pr.pr_num}"
     print(f"Attempting merge of {initial_commit_sha} ({pr_link})")
 
@@ -2367,7 +2234,7 @@ def merge(
             f"Attempting merge of https://github.com/{pr.org}/{pr.project}/pull/{pr.pr_num} ({elapsed_time / 60} minutes elapsed)"
         )
         pr = GitHubPR(pr.org, pr.project, pr.pr_num)
-        if initial_commit_sha != pr.last_commit_sha():
+        if initial_commit_sha != pr.last_commit()["oid"]:
             raise RuntimeError(
                 "New commits were pushed while merging. Please rerun the merge command."
             )
@@ -2534,7 +2401,7 @@ def main() -> None:
     if args.check_mergeability:
         if pr.is_ghstack_pr():
             get_ghstack_prs(repo, pr)  # raises error if out of sync
-        pr.merge_changes_locally(
+        pr.merge_changes(
             repo,
             skip_mandatory_checks=True,
             skip_all_rule_checks=True,
@@ -2549,18 +2416,12 @@ def main() -> None:
         gh_post_pr_comment(org, project, args.pr_num, message, dry_run=args.dry_run)
         return
     try:
-        # Ensure comment id is set, else fail
-        if not args.comment_id:
-            raise ValueError(
-                "Comment ID is required for merging PRs, please provide it using --comment-id"
-            )
-
         merge(
             pr,
             repo,
-            comment_id=args.comment_id,
             dry_run=args.dry_run,
             skip_mandatory_checks=args.force,
+            comment_id=args.comment_id,
             ignore_current=args.ignore_current,
         )
     except Exception as e:
@@ -2582,7 +2443,7 @@ def main() -> None:
                 broken_trunk_checks=[],
                 flaky_checks=[],
                 unstable_checks=[],
-                last_commit_sha=pr.last_commit_sha(default=""),
+                last_commit_sha=pr.last_commit().get("oid", ""),
                 merge_base_sha=pr.get_merge_base(),
                 is_failed=True,
                 skip_mandatory_checks=args.force,

.github/scripts/windows/build_magma.bat

@@ -35,9 +35,6 @@ cd magma
 mkdir build && cd build
 
 set GPU_TARGET=All
-if "%CUVER_NODOT%" == "130" (
-  set CUDA_ARCH_LIST=-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 -gencode arch=compute_100,code=sm_100 -gencode arch=compute_120,code=sm_120
-)
 if "%CUVER_NODOT%" == "129" (
   set CUDA_ARCH_LIST=-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 -gencode arch=compute_100,code=sm_100 -gencode arch=compute_120,code=sm_120
 )

.github/scripts/windows/build_triton.bat

@@ -1,12 +1,18 @@
 @echo on
 
-set DESIRED_PYTHON=%PY_VERS%
-call .ci/pytorch/windows/internal/install_python.bat
-
+set PYTHON_PREFIX=%PY_VERS:.=%
+set PYTHON_PREFIX=py%PYTHON_PREFIX:;=;py%
+call .ci/pytorch/win-test-helpers/installation-helpers/activate_miniconda3.bat
+:: Create a new conda environment
+if "%PY_VERS%" == "3.13t" (
+    call conda create -n %PYTHON_PREFIX% -y -c=conda-forge python-freethreading python=3.13
+) else (
+    call conda create -n %PYTHON_PREFIX% -y -c=conda-forge python=%PY_VERS%
+)
 :: Fix cmake version for issue https://github.com/pytorch/pytorch/issues/150480
-%PYTHON_EXEC% -m pip install wheel pybind11 certifi cython cmake==3.31.6 setuptools==72.1.0 ninja==1.11.1.4
+call conda run -n %PYTHON_PREFIX% pip install wheel pybind11 certifi cython cmake==3.31.6 setuptools==72.1.0 ninja==1.11.1.4
 
 dir "%VC_INSTALL_PATH%"
 
 call "%VC_INSTALL_PATH%\VC\Auxiliary\Build\vcvarsall.bat" x64
-%PYTHON_EXEC% .github/scripts/build_triton_wheel.py --device=%BUILD_DEVICE% %RELEASE%
+call conda run -n %PYTHON_PREFIX% python .github/scripts/build_triton_wheel.py --device=%BUILD_DEVICE% %RELEASE%

.github/templates/common.yml.j2

@@ -4,7 +4,7 @@
 {%- set download_artifact_action = "actions/download-artifact@v4.1.7" -%}
 
 {%- set timeout_minutes = 240 -%}
-{%- set timeout_minutes_windows_binary = 360 -%}
+{%- set timeout_minutes_windows_binary = 300 -%}
 
 {%- macro concurrency(build_environment) -%}
 concurrency:

.github/templates/linux_binary_build_workflow.yml.j2

@@ -114,12 +114,12 @@ jobs:
       ALPINE_IMAGE: "docker.io/s390x/alpine"
       {%- elif config["gpu_arch_type"] == "rocm" %}
       runs_on: linux.rocm.gpu
-      {%- elif config["gpu_arch_type"] == "cuda" and config["gpu_arch_version"] in ["12.6"] %}
+      {%- elif config["gpu_arch_type"] == "cuda" and config["gpu_arch_version"] in ["12.8", "12.9"] %}
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.4xlarge.nvidia.gpu  # 12.6 build can use maxwell (sm_50) runner
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 and 12.9 build need sm_70+ runner
       {%- elif config["gpu_arch_type"] == "cuda" %}
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
+      runs_on: linux.4xlarge.nvidia.gpu # for other cuda versions, we use 4xlarge runner
       {%- else %}
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.4xlarge
@@ -135,7 +135,7 @@ jobs:
       contents: read
     steps:
       - name: Setup XPU
-        uses: pytorch/pytorch/.github/actions/setup-xpu@main
+        uses: ./.github/actions/setup-xpu
       - name: configure aws credentials
         id: aws_creds
         uses: aws-actions/configure-aws-credentials@v4

.github/templates/macos_binary_build_workflow.yml.j2

@@ -68,6 +68,11 @@ jobs:
           chmod +x "${RUNNER_TEMP}/conda.sh"
           /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
           echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
+          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
+            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
+          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
+            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
+          fi
       !{{ common.checkout(deep_clone=False, directory="pytorch") }}
       - name: Populate binary env
         run: |
@@ -105,33 +110,12 @@ jobs:
           # Create new "clean" conda environment for testing
 
           SMOKE_TEST_PARAMS=""
-
-          EXTRA_CONDA_INSTALL_FLAGS=""
-          CONDA_ENV_CREATE_FLAGS=""
-          # shellcheck disable=SC2153
-          case $DESIRED_PYTHON in
-            3.14t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.14)
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
-              desired_python="3.14.0rc1"
-              ;;
-            3.13t)
-              CONDA_ENV_CREATE_FLAGS="python-freethreading"
-              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
-              desired_python="3.13"
-              ;;
-            *)
-              # shellcheck disable=SC2153
-              desired_python=${DESIRED_PYTHON}
-              ;;
-          esac
-
-          # shellcheck disable=SC2086
-          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          if [[ $DESIRED_PYTHON == "3.13t" ]]; then
+            conda create -yn "test_conda_env" python="3.13" python-freethreading -c conda-forge
+            SMOKE_TEST_PARAMS="--torch-compile-check disabled"
+          else
+            conda create -yn "test_conda_env" python="$DESIRED_PYTHON"
+          fi
           conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v