try

ghstack-source-id: b11f5e19e1b622610668017d89feac01750675cf
Pull-Request: https://github.com/pytorch/pytorch/pull/160583

.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

@@ -69,186 +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/"):
@@ -256,8 +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:
@@ -280,7 +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(
@@ -317,20 +205,10 @@ 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
+        build_vars = "MAX_JOBS=5 " + build_vars
 
     override_package_version = os.getenv("OVERRIDE_PACKAGE_VERSION")
     desired_cuda = os.getenv("DESIRED_CUDA")

.ci/aarch64_linux/build_aarch64_wheel.py

@@ -438,7 +438,9 @@ def build_torchvision(
         )
         build_vars += f"BUILD_VERSION={version}.dev{build_date}"
     elif build_version is not None:
-        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
+        build_vars += (
+            f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-')[0]}"
+        )
     if host.using_docker():
         build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
 
@@ -493,7 +495,9 @@ def build_torchdata(
         )
         build_vars += f"BUILD_VERSION={version}.dev{build_date}"
     elif build_version is not None:
-        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
+        build_vars += (
+            f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-')[0]}"
+        )
     if host.using_docker():
         build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
 
@@ -549,7 +553,9 @@ def build_torchtext(
         )
         build_vars += f"BUILD_VERSION={version}.dev{build_date}"
     elif build_version is not None:
-        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
+        build_vars += (
+            f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-')[0]}"
+        )
     if host.using_docker():
         build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
 
@@ -607,7 +613,9 @@ def build_torchaudio(
         )
         build_vars += f"BUILD_VERSION={version}.dev{build_date}"
     elif build_version is not None:
-        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
+        build_vars += (
+            f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-')[0]}"
+        )
     if host.using_docker():
         build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
 

.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

@@ -76,13 +76,10 @@ elif [[ "$image" == *cuda*linter* ]]; then
 elif [[ "$image" == *linter* ]]; then
   # Use a separate Dockerfile for linter to keep a small image size
   DOCKERFILE="linter/Dockerfile"
-elif [[ "$image" == *riscv* ]]; then
-  # Use RISC-V specific Dockerfile
-  DOCKERFILE="ubuntu-cross-riscv/Dockerfile"
 fi
 
-_UCX_COMMIT=7836b165abdbe468a2f607e7254011c07d788152
-_UCC_COMMIT=430e241bf5d38cbc73fc7a6b89155397232e3f96
+_UCX_COMMIT=7bb2722ff2187a0cad557ae4a6afa090569f83fb
+_UCC_COMMIT=20eae37090a4ce1b32bcce6144ccad0b49943e0b
 if [[ "$image" == *rocm* ]]; then
   _UCX_COMMIT=cc312eaa4655c0cc5c2bcd796db938f90563bcf6
   _UCC_COMMIT=0c0fc21559835044ab107199e334f7157d6a0d3d
@@ -114,16 +111,6 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
     ;;
-  pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11)
-    CUDA_VERSION=13.0.0
-    ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=11
-    VISION=yes
-    KATEX=yes
-    UCX_COMMIT=${_UCX_COMMIT}
-    UCC_COMMIT=${_UCC_COMMIT}
-    TRITON=yes
-    ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.8.1
     ANACONDA_PYTHON_VERSION=3.10
@@ -135,6 +122,38 @@ case "$tag" in
     TRITON=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.8.1
+    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
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9)
+    CUDA_VERSION=12.6.3
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-vllm)
     CUDA_VERSION=12.8.1
     ANACONDA_PYTHON_VERSION=3.12
@@ -145,6 +164,39 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
     ;;
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3.12-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6
+    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-cuda12.6-cudnn9-py3.13-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6
+    ANACONDA_PYTHON_VERSION=3.13
+    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)
     CUDA_VERSION=12.8.1
     ANACONDA_PYTHON_VERSION=3.10
@@ -156,18 +208,30 @@ 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
     ;;
-  pytorch-linux-jammy-rocm-n-py3 | pytorch-linux-jammy-rocm-n-py3-benchmarks | pytorch-linux-noble-rocm-n-py3)
+  pytorch-linux-jammy-py3.11-clang12)
+    ANACONDA_PYTHON_VERSION=3.11
+    CLANG_VERSION=12
+    VISION=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-py3.9-gcc9)
+    ANACONDA_PYTHON_VERSION=3.9
+    GCC_VERSION=9
+    VISION=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-rocm-n-py3 | pytorch-linux-noble-rocm-n-py3)
     if [[ $tag =~ "jammy" ]]; then
       ANACONDA_PYTHON_VERSION=3.10
     else
@@ -181,9 +245,7 @@ case "$tag" in
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
-    if [[ $tag =~ "benchmarks" ]]; then
-      INDUCTOR_BENCHMARKS=yes
-    fi
+    INDUCTOR_BENCHMARKS=yes
     ;;
   pytorch-linux-noble-rocm-alpha-py3)
     ANACONDA_PYTHON_VERSION=3.12
@@ -195,26 +257,27 @@ case "$tag" in
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
+    INDUCTOR_BENCHMARKS=yes
     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)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
+    ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
     VISION=yes
     KATEX=yes
@@ -222,8 +285,8 @@ case "$tag" in
     DOCS=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-clang12)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12)
+    ANACONDA_PYTHON_VERSION=3.9
     CUDA_VERSION=12.8.1
     CLANG_VERSION=12
     VISION=yes
@@ -234,8 +297,8 @@ case "$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
@@ -262,10 +325,13 @@ case "$tag" in
     TRITON_CPU=yes
     ;;
   pytorch-linux-jammy-linter)
-    PYTHON_VERSION=3.10
+    # TODO: Use 3.9 here because of this issue https://github.com/python/mypy/issues/13627.
+    # We will need to update mypy version eventually, but that's for another day. The task
+    # would be to upgrade mypy to 1.0.0 with Python 3.11
+    PYTHON_VERSION=3.9
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-linter)
-    PYTHON_VERSION=3.10
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter)
+    PYTHON_VERSION=3.9
     CUDA_VERSION=12.8.1
     ;;
   pytorch-linux-jammy-aarch64-py3.10-gcc11)
@@ -273,6 +339,7 @@ case "$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
@@ -283,15 +350,13 @@ case "$tag" in
     GCC_VERSION=11
     ACL=yes
     VISION=yes
+    CONDA_CMAKE=yes
     OPENBLAS=yes
     # snadampal: skipping llvm src build install because the current version
     # from pytorch/llvm:9.0.1 is x86 specific
     SKIP_LLVM_SRC_BUILD_INSTALL=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-noble-riscv64-py3.12-gcc14)
-    GCC_VERSION=14
-    ;;
   *)
     # Catch-all for builds that are not hardcoded.
     VISION=yes
@@ -412,14 +477,7 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
 fi
 
 if [ -n "$GCC_VERSION" ]; then
-  if [[ "$image" == *riscv* ]]; then
-    # Check RISC-V cross-compilation toolchain version
-    if !(drun riscv64-linux-gnu-gcc-${GCC_VERSION} --version 2>&1 | grep -q " $GCC_VERSION\\W"); then
-      echo "RISC-V GCC_VERSION=$GCC_VERSION, but:"
-      drun riscv64-linux-gnu-gcc-${GCC_VERSION} --version
-      exit 1
-    fi
-  elif !(drun gcc --version 2>&1 | grep -q " $GCC_VERSION\\W"); then
+  if !(drun gcc --version 2>&1 | grep -q " $GCC_VERSION\\W"); then
     echo "GCC_VERSION=$GCC_VERSION, but:"
     drun gcc --version
     exit 1

.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/executorch.txt

@@ -1 +1 @@
-e0dda9059d082537cee36be6c5e4fe3b18c880c0
+56392aa978594cc155fa8af48cd949f5b5f1823a

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

@@ -1,2 +0,0 @@
-transformers==4.56.0
-soxr==0.5.0

.ci/docker/ci_commit_pins/huggingface.txt

@@ -0,0 +1 @@
+243e186efbf7fb93328dd6b34927a4e8c8f24395

.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 +0,0 @@
-74a23feff57432129df84d8099e622773cf77925

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

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

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-bbb06c0334a6772b92d24bde54956e675c8c6604
+f7888497a1eb9e98d4c07537f0d0bcfe180d1363

.ci/docker/common/install_cpython.sh

@@ -66,9 +66,8 @@ function do_cpython_build {
         ln -s pip3 ${prefix}/bin/pip
     fi
     # install setuptools since python 3.12 is required to use distutils
-    # packaging is needed to create symlink since wheel no longer provides needed information
-    ${prefix}/bin/pip install packaging==25.0 wheel==0.45.1 setuptools==80.9.0
-    local abi_tag=$(${prefix}/bin/python -c "from packaging.tags import interpreter_name, interpreter_version; import sysconfig ; from sysconfig import get_config_var; print('{0}{1}-{0}{1}{2}'.format(interpreter_name(), interpreter_version(), 't' if sysconfig.get_config_var('Py_GIL_DISABLED') else ''))")
+    ${prefix}/bin/pip install wheel==0.45.1 setuptools==80.9.0
+    local abi_tag=$(${prefix}/bin/python -c "from wheel.pep425tags import get_abbr_impl, get_impl_ver, get_abi_tag; print('{0}{1}-{2}'.format(get_abbr_impl(), get_impl_ver(), get_abi_tag()))")
     ln -sf ${prefix} /opt/python/${abi_tag}
 }
 
@@ -83,9 +82,9 @@ 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/include/
+  cp -a "libnvshmem/lib/"*     /usr/local/lib/
 
   # 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_executorch.sh

@@ -42,27 +42,22 @@ install_pip_dependencies() {
   # A workaround, ExecuTorch has moved to numpy 2.0 which is not compatible with the current
   # numba and scipy version used in PyTorch CI
   conda_run pip uninstall -y numba scipy
-  # Yaspin is needed for running CI test (get_benchmark_analysis_data.py)
-  pip_install yaspin==3.1.0
 
   popd
 }
 
 setup_executorch() {
+  pushd executorch
+
   export PYTHON_EXECUTABLE=python
-  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON -DEXECUTORCH_BUILD_TESTS=ON"
+  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON"
 
   as_jenkins .ci/scripts/setup-linux.sh --build-tool cmake || true
+  popd
 }
 
-if [ $# -eq 0 ]; then
-  clone_executorch
-  install_buck2
-  install_conda_dependencies
-  install_pip_dependencies
-  pushd executorch
-  setup_executorch
-  popd
-else
-  "$@"
-fi
+clone_executorch
+install_buck2
+install_conda_dependencies
+install_pip_dependencies
+setup_executorch

.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() {
@@ -13,34 +15,11 @@ function install_timm() {
   commit=$(get_pinned_commit timm)
 
   pip_install "git+https://github.com/huggingface/pytorch-image-models@${commit}"
-}
-
-function install_torchbench() {
-  local commit
-  commit=$(get_pinned_commit torchbench)
-  git clone https://github.com/pytorch/benchmark torchbench
-  pushd torchbench
-  git checkout "$commit"
-
-  python install.py --continue_on_fail
-
-  echo "Print all dependencies after TorchBench is installed"
-  python -mpip freeze
-  popd
-
-  chown -R jenkins torchbench
-  chown -R jenkins /opt/conda
+  # Clean up
+  conda_run pip uninstall -y torch torchvision triton
 }
 
 # Pango is needed for weasyprint which is needed for doctr
 conda_install pango
-
-# Stable packages are ok here, just to satisfy TorchBench check
-pip_install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
-
-install_torchbench
 install_huggingface
 install_timm
-
-# Clean up
-conda_run pip uninstall -y torch torchvision torchaudio triton torchao

.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_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

@@ -34,27 +34,18 @@ function install_ubuntu() {
 
     # The xpu-smi packages
     apt-get install -y flex bison xpu-smi
-
-    if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
-        # Compute and Media Runtimes
-        apt-get install -y \
-            intel-opencl-icd intel-level-zero-gpu level-zero \
-            intel-media-va-driver-non-free libmfx1 libmfxgen1 libvpl2 \
-            libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
-            libglapi-mesa libgles2-mesa-dev libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
-            mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo
-        # Development Packages
-        apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev level-zero-dev
-    else # rolling driver
-        apt-get install -y \
-            intel-opencl-icd libze-intel-gpu1 libze1 \
-            intel-media-va-driver-non-free libmfx-gen1 libvpl2 \
-            libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
-            libglapi-mesa libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
-            mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo intel-ocloc
-        apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev libze-dev
+    # Compute and Media Runtimes
+    apt-get install -y \
+        intel-opencl-icd intel-level-zero-gpu level-zero \
+        intel-media-va-driver-non-free libmfx1 libmfxgen1 libvpl2 \
+        libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
+        libglapi-mesa libgles2-mesa-dev libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
+        mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo
+    if [[ "${XPU_DRIVER_TYPE,,}" == "rolling" ]]; then
+        apt-get install -y intel-ocloc
     fi
-
+    # Development Packages
+    apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev level-zero-dev
     # Install Intel Support Packages
     apt-get install -y ${XPU_PACKAGES}
 
@@ -65,14 +56,10 @@ 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)'
@@ -143,18 +130,18 @@ function install_sles() {
 
 }
 
-# Default use GPU driver rolling releases
-XPU_DRIVER_VERSION=""
-if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
-    # Use GPU driver LTS releases
-    XPU_DRIVER_VERSION="/lts/2350"
+# Default use GPU driver LTS releases
+XPU_DRIVER_VERSION="/lts/2350"
+if [[ "${XPU_DRIVER_TYPE,,}" == "rolling" ]]; then
+    # Use GPU driver rolling releases
+    XPU_DRIVER_VERSION=""
 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/build.sh

@@ -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

.ci/docker/requirements-ci.txt

@@ -63,12 +63,11 @@ lark==0.12.0
 #Pinned versions: 0.12.0
 #test that import:
 
-librosa>=0.6.2 ; python_version < "3.11" and platform_machine != "s390x"
-librosa==0.10.2 ; python_version == "3.12" and platform_machine != "s390x"
+librosa>=0.6.2 ; python_version < "3.11"
+librosa==0.10.2 ; python_version == "3.12"
 #Description: A python package for music and audio analysis
 #Pinned versions: >=0.6.2
 #test that import: test_spectral_ops.py
-#librosa depends on numba; disable it for s390x while numba is disabled too
 
 #mkl #this breaks linux-bionic-rocm4.5-py3.7
 #Description: Intel oneAPI Math Kernel Library
@@ -93,9 +92,8 @@ librosa==0.10.2 ; python_version == "3.12" and platform_machine != "s390x"
 #Pinned versions:
 #test that import:
 
-mypy==1.16.0 ; platform_system != "Windows"
+mypy==1.16.0
 # Pin MyPy version because new errors are likely to appear with each release
-# Skip on Windows as lots of type annotations are POSIX specific
 #Description: linter
 #Pinned versions: 1.16.0
 #test that import: test_typing.py, test_type_hints.py
@@ -112,15 +110,14 @@ ninja==1.11.1.3
 #Pinned versions: 1.11.1.3
 #test that import: run_test.py, test_cpp_extensions_aot.py,test_determination.py
 
-numba==0.49.0 ; python_version < "3.9" and platform_machine != "s390x"
-numba==0.55.2 ; python_version == "3.9" and platform_machine != "s390x"
-numba==0.55.2 ; python_version == "3.10" and platform_machine != "s390x"
-numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
+numba==0.49.0 ; python_version < "3.9"
+numba==0.55.2 ; python_version == "3.9"
+numba==0.55.2 ; python_version == "3.10"
+numba==0.60.0 ; python_version == "3.12"
 #Description: Just-In-Time Compiler for Numerical Functions
 #Pinned versions: 0.54.1, 0.49.0, <=0.49.1
 #test that import: test_numba_integration.py
 #For numba issue see https://github.com/pytorch/pytorch/issues/51511
-#Need release > 0.61.2 for s390x due to https://github.com/numba/numba/pull/10073
 
 #numpy
 #Description: Provides N-dimensional arrays and linear algebra
@@ -264,6 +261,11 @@ 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
@@ -305,7 +307,7 @@ pytest-cpp==2.3.0
 #Pinned versions: 2.3.0
 #test that import:
 
-z3-solver==4.15.1.0 ; platform_machine != "s390x"
+z3-solver==4.15.1.0
 #Description: The Z3 Theorem Prover Project
 #Pinned versions:
 #test that import:
@@ -340,7 +342,7 @@ 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:
@@ -359,6 +361,7 @@ pwlf==2.2.1
 #Pinned versions: 2.2.1
 #test that import: test_sac_estimator.py
 
+
 # To build PyTorch itself
 pyyaml
 pyzstd
@@ -380,7 +383,7 @@ 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@d53b0ffb9b1cda68260693ea98f3483823c88d8e#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-cross-riscv/Dockerfile

@@ -1,155 +0,0 @@
-# Cross-compilation Docker container for RISC-V architecture
-ARG UBUNTU_VERSION
-FROM --platform=linux/amd64 ubuntu:${UBUNTU_VERSION} as base
-
-ARG UBUNTU_VERSION
-
-ENV GCC_VERSION=14
-ENV PYTHON_VERSION=3.12.3
-ENV DEBIAN_FRONTEND=noninteractive
-ENV CC=riscv64-linux-gnu-gcc-${GCC_VERSION}
-ENV CXX=riscv64-linux-gnu-g++-${GCC_VERSION}
-ENV QEMU_LD_PREFIX=/usr/riscv64-linux-gnu/
-ENV SYSROOT=/opt/sysroot
-
-# Install basic dependencies
-RUN apt-get update && apt-get install -y \
-    ninja-build \
-    autoconf \
-    automake \
-    libtool \
-    patchelf \
-    ccache \
-    git \
-    wget \
-    python3-pip \
-    python3-venv \
-    python-is-python3 \
-    cmake \
-    sudo \
-    lsb-release \
-    gcc-${GCC_VERSION}-riscv64-linux-gnu \
-    g++-${GCC_VERSION}-riscv64-linux-gnu \
-    pkg-config \
-    && rm -rf /var/lib/apt/lists/*
-
-# Install user
-COPY ./common/install_user.sh install_user.sh
-RUN bash ./install_user.sh && rm install_user.sh
-
-FROM base as python
-ARG ZLIB_VERSION=1.3.1
-ARG FFI_VERSION=3.4.6
-ARG BZ2_VERSION=1.0.8
-ARG XZ_VERSION=5.4.6
-ARG OPENSSL_VERSION=3.2.1
-
-# Set up sysroot directory for dependencies
-ENV PKG_CONFIG_PATH=${SYSROOT}/lib/pkgconfig
-ENV PKG_CONFIG_SYSROOT_DIR=${SYSROOT}
-
-WORKDIR /opt
-
-# Build zlib (for compression)
-RUN echo "--- Building zlib ---" \
-    && wget -c https://www.zlib.net/zlib-${ZLIB_VERSION}.tar.gz \
-    && tar -xf zlib-${ZLIB_VERSION}.tar.gz --no-same-permissions --no-same-owner \
-    && cd zlib-${ZLIB_VERSION}/ \
-    && mkdir build && cd build \
-    && ../configure --prefix=${SYSROOT} \
-    && make -j$(nproc) && make install \
-    && cd ../..
-
-# Build libffi (for ctypes module)
-RUN echo "--- Building libffi ---" \
-    && wget -c https://github.com/libffi/libffi/releases/download/v${FFI_VERSION}/libffi-${FFI_VERSION}.tar.gz \
-    && tar -xf libffi-${FFI_VERSION}.tar.gz --no-same-permissions --no-same-owner \
-    && cd libffi-${FFI_VERSION}/ \
-    && mkdir build && cd build \
-    && ../configure --prefix=${SYSROOT} --host=riscv64-linux-gnu --build=x86_64-linux-gnu \
-    && make -j$(nproc) && make install \
-    && cd ../..
-
-# Build bzip2 (for bz2 module)
-RUN echo "--- Building bzip2 ---" \
-    && wget -c https://sourceware.org/pub/bzip2/bzip2-${BZ2_VERSION}.tar.gz \
-    && tar -xf bzip2-${BZ2_VERSION}.tar.gz --no-same-permissions --no-same-owner \
-    && cd bzip2-${BZ2_VERSION}/ \
-    && make CC=riscv64-linux-gnu-gcc-${GCC_VERSION} bzip2 bzip2recover libbz2.a \
-    && make CC=riscv64-linux-gnu-gcc-${GCC_VERSION} -f Makefile-libbz2_so \
-    && make install PREFIX=${SYSROOT} \
-    && cp libbz2.so.${BZ2_VERSION} ${SYSROOT}/lib/ \
-    && cd ${SYSROOT}/lib/ \
-    && ln -sf libbz2.so.${BZ2_VERSION} libbz2.so.1.0 \
-    && ln -sf libbz2.so.1.0 libbz2.so \
-    && cd /opt/
-
-# Build xz (for lzma module)
-RUN echo "--- Building xz ---" \
-    && wget -c https://github.com/tukaani-project/xz/releases/download/v${XZ_VERSION}/xz-${XZ_VERSION}.tar.gz \
-    && tar -xf xz-${XZ_VERSION}.tar.gz --no-same-permissions --no-same-owner \
-    && cd xz-${XZ_VERSION} \
-    && mkdir build && cd build \
-    && ../configure --prefix=${SYSROOT} --host=riscv64-linux-gnu --build=x86_64-linux-gnu \
-    && make -j$(nproc) && make install \
-    && cd ../..
-
-# Build OpenSSL (for ssl module)
-RUN echo "--- Building OpenSSL ---" \
-    && wget -c https://www.openssl.org/source/openssl-${OPENSSL_VERSION}.tar.gz \
-    && tar -xf openssl-${OPENSSL_VERSION}.tar.gz --no-same-permissions --no-same-owner \
-    && cd openssl-${OPENSSL_VERSION}/ \
-    && mkdir build && cd build \
-    && ../Configure linux64-riscv64 --prefix=${SYSROOT} \
-    && make -j$(nproc) && make install_sw \
-    && cd ../..
-
-# Build SQLite3 (for sqlite3 module)
-RUN echo "--- Building SQLite3 ---" \
-    && wget -c https://www.sqlite.org/2024/sqlite-autoconf-3450200.tar.gz \
-    && tar -xf sqlite-autoconf-3450200.tar.gz --no-same-permissions --no-same-owner \
-    && cd sqlite-autoconf-3450200 \
-    && mkdir build && cd build \
-    && ../configure --prefix=${SYSROOT} --host=riscv64-linux-gnu --build=x86_64-linux-gnu \
-    && make -j$(nproc) && make install \
-    && cd ../..
-
-# Build and install RISC-V Python with all modules
-RUN wget -c https://www.python.org/ftp/python/${PYTHON_VERSION}/Python-${PYTHON_VERSION}.tgz \
-    && tar -xf Python-${PYTHON_VERSION}.tgz --no-same-permissions --no-same-owner \
-    && cd Python-${PYTHON_VERSION} \
-    && mkdir build && cd build \
-    && ../configure \
-        --host=riscv64-linux-gnu \
-        --build=x86_64-linux-gnu \
-        --prefix=${SYSROOT} \
-        --enable-shared \
-        --disable-ipv6 \
-        --with-build-python=/usr/bin/python3 \
-        --with-ensurepip=no \
-        ac_cv_file__dev_ptmx=yes \
-        ac_cv_file__dev_ptc=no \
-    && make -j$(nproc) \
-    && make install
-
-FROM base as final
-COPY --from=python             /opt/sysroot                       /opt/sysroot
-
-# Install crossenv and cmake
-RUN pip install crossenv cmake==4.0.0 --break-system-packages \
-    && /usr/bin/python3 -m crossenv ${SYSROOT}/bin/python3 /opt/riscv-cross-env
-
-# Add pip-installed cmake binaries to PATH
-ENV PATH="/usr/local/bin:${PATH}"
-
-# Set up cross Python environment
-SHELL ["/bin/bash", "-c"]
-RUN source /opt/riscv-cross-env/bin/activate \
-    && pip install setuptools pyyaml typing_extensions wheel
-
-# Set default environment variables for PyTorch build
-ENV Python_ROOT_DIR=${SYSROOT}
-ENV OPENSSL_ROOT_DIR=${SYSROOT}
-
-USER jenkins
-CMD ["bash"]

.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 ${ROCM_VERSION}
 RUN rm install_rocm_magma.sh
@@ -100,11 +96,10 @@ ARG ANACONDA_PYTHON_VERSION
 ENV ANACONDA_PYTHON_VERSION=$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
 
 # (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 $UCC_COMMIT
 ENV UCX_HOME /usr
@@ -97,11 +96,10 @@ RUN rm install_openssl.sh
 ARG INDUCTOR_BENCHMARKS
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
-COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
+COPY ci_commit_pins/huggingface.txt huggingface.txt
 COPY ci_commit_pins/timm.txt timm.txt
-COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${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
 
 ARG TRITON
 ARG TRITON_CPU
@@ -182,6 +180,7 @@ 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="3.10" ${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

@@ -6,6 +6,7 @@ import logging
 from pathlib import Path
 
 from cli.lib.common.path_helper import remove_dir
+from cli.lib.common.utils import run_command
 from git import GitCommandError, RemoteProgress, Repo
 
 
@@ -45,7 +46,7 @@ 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
@@ -55,13 +56,27 @@ 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)
         raise
 
 
+def clone_vllm_pure(commit: str):
+    """
+    cloning vllm and checkout pinned commit
+    """
+    print("clonening vllm....", flush=True)
+    cwd = "vllm"
+    # delete the directory if it exists
+    remove_dir(cwd)
+    # Clone the repo & checkout commit
+    run_command("git clone https://github.com/vllm-project/vllm.git")
+    run_command(f"git checkout {commit}", cwd=cwd)
+    run_command("git submodule update --init --recursive", cwd=cwd)
+
+
 def get_post_build_pinned_commit(name: str, prefix=".github/ci_commit_pins") -> str:
     path = Path(prefix) / f"{name}.txt"
     if not path.exists():

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

@@ -4,7 +4,6 @@ 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
@@ -27,20 +26,36 @@ def pip_install_packages(
         if use_uv
         else [sys.executable, "-m", "pip", "install"]
     )
+
+    if use_uv:
+        logger.info("Installing packages using uv pip")
+
     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)
+    logger.info("Done installing packages")
 
 
 def pip_install_first_match(pattern: str, extras: Optional[str] = None, pref_uv=False):
-    wheel = first_matching_pkg(pattern)
+    """
+    Install the first local whl that matches the given glob pattern.
+
+    Args:
+        pattern (str): Glob pattern for the wheel file(s).
+        extras (str | None): Optional extras (e.g., "opt_einsum") to install with the wheel.
+    """
+    matches = sorted(glob.glob(pattern))
+    if not matches:
+        raise FileNotFoundError(f"No files match: {pattern}")
+    wheel = matches[0]
     target = f"{wheel}[{extras}]" if extras else wheel
-    logger.info("Installing %s...", target)
+    logger.info("Installing wheel: %s", target)
     pip_install_packages([target], prefer_uv=pref_uv)
 
 
@@ -52,20 +67,3 @@ def run_python(args: Union[str, list[str]], env=None):
         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

@@ -8,7 +8,6 @@ import shlex
 import subprocess
 import sys
 from contextlib import contextmanager
-from pathlib import Path
 from typing import Optional
 
 
@@ -116,24 +115,3 @@ def working_directory(path: str):
         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

@@ -0,0 +1,650 @@
+import logging
+import os
+import re
+import subprocess
+import sys
+import textwrap
+from collections.abc import Iterable
+from dataclasses import dataclass
+from enum import Enum
+from pathlib import Path
+from typing import Any, Optional
+
+from cli.lib.common.cli_helper import BaseRunner
+from cli.lib.common.docker_helper import local_image_exists
+from cli.lib.common.envs_helper import (
+    env_bool_field,
+    env_path_field,
+    env_str_field,
+    get_env,
+    with_params_help,
+)
+from cli.lib.common.git_helper import clone_vllm_pure, get_post_build_pinned_commit
+from cli.lib.common.path_helper import (
+    copy,
+    ensure_dir_exists,
+    force_create_dir,
+    get_path,
+    is_path_exist,
+    remove_dir,
+)
+from cli.lib.common.pip_helper import (
+    pip_install_first_match,
+    pip_install_packages,
+    run_python,
+)
+from cli.lib.common.utils import run_command, temp_environ, working_directory
+
+
+logger = logging.getLogger(__name__)
+
+
+# Default path for docker build artifacts
+_DEFAULT_RESULT_PATH = "./shared"
+
+# Temp folder in vllm work place to cp torch whls in vllm work directory for docker build
+_VLLM_TEMP_FOLDER = "tmp"
+
+
+@dataclass
+class VllmBuildParameters:
+    """
+    Parameters defining the vllm external input configurations.
+    Combine with VllmDockerBuildArgs to define the vllm build environment
+    """
+
+    # USE_TORCH_WHEEL: when true, use local Torch wheels; requires TORCH_WHEELS_PATH.
+    #  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")
+
+    # USE_LOCAL_BASE_IMAGE: when true, use an existing local Docker base image; requires BASE_IMAGE
+    # Otherwise, pull dockerfile's default image remotely
+    # BASE_IMAGE: name:tag (only needed when use_local_base_image is True)
+    use_local_base_image: bool = env_bool_field("USE_LOCAL_BASE_IMAGE", True)
+    base_image: str = env_str_field("BASE_IMAGE")
+
+    # USE_LOCAL_DOCKERFILE: when true("1"), use a local Dockerfile; requires DOCKERFILE_PATH.
+    # otherwise, use vllm's default dockerfile.torch_nightly for build
+    # DOCKERFILE_PATH: path to Dockerfile used when use_local_dockerfile is True"
+    use_local_dockerfile: bool = env_bool_field("USE_LOCAL_DOCKERFILE", True)
+    dockerfile_path: Path = env_path_field(
+        "DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile.tmp_vllm"
+    )
+
+    # OUTPUT_DIR: where docker buildx (local exporter) will write artifacts
+    output_dir: Path = env_path_field("OUTPUT_DIR", "shared")
+
+    # --- Build args ----------------------------------------------------------
+    target_stage: str = env_str_field("TARGET_STAGE", "export-wheels")
+
+    tag_name: str = env_str_field("TAG", "vllm-wheels")
+
+    cuda_version: str = env_str_field("CUDA_VERSION", "12.8.1")
+
+    python_version: str = env_str_field("PYTHON_VERSION", "3.12")
+
+    max_jobs: str = env_str_field("MAX_JOBS", "64")
+
+    sccache_bucket: str = env_str_field("SCCACHE_BUCKET")
+
+    sccache_region: str = env_str_field("SCCACHE_REGION")
+
+    torch_cuda_arch_list: str = env_str_field("TORCH_CUDA_ARCH_LIST", "8.9")
+
+    def __post_init__(self):
+        checks = [
+            (
+                self.use_torch_whl,  # flag
+                True,  # trigger_value
+                "torch_whls_path",  # resource
+                is_path_exist,  # check_func
+                "TORCH_WHEELS_PATH is not provided, but USE_TORCH_WHEEL is set to 1",
+            ),
+            (
+                self.use_local_base_image,
+                True,
+                "base_image",
+                local_image_exists,
+                f"BASE_IMAGE {self.base_image} does not found, but USE_LOCAL_BASE_IMAGE is set to 1",
+            ),
+            (
+                self.use_local_dockerfile,
+                True,
+                "dockerfile_path",
+                is_path_exist,
+                " DOCKERFILE_PATH path does not found, but USE_LOCAL_DOCKERFILE is set to 1",
+            ),
+        ]
+        for flag, trigger_value, attr_name, check_func, error_msg in checks:
+            value = getattr(self, attr_name)
+            if flag == trigger_value:
+                if not value or not check_func(value):
+                    raise ValueError(error_msg)
+            else:
+                logger.info("flag  %s is not set", flag)
+        if not self.output_dir:
+            raise ValueError("missing required output_dir")
+
+
+@with_params_help(VllmBuildParameters)
+class VllmBuildRunner(BaseRunner):
+    """
+    Build vLLM using docker buildx.
+
+    Environment variable options:
+        "USE_TORCH_WHEEL":      "1: use local wheels; 0: pull nightly from pypi",
+        "TORCH_WHEELS_PATH":    "Path to local wheels (when USE_TORCH_WHEEL=1)",
+
+        "USE_LOCAL_BASE_IMAGE": "1: use local base image; 0: default image",
+         "BASE_IMAGE":           "name:tag to indicate base image the dockerfile depends on (when USE_LOCAL_BASE_IMAGE=1)",
+
+        "USE_LOCAL_DOCKERFILE": "1: use local Dockerfile; 0: vllm repo default dockerfile.torch_nightly",
+        "DOCKERFILE_PATH":      "Path to Dockerfile (when USE_LOCAL_DOCKERFILE=1)",
+
+        "OUTPUT_DIR":           "e.g. './shared'",
+
+        "TORCH_CUDA_ARCH_LIST": "e.g. '8.0' or '8.0;9.0'",
+        "CUDA_VERSION":         "e.g. '12.8.1'",
+        "PYTHON_VERSION":       "e.g. '3.12'",
+        "MAX_JOBS":             "e.g. '64'",
+        "SCCACHE_BUCKET":       "e.g. 'my-bucket'",
+        "SCCACHE_REGION":       "e.g. 'us-west-2'",
+    """
+
+    def __init__(self, args=None):
+        self.work_directory = "vllm"
+
+    def run(self):
+        """
+        main function to run vllm build
+        1. prepare vllm build environment
+        2. prepare the docker build command args
+        3. run docker build
+        """
+        inputs = VllmBuildParameters()
+        logger.info("Running vllm build with inputs: %s", inputs)
+        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)
+
+        ensure_dir_exists(inputs.output_dir)
+
+        cmd = self._generate_docker_build_cmd(inputs)
+        logger.info("Running docker build: \n %s", cmd)
+        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:
+            return ""
+        tmp_dir = f"./{self.work_directory}/{_VLLM_TEMP_FOLDER}"
+        tmp_path = Path(tmp_dir)
+        force_create_dir(tmp_path)
+        copy(inputs.torch_whls_path, tmp_dir)
+        return tmp_dir
+
+    def cp_dockerfile_if_exist(self, inputs: VllmBuildParameters):
+        if not inputs.use_local_dockerfile:
+            logger.info("using vllm default dockerfile.torch_nightly for build")
+            return
+        dockerfile_path = get_path(inputs.dockerfile_path, resolve=True)
+        vllm_torch_dockerfile = Path(
+            f"./{self.work_directory}/docker/Dockerfile.nightly_torch"
+        )
+        copy(dockerfile_path, vllm_torch_dockerfile)
+
+    def get_result_path(self, path):
+        """
+        Get the absolute path of the result path
+        """
+        if not path:
+            path = _DEFAULT_RESULT_PATH
+        abs_path = get_path(path, resolve=True)
+        return abs_path
+
+    def _get_torch_wheel_path_arg(self, torch_whl_dir: Optional[Path]) -> str:
+        if not torch_whl_dir:
+            return ""
+        return f"--build-arg TORCH_WHEELS_PATH={_VLLM_TEMP_FOLDER}"
+
+    def _get_base_image_args(self, inputs: VllmBuildParameters) -> tuple[str, str, str]:
+        """
+        Returns:
+            - base_image_arg: docker buildx arg string for base image
+            - final_base_image_arg:  docker buildx arg string for vllm-base stage
+            - pull_flag: --pull=true or --pull=false depending on whether the image exists locally
+        """
+        if not inputs.use_local_base_image:
+            return "", "", ""
+
+        base_image = inputs.base_image
+
+        # set both base image and final base image to the same local image
+        base_image_arg = f"--build-arg BUILD_BASE_IMAGE={base_image}"
+        final_base_image_arg = f"--build-arg FINAL_BASE_IMAGE={base_image}"
+
+        if local_image_exists(base_image):
+            pull_flag = "--pull=false"
+            return base_image_arg, final_base_image_arg, pull_flag
+        logger.info(
+            "[INFO] Local image not found:%s will try to pull from remote", {base_image}
+        )
+        return base_image_arg, final_base_image_arg, ""
+
+    def _generate_docker_build_cmd(
+        self,
+        inputs: VllmBuildParameters,
+    ) -> str:
+        base_image_arg, final_base_image_arg, pull_flag = self._get_base_image_args(
+            inputs
+        )
+        torch_arg = self._get_torch_wheel_path_arg(inputs.torch_whls_path)
+
+        return textwrap.dedent(
+            f"""
+            docker buildx build \
+                --output type=local,dest={inputs.output_dir} \
+                -f docker/Dockerfile.nightly_torch \
+                {pull_flag} \
+                {torch_arg} \
+                {base_image_arg} \
+                {final_base_image_arg} \
+                --build-arg max_jobs={inputs.max_jobs} \
+                --build-arg CUDA_VERSION={inputs.cuda_version} \
+                --build-arg PYTHON_VERSION={inputs.python_version} \
+                --build-arg USE_SCCACHE={int(bool(inputs.sccache_bucket and inputs.sccache_region))} \
+                --build-arg SCCACHE_BUCKET_NAME={inputs.sccache_bucket} \
+                --build-arg SCCACHE_REGION_NAME={inputs.sccache_region} \
+                --build-arg torch_cuda_arch_list='{inputs.torch_cuda_arch_list}' \
+                --target {inputs.target_stage} \
+                -t {inputs.tag_name} \
+                --progress=plain .
+        """
+        ).strip()
+
+
+@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("TORCH_WHEELS_PATH", "./dist")
+
+    vllm_whls_path: Path = env_path_field("VLLM_WHEELS_PATH", "./shared")
+
+    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
+
+        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 = [
+            "wheels/xformers/xformers*.whl",
+            "wheels/vllm/vllm*.whl",
+            "wheels/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
+        self.check_versions()
+
+    def run(self):
+        """
+        main function to run vllm test
+        """
+        self.prepare()
+        with working_directory(self.work_directory):
+            if self.test_type == TestInpuType.TEST_PLAN:
+                self.run_test_plan(self.test_plan)
+            else:
+                raise ValueError(f"Unknown test type {self.test_type}")
+
+    def _install_wheels(self, params: VllmTestParameters):
+        logger.info("Running vllm test with inputs: %s", params)
+        logger.info("Installing torch wheel")
+        # torch_p = f"{str(params.torch_whls_path)}/{self.TORCH_WHL_PATH_REGEX}"
+        # pip_install_first_match(torch_p, self.TORCH_WHL_EXTRA)
+
+        logger.info("Installing other torch-related wheels")
+        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):
+        """
+        Install test dependencies for vllm test
+        This method replaces torch dependencies with local torch wheel info in
+        requirements/test.in file from vllm repo.
+
+        Then generates the test.txt file using uv pip compile, along with requirements/test.txt,
+        which is generated by the test.in with torch stable as soft constraint to match
+        packages' version
+
+        """
+        # TODO(elainewy): move this as part of vllm build, to generate the test.txt file
+        logger.info("generate test.txt from requirements/test.in with local torch whls")
+        preprocess_test_in()
+
+        copy(
+            Path("requirements/test.txt"),
+            Path("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"
+        )
+        logger.info("install requirements from test.txt")
+        pip_install_packages(requirements="test.txt", prefer_uv=True)
+        logger.info("Done. install requirements from test.txt")
+
+        # install mambda from source since it does not work now with pip
+        # TODO(elainewy): move this as part of vllm build
+        pip_install_packages(
+            packages=[
+                "--no-build-isolation",
+                "git+https://github.com/state-spaces/mamba@v2.2.4",
+            ],
+            prefer_uv=True,
+        )
+        logger.info("Done. installed requirements from test.txt")
+
+    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 check_versions(self):
+        """
+        check installed packages version
+        """
+        logger.info("double check installed packages")
+        patterns = ["torch", "xformers", "torchvision", "torchaudio", "vllm"]
+        for pkg in patterns:
+            try:
+                module = __import__(pkg)
+                version = getattr(module, "__version__", None)
+                version = version if version else "Unknown version"
+                logger.info("%s: %s", pkg, version)
+            except ImportError:
+                logger.info(" %s: Not installed", pkg)
+        logger.info("Done. checked installed packages")
+
+    def _set_envs(self, inputs: VllmTestParameters):
+        os.environ["TORCH_CUDA_ARCH_LIST"] = inputs.torch_cuda_arch_list
+        if not self.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]"
+            )
+        self.validate_cuda(get_env("TORCH_CUDA_ARCH_LIST"))
+
+        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 run_test_plan(self, test_plan: str):
+        """
+        a method to run list of tests based on the test plan. currently this only
+        used to run vllm tests.
+        """
+        logger.info("run vllm tests.....")
+        tests_map = sample_test_plans()
+        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]
+        logger.info("Running tests: %s", tests["title"])
+
+        pkgs = tests.get("package_install", [])
+        if pkgs:
+            logger.info("Installing packages: %s", pkgs)
+            pip_install_packages(packages=pkgs, prefer_uv=True)
+        with (
+            temp_environ(tests.get("env_var", {})),
+            working_directory(tests.get("working_directory", "tests")),
+        ):
+            failures = []
+            for step in tests["steps"]:
+                with temp_environ(step.get("env_var", {})):
+                    code = run_command(cmd=step["command"], check=False)
+                    if code != 0:
+                        failures.append(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 validate_cuda(self, value: str) -> bool:
+        VALID_VALUES = {"8.0", "8.9", "9.0"}
+        return all(v in VALID_VALUES for v in value.split())
+
+
+def clone_vllm(dst: str = "vllm"):
+    clone_vllm_pure(get_post_build_pinned_commit(dst))
+    """
+    clone_external_repo(
+       target="vllm",
+        repo="https://github.com/vllm-project/vllm.git",
+       dst=dst,
+        update_submodules=True,
+    )
+    """
+
+
+def preprocess_test_in(
+    target_file: str = "requirements/test.in", additional_packages: Iterable[str] = ()
+):
+    """
+    This modifies the target_file file in place. by default, it points to vllm's requirements/test.in
+    It removes torch packages in target_file and replace with local torch whls
+    """
+    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()
+
+    # 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 torch/vision/audio installs 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 = "\n".join(header_lines + [""] + kept_lines) + "\n"
+    target_path.write_text(out)
+    logger.info("[INFO] Updated %s", target_file)
+
+
+def sample_test_plans():
+    """
+    Simple sample to unblock the vllm ci development, which is mimic to
+    https://github.com/vllm-project/vllm/blob/main/.buildkite/test-pipeline.yaml
+    """
+    # TODO(elainewy): Read from yaml file to handle the env and tests for vllm
+    # TODO(elainewy): implement logics to handle package_install
+    return {
+        # test plan:
+        # required id, title, and steps
+        # optional: env_var, package_install, working_directory
+        # by default the working_drectory is "tests/", but it can be changed based on tests, for instance,
+        # vllm sample test happens in samples/
+        "vllm_basic_correctness_test": {
+            "title": "Basic Correctness Test",
+            "id": "vllm_basic_correctness_test",
+            "env_var": {
+                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
+            },
+            # test step:
+            # required: command
+            # available fields: env_var (env_var only set within the scope of the test step), package_install(pip package)
+            "steps": [
+                {
+                    "command": "pytest -v -s basic_correctness/test_cumem.py",
+                },
+                {
+                    "command": "pytest -v -s basic_correctness/test_basic_correctness.py",
+                },
+                {
+                    "command": "pytest -v -s basic_correctness/test_cpu_offload.py",
+                },
+                {
+                    "command": "pytest -v -s basic_correctness/test_preemption.py",
+                    "env_var": {
+                        "VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT": "1",
+                    },
+                },
+            ],
+        },
+        "vllm_basic_models_test": {
+            "title": "Basic models test",
+            "id": "vllm_basic_models_test",
+            "steps": [
+                {"command": "pytest -v -s models/test_transformers.py"},
+                {"command": "pytest -v -s models/test_registry.py"},
+                {"command": "pytest -v -s models/test_utils.py"},
+                {"command": "pytest -v -s models/test_vision.py"},
+                {"command": "pytest -v -s models/test_initialization.py"},
+            ],
+        },
+        "vllm_entrypoints_test": {
+            "title": "Entrypoints Test ",
+            "id": "vllm_entrypoints_test",
+            "env_var": {
+                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
+            },
+            "steps": [
+                {
+                    "command": " ".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",
+                        ]
+                    )
+                },
+                {"command": "pytest -v -s entrypoints/llm/test_lazy_outlines.py"},
+                {"command": "pytest -v -s entrypoints/llm/test_generate.py "},
+                {
+                    "command": "pytest -v -s entrypoints/llm/test_generate_multiple_loras.py"
+                },
+                {
+                    "env_var": {"VLLM_USE_V1": "0"},
+                    "command": "pytest -v -s entrypoints/offline_mode",
+                },
+            ],
+        },
+        "vllm_regression_test": {
+            "title": "Regression Test",
+            "id": "vllm_regression_test",
+            "package_install": ["modelscope"],
+            "steps": [
+                {"command": "pytest -v -s test_regression.py"},
+            ],
+        },
+    }

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

@@ -1,292 +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_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_generate.py",
-                        "--ignore=entrypoints/llm/test_collective_rpc.py",
-                    ]
-                ),
-                "pytest -v -s entrypoints/llm/test_generate.py",
-                "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_build.py

@@ -1,296 +0,0 @@
-import logging
-import os
-import textwrap
-from dataclasses import dataclass
-from pathlib import Path
-from typing import Optional
-
-from cli.lib.common.cli_helper import BaseRunner
-from cli.lib.common.docker_helper import local_image_exists
-from cli.lib.common.envs_helper import (
-    env_bool_field,
-    env_path_field,
-    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.path_helper import (
-    copy,
-    ensure_dir_exists,
-    force_create_dir,
-    get_path,
-    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__)
-
-
-# Default path for docker build artifacts
-_DEFAULT_RESULT_PATH = "./shared"
-
-# Temp folder in vllm work place to cp torch whls in vllm work directory for docker build
-_VLLM_TEMP_FOLDER = "tmp"
-
-
-@dataclass
-class VllmBuildParameters:
-    """
-    Parameters defining the vllm external input configurations.
-    Combine with VllmDockerBuildArgs to define the vllm build environment
-    """
-
-    # USE_TORCH_WHEEL: when true, use local Torch wheels; requires TORCH_WHEELS_PATH.
-    # 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")
-
-    # USE_LOCAL_BASE_IMAGE: when true, use an existing local Docker base image; requires BASE_IMAGE
-    # Otherwise, pull dockerfile's default image remotely
-    # BASE_IMAGE: name:tag (only needed when use_local_base_image is True)
-    use_local_base_image: bool = env_bool_field("USE_LOCAL_BASE_IMAGE", True)
-    base_image: str = env_str_field("BASE_IMAGE")
-
-    # USE_LOCAL_DOCKERFILE: when true("1"), use a local Dockerfile; requires DOCKERFILE_PATH.
-    # otherwise, use vllm's default dockerfile.torch_nightly for build
-    # DOCKERFILE_PATH: path to Dockerfile used when use_local_dockerfile is True"
-    use_local_dockerfile: bool = env_bool_field("USE_LOCAL_DOCKERFILE", True)
-    dockerfile_path: Path = env_path_field(
-        "DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile.tmp_vllm"
-    )
-
-    # the cleaning script to remove torch dependencies from pip
-    cleaning_script: Path = env_path_field(
-        "cleaning_script", ".github/ci_configs/vllm/use_existing_torch.py"
-    )
-
-    # OUTPUT_DIR: where docker buildx (local exporter) will write artifacts
-    output_dir: Path = env_path_field("OUTPUT_DIR", "external/vllm")
-
-    # --- Build args ----------------------------------------------------------
-    target_stage: str = env_str_field("TARGET_STAGE", "export-wheels")
-
-    tag_name: str = env_str_field("TAG", "vllm-wheels")
-
-    cuda_version: str = env_str_field("CUDA_VERSION", "12.8.1")
-
-    python_version: str = env_str_field("PYTHON_VERSION", "3.12")
-
-    max_jobs: str = env_str_field("MAX_JOBS", "64")
-
-    sccache_bucket: str = env_str_field("SCCACHE_BUCKET")
-
-    sccache_region: str = env_str_field("SCCACHE_REGION")
-
-    torch_cuda_arch_list: str = env_str_field("TORCH_CUDA_ARCH_LIST", "8.9")
-
-    def __post_init__(self):
-        checks = [
-            (
-                self.use_torch_whl,  # flag
-                True,  # trigger_value
-                "torch_whls_path",  # resource
-                is_path_exist,  # check_func
-                "TORCH_WHEELS_PATH is not provided, but USE_TORCH_WHEEL is set to 1",
-            ),
-            (
-                self.use_local_base_image,
-                True,
-                "base_image",
-                local_image_exists,
-                f"BASE_IMAGE {self.base_image} does not found, but USE_LOCAL_BASE_IMAGE is set to 1",
-            ),
-            (
-                self.use_local_dockerfile,
-                True,
-                "dockerfile_path",
-                is_path_exist,
-                " DOCKERFILE_PATH path does not found, but USE_LOCAL_DOCKERFILE is set to 1",
-            ),
-        ]
-        for flag, trigger_value, attr_name, check_func, error_msg in checks:
-            value = getattr(self, attr_name)
-            if flag == trigger_value:
-                if not value or not check_func(value):
-                    raise ValueError(error_msg)
-            else:
-                logger.info("flag  %s is not set", flag)
-        if not self.output_dir:
-            raise ValueError("missing required output_dir")
-
-
-@with_params_help(VllmBuildParameters)
-class VllmBuildRunner(BaseRunner):
-    """
-    Build vLLM using docker buildx.
-
-    Environment variable options:
-        "USE_TORCH_WHEEL":      "1: use local wheels; 0: pull nightly from pypi",
-        "TORCH_WHEELS_PATH":    "Path to local wheels (when USE_TORCH_WHEEL=1)",
-
-        "USE_LOCAL_BASE_IMAGE": "1: use local base image; 0: default image",
-         "BASE_IMAGE":           "name:tag to indicate base image the dockerfile depends on (when USE_LOCAL_BASE_IMAGE=1)",
-
-        "USE_LOCAL_DOCKERFILE": "1: use local Dockerfile; 0: vllm repo default dockerfile.torch_nightly",
-        "DOCKERFILE_PATH":      "Path to Dockerfile (when USE_LOCAL_DOCKERFILE=1)",
-
-        "OUTPUT_DIR":           "e.g. './shared'",
-
-        "TORCH_CUDA_ARCH_LIST": "e.g. '8.0' or '8.0;9.0'",
-        "CUDA_VERSION":         "e.g. '12.8.1'",
-        "PYTHON_VERSION":       "e.g. '3.12'",
-        "MAX_JOBS":             "e.g. '64'",
-        "SCCACHE_BUCKET":       "e.g. 'my-bucket'",
-        "SCCACHE_REGION":       "e.g. 'us-west-2'",
-    """
-
-    def __init__(self, args=None):
-        self.work_directory = "vllm"
-
-    def run(self):
-        """
-        main function to run vllm build
-        1. prepare vllm build environment
-        2. prepare the docker build command args
-        3. run docker build
-        """
-        inputs = VllmBuildParameters()
-        logger.info("Running vllm build with inputs: %s", inputs)
-        vllm_commit = clone_vllm()
-
-        self.cp_torch_cleaning_script(inputs)
-        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))
-
-        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")
-
-    def cp_torch_whls_if_exist(self, inputs: VllmBuildParameters) -> str:
-        if not inputs.use_torch_whl:
-            return ""
-        tmp_dir = f"./{self.work_directory}/{_VLLM_TEMP_FOLDER}"
-        tmp_path = Path(tmp_dir)
-        force_create_dir(tmp_path)
-        copy(inputs.torch_whls_path, tmp_dir)
-        return tmp_dir
-
-    def cp_torch_cleaning_script(self, inputs: VllmBuildParameters):
-        script = get_path(inputs.cleaning_script, resolve=True)
-        vllm_script = Path(f"./{self.work_directory}/use_existing_torch.py")
-        copy(script, vllm_script)
-
-    def cp_dockerfile_if_exist(self, inputs: VllmBuildParameters):
-        if not inputs.use_local_dockerfile:
-            logger.info("using vllm default dockerfile.torch_nightly for build")
-            return
-        dockerfile_path = get_path(inputs.dockerfile_path, resolve=True)
-        vllm_torch_dockerfile = Path(
-            f"./{self.work_directory}/docker/Dockerfile.nightly_torch"
-        )
-        copy(dockerfile_path, vllm_torch_dockerfile)
-
-    def get_result_path(self, path):
-        """
-        Get the absolute path of the result path
-        """
-        if not path:
-            path = _DEFAULT_RESULT_PATH
-        abs_path = get_path(path, resolve=True)
-        return abs_path
-
-    def _get_torch_wheel_path_arg(self, torch_whl_dir: Optional[Path]) -> str:
-        if not torch_whl_dir:
-            return ""
-        return f"--build-arg TORCH_WHEELS_PATH={_VLLM_TEMP_FOLDER}"
-
-    def _get_base_image_args(self, inputs: VllmBuildParameters) -> tuple[str, str, str]:
-        """
-        Returns:
-            - base_image_arg: docker buildx arg string for base image
-            - final_base_image_arg:  docker buildx arg string for vllm-base stage
-            - pull_flag: --pull=true or --pull=false depending on whether the image exists locally
-        """
-        if not inputs.use_local_base_image:
-            return "", "", ""
-
-        base_image = inputs.base_image
-
-        # set both base image and final base image to the same local image
-        base_image_arg = f"--build-arg BUILD_BASE_IMAGE={base_image}"
-        final_base_image_arg = f"--build-arg FINAL_BASE_IMAGE={base_image}"
-
-        if local_image_exists(base_image):
-            pull_flag = "--pull=false"
-            return base_image_arg, final_base_image_arg, pull_flag
-        logger.info(
-            "[INFO] Local image not found:%s will try to pull from remote", {base_image}
-        )
-        return base_image_arg, final_base_image_arg, ""
-
-    def _generate_docker_build_cmd(
-        self,
-        inputs: VllmBuildParameters,
-    ) -> str:
-        base_image_arg, final_base_image_arg, pull_flag = self._get_base_image_args(
-            inputs
-        )
-        torch_arg = self._get_torch_wheel_path_arg(inputs.torch_whls_path)
-
-        return textwrap.dedent(
-            f"""
-            docker buildx build \
-                --output type=local,dest={inputs.output_dir} \
-                -f docker/Dockerfile.nightly_torch \
-                {pull_flag} \
-                {torch_arg} \
-                {base_image_arg} \
-                {final_base_image_arg} \
-                --build-arg max_jobs={inputs.max_jobs} \
-                --build-arg CUDA_VERSION={inputs.cuda_version} \
-                --build-arg PYTHON_VERSION={inputs.python_version} \
-                --build-arg USE_SCCACHE={int(bool(inputs.sccache_bucket and inputs.sccache_region))} \
-                --build-arg SCCACHE_BUCKET_NAME={inputs.sccache_bucket} \
-                --build-arg SCCACHE_REGION_NAME={inputs.sccache_region} \
-                --build-arg torch_cuda_arch_list='{inputs.torch_cuda_arch_list}' \
-                --target {inputs.target_stage} \
-                -t {inputs.tag_name} \
-                --progress=plain .
-        """
-        ).strip()

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

@@ -1,280 +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, get_path, 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")
-
-    cleaning_script: Path = env_path_field(
-        "cleaning_script", ".github/ci_configs/vllm/use_existing_torch.py"
-    )
-
-    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)
-        self.cp_torch_cleaning_script(params)
-        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 cp_torch_cleaning_script(self, params: VllmTestParameters):
-        script = get_path(params.cleaning_script, resolve=True)
-        vllm_script = Path(f"./{self.work_directory}/use_existing_torch.py")
-        copy(script, vllm_script)
-
-    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/test_cli/register_test.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_test import VllmTestRunner
+from cli.lib.core.vllm import VllmTestRunner
 
 
 logger = logging.getLogger(__name__)
@@ -12,7 +12,7 @@ logger = logging.getLogger(__name__)
 _TARGETS: dict[str, TargetSpec] = {
     "vllm": {
         "runner": VllmTestRunner,
-        "help": "test vLLM with pytorch main",
+        "help": "test vLLM unittests",
     }
     # add yours ...
 }
@@ -22,18 +22,6 @@ 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",
@@ -41,6 +29,7 @@ def common_args(parser: argparse.ArgumentParser) -> None:
         type=str,
         help="a pre-defined test plan to run, e.g. 'basic_correctness_test'",
     )
+    # TODO(elainewy):add another common option that user can trigger a specific test with test config
 
 
 def register_test_commands(subparsers: argparse._SubParsersAction) -> None:

.ci/lumen_cli/pyproject.toml

@@ -6,7 +6,7 @@ dependencies = [
     "GitPython==3.1.45",
     "docker==7.1.0",
     "pytest==7.3.2",
-    "uv==0.8.6"
+    "uv==0.8.4"
 ]
 
 [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.sh

@@ -5,6 +5,10 @@ set -ex
 SCRIPTPATH="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 
 case "${GPU_ARCH_TYPE:-BLANK}" in
+    BLANK)
+        # Legacy behavior for CircleCI
+        bash "${SCRIPTPATH}/build_cuda.sh"
+        ;;
     cuda)
         bash "${SCRIPTPATH}/build_cuda.sh"
         ;;

.ci/manywheel/build_common.sh

@@ -138,11 +138,28 @@ fi
 
 echo "Calling setup.py bdist at $(date)"
 
-time CMAKE_ARGS=${CMAKE_ARGS[@]} \
-    EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    echo "Calling setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    time EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+    BUILD_LIBTORCH_WHL=1 BUILD_PYTHON_ONLY=0 \
     BUILD_LIBTORCH_CPU_WITH_DEBUG=$BUILD_DEBUG_INFO \
     USE_NCCL=${USE_NCCL} USE_RCCL=${USE_RCCL} USE_KINETO=${USE_KINETO} \
     python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR
+    echo "Finished setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    echo "Calling setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+    time EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+    BUILD_LIBTORCH_WHL=0 BUILD_PYTHON_ONLY=1 \
+    BUILD_LIBTORCH_CPU_WITH_DEBUG=$BUILD_DEBUG_INFO \
+    USE_NCCL=${USE_NCCL} USE_RCCL=${USE_RCCL} USE_KINETO=${USE_KINETO} \
+    CMAKE_FRESH=1 python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR
+    echo "Finished setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+else
+    time CMAKE_ARGS=${CMAKE_ARGS[@]} \
+        EXTRA_CAFFE2_CMAKE_FLAGS=${EXTRA_CAFFE2_CMAKE_FLAGS[@]} \
+        BUILD_LIBTORCH_CPU_WITH_DEBUG=$BUILD_DEBUG_INFO \
+        USE_NCCL=${USE_NCCL} USE_RCCL=${USE_RCCL} USE_KINETO=${USE_KINETO} \
+        python setup.py bdist_wheel -d /tmp/$WHEELHOUSE_DIR
+fi
 echo "Finished setup.py bdist at $(date)"
 
 # Build libtorch packages
@@ -255,6 +272,10 @@ ls /tmp/$WHEELHOUSE_DIR
 mkdir -p "/$WHEELHOUSE_DIR"
 mv /tmp/$WHEELHOUSE_DIR/torch*linux*.whl /$WHEELHOUSE_DIR/
 
+if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    mv /tmp/$WHEELHOUSE_DIR/torch_no_python*.whl /$WHEELHOUSE_DIR/ || true
+fi
+
 if [[ -n "$BUILD_PYTHONLESS" ]]; then
     mkdir -p /$LIBTORCH_HOUSE_DIR
     mv /tmp/$LIBTORCH_HOUSE_DIR/*.zip /$LIBTORCH_HOUSE_DIR
@@ -431,8 +452,16 @@ if [[ -z "$BUILD_PYTHONLESS" ]]; then
   pushd $PYTORCH_ROOT/test
 
   # Install the wheel for this Python version
+  if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    pip uninstall -y "$TORCH_NO_PYTHON_PACKAGE_NAME" || true
+  fi
+
   pip uninstall -y "$TORCH_PACKAGE_NAME"
 
+  if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    pip install "$TORCH_NO_PYTHON_PACKAGE_NAME" --no-index -f /$WHEELHOUSE_DIR --no-dependencies -v
+  fi
+
   pip install "$TORCH_PACKAGE_NAME" --no-index -f /$WHEELHOUSE_DIR --no-dependencies -v
 
   # Print info on the libraries installed in this wheel

.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 +126,15 @@ 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/extras/CUPTI/lib64/libcupti.so.12"
             "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so"
         )
         DEPS_SONAME+=(
@@ -150,83 +146,41 @@ 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

@@ -50,6 +50,9 @@ if [[ ${BUILD_ENVIRONMENT} == *"parallelnative"* ]]; then
   export ATEN_THREADING=NATIVE
 fi
 
+# Enable LLVM dependency for TensorExpr testing
+export USE_LLVM=/opt/llvm
+export LLVM_DIR=/opt/llvm/lib/cmake/llvm
 
 if ! which conda; then
   # In ROCm CIs, we are doing cross compilation on build machines with
@@ -92,27 +95,6 @@ if [[ "$BUILD_ENVIRONMENT" == *aarch64* ]]; then
   export ACL_ROOT_DIR=/ComputeLibrary
 fi
 
-if [[ "$BUILD_ENVIRONMENT" == *riscv64* ]]; then
-  if [[ -f /opt/riscv-cross-env/bin/activate ]]; then
-    # shellcheck disable=SC1091
-    source /opt/riscv-cross-env/bin/activate
-  else
-    echo "Activation file not found"
-    exit 1
-  fi
-
-  export CMAKE_CROSSCOMPILING=TRUE
-  export CMAKE_SYSTEM_NAME=Linux
-  export CMAKE_SYSTEM_PROCESSOR=riscv64
-
-  export USE_CUDA=0
-  export USE_MKLDNN=0
-
-  export SLEEF_TARGET_EXEC_USE_QEMU=ON
-  sudo chown -R jenkins /var/lib/jenkins/workspace /opt
-
-fi
-
 if [[ "$BUILD_ENVIRONMENT" == *libtorch* ]]; then
   POSSIBLE_JAVA_HOMES=()
   POSSIBLE_JAVA_HOMES+=(/usr/local)
@@ -173,7 +155,6 @@ 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 +176,8 @@ 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 ${J}"
+if [[ "$BUILD_ENVIRONMENT" == *cuda* ]] && [[ 1 -eq $(echo "${TORCH_CUDA_ARCH_LIST} >= 8.0" | bc) ]]; then
+  export BUILD_CUSTOM_STEP="ninja -C build flash_attention -j 2"
 fi
 
 if [[ "${BUILD_ENVIRONMENT}" == *clang* ]]; then
@@ -219,6 +192,7 @@ if [[ "$BUILD_ENVIRONMENT" == *-clang*-asan* ]]; then
   export USE_ASAN=1
   export REL_WITH_DEB_INFO=1
   export UBSAN_FLAGS="-fno-sanitize-recover=all"
+  unset USE_LLVM
 fi
 
 if [[ "${BUILD_ENVIRONMENT}" == *no-ops* ]]; then
@@ -239,7 +213,7 @@ fi
 
 # Do not change workspace permissions for ROCm and s390x CI jobs
 # as it can leave workspace with bad permissions for cancelled jobs
-if [[ "$BUILD_ENVIRONMENT" != *rocm* && "$BUILD_ENVIRONMENT" != *s390x* && "$BUILD_ENVIRONMENT" != *riscv64* && -d /var/lib/jenkins/workspace ]]; then
+if [[ "$BUILD_ENVIRONMENT" != *rocm* && "$BUILD_ENVIRONMENT" != *s390x* && -d /var/lib/jenkins/workspace ]]; then
   # Workaround for dind-rootless userid mapping (https://github.com/pytorch/ci-infra/issues/96)
   WORKSPACE_ORIGINAL_OWNER_ID=$(stat -c '%u' "/var/lib/jenkins/workspace")
   cleanup_workspace() {
@@ -284,19 +258,29 @@ else
     # XLA test build fails when WERROR=1
     # set only when building other architectures
     # or building non-XLA tests.
-    if [[ "$BUILD_ENVIRONMENT" != *rocm*  && "$BUILD_ENVIRONMENT" != *xla* && "$BUILD_ENVIRONMENT" != *riscv64* ]]; then
+    if [[ "$BUILD_ENVIRONMENT" != *rocm*  &&
+          "$BUILD_ENVIRONMENT" != *xla* ]]; then
       # Install numpy-2.0.2 for builds which are backward compatible with 1.X
       python -mpip install numpy==2.0.2
 
       WERROR=1 python setup.py clean
 
-      WERROR=1 python setup.py bdist_wheel
+      if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+        python3 tools/packaging/split_wheel.py bdist_wheel
+      else
+        WERROR=1 python setup.py bdist_wheel
+      fi
     else
       python setup.py clean
       if [[ "$BUILD_ENVIRONMENT" == *xla* ]]; then
         source .ci/pytorch/install_cache_xla.sh
       fi
-      python setup.py bdist_wheel
+      if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+        echo "USE_SPLIT_BUILD cannot be used with xla or rocm"
+        exit 1
+      else
+        python setup.py bdist_wheel
+      fi
     fi
     pip_install_whl "$(echo dist/*.whl)"
 
@@ -421,7 +405,7 @@ if [[ "$BUILD_ENVIRONMENT" != *libtorch* && "$BUILD_ENVIRONMENT" != *bazel* ]];
   # don't do this for libtorch as libtorch is C++ only and thus won't have python tests run on its build
   python tools/stats/export_test_times.py
 fi
-# don't do this for bazel or s390x or riscv64 as they don't use sccache
-if [[ "$BUILD_ENVIRONMENT" != *s390x* && "$BUILD_ENVIRONMENT" != *riscv64* && "$BUILD_ENVIRONMENT" != *-bazel-* ]]; then
+# don't do this for bazel or s390x as they don't use sccache
+if [[ "$BUILD_ENVIRONMENT" != *s390x* && "$BUILD_ENVIRONMENT" != *-bazel-* ]]; then
   print_sccache_stats
 fi

.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)
@@ -242,6 +229,7 @@ function install_torchrec_and_fbgemm() {
 
     pip_install tabulate  # needed for newer fbgemm
     pip_install patchelf  # needed for rocm fbgemm
+    pushd /tmp
 
     local wheel_dir=dist/fbgemm_gpu
     local found_whl=0
@@ -257,20 +245,12 @@ function install_torchrec_and_fbgemm() {
     if [ "${found_whl}" == "0" ]; then
       git clone --recursive https://github.com/pytorch/fbgemm
       pushd fbgemm/fbgemm_gpu
-      git checkout "${fbgemm_commit}" --recurse-submodules
-      # until the fbgemm_commit includes the tbb patch
-      patch <<'EOF'
---- a/FbgemmGpu.cmake
-+++ b/FbgemmGpu.cmake
-@@ -184,5 +184,6 @@ gpu_cpp_library(
-     fbgemm_gpu_tbe_cache
-     fbgemm_gpu_tbe_optimizers
-     fbgemm_gpu_tbe_utils
-+    tbb
-   DESTINATION
-     fbgemm_gpu)
-EOF
-      python setup.py bdist_wheel --build-variant=rocm
+      git checkout "${fbgemm_commit}"
+      python setup.py bdist_wheel \
+        --build-variant=rocm \
+        -DHIP_ROOT_DIR="${ROCM_PATH}" \
+        -DCMAKE_C_FLAGS="-DTORCH_USE_HIP_DSA" \
+        -DCMAKE_CXX_FLAGS="-DTORCH_USE_HIP_DSA"
       popd
 
       # Save the wheel before cleaning up
@@ -284,6 +264,7 @@ EOF
     done
 
     rm -rf fbgemm
+    popd
   else
     pip_build_and_install "git+https://github.com/pytorch/torchrec.git@${torchrec_commit}" dist/torchrec
     pip_build_and_install "git+https://github.com/pytorch/FBGEMM.git@${fbgemm_commit}#subdirectory=fbgemm_gpu" dist/fbgemm_gpu
@@ -302,6 +283,30 @@ function clone_pytorch_xla() {
   fi
 }
 
+function checkout_install_torchbench() {
+  local commit
+  commit=$(get_pinned_commit torchbench)
+  git clone https://github.com/pytorch/benchmark torchbench
+  pushd torchbench
+  git checkout "$commit"
+
+  if [ "$1" ]; then
+    python install.py --continue_on_fail models "$@"
+  else
+    # Occasionally the installation may fail on one model but it is ok to continue
+    # to install and test other models
+    python install.py --continue_on_fail
+  fi
+
+  # TODO (huydhn): transformers-4.44.2 added by https://github.com/pytorch/benchmark/pull/2488
+  # is regressing speedup metric. This needs to be investigated further
+  pip install transformers==4.38.1
+
+  echo "Print all dependencies after TorchBench is installed"
+  python -mpip freeze
+  popd
+}
+
 function install_torchao() {
   local commit
   commit=$(get_pinned_commit torchao)

.ci/pytorch/functorch_doc_push_script.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+
+# This is where the local pytorch install in the docker image is located
+pt_checkout="/var/lib/jenkins/workspace"
+source "$pt_checkout/.ci/pytorch/common_utils.sh"
+echo "functorch_doc_push_script.sh: Invoked with $*"
+
+set -ex -o pipefail
+
+version=${DOCS_VERSION:-nightly}
+echo "version: $version"
+
+# Build functorch docs
+pushd $pt_checkout/functorch/docs
+make html
+popd
+
+git clone https://github.com/pytorch/functorch -b gh-pages --depth 1 functorch_ghpages
+pushd functorch_ghpages
+
+if [ "$version" == "main" ]; then
+  version=nightly
+fi
+
+git rm -rf "$version" || true
+mv "$pt_checkout/functorch/docs/build/html" "$version"
+
+git add "$version" || true
+git status
+git config user.email "soumith+bot@pytorch.org"
+git config user.name "pytorchbot"
+# If there aren't changes, don't make a commit; push is no-op
+git commit -m "Generate Python docs from pytorch/pytorch@${GITHUB_SHA}" || true
+git status
+
+if [[ "${WITH_PUSH:-}" == true ]]; then
+  git push -u origin gh-pages
+fi
+
+popd

.ci/pytorch/macos-build.sh

@@ -35,10 +35,11 @@ fi
 
 print_cmake_info
 if [[ ${BUILD_ENVIRONMENT} == *"distributed"* ]]; then
-  USE_OPENMP=1 WERROR=1 python setup.py bdist_wheel
+  # Needed for inductor benchmarks, as lots of HF networks make `torch.distribtued` calls
+  USE_DISTRIBUTED=1 USE_OPENMP=1 WERROR=1 python setup.py bdist_wheel
 else
-  # NB: we always build with distributed; USE_DISTRIBUTED turns off all
-  # backends (specifically the gloo backend), so test that this case works too
+  # Explicitly set USE_DISTRIBUTED=0 to align with the default build config on mac. This also serves as the sole CI config that tests
+  # that building with USE_DISTRIBUTED=0 works at all. See https://github.com/pytorch/pytorch/issues/86448
   USE_DISTRIBUTED=0 USE_OPENMP=1 MACOSX_DEPLOYMENT_TARGET=11.0 WERROR=1 BUILD_TEST=OFF USE_PYTORCH_METAL=1 python setup.py bdist_wheel --plat-name macosx_11_0_arm64
 fi
 if which sccache > /dev/null; then

.ci/pytorch/macos-test.sh

@@ -13,13 +13,9 @@ if [[ ! $(python -c "import torch; print(int(torch.backends.openmp.is_available(
 fi
 popd
 
-python -mpip install -r requirements.txt
-
 # enable debug asserts in serialization
 export TORCH_SERIALIZATION_DEBUG=1
 
-python -mpip install --no-input -r requirements.txt
-
 setup_test_python() {
   # The CircleCI worker hostname doesn't resolve to an address.
   # This environment variable makes ProcessGroupGloo default to
@@ -161,34 +157,6 @@ test_jit_hooks() {
   assert_git_not_dirty
 }
 
-# Shellcheck doesn't like it when you pass no arguments to a function
-# that can take args. See https://www.shellcheck.net/wiki/SC2120
-# shellcheck disable=SC2120
-checkout_install_torchbench() {
-  local commit
-  commit=$(cat .ci/docker/ci_commit_pins/torchbench.txt)
-  git clone https://github.com/pytorch/benchmark torchbench
-  pushd torchbench
-  git checkout "$commit"
-
-  if [ "$1" ]; then
-    python install.py --continue_on_fail models "$@"
-  else
-    # Occasionally the installation may fail on one model but it is ok to continue
-    # 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
-  # https://github.com/pytorch/pytorch/issues/160689 to remove torchao because
-  # its current version 0.12.0 doesn't work with transformers 4.54.0
-  pip uninstall -y torchao
-
-  echo "Print all dependencies after TorchBench is installed"
-  python -mpip freeze
-}
-
 torchbench_setup_macos() {
   git clone --recursive https://github.com/pytorch/vision torchvision
   git clone --recursive https://github.com/pytorch/audio torchaudio
@@ -199,7 +167,7 @@ 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
@@ -208,9 +176,11 @@ 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
 
+  # Shellcheck doesn't like it when you pass no arguments to a function that can take args. See https://www.shellcheck.net/wiki/SC2120
+  # shellcheck disable=SC2119,SC2120
   checkout_install_torchbench
 }
 
@@ -306,47 +276,6 @@ 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"
 }
@@ -395,8 +324,6 @@ 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/numba-cuda-13.patch

@@ -1,25 +0,0 @@
-From 6e08c9d08e9de59c7af28b720289debbbd384764 Mon Sep 17 00:00:00 2001
-From: Michael Wang <13521008+isVoid@users.noreply.github.com>
-Date: Tue, 1 Apr 2025 17:28:05 -0700
-Subject: [PATCH] Avoid bumping certain driver API to avoid future breakage
- (#185)
-
-Co-authored-by: isVoid <isVoid@users.noreply.github.com>
----
- numba_cuda/numba/cuda/cudadrv/driver.py | 3 +++
- 1 file changed, 3 insertions(+)
-
-diff --git a/numba_cuda/numba/cuda/cudadrv/driver.py b/numba_cuda/numba/cuda/cudadrv/driver.py
-index 1641bf77..233e9ed7 100644
---- a/numba_cuda/numba/cuda/cudadrv/driver.py
-+++ b/numba_cuda/numba/cuda/cudadrv/driver.py
-@@ -365,6 +365,9 @@ def _find_api(self, fname):
-         else:
-             variants = ('_v2', '')
- 
-+        if fname in ("cuCtxGetDevice", "cuCtxSynchronize"):
-+            return getattr(self.lib, fname)
-+
-         for variant in variants:
-             try:
-                 return getattr(self.lib, f'{fname}{variant}')

.ci/pytorch/smoke_test/smoke_test.py

@@ -386,8 +386,8 @@ def smoke_test_compile(device: str = "cpu") -> None:
 
 
 def smoke_test_nvshmem() -> None:
-    if not torch.cuda.is_available() or target_os == "windows":
-        print("Windows platform or CUDA is not available, skipping NVSHMEM test")
+    if not torch.cuda.is_available():
+        print("CUDA is not available, skipping NVSHMEM test")
         return
 
     # Check if NVSHMEM is compiled in current build
@@ -396,9 +396,7 @@ def smoke_test_nvshmem() -> None:
     except ImportError:
         # Not built with NVSHMEM support.
         # torch is not compiled with NVSHMEM prior to 2.9
-        from torch.torch_version import TorchVersion
-
-        if TorchVersion(torch.__version__) < (2, 9):
+        if torch.__version__ < "2.9":
             return
         else:
             # After 2.9: NVSHMEM is expected to be compiled in current build

.ci/pytorch/test.sh

@@ -32,16 +32,6 @@ if [[ "$BUILD_ENVIRONMENT" != *rocm* && "$BUILD_ENVIRONMENT" != *s390x* && -d /v
   git config --global --add safe.directory /var/lib/jenkins/workspace
 fi
 
-
-# Patch numba to avoid CUDA-13 crash, see https://github.com/pytorch/pytorch/issues/162878
-NUMBA_CUDA_DIR=$(python -c "import os;import numba.cuda; print(os.path.dirname(numba.cuda.__file__))" 2>/dev/null || true)
-if [ -n "$NUMBA_CUDA_DIR" ]; then
-  NUMBA_PATCH="$(dirname "$(realpath "${BASH_SOURCE[0]}")")/numba-cuda-13.patch"
-  pushd "$NUMBA_CUDA_DIR"
-  patch -p4 <"$NUMBA_PATCH"
-  popd
-fi
-
 echo "Environment variables:"
 env
 
@@ -101,7 +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.
@@ -506,14 +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
@@ -1070,10 +1051,20 @@ test_libtorch_api() {
     mkdir -p $TEST_REPORTS_DIR
 
     OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="${MNIST_DIR}" "$TORCH_BIN_DIR"/test_api --gtest_filter='-IMethodTest.*' --gtest_output=xml:$TEST_REPORTS_DIR/test_api.xml
+    "$TORCH_BIN_DIR"/test_tensorexpr --gtest_output=xml:$TEST_REPORTS_DIR/test_tensorexpr.xml
   else
     # Exclude IMethodTest that relies on torch::deploy, which will instead be ran in test_deploy
     OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="${MNIST_DIR}" python test/run_test.py --cpp --verbose -i cpp/test_api -k "not IMethodTest"
 
+    # On s390x, pytorch is built without llvm.
+    # Even if it would be built with llvm, llvm currently doesn't support used features on s390x and
+    # test fails with errors like:
+    # JIT session error: Unsupported target machine architecture in ELF object pytorch-jitted-objectbuffer
+    # unknown file: Failure
+    # C++ exception with description "valOrErr INTERNAL ASSERT FAILED at "/var/lib/jenkins/workspace/torch/csrc/jit/tensorexpr/llvm_jit.h":34, please report a bug to PyTorch. Unexpected failure in LLVM JIT: Failed to materialize symbols: { (main, { func }) }
+    if [[ "${BUILD_ENVIRONMENT}" != *s390x* ]]; then
+      python test/run_test.py --cpp --verbose -i cpp/test_tensorexpr
+    fi
   fi
 
   # quantization is not fully supported on s390x yet
@@ -1550,10 +1541,14 @@ test_executorch() {
   install_torchvision
   install_torchaudio
 
-  INSTALL_SCRIPT="$(pwd)/.ci/docker/common/install_executorch.sh"
-
   pushd /executorch
-  "${INSTALL_SCRIPT}" setup_executorch
+
+  export PYTHON_EXECUTABLE=python
+  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON"
+
+  # NB: We need to rebuild ExecuTorch runner here because it depends on PyTorch
+  # from the PR
+  bash .ci/scripts/setup-linux.sh --build-tool cmake
 
   echo "Run ExecuTorch unit tests"
   pytest -v -n auto
@@ -1567,6 +1562,10 @@ test_executorch() {
 
   popd
 
+  # Test torchgen generated code for Executorch.
+  echo "Testing ExecuTorch op registration"
+  "$BUILD_BIN_DIR"/test_edge_op_registration
+
   assert_git_not_dirty
 }
 
@@ -1574,7 +1573,6 @@ test_linux_aarch64() {
   python test/run_test.py --include test_modules test_mkldnn test_mkldnn_fusion test_openmp test_torch test_dynamic_shapes \
         test_transformers test_multiprocessing test_numpy_interop test_autograd test_binary_ufuncs test_complex test_spectral_ops \
         test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops \
-        distributed/elastic/timer/api_test distributed/elastic/timer/local_timer_example distributed/elastic/timer/local_timer_test \
         --shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose
 
   # Dynamo tests
@@ -1642,9 +1640,15 @@ elif [[ "${TEST_CONFIG}" == *xla* ]]; then
   build_xla
   test_xla
 elif [[ "$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"
+    if [[ "$BUILD_ENVIRONMENT" == *sm80* ]]; then
+      export TORCH_CUDA_ARCH_LIST="8.0"
+    elif [[ "$BUILD_ENVIRONMENT" == *sm90* ]]; then
+      export TORCH_CUDA_ARCH_LIST="9.0"
+    else
+      export TORCH_CUDA_ARCH_LIST="8.9"
+    fi
+    python -m cli.run test external vllm --test-plan "$TEST_CONFIG"
 elif [[ "${TEST_CONFIG}" == *executorch* ]]; then
   test_executorch
 elif [[ "$TEST_CONFIG" == 'jit_legacy' ]]; then
@@ -1690,40 +1694,54 @@ elif [[ "${TEST_CONFIG}" == *timm* ]]; then
 elif [[ "${TEST_CONFIG}" == cachebench ]]; then
   install_torchaudio
   install_torchvision
-  PYTHONPATH=/torchbench test_cachebench
+  checkout_install_torchbench nanogpt BERT_pytorch resnet50 hf_T5 llama moco
+  PYTHONPATH=$(pwd)/torchbench test_cachebench
 elif [[ "${TEST_CONFIG}" == verify_cachebench ]]; then
   install_torchaudio
   install_torchvision
-  PYTHONPATH=/torchbench test_verify_cachebench
+  checkout_install_torchbench nanogpt
+  PYTHONPATH=$(pwd)/torchbench test_verify_cachebench
 elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
   install_torchaudio
   install_torchvision
+  install_torchao
   id=$((SHARD_NUMBER-1))
   # https://github.com/opencv/opencv-python/issues/885
   pip_install opencv-python==4.8.0.74
   if [[ "${TEST_CONFIG}" == *inductor_torchbench_smoketest_perf* ]]; then
-    PYTHONPATH=/torchbench test_inductor_torchbench_smoketest_perf
+    checkout_install_torchbench hf_Bert hf_Albert timm_vision_transformer
+    PYTHONPATH=$(pwd)/torchbench test_inductor_torchbench_smoketest_perf
   elif [[ "${TEST_CONFIG}" == *inductor_torchbench_cpu_smoketest_perf* ]]; then
-    PYTHONPATH=/torchbench test_inductor_torchbench_cpu_smoketest_perf
+    checkout_install_torchbench timm_vision_transformer phlippe_densenet basic_gnn_edgecnn \
+      llama_v2_7b_16h resnet50 timm_efficientnet mobilenet_v3_large timm_resnest \
+      functorch_maml_omniglot yolov3 mobilenet_v2 resnext50_32x4d densenet121 mnasnet1_0
+    PYTHONPATH=$(pwd)/torchbench test_inductor_torchbench_cpu_smoketest_perf
   elif [[ "${TEST_CONFIG}" == *torchbench_gcp_smoketest* ]]; then
-    TORCHBENCHPATH=/torchbench test_torchbench_gcp_smoketest
+    checkout_install_torchbench
+    TORCHBENCHPATH=$(pwd)/torchbench test_torchbench_gcp_smoketest
   else
+    checkout_install_torchbench
     # Do this after checkout_install_torchbench to ensure we clobber any
     # nightlies that torchbench may pull in
     if [[ "${TEST_CONFIG}" != *cpu* ]]; then
       install_torchrec_and_fbgemm
     fi
-    PYTHONPATH=/torchbench test_dynamo_benchmark torchbench "$id"
+    PYTHONPATH=$(pwd)/torchbench test_dynamo_benchmark torchbench "$id"
   fi
 elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
   install_torchvision
-  PYTHONPATH=/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
+  PYTHONPATH=$(pwd)/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
   if [[ "$SHARD_NUMBER" -eq "1" ]]; then
     test_inductor_aoti
   fi
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision
   test_inductor_shard "${SHARD_NUMBER}"
+  if [[ "${SHARD_NUMBER}" == 1 ]]; then
+    if [[ "${BUILD_ENVIRONMENT}" != linux-jammy-py3.9-gcc11-build ]]; then
+      test_inductor_distributed
+    fi
+  fi
 elif [[ "${TEST_CONFIG}" == *einops* ]]; then
   test_einops
 elif [[ "${TEST_CONFIG}" == *dynamo_wrapped* ]]; then

.ci/pytorch/win-test-helpers/build_pytorch.bat

@@ -61,10 +61,9 @@ if "%USE_XPU%"=="1" (
   call "C:\Program Files (x86)\Intel\oneAPI\compiler\latest\env\vars.bat"
   call "C:\Program Files (x86)\Intel\oneAPI\ocloc\latest\env\vars.bat"
   if errorlevel 1 exit /b 1
-  :: Reduce build time
-  SET TORCH_XPU_ARCH_LIST=bmg
-  :: Re-setup python env for build
-  call pip install -r requirements.txt
+  :: Reduce build time. Only have MTL self-hosted runner now
+  SET TORCH_XPU_ARCH_LIST=xe-lpg
+  SET USE_KINETO=0
 )
 
 @echo on
@@ -137,7 +136,7 @@ sccache --show-stats
 python -c "import os, glob; os.system('python -mpip install --no-index --no-deps ' + glob.glob('dist/*.whl')[0])"
 (
   if "%BUILD_ENVIRONMENT%"=="" (
-    echo NOTE: To run `import torch`, please make sure to activate the conda environment by running `call %CONDA_ROOT_DIR%\Scripts\activate.bat %CONDA_ROOT_DIR%\envs\py_tmp` in Command Prompt before running Git Bash.
+    echo NOTE: To run `import torch`, please make sure to activate the conda environment by running `call %CONDA_PARENT_DIR%\Miniconda3\Scripts\activate.bat %CONDA_PARENT_DIR%\Miniconda3` in Command Prompt before running Git Bash.
   ) else (
     copy /Y "dist\*.whl" "%PYTORCH_FINAL_PACKAGE_DIR%"
 

.ci/pytorch/win-test-helpers/installation-helpers/activate_miniconda3.bat

@@ -3,12 +3,12 @@ if "%BUILD_ENVIRONMENT%"=="" (
 ) else (
   set CONDA_PARENT_DIR=C:\Jenkins
 )
-set CONDA_ROOT_DIR=%CONDA_PARENT_DIR%\Miniconda3
+
 
 :: Be conservative here when rolling out the new AMI with conda. This will try
 :: to install conda as before if it couldn't find the conda installation. This
 :: can be removed eventually after we gain enough confidence in the AMI
-if not exist %CONDA_ROOT_DIR% (
+if not exist %CONDA_PARENT_DIR%\Miniconda3 (
   set INSTALL_FRESH_CONDA=1
 )
 
@@ -17,14 +17,10 @@ if "%INSTALL_FRESH_CONDA%"=="1" (
   if errorlevel 1 exit /b
   if not errorlevel 0 exit /b
 
-  %TMP_DIR_WIN%\Miniconda3-latest-Windows-x86_64.exe /InstallationType=JustMe /RegisterPython=0 /S /AddToPath=0 /D=%CONDA_ROOT_DIR%
+  %TMP_DIR_WIN%\Miniconda3-latest-Windows-x86_64.exe /InstallationType=JustMe /RegisterPython=0 /S /AddToPath=0 /D=%CONDA_PARENT_DIR%\Miniconda3
   if errorlevel 1 exit /b
   if not errorlevel 0 exit /b
 )
 
 :: Activate conda so that we can use its commands, i.e. conda, python, pip
-call %CONDA_ROOT_DIR%\Scripts\activate.bat %CONDA_ROOT_DIR%
-:: Activate conda so that we can use its commands, i.e. conda, python, pip
-call conda activate py_tmp
-
-call pip install -r .ci/docker/requirements-ci.txt
+call %CONDA_PARENT_DIR%\Miniconda3\Scripts\activate.bat %CONDA_PARENT_DIR%\Miniconda3

.ci/pytorch/win-test-helpers/setup_pytorch_env.bat

@@ -14,7 +14,7 @@ if not errorlevel 0 exit /b
 :: build\torch. Rather than changing all these references, making a copy of torch folder
 :: from conda to the current workspace is easier. The workspace will be cleaned up after
 :: the job anyway
-xcopy /s %CONDA_ROOT_DIR%\envs\py_tmp\Lib\site-packages\torch %TMP_DIR_WIN%\build\torch\
+xcopy /s %CONDA_PARENT_DIR%\Miniconda3\Lib\site-packages\torch %TMP_DIR_WIN%\build\torch\
 
 pushd .
 if "%VC_VERSION%" == "" (

.ci/pytorch/win-test.sh

@@ -38,20 +38,13 @@ if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
 fi
 
 # TODO: Move both of them to Windows AMI
-python -m pip install tensorboard==2.13.0 protobuf==5.29.4 pytest-subtests==0.13.1
-
-# Copied from https://github.com/pytorch/test-infra/blob/be01a40157c36cd5a48391fdf44a7bc3ebd4c7e3/aws/ami/windows/scripts/Installers/Install-Pip-Dependencies.ps1#L16 with some adjustments
-# pytest-rerunfailures==10.3 as 10.2 fails with INTERNALERROR> pluggy._manager.PluginValidationError: unknown hook 'pytest_configure_node'
-# scipy from 1.6.3 to 1.10
-# expecttest from 0.1.3 to 0.3.0
-# xdoctest from 1.0.2 to 1.3.0
-python -m pip install "future==0.18.2" "hypothesis==5.35.1" "expecttest==0.3.0" "librosa>=0.6.2" "scipy==1.10.1" "psutil==5.9.1" "pynvml==11.4.1" "pillow==9.2.0" "unittest-xml-reporting<=3.2.0,>=2.0.0" "pytest==7.1.3" "pytest-xdist==2.5.0" "pytest-flakefinder==1.1.0" "pytest-rerunfailures==10.3" "pytest-shard==0.1.2" "sympy==1.11.1" "xdoctest==1.3.0" "pygments==2.12.0" "opt-einsum>=3.3" "networkx==2.8.8" "mpmath==1.2.1" "pytest-cpp==2.3.0" "boto3==1.35.42"
+python -m pip install pytest-rerunfailures==10.3 pytest-cpp==2.3.0 tensorboard==2.13.0 protobuf==5.29.4 pytest-subtests==0.13.1
 
 # Install Z3 optional dependency for Windows builds.
 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
@@ -59,6 +52,9 @@ python -m pip install parameterized==0.8.1
 # Install pulp for testing ilps under torch\distributed\_tools
 python -m pip install pulp==2.9.0
 
+# Install expecttest to merge https://github.com/pytorch/pytorch/pull/155308
+python -m pip install expecttest==0.3.0
+
 run_tests() {
     # Run nvidia-smi if available
     for path in '/c/Program Files/NVIDIA Corporation/NVSMI/nvidia-smi.exe' /c/Windows/System32/nvidia-smi.exe; do

.ci/pytorch/windows/cuda126.bat

@@ -37,7 +37,7 @@ IF "%CUDA_PATH_V126%"=="" (
 )
 
 IF "%BUILD_VISION%" == "" (
-    set TORCH_CUDA_ARCH_LIST=5.0;6.0;6.1;7.0;7.5;8.0;8.6;9.0
+    set TORCH_CUDA_ARCH_LIST=6.1;7.0;7.5;8.0;8.6;9.0
     set TORCH_NVCC_FLAGS=-Xfatbin -compress-all
 ) ELSE (
     set NVCC_FLAGS=-D__CUDA_NO_HALF_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_90,code=compute_90

.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

@@ -85,7 +85,7 @@ mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR" || true
 # Create an isolated directory to store this builds pytorch checkout and conda
 # installation
 if [[ -z "$MAC_PACKAGE_WORK_DIR" ]]; then
-    MAC_PACKAGE_WORK_DIR="$(pwd)/tmp_wheel_${DESIRED_PYTHON}_$(date +%H%M%S)"
+    MAC_PACKAGE_WORK_DIR="$(pwd)/tmp_wheel_conda_${DESIRED_PYTHON}_$(date +%H%M%S)"
 fi
 mkdir -p "$MAC_PACKAGE_WORK_DIR" || true
 if [[ -n ${GITHUB_ACTIONS} ]]; then
@@ -96,11 +96,11 @@ fi
 whl_tmp_dir="${MAC_PACKAGE_WORK_DIR}/dist"
 mkdir -p "$whl_tmp_dir"
 
-mac_version='macosx-11_0-arm64'
+mac_version='macosx_11_0_arm64'
 libtorch_arch='arm64'
 
 # Create a consistent wheel package name to rename the wheel to
-wheel_filename_new="${TORCH_PACKAGE_NAME}-${build_version}${build_number_prefix}-cp${python_nodot}-none-${mac_version//[-,]/_}.whl"
+wheel_filename_new="${TORCH_PACKAGE_NAME}-${build_version}${build_number_prefix}-cp${python_nodot}-none-${mac_version}.whl"
 
 ###########################################################
 
@@ -124,63 +124,77 @@ 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"
-        mac_version='macosx-11.0-arm64'
-        NUMPY_PINNED_VERSION="==2.1.0"
-        RENAME_WHEEL=false
-        ;;
-    3.14)
-        echo "Using 3.14t deps"
-        mac_version='macosx-11.0-arm64'
-        NUMPY_PINNED_VERSION="==2.1.0"
-        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"
         RENAME_WHEEL=false
         ;;
     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
 
-PINNED_PACKAGES=(
-    "numpy${NUMPY_PINNED_VERSION}"
-)
-python -mvenv ~/${desired_python}-build
-source ~/${desired_python}-build/bin/activate
-retry pip install "${PINNED_PACKAGES[@]}" -r "${pytorch_rootdir}/requirements.txt"
+# Install into a fresh env
+tmp_env_name="wheel_py$python_nodot"
+conda create ${EXTRA_CONDA_INSTALL_FLAGS} -yn "$tmp_env_name" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS}
+source activate "$tmp_env_name"
+
+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
 
-# For USE_DISTRIBUTED=1 on macOS, this enables gloo, which needs libuv, which
-# is build as part of tensorpipe submodule
+# For USE_DISTRIBUTED=1 on macOS, need libuv, which is build as part of tensorpipe submodule
 export USE_DISTRIBUTED=1
 
+if [[ -n "$CROSS_COMPILE_ARM64" ]]; then
+    export CMAKE_OSX_ARCHITECTURES=arm64
+fi
 export USE_MKLDNN=OFF
 export USE_QNNPACK=OFF
 export BUILD_TEST=OFF
@@ -188,7 +202,16 @@ export BUILD_TEST=OFF
 pushd "$pytorch_rootdir"
 echo "Calling setup.py bdist_wheel at $(date)"
 
-_PYTHON_HOST_PLATFORM=${mac_version} ARCHFLAGS="-arch arm64" python setup.py bdist_wheel -d "$whl_tmp_dir" --plat-name "${mac_version//[-.]/_}"
+if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    echo "Calling setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    BUILD_LIBTORCH_WHL=1 BUILD_PYTHON_ONLY=0 python setup.py bdist_wheel -d "$whl_tmp_dir"
+    echo "Finished setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    echo "Calling setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+    BUILD_LIBTORCH_WHL=0 BUILD_PYTHON_ONLY=1 CMAKE_FRESH=1 python setup.py bdist_wheel -d "$whl_tmp_dir"
+    echo "Finished setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+else
+    python setup.py bdist_wheel -d "$whl_tmp_dir"
+fi
 
 echo "Finished setup.py bdist_wheel at $(date)"
 

.circleci/scripts/binary_linux_test.sh

@@ -65,8 +65,16 @@ fi
 
 if [[ "$PACKAGE_TYPE" != libtorch ]]; then
   if [[ "\$BUILD_ENVIRONMENT" != *s390x* ]]; then
-    pip install "\$pkg" --index-url "https://download.pytorch.org/whl/\${CHANNEL}/${DESIRED_CUDA}"
-    retry pip install -q numpy protobuf typing-extensions
+    if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+      pkg_no_python="$(ls -1 /final_pkgs/torch_no_python* | sort |tail -1)"
+      pkg_torch="$(ls -1 /final_pkgs/torch-* | sort |tail -1)"
+      # todo: after folder is populated use the pypi_pkg channel instead
+      pip install "\$pkg_no_python" "\$pkg_torch" --index-url "https://download.pytorch.org/whl/\${CHANNEL}/${DESIRED_CUDA}_pypi_pkg"
+      retry pip install -q numpy protobuf typing-extensions
+    else
+      pip install "\$pkg" --index-url "https://download.pytorch.org/whl/\${CHANNEL}/${DESIRED_CUDA}"
+      retry pip install -q numpy protobuf typing-extensions
+    fi
   else
     pip install "\$pkg"
     retry pip install -q numpy protobuf typing-extensions

.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
 
@@ -134,6 +134,7 @@ export DESIRED_PYTHON="${DESIRED_PYTHON:-}"
 export DESIRED_CUDA="$DESIRED_CUDA"
 export LIBTORCH_VARIANT="${LIBTORCH_VARIANT:-}"
 export BUILD_PYTHONLESS="${BUILD_PYTHONLESS:-}"
+export USE_SPLIT_BUILD="${USE_SPLIT_BUILD:-}"
 if [[ "${OSTYPE}" == "msys" ]]; then
   export LIBTORCH_CONFIG="${LIBTORCH_CONFIG:-}"
   if [[ "${LIBTORCH_CONFIG:-}" == 'debug' ]]; then

.circleci/scripts/binary_upload.sh

@@ -23,6 +23,10 @@ if [[ "${DRY_RUN}" = "disabled" ]]; then
   AWS_S3_CP="aws s3 cp"
 fi
 
+if [[ "${USE_SPLIT_BUILD:-false}" == "true" ]]; then
+  UPLOAD_SUBFOLDER="${UPLOAD_SUBFOLDER}_pypi_pkg"
+fi
+
 # this is special build with all dependencies packaged
 if [[ ${BUILD_NAME} == *-full* ]]; then
   UPLOAD_SUBFOLDER="${UPLOAD_SUBFOLDER}_full"
@@ -51,12 +55,16 @@ 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
@@ -73,7 +72,7 @@ exclude =
     ./docs/src,
     ./functorch/docs,
     ./functorch/examples,
-    ./functorch/docs/source/tutorials,
+    ./functorch/notebooks,
     ./scripts,
     ./test/generated_type_hints_smoketest.py,
     ./third_party,

.github/actionlint.yaml

@@ -12,16 +12,13 @@ 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
     - linux.arm64.2xlarge.ephemeral
     - linux.arm64.m7g.4xlarge
     - linux.arm64.m7g.4xlarge.ephemeral
-    - linux.arm64.r7g.12xlarge.memory
     - linux.4xlarge.nvidia.gpu
     - linux.8xlarge.nvidia.gpu
     - linux.16xlarge.nvidia.gpu
@@ -57,7 +54,6 @@ self-hosted-runner:
     - linux.rocm.gpu.2
     - linux.rocm.gpu.4
     # gfx942 runners
-    - linux.rocm.gpu.gfx942.1
     - linux.rocm.gpu.gfx942.2
     - linux.rocm.gpu.gfx942.4
     - rocm-docker

.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/actions/reuse-old-whl/reuse_old_whl.py

@@ -264,7 +264,7 @@ def unzip_artifact_and_replace_files() -> None:
         change_content_to_new_version(f"artifacts/dist/{old_stem}/torch/version.py")
 
         for file in Path(f"artifacts/dist/{old_stem}").glob(
-            "*.dist-info/*",
+            "*.dist-info/**",
         ):
             change_content_to_new_version(file)
 

.github/actions/setup-rocm/action.yml

@@ -59,6 +59,11 @@ runs:
             echo "$msg"
             exit 1
         fi
+        if [[ $ngpu -eq 1 ]]; then
+            echo "Error: only 1 GPU detected, at least 2 GPUs are needed for distributed jobs"
+            echo "$msg"
+            exit 1
+        fi
 
     - name: Runner diskspace health check
       uses: pytorch/pytorch/.github/actions/diskspace-cleanup@main

.github/actions/setup-win/action.yml

@@ -6,12 +6,6 @@ inputs:
   cuda-version:
     description: which cuda version to install, 'cpu' for none
     required: true
-  python-version:
-    required: false
-    type: string
-    default: "3.10"
-    description: |
-      The python version to be used. Will be 3.10 by default
 
 runs:
   using: composite
@@ -44,24 +38,18 @@ runs:
         CONDA="C:\Jenkins\Miniconda3\condabin\conda.bat"
 
         {
-          echo "CONDA=${CONDA}";
           echo "CONDA_RUN=${CONDA} run --no-capture-output";
           echo "CONDA_BUILD=${CONDA} run conda-build";
           echo "CONDA_INSTALL=${CONDA} install";
         } >> "${GITHUB_ENV}"
 
     - name: Setup Python3
-      env:
-          PYTHON_VERSION: ${{ inputs.python-version }}
       shell: bash
       run: |
         set +e
         set -x
 
-        # Create new py_tmp env with python-version
-        ${CONDA} create -y -n py_tmp python=${PYTHON_VERSION} intel-openmp
-
-        PYTHON3=$(${CONDA_RUN} -n py_tmp which python3)
+        PYTHON3=$(${CONDA_RUN} which python3)
         EXIT_CODE=$?
 
         if [[ "${EXIT_CODE}" == "0" ]]; then
@@ -74,7 +62,7 @@ runs:
           # installation, which is Python 3 based. Its Python is default to Python 3. Further, there
           # is also the Miniconda installation that is Python 2 based, and both can be installed if
           # needed. In both cases, Python binary is just called python
-          PYTHON=$(${CONDA_RUN} -n py_tmp which python)
+          PYTHON=$(${CONDA_RUN} which python)
           EXIT_CODE=$?
 
           if [[ "${EXIT_CODE}" == "0" ]]; then

.github/actions/test-pytorch-binary/action.yml

@@ -24,6 +24,7 @@ runs:
           -e PYTORCH_FINAL_PACKAGE_DIR \
           -e PYTORCH_ROOT \
           -e SKIP_ALL_TESTS \
+          -e USE_SPLIT_BUILD \
           --tty \
           --detach \
           -v "${GITHUB_WORKSPACE}/pytorch:/pytorch" \

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-87ff22e49ed0e92576c4935ccb8c143daac4a3cd
+9b57c7bd5ad4db093c5bb31c802df9f04d933ac9

.github/ci_commit_pins/fbgemm_rocm.txt

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

.github/ci_commit_pins/torchbench.txt

@@ -0,0 +1 @@
+e03a63be43e33596f7f0a43b0f530353785e4a59