CUDA aarch64 12.6 and 12.8 builds fix triton constraints (#165022)

CUDA aarch64 12.6 and 12.8 builds fix triton constraints (#165013)

Since we have introduced CUDA aarch64 builds for all cuda versions we need to remove this constraint.
This was missed by https://github.com/pytorch/pytorch/pull/162364

Proper constraint on triton should be:
```
Requires-Dist: triton==3.5.0; platform_system == "Linux"
```

not:
```
Requires-Dist: triton==3.5.0; platform_system == "Linux" and platform_machine == "x86_64"
```

Pull Request resolved: https://github.com/pytorch/pytorch/pull/165013
Approved by: https://github.com/Camyll, https://github.com/nWEIdia, https://github.com/tinglvv

(cherry picked from commit 81dbeb06f4b3eb6c56625ec25d377eb7c7c6c573)

Co-authored-by: atalman <atalman@fb.com>

.bc-linter.yml

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

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -3,8 +3,20 @@ set -eux -o pipefail
 
 GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}
 
-if [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
+# Set CUDA architecture lists to match x86 build_cuda.sh
+if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
+    export TORCH_CUDA_ARCH_LIST="8.0;9.0"
+elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
+elif [[ "$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"
+    # Bundle ptxas into the cu13 wheel, see https://github.com/pytorch/pytorch/issues/163801
+    export BUILD_BUNDLE_PTXAS=1
 fi
 
 SCRIPTPATH="$( cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P )"
@@ -18,7 +30,7 @@ cd /
 # on the mounted pytorch repo
 git config --global --add safe.directory /pytorch
 pip install -r /pytorch/requirements.txt
-pip install auditwheel==6.2.0
+pip install auditwheel==6.2.0 wheel
 if [ "$DESIRED_CUDA" = "cpu" ]; then
     echo "BASE_CUDA_VERSION is not set. Building cpu wheel."
     #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
@@ -26,6 +38,16 @@ if [ "$DESIRED_CUDA" = "cpu" ]; then
 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
+
     #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

@@ -69,61 +69,186 @@ 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")
-    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",
-    ]
+    # Delete original wheel since it will be repackaged
+    os.system(f"rm {wheel_path}")
 
-    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",
+    # 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 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}"
-        )
+        # 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",
+            "/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)
 
     # Make sure the wheel is tagged with manylinux_2_28
     for f in os.scandir(f"{folder}/tmp/"):
@@ -131,14 +256,8 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
             replace_tag(f"{f.path}/WHEEL")
             break
 
-    os.mkdir(f"{folder}/cuda_wheel")
-    os.system(f"cd {folder}/tmp/; zip -r {folder}/cuda_wheel/{wheelname} *")
-    shutil.move(
-        f"{folder}/cuda_wheel/{wheelname}",
-        f"{folder}/{wheelname}",
-        copy_function=shutil.copy2,
-    )
-    os.system(f"rm -rf {folder}/tmp/ {folder}/cuda_wheel/")
+    os.system(f"wheel pack {folder}/tmp/ -d {folder}")
+    os.system(f"rm -rf {folder}/tmp/")
 
 
 def complete_wheel(folder: str) -> str:
@@ -161,14 +280,7 @@ def complete_wheel(folder: str) -> str:
             f"/{folder}/dist/{repaired_wheel_name}",
         )
     else:
-        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}",
-        )
+        repaired_wheel_name = list_dir(f"/{folder}/dist")[0]
 
     print(f"Copying {repaired_wheel_name} to artifacts")
     shutil.copy2(
@@ -208,7 +320,17 @@ if __name__ == "__main__":
     build_vars = "CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000 "
     # MAX_JOB=5 is not required for CPU backend (see commit 465d98b)
     if enable_cuda:
-        build_vars = "MAX_JOBS=5 " + build_vars
+        build_vars += "MAX_JOBS=5 "
+
+        # 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
 
     override_package_version = os.getenv("OVERRIDE_PACKAGE_VERSION")
     desired_cuda = os.getenv("DESIRED_CUDA")

.ci/aarch64_linux/build_aarch64_wheel.py

@@ -438,9 +438,7 @@ 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('-')[0]}"
-        )
+        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
     if host.using_docker():
         build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
 
@@ -495,9 +493,7 @@ 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('-')[0]}"
-        )
+        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
     if host.using_docker():
         build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
 
@@ -553,9 +549,7 @@ 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('-')[0]}"
-        )
+        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[0]}"
     if host.using_docker():
         build_vars += " CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000"
 
@@ -613,9 +607,7 @@ 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('-')[0]}"
-        )
+        build_vars += f"BUILD_VERSION={build_version} PYTORCH_VERSION={branch[1:].split('-', maxsplit=1)[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

@@ -64,6 +64,10 @@ 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
@@ -76,10 +80,10 @@ 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

@@ -76,10 +76,13 @@ 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=7bb2722ff2187a0cad557ae4a6afa090569f83fb
-_UCC_COMMIT=20eae37090a4ce1b32bcce6144ccad0b49943e0b
+_UCX_COMMIT=7836b165abdbe468a2f607e7254011c07d788152
+_UCC_COMMIT=430e241bf5d38cbc73fc7a6b89155397232e3f96
 if [[ "$image" == *rocm* ]]; then
   _UCX_COMMIT=cc312eaa4655c0cc5c2bcd796db938f90563bcf6
   _UCC_COMMIT=0c0fc21559835044ab107199e334f7157d6a0d3d
@@ -111,6 +114,16 @@ case "$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=${_UCC_COMMIT}
+    TRITON=yes
+    ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.8.1
     ANACONDA_PYTHON_VERSION=3.10
@@ -122,38 +135,6 @@ case "$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
@@ -164,39 +145,6 @@ 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
@@ -208,30 +156,18 @@ case "$tag" in
     TRITON=yes
     ;;
   pytorch-linux-jammy-py3-clang12-onnx)
-    ANACONDA_PYTHON_VERSION=3.9
+    ANACONDA_PYTHON_VERSION=3.10
     CLANG_VERSION=12
     VISION=yes
     ONNX=yes
     ;;
-  pytorch-linux-jammy-py3.9-clang12)
-    ANACONDA_PYTHON_VERSION=3.9
+  pytorch-linux-jammy-py3.10-clang12)
+    ANACONDA_PYTHON_VERSION=3.10
     CLANG_VERSION=12
     VISION=yes
     TRITON=yes
     ;;
-  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)
+  pytorch-linux-jammy-rocm-n-py3 | pytorch-linux-jammy-rocm-n-py3-benchmarks | pytorch-linux-noble-rocm-n-py3)
     if [[ $tag =~ "jammy" ]]; then
       ANACONDA_PYTHON_VERSION=3.10
     else
@@ -245,7 +181,9 @@ case "$tag" in
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
-    INDUCTOR_BENCHMARKS=yes
+    if [[ $tag =~ "benchmarks" ]]; then
+      INDUCTOR_BENCHMARKS=yes
+    fi
     ;;
   pytorch-linux-noble-rocm-alpha-py3)
     ANACONDA_PYTHON_VERSION=3.12
@@ -257,27 +195,26 @@ case "$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-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
+  pytorch-linux-jammy-xpu-n-1-py3)
+    ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
     VISION=yes
     XPU_VERSION=2025.1
     NINJA_VERSION=1.9.0
     TRITON=yes
     ;;
-  pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
-    ANACONDA_PYTHON_VERSION=3.9
+  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
     GCC_VERSION=11
     VISION=yes
     KATEX=yes
@@ -285,8 +222,8 @@ case "$tag" in
     DOCS=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12)
-    ANACONDA_PYTHON_VERSION=3.9
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-clang12)
+    ANACONDA_PYTHON_VERSION=3.10
     CUDA_VERSION=12.8.1
     CLANG_VERSION=12
     VISION=yes
@@ -297,8 +234,8 @@ case "$tag" in
     CLANG_VERSION=18
     VISION=yes
     ;;
-  pytorch-linux-jammy-py3.9-gcc11)
-    ANACONDA_PYTHON_VERSION=3.9
+  pytorch-linux-jammy-py3.10-gcc11)
+    ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
     VISION=yes
     KATEX=yes
@@ -325,13 +262,10 @@ case "$tag" in
     TRITON_CPU=yes
     ;;
   pytorch-linux-jammy-linter)
-    # TODO: Use 3.9 here because of this issue https://github.com/python/mypy/issues/13627.
-    # We will need to update mypy version eventually, but that's for another day. The task
-    # would be to upgrade mypy to 1.0.0 with Python 3.11
-    PYTHON_VERSION=3.9
+    PYTHON_VERSION=3.10
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter)
-    PYTHON_VERSION=3.9
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-linter)
+    PYTHON_VERSION=3.10
     CUDA_VERSION=12.8.1
     ;;
   pytorch-linux-jammy-aarch64-py3.10-gcc11)
@@ -339,7 +273,6 @@ case "$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
@@ -350,13 +283,15 @@ case "$tag" in
     GCC_VERSION=11
     ACL=yes
     VISION=yes
-    CONDA_CMAKE=yes
     OPENBLAS=yes
     # snadampal: skipping llvm src build install because the current version
     # from pytorch/llvm:9.0.1 is x86 specific
     SKIP_LLVM_SRC_BUILD_INSTALL=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
+  pytorch-linux-noble-riscv64-py3.12-gcc14)
+    GCC_VERSION=14
+    ;;
   *)
     # Catch-all for builds that are not hardcoded.
     VISION=yes
@@ -477,7 +412,14 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
 fi
 
 if [ -n "$GCC_VERSION" ]; then
-  if !(drun gcc --version 2>&1 | grep -q " $GCC_VERSION\\W"); then
+  if [[ "$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 " $GCC_VERSION\\W"); then
     echo "GCC_VERSION=$GCC_VERSION, but:"
     drun gcc --version
     exit 1

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

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

.ci/docker/ci_commit_pins/huggingface.txt

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

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

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

.ci/docker/ci_commit_pins/torchbench.txt

@@ -0,0 +1 @@
+74a23feff57432129df84d8099e622773cf77925

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

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

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-11ec6354315768a85da41032535e3b7b99c5f706
+bbb06c0334a6772b92d24bde54956e675c8c6604

.ci/docker/common/install_cpython.sh

@@ -66,8 +66,9 @@ function do_cpython_build {
         ln -s pip3 ${prefix}/bin/pip
     fi
     # install setuptools since python 3.12 is required to use distutils
-    ${prefix}/bin/pip install wheel==0.45.1 setuptools==80.9.0
-    local abi_tag=$(${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()))")
+    # 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 ''))")
     ln -sf ${prefix} /opt/python/${abi_tag}
 }
 
@@ -82,9 +83,9 @@ function build_cpython {
         py_suffix=${py_ver::-1}
         py_folder=$py_suffix
     fi
-    # Only b3 is available now
+    # Update to rc2 due to https://github.com/python/cpython/commit/c72699086fe4
     if [ "$py_suffix" == "3.14.0" ]; then
-        py_suffix="3.14.0b3"
+        py_suffix="3.14.0rc2"
     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=3.3.9
+NVSHMEM_VERSION=3.3.24
 
 function install_cuda {
   version=$1
@@ -62,14 +62,16 @@ function install_nvshmem {
   mkdir -p "${tmpdir}" && cd "${tmpdir}"
 
   # nvSHMEM license: https://docs.nvidia.com/nvshmem/api/sla.html
-  filename="libnvshmem_cuda${cuda_major_version}-linux-${arch_path}-${nvshmem_version}"
-  url="https://developer.download.nvidia.com/compute/redist/nvshmem/${nvshmem_version}/builds/cuda${cuda_major_version}/txz/agnostic/${dl_arch}/${filename}.tar.gz"
+  # This pattern is a lie as it is not consistent across versions, for 3.3.9 it was cuda_ver-arch-nvshhem-ver
+  filename="libnvshmem-linux-${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}"
 
   # download, unpack, install
   wget -q "${url}"
-  tar xf "${filename}.tar.gz"
-  cp -a "libnvshmem/include/"* /usr/local/include/
-  cp -a "libnvshmem/lib/"*     /usr/local/lib/
+  tar xf "${filename}${suffix}"
+  cp -a "${filename}/include/"* /usr/local/cuda/include/
+  cp -a "${filename}/lib/"*     /usr/local/cuda/lib64/
 
   # cleanup
   cd ..
@@ -126,74 +128,6 @@ 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"
@@ -212,18 +146,38 @@ 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; prune_124
+    12.4) install_124;
         ;;
-    12.6|12.6.*) install_126; prune_126
+    12.6|12.6.*) install_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

@@ -5,7 +5,15 @@ set -ex
 # cuSPARSELt license: https://docs.nvidia.com/cuda/cusparselt/license.html
 mkdir tmp_cusparselt && cd tmp_cusparselt
 
-if [[ ${CUDA_VERSION:0:4} =~ ^12\.[5-9]$ ]]; then
+if [[ ${CUDA_VERSION:0:4} =~ "13" ]]; then
+    arch_path='sbsa'
+    export TARGETARCH=${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
     arch_path='sbsa'
     export TARGETARCH=${TARGETARCH:-$(uname -m)}
     if [ ${TARGETARCH} = 'amd64' ] || [ "${TARGETARCH}" = 'x86_64' ]; then

.ci/docker/common/install_inductor_benchmark_deps.sh

@@ -5,9 +5,7 @@ set -ex
 source "$(dirname "${BASH_SOURCE[0]}")/common_utils.sh"
 
 function install_huggingface() {
-  local version
-  commit=$(get_pinned_commit huggingface)
-  pip_install "git+https://github.com/huggingface/transformers@${commit}"
+  pip_install -r huggingface-requirements.txt
 }
 
 function install_timm() {
@@ -15,11 +13,34 @@ function install_timm() {
   commit=$(get_pinned_commit timm)
 
   pip_install "git+https://github.com/huggingface/pytorch-image-models@${commit}"
-  # Clean up
-  conda_run pip uninstall -y torch torchvision triton
+}
+
+function install_torchbench() {
+  local commit
+  commit=$(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
 }
 
 # 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

@@ -7,6 +7,8 @@ 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==1.18.1
-pip_install onnxscript==0.3.1
+pip_install onnxruntime==1.22.1
+pip_install onnxscript==0.4.0
 
 # 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_triton.sh

@@ -57,7 +57,7 @@ if [ ! -f setup.py ]; then
   cd python
 fi
 
-pip_install pybind11==2.13.6
+pip_install pybind11==3.0.1
 
 # 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
@@ -103,5 +103,5 @@ fi
 # It depends on torch and triton. We don't want to install
 # triton and torch from production on Docker CI images
 if [[ "$ANACONDA_PYTHON_VERSION" != 3.9* ]]; then
-  pip_install helion==0.0.10 --no-deps
+  pip_install helion --no-deps
 fi

.ci/docker/common/install_ucc.sh

@@ -44,8 +44,12 @@ function install_ucc() {
 
   ./autogen.sh
 
-  # 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 "$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
 
   if [[ -n "$ROCM_VERSION" ]]; then
     if [[ -n "$PYTORCH_ROCM_ARCH" ]]; then

.ci/docker/common/install_xpu.sh

@@ -34,18 +34,27 @@ function install_ubuntu() {
 
     # The xpu-smi packages
     apt-get install -y flex bison xpu-smi
-    # 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 [[ "${XPU_DRIVER_TYPE,,}" == "rolling" ]]; then
-        apt-get install -y intel-ocloc
+
+    if [[ "${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
     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}
 
@@ -56,10 +65,14 @@ function install_ubuntu() {
 
 function install_rhel() {
     . /etc/os-release
-
-    if [[ ! " 8.8 8.10 9.0 9.2 9.3 " =~ " ${VERSION_ID} " ]]; then
-        echo "RHEL version ${VERSION_ID} not supported"
-        exit
+    if [[ "${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"
     fi
 
     dnf install -y 'dnf-command(config-manager)'
@@ -130,18 +143,18 @@ function install_sles() {
 
 }
 
-# Default use GPU driver LTS releases
-XPU_DRIVER_VERSION="/lts/2350"
-if [[ "${XPU_DRIVER_TYPE,,}" == "rolling" ]]; then
-    # Use GPU driver rolling releases
-    XPU_DRIVER_VERSION=""
+# Default use GPU driver rolling releases
+XPU_DRIVER_VERSION=""
+if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
+    # Use GPU driver LTS releases
+    XPU_DRIVER_VERSION="/lts/2350"
 fi
 
-# Default use Intel® oneAPI Deep Learning Essentials 2025.0
-if [[ "$XPU_VERSION" == "2025.1" ]]; then
-    XPU_PACKAGES="intel-deep-learning-essentials-2025.1"
+# Default use Intel® oneAPI Deep Learning Essentials 2025.1
+if [[ "$XPU_VERSION" == "2025.2" ]]; then
+    XPU_PACKAGES="intel-deep-learning-essentials-2025.2"
 else
-    XPU_PACKAGES="intel-deep-learning-essentials-2025.0"
+    XPU_PACKAGES="intel-deep-learning-essentials-2025.1"
 fi
 
 # The installation depends on the base OS

.ci/docker/common/patch_libstdc.sh

@@ -0,0 +1,9 @@
+#!/bin/bash
+set -xe
+# Script used in Linux x86 and aarch64 CD pipeline
+
+# Workaround for exposing statically linked libstdc++ CXX11 ABI symbols.
+# see: https://github.com/pytorch/pytorch/issues/133437
+LIBNONSHARED=$(gcc -print-file-name=libstdc++_nonshared.a)
+nm -g $LIBNONSHARED | grep " T " | grep recursive_directory_iterator | cut -c 20-  > weaken-symbols.txt
+objcopy --weaken-symbols weaken-symbols.txt $LIBNONSHARED $LIBNONSHARED

.ci/docker/libtorch/Dockerfile

@@ -69,6 +69,19 @@ 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

@@ -130,7 +130,8 @@ ENV LD_LIBRARY_PATH=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/lib64:/op
 RUN for cpython_version in "cp312-cp312" "cp313-cp313" "cp313-cp313t"; do \
     /opt/python/${cpython_version}/bin/python -m pip install setuptools wheel; \
     done;
-
+ADD ./common/patch_libstdc.sh patch_libstdc.sh
+RUN bash ./patch_libstdc.sh && rm patch_libstdc.sh
 
 # cmake-3.18.4 from pip; force in case cmake3 already exists
 RUN yum install -y python3-pip && \
@@ -175,6 +176,6 @@ ENV XPU_DRIVER_TYPE ROLLING
 RUN python3 -m pip install --upgrade pip && \
     python3 -mpip install cmake==3.28.4
 ADD ./common/install_xpu.sh install_xpu.sh
-ENV XPU_VERSION 2025.1
+ENV XPU_VERSION 2025.2
 RUN bash ./install_xpu.sh && rm install_xpu.sh
 RUN pushd /opt/_internal && tar -xJf static-libs-for-embedding-only.tar.xz && popd

.ci/docker/manywheel/Dockerfile_2_28_aarch64

@@ -71,3 +71,5 @@ RUN rm -rf /opt/python/cp33-cp33m /opt/_internal/cpython-3.3.6
 RUN rm -rf /opt/python/cp34-cp34m /opt/_internal/cpython-3.4.6
 COPY --from=openblas     /opt/OpenBLAS/  /opt/OpenBLAS/
 ENV LD_LIBRARY_PATH=/opt/OpenBLAS/lib:$LD_LIBRARY_PATH
+ADD ./common/patch_libstdc.sh patch_libstdc.sh
+RUN bash ./patch_libstdc.sh && rm patch_libstdc.sh

.ci/docker/manywheel/Dockerfile_cuda_aarch64

@@ -95,3 +95,5 @@ COPY --from=nvpl /opt/nvpl/lib/  /usr/local/lib/
 COPY --from=nvpl /opt/nvpl/include/  /usr/local/include/
 RUN ln -sf /usr/local/cuda-${BASE_CUDA_VERSION} /usr/local/cuda
 ENV PATH=/usr/local/cuda/bin:$PATH
+ADD ./common/patch_libstdc.sh patch_libstdc.sh
+RUN bash ./patch_libstdc.sh && rm patch_libstdc.sh

.ci/docker/manywheel/build.sh

@@ -67,6 +67,12 @@ 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

@@ -63,11 +63,12 @@ lark==0.12.0
 #Pinned versions: 0.12.0
 #test that import:
 
-librosa>=0.6.2 ; python_version < "3.11"
-librosa==0.10.2 ; python_version == "3.12"
+librosa>=0.6.2 ; python_version < "3.11" and platform_machine != "s390x"
+librosa==0.10.2 ; python_version == "3.12" and platform_machine != "s390x"
 #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
@@ -92,8 +93,9 @@ librosa==0.10.2 ; python_version == "3.12"
 #Pinned versions:
 #test that import:
 
-mypy==1.16.0
+mypy==1.16.0 ; platform_system != "Windows"
 # 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
@@ -110,14 +112,15 @@ 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"
-numba==0.55.2 ; python_version == "3.9"
-numba==0.55.2 ; python_version == "3.10"
-numba==0.60.0 ; python_version == "3.12"
+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"
 #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
@@ -261,11 +264,6 @@ 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
@@ -307,7 +305,7 @@ pytest-cpp==2.3.0
 #Pinned versions: 2.3.0
 #test that import:
 
-z3-solver==4.15.1.0
+z3-solver==4.15.1.0 ; platform_machine != "s390x"
 #Description: The Z3 Theorem Prover Project
 #Pinned versions:
 #test that import:
@@ -342,7 +340,7 @@ onnx==1.18.0
 #Pinned versions:
 #test that import:
 
-onnxscript==0.3.1
+onnxscript==0.4.0
 #Description: Required by mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
@@ -361,7 +359,6 @@ pwlf==2.2.1
 #Pinned versions: 2.2.1
 #test that import: test_sac_estimator.py
 
-
 # To build PyTorch itself
 pyyaml
 pyzstd
@@ -383,7 +380,7 @@ dataclasses_json==0.6.7
 cmake==4.0.0
 #Description: required for building
 
-tlparse==0.3.30
+tlparse==0.4.0
 #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@722b7e6f9ca512fcc526ad07d62b3d28c50bb6cd#egg=pytorch_sphinx_theme2
+-e git+https://github.com/pytorch/pytorch_sphinx_theme.git@71e55749be14ceb56e7f8211a9fb649866b87ad4#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 @@
-3.4.0
+3.5.0

.ci/docker/triton_xpu_version.txt

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

.ci/docker/ubuntu-cross-riscv/Dockerfile

@@ -0,0 +1,155 @@
+# 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

@@ -96,10 +96,11 @@ ARG ANACONDA_PYTHON_VERSION
 ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
-COPY ci_commit_pins/huggingface.txt huggingface.txt
+COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
 COPY ci_commit_pins/timm.txt timm.txt
+COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
 
 # (optional) Install non-default Ninja version
 ARG NINJA_VERSION

.ci/docker/ubuntu-xpu/Dockerfile

@@ -56,10 +56,10 @@ RUN rm install_openssl.sh
 ARG INDUCTOR_BENCHMARKS
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
-COPY ci_commit_pins/huggingface.txt huggingface.txt
+COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
 COPY ci_commit_pins/timm.txt timm.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt
 
 # Install XPU Dependencies
 ARG XPU_VERSION

.ci/docker/ubuntu/Dockerfile

@@ -66,6 +66,7 @@ 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 $UCC_COMMIT
 ENV UCX_HOME /usr
@@ -96,10 +97,11 @@ RUN rm install_openssl.sh
 ARG INDUCTOR_BENCHMARKS
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
-COPY ci_commit_pins/huggingface.txt huggingface.txt
+COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
 COPY ci_commit_pins/timm.txt timm.txt
+COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
 
 ARG TRITON
 ARG TRITON_CPU
@@ -180,7 +182,6 @@ 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_CUSPARSELT=0 BUILD_PYTHONLESS=1 DESIRED_PYTHON="3.9" ${SCRIPTPATH}/../manywheel/build.sh
+USE_NVSHMEM=0 USE_CUSPARSELT=0 BUILD_PYTHONLESS=1 DESIRED_PYTHON="3.10" ${SCRIPTPATH}/../manywheel/build.sh

.ci/lumen_cli/README.md

@@ -0,0 +1,31 @@
+# 🔧 Lumen_cli
+A Python CLI tool for building and testing PyTorch-based components, using a YAML configuration file for structured, repeatable workflows.
+
+
+## Features
+- **Build**
+    - external projects (e.g. vLLM)
+
+## 📦 Installation
+at the root of the pytorch repo
+```bash
+pip install -e .ci/lumen_cli
+```
+
+## Run the cli tool
+The cli tool must be used at root of pytorch repo, as example to run build external vllm:
+```bash
+python -m cli.run build external vllm
+```
+this will run the build steps with default behaviour for vllm project.
+
+to see help messages, run
+```bash
+python3 -m cli.run --help
+```
+
+## Add customized external build logics
+To add a new external build, for instance, add a new external build logics:
+1. create the build function in cli/lib folder
+2. register your target and the main build function at  EXTERNAL_BUILD_TARGET_DISPATCH in `cli/build_cli/register_build.py`
+3. [optional] create your ci config file in .github/ci_configs/${EXTERNAL_PACKAGE_NAME}.yaml

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

@@ -0,0 +1,37 @@
+import argparse
+import logging
+
+from cli.lib.common.cli_helper import register_targets, RichHelp, TargetSpec
+from cli.lib.core.vllm.vllm_build import VllmBuildRunner
+
+
+logger = logging.getLogger(__name__)
+
+# Maps targets to their argparse configuration and runner
+# it adds new target to path python -m cli.run build external {target} with buildrunner
+_TARGETS: dict[str, TargetSpec] = {
+    "vllm": {
+        "runner": VllmBuildRunner,
+        "help": "Build vLLM using docker buildx.",
+    }
+    # add yours ...
+}
+
+
+def register_build_commands(subparsers: argparse._SubParsersAction) -> None:
+    build_parser = subparsers.add_parser(
+        "build",
+        help="Build related commands",
+        formatter_class=RichHelp,
+    )
+    build_subparsers = build_parser.add_subparsers(dest="build_command", required=True)
+    overview = "\n".join(
+        f"  {name:12} {spec.get('help', '')}" for name, spec in _TARGETS.items()
+    )
+    external_parser = build_subparsers.add_parser(
+        "external",
+        help="Build external targets",
+        description="Build third-party targets.\n\nAvailable targets:\n" + overview,
+        formatter_class=RichHelp,
+    )
+    register_targets(external_parser, _TARGETS)

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

@@ -0,0 +1,71 @@
+"""
+Cli Argparser Utility helpers for CLI tasks.
+
+"""
+
+import argparse
+from abc import ABC, abstractmethod
+
+
+try:
+    from typing import Any, Callable, Required, TypedDict  # Python 3.11+
+except ImportError:
+    from typing import Any, Callable, TypedDict
+
+    from typing_extensions import Required  # Fallback for Python <3.11
+
+
+class BaseRunner(ABC):
+    def __init__(self, args: Any) -> None:
+        self.args = args
+
+    @abstractmethod
+    def run(self) -> None:
+        """runs main logics, required"""
+
+
+# Pretty help: keep newlines + show defaults
+class RichHelp(
+    argparse.ArgumentDefaultsHelpFormatter, argparse.RawDescriptionHelpFormatter
+):
+    pass
+
+
+class TargetSpec(TypedDict, total=False):
+    """CLI subcommand specification with bA."""
+
+    runner: Required[type[BaseRunner]]
+    help: str
+    description: str
+    add_arguments: Callable[[argparse.ArgumentParser], None]
+
+
+def register_targets(
+    parser: argparse.ArgumentParser,
+    target_specs: dict[str, TargetSpec],
+    common_args: Callable[[argparse.ArgumentParser], None] = lambda _: None,
+) -> None:
+    """Register target subcommands."""
+    targets = parser.add_subparsers(
+        dest="target",
+        required=True,
+        metavar="{" + ",".join(target_specs.keys()) + "}",
+    )
+
+    for name, spec in target_specs.items():
+        desc = spec.get("description") or spec["runner"].__doc__ or ""
+
+        p = targets.add_parser(
+            name,
+            help=spec.get("help", ""),
+            description=desc.strip(),
+            formatter_class=RichHelp,
+        )
+        p.set_defaults(
+            func=lambda args, cls=spec["runner"]: cls(args).run(),
+            _runner_class=spec["runner"],
+        )
+        if "add_arguments" in spec and callable(spec["add_arguments"]):
+            spec["add_arguments"](p)
+        if common_args:
+            common_args(p)

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

@@ -0,0 +1,42 @@
+"""
+Docker Utility helpers for CLI tasks.
+"""
+
+import logging
+from typing import Optional
+
+import docker
+from docker.errors import APIError, NotFound
+
+
+logger = logging.getLogger(__name__)
+
+# lazy singleton so we don't reconnect every call
+_docker_client: Optional[docker.DockerClient] = None
+
+
+def _get_client() -> docker.DockerClient:
+    global _docker_client
+    if _docker_client is None:
+        _docker_client = docker.from_env()
+    return _docker_client
+
+
+def local_image_exists(
+    image_name: str, client: Optional[docker.DockerClient] = None
+) -> bool:
+    """Return True if a local Docker image exists."""
+    if not image_name:
+        return False
+
+    client = client or _get_client()
+    try:
+        client.images.get(image_name)
+        return True
+    except (NotFound, APIError) as e:
+        logger.error(
+            "Error when checking Docker image '%s': %s",
+            image_name,
+            e.explanation if hasattr(e, "explanation") else str(e),
+        )
+        return False

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

@@ -0,0 +1,110 @@
+"""
+Environment Variables and Dataclasses Utility helpers for CLI tasks.
+"""
+
+import os
+from dataclasses import field, fields, is_dataclass, MISSING
+from pathlib import Path
+from textwrap import indent
+from typing import Optional, Union
+
+from cli.lib.common.utils import str2bool
+
+
+def get_env(name: str, default: str = "") -> str:
+    """Get environment variable with default fallback."""
+    return os.environ.get(name) or default
+
+
+def env_path_optional(
+    name: str,
+    default: Optional[Union[str, Path]] = None,
+    resolve: bool = True,
+) -> Optional[Path]:
+    """Get environment variable as optional Path."""
+    val = get_env(name) or default
+    if not val:
+        return None
+
+    path = Path(val)
+    return path.resolve() if resolve else path
+
+
+def env_path(
+    name: str,
+    default: Optional[Union[str, Path]] = None,
+    resolve: bool = True,
+) -> Path:
+    """Get environment variable as Path, raise if missing."""
+    path = env_path_optional(name, default, resolve)
+    if not path:
+        raise ValueError(f"Missing path value for {name}")
+    return path
+
+
+def env_bool(
+    name: str,
+    default: bool = False,
+) -> bool:
+    val = get_env(name)
+    if not val:
+        return default
+    return str2bool(val)
+
+
+def env_bool_field(
+    name: str,
+    default: bool = False,
+):
+    return field(default_factory=lambda: env_bool(name, default))
+
+
+def env_path_field(
+    name: str,
+    default: Union[str, Path] = "",
+    *,
+    resolve: bool = True,
+) -> Path:
+    return field(default_factory=lambda: env_path(name, default, resolve=resolve))
+
+
+def env_str_field(
+    name: str,
+    default: str = "",
+) -> str:
+    return field(default_factory=lambda: get_env(name, default))
+
+
+def generate_dataclass_help(cls) -> str:
+    """Auto-generate help text for dataclass fields."""
+    if not is_dataclass(cls):
+        raise TypeError(f"{cls} is not a dataclass")
+
+    def get_value(f):
+        if f.default is not MISSING:
+            return f.default
+        if f.default_factory is not MISSING:
+            try:
+                return f.default_factory()
+            except Exception as e:
+                return f"<error: {e}>"
+        return "<required>"
+
+    lines = [f"{f.name:<22} = {repr(get_value(f))}" for f in fields(cls)]
+    return indent("\n".join(lines), "    ")
+
+
+def with_params_help(params_cls: type, title: str = "Parameter defaults"):
+    """
+    Class decorator that appends a help table generated from another dataclass
+    (e.g., VllmParameters) to the decorated class's docstring.
+    """
+    if not is_dataclass(params_cls):
+        raise TypeError(f"{params_cls} must be a dataclass")
+
+    def _decorator(cls: type) -> type:
+        block = generate_dataclass_help(params_cls)
+        cls.__doc__ = (cls.__doc__ or "") + f"\n\n{title}:\n{block}"
+        return cls
+
+    return _decorator

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

@@ -0,0 +1,143 @@
+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

@@ -0,0 +1,69 @@
+"""
+Git Utility helpers for CLI tasks.
+"""
+
+import logging
+from pathlib import Path
+
+from cli.lib.common.path_helper import remove_dir
+from git import GitCommandError, RemoteProgress, Repo
+
+
+logger = logging.getLogger(__name__)
+
+
+class PrintProgress(RemoteProgress):
+    """Simple progress logger for git operations."""
+
+    def __init__(self, interval: int = 5):
+        super().__init__()
+        self._last_percent = -1
+        self._interval = interval
+
+    def update(self, op_code, cur, max=None, message=""):
+        msg = self._cur_line or message
+        if max and cur:
+            percent = int(cur / max * 100)
+            if percent != self._last_percent and percent % self._interval == 0:
+                self._last_percent = percent
+                logger.info("Progress: %d%% - %s", percent, msg)
+        elif msg:
+            logger.info(msg)
+
+
+def clone_external_repo(target: str, repo: str, dst: str = "", update_submodules=False):
+    """Clone repository with pinned commit and optional submodules."""
+    dst = dst or target
+
+    try:
+        logger.info("Cloning %s to %s", target, dst)
+
+        # Clone and fetch
+        remove_dir(dst)
+        r = Repo.clone_from(repo, dst, progress=PrintProgress())
+        r.git.fetch("--all", "--tags")
+
+        # Checkout pinned commit
+        commit = get_post_build_pinned_commit(target)
+        logger.info("Checking out pinned %s commit %s", target, commit)
+        r.git.checkout(commit)
+
+        # Update submodules if requested
+        if update_submodules and r.submodules:
+            logger.info("Updating %d submodule(s)", len(r.submodules))
+            for sm in r.submodules:
+                sm.update(init=True, recursive=True, progress=PrintProgress())
+
+        logger.info("Successfully cloned %s", target)
+        return r, commit
+
+    except GitCommandError as e:
+        logger.error("Git operation failed: %s", e)
+        raise
+
+
+def get_post_build_pinned_commit(name: str, prefix=".github/ci_commit_pins") -> str:
+    path = Path(prefix) / f"{name}.txt"
+    if not path.exists():
+        raise FileNotFoundError(f"Pin file not found: {path}")
+    return path.read_text(encoding="utf-8").strip()

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

@@ -0,0 +1,14 @@
+"""
+Logger Utility helpers for CLI tasks.
+"""
+
+import logging
+import sys
+
+
+def setup_logging(level: int = logging.INFO):
+    logging.basicConfig(
+        level=level,
+        format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
+        stream=sys.stdout,
+    )

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

@@ -0,0 +1,62 @@
+"""Path utility helpers for CLI tasks."""
+
+import logging
+import shutil
+from pathlib import Path
+from typing import Union
+
+
+logger = logging.getLogger(__name__)
+
+
+def get_path(path: Union[str, Path], resolve: bool = False) -> Path:
+    """Convert to Path object, optionally resolving to absolute path."""
+    if not path:
+        raise ValueError("Path cannot be None or empty")
+    result = Path(path)
+    return result.resolve() if resolve else result
+
+
+def ensure_dir_exists(path: Union[str, Path]) -> Path:
+    """Create directory if it doesn't exist."""
+    path_obj = get_path(path)
+    path_obj.mkdir(parents=True, exist_ok=True)
+    return path_obj
+
+
+def remove_dir(path: Union[str, Path, None]) -> None:
+    """Remove directory if it exists."""
+    if not path:
+        return
+    path_obj = get_path(path)
+    if path_obj.exists():
+        shutil.rmtree(path_obj)
+
+
+def force_create_dir(path: Union[str, Path]) -> Path:
+    """Remove directory if exists, then create fresh empty directory."""
+    remove_dir(path)
+    return ensure_dir_exists(path)
+
+
+def copy(src: Union[str, Path], dst: Union[str, Path]) -> None:
+    """Copy file or directory from src to dst."""
+    src_path = get_path(src, resolve=True)
+    dst_path = get_path(dst, resolve=True)
+
+    if not src_path.exists():
+        raise FileNotFoundError(f"Source does not exist: {src_path}")
+
+    dst_path.parent.mkdir(parents=True, exist_ok=True)
+
+    if src_path.is_file():
+        shutil.copy2(src_path, dst_path)
+    elif src_path.is_dir():
+        shutil.copytree(src_path, dst_path, dirs_exist_ok=True)
+    else:
+        raise ValueError(f"Unsupported path type: {src_path}")
+
+
+def is_path_exist(path: Union[str, Path, None]) -> bool:
+    """Check if path exists."""
+    return bool(path and get_path(path).exists())

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

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

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

@@ -0,0 +1,139 @@
+"""
+General Utility helpers for CLI tasks.
+"""
+
+import logging
+import os
+import shlex
+import subprocess
+import sys
+from contextlib import contextmanager
+from pathlib import Path
+from typing import Optional
+
+
+logger = logging.getLogger(__name__)
+
+
+def run_command(
+    cmd: str,
+    use_shell: bool = False,
+    log_cmd: bool = True,
+    cwd: Optional[str] = None,
+    env: Optional[dict] = None,
+    check: bool = True,
+) -> int:
+    """Run a command with optional shell execution."""
+    if use_shell:
+        args = cmd
+        log_prefix = "[shell]"
+        executable = "/bin/bash"
+    else:
+        args = shlex.split(cmd)
+        log_prefix = "[cmd]"
+        executable = None
+
+    if log_cmd:
+        display_cmd = cmd if use_shell else " ".join(args)
+        logger.info("%s %s", log_prefix, display_cmd)
+
+    run_env = {**os.environ, **(env or {})}
+
+    proc = subprocess.run(
+        args,
+        shell=use_shell,
+        executable=executable,
+        stdout=sys.stdout,
+        stderr=sys.stderr,
+        cwd=cwd,
+        env=run_env,
+        check=False,
+    )
+
+    if check and proc.returncode != 0:
+        logger.error(
+            "%s Command failed (exit %s): %s", log_prefix, proc.returncode, cmd
+        )
+        raise subprocess.CalledProcessError(
+            proc.returncode, args if not use_shell else cmd
+        )
+
+    return proc.returncode
+
+
+def str2bool(value: Optional[str]) -> bool:
+    """Convert environment variables to boolean values."""
+    if not value:
+        return False
+    if not isinstance(value, str):
+        raise ValueError(
+            f"Expected a string value for boolean conversion, got {type(value)}"
+        )
+    value = value.strip().lower()
+
+    true_value_set = {"1", "true", "t", "yes", "y", "on", "enable", "enabled", "found"}
+    false_value_set = {"0", "false", "f", "no", "n", "off", "disable"}
+
+    if value in true_value_set:
+        return True
+    if value in false_value_set:
+        return False
+    raise ValueError(f"Invalid string value for boolean conversion: {value}")
+
+
+@contextmanager
+def temp_environ(updates: dict[str, str]):
+    """
+    Temporarily set environment variables and restore them after the block.
+    Args:
+        updates: Dict of environment variables to set.
+    """
+    missing = object()
+    old: dict[str, str | object] = {k: os.environ.get(k, missing) for k in updates}
+    try:
+        os.environ.update(updates)
+        yield
+    finally:
+        for k, v in old.items():
+            if v is missing:
+                os.environ.pop(k, None)
+            else:
+                os.environ[k] = v  # type: ignore[arg-type]
+
+
+@contextmanager
+def working_directory(path: str):
+    """
+    Temporarily change the working directory inside a context.
+    """
+    if not path:
+        # No-op context
+        yield
+        return
+    prev_cwd = os.getcwd()
+    try:
+        os.chdir(path)
+        yield
+    finally:
+        os.chdir(prev_cwd)
+
+
+def get_wheels(
+    output_dir: Path,
+    max_depth: Optional[int] = None,
+) -> list[str]:
+    """Return a list of wheels found in the given output directory."""
+    root = Path(output_dir)
+    if not root.exists():
+        return []
+    items = []
+    for dirpath, _, filenames in os.walk(root):
+        depth = Path(dirpath).relative_to(root).parts
+        if max_depth is not None and len(depth) > max_depth:
+            continue
+        for fname in sorted(filenames):
+            if fname.endswith(".whl"):
+                pkg = fname.split("-")[0]
+                relpath = str((Path(dirpath) / fname).relative_to(root))
+                items.append({"pkg": pkg, "relpath": relpath})
+    return items

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

@@ -0,0 +1,292 @@
+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

@@ -0,0 +1,285 @@
+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"
+    )
+
+    # 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_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_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

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

.ci/lumen_cli/cli/run.py

@@ -0,0 +1,40 @@
+# main.py
+
+import argparse
+import logging
+
+from cli.build_cli.register_build import register_build_commands
+from cli.lib.common.logger import setup_logging
+from cli.test_cli.register_test import register_test_commands
+
+
+logger = logging.getLogger(__name__)
+
+
+def main():
+    # Define top-level parser
+    parser = argparse.ArgumentParser(description="Lumos CLI")
+    subparsers = parser.add_subparsers(dest="command", required=True)
+    parser.add_argument(
+        "--log-level", default="INFO", help="Log level (DEBUG, INFO, WARNING, ERROR)"
+    )
+
+    # registers second-level subcommands
+    register_build_commands(subparsers)
+    register_test_commands(subparsers)
+
+    # parse args after all options are registered
+    args = parser.parse_args()
+
+    # setup global logging
+    setup_logging(getattr(logging, args.log_level.upper(), logging.INFO))
+    logger.debug("Parsed args: %s", args)
+
+    if hasattr(args, "func"):
+        args.func(args)
+    else:
+        parser.print_help()
+
+
+if __name__ == "__main__":
+    main()

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

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

.ci/lumen_cli/pyproject.toml

@@ -0,0 +1,23 @@
+[project]
+name = "lumen-ci"
+version = "0.1.0"
+dependencies = [
+    "pyyaml==6.0.2",
+    "GitPython==3.1.45",
+    "docker==7.1.0",
+    "pytest==7.3.2",
+    "uv==0.8.6"
+]
+
+[tool.setuptools]
+packages = ["cli"]
+
+[tool.setuptools.package-dir]
+cli = "cli"
+
+[tool.ruff.lint]
+# Enable preview mode for linting
+preview = true
+
+# Now you can select your preview rules, like RUF048
+extend-select = ["RUF048"]

.ci/lumen_cli/tests/test_app.py

@@ -0,0 +1,47 @@
+# tests/test_cli.py
+import io
+import sys
+import unittest
+from contextlib import redirect_stderr, redirect_stdout
+from unittest.mock import patch
+
+from cli.run import main
+
+
+class TestArgparseCLI(unittest.TestCase):
+    @patch("cli.build_cli.register_build.VllmBuildRunner.run", return_value=None)
+    @patch("cli.build_cli.register_build.VllmBuildRunner.__init__", return_value=None)
+    def test_cli_run_build_external(self, mock_init, mock_run):
+        from cli.run import main  # import after patches if needed
+
+        test_args = ["cli.run", "build", "external", "vllm"]
+        with patch.object(sys, "argv", test_args):
+            # argparse may call sys.exit on error; capture to avoid test aborts
+            try:
+                main()
+            except SystemExit:
+                pass
+        mock_init.assert_called_once()  # got constructed
+        mock_run.assert_called_once_with()  # run() called
+
+    def test_build_help(self):
+        test_args = ["cli.run", "build", "--help"]
+
+        with patch.object(sys, "argv", test_args):
+            stdout = io.StringIO()
+            stderr = io.StringIO()
+
+            # --help always raises SystemExit(0)
+            with self.assertRaises(SystemExit) as cm:
+                with redirect_stdout(stdout), redirect_stderr(stderr):
+                    main()
+
+            self.assertEqual(cm.exception.code, 0)
+
+            output = stdout.getvalue()
+            self.assertIn("usage", output)
+            self.assertIn("external", output)
+
+
+if __name__ == "__main__":
+    unittest.main()

.ci/lumen_cli/tests/test_cli_helper.py

@@ -0,0 +1,115 @@
+import argparse
+import io
+import unittest
+from contextlib import redirect_stderr
+from unittest.mock import patch
+
+from cli.lib.common.cli_helper import BaseRunner, register_targets, RichHelp, TargetSpec
+
+
+# ---- Dummy runners for unittests----
+class FooRunner(BaseRunner):
+    """Foo description from docstring."""
+
+    def run(self) -> None:  # replaced by mock
+        pass
+
+
+class BarRunner(BaseRunner):
+    def run(self) -> None:  # replaced by mock
+        pass
+
+
+def add_foo_args(p: argparse.ArgumentParser) -> None:
+    p.add_argument("--x", type=int, required=True, help="x value")
+
+
+def common_args(p: argparse.ArgumentParser) -> None:
+    p.add_argument("--verbose", action="store_true", help="verbose flag")
+
+
+def build_parser(specs: dict[str, TargetSpec]) -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(prog="app", formatter_class=RichHelp)
+    register_targets(
+        parser=parser,
+        target_specs=specs,
+        common_args=common_args,
+    )
+    return parser
+
+
+def get_subparser(
+    parser: argparse.ArgumentParser, name: str
+) -> argparse.ArgumentParser:
+    subparsers_action = next(
+        a
+        for a in parser._subparsers._group_actions  # type: ignore[attr-defined]
+        if isinstance(a, argparse._SubParsersAction)
+    )
+    return subparsers_action.choices[name]
+
+
+class TestRegisterTargets(unittest.TestCase):
+    def test_metavar_lists_targets(self):
+        specs: dict[str, TargetSpec] = {
+            "foo": {"runner": FooRunner, "add_arguments": add_foo_args},
+            "bar": {"runner": BarRunner},
+        }
+        parser = build_parser(specs)
+        subparsers_action = next(
+            a
+            for a in parser._subparsers._group_actions  # type: ignore[attr-defined]
+            if isinstance(a, argparse._SubParsersAction)
+        )
+        self.assertEqual(subparsers_action.metavar, "{foo,bar}")
+
+    def test_add_arguments_and_common_args_present(self):
+        specs: dict[str, TargetSpec] = {
+            "foo": {"runner": FooRunner, "add_arguments": add_foo_args},
+        }
+        parser = build_parser(specs)
+        foo = get_subparser(parser, "foo")
+        help_text = foo.format_help()
+        self.assertIn("--x", help_text)
+        self.assertIn("--verbose", help_text)
+
+    def test_runner_constructed_with_ns_and_run_called(self):
+        specs: dict[str, TargetSpec] = {
+            "foo": {"runner": FooRunner, "add_arguments": add_foo_args},
+        }
+        parser = build_parser(specs)
+
+        with (
+            patch.object(FooRunner, "__init__", return_value=None) as mock_init,
+            patch.object(FooRunner, "run", return_value=None) as mock_run,
+        ):
+            ns = parser.parse_args(["foo", "--x", "3", "--verbose"])
+            ns.func(ns)  # set by register_targets
+            # __init__ received the Namespace
+            self.assertEqual(mock_init.call_count, 1)
+            (called_ns,), _ = mock_init.call_args
+            self.assertIsInstance(called_ns, argparse.Namespace)
+            # run() called with no args
+            mock_run.assert_called_once_with()
+
+    def test_runner_docstring_used_as_description_when_missing(self):
+        specs: dict[str, TargetSpec] = {
+            "foo": {"runner": FooRunner, "add_arguments": add_foo_args},
+        }
+        parser = build_parser(specs)
+        foo = get_subparser(parser, "foo")
+        help_text = foo.format_help()
+        self.assertIn("Foo description from docstring.", help_text)
+
+    def test_missing_target_raises_systemexit_with_usage(self):
+        specs: dict[str, TargetSpec] = {"foo": {"runner": FooRunner}}
+        parser = build_parser(specs)
+        buf = io.StringIO()
+        with self.assertRaises(SystemExit), redirect_stderr(buf):
+            parser.parse_args([])
+        err = buf.getvalue()
+        self.assertIn("usage:", err)
+
+
+if __name__ == "__main__":
+    unittest.main()

.ci/lumen_cli/tests/test_docker_helper.py

@@ -0,0 +1,75 @@
+import unittest
+from unittest import mock
+from unittest.mock import MagicMock
+
+import docker.errors as derr
+from cli.lib.common.docker_helper import _get_client, local_image_exists
+
+
+class TestDockerImageHelpers(unittest.TestCase):
+    def setUp(self):
+        # Reset the singleton in the target module
+        patcher = mock.patch("cli.lib.common.docker_helper._docker_client", None)
+        self.addCleanup(patcher.stop)
+        patcher.start()
+
+    def test_local_image_exists_true(self):
+        # Mock a docker client whose images.get returns an object (no exception)
+        mock_client = MagicMock()
+        mock_client.images.get.return_value = object()
+        ok = local_image_exists("repo:tag", client=mock_client)
+        self.assertTrue(ok)
+
+    def test_local_image_exists_not_found_false(self):
+        mock_client = MagicMock()
+        # Raise docker.errors.NotFound
+        mock_client.images.get.side_effect = derr.NotFound("nope")
+        ok = local_image_exists("missing:latest", client=mock_client)
+        self.assertFalse(ok)
+
+    def test_local_image_exists_api_error_false(self):
+        mock_client = MagicMock()
+        mock_client.images.get.side_effect = derr.APIError("boom", None)
+
+        ok = local_image_exists("broken:tag", client=mock_client)
+        self.assertFalse(ok)
+
+    def test_local_image_exists_uses_lazy_singleton(self):
+        # Patch docker.from_env used by _get_client()
+        with mock.patch(
+            "cli.lib.common.docker_helper.docker.from_env"
+        ) as mock_from_env:
+            mock_docker_client = MagicMock()
+            mock_from_env.return_value = mock_docker_client
+
+            # First call should create and cache the client
+            c1 = _get_client()
+            self.assertIs(c1, mock_docker_client)
+            mock_from_env.assert_called_once()
+
+            # Second call should reuse cached client (no extra from_env calls)
+            c2 = _get_client()
+            self.assertIs(c2, mock_docker_client)
+            mock_from_env.assert_called_once()  # still once
+
+    def test_local_image_exists_without_client_param_calls_get_client_once(self):
+        # Ensure _get_client is called and cached; local_image_exists should reuse it
+        with mock.patch("cli.lib.common.docker_helper._get_client") as mock_get_client:
+            mock_client = MagicMock()
+            mock_get_client.return_value = mock_client
+
+            # 1st call
+            local_image_exists("repo:tag")
+            # 2nd call
+            local_image_exists("repo:tag2")
+
+            # local_image_exists should call _get_client each time,
+            # but your _get_client itself caches docker.from_env.
+            self.assertEqual(mock_get_client.call_count, 2)
+            self.assertEqual(mock_client.images.get.call_count, 2)
+            mock_client.images.get.assert_any_call("repo:tag")
+            mock_client.images.get.assert_any_call("repo:tag2")
+
+
+if __name__ == "__main__":
+    unittest.main()

.ci/lumen_cli/tests/test_envs_helper.py

@@ -0,0 +1,149 @@
+import os
+import unittest
+from dataclasses import dataclass
+from pathlib import Path
+from unittest.mock import patch
+
+import cli.lib.common.envs_helper as m
+
+
+class TestEnvHelpers(unittest.TestCase):
+    def setUp(self):
+        # Keep a copy of the original environment to restore later
+        self._env_backup = dict(os.environ)
+
+    def tearDown(self):
+        # Restore environment to original state
+        os.environ.clear()
+        os.environ.update(self._env_backup)
+
+    # -------- get_env --------
+    def test_get_env_unset_returns_default(self):
+        with patch.dict(os.environ, {}, clear=True):
+            self.assertEqual(m.get_env("FOO", "default"), "default")
+
+    def test_get_env_empty_returns_default(self):
+        with patch.dict(os.environ, {"FOO": ""}, clear=True):
+            self.assertEqual(m.get_env("FOO", "default"), "default")
+
+    def test_get_env_set_returns_value(self):
+        with patch.dict(os.environ, {"FOO": "bar"}, clear=True):
+            self.assertEqual(m.get_env("FOO", "default"), "bar")
+
+    def test_get_env_not_exist_returns_default(self):
+        with patch.dict(os.environ, {"FOO": "bar"}, clear=True):
+            self.assertEqual(m.get_env("TEST_NOT_EXIST", "default"), "default")
+
+    def test_get_env_not_exist_without_default(self):
+        with patch.dict(os.environ, {"FOO": "bar"}, clear=True):
+            self.assertEqual(m.get_env("TEST_NOT_EXIST"), "")
+
+    # -------- env_bool --------
+    def test_env_bool_uses_default_when_unset(self):
+        with patch.dict(os.environ, {}, clear=True):
+            self.assertTrue(m.env_bool("FLAG", default=True))
+            self.assertFalse(m.env_bool("FLAG", default=False))
+
+    def test_env_bool_uses_str2bool_when_set(self):
+        # Patch str2bool used by env_bool so we don't depend on its exact behavior
+        def fake_str2bool(s: str) -> bool:
+            return s.lower() in {"1", "true", "yes", "on", "y"}
+
+        with (
+            patch.dict(os.environ, {"FLAG": "yEs"}, clear=True),
+            patch.object(m, "str2bool", fake_str2bool),
+        ):
+            self.assertTrue(m.env_bool("FLAG", default=False))
+
+    # -------- env_path_optional / env_path --------
+    def test_env_path_optional_unset_returns_none_by_default(self):
+        with patch.dict(os.environ, {}, clear=True):
+            self.assertIsNone(m.env_path_optional("P"))
+
+    def test_env_path_optional_unset_returns_none_when_env_var_is_empty(self):
+        with patch.dict(os.environ, {"P": ""}, clear=True):
+            self.assertIsNone(m.env_path_optional("P"))
+
+    def test_env_path_optional_unset_returns_default_str(self):
+        # default as string; resolve=True by default -> absolute path
+        default_str = "x/y"
+        with patch.dict(os.environ, {}, clear=True):
+            p = m.env_path_optional("P", default=default_str)
+            self.assertIsInstance(p, Path)
+            self.assertIsNotNone(p)
+            if p:
+                self.assertTrue(p.is_absolute())
+                self.assertEqual(p.parts[-2:], ("x", "y"))
+
+    def test_env_path_optional_unset_returns_default_path_no_resolve(self):
+        d = Path("z")
+        with patch.dict(os.environ, {}, clear=True):
+            p = m.env_path_optional("P", default=d, resolve=False)
+            self.assertEqual(p, d)
+
+    def test_env_path_optional_respects_resolve_true(self):
+        with patch.dict(os.environ, {"P": "a/b"}, clear=True):
+            p = m.env_path_optional("P", resolve=True)
+            self.assertIsInstance(p, Path)
+            if p:
+                self.assertTrue(p.is_absolute())
+
+    def test_env_path_optional_respects_resolve_false(self):
+        with patch.dict(os.environ, {"P": "rel/dir"}, clear=True):
+            p = m.env_path_optional("P", resolve=False)
+            self.assertEqual(p, Path("rel/dir"))
+            if p:
+                self.assertFalse(p.is_absolute())
+
+    def test_env_path_raises_when_missing_and_default_none(self):
+        with patch.dict(os.environ, {}, clear=True):
+            with self.assertRaises(ValueError):
+                m.env_path("P", None, resolve=True)
+
+    def test_env_path_returns_path_when_present(self):
+        tmp = Path("./b").resolve()
+        with patch.dict(os.environ, {"P": str(tmp)}, clear=True):
+            p = m.env_path("P", None, resolve=True)
+            self.assertEqual(p, tmp)
+
+    # -------- dataclass field helpers --------
+    def test_dataclass_fields_read_env_at_instantiation(self):
+        @dataclass
+        class Cfg:
+            flag: bool = m.env_bool_field("FLAG", default=False)
+            out: Path = m.env_path_field("OUT", default="ab", resolve=True)
+            name: str = m.env_str_field("NAME", default="anon")
+
+        # First instantiation
+        with patch.dict(
+            os.environ, {"FLAG": "true", "OUT": "outdir", "NAME": "alice"}, clear=True
+        ):
+            cfg1 = Cfg()
+            self.assertTrue(cfg1.flag)
+            self.assertIsInstance(cfg1.out, Path)
+            self.assertTrue(cfg1.out.is_absolute())
+            self.assertEqual(cfg1.name, "alice")
+            cfg1.name = "bob"  # change instance value
+            self.assertEqual(cfg1.name, "bob")  # change is reflected
+
+        # Change env; new instance should reflect new values
+        with patch.dict(os.environ, {"FLAG": "false", "NAME": ""}, clear=True):
+            cfg2 = Cfg()
+            self.assertFalse(cfg2.flag)  # str2bool("false") -> False
+            self.assertTrue("ab" in str(cfg2.out))
+            self.assertIsInstance(cfg2.out, Path)
+            self.assertTrue(cfg2.out.is_absolute())
+            self.assertEqual(cfg2.name, "anon")  # empty -> fallback to default
+
+    def test_dataclass_path_field_with_default_value(self):
+        @dataclass
+        class C2:
+            out: Path = m.env_path_field("OUT", default="some/dir", resolve=False)
+
+        with patch.dict(os.environ, {}, clear=True):
+            c = C2()
+            self.assertEqual(c.out, Path("some/dir"))
+
+
+if __name__ == "__main__":
+    unittest.main()

.ci/lumen_cli/tests/test_path_helper.py

@@ -0,0 +1,122 @@
+# test_path_utils.py
+# Run: pytest -q
+
+import os
+import unittest
+from pathlib import Path
+from tempfile import TemporaryDirectory
+
+from cli.lib.common.path_helper import (
+    copy,
+    ensure_dir_exists,
+    force_create_dir,
+    get_path,
+    is_path_exist,
+    remove_dir,
+)
+
+
+class TestPathHelper(unittest.TestCase):
+    def setUp(self):
+        self.tmpdir = TemporaryDirectory()
+        self.tmp_path = Path(self.tmpdir.name)
+
+    def tearDown(self):
+        self.tmpdir.cleanup()
+
+    # -------- get_path --------
+    def test_get_path_returns_path_for_str(self):
+        # Use relative path to avoid absolute-ness
+        rel_str = "sub/f.txt"
+        os.chdir(self.tmp_path)
+        p = get_path(rel_str, resolve=False)
+        self.assertIsInstance(p, Path)
+        self.assertFalse(p.is_absolute())
+        self.assertEqual(str(p), rel_str)
+
+    def test_get_path_resolves(self):
+        rel_str = "sub/f.txt"
+        p = get_path(str(self.tmp_path / rel_str), resolve=True)
+        self.assertTrue(p.is_absolute())
+        self.assertTrue(str(p).endswith(rel_str))
+
+    def test_get_path_with_path_input(self):
+        p_in = self.tmp_path / "sub/f.txt"
+        p_out = get_path(p_in, resolve=False)
+        self.assertTrue(str(p_out) == str(p_in))
+
+    def test_get_path_with_none_raises(self):
+        with self.assertRaises(ValueError):
+            get_path(None)  # type: ignore[arg-type]
+
+    def test_get_path_invalid_type_raises(self):
+        with self.assertRaises(TypeError):
+            get_path(123)  # type: ignore[arg-type]
+
+    # -------- ensure_dir_exists / force_create_dir / remove_dir --------
+    def test_ensure_dir_exists_creates_and_is_idempotent(self):
+        d = self.tmp_path / "made"
+        ensure_dir_exists(d)
+        self.assertTrue(d.exists() and d.is_dir())
+        ensure_dir_exists(d)
+
+    def test_force_create_dir_clears_existing(self):
+        d = self.tmp_path / "fresh"
+        (d / "inner").mkdir(parents=True)
+        (d / "inner" / "f.txt").write_text("x")
+        force_create_dir(d)
+        self.assertTrue(d.exists())
+        self.assertEqual(list(d.iterdir()), [])
+
+    def test_remove_dir_none_is_noop(self):
+        remove_dir(None)  # type: ignore[arg-type]
+
+    def test_remove_dir_nonexistent_is_noop(self):
+        ghost = self.tmp_path / "ghost"
+        remove_dir(ghost)
+
+    def test_remove_dir_accepts_str(self):
+        d = self.tmp_path / "to_rm"
+        d.mkdir()
+        remove_dir(str(d))
+        self.assertFalse(d.exists())
+
+    # -------- copy --------
+    def test_copy_file_to_file(self):
+        src = self.tmp_path / "src.txt"
+        dst = self.tmp_path / "out" / "dst.txt"
+        src.write_text("hello")
+        copy(src, dst)
+        self.assertEqual(dst.read_text(), "hello")
+
+    def test_copy_dir_to_new_dir(self):
+        src = self.tmp_path / "srcdir"
+        (src / "a").mkdir(parents=True)
+        (src / "a" / "f.txt").write_text("content")
+        dst = self.tmp_path / "destdir"
+        copy(src, dst)
+        self.assertEqual((dst / "a" / "f.txt").read_text(), "content")
+
+    def test_copy_dir_into_existing_dir_overwrite_true_merges(self):
+        src = self.tmp_path / "srcdir"
+        dst = self.tmp_path / "destdir"
+        (src / "x").mkdir(parents=True)
+        (src / "x" / "new.txt").write_text("new")
+        dst.mkdir()
+        (dst / "existing.txt").write_text("old")
+        copy(src, dst)
+        self.assertEqual((dst / "existing.txt").read_text(), "old")
+        self.assertEqual((dst / "x" / "new.txt").read_text(), "new")
+
+    def test_is_str_path_exist(self):
+        p = self.tmp_path / "x.txt"
+        p.write_text("1")
+        self.assertTrue(is_path_exist(str(p)))
+        self.assertTrue(is_path_exist(p))
+        self.assertFalse(is_path_exist(str(self.tmp_path / "missing")))
+        self.assertFalse(is_path_exist(self.tmp_path / "missing"))
+        self.assertFalse(is_path_exist(""))
+
+
+if __name__ == "__main__":
+    unittest.main()

.ci/lumen_cli/tests/test_run_plan.py

@@ -0,0 +1,185 @@
+# 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

@@ -0,0 +1,143 @@
+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

@@ -0,0 +1,176 @@
+import os
+import tempfile
+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"
+
+
+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.common.envs_helper.env_path_optional",
+        side_effect=lambda name, default=None, resolve=True: {
+            "DOCKERFILE_PATH": Path("/abs/vllm/Dockerfile"),
+            "TORCH_WHEELS_PATH": Path("/abs/dist"),
+            "OUTPUT_DIR": Path("/abs/shared"),
+        }.get(name, Path(default) if default is not None else None),
+    )
+    @patch.dict(
+        os.environ,
+        {
+            "USE_TORCH_WHEEL": "1",
+            "USE_LOCAL_BASE_IMAGE": "1",
+            "USE_LOCAL_DOCKERFILE": "1",
+            "BASE_IMAGE": "my/image:tag",
+            "DOCKERFILE_PATH": "vllm/Dockerfile",
+            "TORCH_WHEELS_PATH": "dist",
+            "OUTPUT_DIR": "shared",
+        },
+        clear=True,
+    )
+    def test_params_success_normalizes_and_validates(
+        self, mock_env_path, mock_is_path, mock_local_img
+    ):
+        params = vllm_build.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.dict(
+        os.environ, {"USE_TORCH_WHEEL": "1", "TORCH_WHEELS_PATH": "dist"}, clear=True
+    )
+    def test_params_missing_torch_whls_raises(self, _is_path):
+        with tempfile.TemporaryDirectory() as td:
+            os.chdir(td)
+            with self.assertRaises(ValueError) as cm:
+                vllm_build.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.dict(
+        os.environ, {"USE_LOCAL_BASE_IMAGE": "1", "BASE_IMAGE": "img:tag"}, clear=True
+    )
+    def test_params_missing_local_base_image_raises(self, _local_img):
+        with tempfile.TemporaryDirectory() as td:
+            os.chdir(td)
+            with self.assertRaises(ValueError) as cm:
+                vllm_build.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.dict(
+        os.environ,
+        {"USE_LOCAL_DOCKERFILE": "1", "DOCKERFILE_PATH": "Dockerfile"},
+        clear=True,
+    )
+    def test_params_missing_dockerfile_raises(self, _is_path):
+        with tempfile.TemporaryDirectory() as td:
+            os.chdir(td)
+            with self.assertRaises(ValueError) as cm:
+                vllm_build.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.dict(
+        os.environ,
+        {"OUTPUT_DIR": ""},
+        clear=True,
+    )
+    def test_params_missing_output_dir(self, _is_path):
+        with self.assertRaises(FileNotFoundError):
+            vllm_build.VllmBuildParameters()
+
+
+class TestBuildCmdAndRun(unittest.TestCase):
+    @patch(f"{_VLLM_BUILD_MODULE}.local_image_exists", return_value=True)
+    def test_generate_docker_build_cmd_includes_bits(self, _exists):
+        runner = vllm_build.VllmBuildRunner()
+        inputs = MagicMock()
+        inputs.output_dir = Path("/abs/out")
+        inputs.use_local_base_image = True
+        inputs.base_image = "img:tag"
+        inputs.torch_whls_path = Path("./vllm/tmp")
+        inputs.max_jobs = 64
+        inputs.cuda_version = "12.8.1"
+        inputs.python_version = "3.12"
+        inputs.sccache_bucket = "my-bucket"
+        inputs.sccache_region = "us-west-2"
+        inputs.torch_cuda_arch_list = "8.0;9.0"
+        inputs.target_stage = "export-wheels"
+        inputs.tag_name = "vllm-wheels"
+
+        cmd = runner._generate_docker_build_cmd(inputs)
+        squashed = " ".join(cmd.split())
+
+        self.assertIn("--output type=local,dest=/abs/out", squashed)
+        self.assertIn("-f docker/Dockerfile.nightly_torch", squashed)
+        self.assertIn("--pull=false", squashed)
+        self.assertIn("--build-arg TORCH_WHEELS_PATH=tmp", squashed)
+        self.assertIn("--build-arg BUILD_BASE_IMAGE=img:tag", squashed)
+        self.assertIn("--build-arg FINAL_BASE_IMAGE=img:tag", squashed)
+        self.assertIn("--build-arg max_jobs=64", squashed)
+        self.assertIn("--build-arg CUDA_VERSION=12.8.1", squashed)
+        self.assertIn("--build-arg PYTHON_VERSION=3.12", squashed)
+        self.assertIn("--build-arg USE_SCCACHE=1", squashed)
+        self.assertIn("--build-arg SCCACHE_BUCKET_NAME=my-bucket", squashed)
+        self.assertIn("--build-arg SCCACHE_REGION_NAME=us-west-2", squashed)
+        self.assertIn("--build-arg torch_cuda_arch_list='8.0;9.0'", squashed)
+        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.object(
+        vllm_build.VllmBuildRunner,
+        "_generate_docker_build_cmd",
+        return_value="docker buildx ...",
+    )
+    @patch.dict(
+        os.environ,
+        {
+            "USE_TORCH_WHEEL": "0",
+            "USE_LOCAL_BASE_IMAGE": "0",
+            "USE_LOCAL_DOCKERFILE": "0",
+            "OUTPUT_DIR": "shared",
+        },
+        clear=True,
+    )
+    def test_run_calls_clone_prepare_and_build(
+        self, mock_gen, mock_clone, mock_ensure, mock_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()
+            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()
+        _, kwargs = mock_run.call_args
+        assert kwargs.get("cwd") == "vllm"

.ci/magma/Makefile

@@ -16,6 +16,7 @@ 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
@@ -25,6 +26,12 @@ 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

@@ -28,6 +28,7 @@ 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
@@ -37,6 +38,7 @@ 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

@@ -0,0 +1,26 @@
+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

@@ -5,10 +5,6 @@ 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

@@ -138,28 +138,11 @@ fi
 
 echo "Calling setup.py bdist at $(date)"
 
-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 \
+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/$WHEELHOUSE_DIR
-    echo "Finished setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
-    echo "Calling setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
-    time EXTRA_CAFFE2_CMAKE_FLAGS=${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
@@ -272,10 +255,6 @@ 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 /$LIBTORCH_HOUSE_DIR
@@ -452,16 +431,8 @@ if [[ -z "$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

@@ -66,6 +66,9 @@ 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"
         ;;
@@ -110,13 +113,18 @@ DEPS_SONAME=(
 )
 
 
-# CUDA_VERSION 12.6, 12.8, 12.9
-if [[ $CUDA_VERSION == 12* ]]; then
+# CUDA_VERSION 12.*, 13.*
+if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
     export USE_STATIC_CUDNN=0
     # Try parallelizing nvcc as well
-    export TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
+    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
     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"
@@ -126,15 +134,11 @@ if [[ $CUDA_VERSION == 12* ]]; 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/extras/CUPTI/lib64/libcupti.so.12"
+            "/usr/local/cuda/lib64/libnvshmem_host.so.3"
             "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so"
         )
         DEPS_SONAME+=(
@@ -146,41 +150,83 @@ if [[ $CUDA_VERSION == 12* ]]; 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 != 12.9* ]]; then
-            DEPS_LIST+=("/usr/local/cuda/lib64/libnvToolsExt.so.1")
-            DEPS_SONAME+=("libnvToolsExt.so.1")
+        if [[ $CUDA_VERSION == 13* ]]; then
+            DEPS_LIST+=(
+                "/usr/local/cuda/lib64/libcublas.so.13"
+                "/usr/local/cuda/lib64/libcublasLt.so.13"
+                "/usr/local/cuda/lib64/libcudart.so.13"
+                "/usr/local/cuda/lib64/libnvrtc.so.13"
+                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13"
+                "/usr/local/cuda/lib64/libibverbs.so.1"
+                "/usr/local/cuda/lib64/librdmacm.so.1"
+                "/usr/local/cuda/lib64/libmlx5.so.1"
+                "/usr/local/cuda/lib64/libnl-3.so.200"
+                "/usr/local/cuda/lib64/libnl-route-3.so.200")
+            DEPS_SONAME+=(
+                "libcublas.so.13"
+                "libcublasLt.so.13"
+                "libcudart.so.13"
+                "libnvrtc.so.13"
+                "libcupti.so.13"
+                "libibverbs.so.1"
+                "librdmacm.so.1"
+                "libmlx5.so.1"
+                "libnl-3.so.200"
+                "libnl-route-3.so.200")
+            export USE_CUPTI_SO=1
+            export ATEN_STATIC_CUDA=0
+            export USE_CUDA_STATIC_LINK=0
+            export USE_CUFILE=0
+        else
+            DEPS_LIST+=(
+                "/usr/local/cuda/lib64/libnvToolsExt.so.1"
+                "/usr/local/cuda/lib64/libcublas.so.12"
+                "/usr/local/cuda/lib64/libcublasLt.so.12"
+                "/usr/local/cuda/lib64/libcudart.so.12"
+                "/usr/local/cuda/lib64/libnvrtc.so.12"
+                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12")
+            DEPS_SONAME+=(
+                "libnvToolsExt.so.1"
+                "libcublas.so.12"
+                "libcublasLt.so.12"
+                "libcudart.so.12"
+                "libnvrtc.so.12"
+                "libcupti.so.12")
         fi
     else
         echo "Using nvidia libs from pypi."
         CUDA_RPATHS=(
-            '$ORIGIN/../../nvidia/cublas/lib'
-            '$ORIGIN/../../nvidia/cuda_cupti/lib'
-            '$ORIGIN/../../nvidia/cuda_nvrtc/lib'
-            '$ORIGIN/../../nvidia/cuda_runtime/lib'
             '$ORIGIN/../../nvidia/cudnn/lib'
-            '$ORIGIN/../../nvidia/cufft/lib'
-            '$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'
+            '$ORIGIN/../../nvidia/nccl/lib'
+            '$ORIGIN/../../nvidia/cusparselt/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_rocm.sh

@@ -194,7 +194,7 @@ ROCBLAS_LIB_SRC=$ROCM_HOME/lib/rocblas/library
 ROCBLAS_LIB_DST=lib/rocblas/library
 ROCBLAS_ARCH_SPECIFIC_FILES=$(ls $ROCBLAS_LIB_SRC | grep -E $ARCH)
 ROCBLAS_OTHER_FILES=$(ls $ROCBLAS_LIB_SRC | grep -v gfx)
-ROCBLAS_LIB_FILES=($ROCBLAS_ARCH_SPECIFIC_FILES $OTHER_FILES)
+ROCBLAS_LIB_FILES=($ROCBLAS_ARCH_SPECIFIC_FILES $ROCBLAS_OTHER_FILES)
 
 # hipblaslt library files
 HIPBLASLT_LIB_SRC=$ROCM_HOME/lib/hipblaslt/library

.ci/manywheel/build_xpu.sh

@@ -25,6 +25,7 @@ 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

@@ -50,9 +50,6 @@ 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
@@ -95,6 +92,27 @@ 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)
@@ -155,6 +173,7 @@ 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
@@ -176,8 +195,16 @@ fi
 
 # We only build FlashAttention files for CUDA 8.0+, and they require large amounts of
 # memory to build and will OOM
-if [[ "$BUILD_ENVIRONMENT" == *cuda* ]] && [[ 1 -eq $(echo "${TORCH_CUDA_ARCH_LIST} >= 8.0" | bc) ]]; then
-  export BUILD_CUSTOM_STEP="ninja -C build flash_attention -j 2"
+
+if [[ "$BUILD_ENVIRONMENT" == *cuda* ]] && echo "${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}"
 fi
 
 if [[ "${BUILD_ENVIRONMENT}" == *clang* ]]; then
@@ -192,7 +219,6 @@ if [[ "$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
@@ -213,7 +239,7 @@ 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* && -d /var/lib/jenkins/workspace ]]; then
+if [[ "$BUILD_ENVIRONMENT" != *rocm* && "$BUILD_ENVIRONMENT" != *s390x* && "$BUILD_ENVIRONMENT" != *riscv64* && -d /var/lib/jenkins/workspace ]]; then
   # Workaround for dind-rootless userid mapping (https://github.com/pytorch/ci-infra/issues/96)
   WORKSPACE_ORIGINAL_OWNER_ID=$(stat -c '%u' "/var/lib/jenkins/workspace")
   cleanup_workspace() {
@@ -258,29 +284,19 @@ 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* ]]; then
+    if [[ "$BUILD_ENVIRONMENT" != *rocm*  && "$BUILD_ENVIRONMENT" != *xla* && "$BUILD_ENVIRONMENT" != *riscv64* ]]; 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
 
-      if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
-        python3 tools/packaging/split_wheel.py bdist_wheel
-      else
-        WERROR=1 python setup.py bdist_wheel
-      fi
+      WERROR=1 python setup.py bdist_wheel
     else
       python setup.py clean
       if [[ "$BUILD_ENVIRONMENT" == *xla* ]]; then
         source .ci/pytorch/install_cache_xla.sh
       fi
-      if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
-        echo "USE_SPLIT_BUILD cannot be used with xla or rocm"
-        exit 1
-      else
-        python setup.py bdist_wheel
-      fi
+      python setup.py bdist_wheel
     fi
     pip_install_whl "$(echo dist/*.whl)"
 
@@ -405,7 +421,7 @@ 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 as they don't use sccache
-if [[ "$BUILD_ENVIRONMENT" != *s390x* && "$BUILD_ENVIRONMENT" != *-bazel-* ]]; then
+# 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
   print_sccache_stats
 fi

.ci/pytorch/check_binary.sh

@@ -67,7 +67,7 @@ fi
 #       wheels with cxx11-abi
 
 echo "Checking that the gcc ABI is what we expect"
-if [[ "$(uname)" != 'Darwin' ]]; then
+if [[ "$(uname)" != 'Darwin' &&  "$(uname -m)" != "s390x" ]]; then
   # We also check that there are cxx11 symbols in libtorch
   #
   echo "Checking that symbols in libtorch.so have the right gcc abi"
@@ -300,24 +300,3 @@ except RuntimeError as e:
     exit 1
   fi
 fi
-
-###############################################################################
-# Check for C++ ABI compatibility to GCC-11 - GCC 13
-###############################################################################
-if [[ "$(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

@@ -149,6 +149,19 @@ 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)
@@ -229,7 +242,6 @@ 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
@@ -245,7 +257,7 @@ 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}"
+      git checkout "${fbgemm_commit}" --recurse-submodules
       python setup.py bdist_wheel \
         --build-variant=rocm \
         -DHIP_ROOT_DIR="${ROCM_PATH}" \
@@ -264,7 +276,6 @@ function install_torchrec_and_fbgemm() {
     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
@@ -273,7 +284,7 @@ function install_torchrec_and_fbgemm() {
 
 function clone_pytorch_xla() {
   if [[ ! -d ./xla ]]; then
-    git clone --recursive --quiet https://github.com/pytorch/xla.git
+    git clone --recursive -b r2.9 https://github.com/pytorch/xla.git
     pushd xla
     # pin the xla hash so that we don't get broken by changes to xla
     git checkout "$(cat ../.github/ci_commit_pins/xla.txt)"
@@ -283,30 +294,6 @@ 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/cpp_doc_push_script.sh

@@ -58,7 +58,7 @@ time python tools/setup_helpers/generate_code.py \
 
 # Build the docs
 pushd docs/cpp
-time make VERBOSE=1 html -j
+time make VERBOSE=1 html
 
 popd
 popd

.ci/pytorch/macos-test.sh

@@ -157,6 +157,34 @@ 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
@@ -167,7 +195,7 @@ torchbench_setup_macos() {
   git checkout "$(cat ../.github/ci_commit_pins/vision.txt)"
   git submodule update --init --recursive
   python setup.py clean
-  python setup.py develop
+  python -m pip install -e . -v --no-build-isolation
   popd
 
   pushd torchaudio
@@ -176,11 +204,9 @@ 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 setup.py develop
+  USE_OPENMP=0 python -m pip install -e . -v --no-build-isolation
   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
 }
 
@@ -276,6 +302,47 @@ 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"
 }
@@ -324,6 +391,8 @@ 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

@@ -45,6 +45,7 @@ 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

@@ -0,0 +1,25 @@
+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/check_binary_symbols.py

@@ -32,6 +32,9 @@ LIBTORCH_NAMESPACE_LIST = (
     "torch::",
 )
 
+# Patterns for detecting statically linked libstdc++ symbols
+STATICALLY_LINKED_CXX11_ABI = [re.compile(r".*recursive_directory_iterator.*")]
+
 
 def _apply_libtorch_symbols(symbols):
     return [
@@ -53,12 +56,17 @@ def get_symbols(lib: str) -> list[tuple[str, str, str]]:
     return [x.split(" ", 2) for x in lines.decode("latin1").split("\n")[:-1]]
 
 
-def grep_symbols(lib: str, patterns: list[Any]) -> list[str]:
+def grep_symbols(
+    lib: str, patterns: list[Any], symbol_type: str | None = None
+) -> list[str]:
     def _grep_symbols(
         symbols: list[tuple[str, str, str]], patterns: list[Any]
     ) -> list[str]:
         rc = []
         for _s_addr, _s_type, s_name in symbols:
+            # Filter by symbol type if specified
+            if symbol_type and _s_type != symbol_type:
+                continue
             for pattern in patterns:
                 if pattern.match(s_name):
                     rc.append(s_name)
@@ -80,6 +88,18 @@ def grep_symbols(lib: str, patterns: list[Any]) -> list[str]:
         return functools.reduce(list.__add__, (x.result() for x in tasks), [])
 
 
+def check_lib_statically_linked_libstdc_cxx_abi_symbols(lib: str) -> None:
+    cxx11_statically_linked_symbols = grep_symbols(
+        lib, STATICALLY_LINKED_CXX11_ABI, symbol_type="T"
+    )
+    num_statically_linked_symbols = len(cxx11_statically_linked_symbols)
+    print(f"num_statically_linked_symbols (T): {num_statically_linked_symbols}")
+    if num_statically_linked_symbols > 0:
+        raise RuntimeError(
+            f"Found statically linked libstdc++ symbols (recursive_directory_iterator): {cxx11_statically_linked_symbols[:100]}"
+        )
+
+
 def check_lib_symbols_for_abi_correctness(lib: str) -> None:
     print(f"lib: {lib}")
     cxx11_symbols = grep_symbols(lib, LIBTORCH_CXX11_PATTERNS)
@@ -107,6 +127,7 @@ def main() -> None:
 
     libtorch_cpu_path = str(install_root / "lib" / "libtorch_cpu.so")
     check_lib_symbols_for_abi_correctness(libtorch_cpu_path)
+    check_lib_statically_linked_libstdc_cxx_abi_symbols(libtorch_cpu_path)
 
 
 if __name__ == "__main__":

.ci/pytorch/test.sh

@@ -32,6 +32,16 @@ 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
 
@@ -91,6 +101,7 @@ 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.
@@ -495,6 +506,14 @@ 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
@@ -627,6 +646,8 @@ test_perf_for_dashboard() {
     device=cuda_a10g
   elif [[ "${TEST_CONFIG}" == *h100* ]]; then
     device=cuda_h100
+  elif [[ "${TEST_CONFIG}" == *b200* ]]; then
+    device=cuda_b200
   elif [[ "${TEST_CONFIG}" == *rocm* ]]; then
     device=rocm
   fi
@@ -801,6 +822,16 @@ test_dynamo_benchmark() {
   if [[ "${TEST_CONFIG}" == *perf_compare* ]]; then
     test_single_dynamo_benchmark "training" "$suite" "$shard_id" --training --amp "$@"
   elif [[ "${TEST_CONFIG}" == *perf* ]]; then
+    # TODO (huydhn): Just smoke test some sample models
+    if [[ "${TEST_CONFIG}" == *b200* ]]; then
+      if [[ "${suite}" == "huggingface" ]]; then
+        export TORCHBENCH_ONLY_MODELS="DistillGPT2"
+      elif [[ "${suite}" == "timm_models" ]]; then
+        export TORCHBENCH_ONLY_MODELS="inception_v3"
+      elif [[ "${suite}" == "torchbench" ]]; then
+        export TORCHBENCH_ONLY_MODELS="hf_Bert"
+      fi
+    fi
     test_single_dynamo_benchmark "dashboard" "$suite" "$shard_id" "$@"
   else
     if [[ "${TEST_CONFIG}" == *cpu* ]]; then
@@ -1039,20 +1070,10 @@ 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
@@ -1603,6 +1624,25 @@ test_operator_benchmark() {
       --expected "expected_ci_operator_benchmark_eager_float32_cpu.csv"
 }
 
+test_operator_microbenchmark() {
+  TEST_REPORTS_DIR=$(pwd)/test/test-reports
+  mkdir -p "$TEST_REPORTS_DIR"
+  TEST_DIR=$(pwd)
+
+  cd benchmarks/operator_benchmark/pt_extension
+  python -m pip install .
+
+  cd "${TEST_DIR}"/benchmarks/operator_benchmark
+
+  for OP_BENCHMARK_TESTS in matmul mm addmm bmm; do
+    $TASKSET python -m pt.${OP_BENCHMARK_TESTS}_test --tag-filter long \
+      --output-json-for-dashboard "${TEST_REPORTS_DIR}/operator_microbenchmark_${OP_BENCHMARK_TESTS}_compile.json" \
+      --benchmark-name "PyTorch operator microbenchmark" --use-compile
+    $TASKSET python -m pt.${OP_BENCHMARK_TESTS}_test --tag-filter long \
+      --output-json-for-dashboard "${TEST_REPORTS_DIR}/operator_microbenchmark_${OP_BENCHMARK_TESTS}.json" \
+      --benchmark-name "PyTorch operator microbenchmark"
+  done
+}
 
 if ! [[ "${BUILD_ENVIRONMENT}" == *libtorch* || "${BUILD_ENVIRONMENT}" == *-bazel-* ]]; then
   (cd test && python -c "import torch; print(torch.__config__.show())")
@@ -1627,6 +1667,10 @@ elif [[ "${TEST_CONFIG}" == *xla* ]]; then
   install_torchvision
   build_xla
   test_xla
+elif [[ "$TEST_CONFIG" == *vllm* ]]; then
+    echo "vLLM CI uses TORCH_CUDA_ARCH_LIST: $TORCH_CUDA_ARCH_LIST"
+    (cd .ci/lumen_cli && python -m pip install -e .)
+    python -m cli.run test external vllm --test-plan "$TEST_CONFIG" --shard-id "$SHARD_NUMBER" --num-shards "$NUM_TEST_SHARDS"
 elif [[ "${TEST_CONFIG}" == *executorch* ]]; then
   test_executorch
 elif [[ "$TEST_CONFIG" == 'jit_legacy' ]]; then
@@ -1653,6 +1697,8 @@ elif [[ "${TEST_CONFIG}" == *operator_benchmark* ]]; then
     test_operator_benchmark cpu ${TEST_MODE}
 
   fi
+elif [[ "${TEST_CONFIG}" == *operator_microbenchmark* ]]; then
+  test_operator_microbenchmark
 elif [[ "${TEST_CONFIG}" == *inductor_distributed* ]]; then
   test_inductor_distributed
 elif [[ "${TEST_CONFIG}" == *inductor-halide* ]]; then
@@ -1672,54 +1718,40 @@ elif [[ "${TEST_CONFIG}" == *timm* ]]; then
 elif [[ "${TEST_CONFIG}" == cachebench ]]; then
   install_torchaudio
   install_torchvision
-  checkout_install_torchbench nanogpt BERT_pytorch resnet50 hf_T5 llama moco
-  PYTHONPATH=$(pwd)/torchbench test_cachebench
+  PYTHONPATH=/torchbench test_cachebench
 elif [[ "${TEST_CONFIG}" == verify_cachebench ]]; then
   install_torchaudio
   install_torchvision
-  checkout_install_torchbench nanogpt
-  PYTHONPATH=$(pwd)/torchbench test_verify_cachebench
+  PYTHONPATH=/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 [[ "${TEST_CONFIG}" == *inductor_torchbench_smoketest_perf* ]]; then
-    checkout_install_torchbench hf_Bert hf_Albert timm_vision_transformer
-    PYTHONPATH=$(pwd)/torchbench test_inductor_torchbench_smoketest_perf
+    PYTHONPATH=/torchbench test_inductor_torchbench_smoketest_perf
   elif [[ "${TEST_CONFIG}" == *inductor_torchbench_cpu_smoketest_perf* ]]; then
-    checkout_install_torchbench timm_vision_transformer phlippe_densenet basic_gnn_edgecnn \
-      llama_v2_7b_16h resnet50 timm_efficientnet mobilenet_v3_large timm_resnest \
-      functorch_maml_omniglot yolov3 mobilenet_v2 resnext50_32x4d densenet121 mnasnet1_0
-    PYTHONPATH=$(pwd)/torchbench test_inductor_torchbench_cpu_smoketest_perf
+    PYTHONPATH=/torchbench test_inductor_torchbench_cpu_smoketest_perf
   elif [[ "${TEST_CONFIG}" == *torchbench_gcp_smoketest* ]]; then
-    checkout_install_torchbench
-    TORCHBENCHPATH=$(pwd)/torchbench test_torchbench_gcp_smoketest
+    TORCHBENCHPATH=/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 [[ "${TEST_CONFIG}" != *cpu* ]]; then
       install_torchrec_and_fbgemm
     fi
-    PYTHONPATH=$(pwd)/torchbench test_dynamo_benchmark torchbench "$id"
+    PYTHONPATH=/torchbench test_dynamo_benchmark torchbench "$id"
   fi
 elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
   install_torchvision
-  PYTHONPATH=$(pwd)/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
+  PYTHONPATH=/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
   if [[ "$SHARD_NUMBER" -eq "1" ]]; then
     test_inductor_aoti
   fi
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision
   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

@@ -61,9 +61,10 @@ 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. Only have MTL self-hosted runner now
-  SET TORCH_XPU_ARCH_LIST=xe-lpg
-  SET USE_KINETO=0
+  :: Reduce build time
+  SET TORCH_XPU_ARCH_LIST=bmg
+  :: Re-setup python env for build
+  call pip install -r requirements.txt
 )
 
 @echo on
@@ -136,7 +137,7 @@ 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_PARENT_DIR%\Miniconda3\Scripts\activate.bat %CONDA_PARENT_DIR%\Miniconda3` 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_ROOT_DIR%\Scripts\activate.bat %CONDA_ROOT_DIR%\envs\py_tmp` 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_PARENT_DIR%\Miniconda3 (
+if not exist %CONDA_ROOT_DIR% (
   set INSTALL_FRESH_CONDA=1
 )
 
@@ -17,10 +17,14 @@ 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_PARENT_DIR%\Miniconda3
+  %TMP_DIR_WIN%\Miniconda3-latest-Windows-x86_64.exe /InstallationType=JustMe /RegisterPython=0 /S /AddToPath=0 /D=%CONDA_ROOT_DIR%
   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_PARENT_DIR%\Miniconda3\Scripts\activate.bat %CONDA_PARENT_DIR%\Miniconda3
+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

.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_PARENT_DIR%\Miniconda3\Lib\site-packages\torch %TMP_DIR_WIN%\build\torch\
+xcopy /s %CONDA_ROOT_DIR%\envs\py_tmp\Lib\site-packages\torch %TMP_DIR_WIN%\build\torch\
 
 pushd .
 if "%VC_VERSION%" == "" (

.ci/pytorch/win-test.sh

@@ -38,13 +38,20 @@ if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
 fi
 
 # TODO: Move both of them to Windows AMI
-python -m pip install pytest-rerunfailures==10.3 pytest-cpp==2.3.0 tensorboard==2.13.0 protobuf==5.29.4 pytest-subtests==0.13.1
+python -m pip install 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"
 
 # 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==0.3.30
+python -m pip install tlparse==0.4.0
 
 # Install parameterized
 python -m pip install parameterized==0.8.1
@@ -52,9 +59,6 @@ python -m pip install parameterized==0.8.1
 # Install pulp for testing ilps under torch\distributed\_tools
 python -m pip install pulp==2.9.0
 
-# Install expecttest to merge https://github.com/pytorch/pytorch/pull/155308
-python -m pip install expecttest==0.3.0
-
 run_tests() {
     # Run nvidia-smi if available
     for path in '/c/Program Files/NVIDIA Corporation/NVSMI/nvidia-smi.exe' /c/Windows/System32/nvidia-smi.exe; do

.ci/pytorch/windows/cuda126.bat

@@ -37,7 +37,7 @@ IF "%CUDA_PATH_V126%"=="" (
 )
 
 IF "%BUILD_VISION%" == "" (
-    set TORCH_CUDA_ARCH_LIST=6.1;7.0;7.5;8.0;8.6;9.0
+    set TORCH_CUDA_ARCH_LIST=5.0;6.0;6.1;7.0;7.5;8.0;8.6;9.0
     set TORCH_NVCC_FLAGS=-Xfatbin -compress-all
 ) ELSE (
     set NVCC_FLAGS=-D__CUDA_NO_HALF_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_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/cuda128.bat

@@ -37,10 +37,10 @@ IF "%CUDA_PATH_V128%"=="" (
 )
 
 IF "%BUILD_VISION%" == "" (
-    set TORCH_CUDA_ARCH_LIST=6.1;7.0;7.5;8.0;8.6;9.0;10.0;12.0
+    set TORCH_CUDA_ARCH_LIST=7.0;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_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 -gencode=arch=compute_100,code=compute_100 -gencode=arch=compute_120,code=compute_120
+    set NVCC_FLAGS=-D__CUDA_NO_HALF_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_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 -gencode=arch=compute_100,code=compute_100 -gencode=arch=compute_120,code=compute_120
 )
 
 set "CUDA_PATH=%CUDA_PATH_V128%"

.ci/pytorch/windows/cuda130.bat

@@ -0,0 +1,59 @@
+@echo off
+
+set MODULE_NAME=pytorch
+
+IF NOT EXIST "setup.py" IF NOT EXIST "%MODULE_NAME%" (
+    call internal\clone.bat
+    cd %~dp0
+) ELSE (
+    call internal\clean.bat
+)
+IF ERRORLEVEL 1 goto :eof
+
+call internal\check_deps.bat
+IF ERRORLEVEL 1 goto :eof
+
+REM Check for optional components
+
+set USE_CUDA=
+set CMAKE_GENERATOR=Visual Studio 15 2017 Win64
+
+IF "%NVTOOLSEXT_PATH%"=="" (
+    IF EXIST "C:\Program Files\NVIDIA Corporation\NvToolsExt\lib\x64\nvToolsExt64_1.lib"  (
+        set NVTOOLSEXT_PATH=C:\Program Files\NVIDIA Corporation\NvToolsExt
+    ) ELSE (
+        echo NVTX ^(Visual Studio Extension ^for CUDA^) ^not installed, failing
+        exit /b 1
+    )
+)
+
+IF "%CUDA_PATH_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

@@ -1,12 +1,20 @@
-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
+
+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\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
 
@@ -20,8 +28,3 @@ 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

@@ -26,6 +26,7 @@ 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
@@ -113,6 +114,33 @@ 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=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
+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
 if errorlevel 1 exit /b 1
 
-start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe -s -noreboot
+start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe -s -noreboot
 if errorlevel 1 exit /b 1
 
-del %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe || ver > NUL
+del %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe || ver > NUL

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

@@ -1,12 +1,22 @@
 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 DESIRED_PYTHON not defined, Python version is set to %DESIRED_PYTHON%
+    echo 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/9d6d6c17-ca2d-4735-9331-99447e4a1280/intel-deep-learning-essentials-2025.0.1.28_offline.exe
+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_PRODUCT_NAME=intel.oneapi.win.deep-learning-essentials.product
-set XPU_BUNDLE_VERSION=2025.0.1+20
+set XPU_BUNDLE_VERSION=2025.1.3+5
 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%]==[2025.1] (
-    set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/75d4eb97-914a-4a95-852c-7b9733d80f74/intel-deep-learning-essentials-2025.1.3.8_offline.exe
-    set XPU_BUNDLE_VERSION=2025.1.3+5
+if not [%XPU_VERSION%]==[] if [%XPU_VERSION%]==[2025.2] (
+    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
 )
 
 :: Check if XPU bundle is target version or already installed
@@ -90,14 +90,3 @@ 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

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

.ci/wheel/build_wheel.sh

@@ -124,20 +124,31 @@ popd
 
 export TH_BINARY_BUILD=1
 export INSTALL_TEST=0 # dont install test binaries into site-packages
-export MACOSX_DEPLOYMENT_TARGET=10.15
+export MACOSX_DEPLOYMENT_TARGET=11.0
 export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
 
-SETUPTOOLS_PINNED_VERSION="==70.1.0"
-PYYAML_PINNED_VERSION="=5.3"
 EXTRA_CONDA_INSTALL_FLAGS=""
 CONDA_ENV_CREATE_FLAGS=""
 RENAME_WHEEL=true
 case $desired_python in
+    3.14t)
+        echo "Using 3.14 deps"
+        NUMPY_PINNED_VERSION="==2.1.0"
+        CONDA_ENV_CREATE_FLAGS="python-freethreading"
+        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+        desired_python="3.14.0rc1"
+        RENAME_WHEEL=false
+        ;;
+    3.14)
+        echo "Using 3.14t deps"
+        NUMPY_PINNED_VERSION="==2.1.0"
+        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+        desired_python="3.14.0rc1"
+        RENAME_WHEEL=false
+        ;;
     3.13t)
         echo "Using 3.13 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=6.0.1"
-        NUMPY_PINNED_VERSION="=2.1.0"
+        NUMPY_PINNED_VERSION="==2.1.0"
         CONDA_ENV_CREATE_FLAGS="python-freethreading"
         EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
         desired_python="3.13"
@@ -145,37 +156,23 @@ case $desired_python in
         ;;
     3.13)
         echo "Using 3.13 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=6.0.1"
-        NUMPY_PINNED_VERSION="=2.1.0"
+        NUMPY_PINNED_VERSION="==2.1.0"
         ;;
     3.12)
         echo "Using 3.12 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=6.0.1"
-        NUMPY_PINNED_VERSION="=2.0.2"
+        NUMPY_PINNED_VERSION="==2.0.2"
         ;;
     3.11)
         echo "Using 3.11 deps"
-        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
-        PYYAML_PINNED_VERSION=">=5.3"
-        NUMPY_PINNED_VERSION="=2.0.2"
+        NUMPY_PINNED_VERSION="==2.0.2"
         ;;
     3.10)
         echo "Using 3.10 deps"
-        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"
+        NUMPY_PINNED_VERSION="==2.0.2"
         ;;
     *)
-        echo "Using default deps"
-        NUMPY_PINNED_VERSION="=1.11.3"
+        echo "Unsupported version $desired_python"
+        exit 1
         ;;
 esac
 
@@ -184,17 +181,17 @@ 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
+PINNED_PACKAGES=(
+    "numpy${NUMPY_PINNED_VERSION}"
+)
+retry pip install "${PINNED_PACKAGES[@]}" -r "${pytorch_rootdir}/requirements-build.txt"
+pip install requests ninja typing-extensions
 retry pip install -r "${pytorch_rootdir}/requirements.txt" || true
 retry brew install libomp
 
 # For USE_DISTRIBUTED=1 on macOS, need libuv, which is build as part of tensorpipe submodule
 export USE_DISTRIBUTED=1
 
-if [[ -n "$CROSS_COMPILE_ARM64" ]]; then
-    export CMAKE_OSX_ARCHITECTURES=arm64
-fi
 export USE_MKLDNN=OFF
 export USE_QNNPACK=OFF
 export BUILD_TEST=OFF
@@ -202,16 +199,7 @@ export BUILD_TEST=OFF
 pushd "$pytorch_rootdir"
 echo "Calling setup.py bdist_wheel at $(date)"
 
-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
+python setup.py bdist_wheel -d "$whl_tmp_dir" --plat-name ${mac_version}
 
 echo "Finished setup.py bdist_wheel at $(date)"
 

.circleci/scripts/binary_linux_test.sh

@@ -65,16 +65,8 @@ fi
 
 if [[ "$PACKAGE_TYPE" != libtorch ]]; then
   if [[ "\$BUILD_ENVIRONMENT" != *s390x* ]]; then
-    if [[ "$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
+    pip install "\$pkg" --index-url "https://download.pytorch.org/whl/\${CHANNEL}/${DESIRED_CUDA}"
+    retry pip install -q numpy protobuf typing-extensions
   else
     pip install "\$pkg"
     retry pip install -q numpy protobuf typing-extensions

.circleci/scripts/binary_populate_env.sh

@@ -71,14 +71,7 @@ export PYTORCH_BUILD_NUMBER=1
 
 # Set triton version as part of PYTORCH_EXTRA_INSTALL_REQUIREMENTS
 TRITON_VERSION=$(cat $PYTORCH_ROOT/.ci/docker/triton_version.txt)
-
-# Here PYTORCH_EXTRA_INSTALL_REQUIREMENTS is already set for the all the wheel builds hence append TRITON_CONSTRAINT
-TRITON_CONSTRAINT="platform_system == 'Linux' and platform_machine == 'x86_64'"
-
-# CUDA 12.9 builds have triton for Linux and Linux aarch64 binaries.
-if [[ "$DESIRED_CUDA" == "cu129" ]]; then
-  TRITON_CONSTRAINT="platform_system == 'Linux'"
-fi
+TRITON_CONSTRAINT="platform_system == 'Linux'"
 
 if [[ "$PACKAGE_TYPE" =~ .*wheel.* &&  -n "${PYTORCH_EXTRA_INSTALL_REQUIREMENTS:-}" && ! "$PYTORCH_BUILD_VERSION" =~ .*xpu.* ]]; then
   TRITON_REQUIREMENT="triton==${TRITON_VERSION}; ${TRITON_CONSTRAINT}"
@@ -134,7 +127,6 @@ export DESIRED_PYTHON="${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

@@ -23,10 +23,6 @@ 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"
@@ -55,16 +51,12 @@ 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}" ${cache_control_flag}
+          --metadata "checksum-sha256=${shm_id}"
       )
     done
   )

.circleci/scripts/binary_windows_build.sh

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