[DTensor] Assert DTensorSpec has valid placements (#158133)

Summary:
This helped identify buggy sharding rules during debugging, why not
check it in.

Test Plan:
contbuild & OSS CI

Rollback Plan:

Differential Revision: D78929245

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -208,9 +208,7 @@ 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 "
-        # nvshmem is broken for aarch64 see https://github.com/pytorch/pytorch/issues/160425
-        build_vars += "USE_NVSHMEM=OFF "
+        build_vars = "MAX_JOBS=5 " + build_vars
 
     override_package_version = os.getenv("OVERRIDE_PACKAGE_VERSION")
     desired_cuda = os.getenv("DESIRED_CUDA")

.ci/aarch64_linux/build_aarch64_wheel.py

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

.ci/docker/README.md

@@ -104,6 +104,7 @@ If your new Docker image needs a library installed from a specific pinned commit
    ```bash
    pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-new1)
      CUDA_VERSION=12.8.1
+     CUDNN_VERSION=9
      ANACONDA_PYTHON_VERSION=3.12
      GCC_VERSION=11
      VISION=yes

.ci/docker/build.sh

@@ -76,9 +76,6 @@ 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
@@ -96,6 +93,7 @@ tag=$(echo $image | awk -F':' '{print $2}')
 case "$tag" in
   pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11)
     CUDA_VERSION=12.4
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
     VISION=yes
@@ -106,6 +104,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
     VISION=yes
@@ -116,6 +115,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=9
     VISION=yes
@@ -127,6 +127,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.12
     GCC_VERSION=9
     VISION=yes
@@ -138,6 +139,7 @@ case "$tag" in
     ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.13
     GCC_VERSION=9
     VISION=yes
@@ -147,8 +149,20 @@ case "$tag" in
     TRITON=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9)
+    CUDA_VERSION=12.6.3
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-vllm)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.12
     GCC_VERSION=11
     VISION=yes
@@ -157,8 +171,45 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
     ;;
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.10
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3.12-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.12
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-jammy-cuda12.6-cudnn9-py3.13-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.6
+    CUDNN_VERSION=9
+    ANACONDA_PYTHON_VERSION=3.13
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9)
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=9
     VISION=yes
@@ -179,7 +230,19 @@ case "$tag" in
     VISION=yes
     TRITON=yes
     ;;
-  pytorch-linux-jammy-rocm-n-py3 | pytorch-linux-jammy-rocm-n-py3-benchmarks | pytorch-linux-noble-rocm-n-py3)
+  pytorch-linux-jammy-py3.11-clang12)
+    ANACONDA_PYTHON_VERSION=3.11
+    CLANG_VERSION=12
+    VISION=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-py3.9-gcc9)
+    ANACONDA_PYTHON_VERSION=3.9
+    GCC_VERSION=9
+    VISION=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-rocm-n-py3 | pytorch-linux-noble-rocm-n-py3)
     if [[ $tag =~ "jammy" ]]; then
       ANACONDA_PYTHON_VERSION=3.10
     else
@@ -193,9 +256,7 @@ case "$tag" in
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
-    if [[ $tag =~ "benchmarks" ]]; then
-      INDUCTOR_BENCHMARKS=yes
-    fi
+    INDUCTOR_BENCHMARKS=yes
     ;;
   pytorch-linux-noble-rocm-alpha-py3)
     ANACONDA_PYTHON_VERSION=3.12
@@ -207,6 +268,7 @@ case "$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)
@@ -237,6 +299,7 @@ case "$tag" in
   pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12)
     ANACONDA_PYTHON_VERSION=3.9
     CUDA_VERSION=12.8.1
+    CUDNN_VERSION=9
     CLANG_VERSION=12
     VISION=yes
     TRITON=yes
@@ -306,9 +369,6 @@ case "$tag" in
     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
@@ -318,6 +378,7 @@ case "$tag" in
     fi
     if [[ "$image" == *cuda* ]]; then
       extract_version_from_image_name cuda CUDA_VERSION
+      extract_version_from_image_name cudnn CUDNN_VERSION
     fi
     if [[ "$image" == *rocm* ]]; then
       extract_version_from_image_name rocm ROCM_VERSION
@@ -369,6 +430,9 @@ docker build \
        --build-arg "PYTHON_VERSION=${PYTHON_VERSION}" \
        --build-arg "GCC_VERSION=${GCC_VERSION}" \
        --build-arg "CUDA_VERSION=${CUDA_VERSION}" \
+       --build-arg "CUDNN_VERSION=${CUDNN_VERSION}" \
+       --build-arg "TENSORRT_VERSION=${TENSORRT_VERSION}" \
+       --build-arg "GRADLE_VERSION=${GRADLE_VERSION}" \
        --build-arg "NINJA_VERSION=${NINJA_VERSION:-}" \
        --build-arg "KATEX=${KATEX:-}" \
        --build-arg "ROCM_VERSION=${ROCM_VERSION:-}" \
@@ -429,14 +493,7 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
 fi
 
 if [ -n "$GCC_VERSION" ]; then
-  if [[ "$image" == *riscv* ]]; then
-    # Check RISC-V cross-compilation toolchain version
-    if !(drun riscv64-linux-gnu-gcc-${GCC_VERSION} --version 2>&1 | grep -q " $GCC_VERSION\\W"); then
-      echo "RISC-V GCC_VERSION=$GCC_VERSION, but:"
-      drun riscv64-linux-gnu-gcc-${GCC_VERSION} --version
-      exit 1
-    fi
-  elif !(drun gcc --version 2>&1 | grep -q " $GCC_VERSION\\W"); then
+  if !(drun gcc --version 2>&1 | grep -q " $GCC_VERSION\\W"); then
     echo "GCC_VERSION=$GCC_VERSION, but:"
     drun gcc --version
     exit 1

.ci/docker/ci_commit_pins/huggingface.txt

@@ -1 +1 @@
-v4.54.0
+243e186efbf7fb93328dd6b34927a4e8c8f24395

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-f7888497a1eb9e98d4c07537f0d0bcfe180d1363
+11ec6354315768a85da41032535e3b7b99c5f706

.ci/docker/common/install_cpython.sh

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

.ci/docker/common/install_cuda.sh

@@ -68,8 +68,8 @@ function install_nvshmem {
   # download, unpack, install
   wget -q "${url}"
   tar xf "${filename}.tar.gz"
-  cp -a "libnvshmem/include/"* /usr/local/cuda/include/
-  cp -a "libnvshmem/lib/"*     /usr/local/cuda/lib64/
+  cp -a "libnvshmem/include/"* /usr/local/include/
+  cp -a "libnvshmem/lib/"*     /usr/local/lib/
 
   # cleanup
   cd ..

.ci/docker/common/install_cudnn.sh

@@ -0,0 +1,26 @@
+#!/bin/bash
+
+if [[ -n "${CUDNN_VERSION}" ]]; then
+    # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
+    mkdir tmp_cudnn
+    pushd tmp_cudnn
+    if [[ ${CUDA_VERSION:0:4} == "12.9" || ${CUDA_VERSION:0:4} == "12.8" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.10.2.21_cuda12-archive"
+    elif [[ ${CUDA_VERSION:0:4} == "12.6" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.10.2.21_cuda12-archive"
+    elif [[ ${CUDA_VERSION:0:4} == "12.4" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.10.2.21_cuda12-archive"
+    elif [[ ${CUDA_VERSION:0:2} == "11" ]]; then
+        CUDNN_NAME="cudnn-linux-x86_64-9.1.0.70_cuda11-archive"
+    else
+        print "Unsupported CUDA version ${CUDA_VERSION}"
+        exit 1
+    fi
+    curl --retry 3 -OLs https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-x86_64/${CUDNN_NAME}.tar.xz
+    tar xf ${CUDNN_NAME}.tar.xz
+    cp -a ${CUDNN_NAME}/include/* /usr/local/cuda/include/
+    cp -a ${CUDNN_NAME}/lib/* /usr/local/cuda/lib64/
+    popd
+    rm -rf tmp_cudnn
+    ldconfig
+fi

.ci/docker/common/install_inductor_benchmark_deps.sh

@@ -15,37 +15,11 @@ function install_timm() {
   commit=$(get_pinned_commit timm)
 
   pip_install "git+https://github.com/huggingface/pytorch-image-models@${commit}"
-}
-
-function install_torchbench() {
-  local commit
-  commit=$(get_pinned_commit torchbench)
-  git clone https://github.com/pytorch/benchmark torchbench
-  pushd torchbench
-  git checkout "$commit"
-
-  python install.py --continue_on_fail
-
-  # soxr comes from https://github.com/huggingface/transformers/pull/39429
-  pip install transformers==4.54.0 soxr==0.5.0
-
-  echo "Print all dependencies after TorchBench is installed"
-  python -mpip freeze
-  popd
-
-  chown -R jenkins torchbench
-  chown -R jenkins /opt/conda
+  # Clean up
+  conda_run pip uninstall -y torch torchvision triton
 }
 
 # Pango is needed for weasyprint which is needed for doctr
 conda_install pango
-
-# Stable packages are ok here, just to satisfy TorchBench check
-pip_install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
-
-install_torchbench
 install_huggingface
 install_timm
-
-# Clean up
-conda_run pip uninstall -y torch torchvision torchaudio triton torchao

.ci/docker/common/install_rocm.sh

@@ -30,7 +30,7 @@ EOF
 
     # we want the patch version of 6.4 instead
     if [[ $(ver $ROCM_VERSION) -eq $(ver 6.4) ]]; then
-        ROCM_VERSION="${ROCM_VERSION}.2"
+        ROCM_VERSION="${ROCM_VERSION}.1"
     fi
 
     # Default url values
@@ -85,19 +85,16 @@ EOF
     # CI no longer builds for ROCm 6.3, but
     # ROCm 6.4 did not yet fix the regression, also HIP branch names are different
     if [[ $(ver $ROCM_VERSION) -ge $(ver 6.4) ]] && [[ $(ver $ROCM_VERSION) -lt $(ver 7.0) ]]; then
-        if [[ $(ver $ROCM_VERSION) -eq $(ver 6.4.2) ]]; then
-            HIP_TAG=rocm-6.4.2
-            CLR_HASH=74d78ba3ac4bac235d02bcb48511c30b5cfdd457  # branch release/rocm-rel-6.4.2-statco-hotfix
-        elif [[ $(ver $ROCM_VERSION) -eq $(ver 6.4.1) ]]; then
-            HIP_TAG=rocm-6.4.1
+        if [[ $(ver $ROCM_VERSION) -eq $(ver 6.4.1) ]]; then
+            HIP_BRANCH=release/rocm-rel-6.4
             CLR_HASH=efe6c35790b9206923bfeed1209902feff37f386  # branch release/rocm-rel-6.4.1-statco-hotfix
         elif [[ $(ver $ROCM_VERSION) -eq $(ver 6.4) ]]; then
-            HIP_TAG=rocm-6.4.0
+            HIP_BRANCH=release/rocm-rel-6.4
             CLR_HASH=600f5b0d2baed94d5121e2174a9de0851b040b0c  # branch release/rocm-rel-6.4-statco-hotfix
         fi
         # clr build needs CppHeaderParser but can only find it using conda's python
         python -m pip install CppHeaderParser
-        git clone https://github.com/ROCm/HIP -b $HIP_TAG
+        git clone https://github.com/ROCm/HIP -b $HIP_BRANCH
         HIP_COMMON_DIR=$(readlink -f HIP)
         git clone https://github.com/jeffdaily/clr
         pushd clr

.ci/docker/common/install_xpu.sh

@@ -34,27 +34,18 @@ function install_ubuntu() {
 
     # The xpu-smi packages
     apt-get install -y flex bison xpu-smi
-
-    if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
-        # Compute and Media Runtimes
-        apt-get install -y \
-            intel-opencl-icd intel-level-zero-gpu level-zero \
-            intel-media-va-driver-non-free libmfx1 libmfxgen1 libvpl2 \
-            libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
-            libglapi-mesa libgles2-mesa-dev libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
-            mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo
-        # Development Packages
-        apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev level-zero-dev
-    else # rolling driver
-        apt-get install -y \
-            intel-opencl-icd libze-intel-gpu1 libze1 \
-            intel-media-va-driver-non-free libmfx-gen1 libvpl2 \
-            libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
-            libglapi-mesa libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
-            mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo intel-ocloc
-        apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev libze-dev
+    # Compute and Media Runtimes
+    apt-get install -y \
+        intel-opencl-icd intel-level-zero-gpu level-zero \
+        intel-media-va-driver-non-free libmfx1 libmfxgen1 libvpl2 \
+        libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
+        libglapi-mesa libgles2-mesa-dev libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
+        mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo
+    if [[ "${XPU_DRIVER_TYPE,,}" == "rolling" ]]; then
+        apt-get install -y intel-ocloc
     fi
-
+    # Development Packages
+    apt-get install -y libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev level-zero-dev
     # Install Intel Support Packages
     apt-get install -y ${XPU_PACKAGES}
 
@@ -139,11 +130,11 @@ function install_sles() {
 
 }
 
-# Default use GPU driver rolling releases
-XPU_DRIVER_VERSION=""
-if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
-    # Use GPU driver LTS releases
-    XPU_DRIVER_VERSION="/lts/2350"
+# Default use GPU driver LTS releases
+XPU_DRIVER_VERSION="/lts/2350"
+if [[ "${XPU_DRIVER_TYPE,,}" == "rolling" ]]; then
+    # Use GPU driver rolling releases
+    XPU_DRIVER_VERSION=""
 fi
 
 # Default use Intel® oneAPI Deep Learning Essentials 2025.0

.ci/docker/libtorch/build.sh

@@ -41,7 +41,7 @@ case ${DOCKER_TAG_PREFIX} in
     rocm*)
         # we want the patch version of 6.4 instead
         if [[ $(ver $GPU_ARCH_VERSION) -eq $(ver 6.4) ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.1"
         fi
         BASE_TARGET=rocm
         GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete

.ci/docker/manywheel/build.sh

@@ -77,7 +77,7 @@ case ${image} in
     manylinux2_28-builder:rocm*)
         # we want the patch version of 6.4 instead
         if [[ $(ver $GPU_ARCH_VERSION) -eq $(ver 6.4) ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.1"
         fi
         TARGET=rocm_final
         MANY_LINUX_VERSION="2_28"

.ci/docker/requirements-ci.txt

@@ -63,12 +63,11 @@ lark==0.12.0
 #Pinned versions: 0.12.0
 #test that import:
 
-librosa>=0.6.2 ; python_version < "3.11" and platform_machine != "s390x"
-librosa==0.10.2 ; python_version == "3.12" and platform_machine != "s390x"
+librosa>=0.6.2 ; python_version < "3.11"
+librosa==0.10.2 ; python_version == "3.12"
 #Description: A python package for music and audio analysis
 #Pinned versions: >=0.6.2
 #test that import: test_spectral_ops.py
-#librosa depends on numba; disable it for s390x while numba is disabled too
 
 #mkl #this breaks linux-bionic-rocm4.5-py3.7
 #Description: Intel oneAPI Math Kernel Library
@@ -111,15 +110,14 @@ ninja==1.11.1.3
 #Pinned versions: 1.11.1.3
 #test that import: run_test.py, test_cpp_extensions_aot.py,test_determination.py
 
-numba==0.49.0 ; python_version < "3.9" and platform_machine != "s390x"
-numba==0.55.2 ; python_version == "3.9" and platform_machine != "s390x"
-numba==0.55.2 ; python_version == "3.10" and platform_machine != "s390x"
-numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
+numba==0.49.0 ; python_version < "3.9"
+numba==0.55.2 ; python_version == "3.9"
+numba==0.55.2 ; python_version == "3.10"
+numba==0.60.0 ; python_version == "3.12"
 #Description: Just-In-Time Compiler for Numerical Functions
 #Pinned versions: 0.54.1, 0.49.0, <=0.49.1
 #test that import: test_numba_integration.py
 #For numba issue see https://github.com/pytorch/pytorch/issues/51511
-#Need release > 0.61.2 for s390x due to https://github.com/numba/numba/pull/10073
 
 #numpy
 #Description: Provides N-dimensional arrays and linear algebra
@@ -223,9 +221,9 @@ pygments==2.15.0
 #Pinned versions: 2.12.0
 #test that import: the doctests
 
-#pyyaml
+#PyYAML
 #Description: data serialization format
-#Pinned versions: 6.0.2
+#Pinned versions:
 #test that import:
 
 #requests
@@ -235,7 +233,7 @@ pygments==2.15.0
 
 #rich
 #Description: rich text and beautiful formatting in the terminal
-#Pinned versions: 14.1.0
+#Pinned versions: 10.9.0
 #test that import:
 
 scikit-image==0.19.3 ; python_version < "3.10"
@@ -309,7 +307,7 @@ pytest-cpp==2.3.0
 #Pinned versions: 2.3.0
 #test that import:
 
-z3-solver==4.15.1.0 ; platform_machine != "s390x"
+z3-solver==4.15.1.0
 #Description: The Z3 Theorem Prover Project
 #Pinned versions:
 #test that import:
@@ -363,6 +361,7 @@ pwlf==2.2.1
 #Pinned versions: 2.2.1
 #test that import: test_sac_estimator.py
 
+
 # To build PyTorch itself
 pyyaml
 pyzstd

.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@pytorch_sphinx_theme2#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
@@ -50,8 +50,8 @@ IPython==8.12.0
 #Pinned versions: 8.12.0
 
 myst-nb==0.17.2
-#Description: This is used to generate PyTorch functorch and torch.compile docs.
-#Pinned versions: 0.17.2
+#Description: This is used to generate PyTorch functorch docs
+#Pinned versions: 0.13.2
 
 # The following are required to build torch.distributed.elastic.rendezvous.etcd* docs
 python-etcd==0.4.5
@@ -59,3 +59,4 @@ sphinx-copybutton==0.5.0
 sphinx-design==0.4.0
 sphinxcontrib-mermaid==1.0.0
 myst-parser==0.18.1
+myst-nb

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

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

.ci/docker/ubuntu-rocm/Dockerfile

@@ -98,9 +98,8 @@ COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps
 COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface.txt huggingface.txt
 COPY ci_commit_pins/timm.txt timm.txt
-COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt torchbench.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
 
 # (optional) Install non-default Ninja version
 ARG NINJA_VERSION

.ci/docker/ubuntu/Dockerfile

@@ -98,9 +98,8 @@ COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps
 COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface.txt huggingface.txt
 COPY ci_commit_pins/timm.txt timm.txt
-COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt torchbench.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface.txt
 
 ARG TRITON
 ARG TRITON_CPU

.ci/lumen_cli/README.md

@@ -1,31 +0,0 @@
-# 🔧 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

@@ -1,37 +0,0 @@
-import argparse
-import logging
-
-from cli.lib.common.cli_helper import register_targets, RichHelp, TargetSpec
-from cli.lib.core.vllm 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

@@ -1,71 +0,0 @@
-"""
-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

@@ -1,42 +0,0 @@
-"""
-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

@@ -1,110 +0,0 @@
-"""
-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/git_helper.py

@@ -1,69 +0,0 @@
-"""
-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 commit %s", 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
-
-    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

@@ -1,14 +0,0 @@
-"""
-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

@@ -1,62 +0,0 @@
-"""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/utils.py

@@ -1,79 +0,0 @@
-"""
-General Utility helpers for CLI tasks.
-"""
-
-import logging
-import os
-import shlex
-import subprocess
-import sys
-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}")

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

@@ -1,263 +0,0 @@
-import logging
-import os
-import textwrap
-from dataclasses import dataclass
-from pathlib import Path
-from typing import Optional
-
-from cli.lib.common.cli_helper import BaseRunner
-from cli.lib.common.docker_helper import local_image_exists
-from cli.lib.common.envs_helper import (
-    env_bool_field,
-    env_path_field,
-    env_str_field,
-    with_params_help,
-)
-from cli.lib.common.git_helper import clone_external_repo
-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
-
-
-logger = logging.getLogger(__name__)
-
-
-# Default path for docker build artifacts
-_DEFAULT_RESULT_PATH = "./shared"
-
-# Temp folder in vllm work place to cp torch whls in vllm work directory for docker build
-_VLLM_TEMP_FOLDER = "tmp"
-
-
-@dataclass
-class VllmBuildParameters:
-    """
-    Parameters defining the vllm external input configurations.
-    Combine with VllmDockerBuildArgs to define the vllm build environment
-    """
-
-    # USE_TORCH_WHEEL: when true, use local Torch wheels; requires TORCH_WHEELS_PATH.
-    #  Otherwise docker build pull torch nightly during build
-    # TORCH_WHEELS_PATH: directory containing local torch wheels when use_torch_whl is True
-    use_torch_whl: bool = env_bool_field("USE_TORCH_WHEEL", True)
-    torch_whls_path: Path = env_path_field("TORCH_WHEELS_PATH", "./dist")
-
-    # USE_LOCAL_BASE_IMAGE: when true, use an existing local Docker base image; requires BASE_IMAGE
-    # Otherwise, pull dockerfile's default image remotely
-    # BASE_IMAGE: name:tag (only needed when use_local_base_image is True)
-    use_local_base_image: bool = env_bool_field("USE_LOCAL_BASE_IMAGE", True)
-    base_image: str = env_str_field("BASE_IMAGE")
-
-    # USE_LOCAL_DOCKERFILE: when true("1"), use a local Dockerfile; requires DOCKERFILE_PATH.
-    # otherwise, use vllm's default dockerfile.torch_nightly for build
-    # DOCKERFILE_PATH: path to Dockerfile used when use_local_dockerfile is True"
-    use_local_dockerfile: bool = env_bool_field("USE_LOCAL_DOCKERFILE", True)
-    dockerfile_path: Path = env_path_field(
-        "DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile.tmp_vllm"
-    )
-
-    # OUTPUT_DIR: where docker buildx (local exporter) will write artifacts
-    output_dir: Path = env_path_field("OUTPUT_DIR", "shared")
-
-    # --- Build args ----------------------------------------------------------
-    target_stage: str = env_str_field("TARGET_STAGE", "export-wheels")
-
-    tag_name: str = env_str_field("TAG", "vllm-wheels")
-
-    cuda_version: str = env_str_field("CUDA_VERSION", "12.8.1")
-
-    python_version: str = env_str_field("PYTHON_VERSION", "3.12")
-
-    max_jobs: str = env_str_field("MAX_JOBS", "64")
-
-    sccache_bucket: str = env_str_field("SCCACHE_BUCKET")
-
-    sccache_region: str = env_str_field("SCCACHE_REGION")
-
-    torch_cuda_arch_list: str = env_str_field("TORCH_CUDA_ARCH_LIST", "8.9")
-
-    def __post_init__(self):
-        checks = [
-            (
-                self.use_torch_whl,  # flag
-                True,  # trigger_value
-                "torch_whls_path",  # resource
-                is_path_exist,  # check_func
-                "TORCH_WHEELS_PATH is not provided, but USE_TORCH_WHEEL is set to 1",
-            ),
-            (
-                self.use_local_base_image,
-                True,
-                "base_image",
-                local_image_exists,
-                f"BASE_IMAGE {self.base_image} does not found, but USE_LOCAL_BASE_IMAGE is set to 1",
-            ),
-            (
-                self.use_local_dockerfile,
-                True,
-                "dockerfile_path",
-                is_path_exist,
-                " DOCKERFILE_PATH path does not found, but USE_LOCAL_DOCKERFILE is set to 1",
-            ),
-        ]
-        for flag, trigger_value, attr_name, check_func, error_msg in checks:
-            value = getattr(self, attr_name)
-            if flag == trigger_value:
-                if not value or not check_func(value):
-                    raise ValueError(error_msg)
-            else:
-                logger.info("flag  %s is not set", flag)
-        if not self.output_dir:
-            raise ValueError("missing required output_dir")
-
-
-@with_params_help(VllmBuildParameters)
-class VllmBuildRunner(BaseRunner):
-    """
-    Build vLLM using docker buildx.
-
-    Environment variable options:
-        "USE_TORCH_WHEEL":      "1: use local wheels; 0: pull nightly from pypi",
-        "TORCH_WHEELS_PATH":    "Path to local wheels (when USE_TORCH_WHEEL=1)",
-
-        "USE_LOCAL_BASE_IMAGE": "1: use local base image; 0: default image",
-         "BASE_IMAGE":           "name:tag to indicate base image the dockerfile depends on (when USE_LOCAL_BASE_IMAGE=1)",
-
-        "USE_LOCAL_DOCKERFILE": "1: use local Dockerfile; 0: vllm repo default dockerfile.torch_nightly",
-        "DOCKERFILE_PATH":      "Path to Dockerfile (when USE_LOCAL_DOCKERFILE=1)",
-
-        "OUTPUT_DIR":           "e.g. './shared'",
-
-        "TORCH_CUDA_ARCH_LIST": "e.g. '8.0' or '8.0;9.0'",
-        "CUDA_VERSION":         "e.g. '12.8.1'",
-        "PYTHON_VERSION":       "e.g. '3.12'",
-        "MAX_JOBS":             "e.g. '64'",
-        "SCCACHE_BUCKET":       "e.g. 'my-bucket'",
-        "SCCACHE_REGION":       "e.g. 'us-west-2'",
-    """
-
-    def __init__(self, args=None):
-        self.work_directory = "vllm"
-
-    def run(self):
-        """
-        main function to run vllm build
-        1. prepare vllm build environment
-        2. prepare the docker build command args
-        3. run docker build
-        """
-        inputs = VllmBuildParameters()
-        clone_vllm()
-
-        self.cp_dockerfile_if_exist(inputs)
-
-        # cp torch wheels from root direct to vllm workspace if exist
-        self.cp_torch_whls_if_exist(inputs)
-
-        ensure_dir_exists(inputs.output_dir)
-
-        cmd = self._generate_docker_build_cmd(inputs)
-        logger.info("Running docker build: \n %s", cmd)
-        run_command(cmd, cwd="vllm", env=os.environ.copy())
-
-    def cp_torch_whls_if_exist(self, inputs: VllmBuildParameters) -> str:
-        if not inputs.use_torch_whl:
-            return ""
-        tmp_dir = f"./{self.work_directory}/{_VLLM_TEMP_FOLDER}"
-        tmp_path = Path(tmp_dir)
-        force_create_dir(tmp_path)
-        copy(inputs.torch_whls_path, tmp_dir)
-        return tmp_dir
-
-    def cp_dockerfile_if_exist(self, inputs: VllmBuildParameters):
-        if not inputs.use_local_dockerfile:
-            logger.info("using vllm default dockerfile.torch_nightly for build")
-            return
-        dockerfile_path = get_path(inputs.dockerfile_path, resolve=True)
-        vllm_torch_dockerfile = Path(
-            f"./{self.work_directory}/docker/Dockerfile.nightly_torch"
-        )
-        copy(dockerfile_path, vllm_torch_dockerfile)
-
-    def get_result_path(self, path):
-        """
-        Get the absolute path of the result path
-        """
-        if not path:
-            path = _DEFAULT_RESULT_PATH
-        abs_path = get_path(path, resolve=True)
-        return abs_path
-
-    def _get_torch_wheel_path_arg(self, torch_whl_dir: Optional[Path]) -> str:
-        if not torch_whl_dir:
-            return ""
-        return f"--build-arg TORCH_WHEELS_PATH={_VLLM_TEMP_FOLDER}"
-
-    def _get_base_image_args(self, inputs: VllmBuildParameters) -> tuple[str, str, str]:
-        """
-        Returns:
-            - base_image_arg: docker buildx arg string for base image
-            - final_base_image_arg:  docker buildx arg string for vllm-base stage
-            - pull_flag: --pull=true or --pull=false depending on whether the image exists locally
-        """
-        if not inputs.use_local_base_image:
-            return "", "", ""
-
-        base_image = inputs.base_image
-
-        # set both base image and final base image to the same local image
-        base_image_arg = f"--build-arg BUILD_BASE_IMAGE={base_image}"
-        final_base_image_arg = f"--build-arg FINAL_BASE_IMAGE={base_image}"
-
-        if local_image_exists(base_image):
-            pull_flag = "--pull=false"
-            return base_image_arg, final_base_image_arg, pull_flag
-        logger.info(
-            "[INFO] Local image not found:%s will try to pull from remote", {base_image}
-        )
-        return base_image_arg, final_base_image_arg, ""
-
-    def _generate_docker_build_cmd(
-        self,
-        inputs: VllmBuildParameters,
-    ) -> str:
-        base_image_arg, final_base_image_arg, pull_flag = self._get_base_image_args(
-            inputs
-        )
-        torch_arg = self._get_torch_wheel_path_arg(inputs.torch_whls_path)
-
-        return textwrap.dedent(
-            f"""
-            docker buildx build \
-                --output type=local,dest={inputs.output_dir} \
-                -f docker/Dockerfile.nightly_torch \
-                {pull_flag} \
-                {torch_arg} \
-                {base_image_arg} \
-                {final_base_image_arg} \
-                --build-arg max_jobs={inputs.max_jobs} \
-                --build-arg CUDA_VERSION={inputs.cuda_version} \
-                --build-arg PYTHON_VERSION={inputs.python_version} \
-                --build-arg USE_SCCACHE={int(bool(inputs.sccache_bucket and inputs.sccache_region))} \
-                --build-arg SCCACHE_BUCKET_NAME={inputs.sccache_bucket} \
-                --build-arg SCCACHE_REGION_NAME={inputs.sccache_region} \
-                --build-arg torch_cuda_arch_list='{inputs.torch_cuda_arch_list}' \
-                --target {inputs.target_stage} \
-                -t {inputs.tag_name} \
-                --progress=plain .
-        """
-        ).strip()
-
-
-def clone_vllm():
-    clone_external_repo(
-        target="vllm",
-        repo="https://github.com/vllm-project/vllm.git",
-        dst="vllm",
-        update_submodules=True,
-    )

.ci/lumen_cli/cli/run.py

@@ -1,38 +0,0 @@
-# main.py
-
-import argparse
-import logging
-
-from cli.build_cli.register_build import register_build_commands
-from cli.lib.common.logger import setup_logging
-
-
-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)
-
-    # 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/pyproject.toml

@@ -1,22 +0,0 @@
-[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",
-]
-
-[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

@@ -1,47 +0,0 @@
-# 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

@@ -1,115 +0,0 @@
-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

@@ -1,75 +0,0 @@
-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

@@ -1,149 +0,0 @@
-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

@@ -1,122 +0,0 @@
-# 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_vllm.py

@@ -1,181 +0,0 @@
-import os
-import tempfile
-import unittest
-from pathlib import Path
-from unittest.mock import MagicMock, patch
-
-import cli.lib.core.vllm as vllm
-
-
-class TestVllmBuildParameters(unittest.TestCase):
-    @patch("cli.lib.core.vllm.local_image_exists", return_value=True)
-    @patch("cli.lib.core.vllm.is_path_exist", return_value=True)
-    @patch(
-        "cli.lib.common.envs_helper.env_path_optional",
-        side_effect=lambda name, default=None, resolve=True: {
-            "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.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("cli.lib.core.vllm.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.VllmBuildParameters(
-                    use_local_base_image=False,
-                    use_local_dockerfile=False,
-                )
-        err = cm.exception
-        self.assertIn("TORCH_WHEELS_PATH", str(err))
-
-    @patch("cli.lib.core.vllm.local_image_exists", return_value=False)
-    @patch.dict(
-        os.environ, {"USE_LOCAL_BASE_IMAGE": "1", "BASE_IMAGE": "img:tag"}, clear=True
-    )
-    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.VllmBuildParameters(
-                    use_torch_whl=False,
-                    use_local_dockerfile=False,
-                )
-        err = cm.exception
-        self.assertIn("BASE_IMAGE", str(err))
-
-    @patch("cli.lib.core.vllm.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.VllmBuildParameters(
-                    use_torch_whl=False,
-                    use_local_base_image=False,
-                )
-        err = cm.exception
-        self.assertIn("DOCKERFILE_PATH", str(err))
-
-    @patch("cli.lib.core.vllm.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.VllmBuildParameters()
-
-
-class TestBuildCmdAndRun(unittest.TestCase):
-    @patch("cli.lib.core.vllm.local_image_exists", return_value=True)
-    def test_generate_docker_build_cmd_includes_bits(self, _exists):
-        runner = vllm.VllmBuildRunner()
-        # Craft inputs that simulate a prepared build
-        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())  # normalize whitespace for matching
-
-        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("cli.lib.core.vllm.run_command")
-    @patch("cli.lib.core.vllm.ensure_dir_exists")
-    @patch("cli.lib.core.vllm.clone_vllm")
-    @patch.object(
-        vllm.VllmBuildRunner,
-        "_generate_docker_build_cmd",
-        return_value="docker buildx ...",
-    )
-    @patch.dict(
-        os.environ,
-        {
-            # Make __post_init__ validations pass cheaply
-            "USE_TORCH_WHEEL": "0",
-            "USE_LOCAL_BASE_IMAGE": "0",
-            "USE_LOCAL_DOCKERFILE": "0",
-            "OUTPUT_DIR": "shared",
-        },
-        clear=True,
-    )
-    def test_run_calls_clone_prepare_and_build(
-        self, mock_gen, mock_clone, mock_ensure, mock_run
-    ):
-        # Stub parameters instance so we avoid FS/Docker accesses in run()
-        params = MagicMock()
-        params.output_dir = Path("shared")
-        params.use_local_dockerfile = False
-        params.use_torch_whl = False
-
-        with patch("cli.lib.core.vllm.VllmBuildParameters", return_value=params):
-            runner = vllm.VllmBuildRunner()
-            runner.run()
-
-        mock_clone.assert_called_once()
-        mock_ensure.assert_called_once_with(Path("shared"))
-        mock_gen.assert_called_once_with(params)
-        mock_run.assert_called_once()
-        # ensure we run in vllm workdir
-        _, kwargs = mock_run.call_args
-        assert kwargs.get("cwd") == "vllm"
-
-
-if __name__ == "__main__":
-    unittest.main()

.ci/manywheel/build.sh

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

.ci/manywheel/build_common.sh

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

.ci/manywheel/build_cuda.sh

@@ -134,7 +134,6 @@ if [[ $CUDA_VERSION == 12* ]]; then
             "/usr/local/cuda/lib64/libnvrtc-builtins.so"
             "/usr/local/cuda/lib64/libcufile.so.0"
             "/usr/local/cuda/lib64/libcufile_rdma.so.1"
-            "/usr/local/cuda/lib64/libnvshem_host.so.3"
             "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12"
             "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so"
         )
@@ -153,7 +152,6 @@ if [[ $CUDA_VERSION == 12* ]]; then
             "libcudart.so.12"
             "libnvrtc.so.12"
             "libnvrtc-builtins.so"
-            "libnvshmem_host.so.3"
             "libcufile.so.0"
             "libcufile_rdma.so.1"
             "libcupti.so.12"

.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 $ROCBLAS_OTHER_FILES)
+ROCBLAS_LIB_FILES=($ROCBLAS_ARCH_SPECIFIC_FILES $OTHER_FILES)
 
 # hipblaslt library files
 HIPBLASLT_LIB_SRC=$ROCM_HOME/lib/hipblaslt/library

.ci/pytorch/build.sh

@@ -50,6 +50,9 @@ if [[ ${BUILD_ENVIRONMENT} == *"parallelnative"* ]]; then
   export ATEN_THREADING=NATIVE
 fi
 
+# Enable LLVM dependency for TensorExpr testing
+export USE_LLVM=/opt/llvm
+export LLVM_DIR=/opt/llvm/lib/cmake/llvm
 
 if ! which conda; then
   # In ROCm CIs, we are doing cross compilation on build machines with
@@ -92,27 +95,6 @@ if [[ "$BUILD_ENVIRONMENT" == *aarch64* ]]; then
   export ACL_ROOT_DIR=/ComputeLibrary
 fi
 
-if [[ "$BUILD_ENVIRONMENT" == *riscv64* ]]; then
-  if [[ -f /opt/riscv-cross-env/bin/activate ]]; then
-    # shellcheck disable=SC1091
-    source /opt/riscv-cross-env/bin/activate
-  else
-    echo "Activation file not found"
-    exit 1
-  fi
-
-  export CMAKE_CROSSCOMPILING=TRUE
-  export CMAKE_SYSTEM_NAME=Linux
-  export CMAKE_SYSTEM_PROCESSOR=riscv64
-
-  export USE_CUDA=0
-  export USE_MKLDNN=0
-
-  export SLEEF_TARGET_EXEC_USE_QEMU=ON
-  sudo chown -R jenkins /var/lib/jenkins/workspace /opt
-
-fi
-
 if [[ "$BUILD_ENVIRONMENT" == *libtorch* ]]; then
   POSSIBLE_JAVA_HOMES=()
   POSSIBLE_JAVA_HOMES+=(/usr/local)
@@ -194,7 +176,7 @@ fi
 
 # We only build FlashAttention files for CUDA 8.0+, and they require large amounts of
 # memory to build and will OOM
-if [[ "$BUILD_ENVIRONMENT" == *cuda* ]] && echo "${TORCH_CUDA_ARCH_LIST}" | tr ' ' '\n' | sed 's/$/>= 8.0/' | bc | grep -q 1; then
+if [[ "$BUILD_ENVIRONMENT" == *cuda* ]] && [[ 1 -eq $(echo "${TORCH_CUDA_ARCH_LIST} >= 8.0" | bc) ]]; then
   export BUILD_CUSTOM_STEP="ninja -C build flash_attention -j 2"
 fi
 
@@ -210,6 +192,7 @@ if [[ "$BUILD_ENVIRONMENT" == *-clang*-asan* ]]; then
   export USE_ASAN=1
   export REL_WITH_DEB_INFO=1
   export UBSAN_FLAGS="-fno-sanitize-recover=all"
+  unset USE_LLVM
 fi
 
 if [[ "${BUILD_ENVIRONMENT}" == *no-ops* ]]; then
@@ -230,7 +213,7 @@ fi
 
 # Do not change workspace permissions for ROCm and s390x CI jobs
 # as it can leave workspace with bad permissions for cancelled jobs
-if [[ "$BUILD_ENVIRONMENT" != *rocm* && "$BUILD_ENVIRONMENT" != *s390x* && "$BUILD_ENVIRONMENT" != *riscv64* && -d /var/lib/jenkins/workspace ]]; then
+if [[ "$BUILD_ENVIRONMENT" != *rocm* && "$BUILD_ENVIRONMENT" != *s390x* && -d /var/lib/jenkins/workspace ]]; then
   # Workaround for dind-rootless userid mapping (https://github.com/pytorch/ci-infra/issues/96)
   WORKSPACE_ORIGINAL_OWNER_ID=$(stat -c '%u' "/var/lib/jenkins/workspace")
   cleanup_workspace() {
@@ -275,19 +258,29 @@ else
     # XLA test build fails when WERROR=1
     # set only when building other architectures
     # or building non-XLA tests.
-    if [[ "$BUILD_ENVIRONMENT" != *rocm*  && "$BUILD_ENVIRONMENT" != *xla* && "$BUILD_ENVIRONMENT" != *riscv64* ]]; then
+    if [[ "$BUILD_ENVIRONMENT" != *rocm*  &&
+          "$BUILD_ENVIRONMENT" != *xla* ]]; then
       # Install numpy-2.0.2 for builds which are backward compatible with 1.X
       python -mpip install numpy==2.0.2
 
       WERROR=1 python setup.py clean
 
-      WERROR=1 python setup.py bdist_wheel
+      if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+        python3 tools/packaging/split_wheel.py bdist_wheel
+      else
+        WERROR=1 python setup.py bdist_wheel
+      fi
     else
       python setup.py clean
       if [[ "$BUILD_ENVIRONMENT" == *xla* ]]; then
         source .ci/pytorch/install_cache_xla.sh
       fi
-      python setup.py bdist_wheel
+      if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+        echo "USE_SPLIT_BUILD cannot be used with xla or rocm"
+        exit 1
+      else
+        python setup.py bdist_wheel
+      fi
     fi
     pip_install_whl "$(echo dist/*.whl)"
 
@@ -412,7 +405,7 @@ if [[ "$BUILD_ENVIRONMENT" != *libtorch* && "$BUILD_ENVIRONMENT" != *bazel* ]];
   # don't do this for libtorch as libtorch is C++ only and thus won't have python tests run on its build
   python tools/stats/export_test_times.py
 fi
-# don't do this for bazel or s390x or riscv64 as they don't use sccache
-if [[ "$BUILD_ENVIRONMENT" != *s390x* && "$BUILD_ENVIRONMENT" != *riscv64* && "$BUILD_ENVIRONMENT" != *-bazel-* ]]; then
+# don't do this for bazel or s390x as they don't use sccache
+if [[ "$BUILD_ENVIRONMENT" != *s390x* && "$BUILD_ENVIRONMENT" != *-bazel-* ]]; then
   print_sccache_stats
 fi

.ci/pytorch/common_utils.sh

@@ -229,6 +229,7 @@ function install_torchrec_and_fbgemm() {
 
     pip_install tabulate  # needed for newer fbgemm
     pip_install patchelf  # needed for rocm fbgemm
+    pushd /tmp
 
     local wheel_dir=dist/fbgemm_gpu
     local found_whl=0
@@ -244,7 +245,7 @@ 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}" --recurse-submodules
+      git checkout "${fbgemm_commit}"
       python setup.py bdist_wheel \
         --build-variant=rocm \
         -DHIP_ROOT_DIR="${ROCM_PATH}" \
@@ -263,6 +264,7 @@ 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
@@ -281,6 +283,30 @@ function clone_pytorch_xla() {
   fi
 }
 
+function checkout_install_torchbench() {
+  local commit
+  commit=$(get_pinned_commit torchbench)
+  git clone https://github.com/pytorch/benchmark torchbench
+  pushd torchbench
+  git checkout "$commit"
+
+  if [ "$1" ]; then
+    python install.py --continue_on_fail models "$@"
+  else
+    # Occasionally the installation may fail on one model but it is ok to continue
+    # to install and test other models
+    python install.py --continue_on_fail
+  fi
+
+  # TODO (huydhn): transformers-4.44.2 added by https://github.com/pytorch/benchmark/pull/2488
+  # is regressing speedup metric. This needs to be investigated further
+  pip install transformers==4.38.1
+
+  echo "Print all dependencies after TorchBench is installed"
+  python -mpip freeze
+  popd
+}
+
 function install_torchao() {
   local commit
   commit=$(get_pinned_commit torchao)

.ci/pytorch/macos-test.sh

@@ -157,32 +157,6 @@ test_jit_hooks() {
   assert_git_not_dirty
 }
 
-# Shellcheck doesn't like it when you pass no arguments to a function
-# that can take args. See https://www.shellcheck.net/wiki/SC2120
-# shellcheck disable=SC2120
-checkout_install_torchbench() {
-  local commit
-  commit=$(cat .ci/docker/ci_commit_pins/torchbench.txt)
-  git clone https://github.com/pytorch/benchmark torchbench
-  pushd torchbench
-  git checkout "$commit"
-
-  if [ "$1" ]; then
-    python install.py --continue_on_fail models "$@"
-  else
-    # Occasionally the installation may fail on one model but it is ok to continue
-    # to install and test other models
-    python install.py --continue_on_fail
-  fi
-
-  # soxr comes from https://github.com/huggingface/transformers/pull/39429
-  pip install transformers==4.54.0 soxr==0.5.0
-
-  echo "Print all dependencies after TorchBench is installed"
-  python -mpip freeze
-  popd
-}
-
 torchbench_setup_macos() {
   git clone --recursive https://github.com/pytorch/vision torchvision
   git clone --recursive https://github.com/pytorch/audio torchaudio
@@ -205,6 +179,8 @@ torchbench_setup_macos() {
   USE_OPENMP=0 python setup.py develop
   popd
 
+  # Shellcheck doesn't like it when you pass no arguments to a function that can take args. See https://www.shellcheck.net/wiki/SC2120
+  # shellcheck disable=SC2119,SC2120
   checkout_install_torchbench
 }
 

.ci/pytorch/test.sh

@@ -462,7 +462,7 @@ test_inductor_aoti() {
   # rebuild with the build cache with `BUILD_AOT_INDUCTOR_TEST` enabled
   /usr/bin/env CMAKE_FRESH=1 BUILD_AOT_INDUCTOR_TEST=1 "${BUILD_COMMAND[@]}"
 
-  /usr/bin/env "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference cpp/test_vec_half_AVX2 -dist=loadfile
+  /usr/bin/env "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference -dist=loadfile
 }
 
 test_inductor_cpp_wrapper_shard() {
@@ -627,8 +627,6 @@ test_perf_for_dashboard() {
     device=cuda_a10g
   elif [[ "${TEST_CONFIG}" == *h100* ]]; then
     device=cuda_h100
-  elif [[ "${TEST_CONFIG}" == *b200* ]]; then
-    device=cuda_b200
   elif [[ "${TEST_CONFIG}" == *rocm* ]]; then
     device=rocm
   fi
@@ -803,16 +801,6 @@ test_dynamo_benchmark() {
   if [[ "${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
@@ -986,8 +974,6 @@ test_without_numpy() {
   if [[ "${TEST_CONFIG}" == *dynamo_wrapped* ]]; then
     python -c "import sys;sys.path.insert(0, 'fake_numpy');import torch;torch.compile(lambda x:print(x))('Hello World')"
   fi
-  # Regression test for https://github.com/pytorch/pytorch/pull/157734 (torch.onnx should be importable without numpy)
-  python -c "import sys;sys.path.insert(0, 'fake_numpy');import torch; import torch.onnx"
   popd
 }
 
@@ -1051,10 +1037,20 @@ test_libtorch_api() {
     mkdir -p $TEST_REPORTS_DIR
 
     OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="${MNIST_DIR}" "$TORCH_BIN_DIR"/test_api --gtest_filter='-IMethodTest.*' --gtest_output=xml:$TEST_REPORTS_DIR/test_api.xml
+    "$TORCH_BIN_DIR"/test_tensorexpr --gtest_output=xml:$TEST_REPORTS_DIR/test_tensorexpr.xml
   else
     # Exclude IMethodTest that relies on torch::deploy, which will instead be ran in test_deploy
     OMP_NUM_THREADS=2 TORCH_CPP_TEST_MNIST_PATH="${MNIST_DIR}" python test/run_test.py --cpp --verbose -i cpp/test_api -k "not IMethodTest"
 
+    # On s390x, pytorch is built without llvm.
+    # Even if it would be built with llvm, llvm currently doesn't support used features on s390x and
+    # test fails with errors like:
+    # JIT session error: Unsupported target machine architecture in ELF object pytorch-jitted-objectbuffer
+    # unknown file: Failure
+    # C++ exception with description "valOrErr INTERNAL ASSERT FAILED at "/var/lib/jenkins/workspace/torch/csrc/jit/tensorexpr/llvm_jit.h":34, please report a bug to PyTorch. Unexpected failure in LLVM JIT: Failed to materialize symbols: { (main, { func }) }
+    if [[ "${BUILD_ENVIRONMENT}" != *s390x* ]]; then
+      python test/run_test.py --cpp --verbose -i cpp/test_tensorexpr
+    fi
   fi
 
   # quantization is not fully supported on s390x yet
@@ -1322,13 +1318,10 @@ EOF
 
   # Step 2. Make sure that the public API test "test_correct_module_names" fails when an existing
   # file is modified to introduce an invalid public API function.
-  # The filepath here must not have __all__ defined in it, otherwise the test will pass.
-  # If your PR introduces __all__ to torch/cuda/streams.py please point this to another file
-  # that does not have __all__ defined.
-  EXISTING_FILEPATH="${TORCH_INSTALL_DIR}/cuda/streams.py"
+  EXISTING_FILEPATH="${TORCH_INSTALL_DIR}/nn/parameter.py"
   cp -v "${EXISTING_FILEPATH}" "${EXISTING_FILEPATH}.orig"
   echo "${BAD_PUBLIC_FUNC}" >> "${EXISTING_FILEPATH}"
-  invalid_api="torch.cuda.streams.new_public_func"
+  invalid_api="torch.nn.parameter.new_public_func"
   echo "Appended an invalid public API function to existing file ${EXISTING_FILEPATH}..."
 
   check_public_api_test_fails \
@@ -1562,7 +1555,7 @@ test_executorch() {
 test_linux_aarch64() {
   python test/run_test.py --include test_modules test_mkldnn test_mkldnn_fusion test_openmp test_torch test_dynamic_shapes \
         test_transformers test_multiprocessing test_numpy_interop test_autograd test_binary_ufuncs test_complex test_spectral_ops \
-        test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops \
+        test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops test_cpp_extensions_open_device_registration \
         --shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose
 
   # Dynamo tests
@@ -1674,34 +1667,43 @@ elif [[ "${TEST_CONFIG}" == *timm* ]]; then
 elif [[ "${TEST_CONFIG}" == cachebench ]]; then
   install_torchaudio
   install_torchvision
-  PYTHONPATH=/torchbench test_cachebench
+  checkout_install_torchbench nanogpt BERT_pytorch resnet50 hf_T5 llama moco
+  PYTHONPATH=$(pwd)/torchbench test_cachebench
 elif [[ "${TEST_CONFIG}" == verify_cachebench ]]; then
   install_torchaudio
   install_torchvision
-  PYTHONPATH=/torchbench test_verify_cachebench
+  checkout_install_torchbench nanogpt
+  PYTHONPATH=$(pwd)/torchbench test_verify_cachebench
 elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
   install_torchaudio
   install_torchvision
+  install_torchao
   id=$((SHARD_NUMBER-1))
   # https://github.com/opencv/opencv-python/issues/885
   pip_install opencv-python==4.8.0.74
   if [[ "${TEST_CONFIG}" == *inductor_torchbench_smoketest_perf* ]]; then
-    PYTHONPATH=/torchbench test_inductor_torchbench_smoketest_perf
+    checkout_install_torchbench hf_Bert hf_Albert timm_vision_transformer
+    PYTHONPATH=$(pwd)/torchbench test_inductor_torchbench_smoketest_perf
   elif [[ "${TEST_CONFIG}" == *inductor_torchbench_cpu_smoketest_perf* ]]; then
-    PYTHONPATH=/torchbench test_inductor_torchbench_cpu_smoketest_perf
+    checkout_install_torchbench timm_vision_transformer phlippe_densenet basic_gnn_edgecnn \
+      llama_v2_7b_16h resnet50 timm_efficientnet mobilenet_v3_large timm_resnest \
+      functorch_maml_omniglot yolov3 mobilenet_v2 resnext50_32x4d densenet121 mnasnet1_0
+    PYTHONPATH=$(pwd)/torchbench test_inductor_torchbench_cpu_smoketest_perf
   elif [[ "${TEST_CONFIG}" == *torchbench_gcp_smoketest* ]]; then
-    TORCHBENCHPATH=/torchbench test_torchbench_gcp_smoketest
+    checkout_install_torchbench
+    TORCHBENCHPATH=$(pwd)/torchbench test_torchbench_gcp_smoketest
   else
+    checkout_install_torchbench
     # Do this after checkout_install_torchbench to ensure we clobber any
     # nightlies that torchbench may pull in
     if [[ "${TEST_CONFIG}" != *cpu* ]]; then
       install_torchrec_and_fbgemm
     fi
-    PYTHONPATH=/torchbench test_dynamo_benchmark torchbench "$id"
+    PYTHONPATH=$(pwd)/torchbench test_dynamo_benchmark torchbench "$id"
   fi
 elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
   install_torchvision
-  PYTHONPATH=/torchbench:$PYTHONPATH test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
+  PYTHONPATH=$(pwd)/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
   if [[ "$SHARD_NUMBER" -eq "1" ]]; then
     test_inductor_aoti
   fi

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

@@ -61,10 +61,9 @@ if "%USE_XPU%"=="1" (
   call "C:\Program Files (x86)\Intel\oneAPI\compiler\latest\env\vars.bat"
   call "C:\Program Files (x86)\Intel\oneAPI\ocloc\latest\env\vars.bat"
   if errorlevel 1 exit /b 1
-  :: Reduce build time
-  SET TORCH_XPU_ARCH_LIST=bmg
-  :: Re-setup python env for build
-  call pip install -r requirements.txt
+  :: Reduce build time. Only have MTL self-hosted runner now
+  SET TORCH_XPU_ARCH_LIST=xe-lpg
+  SET USE_KINETO=0
 )
 
 @echo on

.ci/pytorch/windows/cuda126.bat

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

.ci/wheel/build_wheel.sh

@@ -192,6 +192,9 @@ 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 "$CROSS_COMPILE_ARM64" ]]; then
+    export CMAKE_OSX_ARCHITECTURES=arm64
+fi
 export USE_MKLDNN=OFF
 export USE_QNNPACK=OFF
 export BUILD_TEST=OFF
@@ -199,7 +202,16 @@ export BUILD_TEST=OFF
 pushd "$pytorch_rootdir"
 echo "Calling setup.py bdist_wheel at $(date)"
 
-python setup.py bdist_wheel -d "$whl_tmp_dir"
+if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
+    echo "Calling setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    BUILD_LIBTORCH_WHL=1 BUILD_PYTHON_ONLY=0 python setup.py bdist_wheel -d "$whl_tmp_dir"
+    echo "Finished setup.py bdist_wheel for split build (BUILD_LIBTORCH_WHL)"
+    echo "Calling setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+    BUILD_LIBTORCH_WHL=0 BUILD_PYTHON_ONLY=1 CMAKE_FRESH=1 python setup.py bdist_wheel -d "$whl_tmp_dir"
+    echo "Finished setup.py bdist_wheel for split build (BUILD_PYTHON_ONLY)"
+else
+    python setup.py bdist_wheel -d "$whl_tmp_dir"
+fi
 
 echo "Finished setup.py bdist_wheel at $(date)"
 

.circleci/scripts/binary_linux_test.sh

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

.circleci/scripts/binary_populate_env.sh

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

.circleci/scripts/binary_upload.sh

@@ -23,6 +23,10 @@ if [[ "${DRY_RUN}" = "disabled" ]]; then
   AWS_S3_CP="aws s3 cp"
 fi
 
+if [[ "${USE_SPLIT_BUILD:-false}" == "true" ]]; then
+  UPLOAD_SUBFOLDER="${UPLOAD_SUBFOLDER}_pypi_pkg"
+fi
+
 # this is special build with all dependencies packaged
 if [[ ${BUILD_NAME} == *-full* ]]; then
   UPLOAD_SUBFOLDER="${UPLOAD_SUBFOLDER}_full"

.flake8

@@ -7,12 +7,12 @@ max-line-length = 120
 # C408 ignored because we like the dict keyword argument syntax
 # E501 is not flexible enough, we're using B950 instead
 ignore =
-    E203,E305,E402,E501,E704,E721,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,F824,
+    E203,E305,E402,E501,E704,E721,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,
     # shebang has extra meaning in fbcode lints, so I think it's not worth trying
     # to line this up with executable bit
     EXE001,
     # these ignores are from flake8-bugbear; please fix!
-    B007,B008,B017,B019,B023,B028,B903,B904,B905,B906,B907,B908,B910
+    B007,B008,B017,B019,B023,B028,B903,B904,B905,B906,B907
     # these ignores are from flake8-comprehensions; please fix!
     C407,
     # these ignores are from flake8-logging-format; please fix!

.github/actionlint.yaml

@@ -53,13 +53,16 @@ self-hosted-runner:
     - linux.rocm.gpu.mi250
     - linux.rocm.gpu.2
     - linux.rocm.gpu.4
-    # gfx942 runners
-    - linux.rocm.gpu.gfx942.1
-    - linux.rocm.gpu.gfx942.2
-    - linux.rocm.gpu.gfx942.4
+    # MI300 runners
+    - linux.rocm.gpu.mi300.2
+    - linux.rocm.gpu.mi300.4
     - rocm-docker
+    # Repo-specific Apple hosted  runners
+    - macos-m1-ultra
+    - macos-m2-14
     # Org wise AWS `mac2.metal` runners (2020 Mac mini hardware powered by Apple silicon M1 processors)
     - macos-m1-stable
+    - macos-m1-13
     - macos-m1-14
     # GitHub-hosted MacOS runners
     - macos-latest-xlarge

.github/actions/filter-test-configs/action.yml

@@ -70,7 +70,7 @@ runs:
           set -eux
           # PyYAML 6.0 doesn't work with MacOS x86 anymore
           # This must run on Python-3.7 (AmazonLinux2) so can't use request=3.32.2
-          python3 -m pip install requests==2.27.1 pyyaml==6.0.2
+          python3 -m pip install requests==2.27.1 pyyaml==6.0.1
 
     - name: Parse ref
       id: parse-ref

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

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

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

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

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-bdb88e1d66f272cad72156c90ac8428ca61a601c
+b6a3368a45aaafe05f1a6a9f10c68adc5e944d9e

.github/ci_commit_pins/vllm.txt

@@ -1 +1 @@
-0ca2393b47e72c4424a49aa3b32c7c5d0e378a72
+f3137cdd81cae3a48282c22130fbcadcfc64ea95

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-095faec1e7b6cc47220181e74ae9cde2605f9b00
+1c00dea2c9adb2137903c86b4191e8c247f8fda9

.github/ci_configs/vllm/Dockerfile.tmp_vllm

@@ -1,414 +0,0 @@
-# TODO(elainwy): remove this file after the torch nightly dockerfile is in sync in vllm repo
-# The vLLM Dockerfile is used to construct vLLM image against torch nightly and torch main that can be directly used for testing
-
-ARG CUDA_VERSION=12.8.1
-ARG PYTHON_VERSION=3.12
-
-# BUILD_BASE_IMAGE: used to setup python build xformers, and vllm wheels, It can be replaced with a different base image from local machine,
-# by default, it uses the torch-nightly-base stage from this docker image
-ARG BUILD_BASE_IMAGE=torch-nightly-base
-
-# FINAL_BASE_IMAGE: used to set up vllm-instaled environment and build flashinfer,
-# by default, it uses devel-ubuntu22.04 official image.
-ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
-
-
-#################### TORCH NIGHTLY  BASE IMAGE ####################
-# A base image for building vLLM with devel ubuntu 22.04, this is mainly used to build vllm in vllm builtkite ci
-From nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 as torch-nightly-base
-ARG CUDA_VERSION=12.8.1
-ARG PYTHON_VERSION=3.12
-ARG TARGETPLATFORM
-ENV DEBIAN_FRONTEND=noninteractive
-
-RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
-    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
-
-# Install Python and other dependencies if it does not existed
-RUN if ! command -v python3 >/dev/null || ! python3 --version | grep -q "${PYTHON_VERSION}"; then \
-      echo "Installing Python ${PYTHON_VERSION}..." && \
-      echo 'tzdata tzdata/Areas select America' | debconf-set-selections && \
-      echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections && \
-      apt-get update -y && \
-      apt-get install -y ccache software-properties-common git curl sudo && \
-      for i in 1 2 3; do \
-        add-apt-repository -y ppa:deadsnakes/ppa && break || \
-        { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
-      done && \
-      apt-get update -y && \
-      apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv && \
-      update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 && \
-      update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} && \
-      ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config && \
-      curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}; \
-   else \
-      echo "Python ${PYTHON_VERSION} already present, skipping setup."; \
-   fi \
-   && python3 --version && python3 -m pip --version
-
-# Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
-# as it was causing spam when compiling the CUTLASS kernels
-# Ensure gcc >= 10 to avoid CUTLASS issues (bug 92519)
-RUN current_gcc_version=$(gcc -dumpversion | cut -f1 -d.) && \
-    if [ "$current_gcc_version" -lt 10 ]; then \
-      echo "GCC version is $current_gcc_version, installing gcc-10..."; \
-      apt-get update && \
-      apt-get install -y gcc-10 g++-10 && \
-      update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 && \
-      update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100; \
-    else \
-      echo "GCC version is $current_gcc_version, no need to install gcc-10."; \
-    fi && \
-    gcc --version && g++ --version
-
-# install uv for faster pip installs
-RUN --mount=type=cache,target=/root/.cache/uv \
-    python3 -m pip install uv==0.8.4
-
-ENV UV_HTTP_TIMEOUT=500
-ENV UV_INDEX_STRATEGY="unsafe-best-match"
-
-#################### TORCH NIGHTLY  BASE IMAGE ####################
-
-
-#################### BASE BUILD IMAGE ####################
-# A base image for building vLLM with torch nightly or torch wheels
-# prepare basic build environment
-FROM ${BUILD_BASE_IMAGE} AS base
-USER root
-
-# Workaround for https://github.com/openai/triton/issues/2507 and
-# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
-# this won't be needed for future versions of this docker image
-# or future versions of triton.
-RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
-
-# Install uv for faster pip installs if not existed
-RUN --mount=type=cache,target=/root/.cache/uv \
-    if ! python3 -m uv --version >/dev/null 2>&1; then \
-        python3 -m pip install uv==0.8.4; \
-    fi
-ENV UV_HTTP_TIMEOUT=500
-ENV UV_INDEX_STRATEGY="unsafe-best-match"
-
-WORKDIR /workspace
-
-# install build and runtime dependencies
-COPY requirements/common.txt requirements/common.txt
-COPY use_existing_torch.py use_existing_torch.py
-COPY pyproject.toml pyproject.toml
-
-# install build and runtime dependencies without stable torch version
-RUN python3 use_existing_torch.py
-
-# default mount file as placeholder, this just avoid the mount error
-# change to a different vllm folder if this does not exist anymore
-ARG TORCH_WHEELS_PATH="./requirements"
-ARG PINNED_TORCH_VERSION
-
-# Install torch, torchaudio and torchvision based on the input
-# if TORCH_WHEELS_PATH is default "./requirements", it will pull thethe nightly versions using pip
-# otherwise, it will use the whls from TORCH_WHEELS_PATH from the host machine
-RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
-    --mount=type=cache,target=/root/.cache/uv \
-    if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
-        torch_whl=$(find /dist -maxdepth 1 -name 'torch-*.whl' -print -quit); \
-        vision_whl=$(find /dist/vision -name 'torchvision*.whl' | head -n1 | xargs); \
-        audio_whl=$(find /dist/audio -name 'torchaudio*.whl' | head -n1 | xargs); \
-        uv pip install --system "${torch_whl}[opt-einsum]"; \
-        uv pip install --system "${vision_whl}"; \
-        uv pip install --system "${audio_whl}"; \
-    elif [ -n "$PINNED_TORCH_VERSION" ]; then \
-        echo "[INFO] Installing pinned torch nightly version: $PINNED_TORCH_VERSION"; \
-        uv pip install --system "$PINNED_TORCH_VERSION" --index-url https://download.pytorch.org/whl/nightly/cu128; \
-    else \
-        echo "[INFO] Installing torch nightly with latest one"; \
-        uv pip install --system torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128; \
-    fi
-
-# Install numba 0.61.2 for cuda environment
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system numba==0.61.2
-
-# Install common dependencies from vllm common.txt
-RUN --mount=type=cache,target=/root/.cache/uv \
-uv pip install --system -r requirements/common.txt
-
-
-# Must put before installing xformers, so it can install the correct version of xfomrers.
-ARG torch_cuda_arch_list='8.0;8.6;8.9;9.0'
-ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
-ARG max_jobs=16
-ENV MAX_JOBS=${max_jobs}
-
-# Build xformers with cuda and torch nightly/wheel
-# following official xformers guidance: https://github.com/facebookresearch/xformers#build
-ARG XFORMERS_COMMIT=f2de641ef670510cadab099ce6954031f52f191c
-ENV CCACHE_DIR=/root/.cache/ccache
-RUN --mount=type=cache,target=/root/.cache/ccache \
-    --mount=type=cache,target=/root/.cache/uv \
-    echo 'git clone xformers...' \
-    && git clone https://github.com/facebookresearch/xformers.git --recursive \
-    && cd xformers \
-    && git checkout ${XFORMERS_COMMIT} \
-    && git submodule update --init --recursive \
-    && echo 'finish git clone xformers...' \
-    && rm -rf build \
-    && python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose \
-    && cd .. \
-    && rm -rf xformers
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system xformers-dist/*.whl --verbose
-
-# Build can take a long time, and the torch nightly version fetched from url can be different in next docker stage.
-# track the nightly torch version used in the build, when we set up runtime environment we can make sure the version is the same
-RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio' > torch_build_versions.txt
-RUN cat  torch_build_versions.txt
-
-RUN pip freeze | grep -E 'torch|xformers|torchvision|torchaudio'
-
-#################### BASE BUILD IMAGE ####################
-
-
-#################### WHEEL BUILD IMAGE ####################
-# Image used to build vllm wheel
-FROM base AS build
-ARG TARGETPLATFORM
-
-ENV UV_HTTP_TIMEOUT=500
-ENV UV_INDEX_STRATEGY="unsafe-best-match"
-
-COPY . .
-
-RUN python3 use_existing_torch.py
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/build.txt
-
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != "0" ]; then bash tools/check_repo.sh ; fi
-
-# Max jobs used by Ninja to build extensions
-ARG max_jobs=16
-ENV MAX_JOBS=${max_jobs}
-ARG nvcc_threads=2
-ENV NVCC_THREADS=$nvcc_threads
-ARG torch_cuda_arch_list='8.0;8.6;8.9;9.0'
-ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
-
-ARG USE_SCCACHE
-ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
-ARG SCCACHE_REGION_NAME=us-west-2
-ARG SCCACHE_S3_NO_CREDENTIALS=0
-
-# if USE_SCCACHE is set, use sccache to speed up compilation
-RUN --mount=type=cache,target=/root/.cache/uv \
-    --mount=type=bind,source=.git,target=.git \
-    if [ "$USE_SCCACHE" = "1" ]; then \
-        echo "Installing sccache..." \
-        && curl -L -o sccache.tar.gz https://github.com/mozilla/sccache/releases/download/v0.8.1/sccache-v0.8.1-x86_64-unknown-linux-musl.tar.gz \
-        && tar -xzf sccache.tar.gz \
-        && sudo mv sccache-v0.8.1-x86_64-unknown-linux-musl/sccache /usr/bin/sccache \
-        && rm -rf sccache.tar.gz sccache-v0.8.1-x86_64-unknown-linux-musl \
-        && export SCCACHE_BUCKET=${SCCACHE_BUCKET_NAME} \
-        && export SCCACHE_REGION=${SCCACHE_REGION_NAME} \
-        && export SCCACHE_S3_NO_CREDENTIALS=${SCCACHE_S3_NO_CREDENTIALS} \
-        && export SCCACHE_IDLE_TIMEOUT=0 \
-        && export CMAKE_BUILD_TYPE=Release \
-        && sccache --show-stats \
-        && python3 setup.py bdist_wheel --dist-dir=vllm-dist --py-limited-api=cp38 \
-        && sccache --show-stats; \
-    fi
-
-ENV CCACHE_DIR=/root/.cache/ccache
-RUN --mount=type=cache,target=/root/.cache/ccache \
-    --mount=type=cache,target=/root/.cache/uv \
-    --mount=type=bind,source=.git,target=.git  \
-    if [ "$USE_SCCACHE" != "1" ]; then \
-        # Clean any existing CMake artifacts
-        rm -rf .deps && \
-        mkdir -p .deps && \
-        python3 setup.py bdist_wheel --dist-dir=vllm-dist --py-limited-api=cp38; \
-    fi
-
-RUN echo "[DEBUG] Listing  current directory:" && \
-    ls -al && \
-    echo "[DEBUG] Showing torch_build_versions.txt content:" && \
-    cat torch_build_versions.txt
-
-#################### WHEEL BUILD IMAGE ####################
-
-
-################### VLLM INSTALLED IMAGE ####################
-# Setup clean environment for vLLM for test and api server using ubuntu22.04 with AOT flashinfer
-FROM ${FINAL_BASE_IMAGE} AS vllm-base
-USER root
-# prepare for environment starts
-WORKDIR /workspace
-
-RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
-    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
-
-# Install Python and other dependencies if it does not existed
-RUN if ! command -v python3 >/dev/null || ! python3 --version | grep -q "${PYTHON_VERSION}"; then \
-      echo "Installing Python ${PYTHON_VERSION}..." && \
-      echo 'tzdata tzdata/Areas select America' | debconf-set-selections && \
-      echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections && \
-      apt-get update -y && \
-      apt-get install -y ccache software-properties-common git curl sudo && \
-      for i in 1 2 3; do \
-        add-apt-repository -y ppa:deadsnakes/ppa && break || \
-        { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
-      done && \
-      apt-get update -y && \
-      apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv && \
-      update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 && \
-      update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} && \
-      ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config && \
-      curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}; \
-   else \
-      echo "Python ${PYTHON_VERSION} already present, skipping setup."; \
-   fi \
-   && python3 --version && python3 -m pip --version
-
-
-# Get the torch versions, and whls used in previous stagtes for consistency
-COPY --from=base /workspace/torch_build_versions.txt ./torch_build_versions.txt
-COPY --from=base /workspace/xformers-dist /wheels/xformers
-COPY --from=build /workspace/vllm-dist /wheels/vllm
-RUN echo "[DEBUG] Listing current directory before torch install step:" && \
-    ls -al && \
-    echo "[DEBUG] Showing torch_build_versions.txt content:" && \
-    cat torch_build_versions.txt
-
-# Workaround for https://github.com/openai/triton/issues/2507 and
-# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
-# this won't be needed for future versions of this docker image
-# or future versions of triton.
-RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
-
-
-# Install uv for faster pip installs if not existed
-RUN --mount=type=cache,target=/root/.cache/uv \
-    if ! python3 -m uv --version > /dev/null 2>&1; then \
-        python3 -m pip install uv==0.8.4; \
-    fi
-ENV UV_HTTP_TIMEOUT=500
-ENV UV_INDEX_STRATEGY="unsafe-best-match"
-
-# Default mount file as placeholder, this just avoid the mount error
-ARG TORCH_WHEELS_PATH="./requirements"
-# Install torch, torchaudio and torchvision
-# if TORCH_WHEELS_PATH is default "./requirements", it will pull the nightly versions using pip using torch_build_versions.txt
-# otherwise, it will use the whls from TORCH_WHEELS_PATH from the host machine
-RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
-    --mount=type=cache,target=/root/.cache/uv \
-    if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
-        torch_whl=$(find /dist -maxdepth 1 -name 'torch-*.whl' -print -quit); \
-        vision_whl=$(find /dist/vision -name 'torchvision*.whl' | head -n1 | xargs); \
-        audio_whl=$(find /dist/audio -name 'torchaudio*.whl' | head -n1 | xargs); \
-        echo "Found: '${torch_whl}' '${audio_whl}' '${vision_whl}'"; \
-        uv pip install --system "${torch_whl}[opt-einsum]"; \
-        uv pip install --system "${vision_whl}"; \
-        uv pip install --system "${audio_whl}"; \
-    else \
-        echo "[INFO] Installing torch versions from torch_build_versions.txt"; \
-        uv pip install --system $(cat torch_build_versions.txt | xargs) --index-url https://download.pytorch.org/whl/nightly/cu128; \
-    fi
-
-# Install the vllm wheel from previous stage
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system /wheels/vllm/*.whl --verbose
-
-# Install xformers wheel from previous stage
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system /wheels/xformers/*.whl --verbose
-
-
-# Build flashinfer from source.
-ARG torch_cuda_arch_list='8.0;8.9;9.0a'
-# install package for build flashinfer
-# see issue: https://github.com/flashinfer-ai/flashinfer/issues/738
-
-RUN pip install build==1.3.0
-RUN pip freeze | grep -E 'setuptools|packaging|build'
-
-ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
-# Build flashinfer for torch nightly from source around 10 mins
-ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
-# Keep this in sync with https://github.com/vllm-project/vllm/blob/main/requirements/cuda.txt
-ARG FLASHINFER_GIT_REF="v0.2.9rc2"
-RUN --mount=type=cache,target=/root/.cache/uv \
-    git clone --depth 1 --recursive --shallow-submodules \
-        --branch ${FLASHINFER_GIT_REF} \
-        ${FLASHINFER_GIT_REPO} flashinfer \
-    && echo "Building FlashInfer with AOT for arches: ${torch_cuda_arch_list}" \
-    && cd flashinfer \
-    && python3 -m flashinfer.aot \
-    && python3 -m build --no-isolation --wheel --outdir ../wheels/flashinfer \
-    && cd .. \
-    && rm -rf flashinfer
-
-# install flashinfer python
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system wheels/flashinfer/*.whl --verbose
-
-# Logging to confirm the torch versions
-RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'
-################### VLLM INSTALLED IMAGE ####################
-
-
-#################### UNITTEST IMAGE #############################
-FROM vllm-base as test
-
-ENV UV_HTTP_TIMEOUT=500
-ENV UV_INDEX_STRATEGY="unsafe-best-match"
-
-COPY tests/ tests/
-COPY examples examples
-COPY benchmarks benchmarks
-COPY ./vllm/collect_env.py .
-COPY requirements/common.txt requirements/common.txt
-COPY use_existing_torch.py use_existing_torch.py
-COPY pyproject.toml pyproject.toml
-# Install build and runtime dependencies without stable torch version
-COPY requirements/nightly_torch_test.txt requirements/nightly_torch_test.txt
-
-RUN python3 use_existing_torch.py
-
-# install packages
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/common.txt
-# enable fast downloads from hf (for testing)
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system hf_transfer
-ENV HF_HUB_ENABLE_HF_TRANSFER 1
-
-# install development dependencies (for testing)
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -e tests/vllm_test_utils
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/nightly_torch_test.txt
-
-# Workaround for #17068
-# pinned commit for v2.2.4
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system --no-build-isolation "git+https://github.com/state-spaces/mamba@95d8aba8a8c75aedcaa6143713b11e745e7cd0d9#egg=mamba-ssm"
-
-# Logging to confirm the torch versions
-RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'
-
-# Logging to confirm all the packages are installed
-RUN pip freeze
-
-#################### UNITTEST IMAGE #############################
-
-#################### EXPORT STAGE ####################
-FROM scratch as export-wheels
-
-# Just copy the wheels we prepared in previous stages
-COPY --from=base /workspace/xformers-dist /wheels/xformers
-COPY --from=build /workspace/vllm-dist /wheels/vllm
-COPY --from=vllm-base /workspace/wheels/flashinfer /wheels/flashinfer-python

.github/merge_rules.yaml

@@ -488,10 +488,6 @@
   - torch/_dynamo/**
   - torch/csrc/dynamo/**
   - test/dynamo/**
-  - test/dynamo_expected_failures/**
-  - test/dynamo_skips/**
-  - test/inductor_expected_failures/**
-  - test/inductor_skips/**
   approved_by:
   - guilhermeleobas
   mandatory_checks_name:

.github/pytorch-probot.yml

@@ -22,7 +22,6 @@ ciflow_push_tags:
 - ciflow/rocm
 - ciflow/rocm-mi300
 - ciflow/s390
-- ciflow/riscv64
 - ciflow/slow
 - ciflow/trunk
 - ciflow/unstable

.github/requirements-gha-cache.txt

@@ -7,9 +7,9 @@
 #   .ci/docker/requirements-ci.txt
 boto3==1.35.42
 jinja2==3.1.6
-lintrunner==0.12.7
+lintrunner==0.10.7
 ninja==1.10.0.post1
 nvidia-ml-py==11.525.84
-pyyaml==6.0.2
+pyyaml==6.0
 requests==2.32.4
-rich==14.1.0
+rich==10.9.0

.github/requirements/conda-env-macOS-ARM64

@@ -0,0 +1,5 @@
+# Not pinning certifi so that we can always get the latest certificates
+certifi
+pip=23.2.1
+pkg-config=0.29.2
+wheel=0.37.1

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

@@ -2,7 +2,7 @@ boto3==1.35.42
 cmake==3.27.*
 expecttest==0.3.0
 fbscribelogger==0.1.7
-filelock==3.18.0
+filelock==3.6.0
 hypothesis==6.56.4
 librosa>=0.6.2
 mpmath==1.3.0

.github/scripts/amd/package_triton_wheel.sh

@@ -1,4 +1,3 @@
-#!/bin/bash
 set -ex
 
 # Set ROCM_HOME isn't available, use ROCM_PATH if set or /opt/rocm
@@ -51,15 +50,29 @@ do
     cp $lib $TRITON_ROCM_DIR/lib/
 done
 
+# Required ROCm libraries
+if [[ "${MAJOR_VERSION}" == "6" ]]; then
+    libamdhip="libamdhip64.so.6"
+else
+    libamdhip="libamdhip64.so.5"
+fi
+
 # Required ROCm libraries - ROCm 6.0
 ROCM_SO=(
-    "libamdhip64.so"
-    "libhsa-runtime64.so"
-    "libdrm.so"
-    "libdrm_amdgpu.so"
-    "libamd_comgr.so"
-    "librocprofiler-register.so"
+    "${libamdhip}"
+    "libhsa-runtime64.so.1"
+    "libdrm.so.2"
+    "libdrm_amdgpu.so.1"
 )
+if [[ $ROCM_INT -ge 60400 ]]; then
+    ROCM_SO+=("libamd_comgr.so.3")
+else
+    ROCM_SO+=("libamd_comgr.so.2")
+fi
+
+if [[ $ROCM_INT -ge 60100 ]]; then
+    ROCM_SO+=("librocprofiler-register.so.0")
+fi
 
 for lib in "${ROCM_SO[@]}"
 do
@@ -81,6 +94,10 @@ do
     fi
 
     cp $file_path $TRITON_ROCM_DIR/lib
+    # When running locally, and not building a wheel, we need to satisfy shared objects requests that don't look for versions
+    LINKNAME=$(echo $lib | sed -e 's/\.so.*/.so/g')
+    ln -sf $lib $TRITON_ROCM_DIR/lib/$LINKNAME
+
 done
 
 # Copy Include Files

.github/scripts/amd/patch_triton_wheel.sh

@@ -19,13 +19,15 @@ replace_needed_sofiles() {
     find $1 -name '*.so*' -o -name 'ld.lld' | while read sofile; do
         origname=$2
         patchedname=$3
-        set +e
-        origname=$($PATCHELF_BIN --print-needed $sofile | grep "$origname.*")
-        ERRCODE=$?
-        set -e
-        if [ "$ERRCODE" -eq "0" ]; then
-            echo "patching $sofile entry $origname to $patchedname"
-            $PATCHELF_BIN --replace-needed $origname $patchedname $sofile
+        if [[ "$origname" != "$patchedname" ]]; then
+            set +e
+            origname=$($PATCHELF_BIN --print-needed $sofile | grep "$origname.*")
+            ERRCODE=$?
+            set -e
+            if [ "$ERRCODE" -eq "0" ]; then
+                echo "patching $sofile entry $origname to $patchedname"
+                $PATCHELF_BIN --replace-needed $origname $patchedname $sofile
+            fi
         fi
     done
 }

.github/scripts/generate_binary_build_matrix.py

@@ -193,7 +193,7 @@ LIBTORCH_CONTAINER_IMAGES: dict[str, str] = {
     "cpu": "libtorch-cxx11-builder:cpu",
 }
 
-FULL_PYTHON_VERSIONS = ["3.9", "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t"]
+FULL_PYTHON_VERSIONS = ["3.9", "3.10", "3.11", "3.12", "3.13", "3.13t"]
 
 
 def translate_desired_cuda(gpu_arch_type: str, gpu_arch_version: str) -> str:
@@ -273,6 +273,7 @@ def generate_wheels_matrix(
     os: str,
     arches: Optional[list[str]] = None,
     python_versions: Optional[list[str]] = None,
+    use_split_build: bool = False,
 ) -> list[dict[str, str]]:
     package_type = "wheel"
     if os == "linux" or os == "linux-aarch64" or os == "linux-s390x":
@@ -314,11 +315,15 @@ def generate_wheels_matrix(
             # TODO: Enable python 3.13t on cpu-s390x
             if gpu_arch_type == "cpu-s390x" and python_version == "3.13t":
                 continue
-            # TODO: Enable python 3.14 on non linux OSes
-            if os != "linux" and (
-                python_version == "3.14" or python_version == "3.14t"
+
+            if use_split_build and (
+                arch_version not in ["12.6", "12.8", "12.9", "cpu"] or os != "linux"
             ):
-                continue
+                raise RuntimeError(
+                    "Split build is only supported on linux with cuda 12* and cpu.\n"
+                    f"Currently attempting to build on arch version {arch_version} and os {os}.\n"
+                    "Please modify the matrix generation to exclude this combination."
+                )
 
             # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
 
@@ -334,6 +339,7 @@ def generate_wheels_matrix(
                         "gpu_arch_type": gpu_arch_type,
                         "gpu_arch_version": gpu_arch_version,
                         "desired_cuda": desired_cuda,
+                        "use_split_build": "True" if use_split_build else "False",
                         "container_image": WHEEL_CONTAINER_IMAGES[arch_version].split(
                             ":"
                         )[0],
@@ -366,6 +372,7 @@ def generate_wheels_matrix(
                             "desired_cuda": translate_desired_cuda(
                                 gpu_arch_type, gpu_arch_version
                             ),
+                            "use_split_build": "True" if use_split_build else "False",
                             "container_image": WHEEL_CONTAINER_IMAGES[
                                 arch_version
                             ].split(":")[0],
@@ -388,6 +395,7 @@ def generate_wheels_matrix(
                         "desired_cuda": translate_desired_cuda(
                             gpu_arch_type, gpu_arch_version
                         ),
+                        "use_split_build": "True" if use_split_build else "False",
                         "container_image": WHEEL_CONTAINER_IMAGES[arch_version].split(
                             ":"
                         )[0],

.github/scripts/generate_ci_workflows.py

@@ -59,7 +59,9 @@ class BinaryBuildWorkflow:
     is_scheduled: str = ""
     branches: str = "nightly"
     # Mainly for macos
+    cross_compile_arm64: bool = False
     macos_runner: str = "macos-14-xlarge"
+    use_split_build: bool = False
     # Mainly used for libtorch builds
     build_variant: str = ""
 
@@ -70,6 +72,9 @@ class BinaryBuildWorkflow:
                 for item in [self.os, "binary", self.package_type, self.build_variant]
                 if item != ""
             )
+        if self.use_split_build:
+            # added to distinguish concurrency groups
+            self.build_environment += "-split"
 
     def generate_workflow_file(self, workflow_template: jinja2.Template) -> None:
         output_file_path = (
@@ -112,6 +117,21 @@ LINUX_BINARY_BUILD_WORFKLOWS = [
             isolated_workflow=True,
         ),
     ),
+    # See https://github.com/pytorch/pytorch/issues/138750
+    #   BinaryBuildWorkflow(
+    #     os=OperatingSystem.LINUX,
+    #     package_type="manywheel",
+    #     build_configs=generate_binary_build_matrix.generate_wheels_matrix(
+    #         OperatingSystem.LINUX,
+    #         use_split_build=True,
+    #         arches=["11.8", "12.1", "12.4", "cpu"],
+    #     ),
+    #     ciflow_config=CIFlowConfig(
+    #         labels={LABEL_CIFLOW_BINARIES, LABEL_CIFLOW_BINARIES_WHEEL},
+    #         isolated_workflow=True,
+    #     ),
+    #     use_split_build=True,
+    # ),
     BinaryBuildWorkflow(
         os=OperatingSystem.LINUX,
         package_type="libtorch",
@@ -155,11 +175,27 @@ LINUX_BINARY_SMOKE_WORKFLOWS = [
         package_type="manywheel",
         build_configs=generate_binary_build_matrix.generate_wheels_matrix(
             OperatingSystem.LINUX,
-            arches=["12.8"],
-            python_versions=["3.12"],
+            arches=["12.6", "12.8", "12.9"],
+            python_versions=["3.9"],
         ),
         branches="main",
     ),
+    # See https://github.com/pytorch/pytorch/issues/138750
+    # BinaryBuildWorkflow(
+    #     os=OperatingSystem.LINUX,
+    #     package_type="manywheel",
+    #     build_configs=generate_binary_build_matrix.generate_wheels_matrix(
+    #         OperatingSystem.LINUX,
+    #         arches=["11.8", "12.1", "12.4"],
+    #         python_versions=["3.9"],
+    #         use_split_build=True,
+    #     ),
+    #     ciflow_config=CIFlowConfig(
+    #         labels={LABEL_CIFLOW_PERIODIC},
+    #     ),
+    #     branches="main",
+    #     use_split_build=True,
+    # ),
     BinaryBuildWorkflow(
         os=OperatingSystem.LINUX,
         package_type="libtorch",
@@ -302,6 +338,7 @@ MACOS_BINARY_BUILD_WORKFLOWS = [
             generate_binary_build_matrix.RELEASE,
             libtorch_variants=["shared-with-deps"],
         ),
+        cross_compile_arm64=False,
         macos_runner="macos-14-xlarge",
         ciflow_config=CIFlowConfig(
             labels={LABEL_CIFLOW_BINARIES, LABEL_CIFLOW_BINARIES_LIBTORCH},
@@ -314,6 +351,7 @@ MACOS_BINARY_BUILD_WORKFLOWS = [
         build_configs=generate_binary_build_matrix.generate_wheels_matrix(
             OperatingSystem.MACOS_ARM64
         ),
+        cross_compile_arm64=False,
         macos_runner="macos-14-xlarge",
         ciflow_config=CIFlowConfig(
             labels={LABEL_CIFLOW_BINARIES, LABEL_CIFLOW_BINARIES_WHEEL},

.github/scripts/lintrunner.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # Use uv to speed up lintrunner init
-python3 -m pip install -U uv==0.8.* setuptools
+python3 -m pip install uv==0.1.45 setuptools
 
 CACHE_DIRECTORY="/tmp/.lintbin"
 # Try to recover the cached binaries

.github/scripts/runner_determinator.py

@@ -262,12 +262,7 @@ def is_exception_branch(branch: str) -> bool:
     """
     Branches that get opted out of experiments by default, until they're explicitly enabled.
     """
-    return branch.split("/", maxsplit=1)[0] in {
-        "main",
-        "nightly",
-        "release",
-        "landchecks",
-    }
+    return branch.split("/")[0] in {"main", "nightly", "release", "landchecks"}
 
 
 def load_yaml(yaml_text: str) -> Any:

.github/scripts/test_trymerge.py

@@ -70,9 +70,6 @@ def mock_query(
     if key in mocked_queries:
         return mocked_queries[key]
 
-    # TODO: Remove me once https://github.com/pytorch/pytorch/issues/160489 is resolved
-    raise ValueError(f"Key {key} could not be found in gql_mocks")
-
     try:
         rc = fallback_function(*args)
     except HTTPError as err:

.github/scripts/trymerge.py

@@ -108,6 +108,10 @@ GH_CHECKSUITES_FRAGMENT = """
 fragment PRCheckSuites on CheckSuiteConnection {
   edges {
     node {
+      app {
+        name
+        databaseId
+      }
       workflowRun {
         workflow {
           name
@@ -1887,9 +1891,7 @@ def validate_revert(
         else pr.get_comment_by_id(comment_id)
     )
     if comment.editor_login is not None:
-        raise PostCommentError(
-            "Halting the revert as the revert comment has been edited."
-        )
+        raise PostCommentError("Don't want to revert based on edited command")
     author_association = comment.author_association
     author_login = comment.author_login
     allowed_reverters = ["COLLABORATOR", "MEMBER", "OWNER"]

.github/scripts/windows/build_triton.bat

@@ -10,7 +10,7 @@ if "%PY_VERS%" == "3.13t" (
     call conda create -n %PYTHON_PREFIX% -y -c=conda-forge python=%PY_VERS%
 )
 :: Fix cmake version for issue https://github.com/pytorch/pytorch/issues/150480
-call conda run -n %PYTHON_PREFIX% pip install wheel pybind11 certifi cython cmake==3.31.6 setuptools==72.1.0 ninja==1.11.1.4
+call conda run -n %PYTHON_PREFIX% pip install wheel pybind11 certifi cython cmake==3.31.6 setuptools==72.1.0 ninja
 
 dir "%VC_INSTALL_PATH%"
 

.github/templates/macos_binary_build_workflow.yml.j2

@@ -47,6 +47,9 @@ env:
   GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
   PR_NUMBER: ${{ github.event.pull_request.number }}
   SKIP_ALL_TESTS: 0
+{%- if cross_compile_arm64 %}
+  CROSS_COMPILE_ARM64: 1
+{% endif %}
 !{{ common.concurrency(build_environment) }}
 
 jobs:

.github/templates/upload.yml.j2

@@ -25,6 +25,11 @@
       DOCKER_IMAGE: !{{ config["container_image"] }}
       DOCKER_IMAGE_TAG_PREFIX: !{{ config["container_image_tag_prefix"] }}
 {%- endif %}
+{%- if config["package_type"] == "manywheel" %}
+  {%- if config.use_split_build is defined %}
+      use_split_build: !{{ config["use_split_build"] }}
+  {%- endif %}
+{%- endif %}
 {%- if config["package_type"] == "libtorch" %}
   {%- if config["libtorch_config"] %}
       LIBTORCH_CONFIG: !{{ config["libtorch_config"] }}

.github/workflows/_binary-build-linux.yml

@@ -26,6 +26,13 @@ on:
         default: 240
         type: number
         description: timeout for the job
+      use_split_build:
+        description: |
+          [Experimental] Build a libtorch only wheel and build pytorch such that
+          are built from the libtorch wheel.
+        required: false
+        type: boolean
+        default: false
       ALPINE_IMAGE:
         required: false
         type: string
@@ -110,6 +117,7 @@ jobs:
       PR_NUMBER: ${{ github.event.pull_request.number }}
       PYTORCH_FINAL_PACKAGE_DIR: /artifacts
       SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
+      USE_SPLIT_BUILD: ${{ inputs.use_split_build }}
     steps:
       - name: Make the env permanent during this workflow (but not the secrets)
         shell: bash
@@ -134,6 +142,7 @@ jobs:
             echo "PR_NUMBER=${{ env.PR_NUMBER }}"
             echo "PYTORCH_FINAL_PACKAGE_DIR=${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
             echo "SHA1=${{ env.SHA1 }}"
+            echo "USE_SPLIT_BUILD=${{ env.use_split_build }}"
           } >> "${GITHUB_ENV} }}"
 
       - name: List the env
@@ -252,6 +261,7 @@ jobs:
             -e PYTORCH_ROOT \
             -e SKIP_ALL_TESTS \
             -e PYTORCH_EXTRA_INSTALL_REQUIREMENTS \
+            -e USE_SPLIT_BUILD \
             --tty \
             --detach \
             -v "${GITHUB_WORKSPACE}/pytorch:/pytorch" \

.github/workflows/_binary-test-linux.yml

@@ -64,6 +64,13 @@ on:
         required: true
         type: string
         description: Hardware to run this job on. Valid values are linux.4xlarge, linux.4xlarge.nvidia.gpu, linux.arm64.2xlarge, and linux.rocm.gpu
+      use_split_build:
+        description: |
+          [Experimental] Build a libtorch only wheel and build pytorch such that
+          are built from the libtorch wheel.
+        required: false
+        type: boolean
+        default: false
     secrets:
       github-token:
         required: true
@@ -97,6 +104,7 @@ jobs:
       PR_NUMBER: ${{ github.event.pull_request.number }}
       PYTORCH_FINAL_PACKAGE_DIR: /artifacts
       SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
+      USE_SPLIT_BUILD: ${{ inputs.use_split_build }}
     steps:
       - name: Make the env permanent during this workflow (but not the secrets)
         shell: bash
@@ -121,6 +129,7 @@ jobs:
             echo "PR_NUMBER=${{ env.PR_NUMBER }}"
             echo "PYTORCH_FINAL_PACKAGE_DIR=${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
             echo "SHA1=${{ env.SHA1 }}"
+            echo "USE_SPLIT_BUILD=${{ env.USE_SPLIT_BUILD }}"
           } >> "${GITHUB_ENV} }}"
 
       - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"

.github/workflows/_binary-upload.yml

@@ -51,6 +51,13 @@ on:
         required: false
         type: string
         description: Desired python version
+      use_split_build:
+        description: |
+          [Experimental] Build a libtorch only wheel and build pytorch such that
+          are built from the libtorch wheel.
+        required: false
+        type: boolean
+        default: false
     secrets:
       github-token:
         required: true
@@ -79,6 +86,7 @@ jobs:
       PR_NUMBER: ${{ github.event.pull_request.number }}
       PYTORCH_FINAL_PACKAGE_DIR: /artifacts
       SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
+      USE_SPLIT_BUILD: ${{ inputs.use_split_build }}
     steps:
       - name: Checkout PyTorch
         uses: pytorch/pytorch/.github/actions/checkout-pytorch@main

.github/workflows/_linux-build.yml

@@ -287,36 +287,10 @@ jobs:
           # comes from https://github.com/pytorch/test-infra/pull/6058
           TOTAL_MEMORY_WITH_SWAP=$(("${TOTAL_AVAILABLE_MEMORY_IN_GB%.*}" + 3))
 
-          if [[ ${BUILD_ENVIRONMENT} == *"riscv64"* ]]; then
-            # EC2 specific setup for RISC-V emulation
-            # Ensure binfmt_misc is available
-            echo "Mounting binfmt_misc filesystem"
-            sudo mount binfmt_misc -t binfmt_misc /proc/sys/fs/binfmt_misc 2>/dev/null || true
-
-            echo "QEMU registration: multiarch/qemu-user-static"
-            docker run --rm --privileged multiarch/qemu-user-static --reset -p yes || true
-
-            # Final verification
-            echo "Checking binfmt_misc status:"
-            ls -la /proc/sys/fs/binfmt_misc/ 2>/dev/null || echo "Cannot access binfmt_misc directory"
-
-            if [ -f /proc/sys/fs/binfmt_misc/qemu-riscv64 ]; then
-              echo "qemu-riscv64 registration successful"
-            else
-              echo "qemu-riscv64 registration failed - proceeding without emulation"
-              echo "This may cause RISC-V builds to fail"
-            fi
-
-            RISCV_DOCKER_ARGS="--privileged"
-          else
-            RISCV_DOCKER_ARGS=
-          fi
-
           # detached container should get cleaned up by teardown_ec2_linux
           # Used for JENKINS_USER and DOCKER_SHELL_CMD, which can be empty
           # shellcheck disable=SC2086
           container_name=$(docker run \
-            ${RISCV_DOCKER_ARGS} \
             -e BUILD_ENVIRONMENT \
             -e MAX_JOBS="$(nproc --ignore=2)" \
             -e PR_NUMBER \
@@ -332,6 +306,7 @@ jobs:
             -e OUR_GITHUB_JOB_ID \
             -e HUGGING_FACE_HUB_TOKEN \
             -e SCRIBE_GRAPHQL_ACCESS_TOKEN \
+            -e USE_SPLIT_BUILD \
             -e BUILD_ADDITIONAL_PACKAGES \
             --memory="${TOTAL_AVAILABLE_MEMORY_IN_GB%.*}g" \
             --memory-swap="${TOTAL_MEMORY_WITH_SWAP}g" \

.github/workflows/_linux-test.yml

@@ -96,7 +96,7 @@ jobs:
     steps:
       - name: Setup SSH (Click me for login details)
         uses: pytorch/test-infra/.github/actions/setup-ssh@main
-        if: ${{ !contains(matrix.runner, 'b200') && inputs.build-environment != 'linux-s390x-binary-manywheel' }}
+        if: ${{ matrix.runner != 'B200' && inputs.build-environment != 'linux-s390x-binary-manywheel' }}
         with:
           github-secret: ${{ secrets.GITHUB_TOKEN }}
           instructions: |
@@ -109,7 +109,7 @@ jobs:
           no-sudo: true
 
       - name: Setup Python
-        if: contains(matrix.runner, 'b200')
+        if: matrix.runner == 'B200'
         uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
         with:
           python-version: '3.12'
@@ -117,7 +117,7 @@ jobs:
 
       - name: Setup Linux
         uses: ./.github/actions/setup-linux
-        if: inputs.build-environment != 'linux-s390x-binary-manywheel' && !contains(matrix.runner, 'b200')
+        if: inputs.build-environment != 'linux-s390x-binary-manywheel' && matrix.runner != 'B200'
 
       - name: configure aws credentials
         if: ${{ inputs.aws-role-to-assume != '' && inputs.build-environment != 'linux-s390x-binary-manywheel' }}
@@ -128,7 +128,7 @@ jobs:
           aws-region: us-east-1
 
       - name: Login to Amazon ECR
-        if: ${{ inputs.aws-role-to-assume != '' && contains(matrix.runner, 'b200') }}
+        if: ${{ inputs.aws-role-to-assume != '' && matrix.runner == 'B200' }}
         id: login-ecr
         continue-on-error: true
         uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1
@@ -166,17 +166,17 @@ jobs:
         uses: pytorch/test-infra/.github/actions/setup-nvidia@main
         with:
           driver-version: ${{ matrix.config == 'legacy_nvidia_driver' && '525.105.17' || '570.133.07' }}
-        if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' && !contains(matrix.runner, 'b200') }}
+        if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' && matrix.runner != 'B200' }}
 
       - name: Setup GPU_FLAG for docker run
         id: setup-gpu-flag
         run: echo "GPU_FLAG=--gpus all -e NVIDIA_DRIVER_CAPABILITIES=all" >> "${GITHUB_ENV}"
-        if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && (steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'true' || contains(matrix.runner, 'b200')) }}
+        if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && (steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'true' || matrix.runner == 'B200') }}
 
       - name: Setup SCCACHE_SERVER_PORT environment for docker run when on container
         id: setup-sscache-port-flag
         run: echo "SCCACHE_SERVER_PORT_DOCKER_FLAG=-e SCCACHE_SERVER_PORT=$((RUNNER_UID + 4226))" >> "${GITHUB_ENV}"
-        if: ${{ steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'true' && !contains(matrix.runner, 'b200') }}
+        if: ${{ steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'true' && matrix.runner != 'B200' }}
 
       - name: Lock NVIDIA A100 40GB Frequency
         run: |
@@ -277,8 +277,8 @@ jobs:
           NO_TD: ${{ steps.keep-going.outputs.ci-no-td }}
           TD_DISTRIBUTED: ${{ steps.keep-going.outputs.ci-td-distributed }}
           # Do not set SCCACHE_S3_KEY_PREFIX to share the cache between all build jobs
-          SCCACHE_BUCKET: ${{ !contains(matrix.runner, 'b200') && 'ossci-compiler-cache-circleci-v2' || '' }}
-          SCCACHE_REGION: ${{ !contains(matrix.runner, 'b200') && 'us-east-1' || '' }}
+          SCCACHE_BUCKET: ${{ matrix.runner != 'B200' && 'ossci-compiler-cache-circleci-v2' || '' }}
+          SCCACHE_REGION: ${{ matrix.runner != 'B200' && 'us-east-1' || '' }}
           SHM_SIZE: ${{ contains(inputs.build-environment, 'cuda') && '2g' || '1g' }}
           DOCKER_IMAGE: ${{ inputs.docker-image }}
           XLA_CUDA: ${{ contains(inputs.build-environment, 'xla') && '0' || '' }}
@@ -403,7 +403,7 @@ jobs:
           job_identifier: ${{ github.workflow }}_${{ inputs.build-environment }}
 
       - name: Authenticate with AWS
-        if: ${{ contains(matrix.runner, 'b200') }}
+        if: ${{ matrix.runner == 'B200' }}
         uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
         with:
           role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_upload-benchmark-results

.github/workflows/_rocm-test.yml

@@ -88,16 +88,6 @@ jobs:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
 
-      - name: Runner check GPU count (distributed jobs)
-        if: ${{ contains(matrix.config, 'distributed') }}
-        shell: bash
-        run: |
-          ngpu=$(rocminfo | grep -c -E 'Name:.*\sgfx')
-          if [[ $ngpu -lt 4 ]]; then
-            echo "Error: only $ngpu GPU(s) detected, at least 4 GPUs are needed for distributed jobs"
-            exit 1
-          fi
-
       - name: configure aws credentials
         id: aws_creds
         uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
@@ -279,8 +269,8 @@ jobs:
           # copy test results back to the mounted workspace, needed sudo, resulting permissions were correct
           docker exec -t "${{ env.CONTAINER_NAME }}" sh -c "cd ../pytorch && sudo cp -R test/test-reports ../workspace/test"
 
-      - name: Change permissions (only needed for kubernetes runners for now)
-        if: ${{ always() && steps.test.conclusion && (contains(matrix.runner, 'gfx942') || contains(matrix.runner, 'mi355')) }}
+      - name: Change permissions (only needed for MI300 and MI355 kubernetes runners for now)
+        if: ${{ always() && steps.test.conclusion && (contains(matrix.runner, 'mi300') || contains(matrix.runner, 'mi355')) }}
         run: |
           docker exec -t "${{ env.CONTAINER_NAME }}" sh -c "sudo chown -R 1001:1001 test"
 

.github/workflows/build-triton-wheel.yml

@@ -50,7 +50,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        py_vers: [ "3.9", "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t" ]
+        py_vers: [ "3.9", "3.10", "3.11", "3.12", "3.13", "3.13t" ]
         device: ["cuda", "rocm", "xpu", "aarch64"]
         docker-image: ["pytorch/manylinux2_28-builder:cpu"]
         include:
@@ -126,12 +126,6 @@ jobs:
           3.13t)
             PYTHON_EXECUTABLE=/opt/python/cp313-cp313t/bin/python
             ;;
-          3.14)
-            PYTHON_EXECUTABLE=/opt/python/cp314-cp314/bin/python
-            ;;
-          3.14t)
-            PYTHON_EXECUTABLE=/opt/python/cp314-cp314t/bin/python
-            ;;
           *)
             echo "Unsupported python version ${PY_VERS}"
             exit 1

.github/workflows/check_mergeability_ghstack.yml

@@ -56,7 +56,7 @@ jobs:
           cache: pip
           architecture: x64
 
-      - run: pip install pyyaml==6.0.2
+      - run: pip install pyyaml==6.0
         shell: bash
 
       - name: Verify mergeability

.github/workflows/cherry-pick.yml

@@ -26,7 +26,7 @@ jobs:
           cache: pip
 
       # Not the direct dependencies but the script uses trymerge
-      - run: pip install pyyaml==6.0.2
+      - run: pip install pyyaml==6.0
 
       - name: Setup committer id
         run: |

.github/workflows/docker-builds.yml

@@ -51,17 +51,21 @@ jobs:
         docker-image-name: [
           pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-vllm,
+          pytorch-linux-jammy-cuda12.6-cudnn9-py3-gcc9-inductor-benchmarks,
+          pytorch-linux-jammy-cuda12.6-cudnn9-py3.12-gcc9-inductor-benchmarks,
+          pytorch-linux-jammy-cuda12.6-cudnn9-py3.13-gcc9-inductor-benchmarks,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9,
           pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11,
           pytorch-linux-jammy-py3.9-clang12,
+          pytorch-linux-jammy-py3.11-clang12,
+          pytorch-linux-jammy-py3.12-clang12,
           pytorch-linux-jammy-py3.13-clang12,
           pytorch-linux-jammy-rocm-n-py3,
           pytorch-linux-noble-rocm-n-py3,
           pytorch-linux-noble-rocm-alpha-py3,
-          pytorch-linux-jammy-rocm-n-py3-benchmarks,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12,
           pytorch-linux-jammy-py3.9-gcc11,
           pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks,
@@ -72,10 +76,8 @@ jobs:
           pytorch-linux-jammy-py3-clang12-onnx,
           pytorch-linux-jammy-linter,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter,
-          # Executorch pin needs update
-          # pytorch-linux-jammy-py3-clang12-executorch,
-          pytorch-linux-jammy-py3.12-triton-cpu,
-          pytorch-linux-noble-riscv64-py3.12-gcc14
+          pytorch-linux-jammy-py3-clang12-executorch,
+          pytorch-linux-jammy-py3.12-triton-cpu
         ]
         include:
           - docker-image-name: pytorch-linux-jammy-aarch64-py3.10-gcc11

.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml

@@ -60,6 +60,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -83,6 +84,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       build_name: manywheel-py3_9-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -106,6 +108,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       build_name: manywheel-py3_9-cpu-aarch64
     secrets:
@@ -126,6 +129,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -152,6 +156,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       build_name: manywheel-py3_9-cuda-aarch64-12_9
     secrets:
@@ -171,6 +176,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -194,6 +200,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       build_name: manywheel-py3_10-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -217,6 +224,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       build_name: manywheel-py3_10-cpu-aarch64
     secrets:
@@ -237,6 +245,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -263,6 +272,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       build_name: manywheel-py3_10-cuda-aarch64-12_9
     secrets:
@@ -282,6 +292,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -305,6 +316,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       build_name: manywheel-py3_11-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -328,6 +340,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       build_name: manywheel-py3_11-cpu-aarch64
     secrets:
@@ -348,6 +361,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -374,6 +388,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       build_name: manywheel-py3_11-cuda-aarch64-12_9
     secrets:
@@ -393,6 +408,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -416,6 +432,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       build_name: manywheel-py3_12-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -439,6 +456,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       build_name: manywheel-py3_12-cpu-aarch64
     secrets:
@@ -459,6 +477,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -485,6 +504,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       build_name: manywheel-py3_12-cuda-aarch64-12_9
     secrets:
@@ -504,6 +524,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -527,6 +548,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       build_name: manywheel-py3_13-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -550,6 +572,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       build_name: manywheel-py3_13-cpu-aarch64
     secrets:
@@ -570,6 +593,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -596,6 +620,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       build_name: manywheel-py3_13-cuda-aarch64-12_9
     secrets:
@@ -615,6 +640,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.13t"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -638,6 +664,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.13t"
       build_name: manywheel-py3_13t-cpu-aarch64
       build_environment: linux-aarch64-binary-manywheel
@@ -661,6 +688,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-aarch64
       DOCKER_IMAGE: manylinux2_28_aarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
+      use_split_build: False
       DESIRED_PYTHON: "3.13t"
       build_name: manywheel-py3_13t-cpu-aarch64
     secrets:
@@ -681,6 +709,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.13t"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.arm64.m7g.4xlarge.ephemeral
@@ -707,6 +736,7 @@ jobs:
       GPU_ARCH_TYPE: cuda-aarch64
       DOCKER_IMAGE: manylinuxaarch64-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
       DESIRED_PYTHON: "3.13t"
       build_name: manywheel-py3_13t-cuda-aarch64-12_9
     secrets:

.github/workflows/generated-linux-binary-manywheel-main.yml

@@ -42,7 +42,54 @@ jobs:
       issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
       curr_branch: ${{ github.head_ref || github.ref_name }}
       curr_ref_type: ${{ github.ref_type }}
-  manywheel-py3_12-cuda12_8-build:
+  manywheel-py3_9-cuda12_6-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: 12.6
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      use_split_build: False
+      DESIRED_PYTHON: "3.9"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_9-cuda12_6
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_9-cuda12_6-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_9-cuda12_6-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu126
+      GPU_ARCH_VERSION: 12.6
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.6
+      use_split_build: False
+      DESIRED_PYTHON: "3.9"
+      build_name: manywheel-py3_9-cuda12_6
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.4xlarge.nvidia.gpu # for other cuda versions, we use 4xlarge runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+
+  manywheel-py3_9-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -56,17 +103,18 @@ jobs:
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.12"
+      use_split_build: False
+      DESIRED_PYTHON: "3.9"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_12-cuda12_8
+      build_name: manywheel-py3_9-cuda12_8
       build_environment: linux-binary-manywheel
       PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_12-cuda12_8-test:  # Testing
+  manywheel-py3_9-cuda12_8-test:  # Testing
     if: ${{ github.repository_owner == 'pytorch' }}
     needs:
-      - manywheel-py3_12-cuda12_8-build
+      - manywheel-py3_9-cuda12_8-build
       - get-label-type
     uses: ./.github/workflows/_binary-test-linux.yml
     with:
@@ -79,8 +127,56 @@ jobs:
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: cuda12.8
-      DESIRED_PYTHON: "3.12"
-      build_name: manywheel-py3_12-cuda12_8
+      use_split_build: False
+      DESIRED_PYTHON: "3.9"
+      build_name: manywheel-py3_9-cuda12_8
+      build_environment: linux-binary-manywheel
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 and 12.9 build need sm_70+ runner
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+
+  manywheel-py3_9-cuda12_9-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: 12.9
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
+      DESIRED_PYTHON: "3.9"
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      build_name: manywheel-py3_9-cuda12_9
+      build_environment: linux-binary-manywheel
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.9; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  manywheel-py3_9-cuda12_9-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - manywheel-py3_9-cuda12_9-build
+      - get-label-type
+    uses: ./.github/workflows/_binary-test-linux.yml
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: manywheel
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: 12.9
+      GPU_ARCH_TYPE: cuda
+      DOCKER_IMAGE: manylinux2_28-builder
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
+      use_split_build: False
+      DESIRED_PYTHON: "3.9"
+      build_name: manywheel-py3_9-cuda12_9
       build_environment: linux-binary-manywheel
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.g4dn.4xlarge.nvidia.gpu  # 12.8 and 12.9 build need sm_70+ runner

.github/workflows/generated-linux-binary-manywheel-rocm-main.yml

@@ -58,6 +58,7 @@ jobs:
       GPU_ARCH_TYPE: rocm
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_9-rocm6_4
@@ -82,6 +83,7 @@ jobs:
       SKIP_ALL_TESTS: 1
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
     steps:
       - name: Setup ROCm

.github/workflows/generated-linux-s390x-binary-manywheel-nightly.yml

@@ -60,6 +60,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       runs_on: linux.s390x
       ALPINE_IMAGE: "docker.io/s390x/alpine"
@@ -83,6 +84,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       build_name: manywheel-py3_9-cpu-s390x
       build_environment: linux-s390x-binary-manywheel
@@ -105,6 +107,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.9"
       build_name: manywheel-py3_9-cpu-s390x
     secrets:
@@ -124,6 +127,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       runs_on: linux.s390x
       ALPINE_IMAGE: "docker.io/s390x/alpine"
@@ -147,6 +151,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       build_name: manywheel-py3_10-cpu-s390x
       build_environment: linux-s390x-binary-manywheel
@@ -169,6 +174,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.10"
       build_name: manywheel-py3_10-cpu-s390x
     secrets:
@@ -188,6 +194,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       runs_on: linux.s390x
       ALPINE_IMAGE: "docker.io/s390x/alpine"
@@ -211,6 +218,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       build_name: manywheel-py3_11-cpu-s390x
       build_environment: linux-s390x-binary-manywheel
@@ -233,6 +241,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.11"
       build_name: manywheel-py3_11-cpu-s390x
     secrets:
@@ -252,6 +261,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       runs_on: linux.s390x
       ALPINE_IMAGE: "docker.io/s390x/alpine"
@@ -275,6 +285,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       build_name: manywheel-py3_12-cpu-s390x
       build_environment: linux-s390x-binary-manywheel
@@ -297,6 +308,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.12"
       build_name: manywheel-py3_12-cpu-s390x
     secrets:
@@ -316,6 +328,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       runs_on: linux.s390x
       ALPINE_IMAGE: "docker.io/s390x/alpine"
@@ -339,6 +352,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       build_name: manywheel-py3_13-cpu-s390x
       build_environment: linux-s390x-binary-manywheel
@@ -361,6 +375,7 @@ jobs:
       GPU_ARCH_TYPE: cpu-s390x
       DOCKER_IMAGE: pytorch/manylinuxs390x-builder
       DOCKER_IMAGE_TAG_PREFIX: cpu-s390x
+      use_split_build: False
       DESIRED_PYTHON: "3.13"
       build_name: manywheel-py3_13-cpu-s390x
     secrets:

.github/workflows/inductor-perf-test-b200.yml

@@ -1,154 +0,0 @@
-name: inductor-perf-b200
-
-on:
-  schedule:
-    - cron: 0 7 * * 1-6
-    - cron: 0 7 * * 0
-  # NB: GitHub has an upper limit of 10 inputs here, so before we can sort it
-  # out, let try to run torchao cudagraphs_low_precision as part of cudagraphs
-  workflow_dispatch:
-    inputs:
-      training:
-        description: Run training (on by default)?
-        required: false
-        type: boolean
-        default: true
-      inference:
-        description: Run inference (on by default)?
-        required: false
-        type: boolean
-        default: true
-      default:
-        description: Run inductor_default?
-        required: false
-        type: boolean
-        default: false
-      dynamic:
-        description: Run inductor_dynamic_shapes?
-        required: false
-        type: boolean
-        default: false
-      cppwrapper:
-        description: Run inductor_cpp_wrapper?
-        required: false
-        type: boolean
-        default: false
-      cudagraphs:
-        description: Run inductor_cudagraphs?
-        required: false
-        type: boolean
-        default: true
-      freezing_cudagraphs:
-        description: Run inductor_cudagraphs with freezing for inference?
-        required: false
-        type: boolean
-        default: false
-      aotinductor:
-        description: Run aot_inductor for inference?
-        required: false
-        type: boolean
-        default: false
-      maxautotune:
-        description: Run inductor_max_autotune?
-        required: false
-        type: boolean
-        default: false
-      benchmark_configs:
-        description: The list of configs used the benchmark
-        required: false
-        type: string
-        default: inductor_huggingface_perf_cuda_b200,inductor_timm_perf_cuda_b200,inductor_torchbench_perf_cuda_b200
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
-  cancel-in-progress: true
-
-permissions:
-  id-token: write
-  contents: read
-
-jobs:
-  get-label-type:
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-      opt_out_experiments: lf
-
-  build:
-    name: cuda12.8-py3.10-gcc9-sm100
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      # Use a bigger runner here because CUDA_ARCH 9.0 is only built for H100
-      # or newer GPUs, so it doesn't benefit much from existing compiler cache
-      # from trunk. Also use a memory-intensive runner here because memory is
-      # usually the bottleneck
-      runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
-      cuda-arch-list: '10.0'
-      test-matrix: |
-        { include: [
-          { config: "inductor_huggingface_perf_cuda_b200", shard: 1, num_shards: 1, runner: "linux.dgx.b200" },
-          { config: "inductor_timm_perf_cuda_b200", shard: 1, num_shards: 1, runner: "linux.dgx.b200" },
-          { config: "inductor_torchbench_perf_cuda_b200", shard: 1, num_shards: 1, runner: "linux.dgx.b200" },
-        ]}
-      selected-test-configs: ${{ inputs.benchmark_configs }}
-      build-additional-packages: "vision audio fbgemm torchao"
-    secrets: inherit
-
-  test-periodically:
-    name: cuda12.8-py3.10-gcc9-sm100
-    uses: ./.github/workflows/_linux-test.yml
-    needs: build
-    if: github.event.schedule == '0 7 * * 1-6'
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
-      dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true
-      docker-image: ${{ needs.build.outputs.docker-image }}
-      test-matrix: ${{ needs.build.outputs.test-matrix }}
-      aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-      timeout-minutes: 720
-      disable-monitor: false
-      monitor-log-interval: 15
-      monitor-data-collect-interval: 4
-    secrets: inherit
-
-  test-weekly:
-    name: cuda12.8-py3.10-gcc9-sm100
-    uses: ./.github/workflows/_linux-test.yml
-    needs: build
-    if: github.event.schedule == '0 7 * * 0'
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
-      dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-maxautotune-true-freeze_autotune_cudagraphs-true-cudagraphs_low_precision-true
-      docker-image: ${{ needs.build.outputs.docker-image }}
-      test-matrix: ${{ needs.build.outputs.test-matrix }}
-      timeout-minutes: 1440
-      aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-      disable-monitor: false
-      monitor-log-interval: 15
-      monitor-data-collect-interval: 4
-    secrets: inherit
-
-  test:
-    name: cuda12.8-py3.10-gcc9-sm100
-    uses: ./.github/workflows/_linux-test.yml
-    needs: build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm100
-      dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cudagraphs-${{ inputs.cudagraphs }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}-maxautotune-${{ inputs.maxautotune }}-freezing_cudagraphs-${{ inputs.freezing_cudagraphs }}-cudagraphs_low_precision-${{ inputs.cudagraphs }}
-      docker-image: ${{ needs.build.outputs.docker-image }}
-      test-matrix: ${{ needs.build.outputs.test-matrix }}
-      aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-      timeout-minutes: 720
-      disable-monitor: false
-      monitor-log-interval: 15
-      monitor-data-collect-interval: 4
-    secrets: inherit

.github/workflows/inductor-perf-test-nightly-h100.yml

@@ -2,7 +2,7 @@ name: inductor-perf-nightly-h100
 
 on:
   schedule:
-    - cron: 15 0,12 * * 1-6
+    - cron: 15 0,4,8,12,16,20 * * 1-6
     - cron: 0 7 * * 0
   # NB: GitHub has an upper limit of 10 inputs here, so before we can sort it
   # out, let try to run torchao cudagraphs_low_precision as part of cudagraphs
@@ -126,7 +126,7 @@ jobs:
     name: cuda12.8-py3.10-gcc9-sm90
     uses: ./.github/workflows/_linux-test.yml
     needs: build
-    if: github.event.schedule == '15 0,12 * * 1-6'
+    if: github.event.schedule == '15 0,4,8,12,16,20 * * 1-6'
     with:
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm90
       dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true

.github/workflows/inductor-perf-test-nightly-rocm.yml

@@ -85,26 +85,26 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     with:
       build-environment: linux-jammy-rocm-py3_10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
+      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
       test-matrix: |
         { include: [
-          { config: "inductor_huggingface_perf_rocm", shard: 1, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_huggingface_perf_rocm", shard: 2, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_huggingface_perf_rocm", shard: 3, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_huggingface_perf_rocm", shard: 4, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_timm_perf_rocm", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 1, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 2, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 3, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 4, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 5, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 6, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 7, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor_torchbench_perf_rocm", shard: 8, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor_huggingface_perf_rocm", shard: 1, num_shards: 4, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_huggingface_perf_rocm", shard: 2, num_shards: 4, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_huggingface_perf_rocm", shard: 3, num_shards: 4, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_huggingface_perf_rocm", shard: 4, num_shards: 4, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_timm_perf_rocm", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_timm_perf_rocm", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_timm_perf_rocm", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_timm_perf_rocm", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_timm_perf_rocm", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 1, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 2, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 3, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 4, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 5, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 6, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 7, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor_torchbench_perf_rocm", shard: 8, num_shards: 8, runner: "linux.rocm.gpu.mi300.2" },
         ]}
     secrets: inherit
 

.github/workflows/inductor-periodic.yml

@@ -77,25 +77,25 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     with:
       build-environment: linux-jammy-rocm-py3_10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
+      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
       sync-tag: rocm-build
       test-matrix: |
         { include: [
-          { config: "dynamo_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamo_eager_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamo_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamo_eager_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamo_eager_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "aot_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "aot_eager_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "aot_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "aot_eager_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "aot_eager_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamic_aot_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamic_aot_eager_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamic_aot_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamic_aot_eager_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "dynamic_aot_eager_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "dynamo_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamo_eager_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamo_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamo_eager_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamo_eager_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "aot_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "aot_eager_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "aot_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "aot_eager_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "aot_eager_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamic_aot_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamic_aot_eager_torchbench", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamic_aot_eager_huggingface", shard: 1, num_shards: 1, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamic_aot_eager_timm", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "dynamic_aot_eager_timm", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
         ]}
     secrets: inherit
 

.github/workflows/inductor-rocm-mi300.yml

@@ -47,8 +47,8 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
       test-matrix: |
         { include: [
-          { config: "inductor", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "inductor", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
+          { config: "inductor", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
+          { config: "inductor", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.mi300.2" },
         ]}
     secrets: inherit
 

.github/workflows/mac-mps.yml

@@ -28,6 +28,7 @@ jobs:
       # than our AWS macos-m1-14 runners
       test-matrix: |
         { include: [
+          { config: "test_mps", shard: 1, num_shards: 1, runner: "macos-m1-13" },
           { config: "test_mps", shard: 1, num_shards: 1, runner: "macos-m1-14" },
           { config: "test_mps", shard: 1, num_shards: 1, runner: "macos-m2-15" },
         ]}

.github/workflows/nightly.yml

@@ -75,11 +75,10 @@ jobs:
             repo-owner: pytorch
             branch: main
             pin-folder: .github/ci_commit_pins
-          # executorch jobs are disabled since it needs some manual work for the hash update
-          # - repo-name: executorch
-          #   repo-owner: pytorch
-          #   branch: main
-          #   pin-folder: .ci/docker/ci_commit_pins
+          - repo-name: executorch
+            repo-owner: pytorch
+            branch: main
+            pin-folder: .ci/docker/ci_commit_pins
           - repo-name: triton
             repo-owner: triton-lang
             branch: main