[XPU] Add _make_xccl_premul_sum binding for XCCL backend

.bazelrc

@@ -17,9 +17,11 @@ build:no-tty --show_progress_rate_limit 10
 build --define=cuda=true
 # rules_cuda configuration
 build --@rules_cuda//cuda:enable_cuda
-build --@rules_cuda//cuda:cuda_targets=sm_52
+build --@rules_cuda//cuda:cuda_targets=sm_75
 build --@rules_cuda//cuda:compiler=nvcc
 build --repo_env=CUDA_PATH=/usr/local/cuda
+common --repo_env=PIP_INDEX_URL=https://pypi.org/simple/
+common --repo_env=PIP_TRUSTED_HOST=pypi.org
 
 # Configuration to build without GPU support
 build:cpu-only --define=cuda=false

.ci/docker/almalinux/Dockerfile

@@ -19,15 +19,12 @@ RUN git config --global --add safe.directory '*'
 ENV PATH=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/bin:$PATH
 
 # cmake-3.18.4 from pip
+# NS: Apr 1 2026 3.18.4 is gone, reported here https://github.com/scikit-build/cmake-python-distributions/issues/693
 RUN yum install -y python3-pip && \
-    python3 -mpip install cmake==3.18.4 && \
+    python3 -mpip install cmake==3.18.4.post1 && \
     ln -s /usr/local/bin/cmake /usr/bin/cmake3
 RUN rm -rf /usr/local/cuda-*
 
-FROM base as openssl
-ADD ./common/install_openssl.sh install_openssl.sh
-RUN bash ./install_openssl.sh && rm install_openssl.sh
-
 FROM base as patchelf
 # Install patchelf
 ADD ./common/install_patchelf.sh install_patchelf.sh
@@ -84,6 +81,8 @@ RUN yum -y update && \
     yum -y install glibc-langpack-en && \
     yum install -y sudo wget curl perl util-linux xz bzip2 git patch which perl zlib-devel openssl-devel yum-utils autoconf automake make gcc-toolset-${DEVTOOLSET_VERSION}-gcc gcc-toolset-${DEVTOOLSET_VERSION}-gcc-c++ gcc-toolset-${DEVTOOLSET_VERSION}-gcc-gfortran gcc-toolset-${DEVTOOLSET_VERSION}-gdb
 RUN git config --global --add safe.directory '*'
+# All rocm clang cfg files load the same rocm.cfg, make sure it points to the right toolchain.
+RUN echo "--gcc-toolchain=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr" >> /opt/rocm/llvm/bin/rocm.cfg
 ENV PATH=/opt/rh/gcc-toolset-${DEVTOOLSET_VERSION}/root/usr/bin:$PATH
 
 FROM rocm_base as rocm
@@ -109,7 +108,6 @@ COPY --from=cuda13.2  /usr/local/cuda-13.2 /usr/local/cuda-13.2
 # Final step
 FROM ${BASE_TARGET} as final
 ARG DEVTOOLSET_VERSION=13
-COPY --from=openssl            /opt/openssl           /opt/openssl
 COPY --from=patchelf           /patchelf              /usr/local/bin/patchelf
 COPY --from=conda              /opt/conda             /opt/conda
 

.ci/docker/build.sh

@@ -95,15 +95,13 @@ case "$tag" in
     CUDA_VERSION=12.4
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
-    VISION=yes
     KATEX=yes
     TRITON=yes
     ;;
   pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11)
     CUDA_VERSION=12.8.1
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
-    VISION=yes
     KATEX=yes
     TRITON=yes
     INSTALL_MINGW=yes
@@ -112,15 +110,13 @@ case "$tag" in
     CUDA_VERSION=13.0.2
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
-    VISION=yes
     KATEX=yes
     TRITON=yes
     ;;
   pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11-inductor-benchmarks)
     CUDA_VERSION=13.0.2
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
-    VISION=yes
     KATEX=yes
     TRITON=yes
     INDUCTOR_BENCHMARKS=yes
@@ -129,35 +125,32 @@ case "$tag" in
     CUDA_VERSION=13.0.2
     ANACONDA_PYTHON_VERSION=3.12
     GCC_VERSION=11
-    VISION=yes
     KATEX=yes
     TRITON=yes
     ;;
-  pytorch-linux-jammy-py3-clang15-onnx)
+  pytorch-linux-jammy-py3.10-clang18)
     ANACONDA_PYTHON_VERSION=3.10
-    CLANG_VERSION=15
-    VISION=yes
+    CLANG_VERSION=18
+    GCC_VERSION=11
+    KATEX=yes
+    DOCS=yes
     ONNX=yes
     ;;
-  pytorch-linux-jammy-py3.10-clang15)
-    ANACONDA_PYTHON_VERSION=3.10
-    CLANG_VERSION=15
-    ;;
-  pytorch-linux-jammy-py3.11-clang15)
+  pytorch-linux-jammy-py3.11-clang18)
     ANACONDA_PYTHON_VERSION=3.11
-    CLANG_VERSION=15
+    CLANG_VERSION=18
     ;;
-  pytorch-linux-jammy-py3.12-clang15)
+  pytorch-linux-jammy-py3.12-clang18)
     ANACONDA_PYTHON_VERSION=3.12
-    CLANG_VERSION=15
+    CLANG_VERSION=18
     ;;
-  pytorch-linux-jammy-py3.13-clang15)
+  pytorch-linux-jammy-py3.13-clang18)
     ANACONDA_PYTHON_VERSION=3.13
-    CLANG_VERSION=15
+    CLANG_VERSION=18
     ;;
-  pytorch-linux-jammy-py3.14-clang15)
+  pytorch-linux-jammy-py3.14-clang18)
     ANACONDA_PYTHON_VERSION=3.14
-    CLANG_VERSION=15
+    CLANG_VERSION=18
     ;;
   pytorch-linux-jammy-rocm-n-py3 | pytorch-linux-jammy-rocm-n-py3-benchmarks | pytorch-linux-noble-rocm-n-py3)
     if [[ $tag =~ "jammy" ]]; then
@@ -166,9 +159,7 @@ case "$tag" in
       ANACONDA_PYTHON_VERSION=3.12
     fi
     GCC_VERSION=13
-    VISION=yes
     ROCM_VERSION=7.2
-    NINJA_VERSION=1.9.0
     TRITON=yes
     KATEX=yes
     PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950;gfx1100"
@@ -179,33 +170,27 @@ case "$tag" in
   pytorch-linux-noble-rocm-nightly-py3)
     ANACONDA_PYTHON_VERSION=3.12
     GCC_VERSION=13
-    VISION=yes
     ROCM_VERSION=nightly
-    NINJA_VERSION=1.9.0
     TRITON=yes
     KATEX=yes
     PYTORCH_ROCM_ARCH="gfx942"
     ;;
   pytorch-linux-jammy-xpu-n-1-py3)
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
-    VISION=yes
     XPU_VERSION=2025.2
     XPU_DRIVER_TYPE=LTS
-    NINJA_VERSION=1.9.0
     TRITON=yes
     ;;
   pytorch-linux-noble-xpu-n-py3 | pytorch-linux-noble-xpu-n-py3-client | pytorch-linux-noble-xpu-n-py3-inductor-benchmarks)
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=13
-    VISION=yes
     XPU_VERSION=2025.3
     if [[ $tag =~ "client" ]]; then
       XPU_DRIVER_TYPE=CLIENT
     else
       XPU_DRIVER_TYPE=LTS
     fi
-    NINJA_VERSION=1.9.0
     TRITON=yes
     if [[ $tag =~ "benchmarks" ]]; then
       INDUCTOR_BENCHMARKS=yes
@@ -214,36 +199,19 @@ case "$tag" in
   pytorch-linux-jammy-py3-gcc11-inductor-benchmarks)
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
-    VISION=yes
     KATEX=yes
-    TRITON=yes
     DOCS=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-clang15)
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-clang18)
     ANACONDA_PYTHON_VERSION=3.10
     CUDA_VERSION=12.8.1
-    CLANG_VERSION=15
-    VISION=yes
-    TRITON=yes
-    ;;
-  pytorch-linux-jammy-py3-clang18-asan)
-    ANACONDA_PYTHON_VERSION=3.10
     CLANG_VERSION=18
-    VISION=yes
-    ;;
-  pytorch-linux-jammy-py3.10-gcc11)
-    ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=11
-    VISION=yes
-    KATEX=yes
     TRITON=yes
-    DOCS=yes
-    UNINSTALL_DILL=yes
     ;;
-  pytorch-linux-jammy-py3-clang15-executorch)
+  pytorch-linux-jammy-py3-clang18-executorch)
     ANACONDA_PYTHON_VERSION=3.10
-    CLANG_VERSION=15
+    CLANG_VERSION=18
     EXECUTORCH=yes
     ;;
   pytorch-linux-jammy-py3.12-halide)
@@ -288,29 +256,20 @@ case "$tag" in
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=13
     ACL=yes
-    VISION=yes
     OPENBLAS=yes
-    # snadampal: skipping llvm src build install because the current version
-    # from pytorch/llvm:9.0.1 is x86 specific
-    SKIP_LLVM_SRC_BUILD_INSTALL=yes
     ;;
   pytorch-linux-jammy-aarch64-py3.10-gcc13-inductor-benchmarks)
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=13
     ACL=yes
-    VISION=yes
     OPENBLAS=yes
-    # snadampal: skipping llvm src build install because the current version
-    # from pytorch/llvm:9.0.1 is x86 specific
-    SKIP_LLVM_SRC_BUILD_INSTALL=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
   pytorch-linux-noble-riscv64-py3.12-gcc14)
     GCC_VERSION=14
     ;;
   *)
     # Catch-all for builds that are not hardcoded.
-    VISION=yes
     echo "image '$image' did not match an existing build configuration"
     if [[ "$image" == *py* ]]; then
       extract_version_from_image_name py ANACONDA_PYTHON_VERSION
@@ -327,14 +286,10 @@ case "$tag" in
       if [[ -z "$ROCM_VERSION" ]]; then
         extract_version_from_image_name rocm ROCM_VERSION
       fi
-      NINJA_VERSION=1.9.0
       TRITON=yes
       # To ensure that any ROCm config will build using conda cmake
       # and thus have LAPACK/MKL enabled
       fi
-    if [[ "$image" == *centos7* ]]; then
-      NINJA_VERSION=1.10.2
-    fi
     if [[ "$image" == *gcc* ]]; then
       extract_version_from_image_name gcc GCC_VERSION
     fi
@@ -366,7 +321,6 @@ docker buildx build \
        ${progress_flag} \
        --build-arg "BUILD_ENVIRONMENT=${image}" \
        --build-arg "LLVMDEV=${LLVMDEV:-}" \
-       --build-arg "VISION=${VISION:-}" \
        --build-arg "UBUNTU_VERSION=${UBUNTU_VERSION}" \
        --build-arg "DEVTOOLSET_VERSION=${DEVTOOLSET_VERSION}" \
        --build-arg "GLIBC_VERSION=${GLIBC_VERSION}" \
@@ -376,7 +330,6 @@ docker buildx build \
        --build-arg "PYTHON_VERSION=${PYTHON_VERSION}" \
        --build-arg "GCC_VERSION=${GCC_VERSION}" \
        --build-arg "CUDA_VERSION=${CUDA_VERSION}" \
-       --build-arg "NINJA_VERSION=${NINJA_VERSION:-}" \
        --build-arg "KATEX=${KATEX:-}" \
        --build-arg "ROCM_VERSION=${ROCM_VERSION:-}" \
        --build-arg "PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}" \
@@ -392,11 +345,9 @@ docker buildx build \
        --build-arg "TPU=${TPU}" \
        --build-arg "XPU_VERSION=${XPU_VERSION}" \
        --build-arg "XPU_DRIVER_TYPE=${XPU_DRIVER_TYPE}" \
-       --build-arg "UNINSTALL_DILL=${UNINSTALL_DILL}" \
        --build-arg "ACL=${ACL:-}" \
        --build-arg "OPENBLAS=${OPENBLAS:-}" \
        --build-arg "SKIP_SCCACHE_INSTALL=${SKIP_SCCACHE_INSTALL:-}" \
-       --build-arg "SKIP_LLVM_SRC_BUILD_INSTALL=${SKIP_LLVM_SRC_BUILD_INSTALL:-}" \
        --build-arg "INSTALL_MINGW=${INSTALL_MINGW:-}" \
        -f $(dirname ${DOCKERFILE})/Dockerfile \
        --load \

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

@@ -1,2 +1,2 @@
-transformers==5.2.0
+transformers==5.5.3
 soxr==0.5.0

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

@@ -1 +1 @@
-33f782efa9464adebb448ea1f1df1a64ec37ceb0
+21033c4e2be9b42c9e6ce7a39a70ead2aba279b4

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-9844da955a9db14ec69c9aac828ee9803085e288
+b4e20bbe55617cc798b986c2555a2bc7b303c737

.ci/docker/common/install_amdsmi.sh

@@ -7,7 +7,7 @@ source /etc/rocm_env.sh
 # For theRock nightly, amd_smi may already be installed or in a different location
 if [ -d "${ROCM_PATH}/share/amd_smi" ]; then
   echo "Installing amdsmi from: ${ROCM_PATH}/share/amd_smi"
-  cd ${ROCM_PATH}/share/amd_smi && pip install .
+  cd ${ROCM_PATH}/share/amd_smi && python3 -m pip install .
 else
   echo "AMD SMI not found at ${ROCM_PATH}/share/amd_smi - skipping (may already be installed via pip)"
 fi