Patch for multi device CUDA (#3100 )

Fix qmv_impl for small N (#3096 )
JACCL update (#3094 )
2026-02-05 17:33:51 -08:00 · 2026-02-05 17:33:36 -08:00 · 2026-02-05 15:16:07 -08:00 · 2026-02-05 08:51:29 -08:00 · 2026-02-05 07:04:02 -08:00 · 2026-02-05 10:58:49 +09:00
427 changed files with 31326 additions and 7868 deletions
@@ -1,579 +0,0 @@
-version: 2.1
-
-orbs:
-  apple: ml-explore/pr-approval@0.1.0
-
-parameters:
-  nightly_build:
-    type: boolean
-    default: false
-  test_release:
-    type: boolean
-    default: false
-
-jobs:
-  build_documentation:
-    parameters:
-      upload-docs:
-        type: boolean
-        default: false
-    macos:
-      xcode: "26.0.0"
-    resource_class: m4pro.medium
-    steps:
-      - checkout
-      - run:
-          name: Install
-          command: |
-            xcodebuild -downloadComponent MetalToolchain
-            brew install python@3.10
-            brew install doxygen
-            python3.10 -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            pip install --upgrade cmake
-            pip install -r docs/requirements.txt
-            pip install . -v
-      - when:
-          condition:
-            not: << parameters.upload-docs >>
-          steps:
-            - run:
-               name: Build documentation
-               command: |
-                 source env/bin/activate
-                 cd docs && doxygen && make html O=-W
-      - when:
-          condition: << parameters.upload-docs >>
-          steps:
-            - add_ssh_keys:
-                fingerprints:
-                  - "SHA256:OhcVVMovbT0pkgMeiVRyxMnjV9R2t+hKBsNcuxq9h+0"
-            - run:
-               name: Upload documentation
-               command: |
-                 source env/bin/activate
-                 git config user.email "mlx@group.apple.com"
-                 git config user.name "CircleCI Docs"
-                 git checkout gh-pages
-                 git rebase main
-                 cd docs
-                 git rm -rf build/html
-                 doxygen && make html O=-W
-                 git add -f build/html
-                 git commit -m "rebase"
-                 git push -f origin gh-pages
-
-  linux_build_and_test:
-    machine:
-      image: ubuntu-2204:current
-      resource_class: large
-    steps:
-      - checkout
-      - run:
-          name: Run style checks
-          command: |
-            pip install pre-commit
-            pre-commit run --all
-            if ! git diff --quiet; then echo 'Style checks failed, please install pre-commit and run pre-commit run --all and push the change'; exit 1; fi
-      - run:
-          name: Install dependencies
-          command: |
-            export DEBIAN_FRONTEND=noninteractive
-            export NEEDRESTART_MODE=a
-            sudo apt-get update
-            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
-            curl -LsSf https://astral.sh/uv/install.sh | sh
-      - run:
-          name: Install Python package
-          command: |
-            uv venv
-            uv pip install cmake
-            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              uv pip install -e ".[dev]" -v
-      - run:
-          name: Generate package stubs
-          command: |
-            uv pip install typing_extensions
-            uv run --no-project setup.py generate_stubs
-      - run:
-          name: Run Python tests
-          command: |
-            source .venv/bin/activate
-            python -m unittest discover python/tests -v
-            mpirun --bind-to none -host localhost:8 -np 8 python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
-            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
-      - run:
-          name: Build CPP only
-          command: |
-            source .venv/bin/activate
-            mkdir -p build && cd build
-            cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
-            make -j `nproc`
-      - run:
-          name: Run CPP tests
-          command: ./build/tests/tests
-
-  mac_build_and_test:
-    parameters:
-      xcode_version:
-        type: string
-        default: "26.0.0"
-      macosx_deployment_target:
-        type: string
-        default: ""
-    macos:
-      xcode: << parameters.xcode_version >>
-    environment:
-      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    resource_class: m4pro.medium
-    steps:
-      - checkout
-      - run:
-          name: Install dependencies
-          command: |
-            xcodebuild -downloadComponent MetalToolchain
-            HOMEBREW_NO_AUTO_UPDATE=1 HOMEBREW_NO_INSTALL_CLEANUP=1 \
-              brew install openmpi uv
-      - run:
-          name: Install Python package
-          command: |
-            uv venv --python 3.10
-            uv pip install \
-              nanobind==2.4.0 \
-              cmake \
-              numpy \
-              torch \
-              tensorflow \
-              unittest-xml-reporting
-            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              uv pip install -e . -v
-      - run:
-          name: Generate package stubs
-          command: |
-            uv pip install typing_extensions
-            uv run --no-project setup.py generate_stubs
-      - run:
-          name: Run Python tests
-          command: |
-            source .venv/bin/activate
-            LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
-            LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
-            mpirun --bind-to none -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
-            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
-      - run:
-          name: Build example extension
-          command: |
-            source .venv/bin/activate
-            cd examples/extensions
-            uv pip install -r requirements.txt
-            uv run --no-project setup.py build_ext --inplace
-            uv run --no-project python test.py
-      - store_test_results:
-          path: test-results
-      - run:
-          name: Build CPP only
-          command: |
-            source .venv/bin/activate
-            mkdir -p build && cd build && cmake .. && make -j `sysctl -n hw.ncpu`
-      - run:
-          name: Run CPP tests
-          command: |
-            DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 ./build/tests/tests
-      - run:
-          name: Build small binary
-          command: |
-            source .venv/bin/activate
-            cd build/
-            cmake .. -DCMAKE_BUILD_TYPE=MinSizeRel \
-              -DBUILD_SHARED_LIBS=ON \
-              -DMLX_BUILD_CPU=OFF \
-              -DMLX_BUILD_SAFETENSORS=OFF \
-              -DMLX_BUILD_GGUF=OFF \
-              -DMLX_METAL_JIT=ON
-            make -j `sysctl -n hw.ncpu`
-      - run:
-          name: Run Python tests with JIT
-          command: |
-            CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-              uv pip install -e . -v
-            LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
-              METAL_DEBUG_ERROR_MODE=0 \
-              uv run --no-project python -m xmlrunner discover \
-                -v python/tests \
-                -o test-results/gpu_jit
-
-  cuda_build_and_test:
-    parameters:
-      image_date:
-        type: string
-        default: "2023.11.1"
-    machine:
-      image: "linux-cuda-12:<< parameters.image_date >>"
-      resource_class: gpu.nvidia.small.gen2
-    steps:
-      - checkout
-      - restore_cache:
-          keys:
-            - cuda-<< parameters.image_date >>-{{ arch }}-
-      - run:
-          name: Install dependencies
-          command: |
-            sudo apt-get update
-            sudo apt-get install libcudnn9-dev-cuda-12
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install libnccl2 libnccl-dev
-            curl -sL https://github.com/ccache/ccache/releases/download/v4.11.3/ccache-4.11.3-linux-x86_64.tar.xz | tar xJf -
-            sudo mv ccache-4.11.3-linux-x86_64/ccache /usr/bin/ccache
-            rm -rf ccache-4.11.3-linux-x86_64
-            curl -LsSf https://astral.sh/uv/install.sh | sh
-      - run:
-          name: Set CCache size
-          command: ccache --max-size 1G
-      - run:
-          name: Install Python package
-          command: |
-            uv venv
-            uv pip install cmake
-            DEBUG=1 CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_COMPILE_WARNING_AS_ERROR=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-              uv pip install -e ".[dev]" -v
-      - run:
-          name: Run Python tests
-          command: |
-            source .venv/bin/activate
-            LOW_MEMORY=1 DEVICE=cpu python -m unittest discover python/tests -v
-            LOW_MEMORY=1 DEVICE=gpu python -m tests discover python/tests -v
-      - run:
-          name: Build CPP only
-          command: |
-            source .venv/bin/activate
-            cmake . -B build \
-              -DMLX_BUILD_CUDA=ON \
-              -DCMAKE_CUDA_COMPILER=`which nvcc` \
-              -DCMAKE_BUILD_TYPE=DEBUG
-            cmake --build build -j `nproc`
-      - run:
-          name: Run CPP tests
-          command: ./build/tests/tests -sfe="*fft_tests.cpp,*linalg_tests.cpp"
-      - run:
-          name: CCache report
-          command: |
-            ccache --show-stats
-            ccache --zero-stats
-            ccache --cleanup
-      - save_cache:
-          key: cuda-<< parameters.image_date >>-{{ arch }}-{{ epoch }}
-          paths:
-            - /home/circleci/.cache/ccache
-
-  build_release:
-    parameters:
-      python_version:
-        type: string
-        default: "3.10"
-      xcode_version:
-        type: string
-        default: "26.0.0"
-      build_env:
-        type: string
-        default: ""
-      macosx_deployment_target:
-        type: string
-        default: ""
-    macos:
-      xcode: << parameters.xcode_version >>
-    resource_class: m4pro.medium
-    environment:
-      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    steps:
-      - checkout
-      - run:
-          name: Install dependencies
-          command: |
-            xcodebuild -downloadComponent MetalToolchain
-            mkdir -p ~/miniconda3
-            curl https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-arm64.sh -o ~/miniconda3/miniconda.sh
-            bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
-            rm ~/miniconda3/miniconda.sh
-            source ~/miniconda3/bin/activate
-            conda init --all
-            conda create -n env python=<< parameters.python_version >> -y
-            conda activate env
-            pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install --upgrade setuptools
-            pip install numpy
-            pip install twine
-            pip install build
-      - run:
-          name: Install Python package
-          command: |
-            conda activate env
-            env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
-              pip install . -v
-      - run:
-          name: Generate package stubs
-          command: |
-            conda activate env
-            pip install typing_extensions
-            python setup.py generate_stubs
-      - run:
-          name: Build Python package
-          command: |
-            conda activate env
-            python setup.py clean --all
-            << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
-      - when:
-          condition:
-            equal: ["3.10", << parameters.python_version >>]
-          steps:
-            - run:
-                name: Build common package
-                command: |
-                  conda activate env
-                  python setup.py clean --all
-                  << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
-      - when:
-          condition: << parameters.build_env >>
-          steps:
-            - run:
-                name: Upload package
-                command: |
-                  conda activate env
-                  twine upload dist/*
-      - store_artifacts:
-          path: dist/
-
-  build_linux_release:
-    parameters:
-      python_version:
-        type: string
-        default: "3.10"
-      build_env:
-        type: string
-        default: ""
-    machine:
-      image: ubuntu-2204:current
-      resource_class: large
-    steps:
-      - checkout
-      - run:
-          name: Build wheel
-          command: |
-            PYTHON=python<< parameters.python_version >>
-            export DEBIAN_FRONTEND=noninteractive
-            export NEEDRESTART_MODE=a
-            sudo apt-get update
-            TZ=Etc/UTC sudo apt-get -y install tzdata
-            sudo add-apt-repository -y ppa:deadsnakes/ppa
-            sudo apt-get install -y $PYTHON $PYTHON-dev $PYTHON-full
-            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
-            $PYTHON -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            pip install --upgrade cmake
-            pip install auditwheel
-            pip install patchelf
-            pip install build
-            pip install twine
-            << parameters.build_env >> pip install ".[dev]" -v
-            pip install typing_extensions
-            python setup.py generate_stubs
-            python setup.py clean --all
-            MLX_BUILD_STAGE=1 << parameters.build_env >> python -m build -w
-            bash python/scripts/repair_linux.sh
-      - when:
-          condition:
-            equal: ["3.10", << parameters.python_version >>]
-          steps:
-            - run:
-                name: Build common package
-                command: |
-                  source env/bin/activate
-                  python setup.py clean --all
-                  << parameters.build_env >> MLX_BUILD_STAGE=2 \
-                    python -m build -w
-                  auditwheel repair dist/mlx_cpu*.whl --plat manylinux_2_35_x86_64
-      - when:
-          condition: << parameters.build_env >>
-          steps:
-            - run:
-                name: Upload packages
-                command: |
-                  source env/bin/activate
-                  twine upload wheelhouse/*.whl
-      - store_artifacts:
-          path: wheelhouse/
-
-  build_cuda_release:
-    parameters:
-      build_env:
-        type: string
-        default: ""
-    machine:
-      image: ubuntu-2204:current
-      resource_class: xlarge
-    steps:
-      - checkout
-      - run:
-          name: Build wheel
-          command: |
-            export DEBIAN_FRONTEND=noninteractive
-            export NEEDRESTART_MODE=a
-            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
-            sudo dpkg -i cuda-keyring_1.1-1_all.deb
-            sudo apt-get update
-            sudo apt-get install cuda-toolkit-12-9 libcudnn9-dev-cuda-12
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install zip
-            pip install auditwheel
-            pip install patchelf
-            pip install build
-            pip install twine
-            export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}
-            export LD_LIBRARY_PATH=/usr/local/cuda/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-            << parameters.build_env >> MLX_BUILD_STAGE=2 \
-              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-              python -m build -w
-            bash python/scripts/repair_cuda.sh
-      - when:
-          condition: << parameters.build_env >>
-          steps:
-            - run:
-                name: Upload package
-                command: |
-                  twine upload wheelhouse/*.whl
-      - store_artifacts:
-          path: wheelhouse/
-
-workflows:
-  build_and_test:
-    when:
-      and:
-        - matches:
-            pattern: "^(?!pull/)[-\\w]+$"
-            value: << pipeline.git.branch >>
-        - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.test_release >>
-    jobs:
-      - mac_build_and_test:
-          matrix:
-            parameters:
-              macosx_deployment_target: ["13.5", "15.0"]
-      - linux_build_and_test
-      - cuda_build_and_test:
-          matrix:
-            parameters:
-              image_date: ["2023.11.1", "2025.05.1"]
-      - build_documentation 
-
-  build_pypi_release:
-    when:
-      and:
-        - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.test_release >>
-    jobs:
-      - build_release:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          matrix:
-            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              build_env: ["PYPI_RELEASE=1"]
-              xcode_version: ["26.0.0"]
-      - build_documentation:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          upload-docs: true
-      - build_linux_release:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          matrix:
-            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-              build_env: ["PYPI_RELEASE=1"]
-      - build_cuda_release:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          matrix:
-            parameters:
-              build_env: ["PYPI_RELEASE=1"]
-
-  prb:
-    when:
-      matches:
-        pattern: "^pull/\\d+(/head)?$"
-        value: << pipeline.git.branch >>
-    jobs:
-      - hold:
-          type: approval
-      - apple/authenticate:
-          context: pr-approval
-      - mac_build_and_test:
-          requires: [ hold ]
-          matrix:
-            parameters:
-              macosx_deployment_target: ["13.5", "15.0"]
-      - linux_build_and_test:
-          requires: [ hold ]
-      - cuda_build_and_test:
-          requires: [ hold ]
-          matrix:
-            parameters:
-              image_date: ["2023.11.1", "2025.05.1"]
-  nightly_build:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.nightly_build >>
-    jobs:
-      - build_release:
-          matrix:
-            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              xcode_version: ["26.0.0"]
-      - build_linux_release:
-          matrix:
-            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-      - build_cuda_release
-
-  build_dev_release:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.test_release >>
-    jobs:
-      - build_release:
-          matrix:
-            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              build_env: ["DEV_RELEASE=1"]
-              xcode_version: ["26.0.0"]
-      - build_linux_release:
-          matrix:
-            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-              build_env: ["DEV_RELEASE=1"]
-      - build_cuda_release:
-          matrix:
-            parameters:
-              build_env: ["DEV_RELEASE=1"]
@@ -0,0 +1,31 @@
+name: 'Build CUDA wheel'
+description: 'Build CUDA wheel'
+
+inputs:
+  arch:
+    description: 'Platform architecture tag'
+    required: true
+    type: choice
+    options:
+      - x86_64
+      - aarch64
+
+runs:
+  using: "composite"
+  steps:
+    - name: Build package
+      shell: bash
+      env:
+        CMAKE_ARGS: -DMLX_BUILD_CUDA=ON
+      run: |
+        pip install auditwheel build patchelf setuptools
+        python setup.py clean --all
+        MLX_BUILD_STAGE=2 python -m build -w
+
+        auditwheel repair dist/mlx_cuda*.whl \
+          --plat manylinux_2_35_${{ inputs.arch }} \
+          --exclude libcublas* \
+          --exclude libcuda* \
+          --exclude libcudnn* \
+          --exclude libnccl* \
+          --exclude libnvrtc*
@@ -0,0 +1,38 @@
+name: 'Build Documentation'
+description: 'Build documentation'
+
+runs:
+  using: "composite"
+  steps:
+    - name: Setup machine
+      uses: ./.github/actions/setup-linux
+
+    - name: Install dependencies
+      shell: bash
+      run: |
+        sudo apt-get install -y doxygen
+        source .venv/bin/activate
+        pip install -r docs/requirements.txt
+        pip install . -v
+  
+    - name: Build documentation
+      shell: bash
+      run: |
+        source .venv/bin/activate
+        cd docs
+        doxygen
+        make html O=-W
+    
+    - name: Create artifact tar
+      shell: bash
+      run: tar -cf artifact.tar -C docs --dereference build/html index.html
+
+    # Do it manually because upload-pages-artifact requires gtar
+    - name: Upload artifact
+      id: upload-artifact
+      uses: actions/upload-artifact@v5
+      with:
+        name: github-pages
+        path: artifact.tar
+        retention-days: 1
+        if-no-files-found: error
@@ -0,0 +1,42 @@
+name: 'Build Linux wheel'
+description: 'Build Linux wheel'
+
+inputs:
+  build-backend:
+    description: 'Build the backend mlx-cpu package'
+    type: boolean
+    required: false
+    default: false
+  arch:
+    description: 'Platform architecture tag'
+    required: true
+    type: choice
+    options:
+      - x86_64
+      - aarch64
+
+runs:
+  using: "composite"
+  steps:
+    - name: Build MLX
+      shell: bash
+      run: pip install -e . -v
+
+    - name: Build Python package
+      shell: bash
+      run: |
+        pip install auditwheel patchelf build
+        python setup.py clean --all
+        MLX_BUILD_STAGE=1 python -m build -w
+        auditwheel repair dist/mlx-*.whl \
+          --plat manylinux_2_35_${{ inputs.arch }} \
+          --exclude libmlx.so* \
+          --only-plat
+
+    - name: Build backend package
+      if: ${{ inputs.build-backend }}
+      shell: bash
+      run: |
+        python setup.py clean --all
+        MLX_BUILD_STAGE=2 python -m build -w
+        auditwheel repair dist/mlx_cpu*.whl --plat manylinux_2_35_${{ inputs.arch }}
@@ -0,0 +1,38 @@
+name: 'Build and Test on Linux'
+
+inputs:
+  toolkit:
+    description: 'The toolkit to build with'
+    required: false
+    default: 'cpu'
+
+runs:
+  using: "composite"
+  steps:
+
+    - name: Install Python package
+      id: python_build
+      shell: sh
+      env:
+        DEBUG: 1
+        CMAKE_ARGS: >-
+          -DCMAKE_COMPILE_WARNING_AS_ERROR=ON
+          -DMLX_BUILD_CUDA=${{ startsWith(inputs.toolkit, 'cuda') && 'ON' || 'OFF' }}
+      run: |
+        if ${{ startsWith(inputs.toolkit, 'cuda') && runner.arch == 'arm64' }} ; then
+          # There is no GPU in arm64 runner, use a common arch.
+          CMAKE_ARGS="$CMAKE_ARGS -DMLX_CUDA_ARCHITECTURES=90a"
+          # Can not build tests and stubs when the built executables can not run.
+          CMAKE_ARGS="$CMAKE_ARGS -DMLX_BUILD_TESTS=OFF -DMLX_BUILD_PYTHON_STUBS=OFF"
+        fi
+        # Install cpu-only torch to save space
+        pip install torch --index-url https://download.pytorch.org/whl/cpu
+        pip install --no-build-isolation -e ".[dev]" -v
+        # Pass the CMAKE_ARGS to following steps.
+        echo CMAKE_ARGS="$CMAKE_ARGS" >> $GITHUB_OUTPUT
+
+    - name: Build CPP only
+      shell: bash
+      run: |
+        cmake . -B build -DCMAKE_BUILD_TYPE=Debug ${{ steps.python_build.outputs.CMAKE_ARGS }}
+        cmake --build build -j $(nproc)
@@ -0,0 +1,36 @@
+name: 'Build macOS release'
+description: 'Build MLX releases macOS'
+
+inputs:
+  macos-target:
+    description: 'macOS build target'
+    required: false
+    default: '15.0'
+  build-backend:
+    description: 'Build the backend mlx-metal package'
+    type: boolean
+    required: false
+    default: false
+
+runs:
+  using: "composite"
+  steps:
+    - name: Build Python package
+      shell: bash -l {0}
+      env:
+        DEVELOPER_DIR: /Applications/Xcode-latest.app
+        MACOSX_DEPLOYMENT_TARGET: ${{ inputs.macos-target }}
+      run: |
+        pip install build
+        python setup.py clean --all
+        MLX_BUILD_STAGE=1 python -m build -w
+
+    - name: Build backend package
+      if: ${{ inputs.build-backend }}
+      shell: bash -l {0}
+      env:
+        DEVELOPER_DIR: /Applications/Xcode-latest.app
+        MACOSX_DEPLOYMENT_TARGET: ${{ inputs.macos-target }}
+      run: |
+        python setup.py clean --all
+        MLX_BUILD_STAGE=2 python -m build -w
@@ -0,0 +1,80 @@
+name: 'Build and Test on macOS'
+description: 'Build and test MLX on macOS'
+
+runs:
+  using: "composite"
+  steps:
+    - name: Install dependencies
+      env:
+        DEBUG: 1
+        CMAKE_ARGS: "-DCMAKE_COMPILE_WARNING_AS_ERROR=ON"
+      shell: bash -l {0}
+      run: |
+        pip install --upgrade pip
+        pip install cmake setuptools typing_extensions
+        pip install -e . -v
+
+    - name: Install tests dependencies
+      shell: bash -l {0}
+      run: |
+        pip install numpy torch tensorflow
+
+    - name: Run Python tests
+      shell: bash -l {0}
+      env:
+        LOW_MEMORY: 1
+      run: |
+        DEVICE=cpu python -m unittest discover -v python/tests
+        DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m unittest discover -v python/tests
+        mpirun --bind-to none -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
+        mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
+        if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
+    
+    - name: Build example extension
+      shell: bash -l {0}
+      run: |
+        cd examples/extensions
+        pip install -r requirements.txt
+        python setup.py build_ext --inplace
+        python test.py
+    
+    - name: Build CPP only
+      shell: bash -l {0}
+      run: |
+        mkdir -p build
+        cd build
+        cmake ..
+        make -j $(sysctl -n hw.ncpu)
+    
+    - name: Run CPP tests
+      shell: bash -l {0}
+      env:
+        DEVICE: gpu
+        METAL_DEVICE_WRAPPER_TYPE: 1
+        METAL_DEBUG_ERROR_MODE: 0
+      run: ./build/tests/tests
+    
+    - name: Build small binary with JIT
+      shell: bash -l {0}
+      run: |
+        mkdir -p build
+        cd build
+        cmake .. -DCMAKE_BUILD_TYPE=MinSizeRel \
+          -DBUILD_SHARED_LIBS=ON \
+          -DMLX_BUILD_CPU=OFF \
+          -DMLX_BUILD_SAFETENSORS=OFF \
+          -DMLX_BUILD_GGUF=OFF \
+          -DMLX_METAL_JIT=ON
+        make -j $(sysctl -n hw.ncpu)
+    
+    - name: Run Python tests with JIT
+      shell: bash -l {0}
+      env:
+        LOW_MEMORY: 1
+        DEVICE: gpu
+        METAL_DEVICE_WRAPPER_TYPE: 1
+        METAL_DEBUG_ERROR_MODE: 0
+      run: |
+        CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
+          pip install -e . -v
+        python -m unittest discover -v python/tests
@@ -0,0 +1,26 @@
+name: 'Build on Windows'
+
+runs:
+  using: 'composite'
+  steps:
+    - name: Install Python package
+      id: python-build
+      shell: cmd
+      env:
+        # For MSVC, Ninja/Release is the only config supported by ccache.
+        CMAKE_ARGS: >-
+          -G Ninja
+          -DCMAKE_BUILD_TYPE=Release
+          -DCMAKE_C_COMPILER=cl
+          -DCMAKE_CXX_COMPILER=cl
+          -DCMAKE_RC_COMPILER=rc
+      run: |
+        uv pip install ".[dev]" -v
+        :: Pass the CMAKE_ARGS to following steps.
+        >>%GITHUB_OUTPUT% ECHO CMAKE_ARGS=%CMAKE_ARGS%
+
+    - name: Build CPP only
+      shell: cmd
+      run: |
+        cmake . -B build ${{ steps.python-build.outputs.CMAKE_ARGS }}
+        cmake --build build -j %NUMBER_OF_PROCESSORS%
@@ -0,0 +1,93 @@
+name: 'Setup Linux Environment'
+description: 'Install dependencies for Linux builds'
+
+inputs:
+  toolkit:
+    description: 'Which toolkit to install'
+    required: false
+    default: 'cpu'
+  python-version:
+    description: 'Version of python to set up'
+    required: false
+    default: '3.14'
+  use-ccache:
+    description: 'Whether to enable ccache'
+    required: false
+    default: 'true'
+
+runs:
+  using: "composite"
+  steps:
+    - name: Install common dependencies
+      shell: bash
+      run: |
+        echo "::group::Install common dependencies"
+        sudo apt-get update
+        sudo apt-get install -y --no-install-recommends \
+            zip \
+            libblas-dev liblapack-dev liblapacke-dev \
+            openmpi-bin openmpi-common libopenmpi-dev
+        echo "::endgroup::"
+
+    - name: Use ccache
+      if: ${{ inputs.use-ccache == 'true' }}
+      uses: hendrikmuhs/ccache-action@v1.2
+      with:
+        key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ inputs.toolkit }}
+        max-size: 1GB
+        # ccache-action bug: running "apt-get update" fails on large arm runner.
+        update-package-index: false
+
+    - uses: actions/setup-python@v6
+      with:
+        python-version: ${{ inputs.python-version }}
+
+    - name: Setup Python venv
+      shell: bash
+      run: |
+        echo "::group::Setup Python venv"
+        python -m venv .venv
+        source .venv/bin/activate
+        pip install setuptools cmake typing_extensions
+        echo PATH=$PATH >> $GITHUB_ENV
+        # Search python packages in .venv
+        echo PYTHONPATH=`python -c 'import sys; print(sys.path[-1])'` >> $GITHUB_ENV
+        echo "::endgroup::"
+
+    - name: Install CUDA toolkit
+      if: ${{ startsWith(inputs.toolkit, 'cuda') }}
+      shell: bash
+      env:
+        # Note: the CI machine does not meet CUDA 13's driver requirement.
+        # Compatibility matrix:
+        # https://docs.nvidia.com/deeplearning/cudnn/backend/latest/reference/support-matrix.html
+        PACKAGES: |
+          {
+            "cuda-12.6": "libcudnn9-dev-cuda-12 cuda-compiler-12-6 cuda-libraries-dev-12-6",
+            "cuda-12.9": "libcudnn9-dev-cuda-12 cuda-compiler-12-9 cuda-libraries-dev-12-9",
+            "cuda-13.0": "libcudnn9-dev-cuda-13 cuda-compiler-13-0 cuda-libraries-dev-13-0"
+          }
+      run: |
+        echo "::group::Install CUDA toolkit"
+        # The CUDA binaries are hosted in the "sbsa" repo, the "arm64" repo is
+        # Jetson specific. SBSA means Arm Server Base System Architecture.
+        ARCH=${{ runner.arch == 'arm64' && 'sbsa' || 'x86_64' }}
+        wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/$ARCH/cuda-keyring_1.1-1_all.deb
+        sudo dpkg -i cuda-keyring_1.1-1_all.deb
+        sudo apt-get update
+        sudo apt-get install -y --no-install-recommends \
+            libnccl2 libnccl-dev \
+            ${{ fromJson(env.PACKAGES)[inputs.toolkit] }}
+        echo "/usr/local/${{ inputs.toolkit }}/bin" >> $GITHUB_PATH
+        echo "::endgroup::"
+
+    - name: CUDA packages and driver report
+      if: ${{ startsWith(inputs.toolkit, 'cuda') }}
+      shell: bash
+      run: |
+        echo "::group::Installed NVIDIA and CUDA packages"
+        dpkg -l | egrep "cuda|nvidia" -i
+        echo "::endgroup::"
+        echo "::group::NVIDIA-SMI Status"
+        nvidia-smi || true
+        echo "::endgroup::"
@@ -0,0 +1,24 @@
+name: 'Setup macOS Environment'
+description: 'Install dependencies for macOS builds'
+
+inputs:
+  python-version:
+    description: 'Python version to use'
+    required: false
+    default: '3.10'
+
+runs:
+  using: "composite"
+  steps:
+    - name: Install Homebrew packages
+      shell: sh
+      run: /opt/homebrew/bin/brew install openmpi
+    
+    - name: Verify MetalToolchain installed
+      shell: bash
+      run: xcodebuild -showComponent MetalToolchain
+
+    - uses: conda-incubator/setup-miniconda@v3
+      with:
+        miniconda-version: "latest"
+        python-version: ${{ inputs.python-version }}
@@ -0,0 +1,42 @@
+name: 'Setup Windows environment'
+
+inputs:
+  python-version:
+    description: 'Version of python to set up'
+    required: false
+    default: '3.14'
+  use-ccache:
+    description: 'Whether to enable ccache'
+    required: false
+    default: 'true'
+
+runs:
+  using: 'composite'
+  steps:
+    - name: Use ccache
+      if: ${{ inputs.use-ccache == 'true' }}
+      uses: hendrikmuhs/ccache-action@v1.2
+      with:
+        key: ccache-${{ runner.os }}-${{ runner.arch }}-cpu
+        max-size: 1GB
+
+    - name: Setup Visual Studio cmd
+      shell: cmd
+      run: |
+        :: Find out path to VS.
+        pushd "C:\Program Files (x86)\Microsoft Visual Studio\Installer\"
+        for /f "delims=" %%x in ('.\vswhere.exe -latest -property InstallationPath') do set VSPATH=%%x
+        popd
+        :: Import VS vars.
+        call "%VSPATH%\VC\Auxiliary\Build\vcvarsall.bat" x64
+        :: Export to all steps.
+        >>%GITHUB_ENV% set
+
+    - uses: astral-sh/setup-uv@v7
+
+    - name: Setup Python venv
+      shell: cmd
+      run: |
+        uv venv --python ${{ inputs.python-version }}
+        call ".venv/Scripts/activate.bat"
+        >>%GITHUB_ENV% set
@@ -0,0 +1,69 @@
+name: 'Run Linux tests'
+
+inputs:
+  has-gpu:
+    description: 'Run GPU tests'
+    required: false
+    default: false
+
+runs:
+  using: "composite"
+  steps:
+    - name: Run MPI tests
+      shell: bash
+      run: |
+        echo "::group::MPI tests"
+        mpirun --bind-to none --allow-run-as-root -host localhost:8 -np 8 python python/tests/mpi_test_distributed.py
+        echo "::endgroup::"
+
+    - name: Run distributed tests
+      if: ${{ inputs.has-gpu == 'false' }}
+      shell: bash
+      run: |
+        echo "::group::Distributed tests"
+        mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
+        if grep -Fq '[WARN]' stderr.log ; then
+          grep -F '[WARN]' stderr.log
+          echo "Distributed ring test failed";
+          exit 1;
+        fi
+        echo "::endgroup::"
+
+    - name: Run Python tests - CPU
+      if: ${{ inputs.has-gpu == 'false' }}
+      shell: bash
+      env:
+        DEVICE: cpu
+      run: |
+        echo "::group::Python tests - CPU"
+        python -m unittest discover python/tests -v
+        echo "::endgroup::"
+
+    - name: Run Python tests - GPU
+      if: ${{ inputs.has-gpu == 'true' }}
+      shell: bash
+      env:
+        DEVICE: gpu
+      run: |
+        echo "::group::Python tests - GPU"
+        python -m tests discover python/tests -v
+        echo "::endgroup::"
+
+    - name: Run CPP tests - CPU
+      shell: bash
+      env:
+        DEVICE: cpu
+      run: |
+        echo "::group::CPP tests - CPU"
+        ./build/tests/tests
+        echo "::endgroup::"
+
+    - name: Run CPP tests - GPU
+      if: ${{ inputs.has-gpu == 'true' }}
+      shell: bash
+      env:
+        DEVICE: gpu
+      run: |
+        echo "::group::CPP tests - GPU"
+        ./build/tests/tests -sfe="*fft_tests.cpp,*linalg_tests.cpp"
+        echo "::endgroup::"
@@ -0,0 +1,20 @@
+name: 'Run tests on Windows'
+
+runs:
+  using: 'composite'
+  steps:
+    - name: Run Python tests - CPU
+      shell: bash
+      run: |
+        echo "::group::Python tests - CPU"
+        python -m unittest discover python/tests -v
+        echo "::endgroup::"
+
+    - name: Run CPP tests - CPU
+      shell: bash
+      env:
+        DEVICE: cpu
+      run: |
+        echo "::group::CPP tests - CPU"
+        ./build/tests.exe -tce="*gguf*,test random uniform"
+        echo "::endgroup::"
@@ -0,0 +1,6 @@
+version: 2
+updates:
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"
@@ -0,0 +1,48 @@
+#!/bin/bash
+set -ex
+
+export CMAKE_C_COMPILER=/usr/bin/clang
+export CMAKE_CXX_COMPILER=/usr/bin/clang++
+BASE_CMAKE_ARGS="-DCMAKE_BUILD_TYPE=DEBUG -DCMAKE_COMPILE_WARNING_AS_ERROR=ON"
+if [[ "$(uname -s)" != "Darwin" ]]; then
+  BASE_CMAKE_ARGS+=" -DMLX_BUILD_METAL=OFF"
+fi
+
+run_test() {
+  local sanitizer_name=$1
+  local cmake_sanitizer_flag="-DUSE_${sanitizer_name}=ON"
+  echo "  Running tests with: ${sanitizer_name}"
+
+  case "$sanitizer_name" in
+    ASAN)
+      export ASAN_OPTIONS="detect_leaks=0"
+      ;;
+    UBSAN)
+      export UBSAN_OPTIONS="halt_on_error=0:print_stacktrace=1"
+      ;;
+    TSAN)
+      export TSAN_OPTIONS=""
+      ;;
+  esac
+
+  rm -rf build
+  mkdir -p build
+  pushd build > /dev/null
+
+  cmake .. ${BASE_CMAKE_ARGS} ${cmake_sanitizer_flag}
+  make -j $(nproc)
+  ./tests/tests
+
+  popd > /dev/null
+  unset ${sanitizer_name}_OPTIONS
+}
+
+sanitizer_arg=$(echo "$1" | tr '[:lower:]' '[:upper:]')
+
+if [[ "$sanitizer_arg" == "ASAN" || "$sanitizer_arg" == "UBSAN" || "$sanitizer_arg" == "TSAN" ]]; then
+  run_test "$sanitizer_arg"
+  echo "  ${sanitizer_arg} test run completed successfully."
+else
+  echo "Error: Invalid sanitizer '$1'. Please use one of: ASAN, UBSAN, TSAN."
+  exit 1
+fi
@@ -0,0 +1,27 @@
+#!/bin/bash
+set -ex
+
+# [Setup] Install dependencies inside the container.
+dnf update -y
+dnf install -y \
+  blas-devel \
+  lapack-devel \
+  openblas-devel \
+  make \
+  cmake \
+  clang \
+  git
+dnf clean all
+
+# [C++] CI Build Sanity Check: Verifies code compilation, not for release.
+export CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON"
+export DEBUG=1
+export CMAKE_C_COMPILER=/usr/bin/clang
+export CMAKE_CXX_COMPILER=/usr/bin/clang++
+
+mkdir -p build
+pushd build
+cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
+make -j $(nproc)
+./tests/tests
+popd
@@ -0,0 +1,152 @@
+name: Build and Test
+
+on:
+  pull_request:
+  push:
+    branches:
+      - main
+      # For testing CI without starting a pull request:
+      - test/*
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+
+jobs:
+  check_lint:
+    name: Check Lint
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v6
+      - uses: pre-commit/action@v3.0.1
+
+  linux_build_and_test:
+    name: Linux (cpu, ${{ matrix.arch }})
+    needs: check_lint
+    strategy:
+      fail-fast: false
+      matrix:
+        arch: ['x86_64', 'aarch64']
+    runs-on: ${{ matrix.arch == 'x86_64' && 'ubuntu-22.04' || 'ubuntu-22.04-arm' }}
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-linux
+      - uses: ./.github/actions/build-linux
+      - uses: ./.github/actions/test-linux
+      - run: df -h
+
+  cuda_build_and_test:
+    name: Linux (${{ matrix.toolkit }}, ${{ matrix.arch }})
+    if: github.repository == 'ml-explore/mlx'
+    needs: check_lint
+    strategy:
+      fail-fast: false
+      matrix:
+        arch: ['x86_64', 'aarch64']
+        toolkit: ['cuda-12.6', 'cuda-12.9']
+    runs-on: ${{ matrix.arch == 'x86_64' && 'gpu-t4-4-core' || 'ubuntu-22.04-arm' }}
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-linux
+        with:
+          toolkit: ${{ matrix.toolkit }}
+      - uses: ./.github/actions/build-linux
+        with:
+          toolkit: ${{ matrix.toolkit }}
+      - uses: ./.github/actions/test-linux
+        if: matrix.arch == 'x86_64'
+        with:
+          has-gpu: true
+
+  mac_build_and_test:
+    name: macOS (${{ matrix.macos-target }})
+    if: github.repository == 'ml-explore/mlx'
+    strategy:
+      matrix:
+        macos-target: ["14.0", "15.0", "26.0"]
+    runs-on: [self-hosted, macos]
+    env:
+      MACOSX_DEPLOYMENT_TARGET: ${{ matrix.macos-target }}
+    needs: check_lint
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-macos
+      - uses: ./.github/actions/build-macos
+
+  windows_build_and_test:
+    name: Windows (cpu, x86_64)
+    needs: check_lint
+    runs-on: windows-2025
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-windows
+      - uses: ./.github/actions/build-windows
+      - uses: ./.github/actions/test-windows
+
+  build_documentation:
+    name: Build Documentation
+    if: github.repository == 'ml-explore/mlx'
+    runs-on: ubuntu-22.04
+    needs: check_lint
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/build-docs
+
+  linux_sanitizer_build_and_test:
+    name: Linux Sanitizer Tests (${{ matrix.sanitizer }})
+    needs: check_lint
+    strategy:
+      fail-fast: false
+      matrix:
+        sanitizer: [ASAN, UBSAN]
+        # todo 12/16/2025: enable TSAN later + consider enabling ASAN for GPU backend tests.
+        # sanitizer: [ASAN, UBSAN, TSAN]
+    runs-on: ubuntu-22.04-arm
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v6
+
+      - name: Install Dependencies
+        run: |
+          export DEBIAN_FRONTEND=noninteractive
+          sudo apt-get update -y
+          sudo apt-get install -y \
+            build-essential \
+            libblas-dev \
+            liblapacke-dev \
+            libopenblas-dev \
+            cmake \
+            clang \
+            git
+          sudo apt-get clean
+          sudo rm -rf /var/lib/apt/lists/*
+
+      - name: Linux Build and Test with ${{ matrix.sanitizer }}
+        run: |
+          bash .github/scripts/build-sanitizer-tests.sh ${{ matrix.sanitizer }}
+
+  linux_fedora_build_cpp:
+    name: Linux Fedora (${{ matrix.arch }})
+    needs: check_lint
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - host: ubuntu-22.04
+            arch: x86_64
+          - host: ubuntu-22.04-arm
+            arch: aarch64
+
+    runs-on: ${{ matrix.host }}
+    container:
+      image: fedora:42
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v6
+
+      - name: CPP Build Test - No Release
+        run: |
+          bash ./.github/scripts/setup+build-cpp-linux-fedora-container.sh
@@ -0,0 +1,28 @@
+name: Documentation
+
+on:
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  build:
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/build-docs
+      
+  deploy:
+    needs: build
+    permissions:
+      pages: write
+      id-token: write
+    runs-on: ubuntu-latest
+    environment:
+      name: github-pages
+      url: ${{ steps.deployment.outputs.page_url }}
+    steps:
+      - name: Deploy to GitHub Pages
+        id: deployment
+        uses: actions/deploy-pages@v4
@@ -0,0 +1,104 @@
+name: Nightly Build
+
+on:
+  schedule:
+    - cron: 33 6 * * 1-5
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  build_linux_release:
+    strategy:
+      fail-fast: false
+      matrix:
+        python_version: ["3.10", "3.14"]
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-linux
+      - uses: ./.github/actions/build-linux-release
+        with:
+          build-backend: ${{ matrix.python-version == '3.10' }}
+          arch: "x86_64"
+      - name: Upload mlx artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          name: linux-wheels-${{ matrix.python_version }}
+          path: wheelhouse/mlx-*.whl
+          retention-days: 7
+      - name: Upload mlx-cpu artifacts
+        if: matrix.python_version == '3.10'
+        uses: actions/upload-artifact@v6
+        with:
+          name: mlx-cpu
+          path: wheelhouse/mlx_cpu-*.whl
+          retention-days: 7
+      - run: df -h
+
+  build_linux_with_tests:
+    strategy:
+      fail-fast: false
+      matrix:
+        python_version: ["3.11", "3.12", "3.13", "3.14"]
+        runner:
+          - ubuntu-22.04
+          - ubuntu-22.04-arm
+    runs-on: ${{ matrix.runner }}
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-linux
+        with:
+          python-version: ${{ matrix.python_version }}
+      - uses: ./.github/actions/build-linux
+      - uses: ./.github/actions/test-linux
+      - run: df -h
+
+  build_mac_release:
+    if: github.repository == 'ml-explore/mlx'
+    strategy:
+      matrix:
+        python-version: ["3.10", "3.13"]
+    runs-on: [self-hosted, macos]
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-macos
+        with:
+          python-version: ${{ matrix.python-version }}
+      - uses: ./.github/actions/build-macos
+      - name: Build macOS 26 package
+        uses: ./.github/actions/build-macos-release
+        with:
+          macos-target: 26.0
+          build-backend: ${{ matrix.python-version == '3.10' }}
+      - name: Build macOS 15 package
+        uses: ./.github/actions/build-macos-release
+        with:
+          macos-target: 15.0
+          build-backend: ${{ matrix.python-version == '3.10' }}
+      - name: Build macOS 14 package
+        uses: ./.github/actions/build-macos-release
+        with:
+          macos-target: 14.0
+          build-backend: ${{ matrix.python-version == '3.10' }}
+
+  build_cuda_release:
+    if: github.repository == 'ml-explore/mlx'
+    runs-on: ubuntu-22-large
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-linux
+        with:
+          toolkit: 'cuda-12.9'
+      - name: Build Python package
+        uses: ./.github/actions/build-cuda-release
+        with:
+          toolkit: 'cuda-12.9'
+          arch: 'x86_64'
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          name: mlx-cuda
+          path: wheelhouse/mlx_cuda_*.whl
+          retention-days: 7
@@ -1,20 +0,0 @@
-on:
-  pull_request:
-    branches:
-      - main
-
-jobs:
-  check_lint:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-      - uses: actions/setup-python@v4
-        with:
-          python-version: 3.8
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install pre-commit black isort clang-format
-      - name: Run lint
-        run: |
-          pre-commit run --all-files
@@ -0,0 +1,256 @@
+name: PyPI Release
+
+on:
+  push:
+    tags:
+      - 'v*'
+    branches:
+      - 'test-publish/*'
+  workflow_dispatch:
+    inputs:
+      dry_run:
+        description: 'Dry run (do not publish to PyPi)'
+        required: false
+        type: boolean
+      dev_release:
+        description: 'Development release (DEV_RELEASE=1)'
+        required: false
+        type: boolean
+
+permissions:
+  contents: read
+
+jobs:
+  build_documentation:
+    if: github.repository == 'ml-explore/mlx'
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/build-docs
+
+  deploy_documentation:
+    if: ${{ !inputs.dry_run }}
+    needs: build_documentation
+    permissions:
+      pages: write
+      id-token: write
+    runs-on: ubuntu-latest
+    environment:
+      name: github-pages
+      url: ${{ steps.deployment.outputs.page_url }}
+    steps:
+      - name: Deploy to GitHub Pages
+        id: deployment
+        uses: actions/deploy-pages@v4
+
+  build_linux_release:
+    if: github.repository == 'ml-explore/mlx'
+    strategy:
+      matrix:
+        python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+        arch: ['x86_64', 'aarch64']
+    runs-on: ${{ matrix.arch == 'x86_64' && 'ubuntu-22.04' || 'ubuntu-22.04-arm' }}
+    env:
+      PYPI_RELEASE: 1
+      DEV_RELEASE: ${{ inputs.dev_release && 1 || 0 }}
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-linux
+        with:
+          python-version: ${{ matrix.python_version }}
+          use-ccache: false
+      - uses: ./.github/actions/build-linux-release
+        with:
+          build-backend: ${{ matrix.python_version == '3.10' }}
+          arch: ${{ matrix.arch }}
+      - name: Upload MLX artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          overwrite: true
+          name: linux-wheels-${{ matrix.python_version }}-${{ matrix.arch }}
+          path: wheelhouse/mlx-*.whl
+          if-no-files-found: error
+      - name: Upload CPU artifacts
+        if: matrix.python_version == '3.10'
+        uses: actions/upload-artifact@v6
+        with:
+          overwrite: true
+          name: mlx-cpu-${{ matrix.arch }}
+          path: wheelhouse/mlx_cpu-*.whl
+          if-no-files-found: error
+
+  build_mac_release:
+    if: github.repository == 'ml-explore/mlx'
+    strategy:
+      matrix:
+        python-version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+    runs-on: [self-hosted, macos]
+    env:
+      PYPI_RELEASE: 1
+      DEV_RELEASE: ${{ inputs.dev_release && 1 || 0 }}
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-macos
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install dependencies
+        shell: bash -l {0}
+        run: |
+          pip install --upgrade pip
+          pip install cmake setuptools typing_extensions
+          pip install -e . -v
+      - name: Build macOS 14 package
+        uses: ./.github/actions/build-macos-release
+        with:
+          macos-target: 14.0
+          build-backend: ${{ matrix.python-version == '3.10' }}
+      - name: Build macOS 15 package
+        uses: ./.github/actions/build-macos-release
+        with:
+          macos-target: 15.0
+          build-backend: ${{ matrix.python-version == '3.10' }}
+      - name: Build macOS 26 package
+        uses: ./.github/actions/build-macos-release
+        with:
+          macos-target: 26.0
+          build-backend: ${{ matrix.python-version == '3.10' }}
+      - name: Upload MLX artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          overwrite: true
+          name: mac-wheels-${{ matrix.python-version }}
+          path: dist/mlx-*.whl
+          if-no-files-found: error
+      - name: Upload Metal artifacts
+        if: matrix.python-version == '3.10'
+        uses: actions/upload-artifact@v6
+        with:
+          overwrite: true
+          name: mlx-metal
+          path: dist/mlx_metal-*.whl
+          if-no-files-found: error
+
+  build_cuda_release:
+    if: github.repository == 'ml-explore/mlx'
+    strategy:
+      matrix:
+        arch: ['x86_64', 'aarch64']
+        toolkit: ['cuda-12.9', 'cuda-13.0']
+    runs-on: ${{ matrix.arch == 'x86_64' && 'ubuntu-22-large' || 'ubuntu-22-large-arm' }}
+    env:
+      PYPI_RELEASE: 1
+      DEV_RELEASE: ${{ inputs.dev_release && 1 || 0 }}
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-linux
+        with:
+          toolkit: ${{ matrix.toolkit }}
+          use-ccache: false
+      - name: Build Python package
+        uses: ./.github/actions/build-cuda-release
+        with:
+          arch: ${{ matrix.arch }}
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          overwrite: true
+          name: mlx-${{ matrix.toolkit }}-${{ matrix.arch }}
+          path: wheelhouse/mlx_cuda_*.whl
+          if-no-files-found: error
+
+  pypi-publish:
+    name: Upload release to PyPI
+    runs-on: ubuntu-latest
+    needs: [build_linux_release, build_mac_release]
+    permissions:
+      id-token: write
+    environment:
+      name: ${{ inputs.dry_run && 'dry-run' || 'pypi' }}
+      url: https://pypi.org/p/mlx
+    steps:
+      - uses: actions/download-artifact@v7
+        with:
+          pattern: linux-wheels-*
+          merge-multiple: true
+          path: dist
+      - uses: actions/download-artifact@v7
+        with:
+          pattern: mac-wheels-*
+          merge-multiple: true
+          path: dist
+      - name: Display structure of downloaded files
+        run: du -ah dist
+      - name: Publish package distributions to PyPI
+        if: ${{ !inputs.dry_run }}
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          repository-url: https://upload.pypi.org/legacy/
+
+  pypi-publish-cuda:
+    name: Upload CUDA release to PyPI
+    runs-on: ubuntu-latest
+    needs: [build_cuda_release]
+    permissions:
+      id-token: write
+    environment:
+      name: ${{ inputs.dry_run && 'dry-run' || 'pypi' }}
+      url: https://pypi.org/p/mlx-cuda
+    steps:
+      - uses: actions/download-artifact@v7
+        with:
+          pattern: mlx-cuda-*
+          merge-multiple: true
+          path: dist
+      - name: Display structure of downloaded files
+        run: du -ah dist
+      - name: Publish package distributions to PyPI
+        if: ${{ !inputs.dry_run }}
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          repository-url: https://upload.pypi.org/legacy/
+
+  pypi-publish-cpu:
+    name: Upload CPU release to PyPI
+    runs-on: ubuntu-latest
+    needs: [build_linux_release]
+    permissions:
+      id-token: write
+    environment:
+      name: ${{ inputs.dry_run && 'dry-run' || 'pypi' }}
+      url: https://pypi.org/p/mlx-cpu
+    steps:
+      - uses: actions/download-artifact@v7
+        with:
+          pattern: mlx-cpu-*
+          merge-multiple: true
+          path: dist
+      - name: Display structure of downloaded files
+        run: du -ah dist
+      - name: Publish package distributions to PyPI
+        if: ${{ !inputs.dry_run }}
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          repository-url: https://upload.pypi.org/legacy/
+
+  pypi-publish-metal:
+    name: Upload Metal release to PyPI
+    runs-on: ubuntu-latest
+    needs: [build_mac_release]
+    permissions:
+      id-token: write
+    environment:
+      name: ${{ inputs.dry_run && 'dry-run' || 'pypi' }}
+      url: https://pypi.org/p/mlx-metal
+    steps:
+      - uses: actions/download-artifact@v7
+        with:
+          name: mlx-metal
+          path: dist
+      - name: Display structure of downloaded files
+        run: du -ah dist
+      - name: Publish package distributions to PyPI
+        if: ${{ !inputs.dry_run }}
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          repository-url: https://upload.pypi.org/legacy/
@@ -3,16 +3,12 @@ __pycache__/
 *.py[cod]
 *$py.class

-# C extensions
-*.so
-
 # tensor files
 *.safe
 *.safetensors

 # Metal libraries
 *.metallib
-venv/

 # Distribution / packaging
 python/mlx/core
@@ -30,6 +26,7 @@ lib64/
 parts/
 sdist/
 var/
+venv/
 wheels/
 share/python-wheels/
 *.egg-info/
@@ -37,12 +34,7 @@ share/python-wheels/
 *.egg
 MANIFEST
 uv.lock
-
-# vim
-*.swp
-
-# Ignore build dir
-build/
+.DS_Store

 # Prerequisites
 *.d
@@ -52,6 +44,7 @@ build/
 *.lo
 *.o
 *.obj
+*.ilk

 # Precompiled Headers
 *.gch
@@ -80,9 +73,9 @@ build/
 # Debug symbols
 *.pdb

-# VSCode 
+# VSCode
 .vscode/
-.DS_Store
-
 # Jetbrains
-.cache
+.cache/
+# vim
+*.swp
@@ -1,16 +1,22 @@
 repos:
+-   repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v6.0.0
+    hooks:
+    -   id: check-yaml
+    # -   id: end-of-file-fixer
+    # -   id: trailing-whitespace
 -   repo: https://github.com/pre-commit/mirrors-clang-format
-    rev: v19.1.7
+    rev: v21.1.8
    hooks:
    -   id: clang-format
 # Using this mirror lets us use mypyc-compiled black, which is about 2x faster
 -   repo: https://github.com/psf/black-pre-commit-mirror
-    rev: 25.1.0
+    rev: 26.1.0
    hooks:
    -   id: black
    
 -   repo: https://github.com/pycqa/isort
-    rev: 6.0.0
+    rev: 7.0.0
    hooks:
    -   id: isort
        args:
@@ -22,7 +22,7 @@ project(

 # ----------------------------- Setup -----------------------------
 set(CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
-set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 set(CMAKE_INSTALL_MESSAGE NEVER)
@@ -40,11 +40,14 @@ option(MLX_METAL_DEBUG "Enhance metal debug workflow" OFF)
 option(MLX_ENABLE_X64_MAC "Enable building for x64 macOS" OFF)
 option(MLX_BUILD_GGUF "Include support for GGUF format" ON)
 option(MLX_BUILD_SAFETENSORS "Include support for safetensors format" ON)
-option(MLX_BUILD_BLAS_FROM_SOURCE "Build OpenBLAS from source code" OFF)
+option(MLX_BUILD_PYTHON_STUBS "Build stub files for python bindings" ON)
 option(MLX_METAL_JIT "Use JIT compilation for Metal kernels" OFF)
 option(MLX_USE_CCACHE "Use CCache for compilation cache when available" ON)
 option(BUILD_SHARED_LIBS "Build mlx as a shared library" OFF)
 option(USE_SYSTEM_FMT "Use system's provided fmt library" OFF)
+option(USE_ASAN "Enable AddressSanitizer (ASan)" OFF)
+option(USE_UBSAN "Enable UndefinedBehaviorSanitizer (UBSan)" OFF)
+option(USE_TSAN "Enable ThreadSanitizer (TSan)" OFF)

 # --------------------- Processor tests -------------------------
 message(
@@ -74,12 +77,70 @@ endif()
 if(MLX_USE_CCACHE)
  find_program(CCACHE_PROGRAM ccache)
  if(CCACHE_PROGRAM)
+    message(STATUS "Found CCache: ${CCACHE_PROGRAM}")
    set(CMAKE_C_COMPILER_LAUNCHER "${CCACHE_PROGRAM}")
    set(CMAKE_CXX_COMPILER_LAUNCHER "${CCACHE_PROGRAM}")
    set(CMAKE_CUDA_COMPILER_LAUNCHER "${CCACHE_PROGRAM}")
  endif()
 endif()

+if(USE_ASAN AND USE_TSAN)
+  message(
+    FATAL_ERROR
+      "AddressSanitizer (ASan) and ThreadSanitizer (TSan) are mutually exclusive and cannot be enabled at the same time."
+  )
+endif()
+
+set(SANITIZER_COMPILE_FLAGS "")
+set(SANITIZER_LINK_FLAGS "")
+
+if(USE_ASAN)
+  if(WIN32 AND MSVC)
+    list(APPEND SANITIZER_COMPILE_FLAGS /fsanitize=address)
+    list(APPEND SANITIZER_LINK_FLAGS /fsanitize=address)
+  else()
+    list(APPEND SANITIZER_COMPILE_FLAGS -fsanitize=address)
+    list(APPEND SANITIZER_LINK_FLAGS -fsanitize=address)
+    if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
+      list(APPEND SANITIZER_LINK_FLAGS -lpthread)
+    endif()
+  endif()
+endif()
+
+if(USE_UBSAN)
+  if(WIN32 AND MSVC)
+    if(CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+      list(APPEND SANITIZER_COMPILE_FLAGS -fsanitize=undefined)
+      list(APPEND SANITIZER_LINK_FLAGS -fsanitize=undefined)
+    else()
+      message(
+        WARNING
+          "UndefinedBehaviorSanitizer (UBSan) is not directly supported via a simple flag in MSVC."
+      )
+    endif()
+  else()
+    list(APPEND SANITIZER_COMPILE_FLAGS -fsanitize=undefined)
+    list(APPEND SANITIZER_LINK_FLAGS -fsanitize=undefined)
+  endif()
+endif()
+
+if(USE_TSAN)
+  if(WIN32 AND MSVC)
+    message(
+      FATAL_ERROR
+        "ThreadSanitizer (TSan) is not supported by the MSVC compiler. Please use Clang or GCC."
+    )
+  elseif(CMAKE_SYSTEM_NAME STREQUAL "Darwin")
+    message(FATAL_ERROR "ThreadSanitizer (TSan) is not supported on macOS.")
+  else()
+    list(APPEND SANITIZER_COMPILE_FLAGS -fsanitize=thread)
+    list(APPEND SANITIZER_LINK_FLAGS -fsanitize=thread)
+    if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
+      list(APPEND SANITIZER_LINK_FLAGS -lpthread)
+    endif()
+  endif()
+endif()
+
 # ----------------------------- Lib -----------------------------

 include(FetchContent)
@@ -88,8 +149,13 @@ cmake_policy(SET CMP0135 NEW)

 add_library(mlx)

+target_compile_options(mlx PUBLIC ${SANITIZER_COMPILE_FLAGS})
+target_link_options(mlx PUBLIC ${SANITIZER_LINK_FLAGS})
+
 if(MLX_BUILD_CUDA)
  enable_language(CUDA)
+  find_package(CUDAToolkit REQUIRED)
+  find_package(CUDNN REQUIRED)
 endif()

 if(MLX_BUILD_METAL)
@@ -122,9 +188,12 @@ if(MLX_BUILD_METAL)
  message(STATUS "Building with macOS SDK version ${MACOS_SDK_VERSION}")

  set(METAL_CPP_URL
-      https://developer.apple.com/metal/cpp/files/metal-cpp_macOS15_iOS18.zip)
+      https://developer.apple.com/metal/cpp/files/metal-cpp_26.zip)

  if(NOT CMAKE_OSX_DEPLOYMENT_TARGET STREQUAL "")
+    if(${CMAKE_OSX_DEPLOYMENT_TARGET} LESS 14.0)
+      message(FATAL_ERROR "MLX requires macOS >= 14.0")
+    endif()
    set(XCRUN_FLAGS "-mmacosx-version-min=${CMAKE_OSX_DEPLOYMENT_TARGET}")
  endif()
  execute_process(
@@ -133,7 +202,6 @@ if(MLX_BUILD_METAL)
      "echo \"__METAL_VERSION__\" | xcrun -sdk macosx metal ${XCRUN_FLAGS} -E -x metal -P - | tail -1 | tr -d '\n'"
    OUTPUT_VARIABLE MLX_METAL_VERSION COMMAND_ERROR_IS_FATAL ANY)
  FetchContent_Declare(metal_cpp URL ${METAL_CPP_URL})
-
  FetchContent_MakeAvailable(metal_cpp)
  target_include_directories(
    mlx PUBLIC $<BUILD_INTERFACE:${metal_cpp_SOURCE_DIR}>
@@ -151,14 +219,17 @@ if(WIN32)
  if(MSVC)
    # GGUF does not build with MSVC.
    set(MLX_BUILD_GGUF OFF)
-    # There is no prebuilt OpenBLAS distribution for MSVC.
-    set(MLX_BUILD_BLAS_FROM_SOURCE ON)
+  endif()
+  # Generate DLL and EXE in the same dir, otherwise EXE will not be able to run.
+  # This is only done when MLX is built as the top project.
+  if(CMAKE_CURRENT_SOURCE_DIR STREQUAL CMAKE_SOURCE_DIR)
+    set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR})
  endif()
  # Windows implementation of dlfcn.h APIs.
  FetchContent_Declare(
    dlfcn-win32
    GIT_REPOSITORY https://github.com/dlfcn-win32/dlfcn-win32.git
-    GIT_TAG v1.4.1
+    GIT_TAG v1.4.2
    EXCLUDE_FROM_ALL)
  block()
  set(BUILD_SHARED_LIBS OFF)
@@ -182,20 +253,25 @@ if(MLX_BUILD_CPU)
    target_link_libraries(mlx PUBLIC ${ACCELERATE_LIBRARY})
    add_compile_definitions(MLX_USE_ACCELERATE)
    add_compile_definitions(ACCELERATE_NEW_LAPACK)
-  elseif(MLX_BUILD_BLAS_FROM_SOURCE)
-    # Download and build OpenBLAS from source code.
+  elseif(WIN32)
+    # Download and link prebuilt binaries of OpenBLAS. Note that we can only
+    # link with the dynamic library, the prebuilt binaries were built with MinGW
+    # so static-linking would require linking with MinGW's runtime.
    FetchContent_Declare(
      openblas
-      GIT_REPOSITORY https://github.com/OpenMathLib/OpenBLAS.git
-      GIT_TAG v0.3.28
-      EXCLUDE_FROM_ALL)
-    set(BUILD_STATIC_LIBS ON) # link statically
-    set(NOFORTRAN ON) # msvc has no fortran compiler
+      URL "https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.31/OpenBLAS-0.3.31-x64.zip"
+    )
    FetchContent_MakeAvailable(openblas)
-    target_link_libraries(mlx PRIVATE openblas)
-    target_include_directories(
-      mlx PRIVATE "${openblas_SOURCE_DIR}/lapack-netlib/LAPACKE/include"
-                  "${CMAKE_BINARY_DIR}/generated" "${CMAKE_BINARY_DIR}")
+    target_link_libraries(mlx
+                          PRIVATE "${openblas_SOURCE_DIR}/lib/libopenblas.lib")
+    target_include_directories(mlx PRIVATE "${openblas_SOURCE_DIR}/include")
+    # Make sure the DLL file is placed in the same dir with executables.
+    set(OPENBLAS_DLL_FILE "${openblas_SOURCE_DIR}/bin/libopenblas.dll")
+    add_custom_command(
+      TARGET mlx
+      POST_BUILD
+      COMMAND ${CMAKE_COMMAND} -E copy_if_different ${OPENBLAS_DLL_FILE}
+              ${CMAKE_BINARY_DIR})
  else()
    if(${CMAKE_HOST_APPLE})
      # The blas shipped in macOS SDK is not supported, search homebrew for
@@ -247,16 +323,13 @@ target_include_directories(
  mlx PUBLIC $<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}>
             $<INSTALL_INTERFACE:include>)

-# Do not add mlx_EXPORTS define for shared library.
-set_target_properties(mlx PROPERTIES DEFINE_SYMBOL "")
-
 if(USE_SYSTEM_FMT)
  find_package(fmt REQUIRED)
 else()
  FetchContent_Declare(
    fmt
    GIT_REPOSITORY https://github.com/fmtlib/fmt.git
-    GIT_TAG 10.2.1
+    GIT_TAG 12.1.0
    EXCLUDE_FROM_ALL)
  FetchContent_MakeAvailable(fmt)
 endif()
@@ -265,14 +338,16 @@ target_link_libraries(mlx PRIVATE $<BUILD_INTERFACE:fmt::fmt-header-only>)
 if(MLX_BUILD_PYTHON_BINDINGS)
  message(STATUS "Building Python bindings.")
  find_package(
-    Python 3.8
+    Python 3.10
    COMPONENTS Interpreter Development.Module
    REQUIRED)
-  execute_process(
-    COMMAND "${Python_EXECUTABLE}" -m nanobind --cmake_dir
-    OUTPUT_STRIP_TRAILING_WHITESPACE
-    OUTPUT_VARIABLE nanobind_ROOT)
-  find_package(nanobind CONFIG REQUIRED)
+  FetchContent_Declare(
+    nanobind
+    GIT_REPOSITORY https://github.com/wjakob/nanobind.git
+    GIT_TAG v2.10.2
+    GIT_SHALLOW TRUE
+    EXCLUDE_FROM_ALL)
+  FetchContent_MakeAvailable(nanobind)
  add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/python/src)
 endif()

@@ -292,6 +367,15 @@ endif()
 # ----------------------------- Installation -----------------------------
 include(GNUInstallDirs)

+if(WIN32)
+  # Install DLLs to the same dir with extension file (core.pyd) on Windows.
+  set(CMAKE_INSTALL_BINDIR ".")
+  if(MLX_BUILD_CPU)
+    # Install OpenBLAS.
+    install(FILES ${OPENBLAS_DLL_FILE} TYPE BIN)
+  endif()
+endif()
+
 # Install library
 install(
  TARGETS mlx
@@ -75,7 +75,7 @@ void time_irregular_binary_ops_3D() {

 void time_irregular_binary_ops_4D() {
  auto device = mx::default_device();
-  std::vector<int> shape = {8, 8, 512, 512};
+  mx::Shape shape = {8, 8, 512, 512};
  auto a = mx::random::uniform(shape);
  auto b = mx::random::uniform(shape);

@@ -115,7 +115,7 @@ void time_irregular_binary_ops_4D() {

 void time_irregular_reshape() {
  auto device = mx::default_device();
-  std::vector<int> shape;
+  mx::Shape shape;
  auto reshape_fn = [&shape, device](const mx::array& a) {
    return mx::reshape(a, shape, device);
  };
@@ -170,7 +170,7 @@ void time_irregular_astype_1D() {
 void time_irregular_astype_2D() {
  auto device = mx::default_device();
  int size = 2048;
-  std::vector<int> shape = {size, size};
+  mx::Shape shape = {size, size};

  auto a = mx::random::uniform(shape);
  TIMEM("2D regular", mx::astype, a, mx::int32, device);
@@ -1,6 +1,5 @@
 # Copyright © 2023 Apple Inc.

-import argparse
 import os
 import subprocess
 import time
@@ -38,10 +38,10 @@ def bench(f, *args):
    for i in range(10):
        f(*args)

-    s = time.time()
+    s = time.perf_counter()
    for i in range(100):
        f(*args)
-    e = time.time()
+    e = time.perf_counter()
    return e - s


@@ -37,10 +37,10 @@ def bench(f, *args):
    for i in range(10):
        f(*args)

-    s = time.time()
+    s = time.perf_counter()
    for i in range(100):
        f(*args)
-    e = time.time()
+    e = time.perf_counter()
    return e - s


@@ -0,0 +1,119 @@
+# Copyright © 2026 Apple Inc.
+
+import math
+import time
+
+import mlx.core as mx
+import numpy as np
+import torch
+
+N_WARMUP = 5
+N_BENCH = 20
+
+
+def bench_mlx(a, b):
+    for _ in range(N_WARMUP):
+        mx.eval(a @ b)
+
+    times = []
+    for _ in range(N_BENCH):
+        start = time.perf_counter_ns()
+        mx.eval(a @ b)
+        end = time.perf_counter_ns()
+        times.append((end - start) * 1e-9)
+
+    return np.mean(times), np.std(times)
+
+
+@torch.no_grad()
+def bench_torch(a, b):
+    for _ in range(N_WARMUP):
+        _ = a @ b
+        torch.mps.synchronize()
+
+    times = []
+    for _ in range(N_BENCH):
+        start = time.perf_counter_ns()
+        _ = a @ b
+        torch.mps.synchronize()
+        end = time.perf_counter_ns()
+        times.append((end - start) * 1e-9)
+
+    return np.mean(times), np.std(times)
+
+
+def check_correctness(out_mx, out_pt, rtol, M, N, K):
+    if not np.allclose(out_pt, out_mx, rtol=rtol, atol=0):
+        abs_diff = np.abs(out_pt - out_mx)
+        rel_diff = abs_diff / np.maximum(np.abs(out_pt), 1e-10)
+
+        print(
+            f"  WARNING: Correctness failed at {M}x{N}x{K}: "
+            f"max_abs={np.max(abs_diff):.6e}, max_rel={np.max(rel_diff):.6e}"
+        )
+
+
+def bench_gemm(M, N, K, dtype, rtol):
+    scale = 0.5 / math.sqrt(K)
+    a_np = np.random.uniform(0, scale, (M, K)).astype(np.float32)
+    b_np = np.random.uniform(0, scale, (K, N)).astype(np.float32)
+
+    a_mx = mx.array(a_np).astype(getattr(mx, dtype))
+    b_mx = mx.array(b_np).astype(getattr(mx, dtype))
+
+    a_pt = torch.from_numpy(a_np).to(dtype=getattr(torch, dtype), device="mps")
+    b_pt = torch.from_numpy(b_np).to(dtype=getattr(torch, dtype), device="mps")
+    torch.mps.synchronize()
+
+    torch_mean, torch_std = bench_torch(a_pt, b_pt)
+    mlx_mean, mlx_std = bench_mlx(a_mx, b_mx)
+
+    out_mx = (a_mx @ b_mx).astype(mx.float32)
+    out_pt = (a_pt @ b_pt).to(torch.float32).to("cpu").numpy(force=True)
+    check_correctness(out_mx, out_pt, rtol, M, N, K)
+
+    return mlx_mean, mlx_std, torch_mean, torch_std
+
+
+if __name__ == "__main__":
+    dtypes = ("bfloat16", "float16", "float32")
+
+    rtols = {
+        "float32": 1e-3,
+        "float16": 5e-3,
+        "bfloat16": 1e-2,
+    }
+
+    shapes = (
+        (2048, 2048, 10240),
+        (2048, 3072, 10240),
+        (3072, 3072, 10240),
+        (3072, 3072, 12288),
+        (3072, 4096, 12288),
+        (4096, 4096, 12288),
+        (4096, 4096, 18432),
+        (4096, 4096, 21504),
+        (4096, 6144, 21504),
+        (6144, 6144, 21504),
+    )
+
+    for dtype in dtypes:
+        print(f"\nPerformance ({dtype}):")
+        print(
+            f"{'M':>5s} {'N':>5s} {'K':>6s}  "
+            f"{'MLX (ms)':>15s}  {'Torch (ms)':>15s}  {'Speedup':>10s}"
+        )
+        print("-" * 80)
+
+        for M, N, K in shapes:
+            mlx_mean, mlx_std, torch_mean, torch_std = bench_gemm(
+                M, N, K, dtype, rtols[dtype]
+            )
+            speedup = torch_mean / mlx_mean
+
+            print(
+                f"{M:5d} {N:5d} {K:6d}  "
+                f"{mlx_mean*1000:7.2f}±{mlx_std*1000:5.2f}  "
+                f"{torch_mean*1000:7.2f}±{torch_std*1000:5.2f}  "
+                f"{speedup:8.2f}x"
+            )
@@ -0,0 +1,212 @@
+import math
+import os
+import subprocess
+import time
+from copy import copy
+from functools import partial
+
+import matplotlib.pyplot as plt
+import mlx.core as mx
+import numpy as np
+import torch
+from matplotlib.ticker import FuncFormatter
+
+RESULTS_DIR = "./results"
+
+
+if not os.path.isdir(RESULTS_DIR):
+    os.mkdir(RESULTS_DIR)
+
+DEVICE_NAME = subprocess.check_output(["sysctl", "-n", "machdep.cpu.brand_string"])
+DEVICE_NAME = DEVICE_NAME.decode("utf-8").strip("\n")
+
+TORCH_DEVICE = torch.device(
+    "mps"
+    if torch.backends.mps.is_available()
+    else ("cuda" if torch.cuda.is_available() else "cpu")
+)
+
+
+N_WARMUP = 5
+N_ITER_BENCH = 50
+N_ITER_FUNC = 20
+
+VECTOR_LENGTHS = [4096 * (2**i) for i in range(10)]
+MASK_DENSITIES = [0.01, 0.1, 0.25, 0.5]
+D_TYPES = ("float32", "float16")
+
+
+def _power_of_two_formatter(value, _position):
+    if value <= 0:
+        return ""
+    exponent = int(round(math.log2(value)))
+    if abs(value - (1 << exponent)) / value > 1e-6:
+        return f"{value:g}"
+    return f"$2^{{{exponent}}}$"
+
+
+def torch_sync():
+    if TORCH_DEVICE.type == "cuda":
+        torch.cuda.synchronize()
+    elif TORCH_DEVICE.type == "mps":
+        torch.mps.synchronize()
+
+
+def masked_scatter_mlx(self_arr, mask_arr, src_arr):
+    outs = []
+    for _ in range(N_ITER_FUNC):
+        out = copy(self_arr)
+        out[mask_arr] = src_arr
+        outs.append(out)
+    mx.eval(outs)
+    return outs
+
+
+@torch.no_grad()
+def masked_scatter_torch(self_tensor, mask_tensor, src_tensor):
+    outs = []
+    for _ in range(N_ITER_FUNC):
+        out = self_tensor.clone()
+        out.masked_scatter_(mask_tensor, src_tensor)
+        outs.append(out)
+    torch_sync()
+    return outs
+
+
+def measure(fn):
+    for _ in range(N_WARMUP):
+        fn()
+    start = time.perf_counter_ns()
+    for _ in range(N_ITER_BENCH):
+        fn()
+    end = time.perf_counter_ns()
+    return (end - start) * 1e-9
+
+
+def bytes_touched(length, true_count, item_size):
+    mask_bytes = length
+    self_bytes = length * item_size * 2  # read + write
+    src_bytes = true_count * item_size
+    return (mask_bytes + self_bytes + src_bytes) * N_ITER_FUNC * N_ITER_BENCH
+
+
+def build_case(length, density, np_dtype, torch_dtype):
+    true_count = max(1, int(round(length * density)))
+
+    rng = np.random.default_rng()
+    self_np = rng.normal(0.0, 1.0, length).astype(np_dtype)
+    mask_np = np.zeros(length, dtype=bool)
+    mask_np[:true_count] = True
+    rng.shuffle(mask_np)
+    src_np = rng.normal(0.0, 1.0, true_count).astype(np_dtype)
+
+    self_mlx = mx.array(self_np)
+    mask_mlx = mx.array(mask_np)
+    src_mlx = mx.array(src_np)
+
+    self_torch = torch.from_numpy(self_np).to(device=TORCH_DEVICE, dtype=torch_dtype)
+    mask_torch = torch.from_numpy(mask_np).to(device=TORCH_DEVICE)
+    src_torch = torch.from_numpy(src_np).to(device=TORCH_DEVICE, dtype=torch_dtype)
+
+    # Correctness check once per configuration
+    mx_out = mx.array(self_np)
+    mx_out[mask_mlx] = src_mlx
+    mx.eval(mx_out)
+    torch_out = self_torch.clone()
+    torch_out.masked_scatter_(mask_torch, src_torch)
+
+    atol = 5e-3 if np_dtype == np.float16 else 1e-5
+    if not np.allclose(np.array(mx_out), torch_out.cpu().numpy(), atol=atol):
+        raise AssertionError("masked_scatter results diverged between MLX and Torch")
+
+    return (self_mlx, mask_mlx, src_mlx, self_torch, mask_torch, src_torch, true_count)
+
+
+def bench_case(length, density, dtype):
+    np_dtype = getattr(np, dtype)
+    torch_dtype = getattr(torch, dtype)
+    (
+        self_mlx,
+        mask_mlx,
+        src_mlx,
+        self_torch,
+        mask_torch,
+        src_torch,
+        true_count,
+    ) = build_case(length, density, np_dtype, torch_dtype)
+
+    time_mlx = measure(partial(masked_scatter_mlx, self_mlx, mask_mlx, src_mlx))
+    time_torch = measure(
+        partial(masked_scatter_torch, self_torch, mask_torch, src_torch)
+    )
+
+    total_bytes = bytes_touched(length, true_count, np_dtype().itemsize)
+    bytes_per_gb = float(1024**3)
+    mlx_gbps = (total_bytes / bytes_per_gb) / time_mlx
+    torch_gbps = (total_bytes / bytes_per_gb) / time_torch
+
+    return time_mlx, time_torch, mlx_gbps, torch_gbps
+
+
+def plot_density(ax_perf, ax_speedup, density, dtype):
+    mlx_gbps = []
+    torch_gbps = []
+    mlx_times = []
+    torch_times = []
+
+    for length in VECTOR_LENGTHS:
+        t_mlx, t_torch, gbps_mlx, gbps_torch = bench_case(length, density, dtype)
+        mlx_gbps.append(gbps_mlx)
+        torch_gbps.append(gbps_torch)
+        mlx_times.append(t_mlx)
+        torch_times.append(t_torch)
+
+    ax_perf.plot(VECTOR_LENGTHS, mlx_gbps, "tab:blue", label="MLX")
+    ax_perf.plot(VECTOR_LENGTHS, torch_gbps, "tab:red", label="Torch")
+    ax_perf.set_xscale("log", base=2)
+    ax_perf.set_xticks(VECTOR_LENGTHS)
+    formatter = FuncFormatter(_power_of_two_formatter)
+    ax_perf.xaxis.set_major_formatter(formatter)
+    ax_perf.set_title(f"density={density:.2f}")
+    ax_perf.set_ylabel("GB/s")
+    ax_perf.grid(True, which="both", linestyle=":", alpha=0.4)
+    ax_perf.legend()
+
+    speedup = np.array(torch_times) / np.array(mlx_times)
+    ax_speedup.plot(VECTOR_LENGTHS, speedup, "tab:green")
+    ax_speedup.axhline(1.0, color="tab:gray", linestyle="--")
+    ax_speedup.set_xscale("log", base=2)
+    ax_speedup.set_xticks(VECTOR_LENGTHS)
+    ax_speedup.xaxis.set_major_formatter(formatter)
+    ax_speedup.set_ylabel("Speedup (Torch_t / MLX_t)")
+    ax_speedup.grid(True, which="both", linestyle=":", alpha=0.4)
+
+
+def main():
+    for dtype in D_TYPES:
+        fig, axs = plt.subplots(
+            len(MASK_DENSITIES),
+            2,
+            figsize=(10, 12),
+            layout="constrained",
+            sharex=True,
+        )
+
+        for i, density in enumerate(MASK_DENSITIES):
+            plot_density(axs[i][0], axs[i][1], density, dtype)
+            axs[i][0].set_xlabel("vector length")
+            axs[i][1].set_xlabel("vector length")
+
+        fig.suptitle(
+            f"{DEVICE_NAME.replace('Apple ', '')} ({TORCH_DEVICE.type}) | dtype={dtype}"
+        )
+        output_path = os.path.join(
+            RESULTS_DIR,
+            f"{DEVICE_NAME.replace(' ', '_')}_masked_scatter_{dtype}.pdf",
+        )
+        fig.savefig(output_path)
+        plt.close(fig)
+
+
+if __name__ == "__main__":
+    main()
@@ -31,8 +31,8 @@ def measure_runtime(fn, **kwargs):
    for _ in range(5):
        fn(**kwargs)

-    tic = time.time()
+    tic = time.perf_counter()
    iters = 100
    for _ in range(iters):
        fn(**kwargs)
-    return (time.time() - tic) * 1000 / iters
+    return (time.perf_counter() - tic) * 1000 / iters
@@ -0,0 +1,177 @@
+# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+# Modified from
+# https://github.com/NVIDIA/cudnn-frontend/blob/main/cmake/cuDNN.cmake
+
+# Return the last file matching the pattern.
+function(find_file_glob VAR PATTERN)
+  file(GLOB _RESULT "${PATTERN}")
+  if(_RESULT)
+    list(LENGTH ${_RESULT} _RESULT_LENGTH)
+    if(_RESULT_LENGTH GREATER 0)
+      list(GET ${_RESULT} -1 _RESULT)
+    endif()
+    set(${VAR}
+        "${_RESULT}"
+        PARENT_SCOPE)
+  endif()
+endfunction()
+
+# Find the dir including the "cudnn.h" file.
+find_path(
+  CUDNN_INCLUDE_DIR cudnn.h
+  HINTS ${CUDNN_INCLUDE_PATH} ${CUDAToolkit_INCLUDE_DIRS}
+  PATH_SUFFIXES include OPTIONAL)
+
+# Glob searching "cudnn.h" for Windows.
+if(WIN32 AND NOT CUDNN_INCLUDE_DIR)
+  find_file_glob(
+    CUDNN_H_PATH
+    "C:/Program Files/NVIDIA/CUDNN/*/include/${CUDAToolkit_VERSION_MAJOR}.*/cudnn.h"
+  )
+  if(CUDNN_H_PATH)
+    get_filename_component(CUDNN_INCLUDE_DIR "${CUDNN_H_PATH}" DIRECTORY)
+  endif()
+endif()
+
+if(NOT CUDNN_INCLUDE_DIR)
+  message(
+    FATAL_ERROR
+      "Unable to find cudnn.h, please make sure cuDNN is installed and pass CUDNN_INCLUDE_PATH to cmake."
+  )
+endif()
+
+# Get cudnn version.
+file(READ "${CUDNN_INCLUDE_DIR}/cudnn_version.h" cudnn_version_header)
+string(REGEX MATCH "#define CUDNN_MAJOR [1-9]+" macrodef
+             "${cudnn_version_header}")
+string(REGEX MATCH "[1-9]+" CUDNN_MAJOR_VERSION "${macrodef}")
+
+# Function for searching library files.
+function(find_cudnn_library NAME)
+  if(NOT "${ARGV1}" STREQUAL "OPTIONAL")
+    set(_CUDNN_REQUIRED TRUE)
+  else()
+    set(_CUDNN_REQUIRED FALSE)
+  endif()
+
+  find_library(
+    ${NAME}_LIBRARY
+    NAMES ${NAME} "lib${NAME}.so.${CUDNN_MAJOR_VERSION}" NAMES_PER_DIR
+    HINTS ${CUDNN_LIBRARY_PATH} ${CUDAToolkit_LIBRARY_DIR}
+    PATH_SUFFIXES lib64 lib/x64 lib OPTIONAL)
+
+  if(WIN32 AND NOT ${NAME}_LIBRARY)
+    find_file_glob(
+      ${NAME}_LIBRARY
+      "C:/Program Files/NVIDIA/CUDNN/*/lib/${CUDAToolkit_VERSION_MAJOR}.*/x64/${NAME}.lib"
+    )
+  endif()
+
+  if(NOT ${NAME}_LIBRARY AND ${_CUDNN_REQUIRED})
+    message(
+      FATAL_ERROR
+        "Unable to find ${NAME}, please make sure cuDNN is installed and pass CUDNN_LIBRARY_PATH to cmake."
+    )
+  endif()
+
+  if(${NAME}_LIBRARY)
+    add_library(CUDNN::${NAME} UNKNOWN IMPORTED)
+    set_target_properties(
+      CUDNN::${NAME}
+      PROPERTIES INTERFACE_INCLUDE_DIRECTORIES ${CUDNN_INCLUDE_DIR}
+                 IMPORTED_LOCATION ${${NAME}_LIBRARY})
+    set(${NAME}_LIBRARY
+        "${${NAME}_LIBRARY}"
+        PARENT_SCOPE)
+  else()
+    message(STATUS "${NAME} not found.")
+  endif()
+endfunction()
+
+# Search for the main cudnn library.
+find_cudnn_library(cudnn)
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(CUDNN REQUIRED_VARS CUDNN_INCLUDE_DIR
+                                                      cudnn_LIBRARY)
+
+if(CUDNN_INCLUDE_DIR AND cudnn_LIBRARY)
+  set(CUDNN_FOUND
+      ON
+      CACHE INTERNAL "cuDNN Library Found")
+else()
+  set(CUDNN_FOUND
+      OFF
+      CACHE INTERNAL "cuDNN Library Not Found")
+endif()
+
+# Find out all the DLL files for Windows.
+if(WIN32 AND cudnn_LIBRARY)
+  get_filename_component(CUDNN_BIN_DIR "${cudnn_LIBRARY}" DIRECTORY)
+  string(REPLACE "/lib/" "/bin/" CUDNN_BIN_DIR "${CUDNN_BIN_DIR}")
+  file(
+    GLOB CUDNN_DLL_NAMES
+    RELATIVE "${CUDNN_BIN_DIR}"
+    "${CUDNN_BIN_DIR}/*.dll")
+endif()
+
+# Create an interface library that users can link with.
+add_library(CUDNN::cudnn_all INTERFACE IMPORTED)
+target_link_libraries(CUDNN::cudnn_all INTERFACE CUDNN::cudnn)
+target_include_directories(
+  CUDNN::cudnn_all INTERFACE $<INSTALL_INTERFACE:include>
+                             $<BUILD_INTERFACE:${CUDNN_INCLUDE_DIR}>)
+
+# Add other components of cudnn.
+if(CUDNN_MAJOR_VERSION EQUAL 8)
+  find_cudnn_library(cudnn_adv_infer)
+  find_cudnn_library(cudnn_adv_train)
+  find_cudnn_library(cudnn_cnn_infer)
+  find_cudnn_library(cudnn_cnn_train)
+  find_cudnn_library(cudnn_ops_infer)
+  find_cudnn_library(cudnn_ops_train)
+
+  target_link_libraries(
+    CUDNN::cudnn_all
+    INTERFACE CUDNN::cudnn_adv_train CUDNN::cudnn_ops_train
+              CUDNN::cudnn_cnn_train CUDNN::cudnn_adv_infer
+              CUDNN::cudnn_cnn_infer CUDNN::cudnn_ops_infer)
+
+elseif(CUDNN_MAJOR_VERSION EQUAL 9)
+  find_cudnn_library(cudnn_graph)
+  find_cudnn_library(cudnn_engines_runtime_compiled)
+  find_cudnn_library(cudnn_ops OPTIONAL)
+  find_cudnn_library(cudnn_cnn OPTIONAL)
+  find_cudnn_library(cudnn_adv OPTIONAL)
+  find_cudnn_library(cudnn_engines_precompiled OPTIONAL)
+  find_cudnn_library(cudnn_heuristic OPTIONAL)
+
+  target_link_libraries(
+    CUDNN::cudnn_all
+    INTERFACE CUDNN::cudnn_graph
+              CUDNN::cudnn_engines_runtime_compiled
+              CUDNN::cudnn_ops
+              CUDNN::cudnn_cnn
+              CUDNN::cudnn_adv
+              CUDNN::cudnn_engines_precompiled
+              CUDNN::cudnn_heuristic)
+endif()
@@ -0,0 +1,3 @@
+# This file does nothing but to suppress the cmake warning: "By not providing
+# Findnvpl.cmake in CMAKE_MODULE_PATH...", which is caused by the
+# find_package(nvpl) from cmake's builtin FindLAPACK.cmake module.
@@ -26,6 +26,7 @@ ENABLE_PREPROCESSING   = YES
 MACRO_EXPANSION        = YES
 EXPAND_ONLY_PREDEF     = NO
 SKIP_FUNCTION_MACROS   = NO
+PREDEFINED             = MLX_API=

 ################################################################################
 # Compound extraction control.                                                 #
@@ -38,3 +38,17 @@ the docs. Then force add the `build/html` directory:
 `git add -f build/html`

 Commit and push the changes to the `gh-pages` branch.
+
+## Doc Development Setup
+
+To enable live refresh of docs while writing:
+
+Install sphinx autobuild
+```
+pip install sphinx-autobuild
+```
+
+Run auto build on docs/src folder
+```
+sphinx-autobuild ./src ./build/html
+```
@@ -777,11 +777,11 @@ with the naive :meth:`simple_axpby` we first defined.
            mx.eval(z)

        # Timed run
-        s = time.time()
+        s = time.perf_counter()
        for i in range(100):
            z = f(x, y, alpha, beta)
            mx.eval(z)
-        e = time.time()
+        e = time.perf_counter()
        return 1000 * (e - s) / 100

    simple_time = bench(simple_axpby)
@@ -0,0 +1,40 @@
+Metal Logging
+=============
+
+In debug builds, MLX compiles Metal kernels with ``os_log`` enabled so shader
+warnings and debug messages are visible during development.
+
+.. note::
+    Metal logging is only available with Metal 3.2 or higher (macOS 15 and up,
+    iOS 18 and up).
+
+To enable logging from kernels, first make sure to build in debug mode:
+
+.. code-block:: bash
+
+    DEBUG=1 python -m pip install -e .
+
+Then, in the kernel source code include MLX's logging shim and use
+``mlx::os_log``:
+
+.. code-block::
+
+    #include "mlx/backend/metal/kernels/logging.h"
+
+    constant mlx::os_log logger("mlx", "my_kernel");
+
+    kernel void my_kernel(/* ... */) {
+    // ...
+      logger.log_debug("unexpected state: idx=%u", idx);
+    }
+
+When you run the program, set the Metal log level to your desired level and
+forward logs to ``stderr``:
+
+.. code-block:: bash
+
+    MTL_LOG_LEVEL=MTLLogLevelDebug MTL_LOG_TO_STDERR=1 python script.py
+
+See the `Metal logging guide`_ for more details.
+
+.. _`Metal logging guide`: https://developer.apple.com/documentation/metal/logging-shader-debug-messages
@@ -45,7 +45,7 @@ The next step is to setup a CMake file in ``CMakeLists.txt``:

  project(example LANGUAGES CXX)

-  set(CMAKE_CXX_STANDARD 17)
+  set(CMAKE_CXX_STANDARD 20)
  set(CMAKE_CXX_STANDARD_REQUIRED ON)


@@ -0,0 +1,91 @@
+.. _data_parallelism:
+
+Data Parallelism
+================
+
+MLX enables efficient data parallel distributed training through its
+distributed communication primitives.
+
+.. _training_example:
+
+Training Example
+----------------
+
+In this section we will adapt an MLX training loop to support data parallel
+distributed training. Namely, we will average the gradients across a set of
+hosts before applying them to the model.
+
+Our training loop looks like the following code snippet if we omit the model,
+dataset, and optimizer initialization.
+
+.. code:: python
+
+    model = ...
+    optimizer = ...
+    dataset = ...
+
+    def step(model, x, y):
+        loss, grads = loss_grad_fn(model, x, y)
+        optimizer.update(model, grads)
+        return loss
+
+    for x, y in dataset:
+        loss = step(model, x, y)
+        mx.eval(loss, model.parameters())
+
+All we have to do to average the gradients across machines is perform an
+:func:`all_sum` and divide by the size of the :class:`Group`. Namely we
+have to :func:`mlx.utils.tree_map` the gradients with following function.
+
+.. code:: python
+
+    def all_avg(x):
+        return mx.distributed.all_sum(x) / mx.distributed.init().size()
+
+Putting everything together our training loop step looks as follows with
+everything else remaining the same.
+
+.. code:: python
+
+    from mlx.utils import tree_map
+
+    def all_reduce_grads(grads):
+        N = mx.distributed.init().size()
+        if N == 1:
+            return grads
+        return tree_map(
+            lambda x: mx.distributed.all_sum(x) / N,
+            grads
+        )
+
+    def step(model, x, y):
+        loss, grads = loss_grad_fn(model, x, y)
+        grads = all_reduce_grads(grads)  # <--- This line was added
+        optimizer.update(model, grads)
+        return loss
+
+Using ``nn.average_gradients``
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Although the code example above works correctly; it performs one communication
+per gradient. It is significantly more efficient to aggregate several gradients
+together and perform fewer communication steps.
+
+This is the purpose of :func:`mlx.nn.average_gradients`. The final code looks
+almost identical to the example above:
+
+.. code:: python
+
+    model = ...
+    optimizer = ...
+    dataset = ...
+
+    def step(model, x, y):
+        loss, grads = loss_grad_fn(model, x, y)
+        grads = mx.nn.average_gradients(grads)  # <---- This line was added
+        optimizer.update(model, grads)
+        return loss
+
+    for x, y in dataset:
+        loss = step(model, x, y)
+        mx.eval(loss, model.parameters())
@@ -0,0 +1,239 @@
+.. _tensor_parallelism:
+
+Tensor Parallelism
+==================
+
+In this example, we will explore how tensor parallelism (TP) works in MLX.  We
+will start with an overview of the distributed layers in ``mlx.nn`` and then
+show how to do tensor parallelism Llama-style transformer models.
+
+Sharded Layers
+--------------
+
+:class:`AllToShardedLinear <mlx.nn.AllToShardedLinear>`
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+This layer replicates a common input and shards the weight matrix along the
+output dimension across all devices in the :class:`mlx.core.distributed.Group`.
+The layer produces a sharded output.
+
+For example, consider an :class:`mlx.nn.AllToShardedLinear` layer with
+``input_dims=2`` and ``output_dims=2``, a batched input of shape ``(4, 2)``,
+and a device group with 2 devices. The layer shards the weight matrix along the
+output dimension across the two devices, where each device receives the full
+input and computes a partial output.
+
+.. raw:: html
+
+    <div>
+      <img src="../_static/tp_inference/all-to-sharded-linear.png" alt="column-wise tensor parallelism" style="width: 100%">
+    </div>
+
+This layer does not automatically gather all outputs from each device. This is
+an intended and :ref:`useful design choice <useful_design_choices>`.
+
+:class:`QuantizedAllToShardedLinear <mlx.nn.QuantizedAllToShardedLinear>` is
+the quantized equivalent of :class:`mlx.nn.AllToShardedLinear`.  Similar to
+:class:`mlx.nn.QuantizedLinear`, its parameters are frozen and will not be
+included in any gradient computation.
+
+
+:class:`ShardedToAllLinear <mlx.nn.ShardedToAllLinear>`
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+This layer expects inputs that are sharded along the feature dimension and
+shards the weight matrix along the input dimension across all devices in the
+:class:`mlx.core.distributed.Group`. The layer automatically aggregates the
+results using :class:`mlx.core.distributed.all_sum`, so all devices in the
+group will have the same result.
+
+For example, consider an :class:`mlx.nn.ShardedToAllLinear` layer with
+``input_dims=2`` and ``output_dims=2``, a batched input of shape ``(4, 2)``,
+and a device group with 2 devices. The layer shards the weight matrix along the
+input dimension across the two devices. Each device computes a ``(4,2)``
+output, which is then aggregated with all other device outputs to get layer
+output.
+
+   .. raw:: html
+
+    <div>
+      <img src="../_static/tp_inference/sharded-to-all-linear.png" alt="row-wise tensor parallelism" style="width: 100%">
+    </div>
+
+This layer does not automatically shard the inputs along the feature dimension
+for you. It is necessary to create a "partial" input structure to feed into the
+layer. This is an intended and :ref:`useful design choice
+<useful_design_choices>`.
+
+:class:`QuantizedShardedToAllLinear <mlx.nn.QuantizedShardedToAllLinear>` is
+the quantized equivalent of :class:`mlx.nn.ShardedToAllLinear`.  Similar to
+:class:`mlx.nn.QuantizedLinear`, its parameters are frozen and will not be
+included in any gradient computation.
+
+
+Shard Utility Functions
+-----------------------
+
+:func:`shard_linear <mlx.nn.layers.distributed.shard_linear>`
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Converts a regular linear layer into a tensor parallel layer that distributes
+computation across multiple devices. Takes an existing :class:`mlx.nn.Linear`
+or :class:`mlx.nn.QuantizedLinear` layer and returns a new distributed layer
+(either :class:`mlx.nn.AllToShardedLinear` or
+:class:`mlx.nn.ShardedToAllLinear`, depending on the sharding type). The
+original layer is not modified.
+
+:func:`shard_inplace <mlx.nn.layers.distributed.shard_inplace>`
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Splits the parameters of an existing layer across multiple devices by modifying
+the layer in-place. Unlike :func:`shard_linear
+<mlx.nn.layers.distributed.shard_linear>`, this function does not create a new
+layer or add distributed communication. The layer itself must handle
+distributed communication if needed.
+
+
+.. _useful_design_choices:
+
+Useful Design Choices
+---------------------
+
+The design choices above regarding when operations are done automatically are intentional and make model training and inference easier.
+
+All-to-sharded and sharded-to-all layers naturally go together because the
+output of the former layer is exactly the input needed needed for the latter.
+This removes the need for an intermediate gather step between the layers,
+reducing communication overhead.
+
+This is why :class:`mlx.nn.AllToShardedLinear` does not aggregate results
+automatically and why :class:`mlx.nn.ShardedToAllLinear` does not shard inputs
+automatically. It is so that they can be placed in successive order and work
+together easily.
+
+We can demonstrate this through a simple model using our two types of
+distributed layers.
+
+.. code-block:: python
+
+  x = ... # some (4, 2) model input: batch size 4, feature size 2
+
+  l1 = nn.AllToShardedLinear(2, 2, bias=False)   # initialize the layer
+  l1_out = l1(x) # (4, 1) output
+
+  l2 = nn.ShardedToAllLinear(2, 2, bias=False)
+  l2_out = l2(l1_out) # (4, 2) output
+
+.. raw:: html
+
+    <div>
+      <img src="../_static/tp_inference/column-row-tp.png" alt="two layer tensor parallelism" style="width: 100%">
+      <p style="font-size: 0.85em; margin-top: 0.5em;"><small>A visualization of the simple MLX model using all-to-sharded then sharded-to-all tensor parallelism across 2 devices.</small></p>
+    </div>
+
+
+LLM Inference with Tensor Parallelism
+-------------------------------------
+
+We can apply these TP techniques to LLMs in order to enable inference for much
+larger models by sharding parameters from huge layers across multiple devices.
+
+To demonstrate this, let's apply TP to the Transformer block of our :doc:`Llama
+Inference <llama-inference>` example. In this example, we will use the same
+inference script as the Llama Inference example, which can be found in
+`mlx-examples`_.
+
+Our first edit is to initialize the distributed communication group and get the
+current process rank:
+
+.. code-block:: python
+
+  world = mx.distributed.init()
+  rank = world.rank()
+
+Next, let's look at the current architecture of the transformer block and see how we can apply tensor parallelism:
+
+.. raw:: html
+
+    <div>
+      <img src="../_static/tp_inference/llama-transformer.png" alt="llama transformer example" style="width: 100%">
+    </div>
+
+
+This architecture has two natural places where 
+tensor parallelism can be applied: the attention block and the FFN
+block. Both follow the same pattern: multiple parallel linear layers operating
+on the same input, followed by a single output linear layer. In the attention
+block, the Q, K, and V projections are sharded along the output dimension (all-to-sharded), and the output
+projection is sharded along the input dimension (sharded-to-all). Similarly in the FFN block, the gate and up projections
+become all-to-sharded layers, and the down projection becomes an sharded-to-all layer.
+
+The intermediate operations between the linear layers (RoPE, softmax, scaled
+dot-product attention in the attention block, and element-wise multiplication
+in the FFN block) do not impede the use of our TP paradigm. These operations
+are either:
+
+- **Element-wise operations** (RoPE, element-wise multiplication): These
+  operate independently on each element or position, preserving the sharding
+  pattern without requiring cross-device communication.
+
+- **Operations on non-sharded dimensions** (softmax, scaled dot-product
+  attention): These operate along dimensions that are not sharded (such as the
+  sequence length or head dimensions), so they can be computed independently on
+  each device. The attention computation ``Q @ K^T`` and ``scores @ V`` work
+  correctly with sharded Q, K, V tensors because the matrix multiplications are
+  performed along the sharded feature dimension, and the results remain
+  properly sharded for the subsequent sharded-to-all layer.
+
+To implement sharding in our Llama inference, we use :func:`shard_linear
+<mlx.nn.layers.distributed.shard_linear>` to get sharded linear layers with
+distributed communication. This is easier than using :func:`shard_inplace
+<mlx.nn.layers.distributed.shard_inplace>` and implementing the steps manually
+in the :code:`__call__` function.
+
+The following code shows how to shard the Attention block. The Q, K, and V
+projection layers are converted to all-to-sharded layers, while the output
+projection is converted to a sharded-to-all layer. The number of heads are also
+adjusted to account for the sharding:
+
+.. code-block:: python
+
+  # ... in Attention class
+  def shard(self, group: mx.distributed.Group):
+    self.n_heads = self.n_heads // group.size()
+    self.n_kv_heads = self.n_kv_heads // group.size()
+
+    self.wq = nn.layers.distributed.shard_linear(self.wq, "all-to-sharded", group=group)
+    self.wk = nn.layers.distributed.shard_linear(self.wk, "all-to-sharded", group=group)
+    self.wv = nn.layers.distributed.shard_linear(self.wv, "all-to-sharded", group=group)
+    self.wo = nn.layers.distributed.shard_linear(self.wo, "sharded-to-all", group=group)
+
+Similarly, the FeedForward block is sharded by converting the gate (w1) and up
+(w3) projections to all-to-sharded layers, and the down projection (w2) to
+a sharded-to-all layer:
+
+.. code-block:: python
+
+  # ... in FeedForward class
+  def shard(self, group: mx.distributed.Group):
+    self.w1 = nn.layers.distributed.shard_linear(self.w1, "all-to-sharded", group=group)
+    self.w2 = nn.layers.distributed.shard_linear(self.w2, "sharded-to-all", group=group)
+    self.w3 = nn.layers.distributed.shard_linear(self.w3, "all-to-sharded", group=group)
+
+Finally, in our :code:`load_model` function, we need to apply our sharding
+functions to all transformer layers when using multiple devices:
+
+.. code-block:: python
+
+  # ... in load_model function
+  if world.size() > 1:
+    # convert Linear layers in Transformer/FFN to appropriate Sharded Layers
+    for layer in model.layers:
+        layer.attention.shard(group=world)
+        layer.feed_forward.shard(group=world)
+
+This allows us to use the llama inference file as normal when running
+:code:`python llama.py`, but now we can also run it across two (or more)
+devices via :code:`mlx.launch -n 2 llama.py`.
+
+.. _mlx-examples: https://github.com/ml-explore/mlx-examples/tree/main/llms/llama
@@ -54,6 +54,8 @@ are the CPU and GPU.
   examples/linear_regression
   examples/mlp
   examples/llama-inference
+   examples/data_parallelism
+   examples/tensor_parallelism

 .. toctree::
   :caption: Python API Reference
@@ -89,5 +91,6 @@ are the CPU and GPU.

   dev/extensions
   dev/metal_debugger
+   dev/metal_logging
   dev/custom_metal_kernels
   dev/mlx_in_cpp
@@ -17,11 +17,10 @@ To install from PyPI your system must meet the following requirements:

 - Using an M series chip (Apple silicon)
 - Using a native Python >= 3.10
- macOS >= 13.5
+- macOS >= 14.0

 .. note::
-    MLX is only available on devices running macOS >= 13.5
-    It is highly recommended to use macOS 14 (Sonoma)
+    MLX is only available on devices running macOS >= 14.0 and higher.

 CUDA
 ^^^^
@@ -30,17 +29,20 @@ MLX has a CUDA backend which you can install with:

 .. code-block:: shell

-    pip install mlx[cuda]
+    pip install mlx[cuda12]
+

 To install the CUDA package from PyPi your system must meet the following
 requirements:

- Nvidia architecture >= SM 7.0 (Volta)
+- Nvidia architecture >= SM 7.5
 - Nvidia driver >= 550.54.14
 - CUDA toolkit >= 12.0
 - Linux distribution with glibc >= 2.35
 - Python >= 3.10

+For CUDA 13 use ``pip install mlx[cuda13]``. The CUDA 13 package requires
+an Nvidia driver >= 580 or an appropriate CUDA compatibility package.

 CPU-only (Linux)
 ^^^^^^^^^^^^^^^^
@@ -81,7 +83,7 @@ Build from source
 Build Requirements
 ^^^^^^^^^^^^^^^^^^

- A C++ compiler with C++17 support (e.g. Clang >= 5.0)
+- A C++ compiler with C++20 support (e.g. Clang >= 15.0)
 - `cmake <https://cmake.org/>`_ -- version 3.25 or later, and ``make``
 - Xcode >= 15.0 and macOS SDK >= 14.0

@@ -126,13 +128,6 @@ Run the tests with:

  python -m unittest discover python/tests

-Optional: Install stubs to enable auto completions and type checking from your
-IDE:
-
-.. code-block:: shell
-
-  python setup.py generate_stubs
-
 C++ API
 ^^^^^^^

@@ -52,7 +52,7 @@ The default floating point type is ``float32`` and the default integer type is
     - 4 
     - 32-bit float
   * - ``float64``
-     - 4 
+     - 8
     - 64-bit double
   * - ``complex64``
     - 8 
@@ -17,3 +17,5 @@ Devices and Streams
   set_default_stream
   stream
   synchronize
+   device_count
+   device_info
@@ -183,3 +183,4 @@ In detail:
   nn/functions
   nn/losses
   nn/init
+   nn/distributed
@@ -0,0 +1,30 @@
+.. _nn_distributed:
+
+Distributed
+-----------
+
+Helper Routines
+^^^^^^^^^^^^^^^
+
+The :code:`mlx.nn.layers.distributed` package contains helpful routines to 
+create sharded layers from existing :class:`Modules <mlx.nn.Module>`.
+
+.. currentmodule:: mlx.nn.layers.distributed
+.. autosummary::
+   :toctree: _autosummary
+
+   shard_linear
+   shard_inplace
+
+Layers
+^^^^^^
+
+.. currentmodule:: mlx.nn
+.. autosummary::
+   :toctree: _autosummary
+   :template: nn-module-template.rst
+
+   AllToShardedLinear
+   ShardedToAllLinear
+   QuantizedAllToShardedLinear
+   QuantizedShardedToAllLinear
@@ -10,6 +10,7 @@ Layers
   :template: nn-module-template.rst

   ALiBi
+   AllToShardedLinear
   AvgPool1d
   AvgPool2d
   AvgPool3d
@@ -46,8 +47,10 @@ Layers
   Mish
   MultiHeadAttention
   PReLU
+   QuantizedAllToShardedLinear
   QuantizedEmbedding
   QuantizedLinear
+   QuantizedShardedToAllLinear
   RMSNorm
   ReLU
   ReLU2
@@ -56,6 +59,7 @@ Layers
   RoPE
   SELU
   Sequential
+   ShardedToAllLinear
   Sigmoid
   SiLU
   SinusoidalPositionalEncoding
@@ -11,6 +11,7 @@ Transforms
   eval
   async_eval
   compile
+   checkpoint
   custom_function
   disable_compile
   enable_compile
@@ -257,7 +257,26 @@ constants. For example:

 In order to have the change of state reflected in the outputs of ``fun`` you
 again have two options. The first option is to simply pass ``state`` as input
-to the function. In some cases this can be pretty inconvenient. Hence,
+to the function.
+
+.. code-block:: python
+
+  state = [mx.array(1.0)]
+
+  @mx.compile
+  def fun(x, state):
+      return x + state[0]
+
+  # Prints array(2, dtype=float32)
+  print(fun(mx.array(1.0), state))
+
+  # Update state
+  state[0] = mx.array(5.0)
+
+  # Prints array(6, dtype=float32)
+  print(fun(mx.array(1.0), state))
+
+In some cases this can be pretty inconvenient. Hence,
 :func:`compile` also has a parameter to capture implicit inputs:

 .. code-block:: python
@@ -7,21 +7,29 @@ Distributed Communication

 MLX supports distributed communication operations that allow the computational cost
 of training or inference to be shared across many physical machines. At the
-moment we support two different communication backends:
+moment we support several different communication backends introduced below.
+
+.. list-table::
+   :widths: 20 80
+   :header-rows: 1
+
+   * - Backend
+     - Description
+   * - :ref:`MPI <mpi_section>`
+     - A full featured and mature distributed communications library.
+   * - :ref:`RING <ring_section>`
+     - Ring all reduce and all gather over TCP sockets. Always available and
+       usually faster than MPI.
+   * - :ref:`JACCL <jaccl_section>`
+     - Low latency communication with RDMA over thunderbolt. Necessary for
+       things like tensor parallelism.
+   * - :ref:`NCCL <nccl_section>`
+     - The backend of choice for CUDA environments.

-* `MPI <https://en.wikipedia.org/wiki/Message_Passing_Interface>`_ a
-  full-featured and mature distributed communications library
-* A **ring** backend of our own that uses native TCP sockets and should be
-  faster for thunderbolt connections.

 The list of all currently supported operations and their documentation can be
 seen in the :ref:`API docs<distributed>`.

-.. note::
-   Some operations may not be supported or not as fast as they should be.
-   We are adding more and tuning the ones we have as we are figuring out the
-   best way to do distributed computing on Macs using MLX.
-
 Getting Started
 ---------------

@@ -84,9 +92,8 @@ Selecting Backend
 ^^^^^^^^^^^^^^^^^

 You can select the backend you want to use when calling :func:`init` by passing
-one of ``{'any', 'ring', 'mpi'}``. When passing ``any``, MLX will try to
-initialize the ``ring`` backend and if it fails the ``mpi`` backend. If they
-both fail then a singleton group is created.
+one of ``{'any', 'ring', 'jaccl', 'mpi', 'nccl'}``. When passing ``any``, MLX will try all
+available backends. If they all fail then a singleton group is created.

 .. note::
   After a distributed backend is successfully initialized :func:`init` will
@@ -110,162 +117,13 @@ The following examples aim to clarify the backend initialization logic in MLX:
    world_ring = mx.distributed.init(backend="ring")
    world_any = mx.distributed.init()  # same as MPI because it was initialized first!

-Training Example
----------------
+Distributed Program Examples
+----------------------------

-In this section we will adapt an MLX training loop to support data parallel
-distributed training. Namely, we will average the gradients across a set of
-hosts before applying them to the model.
+- :ref:`Data Parallelism <data_parallelism>`
+- :ref:`Tensor Parallelism <tensor_parallelism>`

-Our training loop looks like the following code snippet if we omit the model,
-dataset and optimizer initialization.
-
-.. code:: python
-
-    model = ...
-    optimizer = ...
-    dataset = ...
-
-    def step(model, x, y):
-        loss, grads = loss_grad_fn(model, x, y)
-        optimizer.update(model, grads)
-        return loss
-
-    for x, y in dataset:
-        loss = step(model, x, y)
-        mx.eval(loss, model.parameters())
-
-All we have to do to average the gradients across machines is perform an
-:func:`all_sum` and divide by the size of the :class:`Group`. Namely we
-have to :func:`mlx.utils.tree_map` the gradients with following function.
-
-.. code:: python
-
-    def all_avg(x):
-        return mx.distributed.all_sum(x) / mx.distributed.init().size()
-
-Putting everything together our training loop step looks as follows with
-everything else remaining the same.
-
-.. code:: python
-
-    from mlx.utils import tree_map
-
-    def all_reduce_grads(grads):
-        N = mx.distributed.init().size()
-        if N == 1:
-            return grads
-        return tree_map(
-            lambda x: mx.distributed.all_sum(x) / N,
-            grads
-        )
-
-    def step(model, x, y):
-        loss, grads = loss_grad_fn(model, x, y)
-        grads = all_reduce_grads(grads)  # <--- This line was added
-        optimizer.update(model, grads)
-        return loss
-
-Utilizing ``nn.average_gradients``
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-Although the code example above works correctly; it performs one communication
-per gradient. It is significantly more efficient to aggregate several gradients
-together and perform fewer communication steps.
-
-This is the purpose of :func:`mlx.nn.average_gradients`. The final code looks
-almost identical to the example above:
-
-.. code:: python
-
-    model = ...
-    optimizer = ...
-    dataset = ...
-
-    def step(model, x, y):
-        loss, grads = loss_grad_fn(model, x, y)
-        grads = mx.nn.average_gradients(grads)  # <---- This line was added
-        optimizer.update(model, grads)
-        return loss
-
-    for x, y in dataset:
-        loss = step(model, x, y)
-        mx.eval(loss, model.parameters())
-
-
-Getting Started with MPI
------------------------
-
-MLX already comes with the ability to "talk" to MPI if it is installed on the
-machine. Launching distributed MLX programs that use MPI can be done with
-``mpirun`` as expected. However, in the following examples we will be using
-``mlx.launch --backend mpi`` which takes care of some nuisances such as setting
-absolute paths for the ``mpirun`` executable and the ``libmpi.dyld`` shared
-library.
-
-The simplest possible usage is the following which, assuming the minimal
-example in the beginning of this page, should result in:
-
-.. code:: shell
-
-    $ mlx.launch --backend mpi -n 2 test.py
-    1 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
-    0 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
-
-The above launches two processes on the same (local) machine and we can see
-both standard output streams. The processes send the array of 1s to each other
-and compute the sum which is printed. Launching with ``mlx.launch -n 4 ...`` would
-print 4 etc.
-
-Installing MPI
-^^^^^^^^^^^^^^
-
-MPI can be installed with Homebrew, using the Anaconda package manager or
-compiled from source. Most of our testing is done using ``openmpi`` installed
-with the Anaconda package manager as follows:
-
-.. code:: shell
-
-    $ conda install conda-forge::openmpi
-
-Installing with Homebrew may require specifying the location of ``libmpi.dyld``
-so that MLX can find it and load it at runtime. This can simply be achieved by
-passing the ``DYLD_LIBRARY_PATH`` environment variable to ``mpirun`` and it is
-done automatically by ``mlx.launch``.
-
-.. code:: shell
-
-    $ mpirun -np 2 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python test.py
-    $ # or simply
-    $ mlx.launch -n 2 test.py
-
-Setting up Remote Hosts
-^^^^^^^^^^^^^^^^^^^^^^^
-
-MPI can automatically connect to remote hosts and set up the communication over
-the network if the remote hosts can be accessed via ssh. A good checklist to
-debug connectivity issues is the following:
-
-* ``ssh hostname`` works from all machines to all machines without asking for
-  password or host confirmation
-* ``mpirun`` is accessible on all machines.
-* Ensure that the ``hostname`` used by MPI is the one that you have configured
-  in the ``.ssh/config`` files on all machines.
-
-Tuning MPI All Reduce
-^^^^^^^^^^^^^^^^^^^^^
-
-.. note::
-
-    For faster all reduce consider using the ring backend either with Thunderbolt
-    connections or over Ethernet.
-
-Configure MPI to use N tcp connections between each host to improve bandwidth
-by passing ``--mca btl_tcp_links N``.
-
-Force MPI to use the most performant network interface by setting ``--mca
-btl_tcp_if_include <iface>`` where ``<iface>`` should be the interface you want
-to use.
+.. _ring_section:

 Getting Started with Ring
 -------------------------
@@ -275,7 +133,7 @@ available. It uses TCP sockets so the nodes need to be reachable via a network.
 As the name suggests the nodes are connected in a ring which means that rank 1
 can only communicate with rank 0 and rank 2, rank 2 only with rank 1 and rank 3
 and so on and so forth. As a result :func:`send` and :func:`recv` with
-arbitrary sender and receiver is not supported in the ring backend.
+arbitrary sender and receiver are not supported in the ring backend.

 Defining a Ring
 ^^^^^^^^^^^^^^^
@@ -316,22 +174,13 @@ utility as follows:

 .. code:: shell

-   mlx.distributed_config --verbose --hosts host1,host2,host3,host4
+   mlx.distributed_config --verbose --hosts host1,host2,host3,host4 --backend ring

 By default the script will attempt to discover the thunderbolt ring and provide
 you with the commands to configure each node as well as the ``hostfile.json``
 to use with ``mlx.launch``. If password-less ``sudo`` is available on the nodes
 then ``--auto-setup`` can be used to configure them automatically.

-To validate your connection without configuring anything
-``mlx.distributed_config`` can also plot the ring using DOT format.
-
-.. code:: shell
-
-   mlx.distributed_config --verbose --hosts host1,host2,host3,host4 --dot >ring.dot
-   dot -Tpng ring.dot >ring.png
-   open ring.png
-
 If you want to go through the process manually, the steps are as follows:

 * Disable the thunderbolt bridge interface
@@ -342,3 +191,382 @@ If you want to go through the process manually, the steps are as follows:
  and ``en2`` also on node ``i + 1`` then we may assign IPs ``192.168.0.1`` and
  ``192.168.0.2`` respectively to the two nodes. For more details you can see
  the commands prepared by the utility script.
+
+.. _jaccl_section:
+
+Getting Started with JACCL
+--------------------------
+
+Starting from macOS 26.2, RDMA over thunderbolt is available and
+enables low-latency communication between Macs with thunderbolt 5. MLX provides
+the JACCL backend that uses this functionality to achieve communication latency
+an order of magnitude lower than the ring backend.
+
+.. note::
+
+   The name JACCL (pronounced Jackal) stands for *Jack and Angelos' Collective
+   Communication Library* and it is an obvious pun to Nvidia's NCCL but also
+   tribute to *Jack Beasley* who led the development of RDMA over Thunderbolt
+   at Apple.
+
+Enabling RDMA
+^^^^^^^^^^^^^
+
+Until the feature matures, enabling RDMA over thunderbolt is slightly more
+involved and **cannot** be done remotely even with sudo. In fact, it has to be
+done in macOS recovery:
+
+1. `Start your computer in recovery <https://support.apple.com/en-us/102518>`_.
+2. Open the Terminal by going to Utilities -> Terminal.
+3. Run ``rdma_ctl enable``.
+4. Reboot.
+
+To verify that you have successfully enabled Thunderbolt RDMA you can run
+``ibv_devices`` which should produce something like the following for an M3 Ultra.
+
+.. code-block:: bash
+
+    ~ % ibv_devices
+    device          	   node GUID
+    ------          	----------------
+    rdma_en2        	8096a9d9edbaac05
+    rdma_en3        	8196a9d9edbaac05
+    rdma_en5        	8396a9d9edbaac05
+    rdma_en4        	8296a9d9edbaac05
+    rdma_en6        	8496a9d9edbaac05
+    rdma_en7        	8596a9d9edbaac05
+
+Defining a Mesh
+^^^^^^^^^^^^^^^
+
+The JACCL backend supports only fully connected topologies. Namely, there needs
+to be a thunderbolt cable connecting all pairs of Macs directly. For example, in
+the following topology visualizations, the left one is valid because there is a
+connection from any node to any other node, while for the one on the right M3
+Ultra 1 is not connected to M3 Ultra 2.
+
+.. raw:: html
+
+   <div style="display: flex; text-align: center; align-items: end; font-size: 80%;">
+     <div>
+       <img src="../_static/distributed/m3-ultra-mesh.png" alt="M3 Ultra thunderbolt mesh" style="width: 55%">
+       <p>Fully connected mesh of four M3 Ultra.</p>
+     </div>
+     <div>
+       <img src="../_static/distributed/m3-ultra-mesh-broken.png" alt="M3 Ultra broken thunderbolt mesh" style="width: 55%">
+       <p>Not a valid mesh (M3 Ultra 1 is not connected to M3 Ultra 2).</p>
+     </div>
+   </div>
+
+Similar to the ring backend, the easiest way to use JACCL with MLX is to write
+a JSON hostfile that will be used by ``mlx.launch``. The hostfile needs to contain
+
+- Hostnames to use for launching scripts via ssh
+- An IP for rank 0 that is reachable by all nodes
+- A list of rdma devices that connect each node to each other node
+
+The following JSON defines the valid 4-node mesh from the image above.
+
+.. code-block:: json
+
+    [
+        {
+            "ssh": "m3-ultra-1",
+            "ips": ["123.123.123.1"],
+            "rdma": [null, "rdma_en5", "rdma_en4", "rdma_en3"]
+        },
+        {
+            "ssh": "m3-ultra-2",
+            "ips": [],
+            "rdma": ["rdma_en5", null, "rdma_en3", "rdma_en4"]
+        },
+        {
+            "ssh": "m3-ultra-3",
+            "ips": [],
+            "rdma": ["rdma_en4", "rdma_en3", null, "rdma_en5"]
+        },
+        {
+            "ssh": "m3-ultra-4",
+            "ips": [],
+            "rdma": ["rdma_en3", "rdma_en4", "rdma_en5", null]
+        }
+    ]
+
+Even though TCP/IP is not used when communicating with Thunderbolt RDMA,
+disabling the thunderbolt bridge is still required as well as setting up
+isolated local networks for each thunderbolt connection.
+
+All of the above can be done instead via ``mlx.distributed_config``. This helper
+script will
+
+- ssh into each node
+- extract the thunderbolt connectivity
+- check for a valid mesh
+- provide the commands to configure each node (or run them if sudo is available)
+- generate the hostfile to be used with ``mlx.launch``
+
+Putting It All Together
+^^^^^^^^^^^^^^^^^^^^^^^^
+
+For example launching a distributed MLX script that uses JACCL is fairly simple
+if the nodes are reachable via ssh and have password-less sudo.
+
+First, connect all the thunderbolt cables. Then we can verify the connections
+by using the ``mlx.distributed_config`` script to visualize them.
+
+.. code-block::
+
+   mlx.distributed_config --verbose \
+        --hosts m3-ultra-1,m3-ultra-2,m3-ultra-3,m3-ultra-4 \
+        --over thunderbolt --dot | dot -Tpng | open -f -a Preview
+
+After making sure that everything looks right we can auto-configure the nodes
+and save the hostfile to ``m3-ultra-jaccl.json`` by running:
+
+.. code-block::
+
+   mlx.distributed_config --verbose \
+        --hosts m3-ultra-1,m3-ultra-2,m3-ultra-3,m3-ultra-4 \
+        --over thunderbolt --backend jaccl \
+        --auto-setup --output m3-ultra-jaccl.json
+
+And now we are ready to run a distributed MLX script such as distributed inference
+of a gigantic model using MLX LM.
+
+.. code-block::
+
+   mlx.launch --verbose --backend jaccl --hostfile m3-ultra-jaccl.json \
+        --env MLX_METAL_FAST_SYNCH=1 -- \  # <--- important
+        /path/to/remote/python -m mlx_lm chat --model mlx-community/DeepSeek-R1-0528-4bit
+
+.. note::
+
+   Defining the environment variable ``MLX_METAL_FAST_SYNCH=1`` enables a
+   different, faster way of synchronizing between the GPU and the CPU. It is
+   not specific to the JACCL backend and can be used in all cases where the CPU
+   and GPU need to collaborate for some computation and is pretty critical for
+   low-latency communication since the communication is done by the CPU.
+
+.. _nccl_section:
+
+Getting Started with NCCL
+-------------------------
+
+MLX on CUDA environments ships with the ability to talk to `NCCL
+<https://developer.nvidia.com/nccl>`_ which is a high-performance collective
+communication library that supports both multi-gpu and multi-node setups.
+
+For CUDA environments, NCCL is the default backend for ``mlx.launch`` and all
+it takes to run a distributed job is
+
+.. code-block::
+
+   mlx.launch -n 8 test.py
+
+   # perfect for interactive scripts
+   mlx.launch -n 8 python -m mlx_lm chat --model my-model
+
+You can also use ``mlx.launch`` to ssh to a remote node and launch a script
+with the same ease
+
+.. code-block::
+
+   mlx.launch --hosts my-cuda-node -n 8 test.py
+
+In many cases you may not want to use ``mlx.launch`` with the NCCL backend
+because the cluster scheduler will be the one launching the processes. You can
+:ref:`see which environment variables need to be defined <no_mlx_launch>` in
+order for the MLX NCCL backend to be initialized correctly.
+
+.. _mpi_section:
+
+Getting Started with MPI
+------------------------
+
+MLX already comes with the ability to "talk" to `MPI
+<https://en.wikipedia.org/wiki/Message_Passing_Interface>`_ if it is installed
+on the machine. Launching distributed MLX programs that use MPI can be done
+with ``mpirun`` as expected. However, in the following examples we will be
+using ``mlx.launch --backend mpi`` which takes care of some nuisances such as
+setting absolute paths for the ``mpirun`` executable and the ``libmpi.dyld``
+shared library.
+
+The simplest possible usage is the following which, assuming the minimal
+example in the beginning of this page, should result in:
+
+.. code:: shell
+
+    $ mlx.launch --backend mpi -n 2 test.py
+    1 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
+    0 array([2, 2, 2, ..., 2, 2, 2], dtype=float32)
+
+The above launches two processes on the same (local) machine and we can see
+both standard output streams. The processes send the array of 1s to each other
+and compute the sum which is printed. Launching with ``mlx.launch -n 4 ...`` would
+print 4 etc.
+
+Installing MPI
+^^^^^^^^^^^^^^
+
+MPI can be installed with Homebrew, pip, using the Anaconda package manager, or
+compiled from source. Most of our testing is done using ``openmpi`` installed
+with the Anaconda package manager as follows:
+
+.. code:: shell
+
+    $ conda install conda-forge::openmpi
+
+Installing with Homebrew or pip requires specifying the location of ``libmpi.dyld``
+so that MLX can find it and load it at runtime. This can simply be achieved by
+passing the ``DYLD_LIBRARY_PATH`` environment variable to ``mpirun`` and it is
+done automatically by ``mlx.launch``. Some environments use a non-standard
+library filename that can be specified using the ``MPI_LIBNAME`` environment
+variable. This is automatically taken care of by ``mlx.launch`` as well.
+
+.. code:: shell
+
+    $ mpirun -np 2 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ -x MPI_LIBNAME=libmpi.40.dylib python test.py
+    $ # or simply
+    $ mlx.launch -n 2 test.py
+
+Setting up Remote Hosts
+^^^^^^^^^^^^^^^^^^^^^^^
+
+MPI can automatically connect to remote hosts and set up the communication over
+the network if the remote hosts can be accessed via ssh. A good checklist to
+debug connectivity issues is the following:
+
+* ``ssh hostname`` works from all machines to all machines without asking for
+  password or host confirmation
+* ``mpirun`` is accessible on all machines.
+* Ensure that the ``hostname`` used by MPI is the one that you have configured
+  in the ``.ssh/config`` files on all machines.
+
+Tuning MPI All Reduce
+^^^^^^^^^^^^^^^^^^^^^
+
+.. note::
+
+    For faster all reduce consider using the ring backend either with Thunderbolt
+    connections or over Ethernet.
+
+Configure MPI to use N tcp connections between each host to improve bandwidth
+by passing ``--mca btl_tcp_links N``.
+
+Force MPI to use the most performant network interface by setting ``--mca
+btl_tcp_if_include <iface>`` where ``<iface>`` should be the interface you want
+to use.
+
+.. _no_mlx_launch:
+
+Distributed Without ``mlx.launch``
+----------------------------------
+
+None of the implementations of the distributed backends require launching with
+``mlx.launch``. The script simply connects to each host. Starts a process per
+rank and sets up the necessary environment variables before delegating to your
+MLX script. See the :doc:`dedicated documentation page <launching_distributed>`
+for more details.
+
+For many use-cases this will be the easiest way to perform distributed
+computations in MLX. However, there may be reasons that you cannot or should
+not use ``mlx.launch``. A common such case is the use of a scheduler that
+starts all the processes for you on machines undetermined at the time of
+scheduling the job.
+
+Below we list the environment variables required to use each backend.
+
+Ring
+^^^^^^
+
+**MLX_RANK** should contain a single 0-based integer that defines the rank of
+the process.
+
+**MLX_HOSTFILE** should contain the path to a json file that contains IPs and
+ports for each rank to listen to, something like the following:
+
+.. code-block:: json
+
+   [
+     ["123.123.1.1:5000", "123.123.1.2:5000"],
+     ["123.123.2.1:5000", "123.123.2.2:5000"],
+     ["123.123.3.1:5000", "123.123.3.2:5000"],
+     ["123.123.4.1:5000", "123.123.4.2:5000"]
+   ]
+
+**MLX_RING_VERBOSE** is optional and if set to 1 it enables some more logging
+from the distributed backend.
+
+JACCL
+^^^^^
+
+**MLX_RANK** should contain a single 0-based integer that defines the rank of
+the process.
+
+**MLX_JACCL_COORDINATOR** should contain the IP and port that rank 0 can listen
+to all the other ranks connect to in order to establish the RDMA connections.
+
+**MLX_IBV_DEVICES** should contain the path to a json file that contains the
+ibverbs device names that connect each node to each other node, something like
+the following:
+
+.. code-block:: json
+
+   [
+      [null, "rdma_en5", "rdma_en4", "rdma_en3"],
+      ["rdma_en5", null, "rdma_en3", "rdma_en4"],
+      ["rdma_en4", "rdma_en3", null, "rdma_en5"],
+      ["rdma_en3", "rdma_en4", "rdma_en5", null]
+   ]
+
+
+NCCL
+^^^^^
+
+**MLX_RANK** should contain a single 0-based integer that defines the rank of
+the process.
+
+**MLX_WORLD_SIZE** should contain the total number of processes that will be
+launched.
+
+**NCCL_HOST_IP** and **NCCL_PORT** should contain the IP and port that all
+hosts can connect to to establish the NCCL communication.
+
+**CUDA_VISIBLE_DEVICES** should contain the local index of the gpu that
+corresponds to this process.
+
+Of course any `other environment variable
+<https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html>`_ that is
+used by NCCL can be set.
+
+.. _tips_and_tricks:
+
+Tips and Tricks
+----------------
+
+This is a small collection of tips to help you utilize better the distributed
+communication capabilities of MLX.
+
+- *Test locally first.*
+
+  You can use the pattern ``mlx.launch -n2 -- my_script.py`` to run a small
+  scale test on a single node first.
+
+- *Batch your communication.*
+
+  As described in the :ref:`training example <training_example>`, performing a
+  lot of small communications can hurt performance. Copy the approach of
+  :func:`mlx.nn.average_gradients` to gather many small communications in a
+  single large one.
+
+- *Visualize the connectivity.*
+
+  Use ``mlx.distributed_config --hosts h1,h2,h3 --over thunderbolt --dot`` to
+  visualize the connnections and make sure that the cables are connected
+  correctly. See the :ref:`JACCL section <jaccl_section>` for examples.
+
+- *Use the debugger.*
+
+  ``mlx.launch`` is meant for interactive use. It broadcasts stdin to all
+  processes and gathers stdout from all processes. This makes using ``pdb`` a
+  breeze.
@@ -70,7 +70,8 @@ Differences from NumPy

  * Indexing does not perform bounds checking. Indexing out of bounds is
    undefined behavior.
-  * Boolean mask based indexing is not yet supported.
+  * Boolean mask based indexing is supported for assignment only (see
+    :ref:`boolean-mask-assignment`).

 The reason for the lack of bounds checking is that exceptions cannot propagate
 from the GPU. Performing bounds checking for array indices before launching the
@@ -143,3 +144,51 @@ expected. For example:

 In the above ``dfdx`` will have the correct gradient, namely zeros at ``idx``
 and ones elsewhere.
+
+.. _boolean-mask-assignment:
+
+Boolean Mask Assignment
+-----------------------
+
+MLX supports boolean indices using NumPy syntax. A mask must already be
+a :class:`bool_` MLX :class:`array` or a NumPy ``ndarray`` with ``dtype=bool``.
+Other index types are routed through the standard scatter code.
+
+.. code-block:: shell
+
+   >>> a = mx.array([1.0, 2.0, 3.0])
+   >>> mask = mx.array([True, False, True])
+   >>> updates = mx.array([5.0, 6.0])
+   >>> a[mask] = updates
+   >>> a
+   array([5.0, 2.0, 6.0], dtype=float32)
+
+Scalar assignments broadcast to every ``True`` entry in ``mask``. For non-scalar
+assignments, ``updates`` must provide at least as many elements as there are
+``True`` entries in ``mask``.
+
+.. code-block:: shell
+
+   >>> a = mx.zeros((2, 3))
+   >>> mask = mx.array([[True, False, True],
+                        [False, False, True]])
+   >>> a[mask] = 1.0
+   >>> a
+   array([[1.0, 0.0, 1.0],
+          [0.0, 0.0, 1.0]], dtype=float32)
+
+Boolean masks follow NumPy semantics:
+
+- The mask shape must match the shape of the axes it indexes exactly. The only
+  exception is a scalar boolean mask, which broadcasts to the full array.
+- Any axes not covered by the mask are taken in full.
+
+.. code-block:: shell
+
+   >>> a = mx.arange(1000).reshape(10, 10, 10)
+   >>> a[mx.random.normal((10, 10)) > 0.0] = 0  # valid: mask covers axes 0 and 1
+
+The mask of shape ``(10, 10)`` applies to the first two axes, so ``a[mask]``
+selects the 1-D slices ``a[i, j, :]`` where ``mask[i, j]`` is ``True``.
+Shapes such as ``(1, 10, 10)`` or ``(10, 10, 1)`` do not match the indexed
+axes and therefore raise errors.
@@ -7,13 +7,106 @@ Launching Distributed Programs

 .. currentmodule:: mlx.core.distributed

-Installing the MLX python package provides a helper script ``mlx.launch`` that
-can be used to run python scripts distributed on several nodes. It allows
-launching using either the MPI backend or the ring backend. See the
-:doc:`distributed docs <distributed>` for the different backends.
+The MLX python package provides two utilities to help you configure
+your Macs for distributed computation and also launch distributed programs on
+multiple nodes or with many processes in a single node. These utilities are aptly named

-Usage
-----
+- ``mlx.launch``
+- ``mlx.distributed_config``
+
+See the :doc:`distributed docs <distributed>` for an introduction and
+getting-started guides to the various backends.
+
+``mlx.distributed_config`` 
+---------------------------
+
+Unless you are launching distributed jobs locally for development or multi-gpu
+CUDA environments, then you have several Macs that you need to configure for
+distributed communication with MLX.
+
+``mlx.distributed_config`` aims to automate the process of configuring the
+network interfaces (especially for communication over thunderbolt) and also
+creating the hostfile to be used with ``mlx.launch``.
+
+We will analyse 3 cases of using ``mlx.distributed_config``
+
+1. RDMA over thunderbolt using JACCL
+2. TCP/IP over thunderbolt using the ring backend
+3. TCP/IP over ethernet using the ring backend
+
+JACCL
+^^^^^^^
+
+After following :ref:`the steps to enable RDMA <jaccl_section>` you can run the
+following command to configure the nodes and create the hostfile.
+
+.. code-block::
+
+   mlx.distributed_config --verbose --backend jaccl \
+        --hosts m3-ultra-1,m3-ultra-2,m3-ultra-3,m3-ultra-4 --over thunderbolt \
+        --auto-setup --output m3-ultra-jaccl.json
+
+Let's walk through the steps that the script takes to configure the nodes.
+
+1. ssh to all nodes to verify that they are reachable
+2. Extract the thunderbolt connectivity. Namely run commands on each node to
+   calculate which node is connected to which other node.
+3. Verify that we have a valid fully connected mesh
+4. Check that RDMA is enabled
+5. Extract the ethernet IP from interface en0
+6. Disable the thunderbolt bridge and set up peer to peer networks for each
+   thunderbolt cable
+7. Write the hostfile
+
+Knowing the above steps allows you to manually configure the nodes but also
+debug any configuration issue. For instance changing the Ethernet IP to a
+different interface directly in the config is possible (as long as it is
+reachable from all nodes).
+
+The ``--auto-setup`` argument requires password-less sudo on each node. If it
+isn't available then the configuration script will print commands to be run on
+each node.
+
+Ring over thunderbolt
+^^^^^^^^^^^^^^^^^^^^^
+
+Setting up a ring backend over thunderbolt only requires changing the
+``--backend`` from ``jaccl`` to ``ring``.
+
+The steps are very similar with the main difference being that instead of
+verifying that the nodes are fully connected, the script attempts to identify a
+ring topology (or multiple rings).
+
+Ring over Ethernet
+^^^^^^^^^^^^^^^^^^
+
+Configuring the ring backend over ethernet doesn't require setting up network
+interface and as such it simply extracts the ``en0`` IP from each node and
+writes the hostfile.
+
+Debugging cable connections
+^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+``mlx.distributed_config`` can help you debug the connectivity of your nodes
+over thunderbolt by exporting a graph of the connections.
+
+Running
+
+.. code-block::
+
+   mlx.distributed_config --verbose \
+        --hosts host1,host2,host3,host4 \
+        --over thunderbolt --dot
+
+will export a `GraphViz <https://graphviz.org>`_ representation of the
+connections between the nodes which makes it very easy to figure out which
+cable is not connected correctly.
+
+See :ref:`the JACCL section <jaccl_section>` for an example.
+
+
+``mlx.launch``
+--------------

 The minimal usage example of ``mlx.launch`` is simply

@@ -33,6 +126,10 @@ the rest if one of them fails unexpectedly or if ``mlx.launch`` is terminated.
 It also takes care of forwarding the output of each remote process to stdout
 and stderr respectively.

+Importantly, it also broadcasts stdin to each process which enables interactive
+programs to work in distributed mode as well as debugging using the interactive
+debugger.
+
 Providing Hosts
 ^^^^^^^^^^^^^^^^

@@ -63,10 +160,62 @@ host and on the same path. A good checklist to debug errors is the following:
  ``mlx.launch --print-python`` to see what that path is.
 * the script you want to run is available on all hosts at the same path

+If you are launching from a node with a completely different setup than the
+nodes that the program will run on, you can specify ``--no-verify-script`` so
+that ``mlx.launch`` does not attempt to verify that the executable and script
+exist locally before launching the distributed job.
+
+.. _ring_specifics:
+
+Ring Specifics
+^^^^^^^^^^^^^^
+
+The :ref:`ring <ring_section>` backend, which is also the default
+backend, can be explicitly selected with the argument ``--backend ring``. The
+ring backend has some specific requirements and arguments that are different to
+other backends:
+
+* The argument ``--hosts`` only accepts IPs and not hostnames. If we need to
+  ssh to a hostname that does not correspond to the IP we want to bind to we
+  have to provide a hostfile.
+* ``--starting-port`` defines the port to bind to on the remote hosts.
+  Specifically rank 0 for the first IP will use this port and each subsequent
+  IP or rank will add 1 to this port.
+* ``--connections-per-ip`` allows us to increase the number of connections
+  between neighboring nodes. This corresponds to ``--mca btl_tcp_links 2`` for
+  ``mpirun``.
+
+.. _jaccl_specifics:
+
+JACCL Specifics
+^^^^^^^^^^^^^^^^
+
+The :ref:`JACCL <jaccl_section>` backend can be selected with the argument
+``--backend jaccl``. A hostfile is necessary to launch with this backend
+because it needs to contain the RDMA devices connecting each node to each other
+node.
+
+NCCL Specifics
+^^^^^^^^^^^^^^
+
+The :ref:`NCCL <nccl_section>` backend is the default backend for CUDA
+environments. When launching from a Mac to a Linux machine with CUDA then the
+backend should be selected using ``--backend nccl``.
+
+The ``--repeat-hosts, -n`` argument should be used to launch multi-node and
+multi-gpu jobs. For instance
+
+.. code-block::
+
+   mlx.launch --backend nccl --hosts linux-1,linux-2 -n 8 --no-verify-script -- ./my-job.sh
+
+will attempt to launch 16 processes, 8 on each node that will all run
+``my-job.sh``.
+
 .. _mpi_specifics:

 MPI Specifics
-------------
+^^^^^^^^^^^^^

 One can use MPI by passing ``--backend mpi`` to ``mlx.launch``. In that case,
 ``mlx.launch`` is a thin wrapper over ``mpirun``. Moreover,
@@ -83,23 +232,3 @@ to choose a specific interface for the byte-transfer-layer of MPI we can call
 .. code:: shell

    mlx.launch --backend mpi --mpi-arg '--mca btl_tcp_if_include en0' --hostfile hosts.json my_script.py
-
-
-.. _ring_specifics:
-
-Ring Specifics
--------------
-
-The ring backend, which is also the default backend, can be explicitly selected
-with the argument ``--backend ring``. The ring backend has some specific
-requirements and arguments that are different to MPI:
-
-* The argument ``--hosts`` only accepts IPs and not hostnames. If we need to
-  ssh to a hostname that does not correspond to the IP we want to bind to we
-  have to provide a hostfile.
-* ``--starting-port`` defines the port to bind to on the remote hosts.
-  Specifically rank 0 for the first IP will use this port and each subsequent
-  IP or rank will add 1 to this port.
-* ``--connections-per-ip`` allows us to increase the number of connections
-  between neighboring nodes. This corresponds to ``--mca btl_tcp_links 2`` for
-  ``mpirun``.
@@ -3,6 +3,6 @@ requires = [
  "setuptools>=42",
  "cmake>=3.25",
  "mlx>=0.18.0",
-  "nanobind==2.4.0",
+  "nanobind==2.10.2",
 ]
 build-backend = "setuptools.build_meta"
@@ -1,4 +1,4 @@
 setuptools>=42
 cmake>=3.25
 mlx>=0.21.0
-nanobind==2.4.0
+nanobind==2.10.2
@@ -29,12 +29,12 @@ def loss_fn(w):

 grad_fn = mx.grad(loss_fn)

-tic = time.time()
+tic = time.perf_counter()
 for _ in range(num_iters):
    grad = grad_fn(w)
    w = w - lr * grad
    mx.eval(w)
-toc = time.time()
+toc = time.perf_counter()

 loss = loss_fn(w)
 error_norm = mx.sum(mx.square(w - w_star)).item() ** 0.5
@@ -30,13 +30,13 @@ def loss_fn(w):

 grad_fn = mx.grad(loss_fn)

-tic = time.time()
+tic = time.perf_counter()
 for _ in range(num_iters):
    grad = grad_fn(w)
    w = w - lr * grad
    mx.eval(w)

-toc = time.time()
+toc = time.perf_counter()

 loss = loss_fn(w)
 final_preds = (X @ w) > 0
@@ -0,0 +1,117 @@
+from itertools import product
+
+import mlx.core as mx
+
+
+# In mxfp8 mode, the results do not match exactly:
+# fewer than 1% of output elements differ.
+# This does not appear to be a systematic error.
+# The error can exceed 1 ULP for very small values,
+# and is always below 1 ULP for larger values.
+# For nvfp4, the results match exactly.
+# therefore I suspect that the discrepancy comes from
+# the mxfp8 matmul implementation in cuBLASLt..
+def ulp_bf16_at(x):
+    ax = mx.abs(x)
+    min_normal = mx.array(2.0**-126)
+    ax = mx.where(ax < min_normal, min_normal, ax)
+    e = mx.floor(mx.log2(ax))
+    return mx.power(2.0, e - 7.0)
+
+
+def test_qqmm():
+    key = mx.random.key(0)
+    k1, k2 = mx.random.split(key)
+    dtypes = [mx.bfloat16, mx.float32, mx.float16]
+
+    tests = (
+        (16, "nvfp4", 4),
+        (32, "mxfp8", 8),
+    )
+    shapes = (
+        [64, 65, 33, 128, 256, 1024, 1024 * 8],  # M
+        [64, 128, 256, 1024, 1024 * 8],  # N
+        [64, 128, 256, 1024, 1024 * 8],  # K
+    )
+    for group_size, mode, bits in tests:
+        for M, N, K in product(*shapes):
+            for dtype in dtypes:
+                x = mx.random.normal(shape=(M, K), key=k1, dtype=dtype)
+                w = mx.random.normal(shape=(N, K), key=k2, dtype=dtype)
+                w_q, scales_w = mx.quantize(w, group_size, bits, mode=mode)
+                w_dq = mx.dequantize(
+                    w_q,
+                    scales_w,
+                    group_size=group_size,
+                    bits=bits,
+                    mode=mode,
+                    dtype=dtype,
+                )
+                y_q = mx.qqmm(
+                    x,
+                    w_q,
+                    scales_w,
+                    group_size=group_size,
+                    bits=bits,
+                    mode=mode,
+                )
+                x_q, scales_x = mx.quantize(
+                    x, group_size=group_size, bits=bits, mode=mode
+                )
+                x_dq = mx.dequantize(
+                    x_q,
+                    scales_x,
+                    group_size=group_size,
+                    bits=bits,
+                    mode=mode,
+                    dtype=dtype,
+                )
+                y_hat = mx.matmul(x_dq, mx.transpose(w_dq))
+                ulp = ulp_bf16_at(y_hat)
+                error = (y_q - y_hat).abs()
+                if not (mx.logical_or(error < 1e-3, error <= ulp).all()):
+                    raise AssertionError(
+                        f"qqmm test failed for shape {(M, N, K)}, "
+                        f"group_size={group_size}, bits={bits}, "
+                        f"mode={mode}, dtype={dtype}"
+                    )
+
+
+def test_qqmm_vjp():
+    key = mx.random.key(0)
+    k1, k2 = mx.random.split(key)
+    M = 64
+    N = 1024
+    K = 512
+    tests = (
+        (16, "nvfp4", 4),
+        (32, "mxfp8", 8),
+    )
+    x = mx.random.normal(shape=(M, K), key=k1)
+    c = mx.ones(shape=(M, N))
+
+    for group_size, mode, bits in tests:
+        w = mx.random.normal(shape=(N, K), key=k2)
+
+        def fn(x):
+            return mx.qqmm(x, w, group_size=group_size, bits=bits, mode=mode)
+
+        _, vjp_out = mx.vjp(fn, primals=(x,), cotangents=(c,))
+        w_tq, scales_wt = mx.quantize(
+            mx.transpose(w), group_size=group_size, bits=bits, mode=mode
+        )
+        expected_out = mx.qqmm(
+            c, w_tq, scales_wt, group_size=group_size, bits=bits, mode=mode
+        )
+        ulp = ulp_bf16_at(expected_out)
+        error = (vjp_out[0] - expected_out).abs()
+        if not (mx.logical_or(error < 1e-3, error <= ulp).all()):
+            raise AssertionError(
+                f"qqmm vjp test failed for shape {(M, N, K)}, "
+                f"group_size={group_size}, bits={bits}, mode={mode}"
+            )
+
+
+if __name__ == "__main__":
+    test_qqmm()
+    test_qqmm_vjp()
@@ -1,7 +1,6 @@
 target_sources(
  mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/allocator.cpp
-          ${CMAKE_CURRENT_SOURCE_DIR}/array.cpp
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/array.cpp
          ${CMAKE_CURRENT_SOURCE_DIR}/compile.cpp
          ${CMAKE_CURRENT_SOURCE_DIR}/device.cpp
          ${CMAKE_CURRENT_SOURCE_DIR}/dtype.cpp
@@ -23,16 +22,57 @@ target_sources(
 # Define MLX_VERSION only in the version.cpp file.
 add_library(mlx_version OBJECT ${CMAKE_CURRENT_SOURCE_DIR}/version.cpp)
 target_compile_definitions(mlx_version PRIVATE MLX_VERSION="${MLX_VERSION}")
+target_include_directories(mlx_version PRIVATE ${PROJECT_SOURCE_DIR})
 target_link_libraries(mlx PRIVATE $<BUILD_INTERFACE:mlx_version>)

-if(MSVC)
-  # Disable some MSVC warnings to speed up compilation.
-  target_compile_options(mlx PUBLIC /wd4068 /wd4244 /wd4267 /wd4804)
+# Do not export symbols by default.
+set_target_properties(
+  mlx mlx_version
+  PROPERTIES VISIBILITY_INLINES_HIDDEN ON
+             CXX_VISIBILITY_PRESET hidden
+             CUDA_VISIBILITY_PRESET hidden)
+
+# Define MLX_EXPORT for shared libraries, MLX_STATIC for static libraries.
+set_target_properties(mlx PROPERTIES DEFINE_SYMBOL MLX_EXPORT)
+if(BUILD_SHARED_LIBS)
+  target_compile_definitions(mlx_version PUBLIC MLX_EXPORT)
+else()
+  target_compile_definitions(mlx PUBLIC MLX_STATIC)
+  target_compile_definitions(mlx_version PUBLIC MLX_STATIC)
 endif()

-if(WIN32)
-  # Export symbols by default to behave like macOS/linux.
-  set_target_properties(mlx PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS TRUE)
+if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
+  # Supress warnings: note: parameter passing for argument of type
+  # 'std::pair<float, float>' when C++17 is enabled changed to match C++14 in
+  # GCC 10.1
+  target_compile_options(mlx PRIVATE -Wno-psabi)
+endif()
+
+if(MSVC)
+  # Some of CUDA's headers include windows.h, which defines min/max macros.
+  target_compile_definitions(mlx PRIVATE NOMINMAX WIN32_LEAN_AND_MEAN)
+  # Unicode support in fmt does not compile in .cu files.
+  target_compile_definitions(mlx PRIVATE FMT_UNICODE=0)
+  # Disable some MSVC warnings to speed up compilation.
+  target_compile_options(
+    mlx
+    PUBLIC $<$<COMPILE_LANGUAGE:CXX>:/wd4244 /wd4267>
+    PRIVATE $<$<COMPILE_LANGUAGE:CXX>:/wd4068
+            /wd4146
+            /wd4700
+            /wd4804
+            /wd4805>
+            $<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=/wd4244
+            -Xcompiler=/wd4267>)
+  # Enable /bigobj for heavily templated code (e.g., binary.cpp) that exceeds
+  # the default 65,535 section limit in COFF object files.
+  target_compile_options(
+    mlx PRIVATE $<$<COMPILE_LANGUAGE:CXX>:/bigobj>
+                $<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=/bigobj>)
+  # Use modern preprocessor, otherwise CCCL would complain.
+  target_compile_options(
+    mlx PRIVATE $<$<COMPILE_LANGUAGE:CXX>:/Zc:preprocessor>
+                $<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=/Zc:preprocessor>)
 endif()

 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/common)
@@ -1,24 +0,0 @@
-// Copyright © 2023 Apple Inc.
-
-#include <cstdlib>
-#include <sstream>
-
-#include "mlx/allocator.h"
-
-namespace mlx::core::allocator {
-
-Buffer malloc(size_t size) {
-  auto buffer = allocator().malloc(size);
-  if (size && !buffer.ptr()) {
-    std::ostringstream msg;
-    msg << "[malloc] Unable to allocate " << size << " bytes.";
-    throw std::runtime_error(msg.str());
-  }
-  return buffer;
-}
-
-void free(Buffer buffer) {
-  allocator().free(buffer);
-}
-
-} // namespace mlx::core::allocator
@@ -4,17 +4,19 @@

 #include <cstdlib>

+#include "mlx/api.h"
+
 namespace mlx::core::allocator {

 // Simple wrapper around buffer pointers
 // WARNING: Only Buffer objects constructed from and those that wrap
 //          raw pointers from mlx::allocator are supported.
-class Buffer {
+class MLX_API Buffer {
 private:
  void* ptr_;

 public:
-  Buffer(void* ptr) : ptr_(ptr) {};
+  explicit Buffer(void* ptr) : ptr_(ptr) {};

  // Get the raw data pointer from the buffer
  void* raw_ptr();
@@ -28,16 +30,16 @@ class Buffer {
  };
 };

-Buffer malloc(size_t size);
-
-void free(Buffer buffer);
-
-class Allocator {
+class MLX_API Allocator {
  /** Abstract base class for a memory allocator. */
 public:
  virtual Buffer malloc(size_t size) = 0;
  virtual void free(Buffer buffer) = 0;
  virtual size_t size(Buffer buffer) const = 0;
+  virtual Buffer make_buffer(void* ptr, size_t size) {
+    return Buffer{nullptr};
+  };
+  virtual void release(Buffer buffer) {}

  Allocator() = default;
  Allocator(const Allocator& other) = delete;
@@ -47,6 +49,27 @@ class Allocator {
  virtual ~Allocator() = default;
 };

-Allocator& allocator();
+MLX_API Allocator& allocator();
+
+inline Buffer malloc(size_t size) {
+  return allocator().malloc(size);
+}
+
+inline void free(Buffer buffer) {
+  allocator().free(buffer);
+}
+
+// Make a Buffer from a raw pointer of the given size without a copy.  If a
+// no-copy conversion is not possible then the returned buffer.ptr() will be
+// nullptr. Any buffer created with this function must be released with
+// release(buffer)
+inline Buffer make_buffer(void* ptr, size_t size) {
+  return allocator().make_buffer(ptr, size);
+};
+
+// Release a buffer from the allocator made with make_buffer
+inline void release(Buffer buffer) {
+  allocator().release(buffer);
+}

 } // namespace mlx::core::allocator
@@ -0,0 +1,29 @@
+// Copyright © 2024 Apple Inc.
+
+#pragma once
+
+// MLX_API macro for controlling symbol visibility, must add for public APIs.
+//
+// Usage:
+//   MLX_API void some_function(...);
+//   class MLX_API SomeClass { ... };
+
+#if defined(MLX_STATIC)
+
+// Static library build - no import/export decorations needed
+#define MLX_API
+
+#else
+
+// Shared library build.
+#if defined(_WIN32)
+#if defined(MLX_EXPORT)
+#define MLX_API __declspec(dllexport)
+#else
+#define MLX_API __declspec(dllimport)
+#endif // defined(MLX_EXPORT)
+#else
+#define MLX_API __attribute__((visibility("default")))
+#endif // defined(_WIN32)
+
+#endif // defined(MLX_STATIC)
@@ -21,11 +21,12 @@ array::array(
    Dtype dtype,
    std::shared_ptr<Primitive> primitive,
    std::vector<array> inputs)
-    : array_desc_(std::make_shared<ArrayDesc>(
-          std::move(shape),
-          dtype,
-          std::move(primitive),
-          std::move(inputs))) {
+    : array_desc_(
+          std::make_shared<ArrayDesc>(
+              std::move(shape),
+              dtype,
+              std::move(primitive),
+              std::move(inputs))) {
  if (has_primitive() && this->primitive().stream().device == Device::gpu) {
    for (auto& in : this->inputs()) {
      if (in.dtype() == float64) {
@@ -64,24 +65,48 @@ array array::unsafe_weak_copy(const array& other) {
      other.strides(),
      other.flags(),
      [](auto) {});
-  cpy.array_desc_->data_ptr = other.array_desc_->data_ptr;
+  cpy.array_desc_->offset = other.array_desc_->offset;
  return cpy;
 }

 array::array(std::initializer_list<float> data)
-    : array_desc_(std::make_shared<ArrayDesc>(
-          Shape{static_cast<ShapeElem>(data.size())},
-          float32)) {
+    : array_desc_(
+          std::make_shared<ArrayDesc>(
+              Shape{static_cast<ShapeElem>(data.size())},
+              float32)) {
  init(data.begin());
 }

 array::array(std::initializer_list<int> data, Dtype dtype)
-    : array_desc_(std::make_shared<ArrayDesc>(
-          Shape{static_cast<ShapeElem>(data.size())},
-          dtype)) {
+    : array_desc_(
+          std::make_shared<ArrayDesc>(
+              Shape{static_cast<ShapeElem>(data.size())},
+              dtype)) {
  init(data.begin());
 }

+array::array(
+    void* data,
+    Shape shape,
+    Dtype dtype,
+    const std::function<void(void*)>& deleter)
+    : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
+  auto buffer = allocator::make_buffer(data, nbytes());
+  if (buffer.ptr() == nullptr) {
+    set_data(allocator::malloc(nbytes()));
+    auto ptr = static_cast<char*>(data);
+    std::copy(ptr, ptr + nbytes(), this->data<char>());
+    deleter(data);
+  } else {
+    auto wrapped_deleter = [deleter](allocator::Buffer buffer) {
+      auto ptr = buffer.raw_ptr();
+      allocator::release(buffer);
+      return deleter(ptr);
+    };
+    set_data(buffer, std::move(wrapped_deleter));
+  }
+}
+
 /* Build an array from a shared buffer */
 array::array(allocator::Buffer data, Shape shape, Dtype dtype, Deleter deleter)
    : array_desc_(std::make_shared<ArrayDesc>(std::move(shape), dtype)) {
@@ -141,7 +166,7 @@ bool array::is_tracer() const {

 void array::set_data(allocator::Buffer buffer, Deleter d) {
  array_desc_->data = std::make_shared<Data>(buffer, d);
-  array_desc_->data_ptr = buffer.raw_ptr();
+  array_desc_->offset = 0;
  array_desc_->data_size = size();
  array_desc_->flags.contiguous = true;
  array_desc_->flags.row_contiguous = true;
@@ -156,7 +181,7 @@ void array::set_data(
    Flags flags,
    Deleter d) {
  array_desc_->data = std::make_shared<Data>(buffer, d);
-  array_desc_->data_ptr = buffer.raw_ptr();
+  array_desc_->offset = 0;
  array_desc_->data_size = data_size;
  array_desc_->strides = std::move(strides);
  array_desc_->flags = flags;
@@ -167,14 +192,13 @@ void array::copy_shared_buffer(
    const Strides& strides,
    Flags flags,
    size_t data_size,
-    size_t offset /* = 0 */) {
+    int64_t offset /* = 0 */) {
  array_desc_->data = other.array_desc_->data;
  array_desc_->strides = strides;
  array_desc_->flags = flags;
  array_desc_->data_size = data_size;
-  auto char_offset = sizeof(char) * itemsize() * offset;
-  array_desc_->data_ptr = static_cast<void*>(
-      static_cast<char*>(other.array_desc_->data_ptr) + char_offset);
+  array_desc_->offset =
+      sizeof(char) * itemsize() * offset + other.array_desc_->offset;
 }

 void array::copy_shared_buffer(const array& other) {
@@ -8,6 +8,7 @@
 #include <vector>

 #include "mlx/allocator.h"
+#include "mlx/api.h"
 #include "mlx/dtype.h"
 #include "mlx/event.h"
 #include "mlx/small_vector.h"
@@ -22,7 +23,7 @@ using ShapeElem = int32_t;
 using Shape = SmallVector<ShapeElem>;
 using Strides = SmallVector<int64_t>;

-class array {
+class MLX_API array {
  /* An array is really a node in a graph. It contains a shared ArrayDesc
   * object */

@@ -57,6 +58,16 @@ class array {
      Shape shape,
      Dtype dtype = TypeToDtype<T>());

+  /* Build an array from a raw pointer. The constructor will attempt to use the
+   * input data without a copy. The deleter will be called when the array no
+   * longer needs the underlying memory - after the array is destroyed in the
+   * no-copy case and after the copy otherwise. */
+  explicit array(
+      void* data,
+      Shape shape,
+      Dtype dtype,
+      const std::function<void(void*)>& deleter);
+
  /* Build an array from a buffer */
  explicit array(
      allocator::Buffer data,
@@ -111,7 +122,7 @@ class array {
   *  This function supports negative indexing and provides
   *  bounds checking. */
  auto shape(int dim) const {
-    return shape().at(dim < 0 ? dim + ndim() : dim);
+    return shape().at(dim < 0 ? dim + static_cast<int>(ndim()) : dim);
  }

  /** The strides of the array. */
@@ -125,7 +136,7 @@ class array {
   *  This function supports negative indexing and provides
   *  bounds checking. */
  auto strides(int dim) const {
-    return strides().at(dim < 0 ? dim + ndim() : dim);
+    return strides().at(dim < 0 ? dim + static_cast<int>(ndim()) : dim);
  }

  /** Get the arrays data type. */
@@ -143,7 +154,7 @@ class array {
  template <typename T>
  T item() const;

-  struct ArrayIterator {
+  struct MLX_API ArrayIterator {
    using iterator_category = std::random_access_iterator_tag;
    using difference_type = size_t;
    using value_type = const array;
@@ -294,6 +305,11 @@ class array {
    return array_desc_->siblings;
  }

+  /** The array's position in the sibling list. */
+  int sibling_position() const {
+    return array_desc_->position;
+  }
+
  void set_siblings(std::vector<array> siblings, uint16_t position) {
    array_desc_->siblings = std::move(siblings);
    array_desc_->position = position;
@@ -349,15 +365,23 @@ class array {
    return array_desc_->data;
  }

-  // Return a raw pointer to the arrays data
+  // Return a raw pointer to the arrays data. This function may do a copy if
+  // the underlying buffer is not accessible on the CPU. When accessing the
+  // data for GPU kernels, be sure to use the correct method / function for the
+  // given backend to access the GPU pointer.
  template <typename T>
  T* data() {
-    return static_cast<T*>(array_desc_->data_ptr);
+    return reinterpret_cast<T*>(
+        (static_cast<char*>(buffer().raw_ptr()) + array_desc_->offset));
  }

  template <typename T>
  const T* data() const {
-    return static_cast<T*>(array_desc_->data_ptr);
+    return const_cast<array&>(*this).data<T>();
+  }
+
+  int64_t offset() const {
+    return array_desc_->offset;
  }

  enum Status {
@@ -426,7 +450,7 @@ class array {
      const Strides& strides,
      Flags flags,
      size_t data_size,
-      size_t offset = 0);
+      int64_t offset = 0);

  void copy_shared_buffer(const array& other);

@@ -441,7 +465,7 @@ class array {
  template <typename It>
  void init(const It src);

-  struct ArrayDesc {
+  struct MLX_API ArrayDesc {
    Shape shape;
    Strides strides;
    size_t size;
@@ -461,14 +485,14 @@ class array {
    // can share the underlying data buffer.
    std::shared_ptr<Data> data;

-    // Properly offset data pointer
-    void* data_ptr{nullptr};
+    // Offset from beginning of data pointer
+    int64_t offset{0};

    // The size in elements of the data buffer the array accesses
-    size_t data_size;
+    size_t data_size{0};

    // Contains useful meta data about the array
-    Flags flags;
+    Flags flags{true, true, true};

    std::vector<array> inputs;
    // An array to keep track of the siblings from a multi-output
@@ -518,9 +542,10 @@ template <typename T>
 array::array(
    std::initializer_list<T> data,
    Dtype dtype /* = TypeToDtype<T>() */)
-    : array_desc_(std::make_shared<ArrayDesc>(
-          Shape{static_cast<ShapeElem>(data.size())},
-          dtype)) {
+    : array_desc_(
+          std::make_shared<ArrayDesc>(
+              Shape{static_cast<ShapeElem>(data.size())},
+              dtype)) {
  init(data.begin());
 }

@@ -38,20 +38,20 @@ inline void set_binary_op_output_data(
    const array& a,
    const array& b,
    array& out,
-    BinaryOpType bopt) {
+    BinaryOpType bopt,
+    std::function<allocator::Buffer(size_t)> mallocfn = allocator::malloc) {
  bool b_donatable = is_donatable(b, out);
  bool a_donatable = is_donatable(a, out);
  switch (bopt) {
    case BinaryOpType::ScalarScalar:
-      out.set_data(
-          allocator::malloc(out.itemsize()), 1, a.strides(), a.flags());
+      out.set_data(mallocfn(out.itemsize()), 1, a.strides(), a.flags());
      break;
    case BinaryOpType::ScalarVector:
      if (b_donatable) {
        out.copy_shared_buffer(b);
      } else {
        out.set_data(
-            allocator::malloc(b.data_size() * out.itemsize()),
+            mallocfn(b.data_size() * out.itemsize()),
            b.data_size(),
            b.strides(),
            b.flags());
@@ -62,7 +62,7 @@ inline void set_binary_op_output_data(
        out.copy_shared_buffer(a);
      } else {
        out.set_data(
-            allocator::malloc(a.data_size() * out.itemsize()),
+            mallocfn(a.data_size() * out.itemsize()),
            a.data_size(),
            a.strides(),
            a.flags());
@@ -75,7 +75,7 @@ inline void set_binary_op_output_data(
        out.copy_shared_buffer(b);
      } else {
        out.set_data(
-            allocator::malloc(a.data_size() * out.itemsize()),
+            mallocfn(a.data_size() * out.itemsize()),
            a.data_size(),
            a.strides(),
            a.flags());
@@ -88,7 +88,7 @@ inline void set_binary_op_output_data(
          b_donatable && b.flags().row_contiguous && b.size() == out.size()) {
        out.copy_shared_buffer(b);
      } else {
-        out.set_data(allocator::malloc(out.nbytes()));
+        out.set_data(mallocfn(out.nbytes()));
      }
      break;
  }
@@ -6,7 +6,7 @@ namespace mlx::core {

 void broadcast(const array& in, array& out) {
  if (out.size() == 0) {
-    out.set_data(nullptr);
+    out.set_data(allocator::malloc(0));
    return;
  }
  Strides strides(out.ndim(), 0);
@@ -2,6 +2,7 @@

 #pragma once

+#include <algorithm>
 #include <cassert>
 #include <functional>
 #include <map>
@@ -114,7 +114,9 @@ void compiled_allocate_outputs(
    const std::vector<array>& inputs,
    std::vector<array>& outputs,
    const std::function<bool(size_t)>& is_constant,
-    bool contiguous) {
+    bool contiguous,
+    const std::function<allocator::Buffer(size_t)>&
+        mallocfn /* = allocator::malloc */) {
  if (contiguous) {
    int o = 0;
    Strides strides;
@@ -128,7 +130,7 @@ void compiled_allocate_outputs(
      // - Donatable
      // - Not a constant
      if (in.itemsize() == outputs[o].itemsize() && !is_scalar(in) &&
-          in.is_donatable() && is_constant(i)) {
+          in.is_donatable() && !is_constant(i)) {
        outputs[o++].copy_shared_buffer(in);
      }
      // Get representative input flags to properly set non-donated outputs
@@ -140,7 +142,7 @@ void compiled_allocate_outputs(
    }
    for (; o < outputs.size(); ++o) {
      outputs[o].set_data(
-          allocator::malloc(data_size * outputs[o].itemsize()),
+          mallocfn(data_size * outputs[o].itemsize()),
          data_size,
          strides,
          flags);
@@ -156,14 +158,14 @@ void compiled_allocate_outputs(
      // - Not a constant
      if (in.flags().row_contiguous && in.size() == outputs[o].size() &&
          in.itemsize() == outputs[o].itemsize() && in.is_donatable() &&
-          is_constant(i)) {
+          !is_constant(i)) {
        outputs[o].copy_shared_buffer(
            in, outputs[o].strides(), in.flags(), in.data_size());
        o++;
      }
    }
    for (; o < outputs.size(); ++o) {
-      outputs[o].set_data(allocator::malloc(outputs[o].nbytes()));
+      outputs[o].set_data(mallocfn(outputs[o].nbytes()));
    }
  }
 }
@@ -58,7 +58,9 @@ void compiled_allocate_outputs(
    const std::vector<array>& inputs,
    std::vector<array>& outputs,
    const std::function<bool(size_t)>& is_constant,
-    bool contiguous);
+    bool contiguous,
+    const std::function<allocator::Buffer(size_t)>& mallocfn =
+        allocator::malloc);

 // Collapse contiguous dims ignoring scalars and constants.
 std::tuple<bool, Shape, std::vector<Strides>> compiled_collapse_contiguous_dims(
@@ -22,7 +22,11 @@ enum class CopyType {
  GeneralGeneral
 };

-inline bool set_copy_output_data(const array& in, array& out, CopyType ctype) {
+inline bool set_copy_output_data(
+    const array& in,
+    array& out,
+    CopyType ctype,
+    std::function<allocator::Buffer(size_t)> mallocfn = allocator::malloc) {
  if (ctype == CopyType::Vector) {
    // If the input is donateable, we are doing a vector copy and the types
    // have the same size, then the input buffer can hold the output.
@@ -31,14 +35,14 @@ inline bool set_copy_output_data(const array& in, array& out, CopyType ctype) {
      return true;
    } else {
      out.set_data(
-          allocator::malloc(in.data_size() * out.itemsize()),
+          mallocfn(in.data_size() * out.itemsize()),
          in.data_size(),
          in.strides(),
          in.flags());
      return false;
    }
  } else {
-    out.set_data(allocator::malloc(out.nbytes()));
+    out.set_data(mallocfn(out.nbytes()));
    return false;
  }
 }
@@ -14,17 +14,13 @@ std::tuple<int64_t, Strides> prepare_slice(
    data_offset += start_indices[i] * in.strides()[i];
    inp_strides[i] = in.strides()[i] * strides[i];
  }
-  // Normalize the offset
-  if (data_offset < 0) {
-    data_offset += in.data_size();
-  }
  return std::make_tuple(data_offset, inp_strides);
 }

 void shared_buffer_slice(
    const array& in,
    const Strides& out_strides,
-    size_t data_offset,
+    int64_t data_offset,
    size_t data_size,
    array& out) {
  // Compute row/col contiguity
@@ -45,23 +41,30 @@ void slice(
    const Shape& start_indices,
    const Shape& strides) {
  if (out.size() == 0) {
-    out.set_data(nullptr);
+    out.set_data(allocator::malloc(0));
    return;
  }

  // Calculate out strides, initial offset
  auto [data_offset, inp_strides] = prepare_slice(in, start_indices, strides);
-  int64_t data_end = 1;
-  for (int i = 0; i < start_indices.size(); ++i) {
-    if (in.shape()[i] > 1) {
-      auto end_idx = start_indices[i] + out.shape()[i] * strides[i] - 1;
-      data_end += end_idx * in.strides()[i];
+
+  // Get the location of the end based on the inp strides and out.shape()
+  int64_t low_idx = 0;
+  int64_t high_idx = 0;
+  for (int i = 0; i < inp_strides.size(); ++i) {
+    auto delta = inp_strides[i] * (out.shape()[i] - 1);
+    if (inp_strides[i] > 0) {
+      high_idx += delta;
+    } else {
+      low_idx += delta;
    }
  }
-  if (data_end < 0) {
-    data_end += in.data_size();
+  int64_t data_size = (high_idx - low_idx) + 1;
+  if (data_size < 0) {
+    std::ostringstream msg;
+    msg << "[slice] Computed invalid data size: " << data_size << ".";
+    throw std::runtime_error(msg.str());
  }
-  size_t data_size = (data_end - data_offset);
  shared_buffer_slice(in, inp_strides, data_offset, data_size, out);
 }

@@ -46,7 +46,8 @@ inline void set_ternary_op_output_data(
    const array& b,
    const array& c,
    array& out,
-    TernaryOpType topt) {
+    TernaryOpType topt,
+    std::function<allocator::Buffer(size_t)> mallocfn = allocator::malloc) {
  auto maybe_donate = [&out](const array& x) {
    if (is_donatable(x, out)) {
      out.copy_shared_buffer(x);
@@ -57,13 +58,12 @@ inline void set_ternary_op_output_data(

  switch (topt) {
    case TernaryOpType::ScalarScalarScalar:
-      out.set_data(
-          allocator::malloc(out.itemsize()), 1, b.strides(), b.flags());
+      out.set_data(mallocfn(out.itemsize()), 1, b.strides(), b.flags());
      break;
    case TernaryOpType::VectorVectorVector:
      if (!(maybe_donate(a) || maybe_donate(b) || maybe_donate(c))) {
        out.set_data(
-            allocator::malloc(out.itemsize() * b.data_size()),
+            mallocfn(out.itemsize() * b.data_size()),
            b.data_size(),
            b.strides(),
            b.flags());
@@ -76,7 +76,7 @@ inline void set_ternary_op_output_data(
      if (!((a.flags().row_contiguous && maybe_donate(a)) ||
            (b.flags().row_contiguous && maybe_donate(b)) ||
            (c.flags().row_contiguous && maybe_donate(c)))) {
-        out.set_data(allocator::malloc(out.nbytes()));
+        out.set_data(mallocfn(out.nbytes()));
      }
      break;
  }
@@ -7,19 +7,22 @@

 namespace mlx::core {

-inline void set_unary_output_data(const array& in, array& out) {
+inline void set_unary_output_data(
+    const array& in,
+    array& out,
+    std::function<allocator::Buffer(size_t)> mallocfn = allocator::malloc) {
  if (in.flags().contiguous) {
    if (is_donatable(in, out)) {
      out.copy_shared_buffer(in);
    } else {
      out.set_data(
-          allocator::malloc(in.data_size() * out.itemsize()),
+          mallocfn(in.data_size() * out.itemsize()),
          in.data_size(),
          in.strides(),
          in.flags());
    }
  } else {
-    out.set_data(allocator::malloc(out.nbytes()));
+    out.set_data(mallocfn(out.nbytes()));
  }
 }

@@ -40,7 +40,7 @@ add_dependencies(mlx cpu_compiled_preamble)

 target_sources(
  mlx
-  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/available.cpp
+  PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/device_info.cpp
          ${CMAKE_CURRENT_SOURCE_DIR}/arg_reduce.cpp
          ${CMAKE_CURRENT_SOURCE_DIR}/binary.cpp
          ${CMAKE_CURRENT_SOURCE_DIR}/conv.cpp
@@ -1,11 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#include "mlx/backend/cpu/available.h"
-
-namespace mlx::core::cpu {
-
-bool is_available() {
-  return true;
-}
-
-} // namespace mlx::core::cpu
@@ -1,9 +0,0 @@
-// Copyright © 2025 Apple Inc.
-
-#pragma once
-
-namespace mlx::core::cpu {
-
-bool is_available();
-
-} // namespace mlx::core::cpu
@@ -14,233 +14,11 @@

 namespace mlx::core {

-namespace {
-
-template <typename Op>
-void binary(const array& a, const array& b, array& out, Op op, Stream stream) {
-  auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, out, bopt);
-
-  auto& encoder = cpu::get_command_encoder(stream);
-  encoder.set_input_array(a);
-  encoder.set_input_array(b);
-  encoder.set_output_array(out);
-  encoder.dispatch([a = array::unsafe_weak_copy(a),
-                    b = array::unsafe_weak_copy(b),
-                    out = array::unsafe_weak_copy(out),
-                    bopt]() mutable {
-    switch (out.dtype()) {
-      case bool_:
-        binary_op<bool, Op>(a, b, out, bopt);
-        break;
-      case uint8:
-        binary_op<uint8_t, Op>(a, b, out, bopt);
-        break;
-      case uint16:
-        binary_op<uint16_t, Op>(a, b, out, bopt);
-        break;
-      case uint32:
-        binary_op<uint32_t, Op>(a, b, out, bopt);
-        break;
-      case uint64:
-        binary_op<uint64_t, Op>(a, b, out, bopt);
-        break;
-      case int8:
-        binary_op<int8_t, Op>(a, b, out, bopt);
-        break;
-      case int16:
-        binary_op<int16_t, Op>(a, b, out, bopt);
-        break;
-      case int32:
-        binary_op<int32_t, Op>(a, b, out, bopt);
-        break;
-      case int64:
-        binary_op<int64_t, Op>(a, b, out, bopt);
-        break;
-      case float16:
-        binary_op<float16_t, Op>(a, b, out, bopt);
-        break;
-      case float32:
-        binary_op<float, Op>(a, b, out, bopt);
-        break;
-      case float64:
-        binary_op<double, Op>(a, b, out, bopt);
-        break;
-      case bfloat16:
-        binary_op<bfloat16_t, Op>(a, b, out, bopt);
-        break;
-      case complex64:
-        binary_op<complex64_t, Op>(a, b, out, bopt);
-        break;
-    }
-  });
-}
-
-template <typename Op>
-void comparison_op(
-    const array& a,
-    const array& b,
-    array& out,
-    Op op,
-    Stream stream) {
-  auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, out, bopt);
-
-  auto& encoder = cpu::get_command_encoder(stream);
-  encoder.set_input_array(a);
-  encoder.set_input_array(b);
-  encoder.set_output_array(out);
-  encoder.dispatch([a = array::unsafe_weak_copy(a),
-                    b = array::unsafe_weak_copy(b),
-                    out = array::unsafe_weak_copy(out),
-                    bopt]() mutable {
-    switch (a.dtype()) {
-      case bool_:
-        binary_op<bool, bool, Op>(a, b, out, bopt);
-        break;
-      case uint8:
-        binary_op<uint8_t, bool, Op>(a, b, out, bopt);
-        break;
-      case uint16:
-        binary_op<uint16_t, bool, Op>(a, b, out, bopt);
-        break;
-      case uint32:
-        binary_op<uint32_t, bool, Op>(a, b, out, bopt);
-        break;
-      case uint64:
-        binary_op<uint64_t, bool, Op>(a, b, out, bopt);
-        break;
-      case int8:
-        binary_op<int8_t, bool, Op>(a, b, out, bopt);
-        break;
-      case int16:
-        binary_op<int16_t, bool, Op>(a, b, out, bopt);
-        break;
-      case int32:
-        binary_op<int32_t, bool, Op>(a, b, out, bopt);
-        break;
-      case int64:
-        binary_op<int64_t, bool, Op>(a, b, out, bopt);
-        break;
-      case float16:
-        binary_op<float16_t, bool, Op>(a, b, out, bopt);
-        break;
-      case float32:
-        binary_op<float, bool, Op>(a, b, out, bopt);
-        break;
-      case float64:
-        binary_op<double, bool, Op>(a, b, out, bopt);
-        break;
-      case bfloat16:
-        binary_op<bfloat16_t, bool, Op>(a, b, out, bopt);
-        break;
-      case complex64:
-        binary_op<complex64_t, bool, Op>(a, b, out, bopt);
-        break;
-    }
-  });
-}
-
-template <typename Op>
-void binary_float(
-    const array& a,
-    const array& b,
-    array& out,
-    Op op,
-    Stream stream) {
-  auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, out, bopt);
-
-  auto& encoder = cpu::get_command_encoder(stream);
-  encoder.set_input_array(a);
-  encoder.set_input_array(b);
-  encoder.set_output_array(out);
-  encoder.dispatch([a = array::unsafe_weak_copy(a),
-                    b = array::unsafe_weak_copy(b),
-                    out = array::unsafe_weak_copy(out),
-                    bopt]() mutable {
-    switch (out.dtype()) {
-      case float16:
-        binary_op<float16_t, Op>(a, b, out, bopt);
-        break;
-      case float32:
-        binary_op<float, Op>(a, b, out, bopt);
-        break;
-      case float64:
-        binary_op<double, Op>(a, b, out, bopt);
-        break;
-      case bfloat16:
-        binary_op<bfloat16_t, Op>(a, b, out, bopt);
-        break;
-      case complex64:
-        binary_op<complex64_t, Op>(a, b, out, bopt);
-        break;
-      default:
-        throw std::runtime_error(
-            "[binary_float] Only supports floating point types.");
-    }
-  });
-}
-
-template <typename Op>
-void binary_int(
-    const array& a,
-    const array& b,
-    array& out,
-    Op op,
-    Stream stream) {
-  auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, out, bopt);
-
-  auto& encoder = cpu::get_command_encoder(stream);
-  encoder.set_input_array(a);
-  encoder.set_input_array(b);
-  encoder.set_output_array(out);
-  encoder.dispatch([a = array::unsafe_weak_copy(a),
-                    b = array::unsafe_weak_copy(b),
-                    out = array::unsafe_weak_copy(out),
-                    bopt]() mutable {
-    switch (out.dtype()) {
-      case bool_:
-        binary_op<bool, Op>(a, b, out, bopt);
-      case uint8:
-        binary_op<uint8_t, Op>(a, b, out, bopt);
-        break;
-      case uint16:
-        binary_op<uint16_t, Op>(a, b, out, bopt);
-        break;
-      case uint32:
-        binary_op<uint32_t, Op>(a, b, out, bopt);
-        break;
-      case uint64:
-        binary_op<uint64_t, Op>(a, b, out, bopt);
-        break;
-      case int8:
-        binary_op<int8_t, Op>(a, b, out, bopt);
-        break;
-      case int16:
-        binary_op<int16_t, Op>(a, b, out, bopt);
-        break;
-      case int32:
-        binary_op<int32_t, Op>(a, b, out, bopt);
-        break;
-      case int64:
-        binary_op<int64_t, Op>(a, b, out, bopt);
-        break;
-      default:
-        throw std::runtime_error("[binary_int] Type not supported");
-        break;
-    }
-  });
-}
-
-} // namespace
-
 void Add::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
  auto& a = inputs[0];
  auto& b = inputs[1];
-  binary(a, b, out, detail::Add(), stream());
+  binary_op_cpu(a, b, out, detail::Add(), stream());
 }

 void DivMod::eval_cpu(
@@ -324,14 +102,14 @@ void Divide::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
  auto& a = inputs[0];
  auto& b = inputs[1];
-  binary(a, b, out, detail::Divide(), stream());
+  binary_op_cpu(a, b, out, detail::Divide(), stream());
 }

 void Remainder::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
  auto& a = inputs[0];
  auto& b = inputs[1];
-  binary(a, b, out, detail::Remainder(), stream());
+  binary_op_cpu(a, b, out, detail::Remainder(), stream());
 }

 void Equal::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -372,89 +150,90 @@ void Equal::eval_cpu(const std::vector<array>& inputs, array& out) {
      }
    });
  } else {
-    comparison_op(a, b, out, detail::Equal(), stream());
+    comparison_op_cpu(a, b, out, detail::Equal(), stream());
  }
 }

 void Greater::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::Greater(), stream());
+  comparison_op_cpu(inputs[0], inputs[1], out, detail::Greater(), stream());
 }

 void GreaterEqual::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::GreaterEqual(), stream());
+  comparison_op_cpu(
+      inputs[0], inputs[1], out, detail::GreaterEqual(), stream());
 }

 void Less::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::Less(), stream());
+  comparison_op_cpu(inputs[0], inputs[1], out, detail::Less(), stream());
 }

 void LessEqual::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::LessEqual(), stream());
+  comparison_op_cpu(inputs[0], inputs[1], out, detail::LessEqual(), stream());
 }

 void LogAddExp::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
  auto& a = inputs[0];
  auto& b = inputs[1];
-  binary_float(a, b, out, detail::LogAddExp(), stream());
+  binary_float_op_cpu(a, b, out, detail::LogAddExp(), stream());
 }

 void LogicalAnd::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2); // LogicalAnd requires two input arrays
  auto& in1 = inputs[0];
  auto& in2 = inputs[1];
-  binary(in1, in2, out, detail::LogicalAnd(), stream());
+  binary_op_cpu(in1, in2, out, detail::LogicalAnd(), stream());
 }

 void LogicalOr::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2); // LogicalOr requires two input arrays
  auto& in1 = inputs[0];
  auto& in2 = inputs[1];
-  binary(in1, in2, out, detail::LogicalOr(), stream());
+  binary_op_cpu(in1, in2, out, detail::LogicalOr(), stream());
 }

 void Maximum::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
  auto& a = inputs[0];
  auto& b = inputs[1];
-  binary(a, b, out, detail::Maximum(), stream());
+  binary_op_cpu(a, b, out, detail::Maximum(), stream());
 }

 void Minimum::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
  auto& a = inputs[0];
  auto& b = inputs[1];
-  binary(a, b, out, detail::Minimum(), stream());
+  binary_op_cpu(a, b, out, detail::Minimum(), stream());
 }

 void Multiply::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
  auto& a = inputs[0];
  auto& b = inputs[1];
-  binary(a, b, out, detail::Multiply(), stream());
+  binary_op_cpu(a, b, out, detail::Multiply(), stream());
 }

 void NotEqual::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
-  comparison_op(inputs[0], inputs[1], out, detail::NotEqual(), stream());
+  comparison_op_cpu(inputs[0], inputs[1], out, detail::NotEqual(), stream());
 }

 void Power::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
  auto& a = inputs[0];
  auto& b = inputs[1];
-  binary(a, b, out, detail::Power(), stream());
+  binary_op_cpu(a, b, out, detail::Power(), stream());
 }

 void Subtract::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
  auto& a = inputs[0];
  auto& b = inputs[1];
-  binary(a, b, out, detail::Subtract(), stream());
+  binary_op_cpu(a, b, out, detail::Subtract(), stream());
 }

 void BitwiseBinary::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -463,19 +242,19 @@ void BitwiseBinary::eval_cpu(const std::vector<array>& inputs, array& out) {
  auto& b = inputs[1];
  switch (op_) {
    case BitwiseBinary::And:
-      binary_int(a, b, out, detail::BitwiseAnd(), stream());
+      binary_int_op_cpu(a, b, out, detail::BitwiseAnd(), stream());
      break;
    case BitwiseBinary::Or:
-      binary_int(a, b, out, detail::BitwiseOr(), stream());
+      binary_int_op_cpu(a, b, out, detail::BitwiseOr(), stream());
      break;
    case BitwiseBinary::Xor:
-      binary_int(a, b, out, detail::BitwiseXor(), stream());
+      binary_int_op_cpu(a, b, out, detail::BitwiseXor(), stream());
      break;
    case BitwiseBinary::LeftShift:
-      binary_int(a, b, out, detail::LeftShift(), stream());
+      binary_int_op_cpu(a, b, out, detail::LeftShift(), stream());
      break;
    case BitwiseBinary::RightShift:
-      binary_int(a, b, out, detail::RightShift(), stream());
+      binary_int_op_cpu(a, b, out, detail::RightShift(), stream());
      break;
  }
 }
@@ -484,7 +263,7 @@ void ArcTan2::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
  const auto& a = inputs[0];
  const auto& b = inputs[1];
-  binary_float(a, b, out, detail::ArcTan2(), stream());
+  binary_float_op_cpu(a, b, out, detail::ArcTan2(), stream());
 }

 } // namespace mlx::core
@@ -7,6 +7,7 @@
 #include "mlx/backend/common/binary.h"
 #include "mlx/backend/common/utils.h"

+#include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"

 namespace mlx::core {
@@ -290,4 +291,227 @@ void binary_op(const array& a, const array& b, array& out, BinaryOpType bopt) {
  binary_op<T, T, Op>(a, b, out, bopt);
 }

+template <typename Op>
+void binary_op_cpu(
+    const array& a,
+    const array& b,
+    array& out,
+    Op op,
+    Stream stream) {
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out = array::unsafe_weak_copy(out),
+                    bopt]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        binary_op<bool, Op>(a, b, out, bopt);
+        break;
+      case uint8:
+        binary_op<uint8_t, Op>(a, b, out, bopt);
+        break;
+      case uint16:
+        binary_op<uint16_t, Op>(a, b, out, bopt);
+        break;
+      case uint32:
+        binary_op<uint32_t, Op>(a, b, out, bopt);
+        break;
+      case uint64:
+        binary_op<uint64_t, Op>(a, b, out, bopt);
+        break;
+      case int8:
+        binary_op<int8_t, Op>(a, b, out, bopt);
+        break;
+      case int16:
+        binary_op<int16_t, Op>(a, b, out, bopt);
+        break;
+      case int32:
+        binary_op<int32_t, Op>(a, b, out, bopt);
+        break;
+      case int64:
+        binary_op<int64_t, Op>(a, b, out, bopt);
+        break;
+      case float16:
+        binary_op<float16_t, Op>(a, b, out, bopt);
+        break;
+      case float32:
+        binary_op<float, Op>(a, b, out, bopt);
+        break;
+      case float64:
+        binary_op<double, Op>(a, b, out, bopt);
+        break;
+      case bfloat16:
+        binary_op<bfloat16_t, Op>(a, b, out, bopt);
+        break;
+      case complex64:
+        binary_op<complex64_t, Op>(a, b, out, bopt);
+        break;
+    }
+  });
+}
+
+template <typename Op>
+void comparison_op_cpu(
+    const array& a,
+    const array& b,
+    array& out,
+    Op op,
+    Stream stream) {
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out = array::unsafe_weak_copy(out),
+                    bopt]() mutable {
+    switch (a.dtype()) {
+      case bool_:
+        binary_op<bool, bool, Op>(a, b, out, bopt);
+        break;
+      case uint8:
+        binary_op<uint8_t, bool, Op>(a, b, out, bopt);
+        break;
+      case uint16:
+        binary_op<uint16_t, bool, Op>(a, b, out, bopt);
+        break;
+      case uint32:
+        binary_op<uint32_t, bool, Op>(a, b, out, bopt);
+        break;
+      case uint64:
+        binary_op<uint64_t, bool, Op>(a, b, out, bopt);
+        break;
+      case int8:
+        binary_op<int8_t, bool, Op>(a, b, out, bopt);
+        break;
+      case int16:
+        binary_op<int16_t, bool, Op>(a, b, out, bopt);
+        break;
+      case int32:
+        binary_op<int32_t, bool, Op>(a, b, out, bopt);
+        break;
+      case int64:
+        binary_op<int64_t, bool, Op>(a, b, out, bopt);
+        break;
+      case float16:
+        binary_op<float16_t, bool, Op>(a, b, out, bopt);
+        break;
+      case float32:
+        binary_op<float, bool, Op>(a, b, out, bopt);
+        break;
+      case float64:
+        binary_op<double, bool, Op>(a, b, out, bopt);
+        break;
+      case bfloat16:
+        binary_op<bfloat16_t, bool, Op>(a, b, out, bopt);
+        break;
+      case complex64:
+        binary_op<complex64_t, bool, Op>(a, b, out, bopt);
+        break;
+    }
+  });
+}
+
+template <typename Op>
+void binary_float_op_cpu(
+    const array& a,
+    const array& b,
+    array& out,
+    Op op,
+    Stream stream) {
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out = array::unsafe_weak_copy(out),
+                    bopt]() mutable {
+    switch (out.dtype()) {
+      case float16:
+        binary_op<float16_t, Op>(a, b, out, bopt);
+        break;
+      case float32:
+        binary_op<float, Op>(a, b, out, bopt);
+        break;
+      case float64:
+        binary_op<double, Op>(a, b, out, bopt);
+        break;
+      case bfloat16:
+        binary_op<bfloat16_t, Op>(a, b, out, bopt);
+        break;
+      case complex64:
+        binary_op<complex64_t, Op>(a, b, out, bopt);
+        break;
+      default:
+        throw std::runtime_error(
+            "[binary_float] Only supports floating point types.");
+    }
+  });
+}
+
+template <typename Op>
+void binary_int_op_cpu(
+    const array& a,
+    const array& b,
+    array& out,
+    Op op,
+    Stream stream) {
+  auto bopt = get_binary_op_type(a, b);
+  set_binary_op_output_data(a, b, out, bopt);
+
+  auto& encoder = cpu::get_command_encoder(stream);
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    out = array::unsafe_weak_copy(out),
+                    bopt]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        binary_op<bool, Op>(a, b, out, bopt);
+      case uint8:
+        binary_op<uint8_t, Op>(a, b, out, bopt);
+        break;
+      case uint16:
+        binary_op<uint16_t, Op>(a, b, out, bopt);
+        break;
+      case uint32:
+        binary_op<uint32_t, Op>(a, b, out, bopt);
+        break;
+      case uint64:
+        binary_op<uint64_t, Op>(a, b, out, bopt);
+        break;
+      case int8:
+        binary_op<int8_t, Op>(a, b, out, bopt);
+        break;
+      case int16:
+        binary_op<int16_t, Op>(a, b, out, bopt);
+        break;
+      case int32:
+        binary_op<int32_t, Op>(a, b, out, bopt);
+        break;
+      case int64:
+        binary_op<int64_t, Op>(a, b, out, bopt);
+        break;
+      default:
+        throw std::runtime_error("[binary_int] Type not supported");
+        break;
+    }
+  });
+}
+
 } // namespace mlx::core
@@ -119,13 +119,15 @@ void* compile(
    source_file.close();

    try {
-      JitCompiler::exec(JitCompiler::build_command(
-          output_dir, source_file_name, shared_lib_name));
+      JitCompiler::exec(
+          JitCompiler::build_command(
+              output_dir, source_file_name, shared_lib_name));
    } catch (const std::exception& error) {
-      throw std::runtime_error(fmt::format(
-          "[Compile::eval_cpu] Failed to compile function {0}: {1}",
-          kernel_name,
-          error.what()));
+      throw std::runtime_error(
+          fmt::format(
+              "[Compile::eval_cpu] Failed to compile function {0}: {1}",
+              kernel_name,
+              error.what()));
    }
  }

@@ -0,0 +1,113 @@
+// Copyright © 2026 Apple Inc.
+
+#include "mlx/backend/cpu/device_info.h"
+
+#ifdef __APPLE__
+#include <sys/sysctl.h>
+#include <sys/utsname.h>
+#elif defined(_WIN32)
+#include <windows.h>
+#else
+#include <sys/utsname.h>
+#include <fstream>
+#endif
+
+namespace mlx::core::cpu {
+
+namespace {
+
+// Get CPU architecture string at runtime
+std::string get_cpu_architecture() {
+#ifdef _WIN32
+  // Use GetNativeSystemInfo to get the actual hardware architecture,
+  // even when running under WoW64 emulation
+  SYSTEM_INFO sysInfo;
+  GetNativeSystemInfo(&sysInfo);
+  switch (sysInfo.wProcessorArchitecture) {
+    case PROCESSOR_ARCHITECTURE_AMD64:
+      return "x86_64";
+    case PROCESSOR_ARCHITECTURE_ARM64:
+      return "arm64";
+    case PROCESSOR_ARCHITECTURE_INTEL:
+      return "x86";
+    case PROCESSOR_ARCHITECTURE_ARM:
+      return "arm";
+    default:
+      return "unknown";
+  }
+#else
+  // Use uname() for runtime detection on Unix-like systems.
+  // This returns the actual hardware architecture (e.g., "arm64" on Apple
+  // Silicon even when running x86_64 binaries via Rosetta 2)
+  struct utsname info;
+  if (uname(&info) == 0) {
+    return std::string(info.machine);
+  }
+  return "unknown";
+#endif
+}
+
+// Get CPU device name (brand string)
+std::string get_cpu_name() {
+#ifdef __APPLE__
+  char model[256];
+  size_t len = sizeof(model);
+  if (sysctlbyname("machdep.cpu.brand_string", &model, &len, NULL, 0) == 0) {
+    return std::string(model);
+  }
+#elif defined(_WIN32)
+  // Read CPU brand string from registry
+  HKEY hKey;
+  if (RegOpenKeyExA(
+          HKEY_LOCAL_MACHINE,
+          "HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0",
+          0,
+          KEY_READ,
+          &hKey) == ERROR_SUCCESS) {
+    char brand[256];
+    DWORD size = sizeof(brand);
+    if (RegQueryValueExA(
+            hKey, "ProcessorNameString", NULL, NULL, (LPBYTE)brand, &size) ==
+        ERROR_SUCCESS) {
+      RegCloseKey(hKey);
+      return std::string(brand);
+    }
+    RegCloseKey(hKey);
+  }
+#else
+  // Try reading from /proc/cpuinfo on Linux
+  std::ifstream cpuinfo("/proc/cpuinfo");
+  if (cpuinfo.is_open()) {
+    std::string line;
+    while (std::getline(cpuinfo, line)) {
+      if (line.starts_with("model name")) {
+        if (auto n = line.find(": "); n != std::string::npos) {
+          return line.substr(n + 2);
+        }
+      }
+    }
+  }
+#endif
+  return get_cpu_architecture();
+}
+
+} // anonymous namespace
+
+bool is_available() {
+  return true;
+}
+
+int device_count() {
+  return 1;
+}
+
+const std::unordered_map<std::string, std::variant<std::string, size_t>>&
+device_info(int /* device_index */) {
+  static auto info =
+      std::unordered_map<std::string, std::variant<std::string, size_t>>{
+          {"device_name", get_cpu_name()},
+          {"architecture", get_cpu_architecture()}};
+  return info;
+}
+
+} // namespace mlx::core::cpu
@@ -0,0 +1,28 @@
+// Copyright © 2026 Apple Inc.
+
+#pragma once
+
+#include <string>
+#include <unordered_map>
+#include <variant>
+
+namespace mlx::core::cpu {
+
+bool is_available();
+
+/**
+ * Get the number of available CPU devices.
+ *
+ * For CPU, always returns 1.
+ */
+int device_count();
+
+/**
+ * Get CPU device information.
+ *
+ * Returns a map with basic CPU device properties.
+ */
+const std::unordered_map<std::string, std::variant<std::string, size_t>>&
+device_info(int device_index = 0);
+
+} // namespace mlx::core::cpu
@@ -95,4 +95,9 @@ void Recv::eval_cpu(
  distributed::detail::recv(group(), outputs[0], src_, stream());
 }

+void ReduceScatter::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  throw std::runtime_error("[ReduceScatter] Not implemented yet.");
+}
 } // namespace mlx::core::distributed
@@ -12,6 +12,167 @@ namespace mlx::core {

 namespace {

+template <typename T>
+complex64_t to_complex(T r, T i) {
+  return {static_cast<float>(r), static_cast<float>(i)};
+}
+
+template <typename T, class Enable = void>
+struct EigWork {};
+
+template <typename T>
+struct EigWork<
+    T,
+    typename std::enable_if<std::is_floating_point<T>::value>::type> {
+  using O = complex64_t;
+
+  char jobl;
+  char jobr;
+  int N;
+  int lwork;
+  int info;
+  std::vector<array::Data> buffers;
+
+  EigWork(char jobl_, char jobr_, int N_, bool compute_eigenvectors)
+      : jobl(jobl_), jobr(jobr_), N(N_), lwork(-1) {
+    T work;
+    int n_vecs_l = compute_eigenvectors ? N_ : 1;
+    int n_vecs_r = 1;
+    geev<T>(
+        &jobl,
+        &jobr,
+        &N,
+        nullptr,
+        &N,
+        nullptr,
+        nullptr,
+        nullptr,
+        &n_vecs_l,
+        nullptr,
+        &n_vecs_r,
+        &work,
+        &lwork,
+        &info);
+    lwork = static_cast<int>(work);
+
+    buffers.emplace_back(allocator::malloc(sizeof(T) * N * 2));
+    if (compute_eigenvectors) {
+      buffers.emplace_back(allocator::malloc(sizeof(T) * N * N * 2));
+    }
+    buffers.emplace_back(allocator::malloc(sizeof(T) * lwork));
+  }
+
+  void run(T* a, O* values, O* vectors) {
+    auto eig_tmp = static_cast<T*>(buffers[0].buffer.raw_ptr());
+    T* vec_tmp = nullptr;
+    if (vectors) {
+      vec_tmp = static_cast<T*>(buffers[1].buffer.raw_ptr());
+    }
+    auto work = static_cast<T*>(buffers.back().buffer.raw_ptr());
+
+    int n_vecs_l = vectors ? N : 1;
+    int n_vecs_r = 1;
+    geev<T>(
+        &jobl,
+        &jobr,
+        &N,
+        a,
+        &N,
+        eig_tmp,
+        eig_tmp + N,
+        vectors ? vec_tmp : nullptr,
+        &n_vecs_l,
+        nullptr,
+        &n_vecs_r,
+        work,
+        &lwork,
+        &info);
+
+    for (int i = 0; i < N; ++i) {
+      values[i] = to_complex(eig_tmp[i], eig_tmp[N + i]);
+    }
+
+    if (vectors) {
+      for (int i = 0; i < N; ++i) {
+        if (values[i].imag() != 0) {
+          for (int j = 0; j < N; ++j) {
+            vectors[i * N + j] =
+                to_complex(vec_tmp[i * N + j], -vec_tmp[(i + 1) * N + j]);
+            vectors[(i + 1) * N + j] =
+                to_complex(vec_tmp[i * N + j], vec_tmp[(i + 1) * N + j]);
+          }
+          i += 1;
+        } else {
+          for (int j = 0; j < N; ++j) {
+            vectors[i * N + j] = to_complex(vec_tmp[i * N + j], T(0.0));
+          }
+        }
+      }
+    }
+  }
+};
+
+template <>
+struct EigWork<std::complex<float>> {
+  using T = std::complex<float>;
+  using R = float;
+  using O = T;
+
+  char jobl;
+  char jobr;
+  int N;
+  int lwork;
+  int lrwork;
+  int info;
+  std::vector<array::Data> buffers;
+
+  EigWork(char jobl_, char jobr_, int N_, bool compute_eigenvectors)
+      : jobl(jobl_), jobr(jobr_), N(N_), lwork(-1), lrwork(2 * N_) {
+    T work;
+    R rwork;
+    int n_vecs_l = compute_eigenvectors ? N_ : 1;
+    int n_vecs_r = 1;
+    geev<T>(
+        &jobl,
+        &jobr,
+        &N,
+        nullptr,
+        &N,
+        nullptr,
+        nullptr,
+        &n_vecs_l,
+        nullptr,
+        &n_vecs_r,
+        &work,
+        &lwork,
+        &rwork,
+        &info);
+    lwork = static_cast<int>(work.real());
+    buffers.emplace_back(allocator::malloc(sizeof(T) * lwork));
+    buffers.emplace_back(allocator::malloc(sizeof(R) * lrwork));
+  }
+
+  void run(T* a, T* values, T* vectors) {
+    int n_vecs_l = vectors ? N : 1;
+    int n_vecs_r = 1;
+    geev<T>(
+        &jobl,
+        &jobr,
+        &N,
+        a,
+        &N,
+        values,
+        vectors,
+        &n_vecs_l,
+        nullptr,
+        &n_vecs_r,
+        static_cast<T*>(buffers[0].buffer.raw_ptr()),
+        &lwork,
+        static_cast<R*>(buffers[1].buffer.raw_ptr()),
+        &info);
+  }
+};
+
 template <typename T>
 void eig_impl(
    array& a,
@@ -19,101 +180,39 @@ void eig_impl(
    array& values,
    bool compute_eigenvectors,
    Stream stream) {
-  using OT = std::complex<T>;
  auto a_ptr = a.data<T>();
-  auto eig_ptr = values.data<OT>();
+  auto val_ptr = values.data<complex64_t>();

  auto& encoder = cpu::get_command_encoder(stream);
  encoder.set_input_array(a);
  encoder.set_output_array(values);
-  OT* vec_ptr = nullptr;
+  complex64_t* vec_ptr = nullptr;
  if (compute_eigenvectors) {
    encoder.set_output_array(vectors);
-    vec_ptr = vectors.data<OT>();
+    vec_ptr = vectors.data<complex64_t>();
  }
  encoder.dispatch([a_ptr,
+                    val_ptr,
                    vec_ptr,
-                    eig_ptr,
                    compute_eigenvectors,
                    N = vectors.shape(-1),
                    size = vectors.size()]() mutable {
-    // Work query
    char jobr = 'N';
    char jobl = compute_eigenvectors ? 'V' : 'N';
-    int n_vecs_r = 1;
-    int n_vecs_l = compute_eigenvectors ? N : 1;
-    int lwork = -1;
-    int info;
-    {
-      T work;
-      geev<T>(
-          &jobl,
-          &jobr,
-          &N,
-          nullptr,
-          &N,
-          nullptr,
-          nullptr,
-          nullptr,
-          &n_vecs_l,
-          nullptr,
-          &n_vecs_r,
-          &work,
-          &lwork,
-          &info);
-      lwork = static_cast<int>(work);
-    }

-    auto eig_tmp_data = array::Data{allocator::malloc(sizeof(T) * N * 2)};
-    auto vec_tmp_data =
-        array::Data{allocator::malloc(vec_ptr ? sizeof(T) * N * N * 2 : 0)};
-    auto eig_tmp = static_cast<T*>(eig_tmp_data.buffer.raw_ptr());
-    auto vec_tmp = static_cast<T*>(vec_tmp_data.buffer.raw_ptr());
-    auto work_buf = array::Data{allocator::malloc(sizeof(T) * lwork)};
+    EigWork<T> work(jobl, jobr, N, compute_eigenvectors);
+
    for (size_t i = 0; i < size / (N * N); ++i) {
-      geev<T>(
-          &jobl,
-          &jobr,
-          &N,
-          a_ptr,
-          &N,
-          eig_tmp,
-          eig_tmp + N,
-          vec_tmp,
-          &n_vecs_l,
-          nullptr,
-          &n_vecs_r,
-          static_cast<T*>(work_buf.buffer.raw_ptr()),
-          &lwork,
-          &info);
-      for (int i = 0; i < N; ++i) {
-        eig_ptr[i] = {eig_tmp[i], eig_tmp[N + i]};
-      }
+      work.run(a_ptr, val_ptr, vec_ptr);
+      a_ptr += N * N;
+      val_ptr += N;
      if (vec_ptr) {
-        for (int i = 0; i < N; ++i) {
-          if (eig_ptr[i].imag() != 0) {
-            // This vector and the next are a pair
-            for (int j = 0; j < N; ++j) {
-              vec_ptr[i * N + j] = {
-                  vec_tmp[i * N + j], -vec_tmp[(i + 1) * N + j]};
-              vec_ptr[(i + 1) * N + j] = {
-                  vec_tmp[i * N + j], vec_tmp[(i + 1) * N + j]};
-            }
-            i += 1;
-          } else {
-            for (int j = 0; j < N; ++j) {
-              vec_ptr[i * N + j] = {vec_tmp[i * N + j], 0};
-            }
-          }
-        }
        vec_ptr += N * N;
      }
-      a_ptr += N * N;
-      eig_ptr += N;
-      if (info != 0) {
+      if (work.info != 0) {
        std::stringstream msg;
        msg << "[Eig::eval_cpu] Eigenvalue decomposition failed with error code "
-            << info;
+            << work.info;
        throw std::runtime_error(msg.str());
      }
    }
@@ -165,8 +264,17 @@ void Eig::eval_cpu(
    case float32:
      eig_impl<float>(a_copy, vectors, values, compute_eigenvectors_, stream());
      break;
+    case float64:
+      eig_impl<double>(
+          a_copy, vectors, values, compute_eigenvectors_, stream());
+      break;
+    case complex64:
+      eig_impl<std::complex<float>>(
+          a_copy, vectors, values, compute_eigenvectors_, stream());
+      break;
    default:
-      throw std::runtime_error("[Eig::eval_cpu] only supports float32.");
+      throw std::runtime_error(
+          "[Eig::eval_cpu] only supports float32, float64, or complex64.");
  }
 }

@@ -12,7 +12,7 @@ namespace mlx::core::cpu {
 // Number of dispatches per scheduler task
 constexpr int DISPATCHES_PER_TASK = 10;

-struct CommandEncoder {
+struct MLX_API CommandEncoder {
  CommandEncoder(Stream stream) : stream_(stream) {}

  CommandEncoder(const CommandEncoder&) = delete;
@@ -62,6 +62,6 @@ struct CommandEncoder {
  int num_ops_{0};
 };

-CommandEncoder& get_command_encoder(Stream stream);
+MLX_API CommandEncoder& get_command_encoder(Stream stream);

 } // namespace mlx::core::cpu
@@ -747,4 +747,108 @@ void ScatterAxis::eval_cpu(const std::vector<array>& inputs, array& out) {
  });
 }

+template <typename T>
+void masked_scatter_impl(const array& mask, const array& src, array& out) {
+  ContiguousIterator mask_it(mask);
+  ContiguousIterator src_it(src);
+  ContiguousIterator out_it(out);
+
+  const bool* mask_ptr = mask.data<bool>();
+  const T* src_ptr = src.data<T>();
+  T* dst_ptr = out.data<T>();
+
+  const size_t batch_count = mask.shape(0);
+  const size_t mask_batch_size = mask.size() / batch_count;
+  const size_t src_batch_size = src.size() / batch_count;
+
+  for (size_t b = 0; b < batch_count; ++b) {
+    size_t src_consumed = 0;
+    src_it.seek(b * src_batch_size);
+
+    for (size_t i = 0; i < mask_batch_size; ++i) {
+      if (mask_ptr[mask_it.loc]) {
+        if (src_consumed >= src_batch_size) {
+          throw std::runtime_error(
+              "[MaskedScatter::eval_cpu] Source does not have enough elements for mask.");
+        }
+        dst_ptr[out_it.loc] = src_ptr[src_it.loc];
+        src_it.step();
+        ++src_consumed;
+      }
+      mask_it.step();
+      out_it.step();
+    }
+  }
+}
+
+void MaskedScatter::eval_cpu(const std::vector<array>& inputs, array& out) {
+  assert(inputs.size() == 3);
+
+  auto& dst = inputs[0];
+  auto& mask = inputs[1];
+  auto& src = inputs[2];
+
+  // Copy src into out (copy allocates memory for out)
+  auto ctype =
+      dst.flags().row_contiguous ? CopyType::Vector : CopyType::General;
+  copy_cpu(dst, out, ctype, stream());
+
+  if (mask.size() == 0) {
+    return;
+  }
+
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(mask);
+  encoder.set_input_array(src);
+  encoder.set_output_array(out);
+  encoder.dispatch([mask = array::unsafe_weak_copy(mask),
+                    src = array::unsafe_weak_copy(src),
+                    out = array::unsafe_weak_copy(out)]() mutable {
+    switch (out.dtype()) {
+      case bool_:
+        masked_scatter_impl<bool>(mask, src, out);
+        break;
+      case uint8:
+        masked_scatter_impl<uint8_t>(mask, src, out);
+        break;
+      case uint16:
+        masked_scatter_impl<uint16_t>(mask, src, out);
+        break;
+      case uint32:
+        masked_scatter_impl<uint32_t>(mask, src, out);
+        break;
+      case uint64:
+        masked_scatter_impl<uint64_t>(mask, src, out);
+        break;
+      case int8:
+        masked_scatter_impl<int8_t>(mask, src, out);
+        break;
+      case int16:
+        masked_scatter_impl<int16_t>(mask, src, out);
+        break;
+      case int32:
+        masked_scatter_impl<int32_t>(mask, src, out);
+        break;
+      case int64:
+        masked_scatter_impl<int64_t>(mask, src, out);
+        break;
+      case float16:
+        masked_scatter_impl<float16_t>(mask, src, out);
+        break;
+      case float32:
+        masked_scatter_impl<float>(mask, src, out);
+        break;
+      case float64:
+        masked_scatter_impl<double>(mask, src, out);
+        break;
+      case bfloat16:
+        masked_scatter_impl<bfloat16_t>(mask, src, out);
+        break;
+      case complex64:
+        masked_scatter_impl<complex64_t>(mask, src, out);
+        break;
+    }
+  });
+}
+
 } // namespace mlx::core
@@ -34,18 +34,30 @@ struct VisualStudioInfo {
    arch = "x64";
 #endif
    // Get path of Visual Studio.
-    std::string vs_path = JitCompiler::exec(fmt::format(
-        "\"{0}\\Microsoft Visual Studio\\Installer\\vswhere.exe\""
-        " -property installationPath",
-        std::getenv("ProgramFiles(x86)")));
+    // Use -latest to get only the most recent installation when multiple
+    // versions are installed, avoiding path concatenation issues.
+    std::string vs_path = JitCompiler::exec(
+        fmt::format(
+            "\"{0}\\Microsoft Visual Studio\\Installer\\vswhere.exe\""
+            " -latest -property installationPath",
+            std::getenv("ProgramFiles(x86)")));
    if (vs_path.empty()) {
      throw std::runtime_error("Can not find Visual Studio.");
    }
+    // Trim any trailing whitespace/newlines from the path
+    vs_path.erase(
+        std::find_if(
+            vs_path.rbegin(),
+            vs_path.rend(),
+            [](unsigned char ch) { return !std::isspace(ch); })
+            .base(),
+        vs_path.end());
    // Read the envs from vcvarsall.
-    std::string envs = JitCompiler::exec(fmt::format(
-        "\"{0}\\VC\\Auxiliary\\Build\\vcvarsall.bat\" {1} >NUL && set",
-        vs_path,
-        arch));
+    std::string envs = JitCompiler::exec(
+        fmt::format(
+            "\"{0}\\VC\\Auxiliary\\Build\\vcvarsall.bat\" {1} >NUL && set",
+            vs_path,
+            arch));
    for (const std::string& line : str_split(envs, '\n')) {
      // Each line is in the format "ENV_NAME=values".
      auto pos = line.find_first_of('=');
@@ -140,12 +152,13 @@ std::string JitCompiler::exec(const std::string& cmd) {
  int code = WEXITSTATUS(status);
 #endif
  if (code != 0) {
-    throw std::runtime_error(fmt::format(
-        "Failed to execute command with return code {0}: \"{1}\", "
-        "the output is: {2}",
-        code,
-        cmd,
-        ret));
+    throw std::runtime_error(
+        fmt::format(
+            "Failed to execute command with return code {0}: \"{1}\", "
+            "the output is: {2}",
+            code,
+            cmd,
+            ret));
  }
  return ret;
 }
@@ -45,9 +45,7 @@
 INSTANTIATE_LAPACK_REAL(geqrf)
 INSTANTIATE_LAPACK_REAL(orgqr)
 INSTANTIATE_LAPACK_REAL(syevd)
-INSTANTIATE_LAPACK_REAL(geev)
 INSTANTIATE_LAPACK_REAL(potrf)
-INSTANTIATE_LAPACK_REAL(gesdd)
 INSTANTIATE_LAPACK_REAL(getrf)
 INSTANTIATE_LAPACK_REAL(getri)
 INSTANTIATE_LAPACK_REAL(trtri)
@@ -63,3 +61,20 @@ INSTANTIATE_LAPACK_REAL(trtri)
  }

 INSTANTIATE_LAPACK_COMPLEX(heevd)
+
+#define INSTANTIATE_LAPACK_ALL(FUNC)                                \
+  template <typename T, typename... Args>                           \
+  void FUNC(Args... args) {                                         \
+    if constexpr (std::is_same_v<T, float>) {                       \
+      MLX_LAPACK_FUNC(s##FUNC)(std::forward<Args>(args)...);        \
+    } else if constexpr (std::is_same_v<T, double>) {               \
+      MLX_LAPACK_FUNC(d##FUNC)(std::forward<Args>(args)...);        \
+    } else if constexpr (std::is_same_v<T, std::complex<float>>) {  \
+      MLX_LAPACK_FUNC(c##FUNC)(std::forward<Args>(args)...);        \
+    } else if constexpr (std::is_same_v<T, std::complex<double>>) { \
+      MLX_LAPACK_FUNC(z##FUNC)(std::forward<Args>(args)...);        \
+    }                                                               \
+  }
+
+INSTANTIATE_LAPACK_ALL(geev)
+INSTANTIATE_LAPACK_ALL(gesdd)
@@ -2,6 +2,8 @@

 #include <cstring>
 #include "mlx/array.h"
+#include "mlx/backend/cpu/binary.h"
+#include "mlx/backend/cpu/binary_ops.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/gemm.h"
@@ -135,15 +137,29 @@ void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
    return;
  }

+  // Handle empty matrix case (K=0)
+  if (inputs[0].shape(-1) == 0) {
+    auto& c = inputs[2];
+    if (beta_ == 1.0f) {
+      CopyType ctype = c.data_size() == 1
+          ? CopyType::Scalar
+          : (c.flags().row_contiguous ? CopyType::Vector : CopyType::General);
+      copy_cpu(c, out, ctype, stream());
+    } else {
+      array beta_scalar = array(beta_, c.dtype());
+      auto& encoder = cpu::get_command_encoder(stream());
+      binary_float_op_cpu(c, beta_scalar, out, detail::Multiply(), stream());
+      encoder.add_temporary(std::move(beta_scalar));
+    }
+    return;
+  }
+
  // Fill output with C
  auto& c = inputs[2];
  CopyType ctype = c.data_size() == 1
      ? CopyType::Scalar
      : (c.flags().row_contiguous ? CopyType::Vector : CopyType::General);
  copy_cpu(c, out, ctype, stream());
-  if (inputs[0].shape(-1) == 0) {
-    return;
-  }
  matmul_general(inputs[0], inputs[1], out, stream(), alpha_, beta_);
 }

@@ -291,6 +291,17 @@ void RandomBits::eval_cpu(const std::vector<array>& inputs, array& out) {
                    num_keys,
                    kshape = keys.shape(),
                    kstrides = keys.strides()]() mutable {
+    auto copy_remaining = [&](char* cptr, size_t loc, uint32_t v) {
+      if (4 * loc + 4 <= bytes_per_key) {
+        reinterpret_cast<uint32_t*>(cptr)[loc] = v;
+      } else {
+        std::copy(
+            reinterpret_cast<char*>(&v),
+            reinterpret_cast<char*>(&v) + bytes_per_key - 4 * loc,
+            cptr + 4 * loc);
+      }
+    };
+
    size_t out_skip = (bytes_per_key + 4 - 1) / 4;
    auto half_size = out_skip / 2;
    bool even = out_skip % 2 == 0;
@@ -310,18 +321,12 @@ void RandomBits::eval_cpu(const std::vector<array>& inputs, array& out) {
      if (count.first < half_size) {
        auto rb = random::threefry2x32_hash(key, count);
        ptr[count.first++] = rb.first;
-        if (bytes_per_key % 4 > 0) {
-          std::copy(
-              reinterpret_cast<char*>(&rb.second),
-              reinterpret_cast<char*>(&rb.second) + bytes_per_key % 4,
-              cptr + 4 * count.second);
-        } else {
-          ptr[count.second] = rb.second;
-        }
+        copy_remaining(cptr, count.second, rb.second);
      }
      if (!even) {
        count.second = 0;
-        ptr[half_size] = random::threefry2x32_hash(key, count).first;
+        copy_remaining(
+            cptr, half_size, random::threefry2x32_hash(key, count).first);
      }
    }
  });
@@ -333,7 +338,7 @@ void Reshape::eval_cpu(const std::vector<array>& inputs, array& out) {

 void DynamicSlice::eval_cpu(const std::vector<array>& inputs, array& out) {
  if (out.size() == 0) {
-    out.set_data(nullptr);
+    out.set_data(allocator::malloc(0));
    return;
  }
  auto& in = inputs[0];
@@ -361,7 +366,7 @@ void DynamicSliceUpdate::eval_cpu(
    const std::vector<array>& inputs,
    array& out) {
  if (out.size() == 0) {
-    out.set_data(nullptr);
+    out.set_data(allocator::malloc(0));
    return;
  }

@@ -396,7 +401,7 @@ void DynamicSliceUpdate::eval_cpu(
 void SliceUpdate::eval_cpu(const std::vector<array>& inputs, array& out) {
  assert(inputs.size() == 2);
  if (out.size() == 0) {
-    out.set_data(nullptr);
+    out.set_data(allocator::malloc(0));
    return;
  }

@@ -1,8 +1,11 @@
 // Copyright © 2023 Apple Inc.

+#include "mlx/backend/common/unary.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
+#include "mlx/backend/cpu/unary.h"
+#include "mlx/backend/cpu/unary_ops.h"
 #include "mlx/fast_primitives.h"
 #include "mlx/primitives.h"
 #include "mlx/utils.h"
@@ -11,7 +14,20 @@ namespace mlx::core {

 namespace {

-const static float MXFP4_LUT[16] = {
+array ensure_row_contiguous(
+    const array& arr,
+    cpu::CommandEncoder& encoder,
+    Stream s) {
+  if (arr.flags().row_contiguous) {
+    return arr;
+  } else {
+    auto arr_cpy = contiguous_copy_cpu(arr, s);
+    encoder.add_temporary(arr_cpy);
+    return arr_cpy;
+  }
+};
+
+const static float FP4_LUT[16] = {
    +0.0f,
    +0.5f,
    +1.0f,
@@ -29,15 +45,19 @@ const static float MXFP4_LUT[16] = {
    -4.0f,
    -6.0f};

-template <typename T>
+template <typename T, int group_size>
 static inline T dequantize_scale(uint8_t s) {
-  using FOrI = union {
-    bfloat16_t f;
-    uint16_t i;
-  };
-  FOrI out;
-  out.i = (s == 0 ? 0x40 : (static_cast<uint16_t>(s) << 7));
-  return static_cast<T>(out.f);
+  if constexpr (group_size == 16) {
+    return static_cast<T>(detail::FromFP8{}(s));
+  } else {
+    using FOrI = union {
+      bfloat16_t f;
+      uint16_t i;
+    };
+    FOrI out;
+    out.i = (s == 0 ? 0x40 : (static_cast<uint16_t>(s) << 7));
+    return static_cast<T>(out.f);
+  }
 }

 inline constexpr short get_pack_factor(int bits, int wsize = 8) {
@@ -434,8 +454,8 @@ void _qmm_dispatch(
  }
 }

-template <typename T>
-void mxfp4_qmm(
+template <typename T, int group_size, int bits>
+void fp_qmm(
    T* result,
    const T* x,
    const uint32_t* w,
@@ -443,8 +463,7 @@ void mxfp4_qmm(
    int M,
    int N,
    int K) {
-  constexpr int group_size = 32;
-  constexpr int pack_factor = get_pack_factor(4, 8);
+  constexpr int pack_factor = get_pack_factor(bits, 8);
  constexpr int packs_in_group = group_size / pack_factor;

  for (int m = 0; m < M; m++) {
@@ -458,25 +477,27 @@ void mxfp4_qmm(
      T xi = *x++;

      for (int n = 0; n < N; n += group_size) {
-        T scale = dequantize_scale<T>(*scales_local++);
+        T scale = dequantize_scale<T, group_size>(*scales_local++);
        for (int ng = 0; ng < packs_in_group; ng++) {
-          uint8_t wi = *w_local++;
-#pragma clang loop unroll(full)
-          for (int p = 0; p < pack_factor; p++) {
+          if constexpr (bits == 4) {
            (*result_local++) +=
-                xi * scale * static_cast<T>(MXFP4_LUT[wi & 0xf]);
-            wi >>= 4;
+                xi * scale * static_cast<T>(FP4_LUT[w_local[0] & 0xf]);
+            (*result_local++) +=
+                xi * scale * static_cast<T>(FP4_LUT[(w_local[0] >> 4) & 0xf]);
+          } else {
+            (*result_local++) +=
+                xi * scale * static_cast<T>(detail::FromFP8{}(w_local[0]));
          }
+          w_local++;
        }
      }
    }
-
    result += N;
  }
 }

-template <typename T>
-void mxfp4_qmm_t(
+template <typename T, int group_size, int bits>
+void fp_qmm_t(
    T* result,
    const T* x,
    const uint32_t* w,
@@ -484,8 +505,7 @@ void mxfp4_qmm_t(
    int M,
    int N,
    int K) {
-  constexpr int group_size = 32;
-  constexpr int pack_factor = get_pack_factor(4, 8);
+  constexpr int pack_factor = get_pack_factor(bits, 8);
  constexpr int packs_in_group = group_size / pack_factor;

  for (int m = 0; m < M; m++) {
@@ -496,16 +516,19 @@ void mxfp4_qmm_t(
      const T* x_local = x;
      T sum = 0;
      for (int k = 0; k < K; k += group_size) {
-        T scale = dequantize_scale<T>(*scales_local++);
+        T scale = dequantize_scale<T, group_size>(*scales_local++);

        T gsum = 0;
        for (int kw = 0; kw < packs_in_group; kw++) {
-          uint8_t wi = *w_local++;
-#pragma clang loop unroll(full)
-          for (int p = 0; p < pack_factor; p++) {
-            gsum += (*x_local++) * static_cast<T>(MXFP4_LUT[wi & 0xf]);
-            wi >>= 4;
+          if constexpr (bits == 4) {
+            gsum += (*x_local++) * static_cast<T>(FP4_LUT[w_local[0] & 0xf]);
+            gsum +=
+                (*x_local++) * static_cast<T>(FP4_LUT[(w_local[0] >> 4) & 0xf]);
+          } else {
+            gsum +=
+                (*x_local++) * static_cast<T>(detail::FromFP8{}(w_local[0]));
          }
+          w_local++;
        }
        sum += scale * gsum;
      }
@@ -517,9 +540,9 @@ void mxfp4_qmm_t(
  }
 }

-template <int S>
-simd::Simd<float, S> mxfp4_extract_bits_simd(const uint32_t* w) {
-  if constexpr (S == 8) {
+template <int S, int bits>
+simd::Simd<float, S> fp_extract_bits_simd(const uint32_t* w) {
+  if constexpr (S == 8 && bits == 4) {
    constexpr std::array<uint32_t, 8> shifts_ = {{0, 4, 8, 12, 16, 20, 24, 28}};
    auto shifts(*(simd::Simd<uint32_t, S>*)&shifts_);
    auto wi = simd::Simd<uint32_t, S>(*w);
@@ -527,17 +550,20 @@ simd::Simd<float, S> mxfp4_extract_bits_simd(const uint32_t* w) {
    wi = wi & 0xf;
    simd::Simd<float, S> w_out;
    for (int i = 0; i < S; ++i) {
-      w_out[i] = MXFP4_LUT[wi[i]];
+      w_out[i] = FP4_LUT[wi[i]];
    }
    return w_out;
+  } else if constexpr (S == 8 && bits == 8) {
+    auto w_out = simd::load<uint8_t, S>(reinterpret_cast<const uint8_t*>(w));
+    return detail::FromFP8{}(w_out);
  } else {
    // Appease compiler.. but should never get here
    throw std::runtime_error("Unsupported combination for simd qmm.");
  }
 }

-template <typename T>
-void mxfp4_qmm_t_simd(
+template <typename T, int group_size, int bits>
+void fp_qmm_t_simd(
    T* result,
    const T* x,
    const uint32_t* w,
@@ -545,8 +571,7 @@ void mxfp4_qmm_t_simd(
    int M,
    int N,
    int K) {
-  constexpr int group_size = 32;
-  constexpr int pack_factor = 32 / 4;
+  constexpr int pack_factor = get_pack_factor(bits, 32);
  constexpr int packs_in_group = group_size / pack_factor;
  constexpr int S = simd::max_size<T>;
  static_assert(
@@ -561,12 +586,12 @@ void mxfp4_qmm_t_simd(
      simd::Simd<float, S> acc(0);
      auto x_local = x;
      for (int k = 0; k < K; k += group_size) {
-        T scale = dequantize_scale<T>(*scales_local++);
+        T scale = dequantize_scale<T, group_size>(*scales_local++);

        simd::Simd<float, S> g_acc(0);
        for (int kw = 0; kw < packs_in_group; kw += packs_per_simd) {
          // Extract bits
-          auto wf = mxfp4_extract_bits_simd<S>(w_local);
+          auto wf = fp_extract_bits_simd<S, bits>(w_local);
          w_local += packs_per_simd;
          simd::Simd<float, S> x_simd = simd::load<T, S>(x_local);
          g_acc = g_acc + x_simd * wf;
@@ -582,8 +607,8 @@ void mxfp4_qmm_t_simd(
  }
 }

-template <typename T>
-void mxfp4_qmm_dispatch_transpose(
+template <typename T, int group_size, int bits>
+void fp_qmm_dispatch_transpose(
    T* result,
    const T* x,
    const uint32_t* w,
@@ -595,17 +620,17 @@ void mxfp4_qmm_dispatch_transpose(
  if (transposed_w) {
    // the simd size must be a multiple of the number of elements per word
    if constexpr (simd::max_size<T> % 8 == 0) {
-      mxfp4_qmm_t_simd<T>(result, x, w, scales, M, N, K);
+      fp_qmm_t_simd<T, group_size, bits>(result, x, w, scales, M, N, K);
    } else {
-      mxfp4_qmm_t<T>(result, x, w, scales, M, N, K);
+      fp_qmm_t<T, group_size, bits>(result, x, w, scales, M, N, K);
    }
  } else {
-    mxfp4_qmm<T>(result, x, w, scales, M, N, K);
+    fp_qmm<T, group_size, bits>(result, x, w, scales, M, N, K);
  }
 }

-template <typename T>
-void mxfp4_qmm_dispatch_typed(
+template <typename T, int group_size, int bits>
+void fp_qmm_dispatch_mode(
    array& out,
    const array& x,
    const array& w,
@@ -623,7 +648,7 @@ void mxfp4_qmm_dispatch_typed(
  auto w_ptr = w.data<uint32_t>();
  auto scales_ptr = scales.data<uint8_t>();
  for (int i = 0; i < batch_size; i++) {
-    mxfp4_qmm_dispatch_transpose<T>(
+    fp_qmm_dispatch_transpose<T, group_size, bits>(
        out_ptr + i * M * N,
        x_ptr + elem_to_loc(i * M * K, x.shape(), x.strides()),
        w_ptr + elem_to_loc(i * w_els, w.shape(), w.strides()),
@@ -635,21 +660,44 @@ void mxfp4_qmm_dispatch_typed(
  }
 }

-void mxfp4_qmm_dispatch(
+template <typename T>
+void fp_qmm_dispatch_typed(
    array& out,
    const array& x,
    const array& w,
    const array& scales,
+    int group_size,
+    int bits,
+    bool transposed_w) {
+  if (bits == 8) {
+    fp_qmm_dispatch_mode<T, 32, 8>(out, x, w, scales, transposed_w);
+  } else if (group_size == 32) {
+    fp_qmm_dispatch_mode<T, 32, 4>(out, x, w, scales, transposed_w);
+  } else {
+    fp_qmm_dispatch_mode<T, 16, 4>(out, x, w, scales, transposed_w);
+  }
+}
+
+void fp_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    int group_size,
+    int bits,
    bool transposed_w) {
  switch (x.dtype()) {
    case bfloat16:
-      mxfp4_qmm_dispatch_typed<bfloat16_t>(out, x, w, scales, transposed_w);
+      fp_qmm_dispatch_typed<bfloat16_t>(
+          out, x, w, scales, group_size, bits, transposed_w);
      break;
    case float16:
-      mxfp4_qmm_dispatch_typed<float16_t>(out, x, w, scales, transposed_w);
+      fp_qmm_dispatch_typed<float16_t>(
+          out, x, w, scales, group_size, bits, transposed_w);
      break;
    case float32:
-      mxfp4_qmm_dispatch_typed<float>(out, x, w, scales, transposed_w);
+      fp_qmm_dispatch_typed<float>(
+          out, x, w, scales, group_size, bits, transposed_w);
      break;
    default:
      throw std::invalid_argument(
@@ -762,9 +810,8 @@ void _bs_qmm_dispatch(
          "[quantized_matmul] only floating types are supported");
  }
 }
-
-template <typename T>
-void mxfp4_bs_qmm_dispatch_typed(
+template <typename T, int group_size, int bits>
+void fp_bs_qmm_dispatch_mode(
    array& out,
    const array& x,
    const array& w,
@@ -791,7 +838,7 @@ void mxfp4_bs_qmm_dispatch_typed(
        i, lhs_indices.shape(), lhs_indices.strides())];
    int w_idx = rhs_indices_ptr[elem_to_loc(
        i, rhs_indices.shape(), rhs_indices.strides())];
-    mxfp4_qmm_dispatch_transpose<T>(
+    fp_qmm_dispatch_transpose<T, group_size, bits>(
        out_ptr + i * M * N,
        x_ptr + elem_to_loc(x_idx * M * K, x.shape(), x.strides()),
        w_ptr + elem_to_loc(w_idx * w_els, w.shape(), w.strides()),
@@ -804,26 +851,75 @@ void mxfp4_bs_qmm_dispatch_typed(
  }
 }

-void mxfp4_bs_qmm_dispatch(
+template <typename T>
+void fp_bs_qmm_dispatch_typed(
    array& out,
    const array& x,
    const array& w,
    const array& scales,
    const array& lhs_indices,
    const array& rhs_indices,
+    int group_size,
+    int bits,
+    bool transposed_w) {
+  if (bits == 8) {
+    fp_bs_qmm_dispatch_mode<T, 32, 8>(
+        out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+  } else if (group_size == 32) {
+    fp_bs_qmm_dispatch_mode<T, 32, 4>(
+        out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+  } else {
+    fp_bs_qmm_dispatch_mode<T, 16, 4>(
+        out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+  }
+}
+
+void fp_bs_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    int group_size,
+    int bits,
    bool transposed_w) {
  switch (x.dtype()) {
    case float32:
-      mxfp4_bs_qmm_dispatch_typed<float>(
-          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      fp_bs_qmm_dispatch_typed<float>(
+          out,
+          x,
+          w,
+          scales,
+          lhs_indices,
+          rhs_indices,
+          group_size,
+          bits,
+          transposed_w);
      break;
    case float16:
-      mxfp4_bs_qmm_dispatch_typed<float16_t>(
-          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      fp_bs_qmm_dispatch_typed<float16_t>(
+          out,
+          x,
+          w,
+          scales,
+          lhs_indices,
+          rhs_indices,
+          group_size,
+          bits,
+          transposed_w);
      break;
    case bfloat16:
-      mxfp4_bs_qmm_dispatch_typed<bfloat16_t>(
-          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      fp_bs_qmm_dispatch_typed<bfloat16_t>(
+          out,
+          x,
+          w,
+          scales,
+          lhs_indices,
+          rhs_indices,
+          group_size,
+          bits,
+          transposed_w);
      break;
    default:
      throw std::invalid_argument(
@@ -839,20 +935,9 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
  auto& scales_pre = inputs[2];

  auto& encoder = cpu::get_command_encoder(stream());
-  auto ensure_row_contiguous = [s = stream(), &encoder](const array& arr) {
-    if (arr.flags().row_contiguous) {
-      return arr;
-    } else {
-      auto arr_cpy = array(arr.shape(), arr.dtype(), nullptr, {});
-      copy_cpu(arr, arr_cpy, CopyType::General, s);
-      encoder.add_temporary(arr_cpy);
-      return arr_cpy;
-    }
-  };
-
-  auto x = ensure_row_contiguous(x_pre);
-  auto w = ensure_row_contiguous(w_pre);
-  auto scales = ensure_row_contiguous(scales_pre);
+  auto x = ensure_row_contiguous(x_pre, encoder, stream());
+  auto w = ensure_row_contiguous(w_pre, encoder, stream());
+  auto scales = ensure_row_contiguous(scales_pre, encoder, stream());

  out.set_data(allocator::malloc(out.nbytes()));

@@ -861,7 +946,7 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
  encoder.set_input_array(scales);
  encoder.set_output_array(out);
  if (mode_ == QuantizationMode::Affine) {
-    auto biases = ensure_row_contiguous(inputs[3]);
+    auto biases = ensure_row_contiguous(inputs[3], encoder, stream());
    encoder.set_input_array(biases);
    encoder.dispatch([out = array::unsafe_weak_copy(out),
                      x = array::unsafe_weak_copy(x),
@@ -878,8 +963,10 @@ void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
                      x = array::unsafe_weak_copy(x),
                      w = array::unsafe_weak_copy(w),
                      scales = array::unsafe_weak_copy(scales),
+                      group_size_ = group_size_,
+                      bits_ = bits_,
                      transpose_ = transpose_]() mutable {
-      mxfp4_qmm_dispatch(out, x, w, scales, transpose_);
+      fp_qmm_dispatch(out, x, w, scales, group_size_, bits_, transpose_);
    });
  }
 }
@@ -950,13 +1037,122 @@ void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
                      scales = array::unsafe_weak_copy(scales),
                      lhs_indices = array::unsafe_weak_copy(lhs_indices),
                      rhs_indices = array::unsafe_weak_copy(rhs_indices),
+                      group_size_ = group_size_,
+                      bits_ = bits_,
                      transpose_ = transpose_]() mutable {
-      mxfp4_bs_qmm_dispatch(
-          out, x, w, scales, lhs_indices, rhs_indices, transpose_);
+      fp_bs_qmm_dispatch(
+          out,
+          x,
+          w,
+          scales,
+          lhs_indices,
+          rhs_indices,
+          group_size_,
+          bits_,
+          transpose_);
    });
  }
 }

+uint8_t to_fp8_e8m0(float x) {
+  if (!std::isfinite(x)) {
+    return 0xFF;
+  }
+  if (x < 0.0f) {
+    return 0x00;
+  }
+  float le = std::log2(x);
+  int n = int(std::round(le));
+
+  n = n < -127 ? -127 : n;
+  n = n > 127 ? 127 : n;
+  return static_cast<uint8_t>(n + 127);
+}
+
+uint8_t to_fp4_e2m1(float x) {
+  if (std::isnan(x)) {
+    return 0x7;
+  }
+
+  const uint8_t sign_bit = (std::signbit(x)) ? 0x8 : 0x0;
+  x = std::abs(x);
+
+  uint8_t bits;
+  if (x > 5.0f) {
+    bits = 0x7;
+  } else if (x >= 3.5f) {
+    bits = 0x6;
+  } else if (x > 2.5f) {
+    bits = 0x5;
+  } else if (x >= 1.75f) {
+    bits = 0x4;
+  } else if (x > 1.25f) {
+    bits = 0x3;
+  } else if (x >= 0.75f) {
+    bits = 0x2;
+  } else if (x > 0.25f) {
+    bits = 0x1;
+  } else {
+    bits = 0x0;
+  }
+  return bits | sign_bit;
+}
+
+template <typename T>
+void fp_quantize_dequantize(
+    const array& w_arr,
+    array& out_arr,
+    int bits,
+    int group_size,
+    size_t w_size) {
+  auto w = w_arr.data<T>();
+  auto out = out_arr.data<T>();
+
+  size_t n_groups = w_size / group_size;
+
+  for (size_t i = 0; i < n_groups; ++i) {
+    size_t idx = i * group_size;
+    float scale = -std::numeric_limits<float>::infinity();
+    for (int j = 0; j < group_size; ++j) {
+      scale = std::max(scale, std::abs(w[idx + j]));
+    }
+    scale /= bits == 4 ? 6.0f : 448.0f;
+    if (group_size == 16) {
+      scale = dequantize_scale<float, 16>(detail::ToFP8()(scale));
+    } else {
+      scale = dequantize_scale<float, 32>(to_fp8_e8m0(scale));
+    }
+
+    for (int j = 0; j < group_size; ++j) {
+      float w_el = scale == 0 ? 0.0f : w[idx + j] / scale;
+      float output;
+      if (bits == 8) {
+        output = detail::FromFP8()(detail::ToFP8()(w_el));
+      } else {
+        output = FP4_LUT[to_fp4_e2m1(w_el)];
+      }
+      out[idx + j] = static_cast<T>(scale * output);
+    }
+  }
+}
+
+void dispatch_quantize_dequantize(
+    const array& w,
+    array& out,
+    int bits,
+    int group_size) {
+  if (w.dtype() == float16) {
+    fp_quantize_dequantize<float16_t>(w, out, bits, group_size, w.size());
+  } else if (w.dtype() == bfloat16) {
+    fp_quantize_dequantize<bfloat16_t>(w, out, bits, group_size, w.size());
+  } else if (w.dtype() == float32) {
+    fp_quantize_dequantize<float>(w, out, bits, group_size, w.size());
+  } else {
+    throw std::runtime_error(
+        "[quantize_dequantize] Only supports floating point inputs");
+  }
+}
+
 template <typename T, typename U>
 void quantize(
    const T* w,
@@ -1041,15 +1237,8 @@ void dispatch_quantize(
 void fast::Quantize::eval_cpu(
    const std::vector<array>& inputs,
    std::vector<array>& outputs) {
-  auto ensure_row_contiguous = [s = stream()](const array& arr) {
-    if (arr.flags().row_contiguous) {
-      return std::make_pair(arr, false);
-    } else {
-      return std::make_pair(contiguous_copy_cpu(arr, s), true);
-    }
-  };
-
-  auto [w, copied] = ensure_row_contiguous(inputs[0]);
+  auto& encoder = cpu::get_command_encoder(stream());
+  auto w = ensure_row_contiguous(inputs[0], encoder, stream());
  auto& out = outputs[0];
  out.set_data(allocator::malloc(out.nbytes()));

@@ -1057,10 +1246,6 @@ void fast::Quantize::eval_cpu(
  auto& biases = outputs[2];
  scales.set_data(allocator::malloc(scales.nbytes()));
  biases.set_data(allocator::malloc(biases.nbytes()));
-  auto& encoder = cpu::get_command_encoder(stream());
-  if (copied) {
-    encoder.add_temporary(w);
-  }
  encoder.set_input_array(w);
  encoder.set_input_array(scales);
  encoder.set_input_array(biases);
@@ -1102,4 +1287,84 @@ void fast::Quantize::eval_cpu(
  });
 }

+void fast::ConvertFP8::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  auto& in = inputs[0];
+  auto& out = outputs[0];
+  set_unary_output_data(in, out);
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  encoder.dispatch([in = array::unsafe_weak_copy(in),
+                    out = array::unsafe_weak_copy(out),
+                    to_fp8 = to_fp8_]() mutable {
+    if (to_fp8) {
+      switch (in.dtype()) {
+        case float16:
+          unary_op<float16_t, uint8_t>(in, out, detail::ToFP8());
+          break;
+        case bfloat16:
+          unary_op<bfloat16_t, uint8_t>(in, out, detail::ToFP8());
+          break;
+        default:
+          unary_op<float, uint8_t>(in, out, detail::ToFP8());
+          break;
+      }
+    } else {
+      switch (out.dtype()) {
+        case float16:
+          unary_op<uint8_t, float16_t>(in, out, detail::FromFP8());
+          break;
+        case bfloat16:
+          unary_op<uint8_t, bfloat16_t>(in, out, detail::FromFP8());
+          break;
+        default:
+          unary_op<uint8_t, float>(in, out, detail::FromFP8());
+          break;
+      }
+    }
+  });
+}
+
+void QQMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
+  auto& encoder = cpu::get_command_encoder(stream());
+
+  bool w_quantized = (inputs[1].dtype() == uint32);
+  if (w_quantized && inputs[0].shape(-2) == 1) {
+    bool donate_x = inputs[0].is_donatable();
+    auto x = ensure_row_contiguous(inputs[0], encoder, stream());
+    auto w = ensure_row_contiguous(inputs[1], encoder, stream());
+    auto scales = ensure_row_contiguous(inputs[2], encoder, stream());
+
+    out.set_data(allocator::malloc(out.nbytes()));
+
+    // If x is a copy it should be donatable
+    donate_x |= x.is_donatable();
+    auto xhat = donate_x
+        ? x
+        : array(allocator::malloc(x.nbytes()), x.shape(), x.dtype());
+    if (!donate_x) {
+      encoder.add_temporary(xhat);
+    }
+    encoder.set_input_array(x);
+    encoder.set_input_array(w);
+    encoder.set_input_array(scales);
+    encoder.set_output_array(out);
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      xhat = array::unsafe_weak_copy(xhat),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      group_size_ = group_size_,
+                      bits_ = bits_]() mutable {
+      dispatch_quantize_dequantize(x, xhat, bits_, group_size_);
+      fp_qmm_dispatch(out, xhat, w, scales, group_size_, bits_, true);
+    });
+    return;
+  } else {
+    throw std::runtime_error("[QQMatmul] NYI for the general case");
+  }
+}
+
 } // namespace mlx::core
@@ -1,5 +1,6 @@
 #pragma once

+#include <arm_neon.h>
 #include <simd/math.h>
 #include <simd/vector.h>

@@ -200,6 +201,15 @@ SIMD_DEFAULT_COMPARISONS(<=)
 SIMD_DEFAULT_COMPARISONS(==)
 SIMD_DEFAULT_COMPARISONS(!=)

+template <typename T, int N>
+Simd<T, N> clz(Simd<T, N> x) {
+  auto a = *(uint32x4_t*)(&x);
+  auto b = *((uint32x4_t*)(&x) + 1);
+  a = vclzq_u32(a);
+  b = vclzq_u32(b);
+  return asd::make_uint8(a, b);
+}
+
 template <typename T, int N>
 Simd<T, N> atan2(Simd<T, N> a, Simd<T, N> b) {
  return asd::atan2(a.value, b.value);
@@ -207,14 +217,20 @@ Simd<T, N> atan2(Simd<T, N> a, Simd<T, N> b) {

 template <typename T, int N>
 Simd<T, N> maximum(Simd<T, N> a, Simd<T, N> b) {
-  // TODO add isnan
-  return asd::max(a.value, b.value);
+  auto out = Simd<T, N>(asd::max(a.value, b.value));
+  if constexpr (!std::is_integral_v<T>) {
+    out = select(isnan(b), b, select(isnan(a), a, out));
+  }
+  return out;
 }

 template <typename T, int N>
 Simd<T, N> minimum(Simd<T, N> a, Simd<T, N> b) {
-  // TODO add isnan
-  return asd::min(a.value, b.value);
+  auto out = Simd<T, N>(asd::min(a.value, b.value));
+  if constexpr (!std::is_integral_v<T>) {
+    out = select(isnan(b), b, select(isnan(a), a, out));
+  }
+  return out;
 }

 template <typename T, int N>
@@ -1,11 +1,18 @@
 #pragma once

+// Required for using M_LN2 in MSVC.
+#define _USE_MATH_DEFINES
+
+#include <math.h>
 #include <stdint.h>
 #include <algorithm>
-#include <cmath>
 #include <complex>
 #include <functional>

+#ifdef _MSC_VER
+#include <intrin.h> // For _BitScanReverse
+#endif
+
 namespace mlx::core::simd {
 template <typename T, int N>
 struct Simd;
@@ -22,6 +29,14 @@ struct Simd<T, 1> {
  Simd(Simd<U, 1> v) : value(v.value) {}
  template <typename U>
  Simd(U v) : value(v) {}
+
+  T operator[](int) const {
+    return value;
+  }
+
+  T& operator[](int) {
+    return value;
+  }
 };

 template <typename T, int N>
@@ -105,7 +120,7 @@ Simd<T, 1> log1p(Simd<T, 1> in) {
      if (r == 0) { // handle underflow
        return Simd<T, 1>{T{x, theta}};
      }
-      return Simd<T, 1>{T{((typeof(x))(0.5)) * std::log1p(r), theta}};
+      return Simd<T, 1>{T{((decltype(x))(0.5)) * std::log1p(r), theta}};
    } else {
      auto z0 = std::hypot(x + 1, y);
      return Simd<T, 1>{T{std::log(z0), theta}};
@@ -171,6 +186,20 @@ DEFAULT_BINARY(&)
 DEFAULT_BINARY(&&)
 DEFAULT_BINARY(||)

+template <typename T>
+Simd<T, 1> clz(Simd<T, 1> x_) {
+#ifdef _MSC_VER
+  // MSVC doesn't have __builtin_clz, use _BitScanReverse instead
+  unsigned long index;
+  if (_BitScanReverse(&index, static_cast<unsigned long>(x_.value))) {
+    return static_cast<T>(31 - index);
+  }
+  return static_cast<T>(32); // All zeros case
+#else
+  return __builtin_clz(x_.value);
+#endif
+}
+
 template <typename T>
 Simd<T, 1> remainder(Simd<T, 1> a_, Simd<T, 1> b_) {
  T a = a_.value;
@@ -3,5 +3,9 @@
 #include "mlx/backend/cpu/simd/base_simd.h"

 #ifdef MLX_USE_ACCELERATE
+#if defined(__x86_64__)
+// the accelerate_simd implementation require neon -- use base implementation
+#else
 #include "mlx/backend/cpu/simd/accelerate_simd.h"
 #endif
+#endif
@@ -8,6 +8,183 @@

 namespace mlx::core {

+template <typename T, class Enable = void>
+struct SVDWork {};
+
+template <typename T>
+struct SVDWork<
+    T,
+    typename std::enable_if<std::is_floating_point<T>::value>::type> {
+  using R = T;
+
+  int N;
+  int M;
+  int K;
+  int lda;
+  int ldu;
+  int ldvt;
+  char jobz;
+  std::vector<array::Data> buffers;
+  int lwork;
+
+  SVDWork(int N, int M, int K, char jobz)
+      : N(N), M(M), K(K), lda(N), ldu(N), ldvt(M), jobz(jobz) {
+    T workspace_dimension = 0;
+
+    // Will contain the indices of eigenvectors that failed to converge (not
+    // used here but required by lapack).
+    buffers.emplace_back(allocator::malloc(sizeof(int) * 8 * K));
+
+    int lwork_query = -1;
+    int info;
+
+    // Compute workspace size.
+    gesdd<T>(
+        /* jobz = */ &jobz,
+        // M and N are swapped since lapack expects column-major.
+        /* m = */ &N,
+        /* n = */ &M,
+        /* a = */ nullptr,
+        /* lda = */ &lda,
+        /* s = */ nullptr,
+        /* u = */ nullptr,
+        /* ldu = */ &ldu,
+        /* vt = */ nullptr,
+        /* ldvt = */ &ldvt,
+        /* work = */ &workspace_dimension,
+        /* lwork = */ &lwork_query,
+        /* iwork = */ static_cast<int*>(buffers[0].buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "[SVD::eval_cpu] workspace calculation failed with code " << info;
+      throw std::runtime_error(ss.str());
+    }
+
+    lwork = workspace_dimension;
+    buffers.emplace_back(allocator::malloc(sizeof(T) * lwork));
+  }
+
+  void run(T* a, R* s, T* u, T* vt) {
+    int info;
+    gesdd<T>(
+        /* jobz = */ &jobz,
+        // M and N are swapped since lapack expects column-major.
+        /* m = */ &N,
+        /* n = */ &M,
+        /* a = */ a,
+        /* lda = */ &lda,
+        /* s = */ s,
+        // According to the identity above, lapack will write Vᵀᵀ as U.
+        /* u = */ u,
+        /* ldu = */ &ldu,
+        // According to the identity above, lapack will write Uᵀ as Vᵀ.
+        /* vt = */ vt,
+        /* ldvt = */ &ldvt,
+        /* work = */ static_cast<T*>(buffers[1].buffer.raw_ptr()),
+        /* lwork = */ &lwork,
+        /* iwork = */ static_cast<int*>(buffers[0].buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "svd_impl: sgesvdx_ failed with code " << info;
+      throw std::runtime_error(ss.str());
+    }
+  }
+};
+
+template <>
+struct SVDWork<std::complex<float>> {
+  using T = std::complex<float>;
+  using R = float;
+
+  int N;
+  int M;
+  int K;
+  int lda;
+  int ldu;
+  int ldvt;
+  char jobz;
+  std::vector<array::Data> buffers;
+  int lwork;
+
+  SVDWork(int N, int M, int K, char jobz)
+      : N(N), M(M), K(K), lda(N), ldu(N), ldvt(M), jobz(jobz) {
+    T workspace_dimension = 0;
+
+    // Will contain the indices of eigenvectors that failed to converge (not
+    // used here but required by lapack).
+    buffers.emplace_back(allocator::malloc(sizeof(int) * 8 * K));
+
+    const int lrwork =
+        jobz == 'A' ? std::max(1, 5 * K * K + 5 * K) : std::max(1, 7 * K);
+    buffers.emplace_back(allocator::malloc(sizeof(float) * lrwork));
+
+    int lwork_query = -1;
+    int work_query = -1;
+    int info;
+
+    // Compute workspace size.
+    gesdd<T>(
+        /* jobz = */ &jobz,
+        // M and N are swapped since lapack expects column-major.
+        /* m = */ &N,
+        /* n = */ &M,
+        /* a = */ nullptr,
+        /* lda = */ &lda,
+        /* s = */ nullptr,
+        /* u = */ nullptr,
+        /* ldu = */ &ldu,
+        /* vt = */ nullptr,
+        /* ldvt = */ &ldvt,
+        /* work = */ &workspace_dimension,
+        /* lwork = */ &lwork_query,
+        /* rwork = */ static_cast<float*>(buffers[1].buffer.raw_ptr()),
+        /* iwork = */ static_cast<int*>(buffers[0].buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "[SVD::eval_cpu] workspace calculation failed with code " << info;
+      throw std::runtime_error(ss.str());
+    }
+
+    lwork = workspace_dimension.real();
+    buffers.emplace_back(allocator::malloc(sizeof(T) * lwork));
+  }
+
+  void run(T* a, R* s, T* u, T* vt) {
+    int info;
+    gesdd<T>(
+        /* jobz = */ &jobz,
+        // M and N are swapped since lapack expects column-major.
+        /* m = */ &N,
+        /* n = */ &M,
+        /* a = */ a,
+        /* lda = */ &lda,
+        /* s = */ s,
+        // According to the identity above, lapack will write Vᵀᵀ as U.
+        /* u = */ u,
+        /* ldu = */ &ldu,
+        // According to the identity above, lapack will write Uᵀ as Vᵀ.
+        /* vt = */ vt,
+        /* ldvt = */ &ldvt,
+        /* work = */ static_cast<T*>(buffers[2].buffer.raw_ptr()),
+        /* lwork = */ &lwork,
+        /* rwork = */ static_cast<float*>(buffers[1].buffer.raw_ptr()),
+        /* iwork = */ static_cast<int*>(buffers[0].buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "svd_impl: sgesvdx_ failed with code " << info;
+      throw std::runtime_error(ss.str());
+    }
+  }
+};
+
 template <typename T>
 void svd_impl(
    const array& a,
@@ -27,6 +204,8 @@ void svd_impl(
  const int N = a.shape(-1);
  const int K = std::min(M, N);

+  using R = typename SVDWork<T>::R;
+
  size_t num_matrices = a.size() / (M * N);

  // lapack clobbers the input, so we have to make a copy.
@@ -42,7 +221,7 @@ void svd_impl(
  encoder.set_input_array(a);
  auto in_ptr = in.data<T>();
  T* u_ptr;
-  T* s_ptr;
+  R* s_ptr;
  T* vt_ptr;

  if (compute_uv) {
@@ -58,7 +237,7 @@ void svd_impl(
    encoder.set_output_array(s);
    encoder.set_output_array(vt);

-    s_ptr = s.data<T>();
+    s_ptr = s.data<R>();
    u_ptr = u.data<T>();
    vt_ptr = vt.data<T>();
  } else {
@@ -68,96 +247,26 @@ void svd_impl(

    encoder.set_output_array(s);

-    s_ptr = s.data<T>();
+    s_ptr = s.data<R>();
    u_ptr = nullptr;
    vt_ptr = nullptr;
  }

  encoder.dispatch([in_ptr, u_ptr, s_ptr, vt_ptr, M, N, K, num_matrices]() {
-    // A of shape M x N. The leading dimension is N since lapack receives Aᵀ.
-    const int lda = N;
-    // U of shape M x M. (N x N in lapack).
-    const int ldu = N;
-    // Vᵀ of shape N x N. (M x M in lapack).
-    const int ldvt = M;
-
-    auto jobz = (u_ptr) ? "A" : "N";
-
-    T workspace_dimension = 0;
-
-    // Will contain the indices of eigenvectors that failed to converge (not
-    // used here but required by lapack).
-    auto iwork = array::Data{allocator::malloc(sizeof(int) * 8 * K)};
-
-    static const int lwork_query = -1;
-
-    int info;
-
-    // Compute workspace size.
-    gesdd<T>(
-        /* jobz = */ jobz,
-        // M and N are swapped since lapack expects column-major.
-        /* m = */ &N,
-        /* n = */ &M,
-        /* a = */ nullptr,
-        /* lda = */ &lda,
-        /* s = */ nullptr,
-        /* u = */ nullptr,
-        /* ldu = */ &ldu,
-        /* vt = */ nullptr,
-        /* ldvt = */ &ldvt,
-        /* work = */ &workspace_dimension,
-        /* lwork = */ &lwork_query,
-        /* iwork = */ static_cast<int*>(iwork.buffer.raw_ptr()),
-        /* info = */ &info);
-
-    if (info != 0) {
-      std::stringstream ss;
-      ss << "[SVD::eval_cpu] workspace calculation failed with code " << info;
-      throw std::runtime_error(ss.str());
-    }
-
-    const int lwork = workspace_dimension;
-    auto scratch = array::Data{allocator::malloc(sizeof(T) * lwork)};
-
+    auto jobz = (u_ptr) ? 'A' : 'N';
+    SVDWork<T> svd_work(N, M, K, jobz);
    // Loop over matrices.
    for (int i = 0; i < num_matrices; i++) {
-      gesdd<T>(
-          /* jobz = */ jobz,
-          // M and N are swapped since lapack expects column-major.
-          /* m = */ &N,
-          /* n = */ &M,
-          /* a = */ in_ptr + M * N * i,
-          /* lda = */ &lda,
-          /* s = */ s_ptr + K * i,
-          // According to the identity above, lapack will write Vᵀᵀ as U.
-          /* u = */ vt_ptr ? vt_ptr + N * N * i : nullptr,
-          /* ldu = */ &ldu,
-          // According to the identity above, lapack will write Uᵀ as Vᵀ.
-          /* vt = */ u_ptr ? u_ptr + M * M * i : nullptr,
-          /* ldvt = */ &ldvt,
-          /* work = */ static_cast<T*>(scratch.buffer.raw_ptr()),
-          /* lwork = */ &lwork,
-          /* iwork = */ static_cast<int*>(iwork.buffer.raw_ptr()),
-          /* info = */ &info);
-
-      if (info != 0) {
-        std::stringstream ss;
-        ss << "svd_impl: sgesvdx_ failed with code " << info;
-        throw std::runtime_error(ss.str());
-      }
+      svd_work.run(
+          in_ptr + M * N * i,
+          s_ptr + K * i,
+          vt_ptr ? vt_ptr + N * N * i : nullptr,
+          u_ptr ? u_ptr + M * M * i : nullptr);
    }
  });
  encoder.add_temporary(in);
 }

-template <typename T>
-void compute_svd(
-    const array& a,
-    bool compute_uv,
-    std::vector<array>& outputs,
-    Stream stream) {}
-
 void SVD::eval_cpu(
    const std::vector<array>& inputs,
    std::vector<array>& outputs) {
@@ -168,9 +277,12 @@ void SVD::eval_cpu(
    case float64:
      svd_impl<double>(inputs[0], outputs, compute_uv_, stream());
      break;
+    case complex64:
+      svd_impl<std::complex<float>>(inputs[0], outputs, compute_uv_, stream());
+      break;
    default:
      throw std::runtime_error(
-          "[SVD::eval_cpu] only supports float32 or float64.");
+          "[SVD::eval_cpu] only supports float32, float64, or complex64.");
  }
 }

@@ -24,9 +24,9 @@ void unary_op(const array& a, array& out, Op) {
  auto ndim = a.ndim();
  if (a.flags().contiguous) {
    auto size = a.data_size();
-    constexpr int N = simd::max_size<T>;
+    constexpr int N = std::min(simd::max_size<T>, simd::max_size<U>);
    while (size >= N) {
-      simd::store(dst, Op{}(simd::load<T, N>(src)));
+      simd::store(dst, simd::Simd<U, N>(Op{}(simd::load<T, N>(src))));
      size -= N;
      src += N;
      dst += N;
--- a/Show More
+++ b/Show More