Add debug line info

Co-authored-by: Awni Hannun <awni@apple.com>

.circleci/config.yml

@@ -1,600 +0,0 @@
-version: 2.1
-
-orbs:
-  apple: ml-explore/pr-approval@0.1.0
-
-parameters:
-  nightly_build:
-    type: boolean
-    default: false
-  weekly_build:
-    type: boolean
-    default: false
-  test_release:
-    type: boolean
-    default: false
-  linux_release:
-    type: boolean
-    default: false
-
-jobs:
-  build_documentation:
-    parameters:
-      upload-docs:
-        type: boolean
-        default: false
-    macos:
-      xcode: "16.2.0"
-    resource_class: m2pro.medium
-    steps:
-      - checkout
-      - run:
-          name: Install
-          command: |
-            brew install python@3.9
-            brew install doxygen
-            python3.9 -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            pip install --upgrade cmake
-            pip install -r docs/requirements.txt
-            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` 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:
-    docker:
-      - image: cimg/python:3.9
-
-    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: |
-            pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install numpy
-            sudo apt-get update
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
-      - run:
-          name: Install Python package
-          command: |
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python3 setup.py build_ext --inplace
-            CMAKE_ARGS="-DMLX_BUILD_METAL=OFF" \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python3 setup.py develop
-      - run:
-          name: Generate package stubs
-          command: |
-            echo "stubs"
-            pip install typing_extensions
-            python setup.py generate_stubs 
-      - run:
-          name: Run Python tests
-          command: |
-            python3 -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
-      - run:
-          name: Build CPP only
-          command: |
-            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: "16.2.0"
-      macosx_deployment_target:
-        type: string
-        default: ""
-    macos:
-      xcode: << parameters.xcode_version >>
-    environment:
-      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    resource_class: m2pro.medium
-    steps:
-      - checkout
-      - run:
-          name: Install dependencies
-          command: |
-            brew install python@3.9
-            brew install openmpi
-            python3.9 -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install numpy
-            pip install torch
-            pip install tensorflow
-            pip install unittest-xml-reporting
-      - run:
-          name: Install Python package
-          command: |
-            source env/bin/activate
-            DEBUG=1 CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-            CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              pip install -e . -v
-      - run:
-          name: Generate package stubs
-          command: |
-            source env/bin/activate
-            pip install typing_extensions
-            python setup.py generate_stubs 
-      - run:
-          name: Run Python tests
-          command: |
-            source env/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
-      - run:
-          name: Build example extension
-          command: |
-            source env/bin/activate
-            cd examples/extensions
-            pip install -r requirements.txt
-            python setup.py build_ext -j8
-      - store_test_results:
-          path: test-results
-      - run:
-          name: Build CPP only
-          command: |
-            source env/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 env/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: |
-            source env/bin/activate
-            CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-              CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-              pip install -e . -v
-            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_jit
-
-  build_release:
-    parameters:
-      python_version:
-        type: string
-        default: "3.9"
-      xcode_version:
-        type: string
-        default: "16.2.0"
-      build_env:
-        type: string
-        default: ""
-      macosx_deployment_target:
-        type: string
-        default: ""
-    macos:
-      xcode: << parameters.xcode_version >>
-    resource_class: m2pro.medium
-    environment:
-      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    steps:
-      - checkout
-      - run:
-          name: Install dependencies
-          command: |
-            brew install python@<< parameters.python_version >>
-            brew install openmpi
-            python<< parameters.python_version >> -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            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: |
-            source env/bin/activate
-            env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
-              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-              pip install . -v
-      - run:
-          name: Generate package stubs
-          command: |
-            source env/bin/activate
-            pip install typing_extensions
-            python setup.py generate_stubs 
-      - run:
-          name: Build Python package
-          command: |
-            source env/bin/activate
-            << parameters.build_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`sysctl -n hw.ncpu` \
-              python -m build -w
-      - when:
-          condition: << parameters.build_env >>
-          steps:
-            - run:
-                name: Upload package
-                command: |
-                  source env/bin/activate
-                  twine upload dist/*
-      - store_artifacts:
-          path: dist/
-
-  build_linux_release:
-    parameters:
-      python_version:
-        type: string
-        default: "3.9"
-      extra_env:
-        type: string
-        default: "DEV_RELEASE=1"
-    docker:
-      - image: ubuntu:20.04
-    steps:
-      - checkout
-      - run:
-          name: Build wheel
-          command: |
-            PYTHON=python<< parameters.python_version >>
-            apt-get update
-            apt-get upgrade -y
-            DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC apt-get -y install tzdata
-            apt-get install -y apt-utils
-            apt-get install -y software-properties-common
-            add-apt-repository -y ppa:deadsnakes/ppa
-            apt-get install -y $PYTHON $PYTHON-dev $PYTHON-full
-            apt-get install -y libblas-dev liblapack-dev liblapacke-dev
-            apt-get install -y build-essential git
-            $PYTHON -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install --upgrade setuptools
-            pip install numpy
-            pip install auditwheel
-            pip install patchelf
-            pip install build
-            pip install twine
-            << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              pip install . -v
-            pip install typing_extensions
-            python setup.py generate_stubs 
-            << parameters.extra_env >> \
-              CMAKE_BUILD_PARALLEL_LEVEL=`nproc` \
-              python -m build --wheel
-            auditwheel show dist/*
-            auditwheel repair dist/* --plat manylinux_2_31_x86_64
-      - run:
-          name: Upload package
-          command: |
-            source env/bin/activate
-            twine upload wheelhouse/*
-      - store_artifacts:
-          path: wheelhouse/
-
-workflows:
-  build_and_test:
-    when:
-      and:
-        - matches:
-            pattern: "^(?!pull/)[-\\w]+$"
-            value: << pipeline.git.branch >>
-        - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.weekly_build >>
-        - not: << pipeline.parameters.test_release >>
-    jobs:
-      - mac_build_and_test:
-          matrix:
-            parameters:
-              macosx_deployment_target: ["13.5", "14.0"]
-      - linux_build_and_test
-      - build_documentation 
-
-  build_pypi_release:
-    when:
-      and:
-        - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.weekly_build >>
-        - not: << pipeline.parameters.test_release >>
-    jobs:
-      - build_release:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              build_env: ["PYPI_RELEASE=1"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "PYPI_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "PYPI_RELEASE=1"
-      - build_documentation:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          upload-docs: true
-
-  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", "14.0"]
-      - linux_build_and_test:
-          requires: [ hold ]
-  nightly_build:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.nightly_build >>
-    jobs:
-      - build_release:
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-  weekly_build:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.weekly_build >>
-    jobs:
-      - build_release:
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              build_env: ["DEV_RELEASE=1"]
-              xcode_version: ["16.2.0", "15.0.0"]
-            exclude:
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "13.5"
-                xcode_version: "16.2.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "14.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.9"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.10"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.11"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.12"
-                build_env: "DEV_RELEASE=1"
-              - macosx_deployment_target: "15.0"
-                xcode_version: "15.0.0"
-                python_version: "3.13"
-                build_env: "DEV_RELEASE=1"
-  linux_test_release:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.linux_release >>
-    jobs:
-      - build_linux_release:
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              extra_env: ["PYPI_RELEASE=1"]

.github/actions/build-cuda-release/action.yml

@@ -0,0 +1,15 @@
+name: 'Build CUDA wheel'
+description: 'Build CUDA wheel'
+
+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
+        bash python/scripts/repair_cuda.sh

.github/actions/build-docs/action.yml

@@ -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

.github/actions/build-linux-release/action.yml

@@ -0,0 +1,40 @@
+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: Generate package stubs
+      shell: bash
+      run: |
+        pip install -e ".[dev]" -v
+        pip install typing_extensions
+        python setup.py generate_stubs
+    - 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
+        bash python/scripts/repair_linux.sh ${{ inputs.arch }}
+    - 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 }}

.github/actions/build-linux/action.yml

@@ -0,0 +1,41 @@
+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 when the built executables can not run.
+          CMAKE_ARGS="$CMAKE_ARGS -DMLX_BUILD_TESTS=OFF"
+        fi
+        pip install --no-build-isolation -e ".[dev]" -v
+        # Pass the CMAKE_ARGS to following steps.
+        echo CMAKE_ARGS="$CMAKE_ARGS" >> $GITHUB_OUTPUT
+
+    - name: Generate package stubs
+      shell: sh
+      run: |
+        pip install typing_extensions
+        python setup.py generate_stubs
+
+    - 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)

.github/actions/build-macos-release/action.yml

@@ -0,0 +1,34 @@
+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:
+        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:
+        MACOSX_DEPLOYMENT_TARGET: ${{ inputs.macos-target }}
+      run: |
+        python setup.py clean --all
+        MLX_BUILD_STAGE=2 python -m build -w

.github/actions/build-macos/action.yml

@@ -0,0 +1,88 @@
+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 nanobind==2.4.0
+        pip install -e . -v
+
+    - name: Generate package stubs
+      shell: bash -l {0}
+      run: |
+        pip install typing_extensions
+        python setup.py generate_stubs
+
+    - name: Install tests dependencies
+      shell: bash -l {0}
+      run: |
+        pip install numpy torch tensorflow unittest-xml-reporting
+
+    - name: Run Python tests
+      shell: bash -l {0}
+      env:
+        LOW_MEMORY: 1
+      run: |
+        DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
+        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
+    
+    - 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 xmlrunner discover \
+            -v python/tests \
+            -o test-results/gpu_jit

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

@@ -0,0 +1,86 @@
+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: '3.10'
+
+runs:
+  using: "composite"
+  steps:
+    - name: Use ccache
+      uses: hendrikmuhs/ccache-action@v1.2
+      with:
+        key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ inputs.toolkit }}-py${{ inputs.python-version }}
+        max-size: 1GB
+
+    - name: Install common dependencies
+      shell: bash
+      run: |
+        sudo apt-get update
+        sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev zip
+
+    - uses: actions/setup-python@v6
+      with:
+        python-version: ${{ inputs.python-version }}
+
+    - name: Setup Python venv
+      shell: bash
+      run: |
+        python -m venv .venv
+        source .venv/bin/activate
+        pip install setuptools cmake nanobind==2.4.0
+        echo PATH=$PATH >> $GITHUB_ENV
+        # Make cmake search .venv for nanobind
+        echo PYTHONPATH=`python -c 'import sys; print(sys.path[-1])'` >> $GITHUB_ENV
+
+    - name: Install MPI
+      shell: bash
+      run: sudo apt-get install -y openmpi-bin openmpi-common libopenmpi-dev
+
+    - name: Install CUDA toolkit
+      if: ${{ startsWith(inputs.toolkit, 'cuda') }}
+      shell: bash
+      env:
+        # Note: the CI machine does not meet CUDA 13's driver requirement.
+        # Compatibility matrix:
+        # https://docs.nvidia.com/deeplearning/cudnn/backend/latest/reference/support-matrix.html
+        PACKAGES: |
+          {
+            "cuda-12.6": "libcudnn9-dev-cuda-12 cuda-toolkit-12-6",
+            "cuda-12.9": "libcudnn9-dev-cuda-12 cuda-toolkit-12-9",
+            "cuda-13.0": "libcudnn9-dev-cuda-13 cuda-toolkit-13-0"
+          }
+      run: |
+        # The CUDA binaries are hosted in the "sbsa" repo, the "arm64" repo is
+        # Jetson specific. SBSA means Arm Server Base System Architecture.
+        ARCH=${{ 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 \
+            libnccl2 libnccl-dev \
+            ${{ fromJson(env.PACKAGES)[inputs.toolkit] }}
+        echo "/usr/local/${{ inputs.toolkit }}/bin" >> $GITHUB_PATH
+
+    - name: CUDA packages and driver report
+      if: ${{ startsWith(inputs.toolkit, 'cuda') }}
+      shell: bash
+      run: |
+        sudo apt-get install -y ubuntu-drivers-common dkms
+        echo "NVIDIA Driver Packages Available:"
+        sudo ubuntu-drivers list --gpgpu
+        echo "NVIDIA Driver Version:"
+        cat /proc/driver/nvidia/version || echo "nvidia driver not found"
+        echo "Installed NVIDIA and CUDA packages:"
+        dpkg -l | egrep "cuda|nvidia" -i
+        echo "DKMS Status:"
+        dkms status || echo "dkms not found"
+        echo "NVIDIA-SMI Status:"
+        nvidia-smi || echo "nvidia-smi not found"

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

@@ -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 }}

.github/actions/test-linux/action.yml

@@ -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::"

.github/dependabot.yml

@@ -0,0 +1,6 @@
+version: 2
+updates:
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"

.github/scripts/setup+build-cpp-linux-fedora-container.sh

@@ -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

.github/workflows/build_and_test.yml

@@ -0,0 +1,108 @@
+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: ${{ matrix.arch == 'x86_64' && 'ubuntu-22.04' || 'ubuntu-22.04-arm' }}
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-linux
+      - uses: ./.github/actions/build-linux
+      - uses: ./.github/actions/test-linux
+
+  cuda_build_and_test:
+    name: Linux (${{ 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"]
+    runs-on: [self-hosted, macos]
+    env:
+      MACOSX_DEPLOYMENT_TARGET: ${{ matrix.macos-target }}
+    needs: check_lint
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-macos
+      - uses: ./.github/actions/build-macos
+
+  build_documentation:
+    name: Build Documentation
+    if: github.repository == 'ml-explore/mlx'
+    runs-on: ubuntu-22.04
+    needs: check_lint
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/build-docs
+
+  linux_fedora_build_cpp:
+    name: Linux Fedora (${{ 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

.github/workflows/documentation.yml

@@ -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

.github/workflows/nightly.yml

@@ -0,0 +1,96 @@
+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: ${{ matrix.python-version == '3.10' }}
+          arch: "x86_64"
+      - name: Upload mlx artifacts
+        uses: actions/upload-artifact@v5
+        with:
+          name: linux-wheels-${{ 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@v5
+        with:
+          name: mlx-cpu
+          path: wheelhouse/mlx_cpu-*.whl
+          retention-days: 7
+
+  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
+
+  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 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'
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v5
+        with:
+          name: mlx-cuda
+          path: wheelhouse/mlx_cuda-*.whl
+          retention-days: 7

.github/workflows/pull_request.yml

@@ -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

.github/workflows/release.yml

@@ -0,0 +1,238 @@
+name: PyPI Release
+
+on:
+  push:
+    tags:
+      - 'v*'
+  workflow_dispatch:
+    inputs:
+      dev_release:
+        description: "Do a dev release or regular release"
+        required: true
+        default: "false"
+
+permissions:
+  contents: read
+
+jobs:
+  setup:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Set publishing variables
+        run: echo "Publishing setup complete"
+
+  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:
+    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: ${{ github.event.inputs.dev_release == 'true' && 1 || 0 }}
+    steps:
+      - uses: actions/checkout@v6
+      - uses: ./.github/actions/setup-linux
+        with:
+          python-version: ${{ matrix.python_version }}
+      - uses: ./.github/actions/build-linux-release
+        with:
+          build-backend: ${{ matrix.python-version == '3.10' }}
+          arch: ${{ matrix.arch }}
+      - name: Upload MLX artifacts
+        uses: actions/upload-artifact@v5
+        with:
+          overwrite: true
+          name: linux-wheels-${{ matrix.python_version }}-${{ matrix.arch }}
+          path: wheelhouse/mlx-*.whl
+      - name: Upload CPU artifacts
+        if: matrix.python_version == '3.10'
+        uses: actions/upload-artifact@v5
+        with:
+          overwrite: true
+          name: mlx-cpu-${{ matrix.arch }}
+          path: wheelhouse/mlx_cpu-*.whl
+  
+  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: ${{ github.event.inputs.dev_release == 'true' && 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 nanobind==2.4.0
+          pip install -e . -v
+      - name: Generate package stubs
+        shell: bash -l {0}
+        run: |
+          pip install typing_extensions
+          python setup.py generate_stubs
+      - 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: ${{ matrix.python-version == '3.10' }}
+      - name: Upload MLX artifacts
+        uses: actions/upload-artifact@v5
+        with:
+          overwrite: true
+          name: mac-wheels-${{ matrix.python-version }}
+          path: dist/mlx-*.whl
+      - name: Upload Metal artifacts
+        if: matrix.python-version == '3.10'
+        uses: actions/upload-artifact@v5
+        with:
+          overwrite: true
+          name: mlx-metal
+          path: dist/mlx_metal-*.whl
+
+  build_cuda_release:
+    if: github.repository == 'ml-explore/mlx'
+    runs-on: ubuntu-22-large
+    env:
+      PYPI_RELEASE: 1
+      DEV_RELEASE: ${{ github.event.inputs.dev_release == 'true' && 1 || 0 }}
+    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
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v5
+        with:
+          overwrite: true
+          name: mlx-cuda
+          path: wheelhouse/mlx_cuda-*.whl
+
+  pypi-publish:
+    name: Upload release to PyPI
+    runs-on: ubuntu-latest
+    needs: [setup, build_linux_release, build_mac_release]
+    permissions:
+      id-token: write
+    environment:
+      name: pypi
+      url: https://pypi.org/p/mlx
+    steps:
+      - uses: actions/download-artifact@v6
+        with:
+          pattern: linux-wheels-*
+          merge-multiple: true
+          path: dist
+      - uses: actions/download-artifact@v6
+        with:
+          pattern: mac-wheels-*
+          merge-multiple: true
+          path: dist
+      - name: Display structure of downloaded files
+        run: ls -R dist
+      - name: Publish package distributions to PyPI
+        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: [setup, build_cuda_release]
+    permissions:
+      id-token: write
+    environment:
+      name: pypi
+      url: https://pypi.org/p/mlx-cuda
+    steps:
+      - uses: actions/download-artifact@v6
+        with:
+          name: mlx-cuda
+          path: dist
+      - name: Display structure of downloaded files
+        run: ls -R dist
+      - name: Publish package distributions to PyPI
+        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: [setup, build_linux_release]
+    permissions:
+      id-token: write
+    environment:
+      name: pypi
+      url: https://pypi.org/p/mlx-cpu
+    steps:
+      - uses: actions/download-artifact@v6
+        with:
+          pattern: mlx-cpu-*
+          merge-multiple: true
+          path: dist
+      - name: Display structure of downloaded files
+        run: ls -R dist
+      - name: Publish package distributions to PyPI
+        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: [setup, build_mac_release]
+    permissions:
+      id-token: write
+    environment:
+      name: pypi
+      url: https://pypi.org/p/mlx-metal
+    steps:
+      - uses: actions/download-artifact@v6
+        with:
+          name: mlx-metal
+          path: dist
+      - name: Display structure of downloaded files
+        run: ls -R dist
+      - name: Publish package distributions to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          repository-url: https://upload.pypi.org/legacy/

.pre-commit-config.yaml

@@ -1,4 +1,10 @@
 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
     hooks:

ACKNOWLEDGMENTS.md

@@ -19,11 +19,17 @@ MLX was developed with contributions from the following individuals:
 - Gleb Pobudzey: Added the `where` primitive, and groups in 1D and 2D convolutions.
 - Paul Paczuski: Improved stability of BCE loss calculation
 - Max-Heinrich Laves: Added `conv_transpose1d`, `conv_transpose2d`, and `conv_transpose3d` ops.
+- Gökdeniz Gülmez: Added the `Muon (MomentUm Orthogonalized by Newton-schulz)` optimizer, and the `ReLU²` activation function.
 
 <a href="https://github.com/ml-explore/mlx/graphs/contributors">
   <img class="dark-light" src="https://contrib.rocks/image?repo=ml-explore/mlx&anon=0&columns=20&max=100&r=true" />
 </a>
 
+# Organizations
+
+MLX has received contributions from the following companies:
+- NVIDIA Corporation & Affiliates
+
 # Third-Party Software
 
 MLX leverages several third-party software, listed here together with

CMakeLists.txt

@@ -26,6 +26,7 @@ set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 set(CMAKE_INSTALL_MESSAGE NEVER)
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 # ----------------------------- Configuration -----------------------------
 option(MLX_BUILD_TESTS "Build tests for mlx" ON)
@@ -41,7 +42,9 @@ 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_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)
 
 # --------------------- Processor tests -------------------------
 message(
@@ -64,10 +67,18 @@ if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
       message(WARNING "Building for x86_64 arch is not officially supported.")
     endif()
   endif()
-
 else()
   set(MLX_BUILD_METAL OFF)
-  message(WARNING "MLX is prioritised for Apple silicon systems using macOS.")
+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()
 
 # ----------------------------- Lib -----------------------------
@@ -78,22 +89,26 @@ cmake_policy(SET CMP0135 NEW)
 
 add_library(mlx)
 
-if(MLX_BUILD_METAL)
-  set(METAL_LIB "-framework Metal")
-  set(FOUNDATION_LIB "-framework Foundation")
-  set(QUARTZ_LIB "-framework QuartzCore")
-endif()
+# 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)
 
 if(MLX_BUILD_CUDA)
   enable_language(CUDA)
 endif()
 
-if(MLX_BUILD_METAL AND NOT METAL_LIB)
-  message(STATUS "Metal not found. Unable to build GPU")
-  set(MLX_BUILD_METAL OFF)
-  set(MLX_METAL_DEBUG OFF)
-elseif(MLX_BUILD_METAL)
-  message(STATUS "Building METAL sources")
+if(MLX_BUILD_METAL)
+  find_library(METAL_LIB Metal)
+  find_library(FOUNDATION_LIB Foundation)
+  find_library(QUARTZ_LIB QuartzCore)
+  if(METAL_LIB)
+    message(STATUS "Metal found ${METAL_LIB}")
+  else()
+    message(
+      FATAL_ERROR
+        "Metal not found. Set MLX_BUILD_METAL=OFF to build without GPU")
+  endif()
 
   if(MLX_METAL_DEBUG)
     add_compile_definitions(MLX_METAL_DEBUG)
@@ -102,7 +117,8 @@ elseif(MLX_BUILD_METAL)
   # Throw an error if xcrun not found
   execute_process(
     COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
-    OUTPUT_VARIABLE MACOS_SDK_VERSION COMMAND_ERROR_IS_FATAL ANY)
+    OUTPUT_VARIABLE MACOS_SDK_VERSION
+    OUTPUT_STRIP_TRAILING_WHITESPACE COMMAND_ERROR_IS_FATAL ANY)
 
   if(${MACOS_SDK_VERSION} LESS 14.0)
     message(
@@ -112,9 +128,12 @@ elseif(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(
@@ -123,7 +142,6 @@ elseif(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}>
@@ -131,6 +149,12 @@ elseif(MLX_BUILD_METAL)
   target_link_libraries(mlx PUBLIC ${METAL_LIB} ${FOUNDATION_LIB} ${QUARTZ_LIB})
 endif()
 
+if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
+  # With newer clang/gcc versions following libs are implicitly linked, but when
+  # building on old distributions they need to be explicitly listed.
+  target_link_libraries(mlx PRIVATE dl pthread)
+endif()
+
 if(WIN32)
   if(MSVC)
     # GGUF does not build with MSVC.
@@ -158,7 +182,7 @@ if(MLX_BUILD_CPU)
     message(STATUS "Accelerate found ${ACCELERATE_LIBRARY}")
     set(MLX_BUILD_ACCELERATE ON)
   else()
-    message(STATUS "Accelerate or arm neon not found, using default backend.")
+    message(STATUS "Accelerate not found, using default backend.")
     set(MLX_BUILD_ACCELERATE OFF)
   endif()
 
@@ -234,12 +258,16 @@ target_include_directories(
 # Do not add mlx_EXPORTS define for shared library.
 set_target_properties(mlx PROPERTIES DEFINE_SYMBOL "")
 
-FetchContent_Declare(
-  fmt
-  GIT_REPOSITORY https://github.com/fmtlib/fmt.git
-  GIT_TAG 10.2.1
-  EXCLUDE_FROM_ALL)
-FetchContent_MakeAvailable(fmt)
+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
+    EXCLUDE_FROM_ALL)
+  FetchContent_MakeAvailable(fmt)
+endif()
 target_link_libraries(mlx PRIVATE $<BUILD_INTERFACE:fmt::fmt-header-only>)
 
 if(MLX_BUILD_PYTHON_BINDINGS)

README.md

@@ -2,7 +2,7 @@
 
 [**Quickstart**](#quickstart) | [**Installation**](#installation) |
 [**Documentation**](https://ml-explore.github.io/mlx/build/html/index.html) |
-[**Examples**](#examples) 
+[**Examples**](#examples)
 
 [![CircleCI](https://circleci.com/gh/ml-explore/mlx.svg?style=svg)](https://circleci.com/gh/ml-explore/mlx)
 
@@ -11,37 +11,37 @@ brought to you by Apple machine learning research.
 
 Some key features of MLX include:
 
- - **Familiar APIs**: MLX has a Python API that closely follows NumPy.  MLX
+- **Familiar APIs**: MLX has a Python API that closely follows NumPy. MLX
    also has fully featured C++, [C](https://github.com/ml-explore/mlx-c), and
    [Swift](https://github.com/ml-explore/mlx-swift/) APIs, which closely mirror
-   the Python API.  MLX has higher-level packages like `mlx.nn` and
+   the Python API. MLX has higher-level packages like `mlx.nn` and
    `mlx.optimizers` with APIs that closely follow PyTorch to simplify building
    more complex models.
 
- - **Composable function transformations**: MLX supports composable function
-   transformations for automatic differentiation, automatic vectorization,
-   and computation graph optimization.
+- **Composable function transformations**: MLX supports composable function
+  transformations for automatic differentiation, automatic vectorization,
+  and computation graph optimization.
 
- - **Lazy computation**: Computations in MLX are lazy. Arrays are only
-   materialized when needed.
+- **Lazy computation**: Computations in MLX are lazy. Arrays are only
+  materialized when needed.
 
- - **Dynamic graph construction**: Computation graphs in MLX are constructed
-   dynamically. Changing the shapes of function arguments does not trigger
-   slow compilations, and debugging is simple and intuitive.
+- **Dynamic graph construction**: Computation graphs in MLX are constructed
+  dynamically. Changing the shapes of function arguments does not trigger
+  slow compilations, and debugging is simple and intuitive.
 
- - **Multi-device**: Operations can run on any of the supported devices
-   (currently the CPU and the GPU).
+- **Multi-device**: Operations can run on any of the supported devices
+  (currently the CPU and the GPU).
 
- - **Unified memory**: A notable difference from MLX and other frameworks
-   is the *unified memory model*. Arrays in MLX live in shared memory.
-   Operations on MLX arrays can be performed on any of the supported
-   device types without transferring data.
+- **Unified memory**: A notable difference from MLX and other frameworks
+  is the *unified memory model*. Arrays in MLX live in shared memory.
+  Operations on MLX arrays can be performed on any of the supported
+  device types without transferring data.
 
 MLX is designed by machine learning researchers for machine learning
 researchers. The framework is intended to be user-friendly, but still efficient
 to train and deploy models. The design of the framework itself is also
 conceptually simple. We intend to make it easy for researchers to extend and
-improve MLX with the goal of quickly exploring new ideas. 
+improve MLX with the goal of quickly exploring new ideas.
 
 The design of MLX is inspired by frameworks like
 [NumPy](https://numpy.org/doc/stable/index.html),
@@ -68,25 +68,30 @@ in the documentation.
 
 ## Installation
 
-MLX is available on [PyPI](https://pypi.org/project/mlx/). To install the Python API, run:
+MLX is available on [PyPI](https://pypi.org/project/mlx/). To install MLX on
+macOS, run:
 
-**With `pip`**:
-
-```
+```bash
 pip install mlx
 ```
 
-**With `conda`**:
+To install the CUDA backend on Linux, run:
 
+```bash
+pip install mlx[cuda]
 ```
-conda install -c conda-forge mlx
+
+To install a CPU-only Linux package, run:
+
+```bash
+pip install mlx[cpu]
 ```
 
 Checkout the
 [documentation](https://ml-explore.github.io/mlx/build/html/install.html#)
 for more information on building the C++ and Python APIs from source.
 
-## Contributing 
+## Contributing
 
 Check out the [contribution guidelines](https://github.com/ml-explore/mlx/tree/main/CONTRIBUTING.md) for more information
 on contributing to MLX. See the
@@ -105,7 +110,7 @@ Hannun, Jagrit Digani, Angelos Katharopoulos, and Ronan Collobert. If you find
 MLX useful in your research and wish to cite it, please use the following
 BibTex entry:
 
-```
+```text
 @software{mlx2023,
   author = {Awni Hannun and Jagrit Digani and Angelos Katharopoulos and Ronan Collobert},
   title = {{MLX}: Efficient and flexible machine learning on Apple silicon},

benchmarks/cpp/irregular_strides.cpp

@@ -1,5 +1,6 @@
 // Copyright © 2023 Apple Inc.
 
+#include <cstring>
 #include <iostream>
 #include <sstream>
 
@@ -74,7 +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);
 
@@ -114,7 +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);
   };
@@ -169,7 +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);

benchmarks/cpp/single_ops.cpp

@@ -192,6 +192,22 @@ void time_reductions() {
 
   auto argmin_along_1 = [&a]() { return mx::argmin(a, 1, false); };
   TIME(argmin_along_1);
+
+  auto indices = mx::array({1});
+  auto updates = mx::reshape(mx::array({NAN}), {1, 1, 1});
+  std::vector<int> axes{0};
+  auto b = scatter(a, {indices}, updates, axes);
+  mx::eval(b);
+
+  auto max_along_0 = [&b]() { return mx::max(b, 0, false); };
+  TIME(max_along_0);
+  auto max_along_1 = [&b]() { return mx::max(b, 1, false); };
+  TIME(max_along_1);
+
+  auto min_along_0 = [&b]() { return mx::min(b, 0, false); };
+  TIME(min_along_0);
+  auto min_along_1 = [&b]() { return mx::min(b, 1, false); };
+  TIME(min_along_1);
 }
 
 void time_gather_scatter() {

benchmarks/python/blas/bench_gemm.py

@@ -142,9 +142,7 @@ def bench_shape(B, M, N, K, np_dtype, transpose="nn"):
     t_b = (0, 1, 2) if transpose[1] == "n" else (0, 2, 1)
 
     c_mlx = a_mx.transpose(t_a) @ b_mx.transpose(t_b)
-    c_npy = a_np.transpose(t_a).astype(np.float32) @ b_np.transpose(t_b).astype(
-        np.float32
-    )
+    c_npy = a_np.transpose(t_a).astype(np_dtype) @ b_np.transpose(t_b).astype(np_dtype)
 
     atol = 1e-5 if np_dtype == np.float32 else 1e-4
 
@@ -163,7 +161,7 @@ def get_gflop_count(B, M, N, K):
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description="Run gemm benchmarks")
 
-    dtypes = ("float32", "float16")
+    dtypes = ("float32", "float16", "complex64")
     transposes = ("nn", "nt", "tn")
     shapes = (
         (16, 234, 768, 3072),
@@ -187,7 +185,7 @@ if __name__ == "__main__":
                 diff = gflops_mx / gflops_pt - 1.0
 
                 print(
-                    f"{B:3d}, {M:4d}, {N:4d}, {K:4d}, {dtype}, {transpose}, {gflops_pt:05.3f}, {gflops_mx:05.3f}, {100. * diff:+5.2f}%"
+                    f"{B:3d}, {M:4d}, {N:4d}, {K:4d}, {dtype}, {transpose}, {gflops_pt:05.3f}, {gflops_mx:05.3f}, {100.0 * diff:+5.2f}%"
                 )
                 if gflops_pt >= 2.0 * gflops_mx:
                     print("ATTENTION ^^^^^^^")

benchmarks/python/blas/bench_gemv.py

@@ -1,6 +1,5 @@
 # Copyright © 2023 Apple Inc.
 
-import argparse
 import os
 import subprocess
 import time
@@ -196,7 +195,7 @@ def bench_with_out_len(ax, out_vec_len, in_vector_lens, dtype, transpose):
 
 
 for transpose in (False, True):
-    for dtype in ("float32", "float16"):
+    for dtype in ("float32", "float16", "complex64"):
         fig, axs = plt.subplots(
             len(in_vec_sizes), 2, figsize=(8.5, 11), layout="constrained"
         )
@@ -215,7 +214,7 @@ for transpose in (False, True):
         fig.suptitle(f"{device_name}: {dtype} {op_name}")
         fig.savefig(
             os.path.join(
-                results_dir, f'{device_name.replace(" ", "_")}_{dtype}_{op_name}.pdf'
+                results_dir, f"{device_name.replace(' ', '_')}_{dtype}_{op_name}.pdf"
             )
         )
         plt.close(fig)

benchmarks/python/comparative/bench_torch.py

@@ -5,6 +5,7 @@ import os
 import time
 
 import torch
+import torch.cuda
 import torch.mps
 
 
@@ -44,8 +45,10 @@ def bench(f, *args):
 
 
 def sync_if_needed(x):
-    if x.device != torch.device("cpu"):
+    if x.device == torch.device("mps"):
         torch.mps.synchronize()
+    elif x.device == torch.device("cuda"):
+        torch.cuda.synchronize()
 
 
 @torch.no_grad()
@@ -99,6 +102,14 @@ def reduction(op, axis, x):
     sync_if_needed(x)
 
 
+@torch.no_grad()
+def sum_and_add(axis, x, y):
+    z = x.sum(axis=axis, keepdims=True)
+    for i in range(50):
+        z = (z + y).sum(axis=axis, keepdims=True)
+    sync_if_needed(x)
+
+
 @torch.no_grad()
 def softmax(axis, x):
     ys = []
@@ -340,7 +351,11 @@ if __name__ == "__main__":
         args.axis.pop(0)
 
     torch.set_num_threads(1)
-    device = "cpu" if args.cpu else "mps"
+    device = "mps"
+    if torch.cuda.is_available():
+        device = "cuda"
+    if args.cpu:
+        device = "cpu"
 
     types = args.dtype
     if not types:
@@ -460,5 +475,8 @@ if __name__ == "__main__":
     elif args.benchmark == "selu":
         print(bench(selu, x))
 
+    elif args.benchmark == "sum_and_add":
+        print(bench(sum_and_add, axis, *xs))
+
     else:
         raise ValueError(f"Unknown benchmark `{args.benchmark}`.")

benchmarks/python/conv_unaligned_bench.py

@@ -0,0 +1,107 @@
+import math
+import time
+
+import mlx.core as mx
+import numpy as np
+import torch
+
+N_warmup = 10
+N_iter_bench = 100
+N_iter_func = 5
+
+
+def bench(f, a, b):
+    for i in range(N_warmup):
+        f(a, b)
+    torch.mps.synchronize()
+
+    s = time.perf_counter_ns()
+    for i in range(N_iter_bench):
+        f(a, b)
+    e = time.perf_counter_ns()
+    return (e - s) * 1e-9
+
+
+def make_mx_conv_2D(strides=(1, 1), padding=(0, 0), groups=1):
+    def mx_conv_2D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = mx.conv2d(a, b, stride=strides, padding=padding, groups=groups)
+            ys.append(y)
+        mx.eval(ys)
+        return ys
+
+    return mx_conv_2D
+
+
+def make_pt_conv_2D(strides=(1, 1), padding=(0, 0), groups=1):
+    @torch.no_grad()
+    def pt_conv_2D(a, b):
+        ys = []
+        for i in range(N_iter_func):
+            y = torch.conv2d(a, b, stride=strides, padding=padding, groups=groups)
+            ys.append(y)
+        torch.mps.synchronize()
+        return ys
+
+    return pt_conv_2D
+
+
+def bench_shape(N, H, W, C, kH, kW, O, strides, padding, groups, np_dtype):
+    scale = 1.0 / math.sqrt(kH * kH * C)
+    a_np = np.random.uniform(0, 0.5, (N, H, W, C)).astype(np_dtype)
+    b_np = np.random.uniform(-scale, scale, (O, kH, kW, int(C / groups))).astype(
+        np_dtype
+    )
+
+    a_mx = mx.array(a_np)
+    b_mx = mx.array(b_np)
+
+    a_pt = torch.from_numpy(a_np.transpose((0, 3, 1, 2))).to("mps")
+    b_pt = torch.from_numpy(b_np.transpose((0, 3, 1, 2))).to("mps")
+
+    torch.mps.synchronize()
+
+    f_mx = make_mx_conv_2D(strides, padding, groups)
+    f_pt = make_pt_conv_2D(strides, padding, groups)
+
+    time_torch = bench(f_pt, a_pt, b_pt)
+    time_mlx = bench(f_mx, a_mx, b_mx)
+
+    out_mx = mx.conv2d(a_mx, b_mx, stride=strides, padding=padding, groups=groups)
+    out_pt = torch.conv2d(
+        a_pt.to("cpu"), b_pt.to("cpu"), stride=strides, padding=padding, groups=groups
+    )
+    out_pt = torch.permute(out_pt, (0, 2, 3, 1))
+    out_pt = out_pt.numpy(force=True)
+
+    atol = 2e-5 if np_dtype == np.float32 else 1e-4
+
+    if not np.allclose(out_pt, out_mx, atol=atol):
+        print(
+            f"Failed at {(N, H, W, C)}, {(O, kH, kW, C)} [strides = {strides}, padding = {padding}, groups = {groups}] with max(|a - b|) = {np.max(np.abs(out_pt - out_mx))}"
+        )
+
+    return time_mlx, time_torch
+
+
+if __name__ == "__main__":
+    dtype = "float32"
+    shapes = (
+        (4, 32, 32, 21, 3, 3, 128),
+        (4, 32, 32, 21, 3, 3, 37),
+        (4, 32, 32, 370, 3, 3, 370),
+        (4, 32, 32, 370, 7, 7, 128),
+        (2, 320, 640, 21, 7, 7, 21),
+    )
+    for N, H, W, C, kh, kw, O in shapes:
+        time_mlx, time_torch = bench_shape(
+            N, H, W, C, kh, kw, O, (1, 1), (0, 0), 1, dtype
+        )
+        diff = time_torch / time_mlx - 1.0
+
+        print(
+            f"({N}, {H:3d}, {W:3d}, {C:3d}), ({O:3d}, {kh:2d}, {kw:2d}, {C:3d}), {dtype}, {100. * diff:+5.2f}%"
+        )
+        if time_mlx >= 2.0 * time_torch:
+            print("ATTENTION ^^^^^^^")

benchmarks/python/layer_norm_bench.py

@@ -1,5 +1,7 @@
 # Copyright © 2023-2024 Apple Inc.
 
+from functools import partial
+
 import mlx.core as mx
 import mlx.nn as nn
 from time_utils import time_fn
@@ -18,51 +20,63 @@ def layer_norm(x, w, b, eps):
     return y
 
 
-def time_layer_norm():
+def time_layer_norm(N, dt):
+    L = 1024
     f1 = lambda x, w, b, y: (layer_norm(x, w, b, 1e-5) * y).sum()
     f2 = lambda x, w, b, y: (mx.fast.layer_norm(x, w, b, 1e-5) * y).sum()
     g1 = mx.grad(f1, argnums=(0, 1, 2))
     g2 = mx.grad(f2, argnums=(0, 1, 2))
 
-    x = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
-    w = mx.random.uniform(shape=(4096,)).astype(mx.float16)
-    b = mx.random.uniform(shape=(4096,)).astype(mx.float16)
-    y = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    x = mx.random.uniform(shape=(8, L, N)).astype(dt)
+    w = mx.random.uniform(shape=(N,)).astype(dt)
+    b = mx.random.uniform(shape=(N,)).astype(dt)
+    y = mx.random.uniform(shape=(8, L, N)).astype(dt)
     mx.eval(x, w, b, y)
 
-    def layer_norm_loop(g, x, w, b):
+    def layer_norm_loop(f, x, w, b):
+        for _ in range(32):
+            x = f(x, w, b)
+        return x
+
+    time_fn(layer_norm_loop, partial(layer_norm, eps=1e-5), x, w, b)
+    time_fn(layer_norm_loop, partial(mx.fast.layer_norm, eps=1e-5), x, w, b)
+
+    def layer_norm_grad_loop(g, x, w, b):
         gx, gw, gb = x, w, b
         for _ in range(32):
             gx, gw, gb = g(gx, gw, gb, y)
         return gx, gw, gb
 
-    time_fn(layer_norm_loop, g1, x, w, b)
-    time_fn(layer_norm_loop, g2, x, w, b)
-    time_fn(layer_norm_loop, mx.compile(g1), x, w, b)
-    time_fn(layer_norm_loop, mx.compile(g2), x, w, b)
+    time_fn(layer_norm_grad_loop, g1, x, w, b)
+    time_fn(layer_norm_grad_loop, g2, x, w, b)
+    time_fn(layer_norm_grad_loop, mx.compile(g1), x, w, b)
+    time_fn(layer_norm_grad_loop, mx.compile(g2), x, w, b)
 
     f1 = lambda x, y: (layer_norm(x, None, None, 1e-5) * y).sum()
     f2 = lambda x, y: (mx.fast.layer_norm(x, None, None, 1e-5) * y).sum()
     g1 = mx.grad(f1, argnums=(0,))
     g2 = mx.grad(f2, argnums=(0,))
 
-    x = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
-    w = mx.random.uniform(shape=(4096,)).astype(mx.float16)
-    b = mx.random.uniform(shape=(4096,)).astype(mx.float16)
-    y = mx.random.uniform(shape=(8, 1024, 4096)).astype(mx.float16)
+    x = mx.random.uniform(shape=(8, L, N)).astype(dt)
+    w = mx.random.uniform(shape=(N,)).astype(dt)
+    b = mx.random.uniform(shape=(N,)).astype(dt)
+    y = mx.random.uniform(shape=(8, L, N)).astype(dt)
     mx.eval(x, w, b, y)
 
-    def layer_norm_loop(g, x):
+    def layer_norm_grad_x_loop(g, x):
         gx = x
         for _ in range(32):
             gx = g(gx, y)
         return gx
 
-    time_fn(layer_norm_loop, g1, x)
-    time_fn(layer_norm_loop, g2, x)
-    time_fn(layer_norm_loop, mx.compile(g1), x)
-    time_fn(layer_norm_loop, mx.compile(g2), x)
+    time_fn(layer_norm_grad_x_loop, g1, x)
+    time_fn(layer_norm_grad_x_loop, g2, x)
+    time_fn(layer_norm_grad_x_loop, mx.compile(g1), x)
+    time_fn(layer_norm_grad_x_loop, mx.compile(g2), x)
 
 
 if __name__ == "__main__":
-    time_layer_norm()
+    for dt in [mx.float32, mx.float16, mx.bfloat16]:
+        for n in [1024, 2048, 4096, 8192, 8192 + 1024]:
+            print(dt, n)
+            time_layer_norm(n, dt)

benchmarks/python/masked_scatter.py

@@ -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()

benchmarks/python/single_ops.py

@@ -51,6 +51,20 @@ def time_maximum():
     time_fn(mx.maximum, a, b)
 
 
+def time_max():
+    a = mx.random.uniform(shape=(32, 1024, 1024))
+    a[1, 1] = mx.nan
+    mx.eval(a)
+    time_fn(mx.max, a, 0)
+
+
+def time_min():
+    a = mx.random.uniform(shape=(32, 1024, 1024))
+    a[1, 1] = mx.nan
+    mx.eval(a)
+    time_fn(mx.min, a, 0)
+
+
 def time_negative():
     a = mx.random.uniform(shape=(10000, 1000))
     mx.eval(a)
@@ -108,6 +122,8 @@ if __name__ == "__main__":
 
     time_add()
     time_matmul()
+    time_min()
+    time_max()
     time_maximum()
     time_exp()
     time_negative()

cmake/FindNCCL.cmake

@@ -0,0 +1,54 @@
+# FindNCCL.cmake This module finds the NVIDIA NCCL library and its include
+# directories.
+
+set(NCCL_ROOT_DIR
+    $ENV{NCCL_ROOT_DIR}
+    CACHE PATH "Folder contains NVIDIA NCCL")
+
+find_path(
+  NCCL_INCLUDE_DIRS
+  NAMES nccl.h
+  HINTS ${NCCL_INCLUDE_DIR} ${NCCL_ROOT_DIR} ${NCCL_ROOT_DIR}/include
+        ${CUDA_TOOLKIT_ROOT_DIR}/include)
+
+if($ENV{USE_STATIC_NCCL})
+  message(
+    STATUS "USE_STATIC_NCCL detected. Linking against static NCCL library")
+  set(NCCL_LIBNAME "libnccl_static.a")
+else()
+  set(NCCL_LIBNAME "nccl")
+endif()
+
+find_library(
+  NCCL_LIBRARIES
+  NAMES ${NCCL_LIBNAME}
+  HINTS ${NCCL_LIB_DIR}
+        ${NCCL_ROOT_DIR}
+        ${NCCL_ROOT_DIR}/lib
+        ${NCCL_ROOT_DIR}/lib/x86_64-linux-gnu
+        ${NCCL_ROOT_DIR}/lib64
+        ${CUDA_TOOLKIT_ROOT_DIR}/lib
+        ${CUDA_TOOLKIT_ROOT_DIR}/lib64)
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(NCCL DEFAULT_MSG NCCL_INCLUDE_DIRS
+                                  NCCL_LIBRARIES)
+
+if(NCCL_FOUND)
+  set(NCCL_HEADER_FILE "${NCCL_INCLUDE_DIRS}/nccl.h")
+  message(
+    STATUS "Determining NCCL version from the header file: ${NCCL_HEADER_FILE}")
+  file(
+    STRINGS ${NCCL_HEADER_FILE} NCCL_MAJOR_VERSION_DEFINED
+    REGEX "^[ \t]*#define[ \t]+NCCL_MAJOR[ \t]+[0-9]+.*$"
+    LIMIT_COUNT 1)
+  if(NCCL_MAJOR_VERSION_DEFINED)
+    string(REGEX REPLACE "^[ \t]*#define[ \t]+NCCL_MAJOR[ \t]+" ""
+                         NCCL_MAJOR_VERSION ${NCCL_MAJOR_VERSION_DEFINED})
+    message(STATUS "NCCL_MAJOR_VERSION: ${NCCL_MAJOR_VERSION}")
+  endif()
+  message(
+    STATUS
+      "Found NCCL (include: ${NCCL_INCLUDE_DIRS}, library: ${NCCL_LIBRARIES})")
+  mark_as_advanced(NCCL_ROOT_DIR NCCL_INCLUDE_DIRS NCCL_LIBRARIES)
+endif()

cmake/Findnvpl.cmake

@@ -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.

docs/requirements.txt

@@ -1,4 +1,5 @@
 sphinx
 breathe
 sphinx-book-theme
+sphinx-copybutton
 mlx

docs/src/conf.py

@@ -18,6 +18,7 @@ release = version
 # -- General configuration ---------------------------------------------------
 
 extensions = [
+    "sphinx_copybutton",
     "sphinx.ext.autodoc",
     "sphinx.ext.autosummary",
     "sphinx.ext.intersphinx",

docs/src/dev/custom_metal_kernels.rst

@@ -8,23 +8,26 @@ MLX supports writing custom Metal kernels through the Python and C++ APIs.
 Simple Example
 --------------
 
+.. currentmodule:: mlx.core
+
 Let's write a custom kernel that computes ``exp`` elementwise:
 
 .. code-block:: python
 
-  def exp_elementwise(a: mx.array):
-      source = """
-          uint elem = thread_position_in_grid.x;
-          T tmp = inp[elem];
-          out[elem] = metal::exp(tmp);
-      """
+  source = """
+      uint elem = thread_position_in_grid.x;
+      T tmp = inp[elem];
+      out[elem] = metal::exp(tmp);
+  """
 
-      kernel = mx.fast.metal_kernel(
-          name="myexp",
-          input_names=["inp"],
-          output_names=["out"],
-          source=source,
-      )
+  kernel = mx.fast.metal_kernel(
+      name="myexp",
+      input_names=["inp"],
+      output_names=["out"],
+      source=source,
+  )
+
+  def exp_elementwise(a: mx.array):
       outputs = kernel(
           inputs=[a],
           template=[("T", mx.float32)],
@@ -39,8 +42,13 @@ Let's write a custom kernel that computes ``exp`` elementwise:
   b = exp_elementwise(a)
   assert mx.allclose(b, mx.exp(a))
 
+Every time you make a kernel, a new Metal library is created and possibly
+JIT compiled. To reduce the overhead from that, build the kernel once with
+:func:`fast.metal_kernel` and then use it many times.
+
 .. note::
-    We are only required to pass the body of the Metal kernel in ``source``.
+   Only pass the body of the Metal kernel in ``source``. The function
+   signature is generated automatically.
 
 The full function signature will be generated using:
 
@@ -78,44 +86,52 @@ Putting this all together, the generated function signature for ``myexp`` is as
 
   template [[host_name("custom_kernel_myexp_float")]] [[kernel]] decltype(custom_kernel_myexp_float<float>) custom_kernel_myexp_float<float>;
 
-Note: ``grid`` and ``threadgroup`` are parameters to the Metal `dispatchThreads <https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/2866532-dispatchthreads>`_ function.
-This means we will launch ``mx.prod(grid)`` threads, subdivided into ``threadgroup`` size threadgroups.
-For optimal performance, each thread group dimension should be less than or equal to the corresponding grid dimension.
+Note: ``grid`` and ``threadgroup`` are parameters to the Metal `dispatchThreads
+<https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/2866532-dispatchthreads>`_
+function. This means we will launch ``mx.prod(grid)`` threads, subdivided into
+``threadgroup`` size threadgroups.  For optimal performance, each thread group
+dimension should be less than or equal to the corresponding grid dimension.
 
-Passing ``verbose=True`` to ``mx.fast.metal_kernel.__call__`` will print the generated code for debugging purposes.
+Passing ``verbose=True`` to :func:`ast.metal_kernel.__call__` will print the
+generated code for debugging purposes.
 
 Using Shape/Strides
 -------------------
 
-``mx.fast.metal_kernel`` supports an argument ``ensure_row_contiguous`` which is ``True`` by default.
-This will copy the ``mx.array`` inputs if needed before the kernel is launched to ensure that the memory layout is row contiguous.
-Generally this makes writing the kernel easier, since we don't have to worry about gaps or the ordering of the dims
-when indexing.
+:func:`fast.metal_kernel` supports an argument ``ensure_row_contiguous`` which
+is ``True`` by default. This will copy the array inputs if needed
+before the kernel is launched to ensure that the memory layout is row
+contiguous.  Generally this makes writing the kernel easier, since we don't
+have to worry about gaps or the ordering of the dims when indexing.
 
-If we want to avoid this copy, ``metal_kernel`` automatically passes ``a_shape``, ``a_strides`` and ``a_ndim`` for each
-input array ``a`` if any are present in ``source``.
-We can then use MLX's built in indexing utils to fetch the right elements for each thread.
+If we want to avoid this copy, :func:`fast.metal_kernel` automatically passes
+``a_shape``, ``a_strides`` and ``a_ndim`` for each input array ``a`` if any are
+present in ``source``. We can then use MLX's built in indexing utils to fetch
+the right elements for each thread.
 
-Let's convert ``myexp`` above to support arbitrarily strided arrays without relying on a copy from ``ensure_row_contiguous``:
+Let's convert ``myexp`` above to support arbitrarily strided arrays without
+relying on a copy from ``ensure_row_contiguous``:
 
 .. code-block:: python
+   
+  source = """
+      uint elem = thread_position_in_grid.x;
+      // Utils from `mlx/backend/metal/kernels/utils.h` are automatically included
+      uint loc = elem_to_loc(elem, inp_shape, inp_strides, inp_ndim);
+      T tmp = inp[loc];
+      // Output arrays are always row contiguous
+      out[elem] = metal::exp(tmp);
+  """
+
+  kernel = mx.fast.metal_kernel(
+      name="myexp_strided",
+      input_names=["inp"],
+      output_names=["out"],
+      source=source,
+      ensure_row_contiguous=False,
+  )
 
   def exp_elementwise(a: mx.array):
-      source = """
-          uint elem = thread_position_in_grid.x;
-          // Utils from `mlx/backend/metal/kernels/utils.h` are automatically included
-          uint loc = elem_to_loc(elem, inp_shape, inp_strides, inp_ndim);
-          T tmp = inp[loc];
-          // Output arrays are always row contiguous
-          out[elem] = metal::exp(tmp);
-      """
-
-      kernel = mx.fast.metal_kernel(
-          name="myexp_strided",
-          input_names=["inp"],
-          output_names=["out"],
-          source=source
-      )
       outputs = kernel(
           inputs=[a],
           template=[("T", mx.float32)],
@@ -123,7 +139,6 @@ Let's convert ``myexp`` above to support arbitrarily strided arrays without rely
           threadgroup=(256, 1, 1),
           output_shapes=[a.shape],
           output_dtypes=[a.dtype],
-          ensure_row_contiguous=False,
       )
       return outputs[0]
 
@@ -142,137 +157,139 @@ We'll start with the following MLX implementation using standard ops:
 
 .. code-block:: python
 
-    def grid_sample_ref(x, grid):
-        N, H_in, W_in, _ = x.shape
-        ix = ((grid[..., 0] + 1) * W_in - 1) / 2
-        iy = ((grid[..., 1] + 1) * H_in - 1) / 2
+  def grid_sample_ref(x, grid):
+      N, H_in, W_in, _ = x.shape
+      ix = ((grid[..., 0] + 1) * W_in - 1) / 2
+      iy = ((grid[..., 1] + 1) * H_in - 1) / 2
 
-        ix_nw = mx.floor(ix).astype(mx.int32)
-        iy_nw = mx.floor(iy).astype(mx.int32)
+      ix_nw = mx.floor(ix).astype(mx.int32)
+      iy_nw = mx.floor(iy).astype(mx.int32)
 
-        ix_ne = ix_nw + 1
-        iy_ne = iy_nw
+      ix_ne = ix_nw + 1
+      iy_ne = iy_nw
 
-        ix_sw = ix_nw
-        iy_sw = iy_nw + 1
+      ix_sw = ix_nw
+      iy_sw = iy_nw + 1
 
-        ix_se = ix_nw + 1
-        iy_se = iy_nw + 1
+      ix_se = ix_nw + 1
+      iy_se = iy_nw + 1
 
-        nw = (ix_se - ix)    * (iy_se - iy)
-        ne = (ix    - ix_sw) * (iy_sw - iy)
-        sw = (ix_ne - ix)    * (iy    - iy_ne)
-        se = (ix    - ix_nw) * (iy    - iy_nw)
+      nw = (ix_se - ix)    * (iy_se - iy)
+      ne = (ix    - ix_sw) * (iy_sw - iy)
+      sw = (ix_ne - ix)    * (iy    - iy_ne)
+      se = (ix    - ix_nw) * (iy    - iy_nw)
 
-        I_nw = x[mx.arange(N)[:, None, None], iy_nw, ix_nw, :]
-        I_ne = x[mx.arange(N)[:, None, None], iy_ne, ix_ne, :]
-        I_sw = x[mx.arange(N)[:, None, None], iy_sw, ix_sw, :]
-        I_se = x[mx.arange(N)[:, None, None], iy_se, ix_se, :]
+      I_nw = x[mx.arange(N)[:, None, None], iy_nw, ix_nw, :]
+      I_ne = x[mx.arange(N)[:, None, None], iy_ne, ix_ne, :]
+      I_sw = x[mx.arange(N)[:, None, None], iy_sw, ix_sw, :]
+      I_se = x[mx.arange(N)[:, None, None], iy_se, ix_se, :]
 
-        mask_nw = (iy_nw >= 0) & (iy_nw <= H_in - 1) & (ix_nw >= 0) & (ix_nw <= W_in - 1)
-        mask_ne = (iy_ne >= 0) & (iy_ne <= H_in - 1) & (ix_ne >= 0) & (ix_ne <= W_in - 1)
-        mask_sw = (iy_sw >= 0) & (iy_sw <= H_in - 1) & (ix_sw >= 0) & (ix_sw <= W_in - 1)
-        mask_se = (iy_se >= 0) & (iy_se <= H_in - 1) & (ix_se >= 0) & (ix_se <= W_in - 1)
+      mask_nw = (iy_nw >= 0) & (iy_nw <= H_in - 1) & (ix_nw >= 0) & (ix_nw <= W_in - 1)
+      mask_ne = (iy_ne >= 0) & (iy_ne <= H_in - 1) & (ix_ne >= 0) & (ix_ne <= W_in - 1)
+      mask_sw = (iy_sw >= 0) & (iy_sw <= H_in - 1) & (ix_sw >= 0) & (ix_sw <= W_in - 1)
+      mask_se = (iy_se >= 0) & (iy_se <= H_in - 1) & (ix_se >= 0) & (ix_se <= W_in - 1)
 
-        I_nw *= mask_nw[..., None]
-        I_ne *= mask_ne[..., None]
-        I_sw *= mask_sw[..., None]
-        I_se *= mask_se[..., None]
+      I_nw *= mask_nw[..., None]
+      I_ne *= mask_ne[..., None]
+      I_sw *= mask_sw[..., None]
+      I_se *= mask_se[..., None]
 
-        output = nw[..., None] * I_nw + ne[..., None] * I_ne + sw[..., None] * I_sw + se[..., None] * I_se
+      output = nw[..., None] * I_nw + ne[..., None] * I_ne + sw[..., None] * I_sw + se[..., None] * I_se
 
-        return output
+      return output
 
-Now let's use ``mx.custom_function`` together with ``mx.fast.metal_kernel``
+Now let's use :func:`custom_function` together with :func:`fast.metal_kernel`
 to write a fast GPU kernel for both the forward and backward passes.
 
 First we'll implement the forward pass as a fused kernel:
 
 .. code-block:: python
 
-    @mx.custom_function
-    def grid_sample(x, grid):
+  source = """
+      uint elem = thread_position_in_grid.x;
+      int H = x_shape[1];
+      int W = x_shape[2];
+      int C = x_shape[3];
+      int gH = grid_shape[1];
+      int gW = grid_shape[2];
 
-        assert x.ndim == 4, "`x` must be 4D."
-        assert grid.ndim == 4, "`grid` must be 4D."
+      int w_stride = C;
+      int h_stride = W * w_stride;
+      int b_stride = H * h_stride;
 
-        B, _, _, C = x.shape
-        _, gN, gM, D = grid.shape
-        out_shape = (B, gN, gM, C)
+      uint grid_idx = elem / C * 2;
+      float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
+      float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
 
-        assert D == 2, "Last dim of `grid` must be size 2."
+      int ix_nw = floor(ix);
+      int iy_nw = floor(iy);
 
-        source = """
-            uint elem = thread_position_in_grid.x;
-            int H = x_shape[1];
-            int W = x_shape[2];
-            int C = x_shape[3];
-            int gH = grid_shape[1];
-            int gW = grid_shape[2];
+      int ix_ne = ix_nw + 1;
+      int iy_ne = iy_nw;
 
-            int w_stride = C;
-            int h_stride = W * w_stride;
-            int b_stride = H * h_stride;
+      int ix_sw = ix_nw;
+      int iy_sw = iy_nw + 1;
 
-            uint grid_idx = elem / C * 2;
-            float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
-            float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
+      int ix_se = ix_nw + 1;
+      int iy_se = iy_nw + 1;
 
-            int ix_nw = floor(ix);
-            int iy_nw = floor(iy);
+      T nw = (ix_se - ix)    * (iy_se - iy);
+      T ne = (ix    - ix_sw) * (iy_sw - iy);
+      T sw = (ix_ne - ix)    * (iy    - iy_ne);
+      T se = (ix    - ix_nw) * (iy    - iy_nw);
 
-            int ix_ne = ix_nw + 1;
-            int iy_ne = iy_nw;
+      int batch_idx = elem / C / gH / gW * b_stride;
+      int channel_idx = elem % C;
+      int base_idx = batch_idx + channel_idx;
 
-            int ix_sw = ix_nw;
-            int iy_sw = iy_nw + 1;
+      T I_nw = x[base_idx + iy_nw * h_stride + ix_nw * w_stride];
+      T I_ne = x[base_idx + iy_ne * h_stride + ix_ne * w_stride];
+      T I_sw = x[base_idx + iy_sw * h_stride + ix_sw * w_stride];
+      T I_se = x[base_idx + iy_se * h_stride + ix_se * w_stride];
 
-            int ix_se = ix_nw + 1;
-            int iy_se = iy_nw + 1;
+      I_nw = iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1 ? I_nw : 0;
+      I_ne = iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1 ? I_ne : 0;
+      I_sw = iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1 ? I_sw : 0;
+      I_se = iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1 ? I_se : 0;
 
-            T nw = (ix_se - ix)    * (iy_se - iy);
-            T ne = (ix    - ix_sw) * (iy_sw - iy);
-            T sw = (ix_ne - ix)    * (iy    - iy_ne);
-            T se = (ix    - ix_nw) * (iy    - iy_nw);
+      out[elem] = nw * I_nw + ne * I_ne + sw * I_sw + se * I_se;
+  """
 
-            int batch_idx = elem / C / gH / gW * b_stride;
-            int channel_idx = elem % C;
-            int base_idx = batch_idx + channel_idx;
+  kernel = mx.fast.metal_kernel(
+      name="grid_sample",
+      input_names=["x", "grid"],
+      output_names=["out"],
+      source=source,
+  )
 
-            T I_nw = x[base_idx + iy_nw * h_stride + ix_nw * w_stride];
-            T I_ne = x[base_idx + iy_ne * h_stride + ix_ne * w_stride];
-            T I_sw = x[base_idx + iy_sw * h_stride + ix_sw * w_stride];
-            T I_se = x[base_idx + iy_se * h_stride + ix_se * w_stride];
+  @mx.custom_function
+  def grid_sample(x, grid):
 
-            I_nw = iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1 ? I_nw : 0;
-            I_ne = iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1 ? I_ne : 0;
-            I_sw = iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1 ? I_sw : 0;
-            I_se = iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1 ? I_se : 0;
+      assert x.ndim == 4, "`x` must be 4D."
+      assert grid.ndim == 4, "`grid` must be 4D."
 
-            out[elem] = nw * I_nw + ne * I_ne + sw * I_sw + se * I_se;
-        """
-        kernel = mx.fast.metal_kernel(
-            name="grid_sample",
-            input_names=["x", "grid"],
-            output_names=["out"],
-            source=source,
-        )
-        outputs = kernel(
-            inputs=[x, grid],
-            template=[("T", x.dtype)],
-            output_shapes=[out_shape],
-            output_dtypes=[x.dtype],
-            grid=(np.prod(out_shape), 1, 1),
-            threadgroup=(256, 1, 1),
-        )
-        return outputs[0]
+      B, _, _, C = x.shape
+      _, gN, gM, D = grid.shape
+      out_shape = (B, gN, gM, C)
+
+      assert D == 2, "Last dim of `grid` must be size 2."
+
+      outputs = kernel(
+          inputs=[x, grid],
+          template=[("T", x.dtype)],
+          output_shapes=[out_shape],
+          output_dtypes=[x.dtype],
+          grid=(np.prod(out_shape), 1, 1),
+          threadgroup=(256, 1, 1),
+      )
+      return outputs[0]
 
 For a reasonably sized input such as:
 
 .. code-block:: python
 
-    x.shape = (8, 1024, 1024, 64)
-    grid.shape = (8, 256, 256, 2)
+  x.shape = (8, 1024, 1024, 64)
+  grid.shape = (8, 256, 256, 2)
 
 On an M1 Max, we see a big performance improvement:
 
@@ -281,11 +298,11 @@ On an M1 Max, we see a big performance improvement:
 Grid Sample VJP
 ---------------
 
-Since we decorated ``grid_sample`` with ``mx.custom_function``, we can now define
-its custom vjp transform so MLX can differentiate it.
+Since we decorated ``grid_sample`` with :func:`custom_function`, we can now
+define its custom vjp transform so MLX can differentiate it.
 
 The backwards pass requires atomically updating ``x_grad``/``grid_grad`` and so
-requires a few extra ``mx.fast.metal_kernel`` features:
+requires a few extra :func:`fast.metal_kernel` features:
 
 * ``init_value=0``
     Initialize all of the kernel's outputs to this value before it runs. This allows us to update only part of the output arrays with the kernel.
@@ -299,128 +316,129 @@ We can then implement the backwards pass as follows:
 
 .. code-block:: python
 
-    @grid_sample.vjp
-    def grid_sample_vjp(primals, cotangent, _):
-        x, grid = primals
-        B, _, _, C = x.shape
-        _, gN, gM, D = grid.shape
+  source = """
+      uint elem = thread_position_in_grid.x;
+      int H = x_shape[1];
+      int W = x_shape[2];
+      int C = x_shape[3];
+      // Pad C to the nearest larger simdgroup size multiple
+      int C_padded = ceildiv(C, threads_per_simdgroup) * threads_per_simdgroup;
 
-        assert D == 2, "Last dim of `grid` must be size 2."
+      int gH = grid_shape[1];
+      int gW = grid_shape[2];
 
-        source = """
-            uint elem = thread_position_in_grid.x;
-            int H = x_shape[1];
-            int W = x_shape[2];
-            int C = x_shape[3];
-            // Pad C to the nearest larger simdgroup size multiple
-            int C_padded = ceildiv(C, threads_per_simdgroup) * threads_per_simdgroup;
+      int w_stride = C;
+      int h_stride = W * w_stride;
+      int b_stride = H * h_stride;
 
-            int gH = grid_shape[1];
-            int gW = grid_shape[2];
+      uint grid_idx = elem / C_padded * 2;
+      float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
+      float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
 
-            int w_stride = C;
-            int h_stride = W * w_stride;
-            int b_stride = H * h_stride;
+      int ix_nw = floor(ix);
+      int iy_nw = floor(iy);
 
-            uint grid_idx = elem / C_padded * 2;
-            float ix = ((grid[grid_idx] + 1) * W - 1) / 2;
-            float iy = ((grid[grid_idx + 1] + 1) * H - 1) / 2;
+      int ix_ne = ix_nw + 1;
+      int iy_ne = iy_nw;
 
-            int ix_nw = floor(ix);
-            int iy_nw = floor(iy);
+      int ix_sw = ix_nw;
+      int iy_sw = iy_nw + 1;
 
-            int ix_ne = ix_nw + 1;
-            int iy_ne = iy_nw;
+      int ix_se = ix_nw + 1;
+      int iy_se = iy_nw + 1;
 
-            int ix_sw = ix_nw;
-            int iy_sw = iy_nw + 1;
+      T nw = (ix_se - ix)    * (iy_se - iy);
+      T ne = (ix    - ix_sw) * (iy_sw - iy);
+      T sw = (ix_ne - ix)    * (iy    - iy_ne);
+      T se = (ix    - ix_nw) * (iy    - iy_nw);
 
-            int ix_se = ix_nw + 1;
-            int iy_se = iy_nw + 1;
+      int batch_idx = elem / C_padded / gH / gW * b_stride;
+      int channel_idx = elem % C_padded;
+      int base_idx = batch_idx + channel_idx;
 
-            T nw = (ix_se - ix)    * (iy_se - iy);
-            T ne = (ix    - ix_sw) * (iy_sw - iy);
-            T sw = (ix_ne - ix)    * (iy    - iy_ne);
-            T se = (ix    - ix_nw) * (iy    - iy_nw);
+      T gix = T(0);
+      T giy = T(0);
+      if (channel_idx < C) {
+          int cot_index = elem / C_padded * C + channel_idx;
+          T cot = cotangent[cot_index];
+          if (iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1) {
+              int offset = base_idx + iy_nw * h_stride + ix_nw * w_stride;
+              atomic_fetch_add_explicit(&x_grad[offset], nw * cot, memory_order_relaxed);
 
-            int batch_idx = elem / C_padded / gH / gW * b_stride;
-            int channel_idx = elem % C_padded;
-            int base_idx = batch_idx + channel_idx;
+              T I_nw = x[offset];
+              gix -= I_nw * (iy_se - iy) * cot;
+              giy -= I_nw * (ix_se - ix) * cot;
+          }
+          if (iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1) {
+              int offset = base_idx + iy_ne * h_stride + ix_ne * w_stride;
+              atomic_fetch_add_explicit(&x_grad[offset], ne * cot, memory_order_relaxed);
 
-            T gix = T(0);
-            T giy = T(0);
-            if (channel_idx < C) {
-                int cot_index = elem / C_padded * C + channel_idx;
-                T cot = cotangent[cot_index];
-                if (iy_nw >= 0 && iy_nw <= H - 1 && ix_nw >= 0 && ix_nw <= W - 1) {
-                    int offset = base_idx + iy_nw * h_stride + ix_nw * w_stride;
-                    atomic_fetch_add_explicit(&x_grad[offset], nw * cot, memory_order_relaxed);
+              T I_ne = x[offset];
+              gix += I_ne * (iy_sw - iy) * cot;
+              giy -= I_ne * (ix - ix_sw) * cot;
+          }
+          if (iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1) {
+              int offset = base_idx + iy_sw * h_stride + ix_sw * w_stride;
+              atomic_fetch_add_explicit(&x_grad[offset], sw * cot, memory_order_relaxed);
 
-                    T I_nw = x[offset];
-                    gix -= I_nw * (iy_se - iy) * cot;
-                    giy -= I_nw * (ix_se - ix) * cot;
-                }
-                if (iy_ne >= 0 && iy_ne <= H - 1 && ix_ne >= 0 && ix_ne <= W - 1) {
-                    int offset = base_idx + iy_ne * h_stride + ix_ne * w_stride;
-                    atomic_fetch_add_explicit(&x_grad[offset], ne * cot, memory_order_relaxed);
+              T I_sw = x[offset];
+              gix -= I_sw * (iy - iy_ne) * cot;
+              giy += I_sw * (ix_ne - ix) * cot;
+          }
+          if (iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1) {
+              int offset = base_idx + iy_se * h_stride + ix_se * w_stride;
+              atomic_fetch_add_explicit(&x_grad[offset], se * cot, memory_order_relaxed);
 
-                    T I_ne = x[offset];
-                    gix += I_ne * (iy_sw - iy) * cot;
-                    giy -= I_ne * (ix - ix_sw) * cot;
-                }
-                if (iy_sw >= 0 && iy_sw <= H - 1 && ix_sw >= 0 && ix_sw <= W - 1) {
-                    int offset = base_idx + iy_sw * h_stride + ix_sw * w_stride;
-                    atomic_fetch_add_explicit(&x_grad[offset], sw * cot, memory_order_relaxed);
+              T I_se = x[offset];
+              gix += I_se * (iy - iy_nw) * cot;
+              giy += I_se * (ix - ix_nw) * cot;
+          }
+      }
 
-                    T I_sw = x[offset];
-                    gix -= I_sw * (iy - iy_ne) * cot;
-                    giy += I_sw * (ix_ne - ix) * cot;
-                }
-                if (iy_se >= 0 && iy_se <= H - 1 && ix_se >= 0 && ix_se <= W - 1) {
-                    int offset = base_idx + iy_se * h_stride + ix_se * w_stride;
-                    atomic_fetch_add_explicit(&x_grad[offset], se * cot, memory_order_relaxed);
+      T gix_mult = W / 2;
+      T giy_mult = H / 2;
 
-                    T I_se = x[offset];
-                    gix += I_se * (iy - iy_nw) * cot;
-                    giy += I_se * (ix - ix_nw) * cot;
-                }
-            }
+      // Reduce across each simdgroup first.
+      // This is much faster than relying purely on atomics.
+      gix = simd_sum(gix);
+      giy = simd_sum(giy);
 
-            T gix_mult = W / 2;
-            T giy_mult = H / 2;
+      if (thread_index_in_simdgroup == 0) {
+          atomic_fetch_add_explicit(&grid_grad[grid_idx], gix * gix_mult, memory_order_relaxed);
+          atomic_fetch_add_explicit(&grid_grad[grid_idx + 1], giy * giy_mult, memory_order_relaxed);
+      }
+  """
+  kernel = mx.fast.metal_kernel(
+      name="grid_sample_grad",
+      input_names=["x", "grid", "cotangent"],
+      output_names=["x_grad", "grid_grad"],
+      source=source,
+      atomic_outputs=True,
+  )
 
-            // Reduce across each simdgroup first.
-            // This is much faster than relying purely on atomics.
-            gix = simd_sum(gix);
-            giy = simd_sum(giy);
+  @grid_sample.vjp
+  def grid_sample_vjp(primals, cotangent, _):
+      x, grid = primals
+      B, _, _, C = x.shape
+      _, gN, gM, D = grid.shape
 
-            if (thread_index_in_simdgroup == 0) {
-                atomic_fetch_add_explicit(&grid_grad[grid_idx], gix * gix_mult, memory_order_relaxed);
-                atomic_fetch_add_explicit(&grid_grad[grid_idx + 1], giy * giy_mult, memory_order_relaxed);
-            }
-        """
-        kernel = mx.fast.metal_kernel(
-            name="grid_sample_grad",
-            input_names=["x", "grid", "cotangent"],
-            output_names=["x_grad", "grid_grad"],
-            source=source,
-            atomic_outputs=True,
-        )
-        # pad the output channels to simd group size
-        # so that our `simd_sum`s don't overlap.
-        simdgroup_size = 32
-        C_padded = (C + simdgroup_size - 1) // simdgroup_size * simdgroup_size
-        grid_size = B * gN * gM * C_padded
-        outputs = kernel(
-            inputs=[x, grid, cotangent],
-            template=[("T", x.dtype)],
-            output_shapes=[x.shape, grid.shape],
-            output_dtypes=[x.dtype, x.dtype],
-            grid=(grid_size, 1, 1),
-            threadgroup=(256, 1, 1),
-            init_value=0,
-        )
-        return outputs[0], outputs[1]
+      assert D == 2, "Last dim of `grid` must be size 2."
+
+      # pad the output channels to simd group size
+      # so that our `simd_sum`s don't overlap.
+      simdgroup_size = 32
+      C_padded = (C + simdgroup_size - 1) // simdgroup_size * simdgroup_size
+      grid_size = B * gN * gM * C_padded
+      outputs = kernel(
+          inputs=[x, grid, cotangent],
+          template=[("T", x.dtype)],
+          output_shapes=[x.shape, grid.shape],
+          output_dtypes=[x.dtype, x.dtype],
+          grid=(grid_size, 1, 1),
+          threadgroup=(256, 1, 1),
+          init_value=0,
+      )
+      return outputs[0], outputs[1]
 
 There's an even larger speed up for the vjp:
 

docs/src/dev/extensions.rst

@@ -138,13 +138,13 @@ more concrete:
         * representing the vectorized computation and the axis which
         * corresponds to the output vectorized dimension.
         */
-        virtual std::pair<std::vector<array>, std::vector<int>> vmap(
+        std::pair<std::vector<array>, std::vector<int>> vmap(
             const std::vector<array>& inputs,
             const std::vector<int>& axes) override;
 
-        /** Print the primitive. */
-        void print(std::ostream& os) override {
-            os << "Axpby";
+        /** The name of primitive. */
+        const char* name() const override {
+          return "Axpby";
         }
 
         /** Equivalence check **/
@@ -394,14 +394,14 @@ below.
         out.set_data(allocator::malloc(out.nbytes()));
 
         // Resolve name of kernel
-        std::ostringstream kname;
-        kname << "axpby_" << "general_" << type_to_name(out);
+        std::stream kname;
+        kname = "axpby_general_" + type_to_name(out);
 
-        // Make sure the metal library is available
-        d.register_library("mlx_ext");
+        // Load the metal library
+        auto lib = d.get_library("mlx_ext", current_binary_dir());
 
         // Make a kernel from this metal library
-        auto kernel = d.get_kernel(kname.str(), "mlx_ext");
+        auto kernel = d.get_kernel(kname, lib);
 
         // Prepare to encode kernel
         auto& compute_encoder = d.get_command_encoder(s.index);

docs/src/index.rst

@@ -70,6 +70,7 @@ are the CPU and GPU.
    python/fft
    python/linalg
    python/metal
+   python/cuda
    python/memory_management
    python/nn
    python/optimizers

docs/src/install.rst

@@ -13,22 +13,48 @@ silicon computer is
 
     pip install mlx
 
-To install from PyPI you must meet the following requirements:
+To install from PyPI your system must meet the following requirements:
 
 - Using an M series chip (Apple silicon)
-- Using a native Python >= 3.9
-- macOS >= 13.5
+- Using a native Python >= 3.10
+- 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
+^^^^
 
-MLX is also available on conda-forge. To install MLX with conda do:
+MLX has a CUDA backend which you can install with:
 
 .. code-block:: shell
 
-   conda install conda-forge::mlx
+    pip install mlx[cuda]
+
+To install the CUDA package from PyPi your system must meet the following
+requirements:
+
+- Nvidia architecture >= SM 7.0 (Volta)
+- Nvidia driver >= 550.54.14
+- CUDA toolkit >= 12.0
+- Linux distribution with glibc >= 2.35
+- Python >= 3.10
+
+
+CPU-only (Linux)
+^^^^^^^^^^^^^^^^
+
+For a CPU-only version of MLX that runs on Linux use:
+
+.. code-block:: shell
+
+    pip install mlx[cpu]
+
+To install the CPU-only package from PyPi your system must meet the following
+requirements:
+
+- Linux distribution with glibc >= 2.35
+- Python >= 3.10
 
 
 Troubleshooting
@@ -65,6 +91,8 @@ Build Requirements
 Python API
 ^^^^^^^^^^
 
+.. _python install:
+
 To build and install the MLX python library from source, first, clone MLX from
 `its GitHub repo <https://github.com/ml-explore/mlx>`_:
 
@@ -76,20 +104,20 @@ Then simply build and install MLX using pip:
 
 .. code-block:: shell
 
-  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install .
+  pip install .
 
 For developing, install the package with development dependencies, and use an
 editable install:
 
 .. code-block:: shell
 
-  CMAKE_BUILD_PARALLEL_LEVEL=8 pip install -e ".[dev]"
+  pip install -e ".[dev]"
 
 Once the development dependencies are installed, you can build faster with:
 
 .. code-block:: shell
 
- CMAKE_BUILD_PARALLEL_LEVEL=8 python setup.py build_ext --inplace
+ python setup.py build_ext --inplace
 
 Run the tests with:
 
@@ -107,6 +135,8 @@ IDE:
 C++ API
 ^^^^^^^
 
+.. _cpp install:
+
 Currently, MLX must be built and installed from source.
 
 Similarly to the python library, to build and install the MLX C++ library start
@@ -185,6 +215,7 @@ should point to the path to the built metal library.
 
       xcrun -sdk macosx --show-sdk-version
 
+
 Binary Size Minimization
 ~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -213,6 +244,50 @@ be anwywhere from a few hundred millisecond to a few seconds depending on the
 application. Once a kernel is compiled, it will be cached by the system. The
 Metal kernel cache persists across reboots.
 
+Linux
+^^^^^
+
+To build from source on Linux (CPU only), install the BLAS and LAPACK headers.
+For example on Ubuntu, run the following:
+
+.. code-block:: shell
+
+   apt-get update -y
+   apt-get install libblas-dev liblapack-dev liblapacke-dev -y
+
+From here follow the instructions to install either the :ref:`Python <python
+install>` or :ref:`C++ <cpp install>` APIs.
+
+CUDA
+^^^^
+
+To build from source on Linux with CUDA, install the BLAS and LAPACK headers
+and the CUDA toolkit. For example on Ubuntu, run the following:
+
+.. code-block:: shell
+
+   wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+   dpkg -i cuda-keyring_1.1-1_all.deb
+   apt-get update -y
+   apt-get -y install cuda-toolkit-12-9
+   apt-get install libblas-dev liblapack-dev liblapacke-dev libcudnn9-dev-cuda-12 -y
+
+
+When building either the Python or C++ APIs make sure to pass the cmake flag
+``MLX_BUILD_CUDA=ON``. For example, to build the Python API run:
+
+.. code-block:: shell
+
+  CMAKE_ARGS="-DMLX_BUILD_CUDA=ON" pip install -e ".[dev]"
+
+To build the C++ package run:
+
+.. code-block:: shell
+
+   mkdir -p build && cd build
+   cmake .. -DMLX_BUILD_CUDA=ON && make -j
+
+
 Troubleshooting
 ^^^^^^^^^^^^^^^
 

docs/src/python/cuda.rst

@@ -0,0 +1,9 @@
+CUDA
+=====
+
+.. currentmodule:: mlx.core.cuda
+
+.. autosummary::
+  :toctree: _autosummary
+
+  is_available

docs/src/python/fast.rst

@@ -13,3 +13,4 @@ Fast
   rope
   scaled_dot_product_attention
   metal_kernel
+  cuda_kernel

docs/src/python/nn/functions.rst

@@ -27,6 +27,7 @@ simple functions.
    mish
    prelu
    relu
+   relu2
    relu6
    selu
    sigmoid

docs/src/python/nn/layers.rst

@@ -50,6 +50,7 @@ Layers
    QuantizedLinear
    RMSNorm
    ReLU
+   ReLU2
    ReLU6
    RNN
    RoPE

docs/src/python/ops.rst

@@ -112,6 +112,7 @@ Operations
    max
    maximum
    mean
+   median
    meshgrid
    min
    minimum

docs/src/python/optimizers.rst

@@ -51,14 +51,14 @@ the saved state. Here's a simple example:
    optimizer.update(model, grads)
 
    # Save the state
-   state = tree_flatten(optimizer.state)
-   mx.save_safetensors("optimizer.safetensors", dict(state))
+   state = tree_flatten(optimizer.state, destination={})
+   mx.save_safetensors("optimizer.safetensors", state)
 
    # Later on, for example when loading from a checkpoint,
    # recreate the optimizer and load the state
    optimizer = optim.Adam(learning_rate=1e-2)
 
-   state = tree_unflatten(list(mx.load("optimizer.safetensors").items()))
+   state = tree_unflatten(mx.load("optimizer.safetensors"))
    optimizer.state = state
 
 Note, not every optimizer configuation parameter is saved in the state. For

docs/src/python/optimizers/common_optimizers.rst

@@ -19,3 +19,4 @@ Common Optimizers
    Adamax
    Lion
    MultiOptimizer
+   Muon

docs/src/usage/compile.rst

@@ -130,8 +130,8 @@ Now make an array, and benchmark both functions:
 .. code-block:: python
 
   x = mx.random.uniform(shape=(32, 1000, 4096))
-  timeit(nn.gelu, x)
-  timeit(mx.compile(nn.gelu), x)
+  timeit(gelu, x)
+  timeit(mx.compile(gelu), x)
 
 On an M1 Max the times are 15.5 and 3.1 milliseconds. The compiled ``gelu`` is
 five times faster.
@@ -225,7 +225,7 @@ In some cases returning updated state can be pretty inconvenient. Hence,
   def fun(x, y):
       z = x + y
       state.append(z)
-      return mx.exp(z), state
+      return mx.exp(z)
 
   fun(mx.array(1.0), mx.array(2.0))
   # Prints [array(3, dtype=float32)]

docs/src/usage/distributed.rst

@@ -7,12 +7,13 @@ 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 three different communication backends:
 
 * `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.
+* A **ring** backend of our own that uses native TCP sockets. It should be
+  faster for thunderbolt connections, but it also works over Ethernet.
+* `nccl <https://developer.nvidia.com/nccl>`_, for use in CUDA environments.
 
 The list of all currently supported operations and their documentation can be
 seen in the :ref:`API docs<distributed>`.
@@ -84,9 +85,8 @@ 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', '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
@@ -184,7 +184,7 @@ almost identical to the example above:
 
     def step(model, x, y):
         loss, grads = loss_grad_fn(model, x, y)
-        grads = mlx.nn.average_gradients(grads) # <---- This line was added
+        grads = mx.nn.average_gradients(grads)  # <---- This line was added
         optimizer.update(model, grads)
         return loss
 
@@ -220,7 +220,7 @@ print 4 etc.
 Installing MPI
 ^^^^^^^^^^^^^^
 
-MPI can be installed with Homebrew, using the Anaconda package manager or
+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:
 
@@ -228,14 +228,16 @@ with the Anaconda package manager as follows:
 
     $ conda install conda-forge::openmpi
 
-Installing with Homebrew may require specifying the location of ``libmpi.dyld``
+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``.
+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/ python test.py
+    $ 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
 

docs/src/usage/export.rst

@@ -7,17 +7,17 @@ Exporting Functions
 
 MLX has an API to export and import functions to and from a file. This lets you
 run computations written in one MLX front-end (e.g. Python) in another MLX
-front-end (e.g. C++). 
+front-end (e.g. C++).
 
 This guide walks through the basics of the MLX export API with some examples.
 To see the full list of functions check-out the :ref:`API documentation
 <export>`.
 
-Basics of Exporting 
+Basics of Exporting
 -------------------
 
 Let's start with a simple example:
- 
+
 .. code-block:: python
 
   def fun(x, y):
@@ -67,7 +67,7 @@ specified as variable positional arguments or as a tuple of arrays:
 
   x = mx.array(1.0)
   y = mx.array(1.0)
-   
+
   # Both arguments to fun are positional
   mx.export_function("add.mlxfn", fun, x, y)
 
@@ -133,7 +133,7 @@ parameters are also saved to the ``model.mlxfn`` file.
    For enclosed arrays inside an exported function, be extra careful to ensure
    they are evaluated. The computation graph that gets exported will include
    the computation that produces enclosed inputs.
-  
+
    If the above example was missing ``mx.eval(model.parameters()``, the
    exported function would include the random initialization of the
    :obj:`mlx.nn.Module` parameters.
@@ -150,8 +150,8 @@ parameters, pass them as inputs to the ``call`` wrapper:
      # Set the model's parameters to the input parameters
      model.update(tree_unflatten(list(params.items())))
      return model(x)
- 
-   params = dict(tree_flatten(model.parameters()))
+
+   params = tree_flatten(model.parameters(), destination={})
    mx.export_function("model.mlxfn", call, (mx.zeros(4),), params)
 
 
@@ -164,13 +164,13 @@ to export a function which can be used for inputs with variable shapes:
 
 .. code-block:: python
 
-  mx.export_function("fun.mlxfn", mx.abs, mx.array(0.0), shapeless=True)
+  mx.export_function("fun.mlxfn", mx.abs, mx.array([0.0]), shapeless=True)
   imported_abs = mx.import_function("fun.mlxfn")
 
   # Ok
-  out, = imported_abs(mx.array(-1.0))
-  
-  # Also ok 
+  out, = imported_abs(mx.array([-1.0]))
+
+  # Also ok
   out, = imported_abs(mx.array([-1.0, -2.0]))
 
 With ``shapeless=False`` (which is the default), the second call to
@@ -197,7 +197,7 @@ a single file by creating an exporting context manager with :func:`exporter`:
   def fun(x, y=None):
       constant = mx.array(3.0)
       if y is not None:
-        x += y 
+        x += y
       return x + constant
 
   with mx.exporter("fun.mlxfn", fun) as exporter:
@@ -215,7 +215,7 @@ a single file by creating an exporting context manager with :func:`exporter`:
   print(out)
 
 In the above example the function constant data, (i.e. ``constant``), is only
-saved once. 
+saved once.
 
 Transformations with Imported Functions
 ---------------------------------------
@@ -238,7 +238,7 @@ on imported functions just like regular Python functions:
   # Prints: array(1, dtype=float32)
   print(dfdx(x))
 
-  # Compile the imported function 
+  # Compile the imported function
   mx.compile(imported_fun)
   # Prints: array(0, dtype=float32)
   print(compiled_fun(x)[0])
@@ -275,7 +275,7 @@ Import and run the function in C++ with only a few lines of code:
   // Prints: array(2, dtype=float32)
   std::cout << outputs[0] << std::endl;
 
-Imported functions can be transformed in C++ just like in Python. Use 
+Imported functions can be transformed in C++ just like in Python. Use
 ``std::vector<mx::array>`` for positional arguments and ``std::map<std::string,
 mx::array>`` for keyword arguments when calling imported functions in C++.
 

docs/src/usage/indexing.rst

@@ -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
@@ -107,6 +108,28 @@ same array:
   >>> a
   array([1, 2, 0], dtype=int32)
 
+Note that unlike NumPy, slicing an array creates a copy, not a view. So
+mutating it does not mutate the original array:
+
+.. code-block:: shell
+
+  >>> a = mx.array([1, 2, 3])
+  >>> b = a[:]
+  >>> b[2] = 0
+  >>> b
+  array([1, 2, 0], dtype=int32)
+  >>> a
+  array([1, 2, 3], dtype=int32)
+
+Also unlike NumPy, updates to the same location are nondeterministic:
+
+.. code-block:: shell
+
+  >>> a = mx.array([1, 2, 3])
+  >>> a[[0, 0]] = mx.array([4, 5])
+
+The first element of ``a`` could be ``4`` or ``5``.
+
 Transformations of functions which use in-place updates are allowed and work as
 expected. For example:
 
@@ -121,3 +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. No mask
+  broadcasting occurs.
+- 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.randn(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.

examples/extensions/axpby/axpby.cpp

@@ -1,5 +1,6 @@
 // Copyright © 2023-2025 Apple Inc.
 
+#include <dlfcn.h>
 #include <iostream>
 #include <sstream>
 
@@ -16,6 +17,19 @@
 
 namespace my_ext {
 
+// A helper function to find the location of the current binary on disk.
+// The Metal library ("mlx_ext.mtllib"), should be in the same directory.
+std::string current_binary_dir() {
+  static std::string binary_dir = []() {
+    Dl_info info;
+    if (!dladdr(reinterpret_cast<void*>(&current_binary_dir), &info)) {
+      throw std::runtime_error("Unable to get current binary dir.");
+    }
+    return std::filesystem::path(info.dli_fname).parent_path().string();
+  }();
+  return binary_dir;
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // Operation Implementation
 ///////////////////////////////////////////////////////////////////////////////
@@ -167,16 +181,15 @@ void Axpby::eval_gpu(
   }
 
   // Resolve name of kernel (corresponds to axpby.metal)
-  std::ostringstream kname;
-  kname << "axpby_";
-  kname << (contiguous_kernel ? "contiguous_" : "general_");
-  kname << type_to_name(out);
+  std::string kname = "axpby_";
+  kname += (contiguous_kernel ? "contiguous_" : "general_");
+  kname += type_to_name(out);
 
-  // Make sure the metal library is available
-  d.register_library("mlx_ext");
+  // Load the metal library
+  auto lib = d.get_library("mlx_ext", current_binary_dir());
 
   // Make a kernel from this metal library
-  auto kernel = d.get_kernel(kname.str(), "mlx_ext");
+  auto kernel = d.get_kernel(kname, lib);
 
   // Prepare to encode kernel
   auto& compute_encoder = d.get_command_encoder(s.index);

examples/extensions/axpby/axpby.h

@@ -74,9 +74,9 @@ class Axpby : public mx::Primitive {
       const std::vector<mx::array>& inputs,
       const std::vector<int>& axes) override;
 
-  /** Print the primitive. */
-  void print(std::ostream& os) override {
-    os << "Axpby";
+  /** The name of primitive. */
+  const char* name() const override {
+    return "Axpby";
   }
 
   /** Equivalence check **/

examples/extensions/requirements.txt

@@ -1,4 +1,4 @@
 setuptools>=42
 cmake>=3.25
 mlx>=0.21.0
-nanobind==2.2.0
+nanobind==2.4.0

examples/extensions/test.py

@@ -3,8 +3,10 @@ from mlx_sample_extensions import axpby
 
 a = mx.ones((3, 4))
 b = mx.ones((3, 4))
-c = axpby(a, b, 4.0, 2.0, stream=mx.cpu)
+c_cpu = axpby(a, b, 4.0, 2.0, stream=mx.cpu)
+c_gpu = axpby(a, b, 4.0, 2.0, stream=mx.gpu)
 
-print(f"c shape: {c.shape}")
-print(f"c dtype: {c.dtype}")
-print(f"c correct: {mx.all(c == 6.0).item()}")
+print(f"c shape: {c_cpu.shape}")
+print(f"c dtype: {c_cpu.dtype}")
+print(f"c_cpu correct: {mx.all(c_cpu == 6.0).item()}")
+print(f"c_gpu correct: {mx.all(c_gpu == 6.0).item()}")

mlx/CMakeLists.txt

@@ -55,6 +55,9 @@ endif()
 
 if(MLX_BUILD_CUDA)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backend/cuda)
+else()
+  target_sources(mlx
+                 PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/backend/cuda/no_cuda.cpp)
 endif()
 
 if(MLX_BUILD_METAL OR MLX_BUILD_CUDA)

mlx/allocator.h

@@ -14,7 +14,7 @@ class Buffer {
   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();

mlx/array.cpp

@@ -64,7 +64,7 @@ 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;
 }
 
@@ -141,7 +141,7 @@ 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 +156,7 @@ 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 +167,13 @@ 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) {
@@ -241,8 +240,8 @@ array::ArrayDesc::ArrayDesc(
     std::vector<array> inputs)
     : shape(std::move(shape)),
       dtype(dtype),
-      status(Status::unscheduled),
       primitive(std::move(primitive)),
+      status(Status::unscheduled),
       inputs(std::move(inputs)) {
   init();
 }

mlx/array.h

@@ -10,6 +10,7 @@
 #include "mlx/allocator.h"
 #include "mlx/dtype.h"
 #include "mlx/event.h"
+#include "mlx/small_vector.h"
 
 namespace mlx::core {
 
@@ -18,8 +19,8 @@ class Primitive;
 
 using Deleter = std::function<void(allocator::Buffer)>;
 using ShapeElem = int32_t;
-using Shape = std::vector<ShapeElem>;
-using Strides = std::vector<int64_t>;
+using Shape = SmallVector<ShapeElem>;
+using Strides = SmallVector<int64_t>;
 
 class array {
   /* An array is really a node in a graph. It contains a shared ArrayDesc
@@ -293,6 +294,11 @@ 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;
@@ -348,15 +354,23 @@ 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 {
@@ -425,7 +439,7 @@ 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);
 
@@ -460,8 +474,8 @@ 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;

mlx/backend/common/binary.h

@@ -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;
   }

mlx/backend/common/broadcasting.cpp

@@ -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);

mlx/backend/common/compiled.cpp

@@ -14,6 +14,8 @@ void print_constant(std::ostream& os, const array& x) {
       return print_float_constant<float16_t>(os, x);
     case bfloat16:
       return print_float_constant<bfloat16_t>(os, x);
+    case float64:
+      return print_float_constant<double>(os, x);
     case complex64:
       return print_complex_constant<complex64_t>(os, x);
     case int8:
@@ -50,6 +52,8 @@ std::string get_type_string(Dtype d) {
       return "float16_t";
     case bfloat16:
       return "bfloat16_t";
+    case float64:
+      return "double";
     case complex64:
       return "complex64_t";
     case bool_:
@@ -110,7 +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;
@@ -136,7 +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);
@@ -159,7 +165,7 @@ void compiled_allocate_outputs(
       }
     }
     for (; o < outputs.size(); ++o) {
-      outputs[o].set_data(allocator::malloc(outputs[o].nbytes()));
+      outputs[o].set_data(mallocfn(outputs[o].nbytes()));
     }
   }
 }

mlx/backend/common/compiled.h

@@ -18,8 +18,12 @@ std::string get_type_string(Dtype d);
 template <typename T>
 void print_float_constant(std::ostream& os, const array& x) {
   auto old_precision = os.precision();
-  os << std::setprecision(std::numeric_limits<float>::digits10 + 1)
-     << x.item<T>() << std::setprecision(old_precision);
+  if constexpr (std::is_same_v<T, double>) {
+    os << std::setprecision(std::numeric_limits<double>::digits10 + 1);
+  } else {
+    os << std::setprecision(std::numeric_limits<float>::digits10 + 1);
+  }
+  os << x.item<T>() << std::setprecision(old_precision);
 }
 
 template <typename T>
@@ -54,7 +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(

mlx/backend/common/copy.h

@@ -2,7 +2,7 @@
 
 #pragma once
 
-#include "mlx/array.h"
+#include "mlx/backend/common/utils.h"
 
 namespace mlx::core {
 
@@ -22,23 +22,27 @@ 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.
-    if (in.is_donatable() && in.itemsize() == out.itemsize()) {
+    if (is_donatable(in, out)) {
       out.copy_shared_buffer(in);
       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;
   }
 }

mlx/backend/common/matmul.h

@@ -0,0 +1,67 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/backend/common/utils.h"
+#include "mlx/utils.h"
+
+#include <sstream>
+
+namespace mlx::core {
+
+inline std::tuple<Shape, Strides, Strides> collapse_batches(
+    const array& a,
+    const array& b) {
+  if (a.ndim() == 2) {
+    return {Shape{1}, Strides{0}, Strides{0}};
+  }
+
+  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
+  Strides A_bstride{a.strides().begin(), a.strides().end() - 2};
+  Strides B_bstride{b.strides().begin(), b.strides().end() - 2};
+
+  auto [batch_shape, batch_strides] =
+      collapse_contiguous_dims(A_bshape, std::vector{A_bstride, B_bstride});
+
+  auto a_batch_strides = batch_strides[0];
+  auto b_batch_strides = batch_strides[1];
+
+  if (batch_shape.empty()) {
+    batch_shape.push_back(1);
+    a_batch_strides.push_back(0);
+    b_batch_strides.push_back(0);
+  }
+
+  return std::make_tuple(batch_shape, a_batch_strides, b_batch_strides);
+}
+
+inline std::tuple<Shape, Strides, Strides, Strides>
+collapse_batches(const array& a, const array& b, const array& c) {
+  if (a.ndim() == 2) {
+    return {Shape{1}, Strides{0}, Strides{0}, Strides{0}};
+  }
+
+  Shape A_bshape{a.shape().begin(), a.shape().end() - 2};
+  Strides A_bstride{a.strides().begin(), a.strides().end() - 2};
+  Strides B_bstride{b.strides().begin(), b.strides().end() - 2};
+  Strides C_bstride{c.strides().begin(), c.strides().end() - 2};
+
+  auto [batch_shape, batch_strides] = collapse_contiguous_dims(
+      A_bshape, std::vector{A_bstride, B_bstride, C_bstride});
+
+  auto A_batch_stride = batch_strides[0];
+  auto B_batch_stride = batch_strides[1];
+  auto C_batch_stride = batch_strides[2];
+
+  if (batch_shape.empty()) {
+    batch_shape.push_back(1);
+    A_batch_stride.push_back(0);
+    B_batch_stride.push_back(0);
+    C_batch_stride.push_back(0);
+  }
+
+  return std::make_tuple(
+      batch_shape, A_batch_stride, B_batch_stride, C_batch_stride);
+}
+
+} // namespace mlx::core

mlx/backend/common/reduce.cpp

@@ -5,11 +5,9 @@
 namespace mlx::core {
 
 std::pair<Shape, Strides> shapes_without_reduction_axes(
-    const array& x,
+    Shape shape,
+    Strides strides,
     const std::vector<int>& axes) {
-  auto shape = x.shape();
-  auto strides = x.strides();
-
   for (int i = axes.size() - 1; i >= 0; i--) {
     int a = axes[i];
     shape.erase(shape.begin() + a);
@@ -19,6 +17,15 @@ std::pair<Shape, Strides> shapes_without_reduction_axes(
   return std::make_pair(shape, strides);
 }
 
+std::pair<Shape, Strides> shapes_without_reduction_axes(
+    const array& x,
+    const std::vector<int>& axes) {
+  auto shape = x.shape();
+  auto strides = x.strides();
+  return shapes_without_reduction_axes(
+      std::move(shape), std::move(strides), axes);
+}
+
 ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes) {
   // The data is all there and we are reducing over everything
   if (x.size() == x.data_size() && axes.size() == x.ndim() &&

mlx/backend/common/reduce.h

@@ -51,5 +51,9 @@ ReductionPlan get_reduction_plan(const array& x, const std::vector<int>& axes);
 std::pair<Shape, Strides> shapes_without_reduction_axes(
     const array& x,
     const std::vector<int>& axes);
+std::pair<Shape, Strides> shapes_without_reduction_axes(
+    Shape shape,
+    Strides strides,
+    const std::vector<int>& axes);
 
 } // namespace mlx::core

mlx/backend/common/slicing.cpp

@@ -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);
 }
 

mlx/backend/common/ternary.h

@@ -11,6 +11,8 @@ namespace mlx::core {
 enum class TernaryOpType {
   ScalarScalarScalar,
   VectorVectorVector,
+  VectorVectorScalar,
+  VectorScalarVector,
   General,
 };
 
@@ -25,6 +27,14 @@ get_ternary_op_type(const array& a, const array& b, const array& c) {
       (a.flags().col_contiguous && b.flags().col_contiguous &&
        c.flags().col_contiguous)) {
     topt = TernaryOpType::VectorVectorVector;
+  } else if (
+      b.data_size() == 1 && a.flags().row_contiguous &&
+      c.flags().row_contiguous) {
+    topt = TernaryOpType::VectorScalarVector;
+  } else if (
+      c.data_size() == 1 && a.flags().row_contiguous &&
+      b.flags().row_contiguous) {
+    topt = TernaryOpType::VectorVectorScalar;
   } else {
     topt = TernaryOpType::General;
   }
@@ -36,7 +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);
@@ -47,24 +58,25 @@ 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());
       }
       break;
+    case TernaryOpType::VectorVectorScalar:
+    case TernaryOpType::VectorScalarVector:
     case TernaryOpType::General:
       // Try to donate an input which is row_contiguous
       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;
   }

mlx/backend/common/unary.h

@@ -0,0 +1,29 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include "mlx/allocator.h"
+#include "mlx/backend/common/utils.h"
+
+namespace mlx::core {
+
+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(
+          mallocfn(in.data_size() * out.itemsize()),
+          in.data_size(),
+          in.strides(),
+          in.flags());
+    }
+  } else {
+    out.set_data(mallocfn(out.nbytes()));
+  }
+}
+
+} // namespace mlx::core

mlx/backend/common/utils.cpp

@@ -1,14 +1,20 @@
 // Copyright © 2023-2024 Apple Inc.
 
+#include <dlfcn.h>
+
 #include "mlx/backend/common/utils.h"
-#include "mlx/primitives.h"
 
 namespace mlx::core {
 
-std::string get_primitive_string(Primitive* primitive) {
-  std::ostringstream op_t;
-  primitive->print(op_t);
-  return op_t.str();
+std::filesystem::path current_binary_dir() {
+  static std::filesystem::path binary_dir = []() {
+    Dl_info info;
+    if (!dladdr(reinterpret_cast<void*>(&current_binary_dir), &info)) {
+      throw std::runtime_error("Unable to get current binary dir.");
+    }
+    return std::filesystem::path(info.dli_fname).parent_path();
+  }();
+  return binary_dir;
 }
 
 std::tuple<Shape, std::vector<Strides>> collapse_contiguous_dims(
@@ -199,14 +205,27 @@ Dims get_2d_grid_dims_common(
       }
     }
   }
-  if (grid_y > UINT32_MAX || grid_x > UINT32_MAX || divisor > 1) {
+  if (grid_y > UINT32_MAX || grid_x > UINT32_MAX) {
     throw std::runtime_error("Unable to safely factor shape.");
   }
   if (grid_y > grid_x) {
     std::swap(grid_x, grid_y);
   }
+  if (divisor > 1) {
+    grid_x = ((grid_x + divisor - 1) / divisor) * divisor;
+  }
   return std::make_tuple(
       static_cast<uint32_t>(grid_x), static_cast<uint32_t>(grid_y), 1);
 }
 
+std::pair<Dims, Dims> get_grid_and_block_common(int dim0, int dim1, int dim2) {
+  auto [bx, by, bz] = get_block_dims_common(dim0, dim1, dim2);
+  auto gx = (dim0 + bx - 1) / bx;
+  auto gy = (dim1 + by - 1) / by;
+  auto gz = (dim2 + bz - 1) / bz;
+
+  return std::make_pair(
+      std::make_tuple(gx, gy, gz), std::make_tuple(bx, by, bz));
+}
+
 } // namespace mlx::core

mlx/backend/common/utils.h

@@ -2,6 +2,7 @@
 
 #pragma once
 
+#include <filesystem>
 #include <tuple>
 #include <vector>
 
@@ -9,7 +10,8 @@
 
 namespace mlx::core {
 
-std::string get_primitive_string(Primitive* primitive);
+// Return the directory that contains current shared library.
+std::filesystem::path current_binary_dir();
 
 inline int64_t
 elem_to_loc(int elem, const Shape& shape, const Strides& strides) {
@@ -95,6 +97,9 @@ Dims get_2d_grid_dims_common(
     const Strides& strides,
     size_t divisor);
 
+// Get both the block and a grid of blocks that covers dim0, dim1 and dim2.
+std::pair<Dims, Dims> get_grid_and_block_common(int dim0, int dim1, int dim2);
+
 struct ContiguousIterator {
   inline void step() {
     int dims = shape_.size();
@@ -192,7 +197,7 @@ void shared_buffer_reshape(
     array& out);
 
 template <typename T>
-inline std::vector<T> remove_index(std::vector<T> vec, size_t index) {
+inline SmallVector<T> remove_index(SmallVector<T> vec, size_t index) {
   vec.erase(std::next(vec.begin(), index));
   return vec;
 }

mlx/backend/cpu/binary.cpp

@@ -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

mlx/backend/cpu/binary.h

@@ -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

mlx/backend/cpu/cholesky.cpp

@@ -20,7 +20,7 @@ void cholesky_impl(const array& a, array& factor, bool upper, Stream stream) {
 
   // The decomposition is computed in place, so just copy the input to the
   // output.
-  copy(
+  copy_cpu(
       a,
       factor,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/compiled.cpp

@@ -15,6 +15,7 @@
 #include "mlx/backend/cpu/jit_compiler.h"
 #include "mlx/device.h"
 #include "mlx/graph_utils.h"
+#include "mlx/version.h"
 
 namespace mlx::core {
 
@@ -94,7 +95,11 @@ void* compile(
     kernel_file_name = kernel_name;
   }
 
-  auto output_dir = std::filesystem::temp_directory_path();
+  auto output_dir =
+      std::filesystem::temp_directory_path() / "mlx" / version() / "cpu";
+  if (!std::filesystem::exists(output_dir)) {
+    std::filesystem::create_directories(output_dir);
+  }
 
   std::string shared_lib_name = "lib" + kernel_file_name + ".so";
   auto shared_lib_path = (output_dir / shared_lib_name).string();
@@ -157,10 +162,12 @@ inline void build_kernel(
 #endif
 
   // Start the kernel
-  os << "void " << kernel_name << "(void** args) {" << std::endl;
+  os << "void " << kernel_name
+     << "(int* shape, int64_t** strides, void** args) {" << std::endl;
 
   // Add the input arguments
   int cnt = 0;
+  int strides_index = 1;
   for (size_t i = 0; i < inputs.size(); ++i) {
     // Skip constants from the input list
     if (is_constant(i)) {
@@ -175,8 +182,8 @@ inline void build_kernel(
        << "];" << std::endl;
     // Scalars and contiguous need no strides
     if (!is_scalar(x) && !contiguous) {
-      os << "  const size_t* " << xname << "_strides = (size_t*)args[" << cnt++
-         << "];" << std::endl;
+      os << "  const int64_t* " << xname << "_strides = strides["
+         << strides_index++ << "];" << std::endl;
     }
   }
 
@@ -186,10 +193,8 @@ inline void build_kernel(
     os << "  " << tstr << "* " << namer.get_name(x) << " = (" << tstr
        << "*)args[" << cnt++ << "];" << std::endl;
   }
-  // Add output strides and shape to extract the indices.
-  if (!contiguous) {
-    os << "  const int* shape = (int*)args[" << cnt++ << "];" << std::endl;
-  } else {
+  // Add output size
+  if (contiguous) {
     os << "  const size_t size = (size_t)args[" << cnt++ << "];" << std::endl;
   }
 
@@ -231,7 +236,7 @@ inline void build_kernel(
       os << "static_cast<" << get_type_string(x.dtype()) << ">(tmp_"
          << namer.get_name(x.inputs()[0]) << ");" << std::endl;
     } else {
-      x.primitive().print(os);
+      os << x.primitive().name();
       os << "()(";
       for (int i = 0; i < x.inputs().size() - 1; i++) {
         os << "tmp_" << namer.get_name(x.inputs()[i]) << ", ";
@@ -290,7 +295,6 @@ void Compiled::eval_cpu(
 
   // Collect function input arguments.
   std::vector<void*> args;
-  int strides_index = 1;
   for (size_t i = 0; i < inputs.size(); ++i) {
     if (is_constant_(i)) {
       continue;
@@ -298,9 +302,6 @@ void Compiled::eval_cpu(
     const auto& x = inputs[i];
     encoder.set_input_array(x);
     args.push_back((void*)x.data<void>());
-    if (!contiguous && !is_scalar(x)) {
-      args.push_back(strides[strides_index++].data());
-    }
   }
 
   // Get the kernel name from the lib
@@ -335,16 +336,20 @@ void Compiled::eval_cpu(
     args.push_back(x.data<void>());
     encoder.set_output_array(x);
   }
-  if (!contiguous) {
-    args.push_back((void*)shape.data());
-  } else {
+  if (contiguous) {
     args.push_back((void*)outputs[0].data_size());
   }
-  auto fun = (void (*)(void**))fn_ptr;
+  auto fun = reinterpret_cast<void (*)(int*, int64_t**, void**)>(fn_ptr);
   encoder.dispatch([fun,
                     args = std::move(args),
                     strides = std::move(strides),
-                    shape = std::move(shape)]() mutable { fun(args.data()); });
+                    shape = std::move(shape)]() mutable {
+    SmallVector<int64_t*> strides_ptrs;
+    for (auto& s : strides) {
+      strides_ptrs.push_back(s.data());
+    }
+    fun(shape.data(), strides_ptrs.data(), args.data());
+  });
 }
 
 } // namespace mlx::core

mlx/backend/cpu/conv.cpp

@@ -883,7 +883,7 @@ void explicit_gemm_conv_1D_cpu(
   // Fill with zeros
   std::vector<array> temps;
   temps.push_back(array(0, conv_dtype));
-  copy(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
 
   // Pick input slice from padded
   size_t data_offset = padding_lo[0] * in_padded.strides()[1];
@@ -895,7 +895,7 @@ void explicit_gemm_conv_1D_cpu(
       in_padded_slice.size(),
       data_offset);
   // Copy input values into the slice
-  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
   temps.push_back(in_padded_slice);
 
   // Make strided view
@@ -920,7 +920,7 @@ void explicit_gemm_conv_1D_cpu(
   // Materialize strided view
   Shape strided_reshape = {N * oH, wH * C};
   array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
-  copy(in_strided_view, in_strided, CopyType::General, stream);
+  copy_cpu(in_strided_view, in_strided, CopyType::General, stream);
   temps.push_back(in_strided);
 
   // Check wt dtype and prepare
@@ -938,13 +938,13 @@ void explicit_gemm_conv_1D_cpu(
         wt.size(),
         0);
     gemm_wt = array(wt_transpose.shape(), float32, nullptr, {});
-    copy(wt_transpose, gemm_wt, CopyType::General, stream);
+    copy_cpu(wt_transpose, gemm_wt, CopyType::General, stream);
     temps.push_back(gemm_wt);
   } else if (wt.dtype() != float32 || !wt.flags().row_contiguous) {
     auto ctype =
         wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
     gemm_wt = array(wt.shape(), float32, nullptr, {});
-    copy(wt, gemm_wt, ctype, stream);
+    copy_cpu(wt, gemm_wt, ctype, stream);
     temps.push_back(gemm_wt);
   }
 
@@ -991,132 +991,7 @@ void explicit_gemm_conv_1D_cpu(
 
   // Copy results if needed
   if (out.dtype() != float32) {
-    copy_inplace(gemm_out, out, CopyType::Vector, stream);
-  }
-  encoder.add_temporaries(std::move(temps));
-}
-
-void explicit_gemm_conv_2D_cpu(
-    const array& in,
-    const array& wt,
-    array out,
-    const std::vector<int>& padding_lo,
-    const std::vector<int>& padding_hi,
-    const std::vector<int>& wt_strides,
-    const std::vector<int>& wt_dilation,
-    Stream stream) {
-  const int N = in.shape(0); // Batch size, should be the same as out.shape(0)
-  const int iH = in.shape(1); // Input spatial dim
-  const int iW = in.shape(2); // Input spatial dim
-  const int oH = out.shape(1); // Output spatial dim
-  const int oW = out.shape(2); // Output spatial dim
-  const int O = wt.shape(0); // Out channels
-  const int C = wt.shape(3); // In channels
-  const int wH = wt.shape(1); // Weight spatial dim
-  const int wW = wt.shape(2); // Weight spatial dim
-
-  auto conv_dtype = out.dtype();
-  auto& encoder = cpu::get_command_encoder(stream);
-
-  // Pad input
-  Shape padded_shape = {
-      N,
-      iH + padding_lo[0] + padding_hi[0],
-      iW + padding_lo[1] + padding_hi[1],
-      C};
-  array in_padded(padded_shape, conv_dtype, nullptr, {});
-
-  // Fill with zeros
-  std::vector<array> temps;
-  temps.push_back(array(0, conv_dtype));
-  copy(temps.back(), in_padded, CopyType::Scalar, stream);
-
-  // Pick input slice from padded
-  size_t data_offset = padding_lo[0] * in_padded.strides()[1] +
-      padding_lo[1] * in_padded.strides()[2];
-  array in_padded_slice(in.shape(), in_padded.dtype(), nullptr, {});
-  in_padded_slice.copy_shared_buffer(
-      in_padded,
-      in_padded.strides(),
-      in_padded.flags(),
-      in_padded_slice.size(),
-      data_offset);
-  temps.push_back(in_padded_slice);
-
-  // Copy input values into the slice
-  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
-
-  // Make strided view
-  Shape strided_shape = {N, oH, oW, wH, wW, C};
-
-  Strides strided_strides = {
-      in_padded.strides()[0],
-      in_padded.strides()[1] * wt_strides[0],
-      in_padded.strides()[2] * wt_strides[1],
-      in_padded.strides()[1],
-      in_padded.strides()[2],
-      in_padded.strides()[3]};
-  auto flags = in_padded.flags();
-
-  array in_strided_view(strided_shape, in_padded.dtype(), nullptr, {});
-  in_strided_view.copy_shared_buffer(
-      in_padded, strided_strides, flags, in_strided_view.size(), 0);
-
-  // Materialize strided view
-  Shape strided_reshape = {N * oH * oW, wH * wW * C};
-  array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
-  copy(in_strided_view, in_strided, CopyType::General, stream);
-  temps.push_back(in_strided);
-
-  // Check wt dtype and prepare
-  auto gemm_wt = wt;
-  auto gemm_out = out;
-
-  if (wt.dtype() != float32 || !wt.flags().row_contiguous) {
-    auto ctype =
-        wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
-    gemm_wt = array(wt.shape(), float32, nullptr, {});
-    copy(wt, gemm_wt, ctype, stream);
-    temps.push_back(gemm_wt);
-  }
-
-  if (out.dtype() != float32) {
-    gemm_out = array(out.shape(), float32, nullptr, {});
-    gemm_out.set_data(allocator::malloc(gemm_out.nbytes()));
-    temps.push_back(gemm_out);
-  }
-
-  encoder.set_input_array(in_strided);
-  encoder.set_input_array(gemm_wt);
-  encoder.set_output_array(gemm_out);
-
-  encoder.dispatch([in_strided_ptr = in_strided.data<float>(),
-                    gemm_wt_ptr = gemm_wt.data<float>(),
-                    gemm_out_ptr = gemm_out.data<float>(),
-                    strided_reshape = std::move(strided_reshape),
-                    O]() {
-    // Perform gemm
-    cblas_sgemm(
-        CblasRowMajor,
-        CblasNoTrans, // no trans A
-        CblasTrans, // transB
-        strided_reshape[0], // M
-        O, // N
-        strided_reshape[1], // K
-        1.0f, // alpha
-        in_strided_ptr,
-        strided_reshape[1], // lda
-        gemm_wt_ptr,
-        strided_reshape[1], // ldb
-        0.0f, // beta
-        gemm_out_ptr,
-        O // ldc
-    );
-  });
-
-  // Copy results if needed
-  if (out.dtype() != float32) {
-    copy_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
   }
   encoder.add_temporaries(std::move(temps));
 }
@@ -1156,7 +1031,7 @@ void explicit_gemm_conv_ND_cpu(
 
   // Fill with zeros
   std::vector<array> temps = {array(0, conv_dtype)};
-  copy(temps.back(), in_padded, CopyType::Scalar, stream);
+  copy_cpu(temps.back(), in_padded, CopyType::Scalar, stream);
 
   // Pick input slice from padded
   size_t data_offset = 0;
@@ -1173,7 +1048,7 @@ void explicit_gemm_conv_ND_cpu(
       data_offset);
 
   // Copy input values into the slice
-  copy_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(in, in_padded_slice, CopyType::GeneralGeneral, stream);
   temps.push_back(in_padded_slice);
 
   // Make strided view
@@ -1212,7 +1087,7 @@ void explicit_gemm_conv_ND_cpu(
   }
 
   array in_strided(strided_reshape, in_strided_view.dtype(), nullptr, {});
-  copy(in_strided_view, in_strided, CopyType::General, stream);
+  copy_cpu(in_strided_view, in_strided, CopyType::General, stream);
   temps.push_back(in_strided);
 
   // Check wt dtype and prepare
@@ -1223,13 +1098,13 @@ void explicit_gemm_conv_ND_cpu(
     auto ctype =
         wt.flags().row_contiguous ? CopyType::Vector : CopyType::General;
     gemm_wt = array(wt.shape(), float32, nullptr, {});
-    copy(wt, gemm_wt, ctype, stream);
+    copy_cpu(wt, gemm_wt, ctype, stream);
     temps.push_back(gemm_wt);
   }
 
   if (flip) {
     auto gemm_wt_ = array(gemm_wt.shape(), float32, nullptr, {});
-    copy(gemm_wt, gemm_wt_, CopyType::Vector, stream);
+    copy_cpu(gemm_wt, gemm_wt_, CopyType::Vector, stream);
     temps.push_back(gemm_wt_);
 
     // Calculate the total size of the spatial dimensions
@@ -1284,7 +1159,7 @@ void explicit_gemm_conv_ND_cpu(
 
   // Copy results if needed
   if (out.dtype() != float32) {
-    copy_inplace(gemm_out, out, CopyType::Vector, stream);
+    copy_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
   }
   encoder.add_temporaries(std::move(temps));
 }

mlx/backend/cpu/copy.cpp

@@ -295,7 +295,11 @@ inline void copy_inplace_dispatch(
 
 } // namespace
 
-void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream) {
+void copy_cpu_inplace(
+    const array& src,
+    array& dst,
+    CopyType ctype,
+    Stream stream) {
   auto& encoder = cpu::get_command_encoder(stream);
   encoder.set_input_array(src);
   encoder.set_output_array(dst);
@@ -305,7 +309,7 @@ void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream) {
        ctype]() mutable { copy_inplace_dispatch(src, dst, ctype); });
 }
 
-void copy(const array& src, array& dst, CopyType ctype, Stream stream) {
+void copy_cpu(const array& src, array& dst, CopyType ctype, Stream stream) {
   bool donated = set_copy_output_data(src, dst, ctype);
   if (donated && src.dtype() == dst.dtype()) {
     // If the output has the same type as the input then there is nothing to
@@ -315,10 +319,10 @@ void copy(const array& src, array& dst, CopyType ctype, Stream stream) {
   if (ctype == CopyType::GeneralGeneral) {
     ctype = CopyType::General;
   }
-  copy_inplace(src, dst, ctype, stream);
+  copy_cpu_inplace(src, dst, ctype, stream);
 }
 
-void copy_inplace(
+void copy_cpu_inplace(
     const array& src,
     array& dst,
     const Shape& data_shape,
@@ -373,4 +377,10 @@ void copy_inplace(
       });
 }
 
+array contiguous_copy_cpu(const array& arr, Stream stream) {
+  array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
+  copy_cpu(arr, arr_copy, CopyType::General, stream);
+  return arr_copy;
+}
+
 } // namespace mlx::core

mlx/backend/cpu/copy.h

@@ -10,10 +10,14 @@
 
 namespace mlx::core {
 
-void copy(const array& src, array& dst, CopyType ctype, Stream stream);
-void copy_inplace(const array& src, array& dst, CopyType ctype, Stream stream);
+void copy_cpu(const array& src, array& dst, CopyType ctype, Stream stream);
+void copy_cpu_inplace(
+    const array& src,
+    array& dst,
+    CopyType ctype,
+    Stream stream);
 
-void copy_inplace(
+void copy_cpu_inplace(
     const array& src,
     array& dst,
     const Shape& data_shape,
@@ -26,4 +30,7 @@ void copy_inplace(
     const std::optional<array>& dynamic_i_offset = std::nullopt,
     const std::optional<array>& dynamic_o_offset = std::nullopt);
 
+// Return a contiguous array with same shape that copies the data of |arr|.
+array contiguous_copy_cpu(const array& arr, Stream stream);
+
 } // namespace mlx::core

mlx/backend/cpu/distributed.cpp

@@ -13,9 +13,7 @@ std::pair<array, bool> ensure_row_contiguous(const array& arr, Stream stream) {
   if (arr.flags().row_contiguous) {
     return {arr, false};
   } else {
-    array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-    copy(arr, arr_copy, CopyType::General, stream);
-    return {arr_copy, true};
+    return {contiguous_copy_cpu(arr, stream), true};
   }
 };
 
@@ -34,8 +32,7 @@ void AllReduce::eval_cpu(
       }
       return in;
     } else {
-      array arr_copy(in.shape(), in.dtype(), nullptr, {});
-      copy(in, arr_copy, CopyType::General, s);
+      array arr_copy = contiguous_copy_cpu(in, s);
       out.copy_shared_buffer(arr_copy);
       return arr_copy;
     }
@@ -98,4 +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

mlx/backend/cpu/eig.cpp

@@ -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,
@@ -19,102 +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;
-      int iwork;
-      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());
       }
     }
@@ -135,7 +233,7 @@ void Eig::eval_cpu(
       : array(a.shape(), complex64, nullptr, {});
 
   auto a_copy = array(a.shape(), a.dtype(), nullptr, {});
-  copy(
+  copy_cpu(
       a,
       a_copy,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,
@@ -166,8 +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.");
   }
 }
 

mlx/backend/cpu/eigh.cpp

@@ -196,7 +196,7 @@ void Eigh::eval_cpu(
 
   values.set_data(allocator::malloc(values.nbytes()));
 
-  copy(
+  copy_cpu(
       a,
       vectors,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/gemms/bnns.cpp

@@ -1,5 +1,4 @@
 // Copyright © 2023-2024 Apple Inc.
-
 #include <Accelerate/Accelerate.h>
 
 #include "mlx/array.h"
@@ -49,9 +48,15 @@ void matmul_bnns(
   size_t K = a_shape[ndim - 1];
 
   BNNSDataType bnns_dtype = to_bnns_dtype<T>();
-
 #pragma GCC diagnostic push
 #pragma GCC diagnostic ignored "-Wdeprecated-declarations"
+  if (beta != 1.0 && beta != 0.0) {
+    // scale the output
+    for (auto i = 0; i < batch_size * M * N; ++i) {
+      out[i] *= beta;
+    }
+    beta = 1.0;
+  }
   const BNNSLayerParametersBroadcastMatMul gemm_params{
       /* float alpha = */ alpha,
       /* float beta = */ beta,

mlx/backend/cpu/gemms/cblas.cpp

@@ -88,4 +88,47 @@ void matmul<double>(
   }
 }
 
+template <>
+void matmul<complex64_t>(
+    const complex64_t* a,
+    const complex64_t* b,
+    complex64_t* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    size_t ldc,
+    float alpha,
+    float beta,
+    size_t batch_size,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides) {
+  auto ndim = a_shape.size();
+  size_t M = a_shape[ndim - 2];
+  size_t N = b_shape[ndim - 1];
+  size_t K = a_shape[ndim - 1];
+  auto calpha = static_cast<complex64_t>(alpha);
+  auto cbeta = static_cast<complex64_t>(beta);
+
+  for (int i = 0; i < batch_size; ++i) {
+    cblas_cgemm(
+        CblasRowMajor,
+        a_transposed ? CblasTrans : CblasNoTrans, // transA
+        b_transposed ? CblasTrans : CblasNoTrans, // transB
+        M,
+        N,
+        K,
+        &calpha,
+        a + elem_to_loc(M * K * i, a_shape, a_strides),
+        lda,
+        b + elem_to_loc(K * N * i, b_shape, b_strides),
+        ldb,
+        &cbeta,
+        out + M * N * i,
+        ldc);
+  }
+}
+
 } // namespace mlx::core

mlx/backend/cpu/hadamard.cpp

@@ -96,7 +96,7 @@ void Hadamard::eval_cpu(const std::vector<array>& inputs, array& out) {
   if (in.flags().row_contiguous && in.is_donatable()) {
     out.copy_shared_buffer(in);
   } else {
-    copy(
+    copy_cpu(
         in,
         out,
         in.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/indexing.cpp

@@ -517,7 +517,7 @@ void Scatter::eval_cpu(const std::vector<array>& inputs, array& out) {
   // Copy src into out (copy allocates memory for out)
   auto ctype =
       src.flags().row_contiguous ? CopyType::Vector : CopyType::General;
-  copy(src, out, ctype, stream());
+  copy_cpu(src, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   std::vector<array> inds;
@@ -686,7 +686,7 @@ void ScatterAxis::eval_cpu(const std::vector<array>& inputs, array& out) {
   // Copy src into out (copy allocates memory for out)
   auto ctype =
       src.flags().row_contiguous ? CopyType::Vector : CopyType::General;
-  copy(src, out, ctype, stream());
+  copy_cpu(src, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_input_array(idx);
@@ -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 (uint 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

mlx/backend/cpu/inverse.cpp

@@ -115,7 +115,7 @@ void inverse_impl(
   //   (A⁻¹)ᵀ = (Aᵀ)⁻¹
 
   // The inverse is computed in place, so just copy the input to the output.
-  copy(
+  copy_cpu(
       a,
       inv,
       a.flags().row_contiguous ? CopyType::Vector : CopyType::General,

mlx/backend/cpu/jit_compiler.cpp

@@ -2,6 +2,7 @@
 
 #include "mlx/backend/cpu/jit_compiler.h"
 
+#include <algorithm>
 #include <sstream>
 #include <vector>
 

mlx/backend/cpu/lapack.h

@@ -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(gesvdx)
 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)

mlx/backend/cpu/logsumexp.cpp

@@ -87,8 +87,7 @@ void LogSumExp::eval_cpu(const std::vector<array>& inputs, array& out) {
     if (x.flags().contiguous && x.strides()[x.ndim() - 1] == 1) {
       return x;
     } else {
-      auto x_copy = array(x.shape(), x.dtype(), nullptr, {});
-      copy(x, x_copy, CopyType::General, s);
+      array x_copy = contiguous_copy_cpu(x, s);
       encoder.add_temporary(x_copy);
       return x_copy;
     }

mlx/backend/cpu/luf.cpp

@@ -31,7 +31,7 @@ void luf_impl(
   strides[ndim - 1] = M;
   strides[ndim - 2] = 1;
   lu.set_data(allocator::malloc(lu.nbytes()), lu.nbytes(), strides, flags);
-  copy_inplace(
+  copy_cpu_inplace(
       a,
       lu,
       a.shape(),

mlx/backend/cpu/masked_mm.cpp

@@ -6,6 +6,7 @@
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
+#include "mlx/backend/cpu/gemm.h"
 #include "mlx/backend/cpu/lapack.h"
 #include "mlx/primitives.h"
 
@@ -52,6 +53,58 @@ inline void mask_matrix(
   }
 }
 
+template <typename T>
+inline void segmented_mm(
+    const T* a,
+    const T* b,
+    const uint32_t* segments,
+    T* out,
+    bool a_transposed,
+    bool b_transposed,
+    size_t lda,
+    size_t ldb,
+    const Shape& a_shape,
+    const Strides& a_strides,
+    const Shape& b_shape,
+    const Strides& b_strides,
+    size_t num_segments,
+    const Shape& segments_shape,
+    const Strides& segments_strides) {
+  int ndim = a_shape.size();
+  Shape a_copy = a_shape;
+  Shape b_copy = b_shape;
+  int32_t M = a_copy[ndim - 2];
+  int32_t N = b_copy[ndim - 1];
+  for (int i = 0; i < num_segments; i++) {
+    uint32_t k_start =
+        segments[elem_to_loc(2 * i, segments_shape, segments_strides)];
+    uint32_t k_end =
+        segments[elem_to_loc(2 * i + 1, segments_shape, segments_strides)];
+    if (k_end <= k_start) {
+      std::fill_n(out + i * M * N, M * N, T(0));
+      continue;
+    }
+    a_copy[ndim - 1] = k_end - k_start;
+    b_copy[ndim - 2] = k_end - k_start;
+    matmul<T>(
+        a + k_start * a_strides[ndim - 1],
+        b + k_start * b_strides[ndim - 2],
+        out + i * M * N,
+        a_transposed,
+        b_transposed,
+        lda,
+        ldb,
+        N,
+        1.0,
+        0.0,
+        1,
+        a_copy,
+        a_strides,
+        b_copy,
+        b_strides);
+  }
+}
+
 } // namespace
 
 void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -71,21 +124,20 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
         if (!expand_all && stx == arr.shape(-1) && sty == 1) {
           if (do_copy) {
             array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-            copy(arr, arr_copy, CopyType::Vector, s);
+            copy_cpu(arr, arr_copy, CopyType::Vector, s);
             return std::make_tuple(false, stx, arr_copy, true);
           }
           return std::make_tuple(false, stx, arr, false);
         } else if (!expand_all && stx == 1 && sty == arr.shape(-2)) {
           if (do_copy) {
             array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-            copy(arr, arr_copy, CopyType::Vector, s);
+            copy_cpu(arr, arr_copy, CopyType::Vector, s);
             return std::make_tuple(true, sty, arr_copy, true);
           }
           return std::make_tuple(true, sty, arr, false);
         } else {
-          array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-          copy(arr, arr_copy, CopyType::General, s);
           int64_t stx = arr.shape(-1);
+          array arr_copy = contiguous_copy_cpu(arr, s);
           return std::make_tuple(false, stx, arr_copy, true);
         }
       };
@@ -163,18 +215,18 @@ void BlockMaskedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
 
   encoder.set_input_array(a);
   encoder.set_input_array(b);
-  const void* a_mask_ptr;
-  const void* b_mask_ptr;
-  const void* out_mask_ptr;
+  const void* a_mask_ptr = nullptr;
+  const void* b_mask_ptr = nullptr;
+  const void* out_mask_ptr = nullptr;
   Shape a_mask_shape;
   Shape b_mask_shape;
   Shape out_mask_shape;
   Strides a_mask_strides;
   Strides b_mask_strides;
   Strides out_mask_strides;
-  bool a_mask_bool;
-  bool b_mask_bool;
-  bool out_mask_bool;
+  bool a_mask_bool = false;
+  bool b_mask_bool = false;
+  bool out_mask_bool = false;
   if (has_op_mask) {
     auto& a_mask = inputs[inputs.size() - 2];
     auto& b_mask = inputs[inputs.size() - 1];
@@ -333,7 +385,7 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
       return std::make_tuple(true, sty, arr);
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
+      copy_cpu(arr, temps.back(), CopyType::General, s);
       int64_t stx = arr.shape(-1);
       return std::make_tuple(false, stx, temps.back());
     }
@@ -371,7 +423,6 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& rhs_indices = inputs[3];
 
   auto batch_shape = get_batch_dims(out.shape());
-  int batch_ndim = batch_shape.size();
 
   auto batch_shape_A = get_batch_dims(a.shape());
   auto batch_strides_A = get_batch_dims(a.strides());
@@ -437,4 +488,121 @@ void GatherMM::eval_cpu(const std::vector<array>& inputs, array& out) {
   encoder.add_temporaries(std::move(temps));
 }
 
+void SegmentedMM::eval_cpu(const std::vector<array>& inputs, array& out) {
+  out.set_data(allocator::malloc(out.nbytes()));
+
+  auto& s = stream();
+  auto& encoder = cpu::get_command_encoder(stream());
+  auto check_transpose = [&s, &encoder](const array& x) {
+    auto stx = x.strides()[x.ndim() - 2];
+    auto sty = x.strides()[x.ndim() - 1];
+    if (stx == x.shape(-1) && sty == 1) {
+      return std::make_tuple(false, stx, x);
+    } else if (stx == 1 && sty == x.shape(-2)) {
+      return std::make_tuple(true, sty, x);
+    } else {
+      array xc(x.shape(), x.dtype(), nullptr, {});
+      copy_cpu(x, xc, CopyType::General, s);
+      encoder.add_temporary(xc);
+      int64_t stx = x.shape(-1);
+      return std::make_tuple(false, stx, xc);
+    }
+  };
+
+  auto [a_transposed, lda, a] = check_transpose(inputs[0]);
+  auto [b_transposed, ldb, b] = check_transpose(inputs[1]);
+  auto& segments = inputs[2];
+
+  encoder.set_input_array(a);
+  encoder.set_input_array(b);
+  encoder.set_input_array(segments);
+  encoder.set_output_array(out);
+  encoder.dispatch([a = array::unsafe_weak_copy(a),
+                    b = array::unsafe_weak_copy(b),
+                    segments = array::unsafe_weak_copy(segments),
+                    out_ptr = out.data<void>(),
+                    a_transposed = a_transposed,
+                    b_transposed = b_transposed,
+                    lda = lda,
+                    ldb = ldb]() {
+    switch (a.dtype()) {
+      case float64:
+        segmented_mm<double>(
+            a.data<double>(),
+            b.data<double>(),
+            segments.data<uint32_t>(),
+            static_cast<double*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case float32:
+        segmented_mm<float>(
+            a.data<float>(),
+            b.data<float>(),
+            segments.data<uint32_t>(),
+            static_cast<float*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case float16:
+        segmented_mm<float16_t>(
+            a.data<float16_t>(),
+            b.data<float16_t>(),
+            segments.data<uint32_t>(),
+            static_cast<float16_t*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      case bfloat16:
+        segmented_mm<bfloat16_t>(
+            a.data<bfloat16_t>(),
+            b.data<bfloat16_t>(),
+            segments.data<uint32_t>(),
+            static_cast<bfloat16_t*>(out_ptr),
+            a_transposed,
+            b_transposed,
+            lda,
+            ldb,
+            a.shape(),
+            a.strides(),
+            b.shape(),
+            b.strides(),
+            segments.size() / 2,
+            segments.shape(),
+            segments.strides());
+        break;
+      default:
+        throw std::invalid_argument(
+            "Segmented mm supports only real float types.");
+    }
+  });
+}
+
 } // namespace mlx::core

mlx/backend/cpu/matmul.cpp

@@ -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"
@@ -81,7 +83,7 @@ void matmul_general(
       return std::make_tuple(true, sty, arr);
     } else {
       temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, stream);
+      copy_cpu(arr, temps.back(), CopyType::General, stream);
       stx = arr.shape(-1);
       return std::make_tuple(false, stx, temps.back());
     }
@@ -91,7 +93,6 @@ void matmul_general(
   auto [b_transposed, ldb, b] = check_transpose(b_pre);
   size_t M = a.shape(-2);
   size_t N = b.shape(-1);
-  size_t K = a.shape(-1);
   if (M == 0 || N == 0) {
     return;
   }
@@ -108,6 +109,9 @@ void matmul_general(
   } else if (out.dtype() == float64) {
     matmul_dispatch<double>(
         a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta, stream);
+  } else if (out.dtype() == complex64) {
+    matmul_dispatch<complex64_t>(
+        a, b, out, a_transposed, b_transposed, lda, ldb, alpha, beta, stream);
   } else {
     throw std::runtime_error("[Matmul::eval_cpu] Invalid type.");
   }
@@ -128,24 +132,34 @@ void Matmul::eval_cpu(const std::vector<array>& inputs, array& out) {
 }
 
 void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
-  if (out.dtype() != float32) {
-    throw std::runtime_error(
-        "[AddMM::eval_cpu] Currently only supports float32.");
-  }
   if (out.size() == 0) {
     out.set_data(allocator::malloc(out.nbytes()));
     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(c, out, ctype, stream());
-  if (inputs[0].shape(-1) == 0) {
-    return;
-  }
+  copy_cpu(c, out, ctype, stream());
   matmul_general(inputs[0], inputs[1], out, stream(), alpha_, beta_);
 }
 

mlx/backend/cpu/primitives.cpp

@@ -22,7 +22,7 @@ void reshape(const array& in, array& out) {
   auto [copy_necessary, out_strides] = prepare_reshape(in, out);
   if (copy_necessary) {
     out.set_data(allocator::malloc(out.nbytes()));
-    copy_inplace(in, out, CopyType::General, out.primitive().stream());
+    copy_cpu_inplace(in, out, CopyType::General, out.primitive().stream());
   } else {
     shared_buffer_reshape(in, out_strides, out);
   }
@@ -175,7 +175,7 @@ void AsType::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
   CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype, stream());
+  copy_cpu(in, out, ctype, stream());
 }
 
 void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -198,7 +198,7 @@ void Concatenate::eval_cpu(const std::vector<array>& inputs, array& out) {
     size_t data_offset = strides[axis_] * sizes[i];
     out_slice.copy_shared_buffer(
         out, strides, flags, out_slice.size(), data_offset);
-    copy_inplace(inputs[i], out_slice, CopyType::GeneralGeneral, stream());
+    copy_cpu_inplace(inputs[i], out_slice, CopyType::GeneralGeneral, stream());
   }
 }
 
@@ -211,7 +211,7 @@ void Contiguous::eval_cpu(const std::vector<array>& inputs, array& out) {
        (allow_col_major_ && in.flags().col_contiguous))) {
     out.copy_shared_buffer(in);
   } else {
-    copy(in, out, CopyType::General, stream());
+    copy_cpu(in, out, CopyType::General, stream());
   }
 }
 
@@ -235,7 +235,7 @@ void Full::eval_cpu(const std::vector<array>& inputs, array& out) {
   } else {
     ctype = CopyType::General;
   }
-  copy(in, out, ctype, stream());
+  copy_cpu(in, out, ctype, stream());
 }
 
 void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -251,7 +251,7 @@ void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(val.dtype() == in.dtype() && in.dtype() == out.dtype());
 
   // Fill output with val
-  copy(val, out, CopyType::Scalar, stream());
+  copy_cpu(val, out, CopyType::Scalar, stream());
 
   // Find offset for start of input values
   size_t data_offset = 0;
@@ -266,7 +266,7 @@ void Pad::eval_cpu(const std::vector<array>& inputs, array& out) {
       out, out.strides(), out.flags(), out_slice.size(), data_offset);
 
   // Copy input values into the slice
-  copy_inplace(in, out_slice, CopyType::GeneralGeneral, stream());
+  copy_cpu_inplace(in, out_slice, CopyType::GeneralGeneral, stream());
 }
 
 void RandomBits::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -333,14 +333,14 @@ 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];
   out.set_data(allocator::malloc(out.nbytes()));
   auto [in_offset, donated] =
       compute_dynamic_offset(inputs[1], in.strides(), axes_, stream());
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ in,
       /* array& dst = */ out,
       /* const Shape& data_shape = */ out.shape(),
@@ -361,7 +361,7 @@ 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;
   }
 
@@ -372,11 +372,11 @@ void DynamicSliceUpdate::eval_cpu(
   auto ctype = in.flags().contiguous && in.size() == in.data_size()
       ? CopyType::Vector
       : CopyType::General;
-  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
+  copy_cpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
 
   auto [out_offset, donated] =
       compute_dynamic_offset(inputs[2], out.strides(), axes_, stream());
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ upd,
       /* array& dst = */ out,
       /* const std::vector<int>& data_shape = */ upd.shape(),
@@ -396,7 +396,7 @@ 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;
   }
 
@@ -412,14 +412,14 @@ void SliceUpdate::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto ctype = in.flags().contiguous && in.size() == in.data_size()
       ? CopyType::Vector
       : CopyType::General;
-  copy(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
+  copy_cpu(in, out, in.data_size() == 1 ? CopyType::Scalar : ctype, stream());
 
   // Calculate out strides, initial offset and if copy needs to be made
   auto [data_offset, out_strides] =
       prepare_slice(out, start_indices_, strides_);
 
   // Do copy
-  copy_inplace(
+  copy_cpu_inplace(
       /* const array& src = */ upd,
       /* array& dst = */ out,
       /* const std::vector<int>& data_shape = */ upd.shape(),
@@ -456,9 +456,9 @@ void View::eval_cpu(const std::vector<array>& inputs, array& out) {
     if (in.dtype() == bool_) {
       auto in_tmp = array(in.shape(), uint8, nullptr, {});
       in_tmp.copy_shared_buffer(in);
-      copy_inplace(in_tmp, tmp, CopyType::General, stream());
+      copy_cpu_inplace(in_tmp, tmp, CopyType::General, stream());
     } else {
-      copy_inplace(in, tmp, CopyType::General, stream());
+      copy_cpu_inplace(in, tmp, CopyType::General, stream());
     }
 
     auto flags = out.flags();

mlx/backend/cpu/qrf.cpp

@@ -26,7 +26,7 @@ void qrf_impl(const array& a, array& q, array& r, Stream stream) {
   strides[in.ndim() - 2] = 1;
   strides[in.ndim() - 1] = M;
   in.set_data(allocator::malloc(in.nbytes()), in.nbytes(), strides, flags);
-  copy_inplace(a, in, CopyType::GeneralGeneral, stream);
+  copy_cpu_inplace(a, in, CopyType::GeneralGeneral, stream);
   auto& encoder = cpu::get_command_encoder(stream);
   q.set_data(allocator::malloc(q.nbytes()));
   r.set_data(allocator::malloc(r.nbytes()));

mlx/backend/cpu/quantized.cpp

@@ -1,10 +1,11 @@
 // Copyright © 2023 Apple Inc.
 
-#include <cassert>
-
+#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"
@@ -13,6 +14,35 @@ namespace mlx::core {
 
 namespace {
 
+const static float MXFP4_LUT[16] = {
+    +0.0f,
+    +0.5f,
+    +1.0f,
+    +1.5f,
+    +2.0f,
+    +3.0f,
+    +4.0f,
+    +6.0f,
+    -0.0f,
+    -0.5f,
+    -1.0f,
+    -1.5f,
+    -2.0f,
+    -3.0f,
+    -4.0f,
+    -6.0f};
+
+template <typename T>
+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);
+}
+
 inline constexpr short get_pack_factor(int bits, int wsize = 8) {
   return (bits == 3 || bits == 5) ? 8 : (bits == 6 ? 4 : wsize / bits);
 }
@@ -407,6 +437,229 @@ void _qmm_dispatch(
   }
 }
 
+template <typename T>
+void mxfp4_qmm(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K) {
+  constexpr int group_size = 32;
+  constexpr int pack_factor = get_pack_factor(4, 8);
+  constexpr int packs_in_group = group_size / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint8_t* w_local = (const uint8_t*)w;
+    const uint8_t* scales_local = scales;
+
+    std::fill(result, result + N, 0);
+
+    for (int k = 0; k < K; k++) {
+      T* result_local = result;
+      T xi = *x++;
+
+      for (int n = 0; n < N; n += group_size) {
+        T scale = dequantize_scale<T>(*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++) {
+            (*result_local++) +=
+                xi * scale * static_cast<T>(MXFP4_LUT[wi & 0xf]);
+            wi >>= 4;
+          }
+        }
+      }
+    }
+
+    result += N;
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_t(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K) {
+  constexpr int group_size = 32;
+  constexpr int pack_factor = get_pack_factor(4, 8);
+  constexpr int packs_in_group = group_size / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint8_t* w_local = (const uint8_t*)w;
+    const uint8_t* scales_local = scales;
+
+    for (int n = 0; n < N; n++) {
+      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 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;
+          }
+        }
+        sum += scale * gsum;
+      }
+      *result = sum;
+      result++;
+    }
+
+    x += K;
+  }
+}
+
+template <int S>
+simd::Simd<float, S> mxfp4_extract_bits_simd(const uint32_t* w) {
+  if constexpr (S == 8) {
+    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);
+    wi = wi >> shifts;
+    wi = wi & 0xf;
+    simd::Simd<float, S> w_out;
+    for (int i = 0; i < S; ++i) {
+      w_out[i] = MXFP4_LUT[wi[i]];
+    }
+    return 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(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K) {
+  constexpr int group_size = 32;
+  constexpr int pack_factor = 32 / 4;
+  constexpr int packs_in_group = group_size / pack_factor;
+  constexpr int S = simd::max_size<T>;
+  static_assert(
+      S % pack_factor == 0, "SIMD size must be divisible by pack factor");
+  constexpr int packs_per_simd = S / pack_factor;
+
+  for (int m = 0; m < M; m++) {
+    const uint32_t* w_local = w;
+    const uint8_t* scales_local = scales;
+
+    for (int n = 0; n < N; n++) {
+      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++);
+
+        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);
+          w_local += packs_per_simd;
+          simd::Simd<float, S> x_simd = simd::load<T, S>(x_local);
+          g_acc = g_acc + x_simd * wf;
+          x_local += S;
+        }
+        acc = acc + scale * g_acc;
+      }
+
+      *result = T(simd::sum(acc));
+      result++;
+    }
+    x += K;
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_dispatch_transpose(
+    T* result,
+    const T* x,
+    const uint32_t* w,
+    const uint8_t* scales,
+    int M,
+    int N,
+    int K,
+    bool transposed_w) {
+  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);
+    } else {
+      mxfp4_qmm_t<T>(result, x, w, scales, M, N, K);
+    }
+  } else {
+    mxfp4_qmm<T>(result, x, w, scales, M, N, K);
+  }
+}
+
+template <typename T>
+void mxfp4_qmm_dispatch_typed(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    bool transposed_w) {
+  int K = x.shape(-1);
+  int M = x.ndim() > 1 ? x.shape(-2) : 1;
+  int N = out.shape(-1);
+  int w_els = w.ndim() > 2 ? w.shape(-1) * w.shape(-2) : 0;
+  int g_els = w.ndim() > 2 ? scales.shape(-1) * scales.shape(-2) : 0;
+  int batch_size = x.size() / (K * M);
+
+  auto out_ptr = out.data<T>();
+  auto x_ptr = x.data<T>();
+  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>(
+        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()),
+        scales_ptr + elem_to_loc(i * g_els, scales.shape(), scales.strides()),
+        M,
+        N,
+        K,
+        transposed_w);
+  }
+}
+
+void mxfp4_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    bool transposed_w) {
+  switch (x.dtype()) {
+    case bfloat16:
+      mxfp4_qmm_dispatch_typed<bfloat16_t>(out, x, w, scales, transposed_w);
+      break;
+    case float16:
+      mxfp4_qmm_dispatch_typed<float16_t>(out, x, w, scales, transposed_w);
+      break;
+    case float32:
+      mxfp4_qmm_dispatch_typed<float>(out, x, w, scales, transposed_w);
+      break;
+    default:
+      throw std::invalid_argument(
+          "[quantized_matmul] only floating types are supported");
+  }
+}
+
 template <typename T>
 void _bs_qmm_dispatch_typed(
     array& out,
@@ -513,115 +766,198 @@ void _bs_qmm_dispatch(
   }
 }
 
+template <typename T>
+void mxfp4_bs_qmm_dispatch_typed(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    bool transposed_w) {
+  int K = x.shape(-1);
+  int M = x.shape(-2);
+  int N = out.shape(-1);
+
+  int w_els = w.shape(-1) * w.shape(-2);
+  int g_els = scales.shape(-1) * scales.shape(-2);
+
+  auto out_ptr = out.data<T>();
+  auto x_ptr = x.data<T>();
+  auto w_ptr = w.data<uint32_t>();
+  auto scales_ptr = scales.data<uint8_t>();
+  auto lhs_indices_ptr = lhs_indices.data<uint32_t>();
+  auto rhs_indices_ptr = rhs_indices.data<uint32_t>();
+
+  for (int i = 0; i < lhs_indices.size(); i++) {
+    int x_idx = lhs_indices_ptr[elem_to_loc(
+        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>(
+        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()),
+        scales_ptr +
+            elem_to_loc(w_idx * g_els, scales.shape(), scales.strides()),
+        M,
+        N,
+        K,
+        transposed_w);
+  }
+}
+
+void mxfp4_bs_qmm_dispatch(
+    array& out,
+    const array& x,
+    const array& w,
+    const array& scales,
+    const array& lhs_indices,
+    const array& rhs_indices,
+    bool transposed_w) {
+  switch (x.dtype()) {
+    case float32:
+      mxfp4_bs_qmm_dispatch_typed<float>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    case float16:
+      mxfp4_bs_qmm_dispatch_typed<float16_t>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    case bfloat16:
+      mxfp4_bs_qmm_dispatch_typed<bfloat16_t>(
+          out, x, w, scales, lhs_indices, rhs_indices, transposed_w);
+      break;
+    default:
+      throw std::invalid_argument(
+          "[quantized_matmul] only floating types are supported");
+  }
+}
+
 } // namespace
 
 void QuantizedMatmul::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 4);
-
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
-  auto& biases_pre = inputs[3];
 
-  std::vector<array> temps;
-  auto ensure_row_contiguous = [s = stream(), &temps](const array& arr) {
+  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 {
-      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
-      return temps.back();
+      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 biases = ensure_row_contiguous(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
-  auto& encoder = cpu::get_command_encoder(stream());
-  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
-  encoder.set_input_array(biases);
   encoder.set_output_array(out);
-  encoder.dispatch([out = array::unsafe_weak_copy(out),
-                    x = array::unsafe_weak_copy(x),
-                    w = array::unsafe_weak_copy(w),
-                    scales = array::unsafe_weak_copy(scales),
-                    biases = array::unsafe_weak_copy(biases),
-                    group_size_ = group_size_,
-                    bits_ = bits_,
-                    transpose_ = transpose_]() mutable {
-    _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
-  });
+  if (mode_ == QuantizationMode::Affine) {
+    auto biases = ensure_row_contiguous(inputs[3]);
+    encoder.set_input_array(biases);
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      biases = array::unsafe_weak_copy(biases),
+                      group_size_ = group_size_,
+                      bits_ = bits_,
+                      transpose_ = transpose_]() mutable {
+      _qmm_dispatch(out, x, w, scales, biases, group_size_, bits_, transpose_);
+    });
+  } else {
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      transpose_ = transpose_]() mutable {
+      mxfp4_qmm_dispatch(out, x, w, scales, transpose_);
+    });
+  }
 }
 
 void GatherQMM::eval_cpu(const std::vector<array>& inputs, array& out) {
-  assert(inputs.size() == 6);
-
   auto& x_pre = inputs[0];
   auto& w_pre = inputs[1];
   auto& scales_pre = inputs[2];
-  auto& biases_pre = inputs[3];
-  auto& lhs_indices = inputs[4];
-  auto& rhs_indices = inputs[5];
+  auto& lhs_indices = inputs[inputs.size() - 2];
+  auto& rhs_indices = inputs[inputs.size() - 1];
 
-  std::vector<array> temps;
+  auto& encoder = cpu::get_command_encoder(stream());
   auto ensure_row_contiguous_last_dims = [s = stream(),
-                                          &temps](const array& arr) {
+                                          &encoder](const array& arr) {
     auto stride_0 = arr.strides()[arr.ndim() - 2];
     auto stride_1 = arr.strides()[arr.ndim() - 1];
     if (stride_0 == arr.shape(-1) && stride_1 == 1) {
       return arr;
     } else {
-      temps.push_back(array(arr.shape(), arr.dtype(), nullptr, {}));
-      copy(arr, temps.back(), CopyType::General, s);
-      return temps.back();
+      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_last_dims(x_pre);
   auto w = ensure_row_contiguous_last_dims(w_pre);
   auto scales = ensure_row_contiguous_last_dims(scales_pre);
-  auto biases = ensure_row_contiguous_last_dims(biases_pre);
 
   out.set_data(allocator::malloc(out.nbytes()));
 
-  auto& encoder = cpu::get_command_encoder(stream());
-  encoder.add_temporaries(std::move(temps));
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(scales);
-  encoder.set_input_array(biases);
   encoder.set_input_array(lhs_indices);
   encoder.set_input_array(rhs_indices);
   encoder.set_output_array(out);
-  encoder.dispatch([out = array::unsafe_weak_copy(out),
-                    x = array::unsafe_weak_copy(x),
-                    w = array::unsafe_weak_copy(w),
-                    scales = array::unsafe_weak_copy(scales),
-                    biases = array::unsafe_weak_copy(biases),
-                    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 {
-    _bs_qmm_dispatch(
-        out,
-        x,
-        w,
-        scales,
-        biases,
-        lhs_indices,
-        rhs_indices,
-        group_size_,
-        bits_,
-        transpose_);
-  });
+  if (mode_ == QuantizationMode::Affine) {
+    auto biases = ensure_row_contiguous_last_dims(inputs[3]);
+    encoder.set_input_array(biases);
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      biases = array::unsafe_weak_copy(biases),
+                      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 {
+      _bs_qmm_dispatch(
+          out,
+          x,
+          w,
+          scales,
+          biases,
+          lhs_indices,
+          rhs_indices,
+          group_size_,
+          bits_,
+          transpose_);
+    });
+  } else {
+    encoder.dispatch([out = array::unsafe_weak_copy(out),
+                      x = array::unsafe_weak_copy(x),
+                      w = array::unsafe_weak_copy(w),
+                      scales = array::unsafe_weak_copy(scales),
+                      lhs_indices = array::unsafe_weak_copy(lhs_indices),
+                      rhs_indices = array::unsafe_weak_copy(rhs_indices),
+                      transpose_ = transpose_]() mutable {
+      mxfp4_bs_qmm_dispatch(
+          out, x, w, scales, lhs_indices, rhs_indices, transpose_);
+    });
+  }
 }
 
 template <typename T, typename U>
@@ -705,16 +1041,14 @@ void dispatch_quantize(
       w_ptr, out_ptr, scales_ptr, biases_ptr, bits, group_size, w.size());
 }
 
-void fast::AffineQuantize::eval_cpu(
+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 {
-      array arr_copy(arr.shape(), arr.dtype(), nullptr, {});
-      copy(arr, arr_copy, CopyType::General, s);
-      return std::make_pair(arr_copy, true);
+      return std::make_pair(contiguous_copy_cpu(arr, s), true);
     }
   };
 
@@ -766,7 +1100,47 @@ void fast::AffineQuantize::eval_cpu(
       }
     } else {
       throw std::runtime_error(
-          "[fast::AffineQuantize::eval_cpu] Only supports floating point inputs");
+          "[fast::Quantize::eval_cpu] Only supports floating point inputs");
+    }
+  });
+}
+
+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;
+      }
     }
   });
 }

mlx/backend/cpu/reduce.cpp

@@ -325,7 +325,15 @@ struct MaxReduce {
   };
 
   template <int N, typename T>
-  T operator()(simd::Simd<T, N> x) {
+  std::enable_if_t<std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    return simd::max(x);
+  };
+
+  template <int N, typename T>
+  std::enable_if_t<!std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    if (simd::any(x != x)) {
+      return static_cast<T>(NAN);
+    }
     return simd::max(x);
   };
 };
@@ -342,7 +350,15 @@ struct MinReduce {
   };
 
   template <int N, typename T>
-  T operator()(simd::Simd<T, N> x) {
+  std::enable_if_t<std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    return simd::min(x);
+  };
+
+  template <int N, typename T>
+  std::enable_if_t<!std::is_integral_v<T>, T> operator()(simd::Simd<T, N> x) {
+    if (simd::any(x != x)) {
+      return static_cast<T>(NAN);
+    }
     return simd::min(x);
   };
 };
@@ -475,19 +491,27 @@ void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
         switch (in.dtype()) {
           case bool_:
           case uint8:
+            reduce_dispatch_sum_prod<uint8_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint16:
+            reduce_dispatch_sum_prod<uint16_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint32:
+            reduce_dispatch_sum_prod<uint32_t>(in, out, reduce_type_, axes_);
+            break;
+          case uint64:
+            reduce_dispatch_sum_prod<uint64_t>(in, out, reduce_type_, axes_);
+            break;
           case int8:
             reduce_dispatch_sum_prod<int8_t>(in, out, reduce_type_, axes_);
             break;
           case int16:
-          case uint16:
             reduce_dispatch_sum_prod<int16_t>(in, out, reduce_type_, axes_);
             break;
           case int32:
-          case uint32:
             reduce_dispatch_sum_prod<int32_t>(in, out, reduce_type_, axes_);
             break;
           case int64:
-          case uint64:
             reduce_dispatch_sum_prod<int64_t>(in, out, reduce_type_, axes_);
             break;
           case float16:
@@ -527,10 +551,10 @@ void Reduce::eval_cpu(const std::vector<array>& inputs, array& out) {
             reduce_dispatch_min_max<uint64_t>(in, out, reduce_type_, axes_);
             break;
           case int8:
-            reduce_dispatch_min_max<uint8_t>(in, out, reduce_type_, axes_);
+            reduce_dispatch_min_max<int8_t>(in, out, reduce_type_, axes_);
             break;
           case int16:
-            reduce_dispatch_min_max<uint16_t>(in, out, reduce_type_, axes_);
+            reduce_dispatch_min_max<int16_t>(in, out, reduce_type_, axes_);
             break;
           case int32:
             reduce_dispatch_min_max<int32_t>(in, out, reduce_type_, axes_);

mlx/backend/cpu/scan.cpp

@@ -250,10 +250,8 @@ void Scan::eval_cpu(const std::vector<array>& inputs, array& out) {
   // Ensure contiguity
   auto in = inputs[0];
   if (!in.flags().row_contiguous) {
-    array arr_copy(in.shape(), in.dtype(), nullptr, {});
-    copy(in, arr_copy, CopyType::General, stream());
-    in = arr_copy;
-    encoder.add_temporary(arr_copy);
+    in = contiguous_copy_cpu(in, stream());
+    encoder.add_temporary(in);
   }
   out.set_data(allocator::malloc(out.nbytes()));
 

mlx/backend/cpu/simd/accelerate_simd.h

@@ -1,5 +1,6 @@
 #pragma once
 
+#include <arm_neon.h>
 #include <simd/math.h>
 #include <simd/vector.h>
 
@@ -9,7 +10,7 @@
 
 #include "mlx/backend/cpu/simd/base_simd.h"
 
-// There seems to be a bug in sims/base.h
+// There seems to be a bug in simd/base_simd.h
 // __XROS_2_0 is not defined, the expression evaluates
 // to true instead of false setting the SIMD library
 // higher than it should be even on macOS < 15
@@ -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>
@@ -234,6 +250,7 @@ Simd<T, N> remainder(Simd<T, N> a, Simd<T, N> b) {
 
 template <typename MaskT, typename T1, typename T2, int N>
 Simd<T1, N> select(Simd<MaskT, N> mask, Simd<T1, N> x, Simd<T2, N> y) {
+  static_assert(std::is_same_v<MaskT, bool>);
   if constexpr (sizeof(T1) == 1) {
     return asd::bitselect(y.value, x.value, asd::convert<char>(mask.value));
   } else if constexpr (sizeof(T1) == 2) {
@@ -251,9 +268,13 @@ Simd<T, N> pow(Simd<T, N> base, Simd<T, N> exp) {
     return asd::pow(base.value, exp.value);
   } else {
     Simd<T, N> res = 1;
-    while (any(exp)) {
-      res = select(exp & 1, res * base, res);
-      base = select(exp, base * base, base);
+    // Raising an integer to a negative power is undefined
+    if (any(exp < 0)) {
+      return 0;
+    }
+    while (any(exp > 0)) {
+      res = select((exp & 1) != 0, res * base, res);
+      base = select(exp > 0, base * base, base);
       exp = exp >> 1;
     }
     return res;

mlx/backend/cpu/simd/base_simd.h

@@ -171,6 +171,11 @@ DEFAULT_BINARY(&)
 DEFAULT_BINARY(&&)
 DEFAULT_BINARY(||)
 
+template <typename T>
+Simd<T, 1> clz(Simd<T, 1> x_) {
+  return __builtin_clz(x_.value);
+}
+
 template <typename T>
 Simd<T, 1> remainder(Simd<T, 1> a_, Simd<T, 1> b_) {
   T a = a_.value;

mlx/backend/cpu/softmax.cpp

@@ -131,8 +131,7 @@ void Softmax::eval_cpu(const std::vector<array>& inputs, array& out) {
       }
       return x;
     } else {
-      array x_copy(x.shape(), x.dtype(), nullptr, {});
-      copy(x, x_copy, CopyType::General, s);
+      array x_copy = contiguous_copy_cpu(x, s);
       out.copy_shared_buffer(x_copy);
       return x_copy;
     }

mlx/backend/cpu/sort.cpp

@@ -8,13 +8,25 @@
 #include "mlx/backend/common/utils.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
-
+#include "mlx/dtype_utils.h"
 #include "mlx/primitives.h"
 
 namespace mlx::core {
 
 namespace {
 
+// NaN-aware comparator that places NaNs at the end
+template <typename T>
+bool nan_aware_less(T a, T b) {
+  if constexpr (std::is_floating_point_v<T> || std::is_same_v<T, complex64_t>) {
+    if (std::isnan(a))
+      return false;
+    if (std::isnan(b))
+      return true;
+  }
+  return a < b;
+}
+
 template <typename T>
 struct StridedIterator {
   using iterator_category = std::random_access_iterator_tag;
@@ -27,7 +39,7 @@ struct StridedIterator {
   StridedIterator() = default;
 
   explicit StridedIterator(T* ptr, int64_t stride, difference_type offset = 0)
-      : ptr_(ptr + offset * stride), stride_(stride) {}
+      : stride_(stride), ptr_(ptr + offset * stride) {}
 
   explicit StridedIterator(array& arr, int axis, difference_type offset = 0)
       : StridedIterator(arr.data<T>(), arr.strides()[axis], offset) {}
@@ -130,7 +142,7 @@ void sort(array& out, int axis) {
     StridedIterator st(data_ptr, axis_stride, 0);
     StridedIterator ed(data_ptr, axis_stride, axis_size);
 
-    std::stable_sort(st, ed);
+    std::stable_sort(st, ed, nan_aware_less<T>);
     src_it.step();
   }
 }
@@ -184,6 +196,15 @@ void argsort(const array& in, array& out, int axis) {
     std::stable_sort(st, ed, [data_ptr, in_stride](IdxT a, IdxT b) {
       auto v1 = data_ptr[a * in_stride];
       auto v2 = data_ptr[b * in_stride];
+
+      // Handle NaNs (place them at the end)
+      if (std::is_floating_point<T>::value) {
+        if (std::isnan(v1))
+          return false;
+        if (std::isnan(v2))
+          return true;
+      }
+
       return v1 < v2 || (v1 == v2 && a < b);
     });
   }
@@ -219,7 +240,7 @@ void partition(array& out, int axis, int kth) {
     StridedIterator md(data_ptr, axis_stride, kth);
     StridedIterator ed(data_ptr, axis_stride, axis_size);
 
-    std::nth_element(st, md, ed);
+    std::nth_element(st, md, ed, nan_aware_less<T>);
   }
 }
 
@@ -276,6 +297,15 @@ void argpartition(const array& in, array& out, int axis, int kth) {
     std::nth_element(st, md, ed, [data_ptr, in_stride](IdxT a, IdxT b) {
       auto v1 = data_ptr[a * in_stride];
       auto v2 = data_ptr[b * in_stride];
+
+      // Handle NaNs (place them at the end)
+      if (std::is_floating_point<T>::value) {
+        if (std::isnan(v1))
+          return false;
+        if (std::isnan(v2))
+          return true;
+      }
+
       return v1 < v2 || (v1 == v2 && a < b);
     });
   }
@@ -333,45 +363,24 @@ void Sort::eval_cpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() == 1);
   auto& in = inputs[0];
 
+  int axis = axis_;
+  if (axis < 0) {
+    axis += in.ndim();
+  }
+
   // Copy input to output
-  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype, stream());
+  CopyType ctype = (in.flags().contiguous && in.strides()[axis] != 0)
+      ? CopyType::Vector
+      : CopyType::General;
+  copy_cpu(in, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_output_array(out);
-  encoder.dispatch(
-      [out = array::unsafe_weak_copy(out), axis_ = axis_]() mutable {
-        switch (out.dtype()) {
-          case bool_:
-            return sort<bool>(out, axis_);
-          case uint8:
-            return sort<uint8_t>(out, axis_);
-          case uint16:
-            return sort<uint16_t>(out, axis_);
-          case uint32:
-            return sort<uint32_t>(out, axis_);
-          case uint64:
-            return sort<uint64_t>(out, axis_);
-          case int8:
-            return sort<int8_t>(out, axis_);
-          case int16:
-            return sort<int16_t>(out, axis_);
-          case int32:
-            return sort<int32_t>(out, axis_);
-          case int64:
-            return sort<int64_t>(out, axis_);
-          case float32:
-            return sort<float>(out, axis_);
-          case float64:
-            return sort<double>(out, axis_);
-          case float16:
-            return sort<float16_t>(out, axis_);
-          case bfloat16:
-            return sort<bfloat16_t>(out, axis_);
-          case complex64:
-            return sort<complex64_t>(out, axis_);
-        }
-      });
+  encoder.dispatch([out = array::unsafe_weak_copy(out), axis]() mutable {
+    dispatch_all_types(out.dtype(), [&](auto type_tag) {
+      sort<MLX_GET_TYPE(type_tag)>(out, axis);
+    });
+  });
 }
 
 void ArgPartition::eval_cpu(const std::vector<array>& inputs, array& out) {
@@ -426,8 +435,10 @@ void Partition::eval_cpu(const std::vector<array>& inputs, array& out) {
   auto& in = inputs[0];
 
   // Copy input to output
-  CopyType ctype = in.flags().contiguous ? CopyType::Vector : CopyType::General;
-  copy(in, out, ctype, stream());
+  CopyType ctype = (in.flags().contiguous && in.strides()[axis_] != 0)
+      ? CopyType::Vector
+      : CopyType::General;
+  copy_cpu(in, out, ctype, stream());
 
   auto& encoder = cpu::get_command_encoder(stream());
   encoder.set_output_array(out);