Refactoring launcher

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

.circleci/config.yml

@@ -1,579 +0,0 @@
-version: 2.1
-
-orbs:
-  apple: ml-explore/pr-approval@0.1.0
-
-parameters:
-  nightly_build:
-    type: boolean
-    default: false
-  test_release:
-    type: boolean
-    default: false
-
-jobs:
-  build_documentation:
-    parameters:
-      upload-docs:
-        type: boolean
-        default: false
-    macos:
-      xcode: "26.0.0"
-    resource_class: m4pro.medium
-    steps:
-      - checkout
-      - run:
-          name: Install
-          command: |
-            xcodebuild -downloadComponent MetalToolchain
-            brew install python@3.9
-            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
-            pip install . -v
-      - when:
-          condition:
-            not: << parameters.upload-docs >>
-          steps:
-            - run:
-               name: Build documentation
-               command: |
-                 source env/bin/activate
-                 cd docs && doxygen && make html O=-W
-      - when:
-          condition: << parameters.upload-docs >>
-          steps:
-            - add_ssh_keys:
-                fingerprints:
-                  - "SHA256:OhcVVMovbT0pkgMeiVRyxMnjV9R2t+hKBsNcuxq9h+0"
-            - run:
-               name: Upload documentation
-               command: |
-                 source env/bin/activate
-                 git config user.email "mlx@group.apple.com"
-                 git config user.name "CircleCI Docs"
-                 git checkout gh-pages
-                 git rebase main
-                 cd docs
-                 git rm -rf build/html
-                 doxygen && make html O=-W
-                 git add -f build/html
-                 git commit -m "rebase"
-                 git push -f origin gh-pages
-
-  linux_build_and_test:
-    machine:
-      image: ubuntu-2204:current
-      resource_class: large
-    steps:
-      - checkout
-      - run:
-          name: Run style checks
-          command: |
-            pip install pre-commit
-            pre-commit run --all
-            if ! git diff --quiet; then echo 'Style checks failed, please install pre-commit and run pre-commit run --all and push the change'; exit 1; fi
-      - run:
-          name: Install dependencies
-          command: |
-            export DEBIAN_FRONTEND=noninteractive
-            export NEEDRESTART_MODE=a
-            sudo apt-get update
-            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
-            curl -LsSf https://astral.sh/uv/install.sh | sh
-      - run:
-          name: Install Python package
-          command: |
-            uv venv
-            uv pip install cmake
-            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              uv pip install -e ".[dev]" -v
-      - run:
-          name: Generate package stubs
-          command: |
-            uv pip install typing_extensions
-            uv run --no-project setup.py generate_stubs
-      - run:
-          name: Run Python tests
-          command: |
-            source .venv/bin/activate
-            python -m unittest discover python/tests -v
-            mpirun --bind-to none -host localhost:8 -np 8 python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
-            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
-      - run:
-          name: Build CPP only
-          command: |
-            source .venv/bin/activate
-            mkdir -p build && cd build
-            cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
-            make -j `nproc`
-      - run:
-          name: Run CPP tests
-          command: ./build/tests/tests
-
-  mac_build_and_test:
-    parameters:
-      xcode_version:
-        type: string
-        default: "26.0.0"
-      macosx_deployment_target:
-        type: string
-        default: ""
-    macos:
-      xcode: << parameters.xcode_version >>
-    environment:
-      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    resource_class: m4pro.medium
-    steps:
-      - checkout
-      - run:
-          name: Install dependencies
-          command: |
-            xcodebuild -downloadComponent MetalToolchain
-            HOMEBREW_NO_AUTO_UPDATE=1 HOMEBREW_NO_INSTALL_CLEANUP=1 \
-              brew install openmpi uv
-      - run:
-          name: Install Python package
-          command: |
-            uv venv --python 3.9
-            uv pip install \
-              nanobind==2.4.0 \
-              cmake \
-              numpy \
-              torch \
-              tensorflow \
-              unittest-xml-reporting
-            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              uv pip install -e . -v
-      - run:
-          name: Generate package stubs
-          command: |
-            uv pip install typing_extensions
-            uv run --no-project setup.py generate_stubs
-      - run:
-          name: Run Python tests
-          command: |
-            source .venv/bin/activate
-            LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
-            LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
-            mpirun --bind-to none -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
-            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
-      - run:
-          name: Build example extension
-          command: |
-            source .venv/bin/activate
-            cd examples/extensions
-            uv pip install -r requirements.txt
-            uv run --no-project setup.py build_ext --inplace
-            uv run --no-project python test.py
-      - store_test_results:
-          path: test-results
-      - run:
-          name: Build CPP only
-          command: |
-            source .venv/bin/activate
-            mkdir -p build && cd build && cmake .. && make -j `sysctl -n hw.ncpu`
-      - run:
-          name: Run CPP tests
-          command: |
-            DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 ./build/tests/tests
-      - run:
-          name: Build small binary
-          command: |
-            source .venv/bin/activate
-            cd build/
-            cmake .. -DCMAKE_BUILD_TYPE=MinSizeRel \
-              -DBUILD_SHARED_LIBS=ON \
-              -DMLX_BUILD_CPU=OFF \
-              -DMLX_BUILD_SAFETENSORS=OFF \
-              -DMLX_BUILD_GGUF=OFF \
-              -DMLX_METAL_JIT=ON
-            make -j `sysctl -n hw.ncpu`
-      - run:
-          name: Run Python tests with JIT
-          command: |
-            CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-              uv pip install -e . -v
-            LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
-              METAL_DEBUG_ERROR_MODE=0 \
-              uv run --no-project python -m xmlrunner discover \
-                -v python/tests \
-                -o test-results/gpu_jit
-
-  cuda_build_and_test:
-    parameters:
-      image_date:
-        type: string
-        default: "2023.11.1"
-    machine:
-      image: "linux-cuda-12:<< parameters.image_date >>"
-      resource_class: gpu.nvidia.small.gen2
-    steps:
-      - checkout
-      - restore_cache:
-          keys:
-            - cuda-<< parameters.image_date >>-{{ arch }}-
-      - run:
-          name: Install dependencies
-          command: |
-            sudo apt-get update
-            sudo apt-get install libcudnn9-dev-cuda-12
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install libnccl2 libnccl-dev
-            curl -sL https://github.com/ccache/ccache/releases/download/v4.11.3/ccache-4.11.3-linux-x86_64.tar.xz | tar xJf -
-            sudo mv ccache-4.11.3-linux-x86_64/ccache /usr/bin/ccache
-            rm -rf ccache-4.11.3-linux-x86_64
-            curl -LsSf https://astral.sh/uv/install.sh | sh
-      - run:
-          name: Set CCache size
-          command: ccache --max-size 1G
-      - run:
-          name: Install Python package
-          command: |
-            uv venv
-            uv pip install cmake
-            DEBUG=1 CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_COMPILE_WARNING_AS_ERROR=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-              uv pip install -e ".[dev]" -v
-      - run:
-          name: Run Python tests
-          command: |
-            source .venv/bin/activate
-            LOW_MEMORY=1 DEVICE=cpu python -m unittest discover python/tests -v
-            LOW_MEMORY=1 DEVICE=gpu python -m tests discover python/tests -v
-      - run:
-          name: Build CPP only
-          command: |
-            source .venv/bin/activate
-            cmake . -B build \
-              -DMLX_BUILD_CUDA=ON \
-              -DCMAKE_CUDA_COMPILER=`which nvcc` \
-              -DCMAKE_BUILD_TYPE=DEBUG
-            cmake --build build -j `nproc`
-      - run:
-          name: Run CPP tests
-          command: ./build/tests/tests -sfe="*fft_tests.cpp,*linalg_tests.cpp"
-      - run:
-          name: CCache report
-          command: |
-            ccache --show-stats
-            ccache --zero-stats
-            ccache --cleanup
-      - save_cache:
-          key: cuda-<< parameters.image_date >>-{{ arch }}-{{ epoch }}
-          paths:
-            - /home/circleci/.cache/ccache
-
-  build_release:
-    parameters:
-      python_version:
-        type: string
-        default: "3.9"
-      xcode_version:
-        type: string
-        default: "26.0.0"
-      build_env:
-        type: string
-        default: ""
-      macosx_deployment_target:
-        type: string
-        default: ""
-    macos:
-      xcode: << parameters.xcode_version >>
-    resource_class: m4pro.medium
-    environment:
-      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    steps:
-      - checkout
-      - run:
-          name: Install dependencies
-          command: |
-            xcodebuild -downloadComponent MetalToolchain
-            mkdir -p ~/miniconda3
-            curl https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-arm64.sh -o ~/miniconda3/miniconda.sh
-            bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
-            rm ~/miniconda3/miniconda.sh
-            source ~/miniconda3/bin/activate
-            conda init --all
-            conda create -n env python=<< parameters.python_version >> -y
-            conda activate env
-            pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install --upgrade setuptools
-            pip install numpy
-            pip install twine
-            pip install build
-      - run:
-          name: Install Python package
-          command: |
-            conda activate env
-            env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
-              pip install . -v
-      - run:
-          name: Generate package stubs
-          command: |
-            conda activate env
-            pip install typing_extensions
-            python setup.py generate_stubs
-      - run:
-          name: Build Python package
-          command: |
-            conda activate env
-            python setup.py clean --all
-            << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
-      - when:
-          condition:
-            equal: ["3.9", << parameters.python_version >>]
-          steps:
-            - run:
-                name: Build common package
-                command: |
-                  conda activate env
-                  python setup.py clean --all
-                  << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
-      - when:
-          condition: << parameters.build_env >>
-          steps:
-            - run:
-                name: Upload package
-                command: |
-                  conda activate env
-                  twine upload dist/*
-      - store_artifacts:
-          path: dist/
-
-  build_linux_release:
-    parameters:
-      python_version:
-        type: string
-        default: "3.9"
-      build_env:
-        type: string
-        default: ""
-    machine:
-      image: ubuntu-2204:current
-      resource_class: large
-    steps:
-      - checkout
-      - run:
-          name: Build wheel
-          command: |
-            PYTHON=python<< parameters.python_version >>
-            export DEBIAN_FRONTEND=noninteractive
-            export NEEDRESTART_MODE=a
-            sudo apt-get update
-            TZ=Etc/UTC sudo apt-get -y install tzdata
-            sudo add-apt-repository -y ppa:deadsnakes/ppa
-            sudo apt-get install -y $PYTHON $PYTHON-dev $PYTHON-full
-            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
-            $PYTHON -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            pip install --upgrade cmake
-            pip install auditwheel
-            pip install patchelf
-            pip install build
-            pip install twine
-            << parameters.build_env >> pip install ".[dev]" -v
-            pip install typing_extensions
-            python setup.py generate_stubs
-            python setup.py clean --all
-            MLX_BUILD_STAGE=1 << parameters.build_env >> python -m build -w
-            bash python/scripts/repair_linux.sh
-      - when:
-          condition:
-            equal: ["3.9", << parameters.python_version >>]
-          steps:
-            - run:
-                name: Build common package
-                command: |
-                  source env/bin/activate
-                  python setup.py clean --all
-                  << parameters.build_env >> MLX_BUILD_STAGE=2 \
-                    python -m build -w
-                  auditwheel repair dist/mlx_cpu*.whl --plat manylinux_2_35_x86_64
-      - when:
-          condition: << parameters.build_env >>
-          steps:
-            - run:
-                name: Upload packages
-                command: |
-                  source env/bin/activate
-                  twine upload wheelhouse/*.whl
-      - store_artifacts:
-          path: wheelhouse/
-
-  build_cuda_release:
-    parameters:
-      build_env:
-        type: string
-        default: ""
-    machine:
-      image: ubuntu-2204:current
-      resource_class: xlarge
-    steps:
-      - checkout
-      - run:
-          name: Build wheel
-          command: |
-            export DEBIAN_FRONTEND=noninteractive
-            export NEEDRESTART_MODE=a
-            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
-            sudo dpkg -i cuda-keyring_1.1-1_all.deb
-            sudo apt-get update
-            sudo apt-get install cuda-toolkit-12-9 libcudnn9-dev-cuda-12
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install zip
-            pip install auditwheel
-            pip install patchelf
-            pip install build
-            pip install twine
-            export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}
-            export LD_LIBRARY_PATH=/usr/local/cuda/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-            << parameters.build_env >> MLX_BUILD_STAGE=2 \
-              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-              python -m build -w
-            bash python/scripts/repair_cuda.sh
-      - when:
-          condition: << parameters.build_env >>
-          steps:
-            - run:
-                name: Upload package
-                command: |
-                  twine upload wheelhouse/*.whl
-      - store_artifacts:
-          path: wheelhouse/
-
-workflows:
-  build_and_test:
-    when:
-      and:
-        - matches:
-            pattern: "^(?!pull/)[-\\w]+$"
-            value: << pipeline.git.branch >>
-        - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.test_release >>
-    jobs:
-      - mac_build_and_test:
-          matrix:
-            parameters:
-              macosx_deployment_target: ["13.5", "15.0"]
-      - linux_build_and_test
-      - cuda_build_and_test:
-          matrix:
-            parameters:
-              image_date: ["2023.11.1", "2025.05.1"]
-      - build_documentation 
-
-  build_pypi_release:
-    when:
-      and:
-        - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.test_release >>
-    jobs:
-      - build_release:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              build_env: ["PYPI_RELEASE=1"]
-              xcode_version: ["26.0.0"]
-      - build_documentation:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          upload-docs: true
-      - build_linux_release:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              build_env: ["PYPI_RELEASE=1"]
-      - build_cuda_release:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          matrix:
-            parameters:
-              build_env: ["PYPI_RELEASE=1"]
-
-  prb:
-    when:
-      matches:
-        pattern: "^pull/\\d+(/head)?$"
-        value: << pipeline.git.branch >>
-    jobs:
-      - hold:
-          type: approval
-      - apple/authenticate:
-          context: pr-approval
-      - mac_build_and_test:
-          requires: [ hold ]
-          matrix:
-            parameters:
-              macosx_deployment_target: ["13.5", "15.0"]
-      - linux_build_and_test:
-          requires: [ hold ]
-      - cuda_build_and_test:
-          requires: [ hold ]
-          matrix:
-            parameters:
-              image_date: ["2023.11.1", "2025.05.1"]
-  nightly_build:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.nightly_build >>
-    jobs:
-      - build_release:
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              xcode_version: ["26.0.0"]
-      - build_linux_release:
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-      - build_cuda_release
-
-  build_dev_release:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.test_release >>
-    jobs:
-      - build_release:
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              build_env: ["DEV_RELEASE=1"]
-              xcode_version: ["26.0.0"]
-      - build_linux_release:
-          matrix:
-            parameters:
-              python_version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
-              build_env: ["DEV_RELEASE=1"]
-      - build_cuda_release:
-          matrix:
-            parameters:
-              build_env: ["DEV_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:

CMakeLists.txt

@@ -74,6 +74,7 @@ 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}")
@@ -88,6 +89,11 @@ cmake_policy(SET CMP0135 NEW)
 
 add_library(mlx)
 
+# 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()
@@ -113,6 +119,10 @@ if(MLX_BUILD_METAL)
     COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-version"
     OUTPUT_VARIABLE MACOS_SDK_VERSION
     OUTPUT_STRIP_TRAILING_WHITESPACE COMMAND_ERROR_IS_FATAL ANY)
+  execute_process(
+    COMMAND zsh "-c" "/usr/bin/xcrun -sdk macosx --show-sdk-path"
+    OUTPUT_VARIABLE CMAKE_OSX_SYSROOT
+    OUTPUT_STRIP_TRAILING_WHITESPACE COMMAND_ERROR_IS_FATAL ANY)
 
   if(${MACOS_SDK_VERSION} LESS 14.0)
     message(
@@ -122,9 +132,12 @@ if(MLX_BUILD_METAL)
   message(STATUS "Building with macOS SDK version ${MACOS_SDK_VERSION}")
 
   set(METAL_CPP_URL
-      https://developer.apple.com/metal/cpp/files/metal-cpp_macOS15_iOS18.zip)
+      https://developer.apple.com/metal/cpp/files/metal-cpp_26.zip)
 
   if(NOT CMAKE_OSX_DEPLOYMENT_TARGET STREQUAL "")
+    if(${CMAKE_OSX_DEPLOYMENT_TARGET} LESS 14.0)
+      message(FATAL_ERROR "MLX requires macOS >= 14.0")
+    endif()
     set(XCRUN_FLAGS "-mmacosx-version-min=${CMAKE_OSX_DEPLOYMENT_TARGET}")
   endif()
   execute_process(
@@ -133,7 +146,6 @@ if(MLX_BUILD_METAL)
       "echo \"__METAL_VERSION__\" | xcrun -sdk macosx metal ${XCRUN_FLAGS} -E -x metal -P - | tail -1 | tr -d '\n'"
     OUTPUT_VARIABLE MLX_METAL_VERSION COMMAND_ERROR_IS_FATAL ANY)
   FetchContent_Declare(metal_cpp URL ${METAL_CPP_URL})
-
   FetchContent_MakeAvailable(metal_cpp)
   target_include_directories(
     mlx PUBLIC $<BUILD_INTERFACE:${metal_cpp_SOURCE_DIR}>

benchmarks/cpp/irregular_strides.cpp

@@ -75,7 +75,7 @@ void time_irregular_binary_ops_3D() {
 
 void time_irregular_binary_ops_4D() {
   auto device = mx::default_device();
-  std::vector<int> shape = {8, 8, 512, 512};
+  mx::Shape shape = {8, 8, 512, 512};
   auto a = mx::random::uniform(shape);
   auto b = mx::random::uniform(shape);
 
@@ -115,7 +115,7 @@ void time_irregular_binary_ops_4D() {
 
 void time_irregular_reshape() {
   auto device = mx::default_device();
-  std::vector<int> shape;
+  mx::Shape shape;
   auto reshape_fn = [&shape, device](const mx::array& a) {
     return mx::reshape(a, shape, device);
   };
@@ -170,7 +170,7 @@ void time_irregular_astype_1D() {
 void time_irregular_astype_2D() {
   auto device = mx::default_device();
   int size = 2048;
-  std::vector<int> shape = {size, size};
+  mx::Shape shape = {size, size};
 
   auto a = mx::random::uniform(shape);
   TIMEM("2D regular", mx::astype, a, mx::int32, device);

benchmarks/python/blas/bench_gemv.py

@@ -1,6 +1,5 @@
 # Copyright © 2023 Apple Inc.
 
-import argparse
 import os
 import subprocess
 import time

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()

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/src/install.rst

@@ -16,12 +16,11 @@ silicon computer is
 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
 ^^^^
@@ -39,7 +38,7 @@ requirements:
 - Nvidia driver >= 550.54.14
 - CUDA toolkit >= 12.0
 - Linux distribution with glibc >= 2.35
-- Python >= 3.9
+- Python >= 3.10
 
 
 CPU-only (Linux)
@@ -55,7 +54,7 @@ To install the CPU-only package from PyPi your system must meet the following
 requirements:
 
 - Linux distribution with glibc >= 2.35
-- Python >= 3.9
+- Python >= 3.10
 
 
 Troubleshooting

docs/src/python/ops.rst

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

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
@@ -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/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
@@ -143,3 +144,51 @@ expected. For example:
 
 In the above ``dfdx`` will have the correct gradient, namely zeros at ``idx``
 and ones elsewhere.
+
+.. _boolean-mask-assignment:
+
+Boolean Mask Assignment
+-----------------------
+
+MLX supports boolean indices using NumPy syntax. A mask must already be
+a :class:`bool_` MLX :class:`array` or a NumPy ``ndarray`` with ``dtype=bool``.
+Other index types are routed through the standard scatter code.
+
+.. code-block:: shell
+
+   >>> a = mx.array([1.0, 2.0, 3.0])
+   >>> mask = mx.array([True, False, True])
+   >>> updates = mx.array([5.0, 6.0])
+   >>> a[mask] = updates
+   >>> a
+   array([5.0, 2.0, 6.0], dtype=float32)
+
+Scalar assignments broadcast to every ``True`` entry in ``mask``. For non-scalar
+assignments, ``updates`` must provide at least as many elements as there are
+``True`` entries in ``mask``.
+
+.. code-block:: shell
+
+   >>> a = mx.zeros((2, 3))
+   >>> mask = mx.array([[True, False, True],
+                        [False, False, True]])
+   >>> a[mask] = 1.0
+   >>> a
+   array([[1.0, 0.0, 1.0],
+          [0.0, 0.0, 1.0]], dtype=float32)
+
+Boolean masks follow NumPy semantics:
+
+- The mask shape must match the shape of the axes it indexes exactly. The only
+  exception is a scalar boolean mask, which broadcasts to the full array.
+- Any axes not covered by the mask are taken in full.
+
+.. code-block:: shell
+
+   >>> a = mx.arange(1000).reshape(10, 10, 10)
+   >>> a[mx.random.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.

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

@@ -294,6 +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;
@@ -349,15 +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 {
@@ -426,7 +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);
 
@@ -461,8 +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

@@ -114,7 +114,9 @@ void compiled_allocate_outputs(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
     const std::function<bool(size_t)>& is_constant,
-    bool contiguous) {
+    bool contiguous,
+    const std::function<allocator::Buffer(size_t)>&
+        mallocfn /* = allocator::malloc */) {
   if (contiguous) {
     int o = 0;
     Strides strides;
@@ -140,7 +142,7 @@ void compiled_allocate_outputs(
     }
     for (; o < outputs.size(); ++o) {
       outputs[o].set_data(
-          allocator::malloc(data_size * outputs[o].itemsize()),
+          mallocfn(data_size * outputs[o].itemsize()),
           data_size,
           strides,
           flags);
@@ -163,7 +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

@@ -58,7 +58,9 @@ void compiled_allocate_outputs(
     const std::vector<array>& inputs,
     std::vector<array>& outputs,
     const std::function<bool(size_t)>& is_constant,
-    bool contiguous);
+    bool contiguous,
+    const std::function<allocator::Buffer(size_t)>& mallocfn =
+        allocator::malloc);
 
 // Collapse contiguous dims ignoring scalars and constants.
 std::tuple<bool, Shape, std::vector<Strides>> compiled_collapse_contiguous_dims(

mlx/backend/common/copy.h

@@ -22,7 +22,11 @@ enum class CopyType {
   GeneralGeneral
 };
 
-inline bool set_copy_output_data(const array& in, array& out, CopyType ctype) {
+inline bool set_copy_output_data(
+    const array& in,
+    array& out,
+    CopyType ctype,
+    std::function<allocator::Buffer(size_t)> mallocfn = allocator::malloc) {
   if (ctype == CopyType::Vector) {
     // If the input is donateable, we are doing a vector copy and the types
     // have the same size, then the input buffer can hold the output.
@@ -31,14 +35,14 @@ inline bool set_copy_output_data(const array& in, array& out, CopyType ctype) {
       return true;
     } else {
       out.set_data(
-          allocator::malloc(in.data_size() * out.itemsize()),
+          mallocfn(in.data_size() * out.itemsize()),
           in.data_size(),
           in.strides(),
           in.flags());
       return false;
     }
   } else {
-    out.set_data(allocator::malloc(out.nbytes()));
+    out.set_data(mallocfn(out.nbytes()));
     return false;
   }
 }

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

@@ -46,7 +46,8 @@ inline void set_ternary_op_output_data(
     const array& b,
     const array& c,
     array& out,
-    TernaryOpType topt) {
+    TernaryOpType topt,
+    std::function<allocator::Buffer(size_t)> mallocfn = allocator::malloc) {
   auto maybe_donate = [&out](const array& x) {
     if (is_donatable(x, out)) {
       out.copy_shared_buffer(x);
@@ -57,13 +58,12 @@ inline void set_ternary_op_output_data(
 
   switch (topt) {
     case TernaryOpType::ScalarScalarScalar:
-      out.set_data(
-          allocator::malloc(out.itemsize()), 1, b.strides(), b.flags());
+      out.set_data(mallocfn(out.itemsize()), 1, b.strides(), b.flags());
       break;
     case TernaryOpType::VectorVectorVector:
       if (!(maybe_donate(a) || maybe_donate(b) || maybe_donate(c))) {
         out.set_data(
-            allocator::malloc(out.itemsize() * b.data_size()),
+            mallocfn(out.itemsize() * b.data_size()),
             b.data_size(),
             b.strides(),
             b.flags());
@@ -76,7 +76,7 @@ inline void set_ternary_op_output_data(
       if (!((a.flags().row_contiguous && maybe_donate(a)) ||
             (b.flags().row_contiguous && maybe_donate(b)) ||
             (c.flags().row_contiguous && maybe_donate(c)))) {
-        out.set_data(allocator::malloc(out.nbytes()));
+        out.set_data(mallocfn(out.nbytes()));
       }
       break;
   }

mlx/backend/common/unary.h

@@ -7,19 +7,22 @@
 
 namespace mlx::core {
 
-inline void set_unary_output_data(const array& in, array& out) {
+inline void set_unary_output_data(
+    const array& in,
+    array& out,
+    std::function<allocator::Buffer(size_t)> mallocfn = allocator::malloc) {
   if (in.flags().contiguous) {
     if (is_donatable(in, out)) {
       out.copy_shared_buffer(in);
     } else {
       out.set_data(
-          allocator::malloc(in.data_size() * out.itemsize()),
+          mallocfn(in.data_size() * out.itemsize()),
           in.data_size(),
           in.strides(),
           in.flags());
     }
   } else {
-    out.set_data(allocator::malloc(out.nbytes()));
+    out.set_data(mallocfn(out.nbytes()));
   }
 }
 

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/conv.cpp

@@ -996,131 +996,6 @@ void explicit_gemm_conv_1D_cpu(
   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_cpu(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_cpu_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_cpu(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_cpu(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_cpu_inplace(gemm_out, out, CopyType::Vector, stream);
-  }
-  encoder.add_temporaries(std::move(temps));
-}
-
 void explicit_gemm_conv_ND_cpu(
     const array& in,
     const array& wt,

mlx/backend/cpu/distributed.cpp

@@ -95,4 +95,9 @@ void Recv::eval_cpu(
   distributed::detail::recv(group(), outputs[0], src_, stream());
 }
 
+void ReduceScatter::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  throw std::runtime_error("[ReduceScatter] Not implemented yet.");
+}
 } // namespace mlx::core::distributed

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());
       }
     }
@@ -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/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/indexing.cpp

@@ -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/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(gesdd)
 INSTANTIATE_LAPACK_REAL(getrf)
 INSTANTIATE_LAPACK_REAL(getri)
 INSTANTIATE_LAPACK_REAL(trtri)
@@ -63,3 +61,20 @@ INSTANTIATE_LAPACK_REAL(trtri)
   }
 
 INSTANTIATE_LAPACK_COMPLEX(heevd)
+
+#define INSTANTIATE_LAPACK_ALL(FUNC)                                \
+  template <typename T, typename... Args>                           \
+  void FUNC(Args... args) {                                         \
+    if constexpr (std::is_same_v<T, float>) {                       \
+      MLX_LAPACK_FUNC(s##FUNC)(std::forward<Args>(args)...);        \
+    } else if constexpr (std::is_same_v<T, double>) {               \
+      MLX_LAPACK_FUNC(d##FUNC)(std::forward<Args>(args)...);        \
+    } else if constexpr (std::is_same_v<T, std::complex<float>>) {  \
+      MLX_LAPACK_FUNC(c##FUNC)(std::forward<Args>(args)...);        \
+    } else if constexpr (std::is_same_v<T, std::complex<double>>) { \
+      MLX_LAPACK_FUNC(z##FUNC)(std::forward<Args>(args)...);        \
+    }                                                               \
+  }
+
+INSTANTIATE_LAPACK_ALL(geev)
+INSTANTIATE_LAPACK_ALL(gesdd)

mlx/backend/cpu/masked_mm.cpp

@@ -215,18 +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];
@@ -423,7 +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());

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"
@@ -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;
   }
@@ -136,15 +137,29 @@ void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
     return;
   }
 
+  // Handle empty matrix case (K=0)
+  if (inputs[0].shape(-1) == 0) {
+    auto& c = inputs[2];
+    if (beta_ == 1.0f) {
+      CopyType ctype = c.data_size() == 1
+          ? CopyType::Scalar
+          : (c.flags().row_contiguous ? CopyType::Vector : CopyType::General);
+      copy_cpu(c, out, ctype, stream());
+    } else {
+      array beta_scalar = array(beta_, c.dtype());
+      auto& encoder = cpu::get_command_encoder(stream());
+      binary_float_op_cpu(c, beta_scalar, out, detail::Multiply(), stream());
+      encoder.add_temporary(std::move(beta_scalar));
+    }
+    return;
+  }
+
   // Fill output with C
   auto& c = inputs[2];
   CopyType ctype = c.data_size() == 1
       ? CopyType::Scalar
       : (c.flags().row_contiguous ? CopyType::Vector : CopyType::General);
   copy_cpu(c, out, ctype, stream());
-  if (inputs[0].shape(-1) == 0) {
-    return;
-  }
   matmul_general(inputs[0], inputs[1], out, stream(), alpha_, beta_);
 }
 

mlx/backend/cpu/primitives.cpp

@@ -333,7 +333,7 @@ void Reshape::eval_cpu(const std::vector<array>& inputs, array& out) {
 
 void DynamicSlice::eval_cpu(const std::vector<array>& inputs, array& out) {
   if (out.size() == 0) {
-    out.set_data(nullptr);
+    out.set_data(allocator::malloc(0));
     return;
   }
   auto& in = inputs[0];
@@ -361,7 +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;
   }
 
@@ -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;
   }
 

mlx/backend/cpu/quantized.cpp

@@ -1,8 +1,11 @@
 // Copyright © 2023 Apple Inc.
 
+#include "mlx/backend/common/unary.h"
 #include "mlx/backend/cpu/copy.h"
 #include "mlx/backend/cpu/encoder.h"
 #include "mlx/backend/cpu/simd/simd.h"
+#include "mlx/backend/cpu/unary.h"
+#include "mlx/backend/cpu/unary_ops.h"
 #include "mlx/fast_primitives.h"
 #include "mlx/primitives.h"
 #include "mlx/utils.h"
@@ -445,7 +448,6 @@ void mxfp4_qmm(
     int K) {
   constexpr int group_size = 32;
   constexpr int pack_factor = get_pack_factor(4, 8);
-  constexpr int bytes_per_pack = get_bytes_per_pack(4);
   constexpr int packs_in_group = group_size / pack_factor;
 
   for (int m = 0; m < M; m++) {
@@ -487,7 +489,6 @@ void mxfp4_qmm_t(
     int K) {
   constexpr int group_size = 32;
   constexpr int pack_factor = get_pack_factor(4, 8);
-  constexpr int bytes_per_pack = get_bytes_per_pack(4);
   constexpr int packs_in_group = group_size / pack_factor;
 
   for (int m = 0; m < M; m++) {
@@ -1104,4 +1105,44 @@ void fast::Quantize::eval_cpu(
   });
 }
 
+void fast::ConvertFP8::eval_cpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  auto& in = inputs[0];
+  auto& out = outputs[0];
+  set_unary_output_data(in, out);
+  auto& encoder = cpu::get_command_encoder(stream());
+  encoder.set_input_array(in);
+  encoder.set_output_array(out);
+  encoder.dispatch([in = array::unsafe_weak_copy(in),
+                    out = array::unsafe_weak_copy(out),
+                    to_fp8 = to_fp8_]() mutable {
+    if (to_fp8) {
+      switch (in.dtype()) {
+        case float16:
+          unary_op<float16_t, uint8_t>(in, out, detail::ToFP8());
+          break;
+        case bfloat16:
+          unary_op<bfloat16_t, uint8_t>(in, out, detail::ToFP8());
+          break;
+        default:
+          unary_op<float, uint8_t>(in, out, detail::ToFP8());
+          break;
+      }
+    } else {
+      switch (out.dtype()) {
+        case float16:
+          unary_op<uint8_t, float16_t>(in, out, detail::FromFP8());
+          break;
+        case bfloat16:
+          unary_op<uint8_t, bfloat16_t>(in, out, detail::FromFP8());
+          break;
+        default:
+          unary_op<uint8_t, float>(in, out, detail::FromFP8());
+          break;
+      }
+    }
+  });
+}
+
 } // namespace mlx::core

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>
 
@@ -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>

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/sort.cpp

@@ -39,7 +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) {}

mlx/backend/cpu/svd.cpp

@@ -8,6 +8,183 @@
 
 namespace mlx::core {
 
+template <typename T, class Enable = void>
+struct SVDWork {};
+
+template <typename T>
+struct SVDWork<
+    T,
+    typename std::enable_if<std::is_floating_point<T>::value>::type> {
+  using R = T;
+
+  int N;
+  int M;
+  int K;
+  int lda;
+  int ldu;
+  int ldvt;
+  char jobz;
+  std::vector<array::Data> buffers;
+  int lwork;
+
+  SVDWork(int N, int M, int K, char jobz)
+      : N(N), M(M), K(K), lda(N), ldu(N), ldvt(M), jobz(jobz) {
+    T workspace_dimension = 0;
+
+    // Will contain the indices of eigenvectors that failed to converge (not
+    // used here but required by lapack).
+    buffers.emplace_back(allocator::malloc(sizeof(int) * 8 * K));
+
+    int lwork_query = -1;
+    int info;
+
+    // Compute workspace size.
+    gesdd<T>(
+        /* jobz = */ &jobz,
+        // M and N are swapped since lapack expects column-major.
+        /* m = */ &N,
+        /* n = */ &M,
+        /* a = */ nullptr,
+        /* lda = */ &lda,
+        /* s = */ nullptr,
+        /* u = */ nullptr,
+        /* ldu = */ &ldu,
+        /* vt = */ nullptr,
+        /* ldvt = */ &ldvt,
+        /* work = */ &workspace_dimension,
+        /* lwork = */ &lwork_query,
+        /* iwork = */ static_cast<int*>(buffers[0].buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "[SVD::eval_cpu] workspace calculation failed with code " << info;
+      throw std::runtime_error(ss.str());
+    }
+
+    lwork = workspace_dimension;
+    buffers.emplace_back(allocator::malloc(sizeof(T) * lwork));
+  }
+
+  void run(T* a, R* s, T* u, T* vt) {
+    int info;
+    gesdd<T>(
+        /* jobz = */ &jobz,
+        // M and N are swapped since lapack expects column-major.
+        /* m = */ &N,
+        /* n = */ &M,
+        /* a = */ a,
+        /* lda = */ &lda,
+        /* s = */ s,
+        // According to the identity above, lapack will write Vᵀᵀ as U.
+        /* u = */ u,
+        /* ldu = */ &ldu,
+        // According to the identity above, lapack will write Uᵀ as Vᵀ.
+        /* vt = */ vt,
+        /* ldvt = */ &ldvt,
+        /* work = */ static_cast<T*>(buffers[1].buffer.raw_ptr()),
+        /* lwork = */ &lwork,
+        /* iwork = */ static_cast<int*>(buffers[0].buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "svd_impl: sgesvdx_ failed with code " << info;
+      throw std::runtime_error(ss.str());
+    }
+  }
+};
+
+template <>
+struct SVDWork<std::complex<float>> {
+  using T = std::complex<float>;
+  using R = float;
+
+  int N;
+  int M;
+  int K;
+  int lda;
+  int ldu;
+  int ldvt;
+  char jobz;
+  std::vector<array::Data> buffers;
+  int lwork;
+
+  SVDWork(int N, int M, int K, char jobz)
+      : N(N), M(M), K(K), lda(N), ldu(N), ldvt(M), jobz(jobz) {
+    T workspace_dimension = 0;
+
+    // Will contain the indices of eigenvectors that failed to converge (not
+    // used here but required by lapack).
+    buffers.emplace_back(allocator::malloc(sizeof(int) * 8 * K));
+
+    const int lrwork =
+        jobz == 'A' ? std::max(1, 5 * K * K + 5 * K) : std::max(1, 7 * K);
+    buffers.emplace_back(allocator::malloc(sizeof(float) * lrwork));
+
+    int lwork_query = -1;
+    int work_query = -1;
+    int info;
+
+    // Compute workspace size.
+    gesdd<T>(
+        /* jobz = */ &jobz,
+        // M and N are swapped since lapack expects column-major.
+        /* m = */ &N,
+        /* n = */ &M,
+        /* a = */ nullptr,
+        /* lda = */ &lda,
+        /* s = */ nullptr,
+        /* u = */ nullptr,
+        /* ldu = */ &ldu,
+        /* vt = */ nullptr,
+        /* ldvt = */ &ldvt,
+        /* work = */ &workspace_dimension,
+        /* lwork = */ &lwork_query,
+        /* rwork = */ static_cast<float*>(buffers[1].buffer.raw_ptr()),
+        /* iwork = */ static_cast<int*>(buffers[0].buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "[SVD::eval_cpu] workspace calculation failed with code " << info;
+      throw std::runtime_error(ss.str());
+    }
+
+    lwork = workspace_dimension.real();
+    buffers.emplace_back(allocator::malloc(sizeof(T) * lwork));
+  }
+
+  void run(T* a, R* s, T* u, T* vt) {
+    int info;
+    gesdd<T>(
+        /* jobz = */ &jobz,
+        // M and N are swapped since lapack expects column-major.
+        /* m = */ &N,
+        /* n = */ &M,
+        /* a = */ a,
+        /* lda = */ &lda,
+        /* s = */ s,
+        // According to the identity above, lapack will write Vᵀᵀ as U.
+        /* u = */ u,
+        /* ldu = */ &ldu,
+        // According to the identity above, lapack will write Uᵀ as Vᵀ.
+        /* vt = */ vt,
+        /* ldvt = */ &ldvt,
+        /* work = */ static_cast<T*>(buffers[2].buffer.raw_ptr()),
+        /* lwork = */ &lwork,
+        /* rwork = */ static_cast<float*>(buffers[1].buffer.raw_ptr()),
+        /* iwork = */ static_cast<int*>(buffers[0].buffer.raw_ptr()),
+        /* info = */ &info);
+
+    if (info != 0) {
+      std::stringstream ss;
+      ss << "svd_impl: sgesvdx_ failed with code " << info;
+      throw std::runtime_error(ss.str());
+    }
+  }
+};
+
 template <typename T>
 void svd_impl(
     const array& a,
@@ -27,6 +204,8 @@ void svd_impl(
   const int N = a.shape(-1);
   const int K = std::min(M, N);
 
+  using R = typename SVDWork<T>::R;
+
   size_t num_matrices = a.size() / (M * N);
 
   // lapack clobbers the input, so we have to make a copy.
@@ -42,7 +221,7 @@ void svd_impl(
   encoder.set_input_array(a);
   auto in_ptr = in.data<T>();
   T* u_ptr;
-  T* s_ptr;
+  R* s_ptr;
   T* vt_ptr;
 
   if (compute_uv) {
@@ -58,7 +237,7 @@ void svd_impl(
     encoder.set_output_array(s);
     encoder.set_output_array(vt);
 
-    s_ptr = s.data<T>();
+    s_ptr = s.data<R>();
     u_ptr = u.data<T>();
     vt_ptr = vt.data<T>();
   } else {
@@ -68,98 +247,26 @@ void svd_impl(
 
     encoder.set_output_array(s);
 
-    s_ptr = s.data<T>();
+    s_ptr = s.data<R>();
     u_ptr = nullptr;
     vt_ptr = nullptr;
   }
 
   encoder.dispatch([in_ptr, u_ptr, s_ptr, vt_ptr, M, N, K, num_matrices]() {
-    // A of shape M x N. The leading dimension is N since lapack receives Aᵀ.
-    const int lda = N;
-    // U of shape M x M. (N x N in lapack).
-    const int ldu = N;
-    // Vᵀ of shape N x N. (M x M in lapack).
-    const int ldvt = M;
-
-    auto jobz = (u_ptr) ? "A" : "N";
-
-    // Will contain the number of singular values after the call has returned.
-    int ns = 0;
-    T workspace_dimension = 0;
-
-    // Will contain the indices of eigenvectors that failed to converge (not
-    // used here but required by lapack).
-    auto iwork = array::Data{allocator::malloc(sizeof(int) * 8 * K)};
-
-    static const int lwork_query = -1;
-
-    int info;
-
-    // Compute workspace size.
-    gesdd<T>(
-        /* jobz = */ jobz,
-        // M and N are swapped since lapack expects column-major.
-        /* m = */ &N,
-        /* n = */ &M,
-        /* a = */ nullptr,
-        /* lda = */ &lda,
-        /* s = */ nullptr,
-        /* u = */ nullptr,
-        /* ldu = */ &ldu,
-        /* vt = */ nullptr,
-        /* ldvt = */ &ldvt,
-        /* work = */ &workspace_dimension,
-        /* lwork = */ &lwork_query,
-        /* iwork = */ static_cast<int*>(iwork.buffer.raw_ptr()),
-        /* info = */ &info);
-
-    if (info != 0) {
-      std::stringstream ss;
-      ss << "[SVD::eval_cpu] workspace calculation failed with code " << info;
-      throw std::runtime_error(ss.str());
-    }
-
-    const int lwork = workspace_dimension;
-    auto scratch = array::Data{allocator::malloc(sizeof(T) * lwork)};
-
+    auto jobz = (u_ptr) ? 'A' : 'N';
+    SVDWork<T> svd_work(N, M, K, jobz);
     // Loop over matrices.
     for (int i = 0; i < num_matrices; i++) {
-      gesdd<T>(
-          /* jobz = */ jobz,
-          // M and N are swapped since lapack expects column-major.
-          /* m = */ &N,
-          /* n = */ &M,
-          /* a = */ in_ptr + M * N * i,
-          /* lda = */ &lda,
-          /* s = */ s_ptr + K * i,
-          // According to the identity above, lapack will write Vᵀᵀ as U.
-          /* u = */ vt_ptr ? vt_ptr + N * N * i : nullptr,
-          /* ldu = */ &ldu,
-          // According to the identity above, lapack will write Uᵀ as Vᵀ.
-          /* vt = */ u_ptr ? u_ptr + M * M * i : nullptr,
-          /* ldvt = */ &ldvt,
-          /* work = */ static_cast<T*>(scratch.buffer.raw_ptr()),
-          /* lwork = */ &lwork,
-          /* iwork = */ static_cast<int*>(iwork.buffer.raw_ptr()),
-          /* info = */ &info);
-
-      if (info != 0) {
-        std::stringstream ss;
-        ss << "svd_impl: sgesvdx_ failed with code " << info;
-        throw std::runtime_error(ss.str());
-      }
+      svd_work.run(
+          in_ptr + M * N * i,
+          s_ptr + K * i,
+          vt_ptr ? vt_ptr + N * N * i : nullptr,
+          u_ptr ? u_ptr + M * M * i : nullptr);
     }
   });
   encoder.add_temporary(in);
 }
 
-template <typename T>
-void compute_svd(
-    const array& a,
-    bool compute_uv,
-    std::vector<array>& outputs,
-    Stream stream) {}
-
 void SVD::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
@@ -170,9 +277,12 @@ void SVD::eval_cpu(
     case float64:
       svd_impl<double>(inputs[0], outputs, compute_uv_, stream());
       break;
+    case complex64:
+      svd_impl<std::complex<float>>(inputs[0], outputs, compute_uv_, stream());
+      break;
     default:
       throw std::runtime_error(
-          "[SVD::eval_cpu] only supports float32 or float64.");
+          "[SVD::eval_cpu] only supports float32, float64, or complex64.");
   }
 }
 

mlx/backend/cpu/unary.h

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

mlx/backend/cpu/unary_ops.h

@@ -108,4 +108,73 @@ struct Square {
   SINGLE()
 };
 
+template <int N>
+Simd<float, N> fp32_from_bits(Simd<uint32_t, N> x) {
+  return *(Simd<float, N>*)(&x);
+}
+template <int N>
+Simd<uint32_t, N> fp32_to_bits(Simd<float, N> x) {
+  return *(Simd<uint32_t, N>*)(&x);
+}
+
+struct ToFP8 {
+  template <typename T, int N>
+  Simd<uint8_t, N> operator()(Simd<T, N> f) {
+    uint32_t fp8_max = 543 << 21;
+    auto denorm_mask = Simd<uint32_t, N>(141 << 23);
+    Simd<uint32_t, N> f_bits;
+    Simd<float, N> f32 = f;
+    f_bits = fp32_to_bits(f32);
+    Simd<uint8_t, N> result = 0u;
+    auto sign = f_bits & 0x80000000;
+    f_bits = f_bits ^ sign;
+
+    auto f_bits_low =
+        fp32_to_bits(fp32_from_bits(f_bits) + fp32_from_bits(denorm_mask));
+    auto result_low = Simd<uint8_t, N>(f_bits_low - denorm_mask);
+
+    auto mant_odd = Simd<uint8_t, N>((f_bits >> 20) & 1);
+    auto f_bits_high = f_bits + (((uint32_t)(7 - 127) << 23) + 0x7FFFF);
+    f_bits_high = f_bits_high + Simd<uint32_t, N>(mant_odd);
+
+    auto result_high = Simd<uint8_t, N>(f_bits_high >> 20);
+    result = select(f_bits < (121 << 23), result_low, result_high);
+
+    auto result_sat = Simd<uint8_t, N>(0x7E);
+    result = select(f_bits >= fp8_max, result_sat, result);
+    return result | Simd<uint8_t, N>(sign >> 24);
+  }
+
+  template <typename T>
+  uint8_t operator()(T x) {
+    return (*this)(Simd<T, 1>(x)).value;
+  }
+};
+
+struct FromFP8 {
+  template <int N>
+  Simd<float, N> operator()(Simd<uint8_t, N> x) {
+    auto w = Simd<uint32_t, N>(x) << 24;
+    auto sign = w & 0x80000000;
+    auto nonsign = w & 0x7FFFFFFF;
+
+    auto renorm_shift = clz(nonsign);
+    renorm_shift = simd::select(
+        renorm_shift > Simd<uint32_t, N>{4},
+        renorm_shift - Simd<uint32_t, N>{4},
+        Simd<uint32_t, N>{0});
+
+    Simd<int32_t, N> inf_nan_mask =
+        (Simd<int32_t, N>(nonsign + 0x01000000) >> 8) & 0x7F800000;
+    auto zero_mask = Simd<int32_t, N>(nonsign - 1) >> 31;
+    auto result = sign |
+        ((((nonsign << renorm_shift >> 4) + ((0x78 - renorm_shift) << 23)) |
+          inf_nan_mask) &
+         ~zero_mask);
+    return fp32_from_bits(result);
+  }
+  float operator()(uint8_t x) {
+    return (*this)(Simd<uint8_t, 1>(x)).value;
+  }
+};
 } // namespace mlx::core::detail

mlx/backend/cuda/CMakeLists.txt

@@ -32,6 +32,7 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/indexing.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/kernel_utils.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/matmul.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/load.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/layer_norm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/logsumexp.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/primitives.cpp
@@ -43,6 +44,7 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/reduce/row_reduce.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rms_norm.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/rope.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/scaled_dot_product_attention.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/scaled_dot_product_attention.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/scan.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
@@ -51,12 +53,19 @@ target_sources(
           ${CMAKE_CURRENT_SOURCE_DIR}/ternary.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/utils.cpp
           ${CMAKE_CURRENT_SOURCE_DIR}/quantized/affine_quantize.cu
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/fp_quantize.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/quantized/quantized.cpp
+          ${CMAKE_CURRENT_SOURCE_DIR}/quantized/convert_fp8.cu
           ${CMAKE_CURRENT_SOURCE_DIR}/worker.cpp)
 
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/binary)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/unary)
 
+# fp4 is not available on < 12.8
+if(CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 12.8.0)
+  target_include_directories(mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/quantized/)
+endif()
+
 if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.9.0)
   target_sources(
     mlx PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/gemms/cublas_gemm_batched_12_9.cu)
@@ -114,10 +123,21 @@ if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8.0)
     mlx PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:--compress-mode=size>")
 endif()
 
-# Compute capability >= 7.0 is required for synchronization between CPU/GPU with
-# managed memory.
+# Use native CUDA arch by default.
 if(NOT DEFINED MLX_CUDA_ARCHITECTURES)
-  set(MLX_CUDA_ARCHITECTURES "native")
+  execute_process(
+    COMMAND __nvcc_device_query
+    OUTPUT_VARIABLE MLX_CUDA_ARCHITECTURES
+    OUTPUT_STRIP_TRAILING_WHITESPACE)
+  set(UPGRADABLE_ARCHITECTURES "90;100;121")
+  if(MLX_CUDA_ARCHITECTURES STREQUAL "")
+    message(
+      FATAL_ERROR
+        "Can not get native CUDA arch, must set MLX_CUDA_ARCHITECTURES")
+  elseif(MLX_CUDA_ARCHITECTURES IN_LIST UPGRADABLE_ARCHITECTURES)
+    # Use arch-specific compute capability whenever possible.
+    set(MLX_CUDA_ARCHITECTURES "${MLX_CUDA_ARCHITECTURES}a")
+  endif()
 endif()
 message(STATUS "CUDA architectures: ${MLX_CUDA_ARCHITECTURES}")
 set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
@@ -129,6 +149,7 @@ FetchContent_Declare(
   URL "https://github.com/NVIDIA/cccl/releases/download/v2.8.1/cccl-v2.8.1.zip")
 FetchContent_MakeAvailable(cccl)
 target_include_directories(mlx BEFORE PRIVATE "${cccl_SOURCE_DIR}/include")
+set_target_properties(mlx PROPERTIES CCCL_DIR "${cccl_SOURCE_DIR}/include")
 
 # Use fixed version of NVTX.
 FetchContent_Declare(
@@ -154,7 +175,7 @@ target_link_libraries(mlx PRIVATE CUDA::nvrtc CUDA::cuda_driver)
 FetchContent_Declare(
   cudnn
   GIT_REPOSITORY https://github.com/NVIDIA/cudnn-frontend.git
-  GIT_TAG v1.14.0
+  GIT_TAG v1.16.0
   GIT_SHALLOW TRUE
   EXCLUDE_FROM_ALL)
 set(CUDNN_FRONTEND_SKIP_JSON_LIB ON)
@@ -170,11 +191,6 @@ target_link_libraries(mlx PRIVATE CUDNN::cudnn_all)
 # Suppress nvcc warnings on MLX headers.
 target_compile_options(mlx PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcudafe
                                    --diag_suppress=997>)
-# 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)
-
 # Install CCCL headers for JIT.
 install(DIRECTORY ${cccl_SOURCE_DIR}/include/cuda
         DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/cccl)

mlx/backend/cuda/allocator.cpp

@@ -1,6 +1,7 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/backend/cuda/allocator.h"
+#include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/utils.h"
 #include "mlx/utils.h"
 
@@ -67,6 +68,7 @@ CudaBuffer* SmallSizePool::malloc() {
   next_free_ = next_free_->next;
   b->buf.data = static_cast<char*>(data_) + i * small_block_size;
   b->buf.size = small_block_size;
+  b->buf.device = -1;
   return &b->buf;
 }
 
@@ -88,16 +90,42 @@ CudaAllocator::CudaAllocator()
           page_size,
           [](CudaBuffer* buf) { return buf->size; },
           [this](CudaBuffer* buf) { cuda_free(buf); }) {
-  // TODO: Set memory limit for multi-device.
   size_t free, total;
   CHECK_CUDA_ERROR(cudaMemGetInfo(&free, &total));
-  memory_limit_ = total * 0.95;
+  memory_limit_ = total * 0.9;
   max_pool_size_ = memory_limit_;
+
+  int device_count = 0;
+  CHECK_CUDA_ERROR(cudaGetDeviceCount(&device_count));
+  int curr;
+  CHECK_CUDA_ERROR(cudaGetDevice(&curr));
+  for (int i = 0; i < device_count; ++i) {
+    CHECK_CUDA_ERROR(cudaSetDevice(i));
+    cudaStream_t s;
+    CHECK_CUDA_ERROR(cudaStreamCreateWithFlags(&s, cudaStreamNonBlocking));
+    free_streams_.push_back(s);
+  }
+  CHECK_CUDA_ERROR(cudaSetDevice(curr));
 }
 
-Buffer CudaAllocator::malloc(size_t size) {
+void copy_to_managed(CudaBuffer& buf) {
+  // TODO maybe make this async on a i/o stream to avoid synchronizing the
+  // device on malloc/and free
+  void* new_data;
+  CHECK_CUDA_ERROR(cudaMallocManaged(&new_data, buf.size));
+  buf.device = -1;
+  CHECK_CUDA_ERROR(cudaMemcpy(new_data, buf.data, buf.size, cudaMemcpyDefault));
+  CHECK_CUDA_ERROR(cudaFree(buf.data));
+  buf.data = new_data;
+}
+
+Buffer
+CudaAllocator::malloc_async(size_t size, int device, cudaStream_t stream) {
+  if (size == 0) {
+    return Buffer{new CudaBuffer{nullptr, 0, -1}};
+  }
+
   // Find available buffer from cache.
-  auto orig_size = size;
   std::unique_lock lock(mutex_);
   if (size <= small_block_size) {
     size = 8;
@@ -107,6 +135,10 @@ Buffer CudaAllocator::malloc(size_t size) {
     size = page_size * ((size + page_size - 1) / page_size);
   }
 
+  if (size <= small_block_size || stream == nullptr) {
+    device = -1;
+  }
+
   CudaBuffer* buf = buffer_cache_.reuse_from_cache(size);
   if (!buf) {
     // If we have a lot of memory pressure try to reclaim memory from the cache.
@@ -122,30 +154,51 @@ Buffer CudaAllocator::malloc(size_t size) {
     }
     lock.unlock();
     if (!buf) {
-      buf = new CudaBuffer{nullptr, size};
-      cudaError_t err = cudaMallocManaged(&buf->data, size);
+      cudaError_t err;
+      void* data = nullptr;
+      if (device == -1) {
+        err = cudaMallocManaged(&data, size);
+      } else {
+        err = cudaMallocAsync(&data, size, stream);
+      }
       if (err != cudaSuccess && err != cudaErrorMemoryAllocation) {
         throw std::runtime_error(fmt::format(
             "cudaMallocManaged failed: {}.", cudaGetErrorString(err)));
       }
+      if (!data) {
+        return Buffer{nullptr};
+      }
+      buf = new CudaBuffer{data, size, device};
     }
     lock.lock();
   }
-  active_memory_ += size;
+  active_memory_ += buf->size;
   peak_memory_ = std::max(active_memory_, peak_memory_);
 
   // Maintain the cache below the requested limit.
   if (get_cache_memory() > max_pool_size_) {
     buffer_cache_.release_cached_buffers(get_cache_memory() - max_pool_size_);
   }
+  // Copy to managed here if the buffer is not on the right device
+  if (buf->device >= 0 && buf->device != device) {
+    copy_to_managed(*buf);
+  }
   return Buffer{buf};
 }
 
+Buffer CudaAllocator::malloc(size_t size) {
+  return malloc_async(size, -1, nullptr);
+}
+
 void CudaAllocator::free(Buffer buffer) {
   auto* buf = static_cast<CudaBuffer*>(buffer.ptr());
   if (!buf) {
     return;
   }
+  if (buf->size == 0) {
+    delete buf;
+    return;
+  }
 
   std::unique_lock lock(mutex_);
   active_memory_ -= buf->size;
@@ -169,7 +222,11 @@ void CudaAllocator::cuda_free(CudaBuffer* buf) {
   if (scalar_pool_.in_pool(buf)) {
     scalar_pool_.free(buf);
   } else {
-    cudaFree(buf->data);
+    if (buf->device >= 0) {
+      CHECK_CUDA_ERROR(cudaFreeAsync(buf->data, free_streams_[buf->device]));
+    } else {
+      CHECK_CUDA_ERROR(cudaFree(buf->data));
+    }
     delete buf;
   }
 }
@@ -220,6 +277,17 @@ CudaAllocator& allocator() {
   return *allocator_;
 }
 
+Buffer malloc_async(size_t size, CommandEncoder& encoder) {
+  auto buffer = allocator().malloc_async(
+      size, encoder.device().cuda_device(), encoder.stream());
+  if (size && !buffer.ptr()) {
+    std::ostringstream msg;
+    msg << "[malloc_async] Unable to allocate " << size << " bytes.";
+    throw std::runtime_error(msg.str());
+  }
+  return buffer;
+}
+
 } // namespace cu
 
 namespace allocator {
@@ -232,7 +300,11 @@ void* Buffer::raw_ptr() {
   if (!ptr_) {
     return nullptr;
   }
-  return static_cast<cu::CudaBuffer*>(ptr_)->data;
+  auto& cbuf = *static_cast<cu::CudaBuffer*>(ptr_);
+  if (cbuf.device != -1) {
+    copy_to_managed(cbuf);
+  }
+  return cbuf.data;
 }
 
 } // namespace allocator

mlx/backend/cuda/allocator.h

@@ -4,19 +4,24 @@
 
 #include "mlx/allocator.h"
 #include "mlx/backend/common/buffer_cache.h"
+#include "mlx/backend/cuda/cuda_utils.h"
 
+#include <cuda_runtime.h>
 #include <mutex>
 #include <set>
 #include <utility>
 
 namespace mlx::core::cu {
 
+class CommandEncoder;
+
 using allocator::Buffer;
 
 // Stores cuda-managed unified memory.
 struct CudaBuffer {
   void* data;
   size_t size;
+  int device; // -1 for managed
 };
 
 class SmallSizePool {
@@ -45,6 +50,7 @@ class SmallSizePool {
 class CudaAllocator : public allocator::Allocator {
  public:
   Buffer malloc(size_t size) override;
+  Buffer malloc_async(size_t size, int device, cudaStream_t stream);
   void free(Buffer buffer) override;
   size_t size(Buffer buffer) const override;
 
@@ -69,9 +75,12 @@ class CudaAllocator : public allocator::Allocator {
   BufferCache<CudaBuffer> buffer_cache_;
   size_t active_memory_{0};
   size_t peak_memory_{0};
+  std::vector<cudaStream_t> free_streams_;
   SmallSizePool scalar_pool_;
 };
 
 CudaAllocator& allocator();
 
+Buffer malloc_async(size_t size, CommandEncoder& encoder);
+
 } // namespace mlx::core::cu

mlx/backend/cuda/arange.cu

@@ -41,9 +41,8 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
   if (out.size() == 0) {
     return;
   }
-  out.set_data(allocator::malloc(out.nbytes()));
-
   auto& encoder = cu::get_command_encoder(stream());
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   encoder.set_output_array(out);
 
   dispatch_int_float_types(out.dtype(), "Arange", [&](auto type_tag) {
@@ -58,7 +57,7 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
           num_blocks,
           block_dims,
           0,
-          out.data<OutType>(),
+          gpu_ptr<OutType>(out),
           out.data_size(),
           static_cast<CTYPE>(start_),
           static_cast<CTYPE>(start_ + step_) - static_cast<CTYPE>(start_));

mlx/backend/cuda/arg_reduce.cu

@@ -140,8 +140,10 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
   nvtx3::scoped_range r("ArgReduce::eval_gpu");
   assert(inputs.size() == 1);
   auto& in = inputs[0];
-  out.set_data(allocator::malloc(out.nbytes()));
+
   auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
 
   // Prepare the shapes, strides and axis arguments.
   Shape shape = remove_index(in.shape(), axis_);
@@ -154,7 +156,6 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
   int32_t ndim = shape.size();
 
   // ArgReduce.
-  auto& encoder = cu::get_command_encoder(s);
   encoder.set_input_array(in);
   encoder.set_output_array(out);
   dispatch_real_types(in.dtype(), "ArgReduce", [&](auto type_tag) {
@@ -172,8 +173,8 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
           num_blocks,
           block_dim(),
           0,
-          in.data<T>(),
-          out.data<uint32_t>(),
+          gpu_ptr<T>(in),
+          gpu_ptr<uint32_t>(out),
           out.size(),
           const_param(shape),
           const_param(in_strides),

mlx/backend/cuda/binary/binary.cuh

@@ -292,9 +292,9 @@ void binary_op_gpu_inplace(
                         {num_blocks_x, num_blocks_y},
                         block_dims,
                         0,
-                        a.data<InType>(),
-                        b.data<InType>(),
-                        out.data<OutType>(),
+                        gpu_ptr<InType>(a),
+                        gpu_ptr<InType>(b),
+                        gpu_ptr<OutType>(out),
                         rest,
                         const_param<dims_constant()>(shape),
                         const_param<dims_constant()>(a_strides),
@@ -310,9 +310,9 @@ void binary_op_gpu_inplace(
                       {num_blocks_x, num_blocks_y},
                       block_dims,
                       0,
-                      a.data<InType>(),
-                      b.data<InType>(),
-                      out.data<OutType>(),
+                      gpu_ptr<InType>(a),
+                      gpu_ptr<InType>(b),
+                      gpu_ptr<OutType>(out),
                       rest,
                       const_param(shape),
                       const_param(a_strides),
@@ -339,9 +339,9 @@ void binary_op_gpu_inplace(
                 num_blocks,
                 block_dims,
                 0,
-                a.data<InType>(),
-                b.data<InType>(),
-                out.data<OutType>(),
+                gpu_ptr<InType>(a),
+                gpu_ptr<InType>(b),
+                gpu_ptr<OutType>(out),
                 out.data_size());
           });
         }
@@ -365,7 +365,10 @@ void binary_op_gpu(
   auto& a = inputs[0];
   auto& b = inputs[1];
   auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, out, bopt);
+  auto& encoder = cu::get_command_encoder(s);
+
+  set_binary_op_output_data(
+      a, b, out, bopt, [&](auto n) { return cu::malloc_async(n, encoder); });
   binary_op_gpu_inplace<Op>(inputs, out, op, s);
 }
 

mlx/backend/cuda/binary_two.cu

@@ -245,14 +245,16 @@ void binary_two_op_gpu_inplace(
   auto& out_a = outputs[0];
   auto& out_b = outputs[1];
   auto bopt = get_binary_op_type(a, b);
-  set_binary_op_output_data(a, b, out_a, bopt);
-  set_binary_op_output_data(a, b, out_b, bopt);
+  auto& encoder = cu::get_command_encoder(s);
+  set_binary_op_output_data(
+      a, b, out_a, bopt, [&](auto n) { return cu::malloc_async(n, encoder); });
+  set_binary_op_output_data(
+      a, b, out_b, bopt, [&](auto n) { return cu::malloc_async(n, encoder); });
 
   if (out_a.size() == 0) {
     return;
   }
 
-  auto& encoder = cu::get_command_encoder(s);
   encoder.set_input_array(a);
   encoder.set_input_array(b);
   encoder.set_output_array(out_a);
@@ -313,10 +315,10 @@ void binary_two_op_gpu_inplace(
                         {num_blocks_x, num_blocks_y},
                         block_dims,
                         0,
-                        a.data<InType>(),
-                        b.data<InType>(),
-                        out_a.data<OutType>(),
-                        out_b.data<OutType>(),
+                        gpu_ptr<InType>(a),
+                        gpu_ptr<InType>(b),
+                        gpu_ptr<OutType>(out_a),
+                        gpu_ptr<OutType>(out_b),
                         rest,
                         const_param<dims_constant()>(shape),
                         const_param<dims_constant()>(a_strides),
@@ -332,10 +334,10 @@ void binary_two_op_gpu_inplace(
                       {num_blocks_x, num_blocks_y},
                       block_dims,
                       0,
-                      a.data<InType>(),
-                      b.data<InType>(),
-                      out_a.data<OutType>(),
-                      out_b.data<OutType>(),
+                      gpu_ptr<InType>(a),
+                      gpu_ptr<InType>(b),
+                      gpu_ptr<OutType>(out_a),
+                      gpu_ptr<OutType>(out_b),
                       rest,
                       const_param(shape),
                       const_param(a_strides),
@@ -366,10 +368,10 @@ void binary_two_op_gpu_inplace(
                 num_blocks,
                 block_dims,
                 0,
-                a.data<InType>(),
-                b.data<InType>(),
-                out_a.data<OutType>(),
-                out_b.data<OutType>(),
+                gpu_ptr<InType>(a),
+                gpu_ptr<InType>(b),
+                gpu_ptr<OutType>(out_a),
+                gpu_ptr<OutType>(out_b),
                 out_a.data_size());
           });
         }

mlx/backend/cuda/compiled.cpp

@@ -293,8 +293,13 @@ void Compiled::eval_gpu(
     }
   }
 
+  auto& encoder = cu::get_command_encoder(s);
+
   // Put outputs.
-  compiled_allocate_outputs(inputs, outputs, is_constant_, contiguous);
+  compiled_allocate_outputs(
+      inputs, outputs, is_constant_, contiguous, [&](auto n) {
+        return cu::malloc_async(n, encoder);
+      });
   for (auto& x : outputs) {
     args.append(x);
   }
@@ -324,7 +329,6 @@ void Compiled::eval_gpu(
     kernel_name += fmt::format(
         "_strided<{}, {}, {}>", shape.size(), index_type, work_per_thread);
   }
-  auto& encoder = cu::get_command_encoder(s);
   for (const auto& in : inputs) {
     encoder.set_input_array(in);
   }

mlx/backend/cuda/conv.cpp

@@ -15,19 +15,16 @@ namespace mlx::core {
 
 namespace {
 
-// Alias for better readability.
-#define CONV_FORWARD CUDNN_BACKEND_OPERATION_CONVOLUTION_FORWARD_DESCRIPTOR
-#define CONV_BACKWARD_INPUT \
-  CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_DATA_DESCRIPTOR
-#define CONV_BACKWARD_WEIGHT \
-  CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_FILTER_DESCRIPTOR
-
-// Custom placeholder representing fallback kernel.
-#define CONV_FALLBACK static_cast<cudnnBackendDescriptorType_t>(-1)
+enum ConvBackendType {
+  CONV_FALLBACK,
+  CONV_FORWARD,
+  CONV_BACKWARD_INPUT,
+  CONV_BACKWARD_WEIGHT,
+};
 
 struct ConvCacheKey {
   int device_id;
-  cudnnDataType_t cudnn_dtype;
+  fe::DataType_t cudnn_dtype;
   std::array<int, MAX_NDIM> input_shape;
   std::array<int, MAX_NDIM> weight_shape;
   std::array<int, MAX_NDIM> stride;
@@ -44,15 +41,13 @@ struct ConvCacheKey {
 auto& conv_cache() {
   static LRUBytesKeyCache<
       ConvCacheKey,
-      std::pair<
-          cudnnBackendDescriptorType_t,
-          std::optional<cudnn_frontend::ExecutionPlan>>>
+      std::pair<ConvBackendType, std::optional<DnnGraph>>>
       cache("MLX_CUDA_CONV_CACHE_SIZE", /* default_capacity */ 128);
   return cache;
 }
 
-auto get_conv_op_settings(
-    cudnnBackendDescriptorType_t backend_type,
+auto get_conv_settings(
+    ConvBackendType backend_type,
     array& x,
     array& w,
     array& y,
@@ -68,8 +63,8 @@ auto get_conv_op_settings(
     for (int i = 0; i < padding_lo.size(); ++i) {
       int wt_size = 1 + kernel_dilation[i] * (w.shape(1 + i) - 1);
       padding_lo[i] = wt_size - padding_lo[i] - 1;
-      int in_size = 1 + kernel_strides[i] * (x.shape(1 + i) - 1);
-      int out_size = 1 + input_dilation[i] * (y.shape(1 + i) - 1);
+      int in_size = 1 + kernel_strides[i] * (y.shape(1 + i) - 1);
+      int out_size = 1 + input_dilation[i] * (x.shape(1 + i) - 1);
       padding_hi[i] = out_size - in_size + padding_hi[i];
     }
     return std::make_tuple(
@@ -95,49 +90,57 @@ auto get_conv_op_settings(
   }
 }
 
-std::optional<cudnn_frontend::OperationGraph> build_conv_op_graph(
+std::optional<DnnGraph> build_conv_graph(
     cu::CommandEncoder& encoder,
-    cudnnBackendDescriptorType_t backend_type,
+    ConvBackendType backend_type,
     Dtype dtype,
     array& x,
     array& w,
     array& y,
-    const SmallVector<int64_t>& stride,
-    const SmallVector<int64_t>& padding_lo,
-    const SmallVector<int64_t>& padding_hi,
-    const SmallVector<int64_t>& dilation) {
-  try {
-    auto compute_dtype = (dtype == float16 || dtype == bfloat16)
-        ? CUDNN_DATA_FLOAT
-        : dtype_to_cudnn_type(dtype);
-    auto conv_desc = cudnn_frontend::ConvDescBuilder()
-                         .setDataType(compute_dtype)
-                         .setMathMode(CUDNN_CROSS_CORRELATION)
-                         .setNDims(stride.size())
-                         .setStrides(stride.size(), stride.data())
-                         .setPrePadding(padding_lo.size(), padding_lo.data())
-                         .setPostPadding(padding_hi.size(), padding_hi.data())
-                         .setDilation(dilation.size(), dilation.data())
-                         .build();
+    const std::vector<int64_t>& stride,
+    const std::vector<int64_t>& padding_lo,
+    const std::vector<int64_t>& padding_hi,
+    const std::vector<int64_t>& dilation) {
+  auto compute_dtype =
+      (dtype == float16 || dtype == bfloat16) ? float32 : dtype;
+  DnnGraph graph(encoder.device().cudnn_handle(), dtype, compute_dtype);
+  auto x_ = graph.tensor_nchw("X", 'x', x);
+  auto w_ = graph.tensor_nchw("W", 'w', w);
 
-    auto op = cudnn_frontend::OperationBuilder(backend_type)
-                  .setxDesc(build_cudnn_tensor_nchw('x', x))
-                  .setwDesc(build_cudnn_tensor_nchw('w', w))
-                  .setyDesc(build_cudnn_tensor_nchw('y', y))
-                  .setcDesc(conv_desc)
-                  .build();
+  auto set_options = [&](auto& options) {
+    options.set_compute_data_type(dtype_to_cudnn_type(compute_dtype))
+        .set_convolution_mode(fe::ConvolutionMode_t::CROSS_CORRELATION)
+        .set_stride(stride)
+        .set_pre_padding(padding_lo)
+        .set_post_padding(padding_hi)
+        .set_dilation(dilation);
+  };
 
-    std::array<cudnn_frontend::Operation const*, 1> ops = {&op};
-    return cudnn_frontend::OperationGraphBuilder()
-        .setHandle(encoder.device().cudnn_handle())
-        .setOperationGraph(ops.size(), ops.data())
-        .build();
-  } catch (cudnn_frontend::cudnnException& error) {
-    if (error.getCudnnStatus() != CUDNN_STATUS_BAD_PARAM) {
-      throw;
-    }
+  std::shared_ptr<fe::graph::Tensor_attributes> y_;
+  if (backend_type == CONV_FORWARD) {
+    auto options = fe::graph::Conv_fprop_attributes();
+    set_options(options);
+    y_ = graph.conv_fprop(x_, w_, options);
+  } else if (backend_type == CONV_BACKWARD_INPUT) {
+    auto options = fe::graph::Conv_dgrad_attributes();
+    set_options(options);
+    y_ = graph.conv_dgrad(x_, w_, options);
+  } else if (backend_type == CONV_BACKWARD_WEIGHT) {
+    auto options = fe::graph::Conv_wgrad_attributes();
+    set_options(options);
+    y_ = graph.conv_wgrad(w_, x_, options);
+  }
+  graph.tensor_nchw(y_, 'y', y)->set_output(true);
+
+  if (graph.prepare().is_bad()) {
     return std::nullopt;
   }
+  graph.deselect_numeric_notes({fe::NumericalNote_t::DOWN_CONVERT_INPUTS});
+  if (dtype == float32 && !env::enable_tf32()) {
+    graph.deselect_numeric_notes({fe::NumericalNote_t::TENSOR_CORE});
+  }
+  CHECK_CUDNN_FE_ERROR(graph.build());
+  return graph;
 }
 
 // Transpose from (C_out, H, W, C_in / groups) to (C_in, H, W, C_out / groups).
@@ -181,7 +184,7 @@ array group_transpose(
 // eval_gpu, with cost of possible redundant copies.
 std::tuple<array, array, array> prepare_args(
     cu::CommandEncoder& encoder,
-    cudnnBackendDescriptorType_t backend_type,
+    ConvBackendType backend_type,
     array in,
     array wt,
     array out,
@@ -221,27 +224,11 @@ std::tuple<array, array, array> prepare_args(
   return {std::move(in), std::move(wt), std::move(out)};
 }
 
-// Get the x/w/y args from the in/wt/out args depending on backend type.
-inline std::tuple<array&, array&, array&> dispatch_args(
-    cudnnBackendDescriptorType_t backend_type,
-    array& in,
-    array& wt,
-    array& out) {
-  switch (backend_type) {
-    case CONV_BACKWARD_INPUT:
-      return {out, wt, in};
-    case CONV_BACKWARD_WEIGHT:
-      return {in, out, wt};
-    default:
-      return {in, wt, out};
-  }
-}
-
 // Register inputs and outputs before actually running conv op. Can only be
 // called once per eval_gpu.
 void register_args(
     cu::CommandEncoder& encoder,
-    cudnnBackendDescriptorType_t backend_type,
+    ConvBackendType backend_type,
     array& in,
     array& wt,
     array& intermediate_out,
@@ -270,19 +257,19 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   if (out_.size() == 0) {
     return;
   }
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
 
   assert(inputs.size() == 2);
   array in = inputs[0];
   array wt = inputs[1];
   array out = out_;
-  out.set_data(allocator::malloc(out.nbytes()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   Dtype dtype = out.dtype();
 
-  auto& s = stream();
-  auto& encoder = cu::get_command_encoder(s);
-
   // Search cache.
-  ConvCacheKey cache_key{
+  BytesKey<ConvCacheKey> cache_key;
+  cache_key.pod = {
       encoder.device().cuda_device(),
       dtype_to_cudnn_type(dtype),
       vector_key(in.shape()),
@@ -297,16 +284,19 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
       get_alignment(wt),
       get_alignment(out)};
   if (auto it = conv_cache().find(cache_key); it != conv_cache().end()) {
-    auto& [backend_type, plan] = it->second;
-    if (plan) {
-      // Run cached plan.
+    auto& [backend_type, graph] = it->second;
+    if (graph) {
+      // Run cached graph.
       std::tie(in, wt, out) =
           prepare_args(encoder, backend_type, in, wt, out, groups_, s);
       register_args(encoder, backend_type, in, wt, out, out_);
-      auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-      if (!encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
-        throw std::runtime_error("[conv] Cached plan failed to execute.");
-      }
+      CHECK_CUDNN_FE_ERROR(graph->encode_capturing(
+          encoder,
+          {
+              {'x', gpu_ptr<void>(in)},
+              {'w', gpu_ptr<void>(wt)},
+              {'y', gpu_ptr<void>(out)},
+          }));
     } else {
       // Run fallback kernel.
       gemm_conv(
@@ -327,7 +317,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
 
   // There is no reliable way to deduce the proper cuDNN backend for the
   // convolution, so we make a best guess and then try.
-  SmallVector<cudnnBackendDescriptorType_t, 2> try_backends;
+  SmallVector<ConvBackendType, 2> try_backends;
   if (flip_) {
     // When weight is flipped, we assume it is backward input convolution.
     try_backends.push_back(CONV_BACKWARD_INPUT);
@@ -345,13 +335,12 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   }
 
   // Try to build op graph.
-  cudnnBackendDescriptorType_t backend_type;
-  std::optional<cudnn_frontend::OperationGraph> op_graph;
+  ConvBackendType backend_type;
+  std::optional<DnnGraph> graph;
   for (auto try_backend : try_backends) {
-    auto [in_copy, wt_copy, out_copy] =
+    auto [x, w, y] =
         prepare_args(encoder, try_backend, in, wt, out, groups_, s);
-    auto [x, w, y] = dispatch_args(try_backend, in_copy, wt_copy, out_copy);
-    auto [stride, padding_lo, padding_hi, dilation] = get_conv_op_settings(
+    auto [stride, padding_lo, padding_hi, dilation] = get_conv_settings(
         try_backend,
         x,
         w,
@@ -361,7 +350,7 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
         padding_hi_,
         kernel_dilation_,
         input_dilation_);
-    op_graph = build_conv_op_graph(
+    graph = build_conv_graph(
         encoder,
         try_backend,
         dtype,
@@ -372,30 +361,27 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
         padding_lo,
         padding_hi,
         dilation);
-    if (op_graph) {
+    if (graph) {
       backend_type = try_backend;
-      in = std::move(in_copy);
-      wt = std::move(wt_copy);
-      out = std::move(out_copy);
+      in = std::move(x);
+      wt = std::move(w);
+      out = std::move(y);
       break;
     }
   }
 
-  if (op_graph) {
-    // Find a plan for the graph and execute it.
-    auto plan = find_cudnn_plan_from_op_graph(
-        encoder.device().cudnn_handle(), backend_type, dtype, *op_graph);
-    if (plan) {
-      // Setup inputs and outputs.
-      register_args(encoder, backend_type, in, wt, out, out_);
-
-      auto [x, w, y] = dispatch_args(backend_type, in, wt, out);
-      if (encode_cudnn_plan(encoder, *plan, {'x', 'w', 'y'}, x, w, y)) {
-        conv_cache().emplace(
-            cache_key, std::make_pair(backend_type, std::move(*plan)));
-        return;
-      }
-    }
+  if (graph) {
+    register_args(encoder, backend_type, in, wt, out, out_);
+    CHECK_CUDNN_FE_ERROR(graph->encode_capturing(
+        encoder,
+        {
+            {'x', gpu_ptr<void>(in)},
+            {'w', gpu_ptr<void>(wt)},
+            {'y', gpu_ptr<void>(out)},
+        }));
+    conv_cache().emplace(
+        cache_key, std::make_pair(backend_type, std::move(*graph)));
+    return;
   }
 
   // Use fallback kernel for settings not supported by cuDNN.

mlx/backend/cuda/conv/gemm_conv.cu

@@ -86,7 +86,7 @@ array unfold_inputs_nd(
     int mat_N,
     ConvParams<NDIM>& params) {
   array unfolded({mat_M, mat_K}, in.dtype(), nullptr, {});
-  unfolded.set_data(allocator::malloc(unfolded.nbytes()));
+  unfolded.set_data(cu::malloc_async(unfolded.nbytes(), encoder));
   encoder.add_temporary(unfolded);
 
   int filter_size = params.C;
@@ -118,8 +118,8 @@ array unfold_inputs_nd(
         num_blocks,
         block_dims,
         0,
-        in.data<DataType>(),
-        unfolded.data<DataType>(),
+        gpu_ptr<DataType>(in),
+        gpu_ptr<DataType>(unfolded),
         filter_size,
         out_pixels,
         params);

mlx/backend/cuda/conv/gemm_grouped_conv.cu

@@ -89,7 +89,7 @@ array grouped_unfold_transpose_inputs_nd(
     int mat_N,
     ConvParams<NDIM>& params) {
   array unfolded({mat_M, mat_K * params.groups}, in.dtype(), nullptr, {});
-  unfolded.set_data(allocator::malloc(unfolded.nbytes()));
+  unfolded.set_data(cu::malloc_async(unfolded.nbytes(), encoder));
   encoder.add_temporary(unfolded);
 
   int filter_size = params.C;
@@ -121,8 +121,8 @@ array grouped_unfold_transpose_inputs_nd(
         num_blocks,
         block_dims,
         0,
-        in.data<DataType>(),
-        unfolded.data<DataType>(),
+        gpu_ptr<DataType>(in),
+        gpu_ptr<DataType>(unfolded),
         filter_size,
         out_pixels,
         params);

mlx/backend/cuda/copy.cu

@@ -5,6 +5,21 @@
 
 namespace mlx::core {
 
+void copy_gpu(const array& in, array& out, CopyType ctype, const Stream& s) {
+  auto& encoder = cu::get_command_encoder(s);
+  bool donated = set_copy_output_data(
+      in, out, ctype, [&](auto n) { return cu::malloc_async(n, encoder); });
+  if (donated && in.dtype() == out.dtype()) {
+    // If the output has the same type as the input then there is nothing to
+    // copy, just use the buffer.
+    return;
+  }
+  if (ctype == CopyType::GeneralGeneral) {
+    ctype = CopyType::General;
+  }
+  copy_gpu_inplace(in, out, ctype, s);
+}
+
 void copy_gpu_inplace(
     const array& in,
     array& out,
@@ -87,11 +102,31 @@ void fill_gpu(const array& in, array& out, const Stream& s) {
   if (out.size() == 0) {
     return;
   }
-  out.set_data(allocator::malloc(out.nbytes()));
   auto& encoder = cu::get_command_encoder(s);
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   encoder.set_input_array(in);
   encoder.set_output_array(out);
   copy_contiguous(encoder, CopyType::Scalar, in, out, 0, 0);
 }
 
+void reshape_gpu(const array& in, array& out, Stream s) {
+  auto [copy_necessary, out_strides] = prepare_reshape(in, out);
+  if (copy_necessary) {
+    auto& encoder = cu::get_command_encoder(s);
+    out.set_data(cu::malloc_async(out.nbytes(), encoder));
+    copy_gpu_inplace(
+        in,
+        out,
+        in.shape(),
+        in.strides(),
+        make_contiguous_strides(in.shape()),
+        0,
+        0,
+        CopyType::General,
+        s);
+  } else {
+    shared_buffer_reshape(in, out_strides, out);
+  }
+}
+
 } // namespace mlx::core

mlx/backend/cuda/copy/copy_contiguous.cu

@@ -77,8 +77,8 @@ void copy_contiguous(
             num_blocks,
             block_dims,
             0,
-            in.data<InType>() + in_offset,
-            out.data<OutType>() + out_offset,
+            gpu_ptr<InType>(in) + in_offset,
+            gpu_ptr<OutType>(out) + out_offset,
             out.data_size());
       });
     });

mlx/backend/cuda/copy/copy_general.cu

@@ -106,8 +106,8 @@ void copy_general(
             using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
             using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
             using IdxT = std::conditional_t<large(), int64_t, int32_t>;
-            const InType* in_ptr = in.data<InType>() + offset_in;
-            OutType* out_ptr = out.data<OutType>() + offset_out;
+            const InType* in_ptr = gpu_ptr<InType>(in) + offset_in;
+            OutType* out_ptr = gpu_ptr<OutType>(out) + offset_out;
             int ndim = shape.size();
             size_t data_size = 1;
             for (auto& s : shape)

mlx/backend/cuda/copy/copy_general_dynamic.cu

@@ -69,8 +69,8 @@ void copy_general_dynamic(
             using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
             using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
             using IdxT = std::conditional_t<large(), int64_t, int32_t>;
-            const InType* in_ptr = in.data<InType>() + offset_in;
-            OutType* out_ptr = out.data<OutType>() + offset_out;
+            const InType* in_ptr = gpu_ptr<InType>(in) + offset_in;
+            OutType* out_ptr = gpu_ptr<OutType>(out) + offset_out;
             int ndim = shape.size();
             if (ndim <= 3) {
               dispatch_1_2_3(ndim, [&](auto dims_constant) {
@@ -90,8 +90,8 @@ void copy_general_dynamic(
                     const_param<dims_constant()>(shape),
                     const_param<dims_constant()>(strides_in),
                     const_param<dims_constant()>(strides_out),
-                    dynamic_offset_in.data<int64_t>(),
-                    dynamic_offset_out.data<int64_t>());
+                    gpu_ptr<int64_t>(dynamic_offset_in),
+                    gpu_ptr<int64_t>(dynamic_offset_out));
               });
             } else { // ndim >= 4
               auto [num_blocks, block_dims] = get_launch_args(out, large());
@@ -107,8 +107,8 @@ void copy_general_dynamic(
                   const_param(strides_in),
                   const_param(strides_out),
                   ndim,
-                  dynamic_offset_in.data<int64_t>(),
-                  dynamic_offset_out.data<int64_t>());
+                  gpu_ptr<int64_t>(dynamic_offset_in),
+                  gpu_ptr<int64_t>(dynamic_offset_out));
             }
           });
     });

mlx/backend/cuda/copy/copy_general_input.cu

@@ -92,8 +92,8 @@ void copy_general_input(
             using InType = cuda_type_t<MLX_GET_TYPE(in_type_tag)>;
             using OutType = cuda_type_t<MLX_GET_TYPE(out_type_tag)>;
             using IdxT = std::conditional_t<large(), int64_t, int32_t>;
-            const InType* in_ptr = in.data<InType>() + offset_in;
-            OutType* out_ptr = out.data<OutType>() + offset_out;
+            const InType* in_ptr = gpu_ptr<InType>(in) + offset_in;
+            OutType* out_ptr = gpu_ptr<OutType>(out) + offset_out;
             int ndim = shape.size();
             int work_per_thread = 1;
             auto dim0 = ndim > 0 ? shape.back() : 1;

mlx/backend/cuda/cuda_utils.h

@@ -0,0 +1,89 @@
+// Copyright © 2025 Apple Inc.
+
+#pragma once
+
+#include <cublasLt.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <cudnn.h>
+
+namespace mlx::core {
+
+// Throw exception if the cuda API does not succeed.
+void check_cublas_error(const char* name, cublasStatus_t err);
+void check_cuda_error(const char* name, cudaError_t err);
+void check_cuda_error(const char* name, CUresult err);
+void check_cudnn_error(const char* name, cudnnStatus_t err);
+
+// The macro version that prints the command that failed.
+#define CHECK_CUBLAS_ERROR(cmd) check_cublas_error(#cmd, (cmd))
+#define CHECK_CUDA_ERROR(cmd) check_cuda_error(#cmd, (cmd))
+#define CHECK_CUDNN_ERROR(cmd) check_cudnn_error(#cmd, (cmd))
+
+// Base class for RAII managed CUDA resources.
+template <typename Handle, cudaError_t (*Destroy)(Handle)>
+class CudaHandle {
+ public:
+  CudaHandle(Handle handle = nullptr) : handle_(handle) {}
+
+  CudaHandle(CudaHandle&& other) : handle_(other.handle_) {
+    assert(this != &other);
+    other.handle_ = nullptr;
+  }
+
+  ~CudaHandle() {
+    // Skip if there was an error to avoid throwing in the destructors
+    if (cudaPeekAtLastError() != cudaSuccess) {
+      return;
+    }
+    reset();
+  }
+
+  CudaHandle(const CudaHandle&) = delete;
+  CudaHandle& operator=(const CudaHandle&) = delete;
+
+  CudaHandle& operator=(CudaHandle&& other) {
+    assert(this != &other);
+    reset();
+    std::swap(handle_, other.handle_);
+    return *this;
+  }
+
+  void reset() {
+    if (handle_ != nullptr) {
+      CHECK_CUDA_ERROR(Destroy(handle_));
+      handle_ = nullptr;
+    }
+  }
+
+  operator Handle() const {
+    return handle_;
+  }
+
+ protected:
+  Handle handle_;
+};
+
+namespace cu {
+class Device;
+}; // namespace cu
+
+// Wrappers of CUDA resources.
+class CudaGraph : public CudaHandle<cudaGraph_t, cudaGraphDestroy> {
+ public:
+  using CudaHandle::CudaHandle;
+  explicit CudaGraph(cu::Device& device);
+  void end_capture(cudaStream_t stream);
+};
+
+class CudaGraphExec : public CudaHandle<cudaGraphExec_t, cudaGraphExecDestroy> {
+ public:
+  void instantiate(cudaGraph_t graph);
+};
+
+class CudaStream : public CudaHandle<cudaStream_t, cudaStreamDestroy> {
+ public:
+  explicit CudaStream(cu::Device& device);
+};
+
+} // namespace mlx::core

mlx/backend/cuda/cudnn_utils.cpp

@@ -7,32 +7,26 @@ namespace mlx::core {
 
 namespace {
 
-// Create a cudnn tensor descriptor.
-template <typename Vec>
-inline cudnn_frontend::Tensor build_cudnn_tensor(
-    int64_t id,
-    const array& x,
-    const Vec& shape,
-    const Vec& strides) {
-  return cudnn_frontend::TensorBuilder()
-      .setDim(shape.size(), shape.data())
-      .setStrides(strides.size(), strides.data())
-      .setId(id)
-      .setAlignment(get_alignment(x))
-      .setDataType(dtype_to_cudnn_type(x.dtype()))
-      .build();
-}
+#define RETURN_IF_ERROR(cmd)          \
+  if (auto ret = cmd; ret.is_bad()) { \
+    return ret;                       \
+  }
 
 // In MLX a singleton dim (shape[dim] == 1) can have any stride, but in cuDNN
 // whether a tensor is contiguous is determined with:
 // shape[dim] == shape[dim + 1] * strides[dim + 1]
 // So a contiguous array with singleton dims in MLX may be mistakenly treated
 // as strided in cuDNN, and we work around it by normalizing the strides.
-Strides normalized_strides(const array& x) {
-  if (!x.flags().row_contiguous || x.ndim() < 2) {
-    return x.strides();
+std::vector<int64_t> normalized_strides(const array& x) {
+  std::vector<int64_t> strides(x.strides().begin(), x.strides().end());
+  if (std::all_of(
+          strides.begin(), strides.end(), [](int64_t s) { return s == 0; })) {
+    strides.back() = 1;
+    return strides;
+  }
+  if (!x.flags().row_contiguous || x.ndim() < 2) {
+    return strides;
   }
-  Strides strides = x.strides();
   for (int i = x.ndim() - 2; i >= 0; --i) {
     if (x.shape(i) == 1) {
       strides[i] = x.shape(i + 1) * strides[i + 1];
@@ -42,7 +36,9 @@ Strides normalized_strides(const array& x) {
 }
 
 // Return the shape and strides after transposing from NHWC to NCHW.
-auto nhwc_to_nchw(SmallVector<int64_t> shape, SmallVector<int64_t> strides) {
+inline auto nhwc_to_nchw(const array& x) {
+  auto shape = convert_vector<int64_t>(x.shape());
+  auto strides = normalized_strides(x);
   assert(shape.size() >= 3);
   shape.insert(shape.begin() + 1, shape.back());
   shape.erase(shape.end() - 1);
@@ -51,225 +47,95 @@ auto nhwc_to_nchw(SmallVector<int64_t> shape, SmallVector<int64_t> strides) {
   return std::make_tuple(std::move(shape), std::move(strides));
 }
 
-inline auto nhwc_to_nchw(const array& x) {
-  return nhwc_to_nchw(
-      convert_vector<int64_t>(x.shape()), normalized_strides(x));
-}
-
-// Return available engines for a |op_graph|.
-cudnn_frontend::EngineConfigList get_cudnn_engine_configs(
-    cudnnBackendDescriptorType_t backend_type,
-    Dtype dtype,
-    cudnn_frontend::OperationGraph& op_graph,
-    bool use_fallback = true) {
-  SmallVector<cudnn_frontend::GeneratorSource, 2> sources;
-  sources.push_back([](auto& op_graph) {
-    auto heuristics = cudnn_frontend::EngineHeuristicsBuilder()
-                          .setOperationGraph(op_graph)
-                          .setHeurMode(CUDNN_HEUR_MODE_A)
-                          .build();
-    return heuristics.getEngineConfig(heuristics.getEngineConfigCount());
-  });
-  if (use_fallback) {
-    sources.push_back([&backend_type](auto& op_graph) {
-      auto fallback = cudnn_frontend::EngineFallbackListBuilder()
-                          .setOperationGraph(op_graph)
-                          .setOperation(backend_type)
-                          .build();
-      return fallback.getFallbackList();
-    });
-  }
-
-  auto configs =
-      cudnn_frontend::EngineConfigGenerator(sources.size(), sources.data())
-          .generate_engine_config(op_graph);
-
-  cudnn_frontend::EngineConfigList filtered_configs;
-  cudnn_frontend::filter(configs, filtered_configs, [dtype](auto c) {
-    if (cudnn_frontend::hasNumericalNote<
-            CUDNN_NUMERICAL_NOTE_DOWN_CONVERT_INPUTS>(c)) {
-      return true;
-    }
-    if (cudnn_frontend::hasNumericalNote<CUDNN_NUMERICAL_NOTE_TENSOR_CORE>(c) &&
-        dtype == float32 && !env::enable_tf32()) {
-      return true;
-    }
-    return false;
-  });
-  return filtered_configs;
-}
-
-// Take |engine_configs| and |op_graph| and find a working execution plans
-// from them.
-std::optional<cudnn_frontend::ExecutionPlan>
-find_cudnn_plan_from_engine_configs(
-    cudnnHandle_t handle,
-    const cudnn_frontend::EngineConfigList& engine_configs,
-    const cudnn_frontend::OperationGraph& op_graph) {
-  auto op_graph_tag = op_graph.getTag();
-  for (const auto& config : engine_configs) {
-    try {
-      return cudnn_frontend::ExecutionPlanBuilder()
-          .setHandle(handle)
-          .setEngineConfig(config, op_graph_tag)
-          .build();
-    } catch (cudnn_frontend::cudnnException& error) {
-      if (error.getCudnnStatus() != CUDNN_STATUS_NOT_SUPPORTED) {
-        throw;
-      }
-    }
-  }
-  return std::nullopt;
-}
-
-// Prepare workspace and args to execute plan.
-template <typename F>
-bool prepare_cudnn_plan(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs,
-    F&& execute) {
-  int workspace_size = plan.getWorkspaceSize();
-  array workspace(
-      workspace_size > 0 ? allocator::malloc(workspace_size)
-                         : allocator::Buffer(nullptr),
-      {workspace_size},
-      uint8);
-
-  auto args = cudnn_frontend::VariantPackBuilder()
-                  .setWorkspacePointer(workspace.data<void>())
-                  .setDataPointers(num_args, data_ptrs)
-                  .setUids(num_args, uids)
-                  .build();
-
-  auto handle = encoder.device().cudnn_handle();
-  cudnnSetStream(handle, encoder.stream());
-
-  if (!execute(handle, plan.get_raw_desc(), args.get_raw_desc())) {
-    return false;
-  }
-
-  encoder.add_temporary(workspace);
-  return true;
-}
-
 } // namespace
 
-cudnn_frontend::Tensor build_cudnn_tensor(int64_t id, const array& x) {
-  auto shape = convert_vector<int64_t>(x.shape());
-  return build_cudnn_tensor(id, x, shape, normalized_strides(x));
+fe::error_t DnnGraph::prepare() {
+  RETURN_IF_ERROR(validate());
+  try {
+    RETURN_IF_ERROR(build_operation_graph(handle_));
+  } catch (cudnn_frontend::cudnnException& error) {
+    // cuDNN bug: they did not catch all exceptions in the API.
+    return {fe::error_code_t::CUDNN_BACKEND_API_FAILED, error.what()};
+  }
+  RETURN_IF_ERROR(create_execution_plans({fe::HeurMode_t::A}));
+  return {};
 }
 
-cudnn_frontend::Tensor build_cudnn_tensor_nchw(int64_t id, const array& x) {
+fe::error_t DnnGraph::build() {
+  RETURN_IF_ERROR(check_support(handle_));
+  RETURN_IF_ERROR(build_plans(handle_));
+  return {};
+}
+
+fe::error_t DnnGraph::encode_graph(
+    cu::CommandEncoder& encoder,
+    std::unordered_map<int64_t, void*> variant_pack) {
+  cudnnSetStream(handle_, encoder.stream());
+  CudaGraph cuda_graph(encoder.device());
+  RETURN_IF_ERROR(populate_cuda_graph(
+      handle_, variant_pack, prepare_workspace(encoder), cuda_graph));
+  encoder.add_graph_node(cuda_graph);
+  return {};
+}
+
+fe::error_t DnnGraph::encode_capturing(
+    cu::CommandEncoder& encoder,
+    std::unordered_map<int64_t, void*> variant_pack) {
+  auto* workspace_ptr = prepare_workspace(encoder);
+  auto capture = encoder.capture_context();
+  cudnnSetStream(handle_, encoder.stream());
+  auto ret = execute(handle_, variant_pack, workspace_ptr);
+  if (ret.is_bad()) {
+    capture.discard = true;
+  }
+  return ret;
+}
+
+void* DnnGraph::prepare_workspace(cu::CommandEncoder& encoder) {
+  int64_t workspace_size = 0;
+  CHECK_CUDNN_FE_ERROR(get_workspace_size(workspace_size));
+  if (workspace_size > 0) {
+    array workspace(
+        cu::malloc_async(workspace_size, encoder),
+        {static_cast<int>(workspace_size)},
+        uint8);
+    encoder.add_temporary(workspace);
+    return gpu_ptr<void>(workspace);
+  }
+  return nullptr;
+}
+
+void DnnGraph::set_tensor_attrs(
+    std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
+    int64_t uid,
+    const array& x,
+    const std::vector<int64_t>& shape,
+    const std::vector<int64_t>& strides) {
+  tensor->set_uid(uid)
+      .set_alignment(get_alignment(x))
+      .set_data_type(dtype_to_cudnn_type(x.dtype()))
+      .set_dim(shape)
+      .set_stride(strides);
+}
+
+void DnnGraph::set_tensor_attrs(
+    std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
+    int64_t uid,
+    const array& x) {
+  set_tensor_attrs(
+      tensor,
+      uid,
+      x,
+      convert_vector<int64_t>(x.shape()),
+      normalized_strides(x));
+}
+
+void DnnGraph::set_tensor_attrs_nchw(
+    std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
+    int64_t uid,
+    const array& x) {
   auto [shape, strides] = nhwc_to_nchw(x);
-  return build_cudnn_tensor(id, x, shape, strides);
+  set_tensor_attrs(tensor, uid, x, shape, strides);
 }
 
-cudnn_frontend::Tensor build_cudnn_tensor_4d_nchw(int64_t id, const array& x) {
-  if (x.ndim() == 0) {
-    SmallVector<int64_t, 4> scalar_dims = {1, 1, 1, 1};
-    return build_cudnn_tensor(id, x, scalar_dims, scalar_dims);
-  }
-  if (x.ndim() == 1) {
-    int64_t s = x.shape(0);
-    SmallVector<int64_t, 4> shape = {1, x.shape(0), 1, 1};
-    SmallVector<int64_t, 4> strides = {s, 1, s, s};
-    return build_cudnn_tensor(id, x, shape, strides);
-  }
-  if (x.ndim() == 2) {
-    int64_t s =
-        x.flags().row_contiguous ? x.shape(1) * x.strides(1) : x.strides(0);
-    SmallVector<int64_t, 4> shape = {x.shape(0), x.shape(1), 1, 1};
-    SmallVector<int64_t, 4> strides = {s, x.strides(1), s, s};
-    return build_cudnn_tensor(id, x, shape, strides);
-  }
-  if (x.ndim() == 3 || x.ndim() == 4) {
-    return build_cudnn_tensor_nchw(id, x);
-  }
-  throw std::runtime_error(
-      fmt::format("Unsupported array with {} dims.", x.ndim()));
-}
-
-cudnn_frontend::Tensor build_cudnn_scalar_4d(int64_t id, Dtype dtype) {
-  SmallVector<int64_t, 4> scalar_dims = {1, 1, 1, 1};
-  return cudnn_frontend::TensorBuilder()
-      .setDim(scalar_dims.size(), scalar_dims.data())
-      .setStrides(scalar_dims.size(), scalar_dims.data())
-      .setId(id)
-      .setAlignment(16)
-      .setDataType(dtype_to_cudnn_type(dtype))
-      .setByValue(true)
-      .build();
-}
-
-std::optional<cudnn_frontend::ExecutionPlan> find_cudnn_plan_from_op_graph(
-    cudnnHandle_t handle,
-    cudnnBackendDescriptorType_t backend_type,
-    Dtype dtype,
-    cudnn_frontend::OperationGraph& op_graph) {
-  auto engine_configs = get_cudnn_engine_configs(backend_type, dtype, op_graph);
-  if (engine_configs.empty()) {
-    return std::nullopt;
-  }
-  return find_cudnn_plan_from_engine_configs(handle, engine_configs, op_graph);
-}
-
-bool encode_cudnn_plan_with_capturing(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs) {
-  return prepare_cudnn_plan(
-      encoder,
-      plan,
-      num_args,
-      uids,
-      data_ptrs,
-      [&](auto handle, auto plan, auto args) {
-        auto capture = encoder.capture_context();
-        if (cudnnBackendExecute(handle, plan, args) != CUDNN_STATUS_SUCCESS) {
-          // Discard the captured graph when failed.
-          capture.discard = true;
-          return false;
-        }
-        return true;
-      });
-}
-
-#if CUDNN_VERSION >= 90500
-bool encode_cudnn_plan_with_graph_api(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    CudaGraph& graph,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs) {
-  return prepare_cudnn_plan(
-      encoder,
-      plan,
-      num_args,
-      uids,
-      data_ptrs,
-      [&](auto handle, auto plan, auto args) {
-        if (!graph) {
-          graph = CudaGraph(encoder.device());
-          if (cudnnBackendPopulateCudaGraph(handle, plan, args, graph) !=
-              CUDNN_STATUS_SUCCESS) {
-            return false;
-          }
-        } else {
-          if (cudnnBackendUpdateCudaGraph(handle, plan, args, graph) !=
-              CUDNN_STATUS_SUCCESS) {
-            return false;
-          }
-        }
-        encoder.add_graph_node(graph);
-        return true;
-      });
-}
-#endif
-
 } // namespace mlx::core

mlx/backend/cuda/cudnn_utils.h

@@ -2,28 +2,34 @@
 
 #pragma once
 
-#include "mlx/array.h"
 #include "mlx/backend/cuda/device/config.h"
 #include "mlx/backend/cuda/utils.h"
 #include "mlx/dtype_utils.h"
 
 #include <cudnn_frontend.h>
-#include <cudnn_frontend_find_plan.h>
 #include <fmt/format.h>
 
-#include <algorithm>
-#include <array>
-
 namespace mlx::core {
 
 namespace cu {
 class CommandEncoder;
 }
 
+namespace fe = cudnn_frontend;
+
+#define CHECK_CUDNN_FE_ERROR(cmd)                                    \
+  do {                                                               \
+    auto error = cmd;                                                \
+    if (!error.is_good()) {                                          \
+      throw std::runtime_error(                                      \
+          fmt::format("{} failed: {}.", #cmd, error.get_message())); \
+    }                                                                \
+  } while (0)
+
 // Return pointer alignment of |x|'s data.
 inline uint8_t get_alignment(const array& x) {
   uint8_t alignment = 1;
-  uintptr_t address = reinterpret_cast<uintptr_t>(x.data<void>());
+  uintptr_t address = reinterpret_cast<uintptr_t>(gpu_ptr<void>(x));
   for (; alignment < 32; alignment *= 2) {
     if (address % (alignment * 2)) {
       return alignment;
@@ -34,8 +40,31 @@ inline uint8_t get_alignment(const array& x) {
 
 // Convert the type of elements in |vec| to |T|.
 template <typename T, typename Vec>
-inline SmallVector<T> convert_vector(const Vec& vec) {
-  return SmallVector<T>(vec.begin(), vec.end());
+inline std::vector<T> convert_vector(const Vec& vec) {
+  return std::vector<T>(vec.begin(), vec.end());
+}
+
+// Map dtype to cudnn data type.
+inline fe::DataType_t dtype_to_cudnn_type(Dtype dtype) {
+  switch (dtype) {
+    case int8:
+      return fe::DataType_t::INT8;
+    case int32:
+      return fe::DataType_t::INT32;
+    case uint8:
+      return fe::DataType_t::UINT8;
+    case float16:
+      return fe::DataType_t::HALF;
+    case bfloat16:
+      return fe::DataType_t::BFLOAT16;
+    case float32:
+      return fe::DataType_t::FLOAT;
+    case float64:
+      return fe::DataType_t::DOUBLE;
+    default:
+      throw std::runtime_error(fmt::format(
+          "Unsupported dtype in cuDNN: {}.", dtype_to_string(dtype)));
+  }
 }
 
 // Return an array that can be used as map key for |vec| with size <= MAX_NDIM.
@@ -43,122 +72,100 @@ inline SmallVector<T> convert_vector(const Vec& vec) {
 // There are 2 differences from the const_param util from kernel_utils.cuh:
 // 1. The rest of array is filled with 0.
 // 2. This util can be used in .cpp files.
-template <typename T, template <typename U> class Vec>
-inline std::array<T, MAX_NDIM> vector_key(const Vec<T>& vec) {
-  if (vec.size() > MAX_NDIM) {
+template <int NDIM = MAX_NDIM, typename T, template <typename U> class Vec>
+inline std::array<T, NDIM> vector_key(const Vec<T>& vec) {
+  if (vec.size() > NDIM) {
     throw std::runtime_error(
-        fmt::format("ndim can not be larger than {}.", MAX_NDIM));
+        fmt::format("ndim can not be larger than {}.", NDIM));
   }
-  std::array<T, MAX_NDIM> result = {};
+  std::array<T, NDIM> result = {};
   std::copy_n(vec.begin(), vec.size(), result.begin());
   return result;
 }
 
-// Helpers used by get_data_ptrs to get pointers.
-inline void* get_data_ptr(const array& arr) {
-  return const_cast<void*>(arr.data<void>());
-}
-
-template <typename T, typename = std::enable_if_t<std::is_scalar_v<T>>>
-inline void* get_data_ptr(T& scalar) {
-  return &scalar;
-}
-
-// Return an array filled with data pointers of args.
-template <typename... Args>
-inline std::array<void*, sizeof...(Args)> get_data_ptrs(Args&... args) {
-  return {get_data_ptr(args)...};
-}
-
-// Map dtype to cudnn data type.
-inline cudnnDataType_t dtype_to_cudnn_type(Dtype dtype) {
-  switch (dtype) {
-    case int8:
-      return CUDNN_DATA_INT8;
-    case int32:
-      return CUDNN_DATA_INT32;
-    case uint8:
-      return CUDNN_DATA_UINT8;
-    case float16:
-      return CUDNN_DATA_HALF;
-    case bfloat16:
-      return CUDNN_DATA_BFLOAT16;
-    case float32:
-      return CUDNN_DATA_FLOAT;
-    case float64:
-      return CUDNN_DATA_DOUBLE;
-    default:
-      throw std::runtime_error(fmt::format(
-          "Unsupported dtype in Convolution: {}.", dtype_to_string(dtype)));
+// Extends cuDNN graph with helpers.
+class DnnGraph : public fe::graph::Graph {
+ public:
+  DnnGraph(cudnnHandle_t handle, Dtype io_dtype, Dtype compute_dtype = float32)
+      : handle_(handle) {
+    set_io_data_type(dtype_to_cudnn_type(io_dtype));
+    set_intermediate_data_type(dtype_to_cudnn_type(compute_dtype));
+    set_compute_data_type(dtype_to_cudnn_type(compute_dtype));
   }
-}
 
-// Create a tensor descriptor from |x|.
-cudnn_frontend::Tensor build_cudnn_tensor(int64_t id, const array& x);
+  // Create a cuDNN tensor description from MLX array |x|.
+  auto& tensor(
+      std::shared_ptr<fe::graph::Tensor_attributes>& attrs,
+      int64_t uid,
+      const array& x) {
+    set_tensor_attrs(attrs, uid, x);
+    return attrs;
+  }
+  auto tensor(const char* name, int64_t uid, const array& x) {
+    auto attrs = Graph::tensor(fe::graph::Tensor_attributes().set_name(name));
+    tensor(attrs, uid, x);
+    return attrs;
+  }
 
-// Create a tensor descriptor from |x|, and transpose from NHWC to NCHW.
-cudnn_frontend::Tensor build_cudnn_tensor_nchw(int64_t id, const array& x);
+  // Create a cuDNN tensor description from MLX array |x|, and transpose it from
+  // NHWC layout to NCHW.
+  auto& tensor_nchw(
+      std::shared_ptr<fe::graph::Tensor_attributes>& attrs,
+      int64_t uid,
+      const array& x) {
+    set_tensor_attrs_nchw(attrs, uid, x);
+    return attrs;
+  }
+  auto tensor_nchw(const char* name, int64_t uid, const array& x) {
+    auto attrs = Graph::tensor(fe::graph::Tensor_attributes().set_name(name));
+    tensor_nchw(attrs, uid, x);
+    return attrs;
+  }
 
-// Create a tensor descriptor from |x|, make sure it is 4D, and transpose it
-// from NHWC to NCHW.
-cudnn_frontend::Tensor build_cudnn_tensor_4d_nchw(int64_t id, const array& x);
+  // Create a cuDNN tensor for scalar.
+  auto scalar(const char* name, int64_t uid, Dtype dtype) {
+    return Graph::tensor(fe::graph::Tensor_attributes()
+                             .set_name(name)
+                             .set_uid(uid)
+                             .set_dim({1, 1, 1, 1})
+                             .set_stride({1, 1, 1, 1})
+                             .set_is_pass_by_value(true)
+                             .set_data_type(dtype_to_cudnn_type(dtype)));
+  }
 
-// Create a 4D scalar tensor descriptor, which is passed by value.
-cudnn_frontend::Tensor build_cudnn_scalar_4d(int64_t id, Dtype dtype);
+  // Call this before setting notes.
+  fe::error_t prepare();
+  // Call this after setting notes.
+  fe::error_t build();
 
-// Find a working plan for |op_graph|.
-std::optional<cudnn_frontend::ExecutionPlan> find_cudnn_plan_from_op_graph(
-    cudnnHandle_t handle,
-    cudnnBackendDescriptorType_t backend_type,
-    Dtype dtype,
-    cudnn_frontend::OperationGraph& op_graph);
+  // Add cuDNN graph to CUDA graph, using native CUDA graph API.
+  fe::error_t encode_graph(
+      cu::CommandEncoder& encoder,
+      std::unordered_map<int64_t, void*> variant_pack);
+  // Add cuDNN graph to CUDA graph, using stream capture.
+  fe::error_t encode_capturing(
+      cu::CommandEncoder& encoder,
+      std::unordered_map<int64_t, void*> variant_pack);
 
-// Encode the plan to command buffer by capturing.
-bool encode_cudnn_plan_with_capturing(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs);
+ private:
+  void* prepare_workspace(cu::CommandEncoder& encoder);
 
-#if CUDNN_VERSION >= 90500
-// Encode the plan to command buffer by using native graph api of cudnn. If the
-// |graph| is empty it will be populated, otherwise it will be updated.
-bool encode_cudnn_plan_with_graph_api(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    CudaGraph& graph,
-    int num_args,
-    const int64_t* uids,
-    void** data_ptrs);
-#endif
+  void set_tensor_attrs(
+      std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
+      int64_t uid,
+      const array& x,
+      const std::vector<int64_t>& shape,
+      const std::vector<int64_t>& strides);
+  void set_tensor_attrs(
+      std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
+      int64_t uid,
+      const array& x);
+  void set_tensor_attrs_nchw(
+      std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
+      int64_t uid,
+      const array& x);
 
-// Helpers to make calls like encode_cudnn_plan(..., {'x', 'y', 'z'}, x, y, z).
-template <typename... Args>
-bool encode_cudnn_plan(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    std::initializer_list<int64_t> uids,
-    Args&... args) {
-  assert(uids.size() == sizeof...(args));
-  auto data_ptrs = get_data_ptrs(args...);
-  return encode_cudnn_plan_with_capturing(
-      encoder, plan, uids.size(), uids.begin(), data_ptrs.data());
-}
-
-#if CUDNN_VERSION >= 90500
-template <typename... Args>
-bool encode_cudnn_plan(
-    cu::CommandEncoder& encoder,
-    cudnn_frontend::ExecutionPlan& plan,
-    CudaGraph& graph,
-    std::initializer_list<int64_t> uids,
-    Args&... args) {
-  assert(uids.size() == sizeof...(args));
-  auto data_ptrs = get_data_ptrs(args...);
-  return encode_cudnn_plan_with_graph_api(
-      encoder, plan, graph, uids.size(), uids.begin(), data_ptrs.data());
-}
-#endif
+  cudnnHandle_t handle_;
+};
 
 } // namespace mlx::core

mlx/backend/cuda/custom_kernel.cpp

@@ -57,7 +57,7 @@ std::string build_kernel(
     const std::vector<std::string>& output_names,
     const std::vector<Dtype>& output_dtypes,
     const std::vector<std::pair<std::string, TemplateArg>>& template_args,
-    const std::vector<CustomKernelShapeInfo>& shape_infos) {
+    const std::vector<std::tuple<bool, bool, bool>>& shape_infos) {
   std::string kernel_source;
   kernel_source.reserve(header.size() + source.size() + 8192);
   kernel_source += default_header;
@@ -81,17 +81,17 @@ std::string build_kernel(
     kernel_source += ",\n";
     // Add input shape, strides and ndim if present in the source
     if (arr.ndim() > 0) {
-      if (shape_infos[i].shape) {
+      if (std::get<0>(shape_infos[i])) {
         kernel_source += "    const __grid_constant__ Shape ";
         kernel_source += name;
         kernel_source += "_shape,\n";
       }
-      if (shape_infos[i].strides) {
+      if (std::get<1>(shape_infos[i])) {
         kernel_source += "    const __grid_constant__ Strides ";
         kernel_source += name;
         kernel_source += "_strides,\n";
       }
-      if (shape_infos[i].ndim) {
+      if (std::get<2>(shape_infos[i])) {
         kernel_source += "    const __grid_constant__ int ";
         kernel_source += name;
         kernel_source += "_ndim,\n";
@@ -154,12 +154,12 @@ CustomKernelFunction cuda_kernel(
         "[custom_kernel] Must specify at least one output.");
   }
 
-  std::vector<CustomKernelShapeInfo> shape_infos;
+  std::vector<std::tuple<bool, bool, bool>> shape_infos;
   for (auto& n : input_names) {
-    CustomKernelShapeInfo shape_info;
-    shape_info.shape = source.find(n + "_shape") != std::string::npos;
-    shape_info.strides = source.find(n + "_strides") != std::string::npos;
-    shape_info.ndim = source.find(n + "_ndim") != std::string::npos;
+    std::tuple<bool, bool, bool> shape_info;
+    std::get<0>(shape_info) = source.find(n + "_shape") != std::string::npos;
+    std::get<1>(shape_info) = source.find(n + "_strides") != std::string::npos;
+    std::get<2>(shape_info) = source.find(n + "_ndim") != std::string::npos;
     shape_infos.push_back(shape_info);
   }
 
@@ -254,8 +254,8 @@ std::vector<array> precompiled_cuda_kernel(
     std::optional<float> init_value,
     bool ensure_row_contiguous,
     StreamOrDevice s) {
-  std::vector<CustomKernelShapeInfo> shape_infos(
-      inputs.size(), CustomKernelShapeInfo{false, false, false});
+  std::vector<std::tuple<bool, bool, bool>> shape_infos(
+      inputs.size(), {false, false, false});
   return array::make_arrays(
       output_shapes,
       output_dtypes,
@@ -279,6 +279,7 @@ void CustomKernel::eval_gpu(
     std::vector<array>& outputs) {
   nvtx3::scoped_range r("CustomKernel::eval_gpu");
   auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
 
   std::vector<array> copies;
 
@@ -288,7 +289,7 @@ void CustomKernel::eval_gpu(
       copies.emplace_back(init_value_.value(), out.dtype());
       fill_gpu(copies.back(), out, s);
     } else {
-      out.set_data(allocator::malloc(out.nbytes()));
+      out.set_data(cu::malloc_async(out.nbytes(), encoder));
     }
   }
 
@@ -326,13 +327,13 @@ void CustomKernel::eval_gpu(
     const array& in = checked_inputs[i];
     auto& shape_info = shape_infos_[i];
     args.append(in);
-    if (shape_info.shape) {
+    if (std::get<0>(shape_info)) {
       args.append_ndim(in.shape());
     }
-    if (shape_info.strides) {
+    if (std::get<1>(shape_info)) {
       args.append_ndim(in.strides());
     }
-    if (shape_info.ndim) {
+    if (std::get<2>(shape_info)) {
       args.append<int32_t>(in.ndim());
     }
   }
@@ -356,7 +357,6 @@ void CustomKernel::eval_gpu(
   dim3 grid((gx + tx - 1) / tx, (gy + ty - 1) / ty, (gz + tz - 1) / tz);
 
   // Call the kernel
-  auto& encoder = cu::get_command_encoder(s);
   for (const auto& in : checked_inputs) {
     encoder.set_input_array(in);
   }

mlx/backend/cuda/device.cpp

@@ -14,20 +14,20 @@ namespace mlx::core::cu {
 
 namespace {
 
-#define CHECK_CUDNN_ERROR(cmd) check_cudnn_error(#cmd, (cmd))
-
-void check_cudnn_error(const char* name, cudnnStatus_t err) {
-  if (err != CUDNN_STATUS_SUCCESS) {
-    throw std::runtime_error(
-        fmt::format("{} failed: {}.", name, cudnnGetErrorString(err)));
-  }
+bool use_cuda_graphs() {
+  static bool use_graphs = env::get_var("MLX_USE_CUDA_GRAPHS", true);
+  return use_graphs;
 }
 
-bool use_cuda_graphs() {
-  static bool use_graphs = []() {
-    return env::get_var("MLX_USE_CUDA_GRAPHS", true);
+const char* save_cuda_graphs_dot_file() {
+  static const char* filename = []() -> const char* {
+    const char* env = std::getenv("MLX_SAVE_CUDA_GRAPHS_DOT_FILE");
+    if (env && std::strlen(env) == 0) {
+      return nullptr;
+    }
+    return env;
   }();
-  return use_graphs;
+  return filename;
 }
 
 } // namespace
@@ -46,6 +46,7 @@ Device::Device(int device) : device_(device) {
         "Device {} does not support synchronization in managed memory.",
         device_));
   }
+
   // The cublasLt handle is used by matmul.
   make_current();
   CHECK_CUBLAS_ERROR(cublasLtCreate(&lt_));
@@ -86,7 +87,7 @@ CommandEncoder::CaptureContext::CaptureContext(CommandEncoder& enc) : enc(enc) {
     return;
   }
   CHECK_CUDA_ERROR(
-      cudaStreamBeginCapture(enc.stream(), cudaStreamCaptureModeGlobal));
+      cudaStreamBeginCapture(enc.stream(), cudaStreamCaptureModeThreadLocal));
 }
 
 CommandEncoder::CaptureContext::~CaptureContext() {
@@ -114,18 +115,17 @@ CommandEncoder::ConcurrentContext::~ConcurrentContext() {
   }
 
   // Use an empty graph node for synchronization
-  CommandEncoder::GraphNode empty{NULL, 'E', std::to_string(enc.node_count_++)};
-  enc.empty_node_count_++;
+  CommandEncoder::GraphNode empty{NULL, "E", std::to_string(enc.node_count_++)};
   CHECK_CUDA_ERROR(cudaGraphAddEmptyNode(&empty.node, enc.graph_, NULL, 0));
 
   // Insert the concurrent -> empty node dependencies
   for (auto& from : enc.concurrent_nodes_) {
     enc.from_nodes_.push_back(from.node);
     enc.to_nodes_.push_back(empty.node);
-    enc.graph_key_ += from.id;
-    enc.graph_key_ += from.node_type;
-    enc.graph_key_ += empty.id;
-    enc.graph_key_ += empty.node_type;
+    enc.graph_deps_key_ += from.id;
+    enc.graph_deps_key_ += "-";
+    enc.graph_deps_key_ += empty.id;
+    enc.graph_deps_key_ += "-";
   }
 
   // Insert the input -> concurrent node dependencies without updating output
@@ -140,9 +140,6 @@ CommandEncoder::ConcurrentContext::~ConcurrentContext() {
 }
 
 void CommandEncoder::insert_graph_dependencies(GraphNode node) {
-  if (node.node_type == 'G') {
-    graph_node_count_++;
-  }
   node.id = std::to_string(node_count_++);
   if (in_concurrent_) {
     concurrent_nodes_.push_back(std::move(node));
@@ -154,6 +151,10 @@ void CommandEncoder::insert_graph_dependencies(GraphNode node) {
 }
 
 void CommandEncoder::insert_graph_dependencies(std::vector<GraphNode> nodes) {
+  for (auto& node : nodes) {
+    graph_nodes_key_ += node.node_type;
+    graph_nodes_key_ += "-";
+  }
   std::vector<GraphNode> deps;
   {
     // Dependencies must be added in the same order to produce a consistent
@@ -181,20 +182,49 @@ void CommandEncoder::insert_graph_dependencies(std::vector<GraphNode> nodes) {
     for (auto& to : nodes) {
       from_nodes_.push_back(from.node);
       to_nodes_.push_back(to.node);
-      graph_key_ += from.id;
-      graph_key_ += from.node_type;
-      graph_key_ += to.id;
-      graph_key_ += to.node_type;
+      graph_deps_key_ += from.id;
+      graph_deps_key_ += "-";
+      graph_deps_key_ += to.id;
+      graph_deps_key_ += "-";
     }
   }
 }
 
+// Can be tuned with MLX_MAX_OPS_PER_BUFFER, MLX_MAX_MB_PER_BUFFER
+std::pair<int, int> get_graph_limits(Device& d) {
+  auto cc =
+      d.compute_capability_major() * 100 + d.compute_capability_minor() * 10;
+  int ops = 20;
+  int mb = 100;
+  switch (cc) {
+    case 800: // A100
+      ops = 20;
+      mb = 400;
+      break;
+    case 900: // H100
+      ops = 30;
+      mb = 400;
+      break;
+    case 1000: // B200
+      ops = 50;
+      mb = 500;
+      break;
+    case 1210: // DGX Spark
+      ops = 20;
+      mb = 25;
+      break;
+  }
+  return {env::max_ops_per_buffer(ops), env::max_mb_per_buffer(mb)};
+}
+
 CommandEncoder::CommandEncoder(Device& d)
     : device_(d),
       stream_(d),
       graph_(d),
       worker_(d),
-      graph_cache_("MLX_CUDA_GRAPH_CACHE_SIZE", /* default_capacity */ 400) {}
+      graph_cache_("MLX_CUDA_GRAPH_CACHE_SIZE", /* default_capacity */ 400) {
+  std::tie(max_ops_per_graph_, max_mb_per_graph_) = get_graph_limits(d);
+}
 
 void CommandEncoder::add_completed_handler(std::function<void()> task) {
   worker_.add_task(std::move(task));
@@ -204,6 +234,7 @@ void CommandEncoder::set_input_array(const array& arr) {
   if (!use_cuda_graphs()) {
     return;
   }
+  bytes_in_graph_ += arr.data_size();
   auto id = reinterpret_cast<std::uintptr_t>(arr.buffer().ptr());
   active_deps_.push_back(id);
 }
@@ -278,13 +309,55 @@ void CommandEncoder::add_kernel_node(
 void CommandEncoder::add_kernel_node(const cudaKernelNodeParams& params) {
   cudaGraphNode_t node;
   CHECK_CUDA_ERROR(cudaGraphAddKernelNode(&node, graph_, NULL, 0, &params));
-  insert_graph_dependencies(GraphNode{node, 'K'});
+  insert_graph_dependencies(GraphNode{node, "K"});
 }
 
 void CommandEncoder::add_kernel_node(const CUDA_KERNEL_NODE_PARAMS& params) {
   CUgraphNode node;
   CHECK_CUDA_ERROR(cuGraphAddKernelNode(&node, graph_, NULL, 0, &params));
-  insert_graph_dependencies(GraphNode{node, 'K'});
+  insert_graph_dependencies(GraphNode{node, "K"});
+}
+
+bool is_graph_updatable(cudaGraph_t graph, int& cluster_dim_x) {
+  // CUDA graphs do not get updated correctly if a kernel node getting updated
+  // has a different cluster shape than the node it's being updated with.
+  size_t num_nodes = 0;
+  CHECK_CUDA_ERROR(cudaGraphGetNodes(graph, nullptr, &num_nodes));
+  if (num_nodes == 0) {
+    return true;
+  }
+
+  std::vector<cudaGraphNode_t> nodes(num_nodes);
+  CHECK_CUDA_ERROR(cudaGraphGetNodes(graph, nodes.data(), &num_nodes));
+  for (const auto& node : nodes) {
+    cudaGraphNodeType type;
+    CHECK_CUDA_ERROR(cudaGraphNodeGetType(node, &type));
+    if (type == cudaGraphNodeTypeGraph) {
+      // Try to be updatable for a structure like graph -> graph -> kernel
+      if (num_nodes > 1) {
+        return false;
+      }
+      cudaGraph_t child;
+      CHECK_CUDA_ERROR(cudaGraphChildGraphNodeGetGraph(node, &child));
+      return is_graph_updatable(child, cluster_dim_x);
+    } else if (type != cudaGraphNodeTypeKernel) {
+      return false;
+    } else {
+      cudaLaunchAttributeValue cluster_dim;
+      CHECK_CUDA_ERROR(cudaGraphKernelNodeGetAttribute(
+          node, cudaLaunchAttributeClusterDimension, &cluster_dim));
+      // Only dim.x can be greater than 1
+      if (cluster_dim.clusterDim.y > 1 || cluster_dim.clusterDim.z > 1) {
+        return false;
+      }
+      // Only one child node allowed when subgraph uses clusters
+      if (cluster_dim.clusterDim.x > 0 && num_nodes > 1) {
+        return false;
+      }
+      cluster_dim_x = cluster_dim.clusterDim.x;
+    }
+  }
+  return true;
 }
 
 void CommandEncoder::add_graph_node(cudaGraph_t child) {
@@ -297,12 +370,16 @@ void CommandEncoder::add_graph_node(cudaGraph_t child) {
     return;
   }
   cudaGraphNode_t node;
+  int cluster_dim_x = 0;
+  is_graph_updatable_ &= is_graph_updatable(child, cluster_dim_x);
   CHECK_CUDA_ERROR(cudaGraphAddChildGraphNode(&node, graph_, NULL, 0, child));
-  insert_graph_dependencies(GraphNode{node, 'G'});
+  insert_graph_dependencies(
+      GraphNode{node, "G" + std::to_string(cluster_dim_x)});
 }
 
-int CommandEncoder::get_num_ops() {
-  return node_count_;
+bool CommandEncoder::needs_commit() {
+  return (node_count_ > max_ops_per_graph_) ||
+      ((bytes_in_graph_ >> 20) > max_mb_per_graph_);
 }
 
 void CommandEncoder::commit() {
@@ -322,53 +399,63 @@ void CommandEncoder::commit() {
           from_nodes_.size()));
     }
 
-    graph_key_ += ".";
-    graph_key_ += std::to_string(node_count_);
-    graph_key_ += ".";
-    graph_key_ += std::to_string(graph_node_count_);
-    graph_key_ += ".";
-    graph_key_ += std::to_string(empty_node_count_);
-
-    CudaGraphExec& graph_exec = graph_cache_[graph_key_];
-
-    if (graph_exec != nullptr) {
-      cudaGraphExecUpdateResult update_result;
-#if CUDART_VERSION >= 12000
-      cudaGraphExecUpdateResultInfo info;
-      cudaGraphExecUpdate(graph_exec, graph_, &info);
-      update_result = info.result;
-#else
-      cudaGraphNode_t error_node;
-      cudaGraphExecUpdate(graph_exec, graph_, &error_node, &update_result);
-#endif // CUDART_VERSION >= 12000
-      if (update_result != cudaGraphExecUpdateSuccess) {
-        cudaGetLastError(); // reset error
-        graph_exec.reset();
-      }
-    }
-    if (graph_exec == nullptr) {
-      graph_exec.instantiate(graph_);
-    }
     device_.make_current();
-    CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
+
+    if (!is_graph_updatable_) {
+      CudaGraphExec graph_exec;
+      graph_exec.instantiate(graph_);
+      CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
+    } else {
+      auto graph_key = graph_nodes_key_ + ":" + graph_deps_key_;
+      auto& graph_exec = graph_cache_[graph_key];
+
+      if (graph_exec != nullptr) {
+        cudaGraphExecUpdateResult update_result;
+#if CUDART_VERSION >= 12000
+        cudaGraphExecUpdateResultInfo info;
+        cudaGraphExecUpdate(graph_exec, graph_, &info);
+        update_result = info.result;
+#else
+        cudaGraphNode_t error_node;
+        cudaGraphExecUpdate(graph_exec, graph_, &error_node, &update_result);
+#endif // CUDART_VERSION >= 12000
+        if (update_result != cudaGraphExecUpdateSuccess) {
+          cudaGetLastError(); // reset error
+          graph_exec.reset();
+        }
+      }
+      if (graph_exec == nullptr) {
+        graph_exec.instantiate(graph_);
+      }
+
+      CHECK_CUDA_ERROR(cudaGraphLaunch(graph_exec, stream_));
+    }
+
+    // Save cuda graph to dot file
+    if (const char* filename = save_cuda_graphs_dot_file(); filename) {
+      static int count = 0;
+      auto path = fmt::format("{}_{}.dot", filename, ++count);
+      CHECK_CUDA_ERROR(cudaGraphDebugDotPrint(graph_, path.c_str(), 0));
+    }
 
     // Reset state
-    graph_node_count_ = 0;
-    empty_node_count_ = 0;
     from_nodes_.clear();
     to_nodes_.clear();
-    graph_key_.clear();
+    graph_deps_key_.clear();
+    graph_nodes_key_.clear();
     node_map_.clear();
     graph_ = CudaGraph(device_);
+    is_graph_updatable_ = true;
   }
 
   // Put completion handlers in a batch.
   worker_.commit(stream_);
   node_count_ = 0;
+  bytes_in_graph_ = 0;
 }
 
 void CommandEncoder::synchronize() {
-  cudaStreamSynchronize(stream_);
+  CHECK_CUDA_ERROR(cudaStreamSynchronize(stream_));
   auto p = std::make_shared<std::promise<void>>();
   std::future<void> f = p->get_future();
   add_completed_handler([p = std::move(p)]() { p->set_value(); });

mlx/backend/cuda/device.h

@@ -3,6 +3,7 @@
 #pragma once
 
 #include "mlx/array.h"
+#include "mlx/backend/cuda/allocator.h"
 #include "mlx/backend/cuda/lru_cache.h"
 #include "mlx/backend/cuda/worker.h"
 #include "mlx/stream.h"
@@ -83,7 +84,7 @@ class CommandEncoder {
   }
 
   void add_completed_handler(std::function<void()> task);
-  int get_num_ops();
+  bool needs_commit();
   void commit();
 
   Device& device() {
@@ -105,8 +106,9 @@ class CommandEncoder {
     cudaGraphNode_t node;
     // K = kernel
     // E = empty
-    // G = subgraph
-    char node_type;
+    // G* = subgraph (with metadata)
+    // Symbols ':', '-' are reserved as separators
+    std::string node_type;
     std::string id;
   };
 
@@ -118,18 +120,21 @@ class CommandEncoder {
   CudaGraph graph_;
   Worker worker_;
   char node_count_{0};
-  char graph_node_count_{0};
-  char empty_node_count_{0};
   bool in_concurrent_{false};
   std::vector<cudaGraphNode_t> from_nodes_;
   std::vector<cudaGraphNode_t> to_nodes_;
-  std::string graph_key_;
+  std::string graph_nodes_key_;
+  std::string graph_deps_key_;
   std::vector<GraphNode> concurrent_nodes_;
   std::vector<std::shared_ptr<array::Data>> temporaries_;
   LRUCache<std::string, CudaGraphExec> graph_cache_;
   std::vector<std::uintptr_t> active_deps_;
   std::vector<std::uintptr_t> active_outputs_;
   std::unordered_map<std::uintptr_t, GraphNode> node_map_;
+  size_t bytes_in_graph_{0};
+  bool is_graph_updatable_{true};
+  int max_ops_per_graph_;
+  int max_mb_per_graph_;
 };
 
 class Device {
@@ -165,6 +170,7 @@ class Device {
   int device_;
   int compute_capability_major_;
   int compute_capability_minor_;
+  std::string device_name_;
   cublasLtHandle_t lt_;
   cudnnHandle_t cudnn_;
   std::unordered_map<int, CommandEncoder> encoders_;

mlx/backend/cuda/device/unary_ops.cuh

@@ -2,6 +2,8 @@
 
 #pragma once
 
+#include <cuda_fp8.h>
+
 #include "mlx/backend/cuda/device/fp16_math.cuh"
 #include "mlx/backend/cuda/device/utils.cuh"
 
@@ -334,4 +336,17 @@ struct Tanh {
   }
 };
 
+struct ToFP8 {
+  template <typename T>
+  __device__ uint8_t operator()(T x) {
+    return __nv_fp8_e4m3(x).__x;
+  }
+};
+
+struct FromFP8 {
+  __device__ float operator()(uint8_t x) {
+    return float(*(__nv_fp8_e4m3*)(&x));
+  }
+};
+
 } // namespace mlx::core::cu

mlx/backend/cuda/distributed.cu

@@ -15,8 +15,10 @@ void AllReduce::eval_gpu(
   assert(inputs.size() == 1);
   assert(outputs.size() == 1);
 
-  auto set_input_output =
-      [s = stream()](const array& in, array& out) -> std::pair<array, array> {
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+  auto set_input_output = [&](const array& in,
+                              array& out) -> std::pair<array, array> {
     if (!in.flags().row_contiguous) {
       copy_gpu(in, out, CopyType::General, s);
       return {out, out};
@@ -24,19 +26,17 @@ void AllReduce::eval_gpu(
       out.copy_shared_buffer(in);
       return {in, out};
     } else {
-      out.set_data(allocator::malloc(out.nbytes()));
+      out.set_data(cu::malloc_async(out.nbytes(), encoder));
       return {in, out};
     }
   };
 
   auto [input, output] = set_input_output(inputs[0], outputs[0]);
 
-  auto& encoder = cu::get_command_encoder(stream());
   encoder.set_input_array(input);
   encoder.set_output_array(output);
 
   auto capture = encoder.capture_context();
-  auto& s = stream();
 
   switch (reduce_type_) {
     case Sum:
@@ -53,4 +53,69 @@ void AllReduce::eval_gpu(
           "Only all reduce sum, max, and min are supported.");
   }
 }
+
+void AllGather::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 1);
+  assert(outputs.size() == 1);
+
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+
+  auto ensure_contiguous = [&s, &encoder](const array& x) {
+    if (x.flags().row_contiguous) {
+      return x;
+    } else {
+      array x_copy = contiguous_copy_gpu(x, s);
+      encoder.add_temporary(x_copy);
+      return x_copy;
+    }
+  };
+
+  auto input = ensure_contiguous(inputs[0]);
+  outputs[0].set_data(cu::malloc_async(outputs[0].nbytes(), encoder));
+
+  encoder.set_input_array(input);
+  encoder.set_output_array(outputs[0]);
+
+  auto capture = encoder.capture_context();
+  distributed::detail::all_gather(group(), input, outputs[0], s);
+}
+
+void ReduceScatter::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  assert(inputs.size() == 1);
+  assert(outputs.size() == 1);
+
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+
+  auto ensure_contiguous = [&s, &encoder](const array& x) {
+    if (x.flags().row_contiguous) {
+      return x;
+    } else {
+      array x_copy = contiguous_copy_gpu(x, s);
+      encoder.add_temporary(x_copy);
+      return x_copy;
+    }
+  };
+
+  auto input = ensure_contiguous(inputs[0]);
+  outputs[0].set_data(cu::malloc_async(outputs[0].nbytes(), encoder));
+
+  encoder.set_input_array(input);
+  encoder.set_output_array(outputs[0]);
+
+  auto capture = encoder.capture_context();
+
+  switch (reduce_type_) {
+    case Sum:
+      distributed::detail::sum_scatter(group(), input, outputs[0], s);
+      break;
+    default:
+      throw std::runtime_error("Only sum scatter is supported. ");
+  }
+}
 } // namespace mlx::core::distributed

mlx/backend/cuda/eval.cpp

@@ -11,9 +11,6 @@
 
 namespace mlx::core::gpu {
 
-// Can be tuned with MLX_MAX_OPS_PER_BUFFER
-constexpr int default_max_nodes_per_graph = 20;
-
 bool is_available() {
   return true;
 }
@@ -53,8 +50,7 @@ void eval(array& arr) {
     encoder.add_temporary(s);
   }
 
-  if (encoder.get_num_ops() >=
-      env::max_ops_per_buffer(default_max_nodes_per_graph)) {
+  if (encoder.needs_commit()) {
     scheduler::notify_new_task(stream);
     encoder.add_completed_handler(
         [stream]() { scheduler::notify_task_completion(stream); });

mlx/backend/cuda/event.cu

@@ -305,6 +305,7 @@ void Event::wait() {
   } else {
     event->atomic->wait(value());
   }
+  CHECK_CUDA_ERROR(cudaPeekAtLastError());
 }
 
 void Event::wait(Stream s) {

mlx/backend/cuda/fence.cpp

@@ -1,6 +1,8 @@
 // Copyright © 2025 Apple Inc.
 
 #include "mlx/fence.h"
+#include "mlx/backend/cuda/allocator.h"
+#include "mlx/backend/cuda/device.h"
 #include "mlx/backend/cuda/event.h"
 
 namespace mlx::core {
@@ -20,8 +22,24 @@ void Fence::wait(Stream s, const array&) {
   fence->event.wait(fence->count);
 }
 
-void Fence::update(Stream s, const array&) {
+void Fence::update(Stream s, const array& a, bool cross_device) {
   auto* fence = static_cast<FenceImpl*>(fence_.get());
+  if (cross_device) {
+    // Move to managed memory if there is a device switch
+    auto& cbuf =
+        *static_cast<cu::CudaBuffer*>(const_cast<array&>(a).buffer().ptr());
+    if (cbuf.device != -1) {
+      void* new_data;
+      CHECK_CUDA_ERROR(cudaMallocManaged(&new_data, cbuf.size));
+      cbuf.device = -1;
+      auto& encoder = cu::device(s.device).get_command_encoder(s);
+      encoder.commit();
+      CHECK_CUDA_ERROR(cudaMemcpyAsync(
+          new_data, cbuf.data, cbuf.size, cudaMemcpyDefault, encoder.stream()));
+      CHECK_CUDA_ERROR(cudaFreeAsync(cbuf.data, encoder.stream()));
+      cbuf.data = new_data;
+    }
+  }
   fence->count++;
   fence->event.signal(s, fence->count);
 }

mlx/backend/cuda/gemms/cublas_gemm.cpp

@@ -241,7 +241,7 @@ void CublasGemm::set_bias(cu::CommandEncoder& encoder, const array& bias) {
       CUBLASLT_MATMUL_DESC_EPILOGUE,
       &epilogue,
       sizeof(epilogue)));
-  auto* bias_ptr = bias.data<void>();
+  auto* bias_ptr = gpu_ptr<void>(bias);
   CHECK_CUBLAS_ERROR(cublasLtMatmulDescSetAttribute(
       matmul_desc_,
       CUBLASLT_MATMUL_DESC_BIAS_POINTER,
@@ -278,9 +278,9 @@ void CublasGemm::run(
 
   execute(
       encoder,
-      out.data<void>(),
-      a.data<void>(),
-      b.data<void>(),
+      gpu_ptr<void>(out),
+      gpu_ptr<void>(a),
+      gpu_ptr<void>(b),
       nullptr,
       alpha);
 }
@@ -321,10 +321,10 @@ void CublasGemm::run(
 
   execute(
       encoder,
-      out.data<void>(),
-      a.data<void>(),
-      b.data<void>(),
-      c.data<void>(),
+      gpu_ptr<void>(out),
+      gpu_ptr<void>(a),
+      gpu_ptr<void>(b),
+      gpu_ptr<void>(c),
       alpha,
       beta);
 }
@@ -370,11 +370,11 @@ void CublasGemm::execute(
     // Ensure workspace is 256-byte aligned
     int nbytes = cuda::ceil_div(heuristic_.workspaceSize, 256) * 256;
     array workspace(
-        allocator::malloc(nbytes),
+        cu::malloc_async(nbytes, encoder),
         {static_cast<int>(heuristic_.workspaceSize)},
         int8);
     encoder.add_temporary(workspace);
-    workspace_ptr = workspace.data<void>();
+    workspace_ptr = gpu_ptr<void>(workspace);
   }
 
   auto capture = encoder.capture_context();

mlx/backend/cuda/gemms/cublas_gemm_batched_12_0.cpp

@@ -25,9 +25,10 @@ void CublasGemm::run_batched(
   for (size_t i = 0; i < nbatch; ++i) {
     execute(
         encoder,
-        out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M_ * N_,
-        a.data<int8_t>() + a.itemsize() * a_it.loc,
-        b.data<int8_t>() + b.itemsize() * b_it.loc,
+        gpu_ptr<int8_t>(out) +
+            out.itemsize() * i * batch_shape.back() * M_ * N_,
+        gpu_ptr<int8_t>(a) + a.itemsize() * a_it.loc,
+        gpu_ptr<int8_t>(b) + b.itemsize() * b_it.loc,
         nullptr,
         alpha);
     a_it.step();
@@ -60,10 +61,11 @@ void CublasGemm::run_batched(
   for (size_t i = 0; i < nbatch; ++i) {
     execute(
         encoder,
-        out.data<int8_t>() + out.itemsize() * i * batch_shape.back() * M_ * N_,
-        a.data<int8_t>() + a.itemsize() * a_it.loc,
-        b.data<int8_t>() + b.itemsize() * b_it.loc,
-        c.data<int8_t>() + c.itemsize() * c_it.loc,
+        gpu_ptr<int8_t>(out) +
+            out.itemsize() * i * batch_shape.back() * M_ * N_,
+        gpu_ptr<int8_t>(a) + a.itemsize() * a_it.loc,
+        gpu_ptr<int8_t>(b) + b.itemsize() * b_it.loc,
+        gpu_ptr<int8_t>(c) + c.itemsize() * c_it.loc,
         alpha,
         beta);
     a_it.step();

mlx/backend/cuda/gemms/cublas_gemm_batched_12_9.cu

@@ -163,7 +163,7 @@ void CublasGemm::run_batched(
 
   // Launch kernel to set device offsets
   auto pointers = array(
-      allocator::malloc(batch_count * sizeof(void*) * 3),
+      cu::malloc_async(batch_count * sizeof(void*) * 3, encoder),
       {batch_count * 3},
       uint64);
 
@@ -183,10 +183,10 @@ void CublasGemm::run_batched(
           num_blocks,
           block_dims,
           0,
-          pointers.data<int8_t*>(),
-          a.data<int8_t>(),
-          b.data<int8_t>(),
-          out.data<int8_t>(),
+          gpu_ptr<int8_t*>(pointers),
+          gpu_ptr<int8_t>(a),
+          gpu_ptr<int8_t>(b),
+          gpu_ptr<int8_t>(out),
           item_size,
           const_param<ndim_constant()>(batch_shape),
           const_param<ndim_constant()>(a_batch_strides),
@@ -200,10 +200,10 @@ void CublasGemm::run_batched(
         num_blocks,
         block_dims,
         0,
-        pointers.data<int8_t*>(),
-        a.data<int8_t>(),
-        b.data<int8_t>(),
-        out.data<int8_t>(),
+        gpu_ptr<int8_t*>(pointers),
+        gpu_ptr<int8_t>(a),
+        gpu_ptr<int8_t>(b),
+        gpu_ptr<int8_t>(out),
         item_size,
         const_param(batch_shape),
         const_param(a_batch_strides),
@@ -219,7 +219,7 @@ void CublasGemm::run_batched(
   encoder.set_input_array(b);
   encoder.set_output_array(out);
 
-  auto a_pointers = pointers.data<int8_t*>();
+  auto a_pointers = gpu_ptr<int8_t*>(pointers);
   auto b_pointers = a_pointers + batch_count;
   auto out_pointers = b_pointers + batch_count;
   execute(
@@ -251,7 +251,7 @@ void CublasGemm::run_batched(
 
   // Launch kernel to set device offsets
   auto pointers = array(
-      allocator::malloc(batch_count * sizeof(uint64_t) * 4),
+      cu::malloc_async(batch_count * sizeof(uint64_t) * 4, encoder),
       {batch_count * 4},
       uint64);
 
@@ -271,11 +271,11 @@ void CublasGemm::run_batched(
           num_blocks,
           block_dims,
           0,
-          pointers.data<int8_t*>(),
-          a.data<int8_t>(),
-          b.data<int8_t>(),
-          c.data<int8_t>(),
-          out.data<int8_t>(),
+          gpu_ptr<int8_t*>(pointers),
+          gpu_ptr<int8_t>(a),
+          gpu_ptr<int8_t>(b),
+          gpu_ptr<int8_t>(c),
+          gpu_ptr<int8_t>(out),
           item_size,
           const_param<ndim_constant()>(batch_shape),
           const_param<ndim_constant()>(a_batch_strides),
@@ -290,11 +290,11 @@ void CublasGemm::run_batched(
         num_blocks,
         block_dims,
         0,
-        pointers.data<int8_t*>(),
-        a.data<int8_t>(),
-        b.data<int8_t>(),
-        c.data<int8_t>(),
-        out.data<int8_t>(),
+        gpu_ptr<int8_t*>(pointers),
+        gpu_ptr<int8_t>(a),
+        gpu_ptr<int8_t>(b),
+        gpu_ptr<int8_t>(c),
+        gpu_ptr<int8_t>(out),
         item_size,
         const_param(batch_shape),
         const_param(a_batch_strides),
@@ -312,7 +312,7 @@ void CublasGemm::run_batched(
   encoder.set_input_array(c);
   encoder.set_output_array(out);
 
-  auto a_pointers = pointers.data<int8_t*>();
+  auto a_pointers = gpu_ptr<int8_t*>(pointers);
   auto b_pointers = a_pointers + batch_count;
   auto c_pointers = b_pointers + batch_count;
   auto out_pointers = c_pointers + batch_count;

mlx/backend/cuda/gemms/gemv.cu

@@ -149,13 +149,13 @@ void gemv(
     auto vec_strides = const_param(b_batch_strides);
 
     if (M == 1) {
-      mat = b.data<DataType>();
-      vec = a.data<DataType>();
+      mat = gpu_ptr<DataType>(b);
+      vec = gpu_ptr<DataType>(a);
       rows = N;
       std::swap(mat_strides, vec_strides);
     } else {
-      mat = a.data<DataType>();
-      vec = b.data<DataType>();
+      mat = gpu_ptr<DataType>(a);
+      vec = gpu_ptr<DataType>(b);
       rows = M;
     }
     uint32_t num_blocks_x = (rows + rows_per_block - 1) / rows_per_block;
@@ -177,7 +177,7 @@ void gemv(
             0,
             mat,
             vec,
-            out.data<DataType>(),
+            gpu_ptr<DataType>(out),
             rows,
             cols);
       } else {
@@ -189,7 +189,7 @@ void gemv(
             0,
             mat,
             vec,
-            out.data<DataType>(),
+            gpu_ptr<DataType>(out),
             rows,
             cols,
             const_param(batch_shape),

mlx/backend/cuda/indexing.cpp

@@ -31,7 +31,7 @@ void append_indices_arg(
     int idx_ndim) {
   SmallVector<const void*> indices(nidx);
   for (int i = 0; i < nidx; ++i) {
-    indices[i] = inputs[i + 1].data<void>();
+    indices[i] = gpu_ptr<void>(inputs[i + 1]);
   }
   args.append(std::move(indices));
   SmallVector<int32_t> indices_shape(nidx * idx_ndim);
@@ -59,7 +59,9 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
   assert(inputs.size() > 0);
   const auto& src = inputs[0];
 
-  out.set_data(allocator::malloc(out.nbytes()));
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   if (out.size() == 0) {
     return;
   }
@@ -80,7 +82,6 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
       dtype_to_string(idx_dtype),
       nidx);
 
-  auto& s = stream();
   cu::JitModule& mod = cu::get_jit_module(s.device, module_name, [&]() {
     std::vector<std::string> kernel_names;
     for (int ndim = 0; ndim <= MAX_NDIM; ++ndim) {
@@ -121,7 +122,6 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
       idx_ndim,
       large ? "int64_t" : "int32_t");
 
-  auto& encoder = cu::get_command_encoder(s);
   for (const auto& in : inputs) {
     encoder.set_input_array(in);
   }
@@ -239,7 +239,9 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
   const auto& src = inputs[0];
   const auto& idx = inputs[1];
 
-  out.set_data(allocator::malloc(out.nbytes()));
+  auto& s = stream();
+  auto& encoder = cu::get_command_encoder(s);
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   if (out.size() == 0) {
     return;
   }
@@ -251,7 +253,6 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
       dtype_to_string(out.dtype()),
       dtype_to_string(idx.dtype()));
 
-  auto& s = stream();
   cu::JitModule& mod = cu::get_jit_module(s.device, module_name, [&]() {
     std::vector<std::string> kernel_names;
     for (int ndim = 0; ndim <= MAX_NDIM; ++ndim) {
@@ -312,7 +313,6 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
       idx.flags().row_contiguous,
       large ? "int64_t" : "int32_t");
 
-  auto& encoder = cu::get_command_encoder(s);
   for (const auto& in : inputs) {
     encoder.set_input_array(in);
   }

mlx/backend/cuda/jit_module.cpp

@@ -279,11 +279,14 @@ void compile(
   // Compile program.
   std::vector<const char*> args;
   bool use_sass = compiler_supports_device_sass(device);
+  auto cc = device.compute_capability_major();
+  std::string arch_tag = (cc == 90 || cc == 100 || cc == 121) ? "a" : "";
   std::string compute = fmt::format(
-      "--gpu-architecture={}_{}{}",
+      "--gpu-architecture={}_{}{}{}",
       use_sass ? "sm" : "compute",
-      device.compute_capability_major(),
-      device.compute_capability_minor());
+      cc,
+      device.compute_capability_minor(),
+      arch_tag);
   args.push_back(compute.c_str());
   std::string cccl_include = cccl_dir();
   if (!cccl_include.empty()) {

mlx/backend/cuda/jit_module.h

@@ -31,7 +31,7 @@ struct KernelArgs {
   }
 
   void append(const array& a) {
-    append(reinterpret_cast<CUdeviceptr>(a.data<void>()));
+    append(reinterpret_cast<CUdeviceptr>(gpu_ptr<void>(a)));
   }
 
   template <typename T>

mlx/backend/cuda/kernel_utils.cuh

@@ -9,6 +9,7 @@
 #include <type_traits>
 
 #include "mlx/array.h"
+#include "mlx/backend/cuda/allocator.h"
 #include "mlx/backend/cuda/device/utils.cuh"
 
 #include <cuda.h>

mlx/backend/cuda/layer_norm.cu

@@ -230,9 +230,10 @@ void LayerNorm::eval_gpu(
   nvtx3::scoped_range r("LayerNorm::eval_gpu");
   auto& s = stream();
   auto& out = outputs[0];
+  auto& encoder = cu::get_command_encoder(s);
 
   // Make sure that the last dimension is contiguous.
-  auto set_output = [&s, &out](const array& x) {
+  auto set_output = [&s, &out, &encoder](const array& x) {
     bool no_copy = x.flags().contiguous && x.strides()[x.ndim() - 1] == 1;
     if (no_copy && x.ndim() > 1) {
       auto s = x.strides()[x.ndim() - 2];
@@ -243,7 +244,7 @@ void LayerNorm::eval_gpu(
         out.copy_shared_buffer(x);
       } else {
         out.set_data(
-            allocator::malloc(x.data_size() * x.itemsize()),
+            cu::malloc_async(x.data_size() * x.itemsize(), encoder),
             x.data_size(),
             x.strides(),
             x.flags());
@@ -265,7 +266,6 @@ void LayerNorm::eval_gpu(
   int64_t w_stride = (w.ndim() == 1) ? w.strides()[0] : 0;
   int64_t b_stride = (b.ndim() == 1) ? b.strides()[0] : 0;
 
-  auto& encoder = cu::get_command_encoder(s);
   encoder.set_input_array(x);
   encoder.set_input_array(w);
   encoder.set_input_array(b);
@@ -280,10 +280,10 @@ void LayerNorm::eval_gpu(
           n_rows,
           block_dim(),
           0,
-          x.data<DataType>(),
-          w.data<DataType>(),
-          b.data<DataType>(),
-          out.data<DataType>(),
+          gpu_ptr<DataType>(x),
+          gpu_ptr<DataType>(w),
+          gpu_ptr<DataType>(b),
+          gpu_ptr<DataType>(out),
           eps_,
           axis_size,
           w_stride,
@@ -335,7 +335,7 @@ void LayerNormVJP::eval_gpu(
     gx.copy_shared_buffer(g);
     g_in_gx = true;
   } else {
-    gx.set_data(allocator::malloc(gx.nbytes()));
+    gx.set_data(cu::malloc_async(gx.nbytes(), encoder));
   }
   if (g_copied && !g_in_gx) {
     encoder.add_temporary(g);
@@ -355,7 +355,7 @@ void LayerNormVJP::eval_gpu(
       g_in_gw = true;
       gw_temp.copy_shared_buffer(g);
     } else {
-      gw_temp.set_data(allocator::malloc(gw_temp.nbytes()));
+      gw_temp.set_data(cu::malloc_async(gw_temp.nbytes(), encoder));
       encoder.add_temporary(gw_temp);
     }
   }
@@ -393,11 +393,11 @@ void LayerNormVJP::eval_gpu(
                 n_rows,
                 block_dim(),
                 0,
-                x.data<DataType>(),
-                w.data<DataType>(),
-                g.data<DataType>(),
-                gx.data<DataType>(),
-                gw_temp.data<DataType>(),
+                gpu_ptr<DataType>(x),
+                gpu_ptr<DataType>(w),
+                gpu_ptr<DataType>(g),
+                gpu_ptr<DataType>(gx),
+                gpu_ptr<DataType>(gw_temp),
                 eps_,
                 axis_size,
                 w_stride);

mlx/backend/cuda/load.cpp

@@ -0,0 +1,60 @@
+// Copyright © 2023 Apple Inc.
+
+#include <algorithm>
+#include <utility>
+
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/utils.h"
+#include "mlx/primitives.h"
+
+namespace {
+
+template <const uint8_t scalar_size>
+void swap_endianness(uint8_t* data_bytes, size_t N) {
+  struct Elem {
+    uint8_t bytes[scalar_size];
+  };
+
+  Elem* data = reinterpret_cast<Elem*>(data_bytes);
+
+  for (size_t i = 0; i < N; i++) {
+    for (size_t j = 0; j < (scalar_size / 2); j++) {
+      std::swap(data[i].bytes[j], data[i].bytes[scalar_size - j - 1]);
+    }
+  }
+}
+
+} // namespace
+
+namespace mlx::core {
+
+void Load::eval_gpu(const std::vector<array>& inputs, array& out) {
+  auto& encoder = cu::get_command_encoder(stream());
+  auto size = out.size();
+  auto nbytes = size * out.itemsize();
+  out.set_data(cu::malloc_async(nbytes, encoder));
+  auto out_ptr = malloc(nbytes);
+  reader_->read(static_cast<char*>(out_ptr), nbytes, offset_);
+  if (swap_endianness_) {
+    switch (out.itemsize()) {
+      case 2:
+        swap_endianness<2>(reinterpret_cast<uint8_t*>(out_ptr), size);
+        break;
+      case 4:
+        swap_endianness<4>(reinterpret_cast<uint8_t*>(out_ptr), size);
+        break;
+      case 8:
+        swap_endianness<8>(reinterpret_cast<uint8_t*>(out_ptr), size);
+        break;
+    }
+  }
+  CHECK_CUDA_ERROR(cudaMemcpyAsync(
+      gpu_ptr<void>(out),
+      out_ptr,
+      nbytes,
+      cudaMemcpyDefault,
+      encoder.stream()));
+  CHECK_CUDA_ERROR(cudaLaunchHostFunc(encoder.stream(), free, out_ptr));
+}
+
+} // namespace mlx::core

mlx/backend/cuda/logsumexp.cu

@@ -115,7 +115,7 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   auto in = ensure_contiguous(inputs[0]);
   if (in.flags().row_contiguous) {
-    out.set_data(allocator::malloc(out.nbytes()));
+    out.set_data(cu::malloc_async(out.nbytes(), encoder));
   } else {
     auto n = in.shape(-1);
     auto flags = in.flags();
@@ -130,7 +130,7 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
     }
     flags.col_contiguous = col_contig;
     out.set_data(
-        allocator::malloc(in.nbytes() / n),
+        cu::malloc_async(in.nbytes() / n, encoder),
         in.data_size() / n,
         std::move(strides),
         flags);
@@ -151,8 +151,8 @@ void LogSumExp::eval_gpu(const std::vector<array>& inputs, array& out) {
           n_rows,
           block_dim(),
           0,
-          in.data<DataType>(),
-          out.data<DataType>(),
+          gpu_ptr<DataType>(in),
+          gpu_ptr<DataType>(out),
           axis_size);
     });
   });

mlx/backend/cuda/lru_cache.h

@@ -135,12 +135,19 @@ class LRUCache {
 };
 
 // Turn a POD struct into a container key by doing bytes compare.
+//
+// Usage:
+//   BytesKey<MyKey> key;
+//   key.pod = { ... };
 template <typename T>
 struct BytesKey {
   T pod;
   static_assert(std::is_standard_layout_v<T>, "T is not POD");
 
-  BytesKey(T pod) : pod(std::move(pod)) {}
+  BytesKey() {
+    // Make sure the paddings between members are filled with 0.
+    memset(&pod, 0, sizeof(T));
+  }
 
   BytesKey(const BytesKey& other) {
     memcpy(&pod, &other.pod, sizeof(T));

mlx/backend/cuda/matmul.cpp

@@ -121,7 +121,7 @@ void Matmul::eval_gpu(const std::vector<array>& inputs, array& out) {
     return;
   }
 
-  out.set_data(allocator::malloc(out.nbytes()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
 
   int M = a_pre.shape(-2);
   int N = b_pre.shape(-1);
@@ -163,7 +163,7 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   if (beta_ == 1 && a.dtype() != complex64 && c.strides(-1) == 1 &&
       c.data_size() == out.shape(-1)) {
-    out.set_data(allocator::malloc(out.nbytes()));
+    out.set_data(cu::malloc_async(out.nbytes(), encoder));
     gemm_and_bias(
         encoder,
         M,
@@ -187,10 +187,10 @@ void AddMM::eval_gpu(const std::vector<array>& inputs, array& out) {
     auto sty = c.strides()[c.ndim() - 1];
     if (sty == 1 && stx == c.shape(-1)) {
       ldc = stx;
-      out.set_data(allocator::malloc(out.nbytes()));
+      out.set_data(cu::malloc_async(out.nbytes(), encoder));
     } else if (sty == 1 && stx == 0) {
       ldc = 0;
-      out.set_data(allocator::malloc(out.nbytes()));
+      out.set_data(cu::malloc_async(out.nbytes(), encoder));
     } else {
       // Copy C into out and set C to out
       ldc = c.shape(-1);

mlx/backend/cuda/primitives.cpp

@@ -28,7 +28,6 @@ NO_GPU(FFT)
 NO_GPU(GatherMM)
 NO_GPU(GatherQMM)
 NO_GPU(Hadamard)
-NO_GPU(Load)
 NO_GPU_MULTI(LUF)
 NO_GPU_MULTI(QRF)
 NO_GPU(QuantizedMatmul)
@@ -38,9 +37,9 @@ NO_GPU(Inverse)
 NO_GPU(Cholesky)
 NO_GPU_MULTI(Eig)
 NO_GPU_MULTI(Eigh)
+NO_GPU(MaskedScatter)
 
 namespace distributed {
-NO_GPU_MULTI(AllGather)
 NO_GPU_MULTI(Send)
 NO_GPU_MULTI(Recv)
 } // namespace distributed

mlx/backend/cuda/quantized/affine_quantize.cu

@@ -262,10 +262,10 @@ void affine_quantize(
             num_blocks,
             block_dims,
             0,
-            w.data<T>(),
-            wq.data<uint8_t>(),
-            scales.data<T>(),
-            biases.data<T>(),
+            gpu_ptr<T>(w),
+            gpu_ptr<uint8_t>(wq),
+            gpu_ptr<T>(scales),
+            gpu_ptr<T>(biases),
             w.size());
       });
     });
@@ -306,7 +306,7 @@ void affine_dequantize(
   enc.set_input_array(scales);
   enc.set_input_array(biases);
   enc.set_output_array(w);
-  dispatch_float_types(w.dtype(), "affine_quantize", [&](auto type_tag) {
+  dispatch_float_types(w.dtype(), "affine_dequantize", [&](auto type_tag) {
     dispatch_groups(group_size_, [&](auto group_size) {
       dispatch_bits(bits_, [&](auto bits) {
         using T = cuda_type_t<MLX_GET_TYPE(type_tag)>;
@@ -318,10 +318,10 @@ void affine_dequantize(
             num_blocks,
             block_dims,
             0,
-            wq.data<uint8_t>(),
-            scales.data<T>(),
-            biases.data<T>(),
-            w.data<T>(),
+            gpu_ptr<uint8_t>(wq),
+            gpu_ptr<T>(scales),
+            gpu_ptr<T>(biases),
+            gpu_ptr<T>(w),
             w.size());
       });
     });

mlx/backend/cuda/quantized/convert_fp8.cu

@@ -0,0 +1,19 @@
+// Copyright © 2025 Apple Inc.
+#include "mlx/backend/cuda/unary/unary.cuh"
+#include "mlx/fast_primitives.h"
+
+namespace mlx::core {
+void fast::ConvertFP8::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  nvtx3::scoped_range r("ConvertFP8::eval_gpu");
+  auto& in = inputs[0];
+  auto& out = outputs[0];
+  auto& s = out.primitive().stream();
+  if (to_fp8_) {
+    unary_op_gpu<cu::ToFP8>(inputs, out, name(), s);
+  } else {
+    unary_op_gpu<cu::FromFP8>(inputs, out, name(), s);
+  }
+}
+} // namespace mlx::core

mlx/backend/cuda/quantized/cuda_fp4.h

@@ -0,0 +1,83 @@
+#pragma once
+
+struct __nv_fp8_e8m0 {
+  __device__ __nv_fp8_e8m0(float x) {
+    if (!std::isfinite(x)) {
+      __x = 0xFF;
+      return;
+    }
+    if (x < 0.0f) {
+      __x = 0x00;
+      return;
+    }
+    float le = std::log2f(x);
+    int n = static_cast<int>(std::nearbyintf(le));
+
+    n = n < -127 ? -127 : n;
+    n = n > 127 ? 127 : n;
+    __x = static_cast<uint8_t>(n + 127);
+  }
+
+  __device__ operator float() {
+    if (__x == 0xFF) {
+      return std::numeric_limits<float>::quiet_NaN();
+    }
+    return std::ldexp(1.0f, static_cast<int>(__x) - 127);
+  }
+
+  uint8_t __x{0};
+};
+
+struct __nv_fp4_e2m1 {
+  __device__ __nv_fp4_e2m1(float x) {
+    if (std::isnan(x)) {
+      __x = 0x7;
+      return;
+    }
+
+    const uint8_t sign_bit = (std::signbit(x)) ? 0x8 : 0x0;
+    x = std::abs(x);
+
+    if (x > 5.0f) {
+      __x = 0x7;
+    } else if (x >= 3.5f) {
+      __x = 0x6;
+    } else if (x > 2.5f) {
+      __x = 0x5;
+    } else if (x >= 1.75f) {
+      __x = 0x4;
+    } else if (x > 1.25f) {
+      __x = 0x3;
+    } else if (x >= 0.75f) {
+      __x = 0x2;
+    } else if (x > 0.25f) {
+      __x = 0x1;
+    } else {
+      __x = 0x0;
+    }
+    __x |= sign_bit;
+  }
+
+  __device__ operator float() {
+    static const float 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};
+
+    return LUT[__x];
+  }
+  uint8_t __x{0};
+};