Add debug line info

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@3.10
-            brew install doxygen
-            python3.10 -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            pip install --upgrade cmake
-            pip install -r docs/requirements.txt
-            pip install . -v
-      - when:
-          condition:
-            not: << parameters.upload-docs >>
-          steps:
-            - run:
-               name: Build documentation
-               command: |
-                 source env/bin/activate
-                 cd docs && doxygen && make html O=-W
-      - when:
-          condition: << parameters.upload-docs >>
-          steps:
-            - add_ssh_keys:
-                fingerprints:
-                  - "SHA256:OhcVVMovbT0pkgMeiVRyxMnjV9R2t+hKBsNcuxq9h+0"
-            - run:
-               name: Upload documentation
-               command: |
-                 source env/bin/activate
-                 git config user.email "mlx@group.apple.com"
-                 git config user.name "CircleCI Docs"
-                 git checkout gh-pages
-                 git rebase main
-                 cd docs
-                 git rm -rf build/html
-                 doxygen && make html O=-W
-                 git add -f build/html
-                 git commit -m "rebase"
-                 git push -f origin gh-pages
-
-  linux_build_and_test:
-    machine:
-      image: ubuntu-2204:current
-      resource_class: large
-    steps:
-      - checkout
-      - run:
-          name: Run style checks
-          command: |
-            pip install pre-commit
-            pre-commit run --all
-            if ! git diff --quiet; then echo 'Style checks failed, please install pre-commit and run pre-commit run --all and push the change'; exit 1; fi
-      - run:
-          name: Install dependencies
-          command: |
-            export DEBIAN_FRONTEND=noninteractive
-            export NEEDRESTART_MODE=a
-            sudo apt-get update
-            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install openmpi-bin openmpi-common libopenmpi-dev
-            curl -LsSf https://astral.sh/uv/install.sh | sh
-      - run:
-          name: Install Python package
-          command: |
-            uv venv
-            uv pip install cmake
-            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              uv pip install -e ".[dev]" -v
-      - run:
-          name: Generate package stubs
-          command: |
-            uv pip install typing_extensions
-            uv run --no-project setup.py generate_stubs
-      - run:
-          name: Run Python tests
-          command: |
-            source .venv/bin/activate
-            python -m unittest discover python/tests -v
-            mpirun --bind-to none -host localhost:8 -np 8 python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
-            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
-      - run:
-          name: Build CPP only
-          command: |
-            source .venv/bin/activate
-            mkdir -p build && cd build
-            cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
-            make -j `nproc`
-      - run:
-          name: Run CPP tests
-          command: ./build/tests/tests
-
-  mac_build_and_test:
-    parameters:
-      xcode_version:
-        type: string
-        default: "26.0.0"
-      macosx_deployment_target:
-        type: string
-        default: ""
-    macos:
-      xcode: << parameters.xcode_version >>
-    environment:
-      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    resource_class: m4pro.medium
-    steps:
-      - checkout
-      - run:
-          name: Install dependencies
-          command: |
-            xcodebuild -downloadComponent MetalToolchain
-            HOMEBREW_NO_AUTO_UPDATE=1 HOMEBREW_NO_INSTALL_CLEANUP=1 \
-              brew install openmpi uv
-      - run:
-          name: Install Python package
-          command: |
-            uv venv --python 3.10
-            uv pip install \
-              nanobind==2.4.0 \
-              cmake \
-              numpy \
-              torch \
-              tensorflow \
-              unittest-xml-reporting
-            DEBUG=1 CMAKE_ARGS="-DCMAKE_COMPILE_WARNING_AS_ERROR=ON" \
-              uv pip install -e . -v
-      - run:
-          name: Generate package stubs
-          command: |
-            uv pip install typing_extensions
-            uv run --no-project setup.py generate_stubs
-      - run:
-          name: Run Python tests
-          command: |
-            source .venv/bin/activate
-            LOW_MEMORY=1 DEVICE=cpu python -m xmlrunner discover -v python/tests -o test-results/cpu
-            LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 python -m xmlrunner discover -v python/tests -o test-results/gpu
-            mpirun --bind-to none -host localhost:8 -np 8 -x DYLD_LIBRARY_PATH=/opt/homebrew/lib/ python python/tests/mpi_test_distributed.py
-            mlx.launch --verbose -n 8 python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
-            if $(grep "\[WARN\]" stderr.log); then echo "Distributed ring test failed"; exit 1; fi
-      - run:
-          name: Build example extension
-          command: |
-            source .venv/bin/activate
-            cd examples/extensions
-            uv pip install -r requirements.txt
-            uv run --no-project setup.py build_ext --inplace
-            uv run --no-project python test.py
-      - store_test_results:
-          path: test-results
-      - run:
-          name: Build CPP only
-          command: |
-            source .venv/bin/activate
-            mkdir -p build && cd build && cmake .. && make -j `sysctl -n hw.ncpu`
-      - run:
-          name: Run CPP tests
-          command: |
-            DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 ./build/tests/tests
-      - run:
-          name: Build small binary
-          command: |
-            source .venv/bin/activate
-            cd build/
-            cmake .. -DCMAKE_BUILD_TYPE=MinSizeRel \
-              -DBUILD_SHARED_LIBS=ON \
-              -DMLX_BUILD_CPU=OFF \
-              -DMLX_BUILD_SAFETENSORS=OFF \
-              -DMLX_BUILD_GGUF=OFF \
-              -DMLX_METAL_JIT=ON
-            make -j `sysctl -n hw.ncpu`
-      - run:
-          name: Run Python tests with JIT
-          command: |
-            CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-              uv pip install -e . -v
-            LOW_MEMORY=1 DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 \
-              METAL_DEBUG_ERROR_MODE=0 \
-              uv run --no-project python -m xmlrunner discover \
-                -v python/tests \
-                -o test-results/gpu_jit
-
-  cuda_build_and_test:
-    parameters:
-      image_date:
-        type: string
-        default: "2023.11.1"
-    machine:
-      image: "linux-cuda-12:<< parameters.image_date >>"
-      resource_class: gpu.nvidia.small.gen2
-    steps:
-      - checkout
-      - restore_cache:
-          keys:
-            - cuda-<< parameters.image_date >>-{{ arch }}-
-      - run:
-          name: Install dependencies
-          command: |
-            sudo apt-get update
-            sudo apt-get install libcudnn9-dev-cuda-12
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install libnccl2 libnccl-dev
-            curl -sL https://github.com/ccache/ccache/releases/download/v4.11.3/ccache-4.11.3-linux-x86_64.tar.xz | tar xJf -
-            sudo mv ccache-4.11.3-linux-x86_64/ccache /usr/bin/ccache
-            rm -rf ccache-4.11.3-linux-x86_64
-            curl -LsSf https://astral.sh/uv/install.sh | sh
-      - run:
-          name: Set CCache size
-          command: ccache --max-size 1G
-      - run:
-          name: Install Python package
-          command: |
-            uv venv
-            uv pip install cmake
-            DEBUG=1 CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_COMPILE_WARNING_AS_ERROR=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-              uv pip install -e ".[dev]" -v
-      - run:
-          name: Run Python tests
-          command: |
-            source .venv/bin/activate
-            LOW_MEMORY=1 DEVICE=cpu python -m unittest discover python/tests -v
-            LOW_MEMORY=1 DEVICE=gpu python -m tests discover python/tests -v
-      - run:
-          name: Build CPP only
-          command: |
-            source .venv/bin/activate
-            cmake . -B build \
-              -DMLX_BUILD_CUDA=ON \
-              -DCMAKE_CUDA_COMPILER=`which nvcc` \
-              -DCMAKE_BUILD_TYPE=DEBUG
-            cmake --build build -j `nproc`
-      - run:
-          name: Run CPP tests
-          command: ./build/tests/tests -sfe="*fft_tests.cpp,*linalg_tests.cpp"
-      - run:
-          name: CCache report
-          command: |
-            ccache --show-stats
-            ccache --zero-stats
-            ccache --cleanup
-      - save_cache:
-          key: cuda-<< parameters.image_date >>-{{ arch }}-{{ epoch }}
-          paths:
-            - /home/circleci/.cache/ccache
-
-  build_release:
-    parameters:
-      python_version:
-        type: string
-        default: "3.10"
-      xcode_version:
-        type: string
-        default: "26.0.0"
-      build_env:
-        type: string
-        default: ""
-      macosx_deployment_target:
-        type: string
-        default: ""
-    macos:
-      xcode: << parameters.xcode_version >>
-    resource_class: m4pro.medium
-    environment:
-      MACOSX_DEPLOYMENT_TARGET: << parameters.macosx_deployment_target >>
-    steps:
-      - checkout
-      - run:
-          name: Install dependencies
-          command: |
-            xcodebuild -downloadComponent MetalToolchain
-            mkdir -p ~/miniconda3
-            curl https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-arm64.sh -o ~/miniconda3/miniconda.sh
-            bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
-            rm ~/miniconda3/miniconda.sh
-            source ~/miniconda3/bin/activate
-            conda init --all
-            conda create -n env python=<< parameters.python_version >> -y
-            conda activate env
-            pip install --upgrade cmake
-            pip install nanobind==2.4.0
-            pip install --upgrade setuptools
-            pip install numpy
-            pip install twine
-            pip install build
-      - run:
-          name: Install Python package
-          command: |
-            conda activate env
-            env -u MACOSX_DEPLOYMENT_TARGET DEV_RELEASE=1 \
-              pip install . -v
-      - run:
-          name: Generate package stubs
-          command: |
-            conda activate env
-            pip install typing_extensions
-            python setup.py generate_stubs
-      - run:
-          name: Build Python package
-          command: |
-            conda activate env
-            python setup.py clean --all
-            << parameters.build_env >> MLX_BUILD_STAGE=1 python -m build -w
-      - when:
-          condition:
-            equal: ["3.10", << parameters.python_version >>]
-          steps:
-            - run:
-                name: Build common package
-                command: |
-                  conda activate env
-                  python setup.py clean --all
-                  << parameters.build_env >> MLX_BUILD_STAGE=2 python -m build -w
-      - when:
-          condition: << parameters.build_env >>
-          steps:
-            - run:
-                name: Upload package
-                command: |
-                  conda activate env
-                  twine upload dist/*
-      - store_artifacts:
-          path: dist/
-
-  build_linux_release:
-    parameters:
-      python_version:
-        type: string
-        default: "3.10"
-      build_env:
-        type: string
-        default: ""
-    machine:
-      image: ubuntu-2204:current
-      resource_class: large
-    steps:
-      - checkout
-      - run:
-          name: Build wheel
-          command: |
-            PYTHON=python<< parameters.python_version >>
-            export DEBIAN_FRONTEND=noninteractive
-            export NEEDRESTART_MODE=a
-            sudo apt-get update
-            TZ=Etc/UTC sudo apt-get -y install tzdata
-            sudo add-apt-repository -y ppa:deadsnakes/ppa
-            sudo apt-get install -y $PYTHON $PYTHON-dev $PYTHON-full
-            sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev
-            $PYTHON -m venv env
-            source env/bin/activate
-            pip install --upgrade pip
-            pip install --upgrade cmake
-            pip install auditwheel
-            pip install patchelf
-            pip install build
-            pip install twine
-            << parameters.build_env >> pip install ".[dev]" -v
-            pip install typing_extensions
-            python setup.py generate_stubs
-            python setup.py clean --all
-            MLX_BUILD_STAGE=1 << parameters.build_env >> python -m build -w
-            bash python/scripts/repair_linux.sh
-      - when:
-          condition:
-            equal: ["3.10", << parameters.python_version >>]
-          steps:
-            - run:
-                name: Build common package
-                command: |
-                  source env/bin/activate
-                  python setup.py clean --all
-                  << parameters.build_env >> MLX_BUILD_STAGE=2 \
-                    python -m build -w
-                  auditwheel repair dist/mlx_cpu*.whl --plat manylinux_2_35_x86_64
-      - when:
-          condition: << parameters.build_env >>
-          steps:
-            - run:
-                name: Upload packages
-                command: |
-                  source env/bin/activate
-                  twine upload wheelhouse/*.whl
-      - store_artifacts:
-          path: wheelhouse/
-
-  build_cuda_release:
-    parameters:
-      build_env:
-        type: string
-        default: ""
-    machine:
-      image: ubuntu-2204:current
-      resource_class: xlarge
-    steps:
-      - checkout
-      - run:
-          name: Build wheel
-          command: |
-            export DEBIAN_FRONTEND=noninteractive
-            export NEEDRESTART_MODE=a
-            wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
-            sudo dpkg -i cuda-keyring_1.1-1_all.deb
-            sudo apt-get update
-            sudo apt-get install cuda-toolkit-12-9 libcudnn9-dev-cuda-12
-            sudo apt-get install libblas-dev liblapack-dev liblapacke-dev
-            sudo apt-get install zip
-            pip install auditwheel
-            pip install patchelf
-            pip install build
-            pip install twine
-            export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}
-            export LD_LIBRARY_PATH=/usr/local/cuda/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
-            << parameters.build_env >> MLX_BUILD_STAGE=2 \
-              CMAKE_ARGS="-DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=`which nvcc`" \
-              python -m build -w
-            bash python/scripts/repair_cuda.sh
-      - when:
-          condition: << parameters.build_env >>
-          steps:
-            - run:
-                name: Upload package
-                command: |
-                  twine upload wheelhouse/*.whl
-      - store_artifacts:
-          path: wheelhouse/
-
-workflows:
-  build_and_test:
-    when:
-      and:
-        - matches:
-            pattern: "^(?!pull/)[-\\w]+$"
-            value: << pipeline.git.branch >>
-        - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.test_release >>
-    jobs:
-      - mac_build_and_test:
-          matrix:
-            parameters:
-              macosx_deployment_target: ["13.5", "15.0"]
-      - linux_build_and_test
-      - cuda_build_and_test:
-          matrix:
-            parameters:
-              image_date: ["2023.11.1", "2025.05.1"]
-      - build_documentation 
-
-  build_pypi_release:
-    when:
-      and:
-        - not: << pipeline.parameters.nightly_build >>
-        - not: << pipeline.parameters.test_release >>
-    jobs:
-      - build_release:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          matrix:
-            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              build_env: ["PYPI_RELEASE=1"]
-              xcode_version: ["26.0.0"]
-      - build_documentation:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          upload-docs: true
-      - build_linux_release:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          matrix:
-            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-              build_env: ["PYPI_RELEASE=1"]
-      - build_cuda_release:
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-          matrix:
-            parameters:
-              build_env: ["PYPI_RELEASE=1"]
-
-  prb:
-    when:
-      matches:
-        pattern: "^pull/\\d+(/head)?$"
-        value: << pipeline.git.branch >>
-    jobs:
-      - hold:
-          type: approval
-      - apple/authenticate:
-          context: pr-approval
-      - mac_build_and_test:
-          requires: [ hold ]
-          matrix:
-            parameters:
-              macosx_deployment_target: ["13.5", "15.0"]
-      - linux_build_and_test:
-          requires: [ hold ]
-      - cuda_build_and_test:
-          requires: [ hold ]
-          matrix:
-            parameters:
-              image_date: ["2023.11.1", "2025.05.1"]
-  nightly_build:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.nightly_build >>
-    jobs:
-      - build_release:
-          matrix:
-            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              xcode_version: ["26.0.0"]
-      - build_linux_release:
-          matrix:
-            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-      - build_cuda_release
-
-  build_dev_release:
-    when:
-      and:
-        - equal: [ main, << pipeline.git.branch >> ]
-        - << pipeline.parameters.test_release >>
-    jobs:
-      - build_release:
-          matrix:
-            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-              macosx_deployment_target: ["13.5", "14.0", "15.0"]
-              build_env: ["DEV_RELEASE=1"]
-              xcode_version: ["26.0.0"]
-      - build_linux_release:
-          matrix:
-            parameters:
-              python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-              build_env: ["DEV_RELEASE=1"]
-      - build_cuda_release:
-          matrix:
-            parameters:
-              build_env: ["DEV_RELEASE=1"]

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

@@ -1,24 +1,15 @@
 name: 'Build CUDA wheel'
 description: 'Build CUDA wheel'
 
-inputs:
-  nvcc-location:
-    description: 'Location of nvcc compiler'
-    required: true
-
 runs:
   using: "composite"
   steps:
     - name: Build package
       shell: bash
       env:
-        MLX_BUILD_STAGE: 2
-        CMAKE_ARGS: -DMLX_BUILD_CUDA=ON -DCMAKE_CUDA_COMPILER=${{ inputs.nvcc-location }}
+        CMAKE_ARGS: -DMLX_BUILD_CUDA=ON
       run: |
         pip install auditwheel build patchelf setuptools
         python setup.py clean --all
-        python -m build -w
-
-        if [ -f "python/scripts/repair_cuda.sh" ]; then
-          bash python/scripts/repair_cuda.sh
-        fi
+        MLX_BUILD_STAGE=2 python -m build -w
+        bash python/scripts/repair_cuda.sh

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

@@ -1,68 +0,0 @@
-name: 'Build and Test with CUDA'
-description: 'Build and test MLX with CUDA'
-
-inputs:
-  build-type:
-    description: 'Build type (debug, release)'
-    required: false
-    default: 'debug'
-  run-tests:
-    description: 'Whether to run tests'
-    required: false
-    default: 'true'
-  nvcc-location:
-    description: 'Location of nvcc compiler'
-    required: true
-    default: '/usr/local/cuda-12.9/bin/nvcc'
-    # this value is dependent on the CUDA tools installed in the setup-linux workflow
-
-runs:
-  using: "composite"
-  steps:
-    - name: Install Python package
-      shell: bash
-      env:
-        DEBUG: 1
-        CMAKE_ARGS: -DMLX_BUILD_CUDA=ON -DCMAKE_COMPILE_WARNING_AS_ERROR=ON -DCMAKE_CUDA_COMPILER=${{ inputs.nvcc-location }}
-      run: pip install -e ".[dev]" -v
-
-    - name: Check if build actually worked
-      shell: bash
-      run: python -c "import mlx.core"
-
-    - name: Run Python tests - CPU
-      if: inputs.run-tests == 'true'
-      shell: bash
-      env:
-        LOW_MEMORY: 1
-        DEVICE: cpu
-      run: python -m unittest discover python/tests -v
-
-    - name: Run Python tests - GPU
-      if: inputs.run-tests == 'true'
-      shell: bash
-      env:
-        LOW_MEMORY: 1
-        DEVICE: gpu
-      run: python -m tests discover python/tests -v
-
-    - name: Build CPP only
-      if: inputs.build-type == 'debug'
-      shell: bash
-      run: |
-        cmake . -B build \
-          -DMLX_BUILD_CUDA=ON \
-          -DCMAKE_CUDA_COMPILER=${{ inputs.nvcc-location }} \
-          -DCMAKE_BUILD_TYPE=DEBUG
-        cmake --build build -j $(nproc)
-    
-    - name: Run CPP tests
-      if: ${{ inputs.build-type == 'debug' && inputs.run-tests == 'true' }}
-      shell: bash
-      run: ./build/tests/tests -sfe="*fft_tests.cpp,*linalg_tests.cpp"
-
-    - name: Build Python package
-      if: inputs.build-type == 'release'
-      uses: ./.github/actions/build-cuda-release
-      with:
-        nvcc-location: ${{ inputs.nvcc-location }}

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

@@ -1,19 +1,19 @@
 name: 'Build Documentation'
-description: 'Build documentation on a mac'
+description: 'Build documentation'
 
 runs:
   using: "composite"
   steps:
     - name: Setup machine
-      uses: ./.github/actions/setup-macos
+      uses: ./.github/actions/setup-linux
 
     - name: Install dependencies
-      shell: sh
+      shell: bash
       run: |
-        brew install doxygen
-        uv pip install --upgrade pip cmake
-        uv pip install -r docs/requirements.txt
-        uv pip install . -v
+        sudo apt-get install -y doxygen
+        source .venv/bin/activate
+        pip install -r docs/requirements.txt
+        pip install . -v
   
     - name: Build documentation
       shell: bash
@@ -24,8 +24,8 @@ runs:
         make html O=-W
     
     - name: Create artifact tar
-      shell: sh
-      run: tar -cf artifact.tar --cd docs/build/html -L .
+      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
@@ -35,4 +35,4 @@ runs:
         name: github-pages
         path: artifact.tar
         retention-days: 1
-        if-no-files-found: error
+        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

@@ -1,78 +1,41 @@
 name: 'Build and Test on Linux'
-description: 'Build and test MLX on Linux'
 
 inputs:
-  build-type:
-    description: 'Build type'
+  toolkit:
+    description: 'The toolkit to build with'
     required: false
-    default: 'debug'
-    type: choice
-    options:
-      - debug
-      - release
-  run-tests:
-    description: 'Whether to run tests'
-    required: false
-    default: 'true'
-    type: boolean
+    default: 'cpu'
 
 runs:
   using: "composite"
   steps:
-    - name: Set DEBUG
-      shell: sh
-      if: inputs.build-type == 'debug'
-      run: echo "DEBUG=1" >> $GITHUB_ENV
-
     - name: Install Python package
+      id: python_build
       shell: sh
       env:
-        CMAKE_ARGS: "-DCMAKE_COMPILE_WARNING_AS_ERROR=ON"
-      run: pip install -e ".[dev]" -v
-    
+        DEBUG: 1
+        CMAKE_ARGS: >-
+          -DCMAKE_COMPILE_WARNING_AS_ERROR=ON
+          -DMLX_BUILD_CUDA=${{ 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: Run Python tests
-      if: inputs.run-tests == 'true'
-      shell: bash
-      run: |
-        python -m unittest discover python/tests -v
-        mpirun --bind-to none --allow-run-as-root -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 -Fq '[WARN]' stderr.log ; then
-          grep -F '[WARN]' stderr.log
-          echo "Distributed ring test failed";
-          exit 1;
-        fi
-    
-    - name: Build CPP only
-      if: inputs.build-type == 'debug'
-      shell: bash
-      run: |
-        mkdir -p build && cd build
-        cmake .. -DMLX_BUILD_METAL=OFF -DCMAKE_BUILD_TYPE=DEBUG
-        make -j $(nproc)
-    
-    - name: Run CPP tests
-      if: ${{ inputs.build-type == 'debug' && inputs.run-tests == 'true' }}
-      shell: sh
-      run: ./build/tests/tests
-    
-    - name: Build Python package
-      if: inputs.build-type == 'release'
-      shell: bash
-      run: |
-        pip install auditwheel patchelf build
-        python setup.py clean --all
-        MLX_BUILD_STAGE=1 python -m build -w
-        if [ -f "python/scripts/repair_linux.sh" ]; then
-          bash python/scripts/repair_linux.sh
-        fi
 
-        python setup.py clean --all
-        MLX_BUILD_STAGE=2 python -m build -w
-        auditwheel repair dist/mlx_cpu*.whl --plat manylinux_2_35_x86_64
+    - 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

@@ -6,17 +6,29 @@ inputs:
     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(s)
-      shell: bash
+    - name: Build Python package
+      shell: bash -l {0}
       env:
         MACOSX_DEPLOYMENT_TARGET: ${{ inputs.macos-target }}
       run: |
-        uv pip install build
-        uv run --no-project setup.py clean --all
-        MLX_BUILD_STAGE=1 uv run -m build -w
-        uv run --no-project setup.py clean --all
-        MLX_BUILD_STAGE=2 uv run -m build -w
+        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

@@ -1,68 +1,51 @@
 name: 'Build and Test on macOS'
 description: 'Build and test MLX on macOS'
 
-inputs:
-  build-type:
-    description: 'Build type (debug, release)'
-    required: false
-    default: 'debug'
-    type: choice
-    options:
-      - debug
-      - release
-  run-tests:
-    description: 'Whether to run tests'
-    required: false
-    default: 'true'
-  build-jit:
-    description: 'Whether to build with JIT'
-    required: false
-    default: 'true'
-
 runs:
   using: "composite"
   steps:
     - name: Install dependencies
-      shell: sh
       env:
         DEBUG: 1
-        DEV_RELEASE: 1
+        CMAKE_ARGS: "-DCMAKE_COMPILE_WARNING_AS_ERROR=ON"
+      shell: bash -l {0}
       run: |
-        uv pip install --upgrade pip cmake setuptools
-        uv pip install nanobind==2.4.0 \
-          numpy torch tensorflow unittest-xml-reporting
-        uv pip install -e . -v
+        pip install --upgrade pip
+        pip install cmake setuptools nanobind==2.4.0
+        pip install -e . -v
 
     - name: Generate package stubs
-      shell: bash
+      shell: bash -l {0}
       run: |
-        uv pip install typing_extensions
-        uv run --no-project setup.py generate_stubs
+        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
-      if: inputs.run-tests == 'true'
-      shell: bash
+      shell: bash -l {0}
       env:
         LOW_MEMORY: 1
       run: |
-        DEVICE=cpu uv run -m xmlrunner discover -v python/tests -o test-results/cpu
-        DEVICE=gpu METAL_DEVICE_WRAPPER_TYPE=1 METAL_DEBUG_ERROR_MODE=0 uv run -m xmlrunner discover -v python/tests -o test-results/gpu
+        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
-      if: inputs.run-tests == 'true'
-      shell: bash
+      shell: bash -l {0}
       run: |
         cd examples/extensions
-        uv pip install -r requirements.txt
-        uv run --no-project setup.py build_ext --inplace
-        uv run --no-project test.py
+        pip install -r requirements.txt
+        python setup.py build_ext --inplace
+        python test.py
     
     - name: Build CPP only
-      if: inputs.build-type == 'debug'
-      shell: bash
+      shell: bash -l {0}
       run: |
         mkdir -p build
         cd build
@@ -70,8 +53,7 @@ runs:
         make -j $(sysctl -n hw.ncpu)
     
     - name: Run CPP tests
-      if: ${{ inputs.build-type == 'debug' && inputs.run-tests == 'true' }}
-      shell: bash
+      shell: bash -l {0}
       env:
         DEVICE: gpu
         METAL_DEVICE_WRAPPER_TYPE: 1
@@ -79,8 +61,7 @@ runs:
       run: ./build/tests/tests
     
     - name: Build small binary with JIT
-      if: inputs.build-jit == 'true'
-      shell: bash
+      shell: bash -l {0}
       run: |
         mkdir -p build
         cd build
@@ -93,8 +74,7 @@ runs:
         make -j $(sysctl -n hw.ncpu)
     
     - name: Run Python tests with JIT
-      if: ${{ inputs.build-jit == 'true' && inputs.run-tests == 'true' }}
-      shell: bash
+      shell: bash -l {0}
       env:
         LOW_MEMORY: 1
         DEVICE: gpu
@@ -102,23 +82,7 @@ runs:
         METAL_DEBUG_ERROR_MODE: 0
       run: |
         CMAKE_ARGS="-DMLX_METAL_JIT=ON" \
-          uv pip install -e . -v
-        uv run -m xmlrunner discover \
+          pip install -e . -v
+        python -m xmlrunner discover \
             -v python/tests \
             -o test-results/gpu_jit
-
-    - name: Build macOS 13 package
-      if: inputs.build-type == 'release'
-      uses: ./.github/actions/build-macos-release
-      with:
-        macos-target: 13.0
-    - name: Build macOS 14 package
-      if: inputs.build-type == 'release'
-      uses: ./.github/actions/build-macos-release
-      with:
-        macos-target: 14.0
-    - name: Build macOS 15 package
-      if: inputs.build-type == 'release'
-      uses: ./.github/actions/build-macos-release
-      with:
-        macos-target: 15.0

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

@@ -2,14 +2,10 @@ name: 'Setup Linux Environment'
 description: 'Install dependencies for Linux builds'
 
 inputs:
-  runner-type:
-    description: 'Whether to set this up as a linux or CUDA runner'
+  toolkit:
+    description: 'Which toolkit to install'
     required: false
-    default: 'linux'
-    type: choice
-    options:
-      - linux
-      - cuda
+    default: 'cpu'
   python-version:
     description: 'Version of python to set up'
     required: false
@@ -18,56 +14,63 @@ inputs:
 runs:
   using: "composite"
   steps:
-    - name: Free disk space
-      shell: sh
-      if: inputs.runner-type == 'linux'
-      run: sudo rm -rf "$AGENT_TOOLSDIRECTORY"
+    - name: Use ccache
+      uses: hendrikmuhs/ccache-action@v1.2
+      with:
+        key: ccache-${{ runner.os }}-${{ runner.arch }}-${{ inputs.toolkit }}-py${{ inputs.python-version }}
+        max-size: 1GB
 
     - name: Install common dependencies
-      env:
-        TZ: Etc/UTC
       shell: bash
       run: |
         sudo apt-get update
-        sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev tzdata zip
-        sudo apt autoremove -y
+        sudo apt-get install -y libblas-dev liblapack-dev liblapacke-dev zip
 
     - uses: actions/setup-python@v6
       with:
         python-version: ${{ inputs.python-version }}
-        cache: 'pip'
 
-    - name: setup python venv
+    - 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
-        pip install --upgrade pip cmake
+        # Make cmake search .venv for nanobind
+        echo PYTHONPATH=`python -c 'import sys; print(sys.path[-1])'` >> $GITHUB_ENV
 
     - name: Install MPI
-      if: inputs.runner-type == 'linux'
       shell: bash
       run: sudo apt-get install -y openmpi-bin openmpi-common libopenmpi-dev
 
-    - name: Network CUDA installation from packages
-      id: install-cuda
-      if: inputs.runner-type == 'cuda'
+    - name: Install CUDA toolkit
+      if: ${{ startsWith(inputs.toolkit, 'cuda') }}
+      shell: bash
       env:
-        TZ: Etc/UTC
-      shell: bash ## Specific to Ubuntu 22.04 & Architecture x86_64
+        # Note: the CI machine does not meet CUDA 13's driver requirement.
+        # Compatibility matrix:
+        # https://docs.nvidia.com/deeplearning/cudnn/backend/latest/reference/support-matrix.html
+        PACKAGES: |
+          {
+            "cuda-12.6": "libcudnn9-dev-cuda-12 cuda-toolkit-12-6",
+            "cuda-12.9": "libcudnn9-dev-cuda-12 cuda-toolkit-12-9",
+            "cuda-13.0": "libcudnn9-dev-cuda-13 cuda-toolkit-13-0"
+          }
       run: |
-        wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+        # The CUDA binaries are hosted in the "sbsa" repo, the "arm64" repo is
+        # Jetson specific. SBSA means Arm Server Base System Architecture.
+        ARCH=${{ 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 libcudnn9-dev-cuda-12 libnccl2 libnccl-dev cuda-toolkit-12-9
-      # Note: This installs CUDA 12.9, which is the latest supported by cuDNN 9.x and works with the NVidia 570 drivers
-      # cuda-toolkit by itself installs version 13 (+) and requires updated drives (580+), which require a reboot to function properly.
-      # Compatibility matrix: https://docs.nvidia.com/deeplearning/cudnn/backend/latest/reference/support-matrix.html
-      # This also drops `nvcc` into `/usr/local/cuda-12.9/bin/nvcc` - but it's *not* on the default PATH
+        sudo apt-get install -y \
+            libnccl2 libnccl-dev \
+            ${{ fromJson(env.PACKAGES)[inputs.toolkit] }}
+        echo "/usr/local/${{ inputs.toolkit }}/bin" >> $GITHUB_PATH
 
-    - name: Package and Driver Report
-      if: inputs.runner-type == 'cuda'
+    - name: CUDA packages and driver report
+      if: ${{ startsWith(inputs.toolkit, 'cuda') }}
       shell: bash
       run: |
         sudo apt-get install -y ubuntu-drivers-common dkms

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

@@ -2,11 +2,6 @@ name: 'Setup macOS Environment'
 description: 'Install dependencies for macOS builds'
 
 inputs:
-  install-mpi:
-    description: 'Whether to install MPI'
-    required: false
-    default: 'true'
-    type: boolean
   python-version:
     description: 'Python version to use'
     required: false
@@ -17,15 +12,13 @@ runs:
   steps:
     - name: Install Homebrew packages
       shell: sh
-      if: inputs.install-mpi == 'true'
       run: /opt/homebrew/bin/brew install openmpi
     
     - name: Verify MetalToolchain installed
       shell: bash
       run: xcodebuild -showComponent MetalToolchain
-    
-    - name: Setup uv
-      uses: astral-sh/setup-uv@v6
+
+    - uses: conda-incubator/setup-miniconda@v3
       with:
-          python-version: ${{ inputs.python-version }}
-          activate-environment: true
+        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/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

@@ -8,9 +8,9 @@ permissions:
 
 jobs:
   build:
-    runs-on: [self-hosted, macos]
+    runs-on: ubuntu-22.04
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/build-docs
       
   deploy:
@@ -25,4 +25,4 @@ jobs:
     steps:
       - name: Deploy to GitHub Pages
         id: deployment
-        uses: actions/deploy-pages@v4
+        uses: actions/deploy-pages@v4

.github/workflows/nightly.yml

@@ -16,12 +16,12 @@ jobs:
         python_version: ["3.10", "3.14"]
     runs-on: ubuntu-22.04
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-linux
-      - uses: ./.github/actions/build-linux
+      - uses: ./.github/actions/build-linux-release
         with:
-          build-type: release
-          run-tests: false
+          build-backend: ${{ matrix.python-version == '3.10' }}
+          arch: "x86_64"
       - name: Upload mlx artifacts
         uses: actions/upload-artifact@v5
         with:
@@ -35,59 +35,62 @@ jobs:
           name: mlx-cpu
           path: wheelhouse/mlx_cpu-*.whl
           retention-days: 7
-  
+
   build_linux_with_tests:
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
-    runs-on: ubuntu-22.04
+        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@v5
+      - 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"]
-        # TODO: 3.14 had issues finding a compatible tensorflow
-    env:
-      MACOSX_DEPLOYMENT_TARGET: "15.0"
     runs-on: [self-hosted, macos]
     steps:
-      - uses: actions/checkout@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-macos
         with:
           python-version: ${{ matrix.python-version }}
       - uses: ./.github/actions/build-macos
-
-  build_cuda_with_tests:
-    runs-on: gpu-t4-4-core
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-linux
+      - name: Build macOS 15 package
+        uses: ./.github/actions/build-macos-release
         with:
-          runner-type: 'cuda'
-      - uses: ./.github/actions/build-cuda
+          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@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-linux
         with:
-          runner-type: 'cuda'
+          toolkit: 'cuda-12.9'
       - name: Build Python package
         uses: ./.github/actions/build-cuda-release
         with:
-          nvcc-location: '/usr/local/cuda-12.9/bin/nvcc'
+          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

@@ -1,46 +0,0 @@
-name: Build and Test
-
-on: pull_request  
-
-permissions:
-  contents: read
-
-jobs:
-  check_lint:
-    runs-on: ubuntu-22.04
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-linux
-      - uses: pre-commit/action@v3.0.1
-
-  linux_build_and_test:
-    runs-on: ubuntu-22.04
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-linux
-      - uses: ./.github/actions/build-linux
-
-  mac_build_and_test:
-    runs-on: [self-hosted, macos]
-    needs: check_lint
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-macos
-      - uses: ./.github/actions/build-macos
-
-  cuda_build_and_test:
-    runs-on: gpu-t4-4-core
-    needs: check_lint
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/setup-linux
-        with:
-          runner-type: 'cuda'
-      - uses: ./.github/actions/build-cuda
-
-  build_documentation:
-    runs-on: [self-hosted, macos]
-    needs: check_lint
-    steps:
-      - uses: actions/checkout@v5
-      - uses: ./.github/actions/build-docs

.github/workflows/release.yml

@@ -5,15 +5,27 @@ on:
     tags:
       - 'v*'
   workflow_dispatch:
+    inputs:
+      dev_release:
+        description: "Do a dev release or regular release"
+        required: true
+        default: "false"
 
 permissions:
   contents: read
 
 jobs:
-  build_documentation:
-    runs-on: [self-hosted, macos]
+  setup:
+    runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v5
+      - 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:
@@ -31,84 +43,113 @@ jobs:
         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"]
-    runs-on: ubuntu-22.04
+        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@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-linux
         with:
           python-version: ${{ matrix.python_version }}
-      - uses: ./.github/actions/build-linux
+      - uses: ./.github/actions/build-linux-release
         with:
-          build-type: release
-          run-tests: false
+          build-backend: ${{ matrix.python-version == '3.10' }}
+          arch: ${{ matrix.arch }}
       - name: Upload MLX artifacts
         uses: actions/upload-artifact@v5
         with:
-          name: linux-wheels-${{ matrix.python_version }}
+          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:
-          name: mlx-cpu
+          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"]
-        # TODO: 3.14 had issues finding a compatible tensorflow
+        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@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-macos
         with:
           python-version: ${{ matrix.python-version }}
-      - uses: ./.github/actions/build-macos
+
+      - 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:
-          build-type: release
+          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@v5
+      - uses: actions/checkout@v6
       - uses: ./.github/actions/setup-linux
         with:
-          runner-type: 'cuda'
+          toolkit: 'cuda-12.9'
       - name: Build Python package
         uses: ./.github/actions/build-cuda-release
-        with:
-          nvcc-location: '/usr/local/cuda-12.9/bin/nvcc'
       - 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: [build_linux_release, build_mac_release]
+    needs: [setup, build_linux_release, build_mac_release]
     permissions:
       id-token: write
     environment:
@@ -118,22 +159,24 @@ jobs:
       - uses: actions/download-artifact@v6
         with:
           pattern: linux-wheels-*
-          merge-multiples: true
-          path: artifacts
+          merge-multiple: true
+          path: dist
       - uses: actions/download-artifact@v6
         with:
           pattern: mac-wheels-*
-          merge-multiples: true
-          path: artifacts
+          merge-multiple: true
+          path: dist
       - name: Display structure of downloaded files
-        run: ls -R artifacts
-      # - name: Publish package distributions to PyPI
-      #  uses: pypa/gh-action-pypi-publish@release/v1
+        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: build_cuda_release
+    needs: [setup, build_cuda_release]
     permissions:
       id-token: write
     environment:
@@ -143,16 +186,18 @@ jobs:
       - uses: actions/download-artifact@v6
         with:
           name: mlx-cuda
-          path: artifacts
+          path: dist
       - name: Display structure of downloaded files
-        run: ls -R artifacts
-      # - name: Publish package distributions to PyPI
-      #  uses: pypa/gh-action-pypi-publish@release/v1
+        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: build_linux_release
+    needs: [setup, build_linux_release]
     permissions:
       id-token: write
     environment:
@@ -161,17 +206,20 @@ jobs:
     steps:
       - uses: actions/download-artifact@v6
         with:
-          name: mlx-cpu
-          path: artifacts
+          pattern: mlx-cpu-*
+          merge-multiple: true
+          path: dist
       - name: Display structure of downloaded files
-        run: ls -R artifacts
-      # - name: Publish package distributions to PyPI
-      #  uses: pypa/gh-action-pypi-publish@release/v1
+        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: build_mac_release
+    needs: [setup, build_mac_release]
     permissions:
       id-token: write
     environment:
@@ -181,8 +229,10 @@ jobs:
       - uses: actions/download-artifact@v6
         with:
           name: mlx-metal
-          path: artifacts
+          path: dist
       - name: Display structure of downloaded files
-        run: ls -R artifacts
-      # - name: Publish package distributions to PyPI
-      #  uses: pypa/gh-action-pypi-publish@release/v1
+        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/

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}")
@@ -127,9 +128,12 @@ 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(
@@ -138,7 +142,6 @@ 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

@@ -17,11 +17,10 @@ To install from PyPI your system must meet the following requirements:
 
 - Using an M series chip (Apple silicon)
 - Using a native Python >= 3.10
-- macOS >= 13.5
+- macOS >= 14.0
 
 .. note::
-    MLX is only available on devices running macOS >= 13.5
-    It is highly recommended to use macOS 14 (Sonoma)
+    MLX is only available on devices running macOS >= 14.0 and higher.
 
 CUDA
 ^^^^

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. No mask
+  broadcasting occurs.
+- Any axes not covered by the mask are taken in full.
+
+.. code-block:: shell
+
+   >>> a = mx.arange(1000).reshape(10, 10, 10)
+   >>> a[mx.random.randn(10, 10) > 0.0] = 0  # valid: mask covers axes 0 and 1
+
+The mask of shape ``(10, 10)`` applies to the first two axes, so ``a[mask]``
+selects the 1-D slices ``a[i, j, :]`` where ``mask[i, j]`` is ``True``.
+Shapes such as ``(1, 10, 10)`` or ``(10, 10, 1)`` do not match the indexed
+axes and therefore raise errors.

mlx/array.cpp

@@ -167,7 +167,7 @@ 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;

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;
@@ -434,7 +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);
 

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
@@ -51,17 +47,24 @@ void slice(
 
   // 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/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/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,
@@ -19,101 +180,39 @@ void eig_impl(
     array& values,
     bool compute_eigenvectors,
     Stream stream) {
-  using OT = std::complex<T>;
   auto a_ptr = a.data<T>();
-  auto eig_ptr = values.data<OT>();
+  auto val_ptr = values.data<complex64_t>();
 
   auto& encoder = cpu::get_command_encoder(stream);
   encoder.set_input_array(a);
   encoder.set_output_array(values);
-  OT* vec_ptr = nullptr;
+  complex64_t* vec_ptr = nullptr;
   if (compute_eigenvectors) {
     encoder.set_output_array(vectors);
-    vec_ptr = vectors.data<OT>();
+    vec_ptr = vectors.data<complex64_t>();
   }
   encoder.dispatch([a_ptr,
+                    val_ptr,
                     vec_ptr,
-                    eig_ptr,
                     compute_eigenvectors,
                     N = vectors.shape(-1),
                     size = vectors.size()]() mutable {
-    // Work query
     char jobr = 'N';
     char jobl = compute_eigenvectors ? 'V' : 'N';
-    int n_vecs_r = 1;
-    int n_vecs_l = compute_eigenvectors ? N : 1;
-    int lwork = -1;
-    int info;
-    {
-      T work;
-      geev<T>(
-          &jobl,
-          &jobr,
-          &N,
-          nullptr,
-          &N,
-          nullptr,
-          nullptr,
-          nullptr,
-          &n_vecs_l,
-          nullptr,
-          &n_vecs_r,
-          &work,
-          &lwork,
-          &info);
-      lwork = static_cast<int>(work);
-    }
 
-    auto eig_tmp_data = array::Data{allocator::malloc(sizeof(T) * N * 2)};
-    auto vec_tmp_data =
-        array::Data{allocator::malloc(vec_ptr ? sizeof(T) * N * N * 2 : 0)};
-    auto eig_tmp = static_cast<T*>(eig_tmp_data.buffer.raw_ptr());
-    auto vec_tmp = static_cast<T*>(vec_tmp_data.buffer.raw_ptr());
-    auto work_buf = array::Data{allocator::malloc(sizeof(T) * lwork)};
+    EigWork<T> work(jobl, jobr, N, compute_eigenvectors);
+
     for (size_t i = 0; i < size / (N * N); ++i) {
-      geev<T>(
-          &jobl,
-          &jobr,
-          &N,
-          a_ptr,
-          &N,
-          eig_tmp,
-          eig_tmp + N,
-          vec_tmp,
-          &n_vecs_l,
-          nullptr,
-          &n_vecs_r,
-          static_cast<T*>(work_buf.buffer.raw_ptr()),
-          &lwork,
-          &info);
-      for (int i = 0; i < N; ++i) {
-        eig_ptr[i] = {eig_tmp[i], eig_tmp[N + i]};
-      }
+      work.run(a_ptr, val_ptr, vec_ptr);
+      a_ptr += N * N;
+      val_ptr += N;
       if (vec_ptr) {
-        for (int i = 0; i < N; ++i) {
-          if (eig_ptr[i].imag() != 0) {
-            // This vector and the next are a pair
-            for (int j = 0; j < N; ++j) {
-              vec_ptr[i * N + j] = {
-                  vec_tmp[i * N + j], -vec_tmp[(i + 1) * N + j]};
-              vec_ptr[(i + 1) * N + j] = {
-                  vec_tmp[i * N + j], vec_tmp[(i + 1) * N + j]};
-            }
-            i += 1;
-          } else {
-            for (int j = 0; j < N; ++j) {
-              vec_ptr[i * N + j] = {vec_tmp[i * N + j], 0};
-            }
-          }
-        }
         vec_ptr += N * N;
       }
-      a_ptr += N * N;
-      eig_ptr += N;
-      if (info != 0) {
+      if (work.info != 0) {
         std::stringstream msg;
         msg << "[Eig::eval_cpu] Eigenvalue decomposition failed with error code "
-            << info;
+            << work.info;
         throw std::runtime_error(msg.str());
       }
     }
@@ -165,8 +264,17 @@ void Eig::eval_cpu(
     case float32:
       eig_impl<float>(a_copy, vectors, values, compute_eigenvectors_, stream());
       break;
+    case float64:
+      eig_impl<double>(
+          a_copy, vectors, values, compute_eigenvectors_, stream());
+      break;
+    case complex64:
+      eig_impl<std::complex<float>>(
+          a_copy, vectors, values, compute_eigenvectors_, stream());
+      break;
     default:
-      throw std::runtime_error("[Eig::eval_cpu] only supports float32.");
+      throw std::runtime_error(
+          "[Eig::eval_cpu] only supports float32, float64, or complex64.");
   }
 }
 

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/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"
@@ -135,15 +137,29 @@ void AddMM::eval_cpu(const std::vector<array>& inputs, array& out) {
     return;
   }
 
+  // Handle empty matrix case (K=0)
+  if (inputs[0].shape(-1) == 0) {
+    auto& c = inputs[2];
+    if (beta_ == 1.0f) {
+      CopyType ctype = c.data_size() == 1
+          ? CopyType::Scalar
+          : (c.flags().row_contiguous ? CopyType::Vector : CopyType::General);
+      copy_cpu(c, out, ctype, stream());
+    } else {
+      array beta_scalar = array(beta_, c.dtype());
+      auto& encoder = cpu::get_command_encoder(stream());
+      binary_float_op_cpu(c, beta_scalar, out, detail::Multiply(), stream());
+      encoder.add_temporary(std::move(beta_scalar));
+    }
+    return;
+  }
+
   // Fill output with C
   auto& c = inputs[2];
   CopyType ctype = c.data_size() == 1
       ? CopyType::Scalar
       : (c.flags().row_contiguous ? CopyType::Vector : CopyType::General);
   copy_cpu(c, out, ctype, stream());
-  if (inputs[0].shape(-1) == 0) {
-    return;
-  }
   matmul_general(inputs[0], inputs[1], out, stream(), alpha_, beta_);
 }
 

mlx/backend/cpu/simd/accelerate_simd.h

@@ -217,14 +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/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 {
@@ -68,96 +247,26 @@ void svd_impl(
 
     encoder.set_output_array(s);
 
-    s_ptr = s.data<T>();
+    s_ptr = s.data<R>();
     u_ptr = nullptr;
     vt_ptr = nullptr;
   }
 
   encoder.dispatch([in_ptr, u_ptr, s_ptr, vt_ptr, M, N, K, num_matrices]() {
-    // A of shape M x N. The leading dimension is N since lapack receives Aᵀ.
-    const int lda = N;
-    // U of shape M x M. (N x N in lapack).
-    const int ldu = N;
-    // Vᵀ of shape N x N. (M x M in lapack).
-    const int ldvt = M;
-
-    auto jobz = (u_ptr) ? "A" : "N";
-
-    T workspace_dimension = 0;
-
-    // Will contain the indices of eigenvectors that failed to converge (not
-    // used here but required by lapack).
-    auto iwork = array::Data{allocator::malloc(sizeof(int) * 8 * K)};
-
-    static const int lwork_query = -1;
-
-    int info;
-
-    // Compute workspace size.
-    gesdd<T>(
-        /* jobz = */ jobz,
-        // M and N are swapped since lapack expects column-major.
-        /* m = */ &N,
-        /* n = */ &M,
-        /* a = */ nullptr,
-        /* lda = */ &lda,
-        /* s = */ nullptr,
-        /* u = */ nullptr,
-        /* ldu = */ &ldu,
-        /* vt = */ nullptr,
-        /* ldvt = */ &ldvt,
-        /* work = */ &workspace_dimension,
-        /* lwork = */ &lwork_query,
-        /* iwork = */ static_cast<int*>(iwork.buffer.raw_ptr()),
-        /* info = */ &info);
-
-    if (info != 0) {
-      std::stringstream ss;
-      ss << "[SVD::eval_cpu] workspace calculation failed with code " << info;
-      throw std::runtime_error(ss.str());
-    }
-
-    const int lwork = workspace_dimension;
-    auto scratch = array::Data{allocator::malloc(sizeof(T) * lwork)};
-
+    auto jobz = (u_ptr) ? 'A' : 'N';
+    SVDWork<T> svd_work(N, M, K, jobz);
     // Loop over matrices.
     for (int i = 0; i < num_matrices; i++) {
-      gesdd<T>(
-          /* jobz = */ jobz,
-          // M and N are swapped since lapack expects column-major.
-          /* m = */ &N,
-          /* n = */ &M,
-          /* a = */ in_ptr + M * N * i,
-          /* lda = */ &lda,
-          /* s = */ s_ptr + K * i,
-          // According to the identity above, lapack will write Vᵀᵀ as U.
-          /* u = */ vt_ptr ? vt_ptr + N * N * i : nullptr,
-          /* ldu = */ &ldu,
-          // According to the identity above, lapack will write Uᵀ as Vᵀ.
-          /* vt = */ u_ptr ? u_ptr + M * M * i : nullptr,
-          /* ldvt = */ &ldvt,
-          /* work = */ static_cast<T*>(scratch.buffer.raw_ptr()),
-          /* lwork = */ &lwork,
-          /* iwork = */ static_cast<int*>(iwork.buffer.raw_ptr()),
-          /* info = */ &info);
-
-      if (info != 0) {
-        std::stringstream ss;
-        ss << "svd_impl: sgesvdx_ failed with code " << info;
-        throw std::runtime_error(ss.str());
-      }
+      svd_work.run(
+          in_ptr + M * N * i,
+          s_ptr + K * i,
+          vt_ptr ? vt_ptr + N * N * i : nullptr,
+          u_ptr ? u_ptr + M * M * i : nullptr);
     }
   });
   encoder.add_temporary(in);
 }
 
-template <typename T>
-void compute_svd(
-    const array& a,
-    bool compute_uv,
-    std::vector<array>& outputs,
-    Stream stream) {}
-
 void SVD::eval_cpu(
     const std::vector<array>& inputs,
     std::vector<array>& outputs) {
@@ -168,9 +277,12 @@ void SVD::eval_cpu(
     case float64:
       svd_impl<double>(inputs[0], outputs, compute_uv_, stream());
       break;
+    case complex64:
+      svd_impl<std::complex<float>>(inputs[0], outputs, compute_uv_, stream());
+      break;
     default:
       throw std::runtime_error(
-          "[SVD::eval_cpu] only supports float32 or float64.");
+          "[SVD::eval_cpu] only supports float32, float64, or complex64.");
   }
 }
 

mlx/backend/cuda/CMakeLists.txt

@@ -44,6 +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
           ${CMAKE_CURRENT_SOURCE_DIR}/slicing.cpp
@@ -122,10 +123,21 @@ if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8.0)
     mlx PRIVATE "$<$<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: ${MLX_CUDA_ARCHITECTURES}")
 set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
@@ -137,6 +149,7 @@ FetchContent_Declare(
   URL "https://github.com/NVIDIA/cccl/releases/download/v2.8.1/cccl-v2.8.1.zip")
 FetchContent_MakeAvailable(cccl)
 target_include_directories(mlx BEFORE PRIVATE "${cccl_SOURCE_DIR}/include")
+set_target_properties(mlx PROPERTIES CCCL_DIR "${cccl_SOURCE_DIR}/include")
 
 # Use fixed version of NVTX.
 FetchContent_Declare(
@@ -162,7 +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)

mlx/backend/cuda/allocator.cpp

@@ -92,22 +92,40 @@ CudaAllocator::CudaAllocator()
           [this](CudaBuffer* buf) { cuda_free(buf); }) {
   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_impl(size_t size, cudaStream_t stream) {
+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;
@@ -117,6 +135,10 @@ Buffer CudaAllocator::malloc_impl(size_t size, cudaStream_t stream) {
     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.
@@ -132,40 +154,40 @@ Buffer CudaAllocator::malloc_impl(size_t size, cudaStream_t stream) {
     }
     lock.unlock();
     if (!buf) {
-      int device = -1;
-      if (stream != nullptr) {
-        cudaStreamGetDevice(stream, &device);
-      }
-      buf = new CudaBuffer{nullptr, size, device};
       cudaError_t err;
+      void* data = nullptr;
       if (device == -1) {
-        err = cudaMallocManaged(&buf->data, size);
+        err = cudaMallocManaged(&data, size);
       } else {
-        err = cudaMallocAsync(&buf->data, size, stream);
+        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_async(size_t size, cudaStream_t stream) {
-  return malloc_impl(size, stream);
-}
-
 Buffer CudaAllocator::malloc(size_t size) {
-  return malloc_impl(size, nullptr);
+  return malloc_async(size, -1, nullptr);
 }
 
 void CudaAllocator::free(Buffer buffer) {
@@ -173,6 +195,10 @@ void CudaAllocator::free(Buffer buffer) {
   if (!buf) {
     return;
   }
+  if (buf->size == 0) {
+    delete buf;
+    return;
+  }
 
   std::unique_lock lock(mutex_);
   active_memory_ -= buf->size;
@@ -197,9 +223,9 @@ void CudaAllocator::cuda_free(CudaBuffer* buf) {
     scalar_pool_.free(buf);
   } else {
     if (buf->device >= 0) {
-      cudaFreeAsync(buf->data, free_streams_[buf->device]);
+      CHECK_CUDA_ERROR(cudaFreeAsync(buf->data, free_streams_[buf->device]));
     } else {
-      cudaFree(buf->data);
+      CHECK_CUDA_ERROR(cudaFree(buf->data));
     }
     delete buf;
   }
@@ -251,8 +277,9 @@ CudaAllocator& allocator() {
   return *allocator_;
 }
 
-Buffer malloc_async(size_t size, cudaStream_t stream) {
-  auto buffer = allocator().malloc_async(size, stream);
+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.";
@@ -275,15 +302,7 @@ void* Buffer::raw_ptr() {
   }
   auto& cbuf = *static_cast<cu::CudaBuffer*>(ptr_);
   if (cbuf.device != -1) {
-    // 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, cbuf.size));
-    cbuf.device = -1;
-    CHECK_CUDA_ERROR(
-        cudaMemcpy(new_data, cbuf.data, cbuf.size, cudaMemcpyDefault));
-    CHECK_CUDA_ERROR(cudaFree(cbuf.data));
-    cbuf.data = new_data;
+    copy_to_managed(cbuf);
   }
   return cbuf.data;
 }

mlx/backend/cuda/allocator.h

@@ -13,6 +13,8 @@
 
 namespace mlx::core::cu {
 
+class CommandEncoder;
+
 using allocator::Buffer;
 
 // Stores cuda-managed unified memory.
@@ -48,7 +50,7 @@ class SmallSizePool {
 class CudaAllocator : public allocator::Allocator {
  public:
   Buffer malloc(size_t size) override;
-  Buffer malloc_async(size_t size, cudaStream_t stream);
+  Buffer malloc_async(size_t size, int device, cudaStream_t stream);
   void free(Buffer buffer) override;
   size_t size(Buffer buffer) const override;
 
@@ -62,7 +64,6 @@ class CudaAllocator : public allocator::Allocator {
   void clear_cache();
 
  private:
-  Buffer malloc_impl(size_t size, cudaStream_t stream);
   void cuda_free(CudaBuffer* buf);
 
   CudaAllocator();
@@ -80,6 +81,6 @@ class CudaAllocator : public allocator::Allocator {
 
 CudaAllocator& allocator();
 
-Buffer malloc_async(size_t size, cudaStream_t stream);
+Buffer malloc_async(size_t size, CommandEncoder& encoder);
 
 } // namespace mlx::core::cu

mlx/backend/cuda/arange.cu

@@ -42,7 +42,7 @@ void Arange::eval_gpu(const std::vector<array>& inputs, array& out) {
     return;
   }
   auto& encoder = cu::get_command_encoder(stream());
-  out.set_data(cu::malloc_async(out.nbytes(), 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) {

mlx/backend/cuda/arg_reduce.cu

@@ -143,7 +143,7 @@ void ArgReduce::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   auto& s = stream();
   auto& encoder = cu::get_command_encoder(s);
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
 
   // Prepare the shapes, strides and axis arguments.
   Shape shape = remove_index(in.shape(), axis_);

mlx/backend/cuda/binary/binary.cuh

@@ -367,9 +367,8 @@ void binary_op_gpu(
   auto bopt = get_binary_op_type(a, b);
   auto& encoder = cu::get_command_encoder(s);
 
-  set_binary_op_output_data(a, b, out, bopt, [&](auto n) {
-    return cu::malloc_async(n, encoder.stream());
-  });
+  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

@@ -246,12 +246,10 @@ void binary_two_op_gpu_inplace(
   auto& out_b = outputs[1];
   auto bopt = get_binary_op_type(a, b);
   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.stream());
-  });
-  set_binary_op_output_data(a, b, out_b, bopt, [&](auto n) {
-    return cu::malloc_async(n, encoder.stream());
-  });
+  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;

mlx/backend/cuda/compiled.cpp

@@ -298,7 +298,7 @@ void Compiled::eval_gpu(
   // Put outputs.
   compiled_allocate_outputs(
       inputs, outputs, is_constant_, contiguous, [&](auto n) {
-        return cu::malloc_async(n, encoder.stream());
+        return cu::malloc_async(n, encoder);
       });
   for (auto& x : outputs) {
     args.append(x);

mlx/backend/cuda/conv.cpp

@@ -277,11 +277,12 @@ void Convolution::eval_gpu(const std::vector<array>& inputs, array& out_) {
   array in = inputs[0];
   array wt = inputs[1];
   array out = out_;
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   Dtype dtype = out.dtype();
 
   // 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()),

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(cu::malloc_async(unfolded.nbytes(), encoder.stream()));
+  unfolded.set_data(cu::malloc_async(unfolded.nbytes(), encoder));
   encoder.add_temporary(unfolded);
 
   int filter_size = params.C;

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(cu::malloc_async(unfolded.nbytes(), encoder.stream()));
+  unfolded.set_data(cu::malloc_async(unfolded.nbytes(), encoder));
   encoder.add_temporary(unfolded);
 
   int filter_size = params.C;

mlx/backend/cuda/copy.cu

@@ -7,9 +7,8 @@ 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.stream());
-  });
+  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.
@@ -104,7 +103,7 @@ void fill_gpu(const array& in, array& out, const Stream& s) {
     return;
   }
   auto& encoder = cu::get_command_encoder(s);
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  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);
@@ -114,7 +113,7 @@ 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.stream()));
+    out.set_data(cu::malloc_async(out.nbytes(), encoder));
     copy_gpu_inplace(
         in,
         out,

mlx/backend/cuda/cuda_utils.h

@@ -29,6 +29,10 @@ class CudaHandle {
   }
 
   ~CudaHandle() {
+    // Skip if there was an error to avoid throwing in the destructors
+    if (cudaPeekAtLastError() != cudaSuccess) {
+      return;
+    }
     reset();
   }
 

mlx/backend/cuda/cudnn_utils.cpp

@@ -135,9 +135,7 @@ bool prepare_cudnn_plan(
   void* workspace_ptr = nullptr;
   if (workspace_size > 0) {
     array workspace(
-        cu::malloc_async(workspace_size, encoder.stream()),
-        {workspace_size},
-        uint8);
+        cu::malloc_async(workspace_size, encoder), {workspace_size}, uint8);
     encoder.add_temporary(workspace);
     workspace_ptr = gpu_ptr<void>(workspace);
   }

mlx/backend/cuda/cudnn_utils.h

@@ -44,13 +44,13 @@ 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;
 }

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,
@@ -289,7 +289,7 @@ void CustomKernel::eval_gpu(
       copies.emplace_back(init_value_.value(), out.dtype());
       fill_gpu(copies.back(), out, s);
     } else {
-      out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+      out.set_data(cu::malloc_async(out.nbytes(), encoder));
     }
   }
 
@@ -327,13 +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());
     }
   }

mlx/backend/cuda/device.cpp

@@ -24,12 +24,21 @@ void check_cudnn_error(const char* name, cudnnStatus_t err) {
 }
 
 bool use_cuda_graphs() {
-  static bool use_graphs = []() {
-    return env::get_var("MLX_USE_CUDA_GRAPHS", true);
-  }();
+  static bool use_graphs = env::get_var("MLX_USE_CUDA_GRAPHS", true);
   return use_graphs;
 }
 
+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 filename;
+}
+
 } // namespace
 
 Device::Device(int device) : device_(device) {
@@ -46,6 +55,7 @@ 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_));
@@ -114,18 +124,17 @@ 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 +149,6 @@ 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 +160,10 @@ 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 +191,49 @@ 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 +243,7 @@ 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 +318,46 @@ 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 != cudaGraphNodeTypeKernel) {
+      return false;
+    }
+    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

@@ -84,7 +84,7 @@ class CommandEncoder {
   }
 
   void add_completed_handler(std::function<void()> task);
-  int get_num_ops();
+  bool needs_commit();
   void commit();
 
   Device& device() {
@@ -106,8 +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;
   };
 
@@ -119,18 +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 {
@@ -166,6 +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/distributed.cu

@@ -26,7 +26,7 @@ void AllReduce::eval_gpu(
       out.copy_shared_buffer(in);
       return {in, out};
     } else {
-      out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+      out.set_data(cu::malloc_async(out.nbytes(), encoder));
       return {in, out};
     }
   };
@@ -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

@@ -34,8 +34,8 @@ void Fence::update(Stream s, const array& a, bool cross_device) {
       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));
+      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;
     }

mlx/backend/cuda/gemms/cublas_gemm.cpp

@@ -370,7 +370,7 @@ void CublasGemm::execute(
     // Ensure workspace is 256-byte aligned
     int nbytes = cuda::ceil_div(heuristic_.workspaceSize, 256) * 256;
     array workspace(
-        cu::malloc_async(nbytes, encoder.stream()),
+        cu::malloc_async(nbytes, encoder),
         {static_cast<int>(heuristic_.workspaceSize)},
         int8);
     encoder.add_temporary(workspace);

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(
-      cu::malloc_async(batch_count * sizeof(void*) * 3, encoder.stream()),
+      cu::malloc_async(batch_count * sizeof(void*) * 3, encoder),
       {batch_count * 3},
       uint64);
 
@@ -251,7 +251,7 @@ void CublasGemm::run_batched(
 
   // Launch kernel to set device offsets
   auto pointers = array(
-      cu::malloc_async(batch_count * sizeof(uint64_t) * 4, encoder.stream()),
+      cu::malloc_async(batch_count * sizeof(uint64_t) * 4, encoder),
       {batch_count * 4},
       uint64);
 

mlx/backend/cuda/indexing.cpp

@@ -61,7 +61,7 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   auto& s = stream();
   auto& encoder = cu::get_command_encoder(s);
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   if (out.size() == 0) {
     return;
   }
@@ -241,7 +241,7 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
 
   auto& s = stream();
   auto& encoder = cu::get_command_encoder(s);
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   if (out.size() == 0) {
     return;
   }

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/layer_norm.cu

@@ -244,7 +244,7 @@ void LayerNorm::eval_gpu(
         out.copy_shared_buffer(x);
       } else {
         out.set_data(
-            cu::malloc_async(x.data_size() * x.itemsize(), encoder.stream()),
+            cu::malloc_async(x.data_size() * x.itemsize(), encoder),
             x.data_size(),
             x.strides(),
             x.flags());
@@ -335,7 +335,7 @@ void LayerNormVJP::eval_gpu(
     gx.copy_shared_buffer(g);
     g_in_gx = true;
   } else {
-    gx.set_data(cu::malloc_async(gx.nbytes(), encoder.stream()));
+    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(cu::malloc_async(gw_temp.nbytes(), encoder.stream()));
+      gw_temp.set_data(cu::malloc_async(gw_temp.nbytes(), encoder));
       encoder.add_temporary(gw_temp);
     }
   }

mlx/backend/cuda/load.cpp

@@ -32,7 +32,7 @@ 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.stream()));
+  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_) {

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(cu::malloc_async(out.nbytes(), encoder.stream()));
+    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(
-        cu::malloc_async(in.nbytes() / n, encoder.stream()),
+        cu::malloc_async(in.nbytes() / n, encoder),
         in.data_size() / n,
         std::move(strides),
         flags);

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(cu::malloc_async(out.nbytes(), encoder.stream()));
+  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(cu::malloc_async(out.nbytes(), encoder.stream()));
+    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(cu::malloc_async(out.nbytes(), encoder.stream()));
+      out.set_data(cu::malloc_async(out.nbytes(), encoder));
     } else if (sty == 1 && stx == 0) {
       ldc = 0;
-      out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+      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

@@ -37,9 +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/quantized.cpp

@@ -59,7 +59,7 @@ void fast::Quantize::eval_gpu(
     auto scales = ensure_row_contiguous(inputs[1], enc, s);
     auto& w = outputs[0];
 
-    w.set_data(cu::malloc_async(w.nbytes(), enc.stream()));
+    w.set_data(cu::malloc_async(w.nbytes(), enc));
 
     if (mode_ == QuantizationMode::Affine) {
       auto biases = ensure_row_contiguous(inputs[2], enc, s);
@@ -72,11 +72,11 @@ void fast::Quantize::eval_gpu(
     auto& wq = outputs[0];
     auto& scales = outputs[1];
 
-    wq.set_data(cu::malloc_async(wq.nbytes(), enc.stream()));
-    scales.set_data(cu::malloc_async(scales.nbytes(), enc.stream()));
+    wq.set_data(cu::malloc_async(wq.nbytes(), enc));
+    scales.set_data(cu::malloc_async(scales.nbytes(), enc));
     if (mode_ == QuantizationMode::Affine) {
       auto& biases = outputs[2];
-      biases.set_data(cu::malloc_async(biases.nbytes(), enc.stream()));
+      biases.set_data(cu::malloc_async(biases.nbytes(), enc));
       affine_quantize(w, wq, scales, biases, group_size_, bits_, enc, s);
     } else {
       fp_quantize(w, wq, scales, group_size_, bits_, enc, s);

mlx/backend/cuda/random.cu

@@ -139,30 +139,36 @@ void RandomBits::eval_gpu(const std::vector<array>& inputs, array& out) {
   // keys has shape (N1, ..., NK, 2)
   // out has shape (N1, ..., NK, M1, M2, ...)
   auto& keys = inputs[0];
-  uint32_t num_keys = keys.size() / 2;
+  size_t num_keys = keys.size() / 2;
 
-  uint32_t elems_per_key = out.size() / num_keys;
-  uint32_t bytes_per_key = out.itemsize() * elems_per_key;
+  size_t elems_per_key = out.size() / num_keys;
+  size_t bytes_per_key = out.itemsize() * elems_per_key;
   auto& s = stream();
   auto& encoder = cu::get_command_encoder(s);
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
   if (out.size() == 0) {
     return;
   }
 
-  uint32_t out_per_key = (bytes_per_key + 4 - 1) / 4;
-  uint32_t half_size = out_per_key / 2;
+  size_t out_per_key = (bytes_per_key + 4 - 1) / 4;
+  size_t half_size = out_per_key / 2;
+
   bool odd = out_per_key % 2;
+  if ((half_size + odd) >= UINT32_MAX || num_keys >= UINT32_MAX) {
+    throw std::runtime_error("[RandomBits::eval_gpu] Large size unsupported");
+  }
 
   encoder.set_input_array(keys);
   encoder.set_output_array(out);
-  dim3 grid_dims{num_keys, half_size + odd};
-  int64_t total = grid_dims.x * grid_dims.y;
-  int32_t threads_y = 1;
-  while ((total / threads_y) >= (1U << 31)) {
+  int64_t total = num_keys * (half_size + odd);
+  uint32_t threads_y = 1;
+  while ((total / threads_y) >= UINT_MAX) {
     threads_y *= 2;
   }
-  int32_t threads_x = cuda::ceil_div(total, threads_y);
+  uint32_t threads_x = cuda::ceil_div(total, threads_y);
+
+  dim3 grid_dims{
+      static_cast<uint32_t>(num_keys), static_cast<uint32_t>(half_size + odd)};
   auto [grid, block] = get_grid_and_block(threads_x, threads_y, 1);
   auto& stream = encoder.stream();
   if (keys.flags().row_contiguous) {

mlx/backend/cuda/reduce/all_reduce.cu

@@ -66,7 +66,7 @@ void all_reduce(
     Reduce::ReduceType reduce_type) {
   constexpr int N_READS = 8;
 
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
 
   auto get_args = [](size_t size, int N) {
     int threads = std::min(512UL, (size + N - 1) / N);
@@ -107,8 +107,7 @@ void all_reduce(
   encoder.set_input_array(in);
   if (blocks > 1) {
     array intermediate({blocks}, out.dtype(), nullptr, {});
-    intermediate.set_data(
-        cu::malloc_async(intermediate.nbytes(), encoder.stream()));
+    intermediate.set_data(cu::malloc_async(intermediate.nbytes(), encoder));
     encoder.add_temporary(intermediate);
     encoder.set_output_array(intermediate);
     dispatch_all_types(dt, [&](auto type_tag) {

mlx/backend/cuda/reduce/init_reduce.cu

@@ -28,7 +28,7 @@ void init_reduce(
     Reduce::ReduceType reduce_type) {
   // Allocate if needed
   if (out.data_shared_ptr() == nullptr) {
-    out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+    out.set_data(cu::malloc_async(out.nbytes(), encoder));
   }
 
   encoder.set_output_array(out);

mlx/backend/cuda/reduce/reduce_utils.cuh

@@ -96,7 +96,7 @@ inline void allocate_same_layout(
     const std::vector<int>& axes,
     cu::CommandEncoder& encoder) {
   if (in.flags().row_contiguous) {
-    out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+    out.set_data(cu::malloc_async(out.nbytes(), encoder));
     return;
   }
 
@@ -135,7 +135,7 @@ inline void allocate_same_layout(
   fl.col_contiguous = cc;
   fl.contiguous = true;
   out.set_data(
-      cu::malloc_async(out.nbytes(), encoder.stream()),
+      cu::malloc_async(out.nbytes(), encoder),
       data_size,
       final_strides,
       fl,

mlx/backend/cuda/rms_norm.cu

@@ -22,26 +22,28 @@ inline __device__ float2 plus_f2(const float2& a, const float2& b) {
 }
 
 // Similar to cub::BlockReduce, but result is broadcasted to every thread.
-template <typename T, int BLOCK_DIM>
+template <typename T, int BLOCK_DIM, int GROUP_DIM = WARP_SIZE>
 struct BlockBroadcastReduce {
-  static_assert(WARP_SIZE <= BLOCK_DIM && BLOCK_DIM <= WARP_SIZE * WARP_SIZE);
-  static_assert(BLOCK_DIM % WARP_SIZE == 0);
-  using TempStorage = T[BLOCK_DIM / WARP_SIZE];
+  using TempStorage = T[std::max(BLOCK_DIM / WARP_SIZE, 1)];
 
   cg::thread_block& block;
   TempStorage& temp;
 
   template <typename Op>
   __device__ T Reduce(const T& input, const Op& op, const T& init_value) {
-    auto warp = cg::tiled_partition<WARP_SIZE>(block);
+    auto warp = cg::tiled_partition<GROUP_DIM>(block);
     T x = cg::reduce(warp, input, op);
-    if (warp.thread_rank() == 0) {
-      temp[warp.meta_group_rank()] = x;
+    if constexpr (BLOCK_DIM > GROUP_DIM) {
+      if (warp.thread_rank() == 0) {
+        temp[warp.meta_group_rank()] = x;
+      }
+      block.sync();
+      x = warp.thread_rank() < warp.meta_group_size() ? temp[warp.thread_rank()]
+                                                      : init_value;
+      return cg::reduce(warp, x, op);
+    } else {
+      return x;
     }
-    block.sync();
-    x = warp.thread_rank() < warp.meta_group_size() ? temp[warp.thread_rank()]
-                                                    : init_value;
-    return cg::reduce(warp, x, op);
   }
 
   __device__ T Sum(const T& input) {
@@ -49,6 +51,52 @@ struct BlockBroadcastReduce {
   }
 };
 
+template <typename T, int BLOCK_DIM, int REDUCE_DIM, int N_READS = 4>
+__global__ void rms_norm_small(
+    const T* x,
+    const T* w,
+    T* out,
+    float eps,
+    uint32_t axis_size,
+    uint32_t n_rows,
+    int64_t w_stride) {
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+
+  using BlockReduceT = BlockBroadcastReduce<float, BLOCK_DIM, REDUCE_DIM>;
+  __shared__ typename BlockReduceT::TempStorage temp;
+
+  auto row =
+      (grid.block_rank() * block.dim_threads().y) + block.thread_index().y;
+  if (row >= n_rows) {
+    return;
+  }
+  x += row * axis_size;
+  out += row * axis_size;
+
+  // Normalizer.
+  float normalizer = 0;
+  auto index = block.thread_index().x;
+  auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    float t = static_cast<float>(xn[i]);
+    normalizer += t * t;
+  }
+
+  normalizer = BlockReduceT{block, temp}.Sum(normalizer);
+  normalizer = rsqrt(normalizer / axis_size + eps);
+
+  // Outputs.
+  auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
+#pragma unroll
+  for (int i = 0; i < N_READS; ++i) {
+    float y = static_cast<float>(xn[i]) * normalizer;
+    xn[i] = wn[i] * static_cast<T>(y);
+  }
+  store_vector<N_READS>(out, index, xn, axis_size);
+}
+
 template <typename T, int BLOCK_DIM, int N_READS = 4>
 __global__ void rms_norm(
     const T* x,
@@ -94,6 +142,74 @@ __global__ void rms_norm(
   }
 }
 
+template <
+    typename T,
+    bool HAS_W,
+    int BLOCK_DIM,
+    int REDUCE_DIM,
+    int N_READS = 4>
+__global__ void rms_norm_vjp_small(
+    const T* x,
+    const T* w,
+    const T* g,
+    T* gx,
+    T* gw,
+    float eps,
+    int32_t axis_size,
+    int32_t n_rows,
+    int64_t w_stride) {
+  auto grid = cg::this_grid();
+  auto block = cg::this_thread_block();
+
+  using BlockReduceF2 = BlockBroadcastReduce<float2, BLOCK_DIM, REDUCE_DIM>;
+  __shared__ typename BlockReduceF2::TempStorage temp;
+
+  auto row =
+      (grid.block_rank() * block.dim_threads().y) + block.thread_index().y;
+  if (row >= n_rows) {
+    return;
+  }
+
+  x += row * axis_size;
+  g += row * axis_size;
+  gx += row * axis_size;
+  gw += row * axis_size;
+
+  // Normalizer.
+  float2 factors = {};
+  auto index = block.thread_index().x;
+  auto xn = load_vector<N_READS>(x, index, axis_size, T(0));
+  auto gn = load_vector<N_READS>(g, index, axis_size, T(0));
+  auto wn = load_vector<N_READS>(w, index, axis_size, w_stride, T(0));
+  for (int i = 0; i < N_READS; i++) {
+    float t = static_cast<float>(xn[i]);
+    float wi = wn[i];
+    float gi = gn[i];
+    float wg = wi * gi;
+    factors = plus_f2(factors, {wg * t, t * t});
+  }
+
+  factors = BlockReduceF2{block, temp}.Reduce(factors, plus_f2, {});
+  float meangwx = factors.x / axis_size;
+  float normalizer = rsqrt(factors.y / axis_size + eps);
+  float normalizer3 = normalizer * normalizer * normalizer;
+
+  // Outputs.
+  for (int i = 0; i < N_READS; i++) {
+    float xi = xn[i];
+    float wi = wn[i];
+    float gi = gn[i];
+    xn[i] = static_cast<T>(normalizer * wi * gi - xi * meangwx * normalizer3);
+    if constexpr (HAS_W) {
+      wn[i] = static_cast<T>(gi * xi * normalizer);
+    }
+  }
+  store_vector<N_READS>(gx, index, xn, axis_size);
+  if constexpr (HAS_W) {
+    store_vector<N_READS>(gw, index, wn, axis_size);
+  }
+}
+
 template <typename T, bool HAS_W, int BLOCK_DIM, int N_READS = 4>
 __global__ void rms_norm_vjp(
     const T* x,
@@ -107,12 +223,8 @@ __global__ void rms_norm_vjp(
   auto grid = cg::this_grid();
   auto block = cg::this_thread_block();
 
-  using BlockReduceF = BlockBroadcastReduce<float, BLOCK_DIM>;
   using BlockReduceF2 = BlockBroadcastReduce<float2, BLOCK_DIM>;
-  __shared__ union {
-    typename BlockReduceF::TempStorage f;
-    typename BlockReduceF2::TempStorage f2;
-  } temp;
+  __shared__ typename BlockReduceF2::TempStorage temp;
 
   x += grid.block_rank() * axis_size;
   g += grid.block_rank() * axis_size;
@@ -134,7 +246,7 @@ __global__ void rms_norm_vjp(
       factors = plus_f2(factors, {wg * t, t * t});
     }
   }
-  factors = BlockReduceF2{block, temp.f2}.Reduce(factors, plus_f2, {});
+  factors = BlockReduceF2{block, temp}.Reduce(factors, plus_f2, {});
   float meangwx = factors.x / axis_size;
   float normalizer = rsqrt(factors.y / axis_size + eps);
   float normalizer3 = normalizer * normalizer * normalizer;
@@ -169,6 +281,43 @@ bool RMSNorm::use_fallback(Stream s) {
   return s.device == Device::cpu;
 }
 
+template <int n_per_thread, typename F>
+void dispatch_group_dim(int axis_size, F&& f) {
+  if (axis_size <= n_per_thread * 8) {
+    f(std::integral_constant<int, 8>{},
+      std::integral_constant<int, 1>(),
+      std::integral_constant<int, 16>());
+  } else if (axis_size <= n_per_thread * 16) {
+    f(std::integral_constant<int, 16>{},
+      std::integral_constant<int, 1>(),
+      std::integral_constant<int, 8>());
+  } else if (axis_size <= n_per_thread * 32) {
+    f(std::integral_constant<int, 32>{},
+      std::integral_constant<int, 1>(),
+      std::integral_constant<int, 4>());
+  } else if (axis_size <= n_per_thread * 32 * 2) {
+    f(std::integral_constant<int, 32>{},
+      std::integral_constant<int, 2>(),
+      std::integral_constant<int, 2>());
+  } else if (axis_size <= n_per_thread * 32 * 4) {
+    f(std::integral_constant<int, 32>{},
+      std::integral_constant<int, 4>(),
+      std::integral_constant<int, 1>());
+  } else if (axis_size <= n_per_thread * 32 * 8) {
+    f(std::integral_constant<int, 32>{},
+      std::integral_constant<int, 8>(),
+      std::integral_constant<int, 1>());
+  } else if (axis_size <= n_per_thread * 32 * 16) {
+    f(std::integral_constant<int, 32>{},
+      std::integral_constant<int, 16>(),
+      std::integral_constant<int, 1>());
+  } else {
+    f(std::integral_constant<int, 32>{},
+      std::integral_constant<int, 32>(),
+      std::integral_constant<int, 1>());
+  }
+}
+
 // TODO: There are duplicate code with backend/metal/normalization.cpp
 void RMSNorm::eval_gpu(
     const std::vector<array>& inputs,
@@ -190,7 +339,7 @@ void RMSNorm::eval_gpu(
         out.copy_shared_buffer(x);
       } else {
         out.set_data(
-            cu::malloc_async(x.data_size() * x.itemsize(), encoder.stream()),
+            cu::malloc_async(x.data_size() * x.itemsize(), encoder),
             x.data_size(),
             x.strides(),
             x.flags());
@@ -216,12 +365,33 @@ void RMSNorm::eval_gpu(
   dispatch_float_types(out.dtype(), "rms_norm", [&](auto type_tag) {
     using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
     constexpr int N_READS = 16 / sizeof(DataType);
-    dispatch_block_dim(cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
-      auto kernel = cu::rms_norm<DataType, block_dim(), N_READS>;
+    if (axis_size <= N_READS * 1024) {
+      dispatch_group_dim<N_READS>(
+          axis_size, [&](auto group_dim, auto n_groups, auto groups_per_block) {
+            constexpr int block_dim = n_groups() * group_dim();
+            auto kernel =
+                cu::rms_norm_small<DataType, block_dim, group_dim(), N_READS>;
+            auto n_blocks =
+                (n_rows + groups_per_block() - 1) / groups_per_block();
+            encoder.add_kernel_node(
+                kernel,
+                n_blocks,
+                {block_dim, groups_per_block()},
+                0,
+                gpu_ptr<DataType>(x),
+                gpu_ptr<DataType>(w),
+                gpu_ptr<DataType>(out),
+                eps_,
+                axis_size,
+                n_rows,
+                w_stride);
+          });
+    } else {
+      auto kernel = cu::rms_norm<DataType, 1024, N_READS>;
       encoder.add_kernel_node(
           kernel,
           n_rows,
-          block_dim(),
+          1024,
           0,
           gpu_ptr<DataType>(x),
           gpu_ptr<DataType>(w),
@@ -229,7 +399,7 @@ void RMSNorm::eval_gpu(
           eps_,
           axis_size,
           w_stride);
-    });
+    }
   });
 }
 
@@ -274,7 +444,7 @@ void RMSNormVJP::eval_gpu(
     gx.copy_shared_buffer(g);
     g_in_gx = true;
   } else {
-    gx.set_data(cu::malloc_async(gx.nbytes(), encoder.stream()));
+    gx.set_data(cu::malloc_async(gx.nbytes(), encoder));
   }
   if (g_copied && !g_in_gx) {
     encoder.add_temporary(g);
@@ -292,7 +462,7 @@ void RMSNormVJP::eval_gpu(
     if (!g_in_gx && donate_g) {
       gw_temp.copy_shared_buffer(g);
     } else {
-      gw_temp.set_data(cu::malloc_async(gw_temp.nbytes(), encoder.stream()));
+      gw_temp.set_data(cu::malloc_async(gw_temp.nbytes(), encoder));
       encoder.add_temporary(gw_temp);
     }
   }
@@ -306,27 +476,51 @@ void RMSNormVJP::eval_gpu(
     dispatch_bool(has_w, [&](auto has_w_constant) {
       using DataType = cuda_type_t<MLX_GET_TYPE(type_tag)>;
       constexpr int N_READS = 16 / sizeof(DataType);
-      dispatch_block_dim(
-          cuda::ceil_div(axis_size, N_READS), [&](auto block_dim) {
-            auto kernel = cu::rms_norm_vjp<
-                DataType,
-                has_w_constant.value,
-                block_dim(),
-                N_READS>;
-            encoder.add_kernel_node(
-                kernel,
-                n_rows,
-                block_dim(),
-                0,
-                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);
-          });
+      if (axis_size <= N_READS * 1024) {
+        dispatch_group_dim<N_READS>(
+            axis_size,
+            [&](auto group_dim, auto n_groups, auto groups_per_block) {
+              constexpr int block_dim = group_dim() * n_groups();
+              auto kernel = cu::rms_norm_vjp_small<
+                  DataType,
+                  has_w_constant.value,
+                  block_dim,
+                  group_dim(),
+                  N_READS>;
+              auto n_blocks =
+                  (n_rows + groups_per_block() - 1) / groups_per_block();
+              encoder.add_kernel_node(
+                  kernel,
+                  n_blocks,
+                  {block_dim, groups_per_block()},
+                  0,
+                  gpu_ptr<DataType>(x),
+                  gpu_ptr<DataType>(w),
+                  gpu_ptr<DataType>(g),
+                  gpu_ptr<DataType>(gx),
+                  gpu_ptr<DataType>(gw_temp),
+                  eps_,
+                  axis_size,
+                  n_rows,
+                  w_stride);
+            });
+      } else {
+        auto kernel =
+            cu::rms_norm_vjp<DataType, has_w_constant.value, 1024, N_READS>;
+        encoder.add_kernel_node(
+            kernel,
+            n_rows,
+            1024,
+            0,
+            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/rope.cu

@@ -292,14 +292,14 @@ void RoPE::eval_gpu(
       donated = true;
       out.copy_shared_buffer(in);
     } else {
-      out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+      out.set_data(cu::malloc_async(out.nbytes(), encoder));
     }
     strides[0] = mat_size;
     strides[1] = in.strides()[ndim - 2];
     strides[2] = in.strides()[ndim - 1];
   } else if (dispatch_ndim == 3) {
     // Handle non-contiguous 3D inputs
-    out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+    out.set_data(cu::malloc_async(out.nbytes(), encoder));
     strides[0] = in.strides()[ndim - 3];
     strides[1] = in.strides()[ndim - 2];
     strides[2] = in.strides()[ndim - 1];

mlx/backend/cuda/scaled_dot_product_attention.cpp

@@ -0,0 +1,629 @@
+// Copyright © 2025 Apple Inc.
+
+#include "mlx/backend/cuda/cudnn_utils.h"
+#include "mlx/backend/cuda/device.h"
+#include "mlx/backend/cuda/lru_cache.h"
+#include "mlx/backend/gpu/copy.h"
+#include "mlx/fast_primitives.h"
+
+#include <nvtx3/nvtx3.hpp>
+
+namespace mlx::core {
+
+namespace fe = cudnn_frontend;
+
+namespace {
+
+#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)
+
+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;
+  }
+  for (int i = x.ndim() - 2; i >= 0; --i) {
+    if (x.shape(i) == 1) {
+      strides[i] = x.shape(i + 1) * strides[i + 1];
+    }
+  }
+  return strides;
+}
+
+void set_tensor_attrs(
+    std::shared_ptr<fe::graph::Tensor_attributes>& tensor,
+    int64_t uid,
+    const array& x) {
+  tensor->set_uid(uid)
+      .set_dim({x.shape().begin(), x.shape().end()})
+      .set_stride(normalized_strides(x));
+}
+
+array prepare_sdpa_input(const array& x, Stream s) {
+  // SDPA kernel's requirements on inputs:
+  // 1. last dim's stride be 1;
+  // 2. pointer be aligned.
+  if (x.strides(-1) != 1 || get_alignment(x) < 16) {
+    array x_copy = contiguous_copy_gpu(x, s);
+    auto& encoder = cu::get_command_encoder(s);
+    encoder.add_temporary(x_copy);
+    return x_copy;
+  }
+  return x;
+}
+
+void malloc_with_same_layout(
+    cu::CommandEncoder& encoder,
+    array& o,
+    const array& q) {
+  if (q.flags().row_contiguous) {
+    o.set_data(cu::malloc_async(o.nbytes(), encoder));
+    return;
+  }
+  // fill_order = argsort(q.strides())
+  Shape fill_order(q.ndim());
+  std::iota(fill_order.begin(), fill_order.end(), 0);
+  std::stable_sort(
+      fill_order.begin(), fill_order.end(), [&q](int idx1, int idx2) {
+        auto s1 = q.strides(idx1) > 0 ? q.strides(idx1) : 1;
+        auto s2 = q.strides(idx2) > 0 ? q.strides(idx2) : 1;
+        return s1 < s2;
+      });
+  // Generate o_strides with fill_order
+  Strides o_strides(q.ndim());
+  int64_t stride = 1;
+  for (int i : fill_order) {
+    o_strides[i] = stride;
+    stride *= o.shape(i);
+  }
+  // o is a transposed contiguous array
+  o.set_data(
+      cu::malloc_async(o.nbytes(), encoder),
+      o.size(),
+      o_strides,
+      {true, false, false});
+}
+
+constexpr int QKV_NDIM = 4;
+
+struct SDPACacheKey {
+  int device_id;
+  cudnnDataType_t cudnn_dtype;
+  std::array<int, QKV_NDIM> q_shape;
+  std::array<int, QKV_NDIM> k_shape;
+  std::array<int, QKV_NDIM> v_shape;
+  std::array<int64_t, QKV_NDIM> q_strides;
+  std::array<int64_t, QKV_NDIM> k_strides;
+  std::array<int64_t, QKV_NDIM> v_strides;
+  bool do_causal;
+  std::array<int, QKV_NDIM> mask_shape;
+  std::array<int64_t, QKV_NDIM> mask_strides;
+  bool output_logsumexp;
+};
+
+inline BytesKey<SDPACacheKey> build_sdpa_cache_key(
+    cu::CommandEncoder& encoder,
+    const array& q,
+    const array& k,
+    const array& v,
+    bool do_causal,
+    const std::optional<array>& mask_arr,
+    bool output_logsumexp = true) {
+  BytesKey<SDPACacheKey> cache_key;
+  cache_key.pod = {
+      encoder.device().cuda_device(),
+      dtype_to_cudnn_type(q.dtype()),
+      vector_key<QKV_NDIM>(q.shape()),
+      vector_key<QKV_NDIM>(k.shape()),
+      vector_key<QKV_NDIM>(v.shape()),
+      vector_key<QKV_NDIM>(q.strides()),
+      vector_key<QKV_NDIM>(k.strides()),
+      vector_key<QKV_NDIM>(v.strides()),
+      do_causal,
+      {},
+      {},
+      output_logsumexp,
+  };
+  if (mask_arr) {
+    cache_key.pod.mask_shape = vector_key<QKV_NDIM>(mask_arr->shape());
+    cache_key.pod.mask_strides = vector_key<QKV_NDIM>(mask_arr->strides());
+  }
+  return cache_key;
+}
+
+auto& sdpa_cache() {
+  static LRUBytesKeyCache<SDPACacheKey, fe::graph::Graph> cache(
+      "MLX_CUDA_SDPA_CACHE_SIZE", /* default_capacity */ 64);
+  return cache;
+}
+
+auto& sdpa_backward_cache() {
+  static LRUBytesKeyCache<SDPACacheKey, fe::graph::Graph> cache(
+      "MLX_CUDA_SDPA_BACKWARD_CACHE_SIZE", /* default_capacity */ 64);
+  return cache;
+}
+
+enum UIDS {
+  Q,
+  K,
+  V,
+  SCALE,
+  BIAS,
+  O,
+  STATS,
+  // Backward graph:
+  D_Q,
+  D_K,
+  D_V,
+  D_O,
+};
+
+fe::graph::Graph build_sdpa_graph(
+    cudnnHandle_t handle,
+    const array& q,
+    const array& k,
+    const array& v,
+    bool do_causal,
+    const std::optional<array>& mask_arr,
+    bool output_logsumexp,
+    const array& o,
+    const array& stats) {
+  auto dtype = fe::DataType_t::HALF;
+  if (q.dtype() == bfloat16) {
+    dtype = fe::DataType_t::BFLOAT16;
+  }
+
+  fe::graph::Graph graph;
+  graph.set_io_data_type(dtype)
+      .set_intermediate_data_type(fe::DataType_t::FLOAT)
+      .set_compute_data_type(fe::DataType_t::FLOAT);
+
+  auto q_ = graph.tensor(fe::graph::Tensor_attributes().set_name("Q"));
+  auto k_ = graph.tensor(fe::graph::Tensor_attributes().set_name("K"));
+  auto v_ = graph.tensor(fe::graph::Tensor_attributes().set_name("V"));
+  set_tensor_attrs(q_, Q, q);
+  set_tensor_attrs(k_, K, k);
+  set_tensor_attrs(v_, V, v);
+
+  auto scale = graph.tensor(fe::graph::Tensor_attributes()
+                                .set_name("Scale")
+                                .set_uid(SCALE)
+                                .set_dim({1, 1, 1, 1})
+                                .set_stride({1, 1, 1, 1})
+                                .set_is_pass_by_value(true)
+                                .set_data_type(fe::DataType_t::FLOAT));
+
+  auto options = fe::graph::SDPA_attributes()
+                     .set_name("sdpa_cudnn")
+                     .set_attn_scale(scale)
+                     .set_generate_stats(output_logsumexp);
+  if (do_causal) {
+    if (q.shape(2) > k.shape(2)) {
+      options.set_causal_mask(do_causal);
+    } else {
+      options.set_causal_mask_bottom_right(do_causal);
+    }
+  }
+  if (mask_arr) {
+    auto bias_ = graph.tensor(fe::graph::Tensor_attributes().set_name("BIAS"));
+    set_tensor_attrs(bias_, BIAS, *mask_arr);
+    options.set_bias(bias_);
+  }
+
+  auto [o_, stats_] = graph.sdpa(q_, k_, v_, options);
+  o_->set_output(true);
+  set_tensor_attrs(o_, O, o);
+  if (output_logsumexp) {
+    stats_->set_output(true).set_data_type(fe::DataType_t::FLOAT);
+    set_tensor_attrs(stats_, STATS, stats);
+  }
+
+  CHECK_CUDNN_FE_ERROR(graph.validate());
+  CHECK_CUDNN_FE_ERROR(graph.build_operation_graph(handle));
+  CHECK_CUDNN_FE_ERROR(graph.create_execution_plans({fe::HeurMode_t::A}));
+  graph.select_behavior_notes(
+      {fe::BehaviorNote_t::SUPPORTS_CUDA_GRAPH_NATIVE_API});
+  CHECK_CUDNN_FE_ERROR(graph.check_support(handle));
+  CHECK_CUDNN_FE_ERROR(graph.build_plans(handle));
+
+  return graph;
+}
+
+fe::graph::Graph build_sdpa_backward_graph(
+    cudnnHandle_t handle,
+    const array& q,
+    const array& k,
+    const array& v,
+    bool do_causal,
+    const std::optional<array>& mask_arr,
+    const array& o,
+    const array& d_o,
+    const array& stats,
+    array& d_q,
+    array& d_k,
+    array& d_v) {
+  auto dtype = fe::DataType_t::HALF;
+  if (q.dtype() == bfloat16) {
+    dtype = fe::DataType_t::BFLOAT16;
+  }
+
+  fe::graph::Graph graph;
+  graph.set_io_data_type(dtype)
+      .set_intermediate_data_type(fe::DataType_t::FLOAT)
+      .set_compute_data_type(fe::DataType_t::FLOAT);
+
+  auto q_ = graph.tensor(fe::graph::Tensor_attributes().set_name("Q"));
+  auto k_ = graph.tensor(fe::graph::Tensor_attributes().set_name("K"));
+  auto v_ = graph.tensor(fe::graph::Tensor_attributes().set_name("V"));
+  auto o_ = graph.tensor(fe::graph::Tensor_attributes().set_name("O"));
+  auto d_o_ = graph.tensor(fe::graph::Tensor_attributes().set_name("D_O"));
+  auto stats_ = graph.tensor(fe::graph::Tensor_attributes().set_name("STATS"));
+  set_tensor_attrs(q_, Q, q);
+  set_tensor_attrs(k_, K, k);
+  set_tensor_attrs(v_, V, v);
+  set_tensor_attrs(o_, O, o);
+  set_tensor_attrs(d_o_, D_O, d_o);
+  set_tensor_attrs(stats_, STATS, stats);
+  stats_->set_data_type(fe::DataType_t::FLOAT);
+
+  auto scale = graph.tensor(fe::graph::Tensor_attributes()
+                                .set_name("Scale")
+                                .set_uid(SCALE)
+                                .set_dim({1, 1, 1, 1})
+                                .set_stride({1, 1, 1, 1})
+                                .set_is_pass_by_value(true)
+                                .set_data_type(fe::DataType_t::FLOAT));
+
+  auto options = fe::graph::SDPA_backward_attributes()
+                     .set_name("sdpa_backward_cudnn")
+                     .set_attn_scale(scale)
+                     .set_attn_scale(scale);
+  if (do_causal) {
+    if (q.shape(2) > k.shape(2)) {
+      options.set_causal_mask(do_causal);
+    } else {
+      options.set_causal_mask_bottom_right(do_causal);
+    }
+  }
+  if (mask_arr) {
+    auto bias_ = graph.tensor(fe::graph::Tensor_attributes().set_name("BIAS"));
+    set_tensor_attrs(bias_, BIAS, *mask_arr);
+    options.set_bias(bias_);
+  }
+
+  auto [d_q_, d_k_, d_v_] =
+      graph.sdpa_backward(q_, k_, v_, o_, d_o_, stats_, options);
+  d_q_->set_output(true);
+  d_k_->set_output(true);
+  d_v_->set_output(true);
+  set_tensor_attrs(d_q_, D_Q, d_q);
+  set_tensor_attrs(d_k_, D_K, d_k);
+  set_tensor_attrs(d_v_, D_V, d_v);
+
+  CHECK_CUDNN_FE_ERROR(graph.validate());
+  CHECK_CUDNN_FE_ERROR(graph.build_operation_graph(handle));
+  CHECK_CUDNN_FE_ERROR(graph.create_execution_plans({fe::HeurMode_t::A}));
+  graph.select_behavior_notes(
+      {fe::BehaviorNote_t::SUPPORTS_CUDA_GRAPH_NATIVE_API});
+  CHECK_CUDNN_FE_ERROR(graph.check_support(handle));
+  CHECK_CUDNN_FE_ERROR(graph.build_plans(handle));
+
+  return graph;
+}
+
+void execute_graph(
+    cu::CommandEncoder& encoder,
+    cudnnHandle_t handle,
+    fe::graph::Graph& graph,
+    std::unordered_map<int64_t, void*>& variant_pack) {
+  int64_t workspace_size = 0;
+  CHECK_CUDNN_FE_ERROR(graph.get_workspace_size(workspace_size));
+  void* workspace_ptr = nullptr;
+  if (workspace_size > 0) {
+    array workspace(
+        cu::malloc_async(workspace_size, encoder),
+        {static_cast<int>(workspace_size)},
+        uint8);
+    encoder.add_temporary(workspace);
+    workspace_ptr = gpu_ptr<void>(workspace);
+  }
+
+  cudnnSetStream(handle, encoder.stream());
+
+  CudaGraph cuda_graph(encoder.device());
+  CHECK_CUDNN_FE_ERROR(graph.populate_cuda_graph(
+      handle, variant_pack, workspace_ptr, cuda_graph));
+  encoder.add_graph_node(cuda_graph);
+}
+
+} // namespace
+
+bool supports_sdpa_cudnn(
+    const array& q,
+    const array& k,
+    const array& v,
+    bool do_causal,
+    Stream s) {
+  static bool enabled = env::get_var("MLX_CUDA_USE_CUDNN_SPDA", 1);
+  if (!enabled) {
+    return false;
+  }
+
+  // cuDNN SDPA requires Ampere and later.
+  if (cu::device(s.device).compute_capability_major() < 8) {
+    return false;
+  }
+
+  // Only use cuDNN for prefilling (T_q > 1) and training (T_q == T_kv).
+  if ((q.shape(2) == 1) && (q.shape(2) != k.shape(2))) {
+    return false;
+  }
+
+  // D_qk and D_v must be a multiple of 8 with maximum value 128.
+  if ((q.shape(-1) % 8 != 0) || (q.shape(-1) > 128) || (v.shape(-1) % 8 != 0) ||
+      (v.shape(-1) > 128)) {
+    return false;
+  }
+
+  Dtype dtype = q.dtype();
+  return dtype == float16 || dtype == bfloat16;
+}
+
+void sdpa_cudnn(
+    const array& q,
+    const array& k,
+    const array& v,
+    float scale,
+    array& o,
+    array& stats,
+    bool do_causal,
+    const std::optional<array>& mask_arr,
+    bool output_logsumexp,
+    Stream s) {
+  auto& encoder = cu::get_command_encoder(s);
+  auto handle = encoder.device().cudnn_handle();
+
+  malloc_with_same_layout(encoder, o, q);
+
+  encoder.set_input_array(q);
+  encoder.set_input_array(k);
+  encoder.set_input_array(v);
+  encoder.set_output_array(o);
+  if (mask_arr) {
+    encoder.set_input_array(*mask_arr);
+  }
+  if (output_logsumexp) {
+    stats.set_data(cu::malloc_async(stats.nbytes(), encoder));
+    encoder.set_output_array(stats);
+  }
+
+  // Search cache.
+  auto cache_key = build_sdpa_cache_key(
+      encoder, q, k, v, do_causal, mask_arr, output_logsumexp);
+  auto it = sdpa_cache().find(cache_key);
+  if (it == sdpa_cache().end()) {
+    auto graph = build_sdpa_graph(
+        handle, q, k, v, do_causal, mask_arr, output_logsumexp, o, stats);
+    it = sdpa_cache().emplace(cache_key, std::move(graph)).first;
+  }
+  auto& graph = it->second;
+
+  std::unordered_map<int64_t, void*> variant_pack{
+      {Q, const_cast<void*>(gpu_ptr<void>(q))},
+      {K, const_cast<void*>(gpu_ptr<void>(k))},
+      {V, const_cast<void*>(gpu_ptr<void>(v))},
+      {SCALE, &scale},
+      {O, gpu_ptr<void>(o)}};
+  if (mask_arr) {
+    variant_pack[BIAS] = const_cast<void*>(gpu_ptr<void>(*mask_arr));
+  }
+  if (output_logsumexp) {
+    variant_pack[STATS] = gpu_ptr<void>(stats);
+  }
+
+  execute_graph(encoder, handle, graph, variant_pack);
+}
+
+void sdpa_backward_cudnn(
+    const array& q,
+    const array& k,
+    const array& v,
+    float scale,
+    const array& o,
+    const array& stats,
+    bool do_causal,
+    const std::optional<array>& mask_arr,
+    const array& d_o,
+    array& d_q,
+    array& d_k,
+    array& d_v,
+    Stream s) {
+  auto& encoder = cu::get_command_encoder(s);
+  auto handle = encoder.device().cudnn_handle();
+
+  malloc_with_same_layout(encoder, d_q, q);
+  malloc_with_same_layout(encoder, d_k, k);
+  malloc_with_same_layout(encoder, d_v, v);
+
+  encoder.set_input_array(q);
+  encoder.set_input_array(k);
+  encoder.set_input_array(v);
+  encoder.set_input_array(o);
+  encoder.set_input_array(stats);
+  encoder.set_input_array(d_o);
+  encoder.set_output_array(d_q);
+  encoder.set_output_array(d_k);
+  encoder.set_output_array(d_v);
+  if (mask_arr) {
+    encoder.set_input_array(*mask_arr);
+  }
+
+  // Search cache.
+  auto cache_key = build_sdpa_cache_key(encoder, q, k, v, do_causal, mask_arr);
+  auto it = sdpa_backward_cache().find(cache_key);
+  if (it == sdpa_backward_cache().end()) {
+    auto graph = build_sdpa_backward_graph(
+        handle, q, k, v, do_causal, mask_arr, o, d_o, stats, d_q, d_k, d_v);
+    it = sdpa_backward_cache().emplace(cache_key, std::move(graph)).first;
+  }
+  auto& graph = it->second;
+
+  std::unordered_map<int64_t, void*> variant_pack{
+      {Q, const_cast<void*>(gpu_ptr<void>(q))},
+      {K, const_cast<void*>(gpu_ptr<void>(k))},
+      {V, const_cast<void*>(gpu_ptr<void>(v))},
+      {SCALE, &scale},
+      {O, const_cast<void*>(gpu_ptr<void>(o))},
+      {STATS, const_cast<void*>(gpu_ptr<void>(stats))},
+      {D_O, const_cast<void*>(gpu_ptr<void>(d_o))},
+      {D_Q, gpu_ptr<void>(d_q)},
+      {D_K, gpu_ptr<void>(d_k)},
+      {D_V, gpu_ptr<void>(d_v)}};
+  if (mask_arr) {
+    variant_pack[BIAS] = const_cast<void*>(gpu_ptr<void>(*mask_arr));
+  }
+
+  execute_graph(encoder, handle, graph, variant_pack);
+}
+
+// Defined in scaled_dot_product_attention.cu file.
+bool supports_sdpa_vector(
+    const array& q,
+    const array& k,
+    const array& v,
+    bool has_mask,
+    bool has_arr_mask,
+    bool do_causal,
+    bool output_logsumexp);
+void sdpa_vector(
+    const array& q,
+    const array& k,
+    const array& v,
+    float scale,
+    array& o,
+    bool do_causal,
+    const std::optional<array>& sinks,
+    Stream s);
+
+namespace fast {
+
+bool ScaledDotProductAttention::use_fallback(
+    const array& q,
+    const array& k,
+    const array& v,
+    bool has_mask,
+    bool has_arr_mask,
+    bool do_causal,
+    bool is_training,
+    bool output_logsumexp,
+    Stream s) {
+  if (s.device == Device::cpu) {
+    return true;
+  }
+
+  return !supports_sdpa_vector(
+             q, k, v, has_mask, has_arr_mask, do_causal, output_logsumexp) &&
+      !supports_sdpa_cudnn(q, k, v, do_causal, s);
+}
+
+bool ScaledDotProductAttention::supports_bool_mask() {
+  return false;
+}
+
+void ScaledDotProductAttention::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  nvtx3::scoped_range r("ScaledDotProductAttention::eval_gpu");
+
+  auto& s = stream();
+
+  array q = prepare_sdpa_input(inputs[0], s);
+  array k = prepare_sdpa_input(inputs[1], s);
+  array v = prepare_sdpa_input(inputs[2], s);
+  auto& out = outputs[0];
+  auto& stats = outputs[1];
+  bool has_mask = inputs.size() - has_sinks_ > 3;
+  bool has_arr_mask = has_mask && !do_causal_;
+
+  std::optional<array> mask_arr;
+  if (has_arr_mask) {
+    mask_arr = prepare_sdpa_input(inputs[3], s);
+  }
+
+  if (supports_sdpa_vector(
+          q, k, v, has_mask, has_arr_mask, do_causal_, output_logsumexp_)) {
+    if (has_sinks_) {
+      sdpa_vector(q, k, v, scale_, out, do_causal_, inputs.back(), s);
+    } else {
+      sdpa_vector(q, k, v, scale_, out, do_causal_, std::nullopt, s);
+    }
+  } else {
+    sdpa_cudnn(
+        q,
+        k,
+        v,
+        scale_,
+        out,
+        stats,
+        do_causal_,
+        mask_arr,
+        output_logsumexp_,
+        s);
+  }
+}
+
+bool ScaledDotProductAttentionVJP::use_fallback(const array& q, Stream s) {
+  // The frontend adds a padding mask when sequence length is not a multiple of
+  // tile size.
+  if (q.shape(2) % 128 != 0) {
+    return true;
+  }
+  return s.device == Device::cpu;
+}
+
+void ScaledDotProductAttentionVJP::eval_gpu(
+    const std::vector<array>& inputs,
+    std::vector<array>& outputs) {
+  nvtx3::scoped_range r("ScaledDotProductAttentionVJP::eval_gpu");
+
+  auto& s = stream();
+
+  assert(inputs.size() >= 6);
+  int primals_size = inputs.size() - 3;
+  bool has_arr_mask = primals_size > 3 + has_sinks_;
+
+  array q = prepare_sdpa_input(inputs[0], s);
+  array k = prepare_sdpa_input(inputs[1], s);
+  array v = prepare_sdpa_input(inputs[2], s);
+  array o = prepare_sdpa_input(inputs[primals_size], s);
+  array stats = prepare_sdpa_input(inputs[primals_size + 1], s);
+  array d_o = prepare_sdpa_input(inputs[primals_size + 2], s);
+
+  std::optional<array> mask_arr;
+  if (has_arr_mask) {
+    mask_arr = prepare_sdpa_input(inputs[3], s);
+  }
+
+  assert(outputs.size() == 3);
+  auto& d_q = outputs[0];
+  auto& d_k = outputs[1];
+  auto& d_v = outputs[2];
+
+  sdpa_backward_cudnn(
+      q, k, v, scale_, o, stats, do_causal_, mask_arr, d_o, d_q, d_k, d_v, s);
+}
+
+} // namespace fast
+
+} // namespace mlx::core

mlx/backend/cuda/scaled_dot_product_attention.cu

@@ -6,10 +6,6 @@
 #include "mlx/backend/cuda/kernel_utils.cuh"
 #include "mlx/backend/gpu/copy.h"
 #include "mlx/dtype_utils.h"
-#include "mlx/fast_primitives.h"
-#include "mlx/transforms_impl.h"
-
-#include <nvtx3/nvtx3.hpp>
 
 #include <cooperative_groups.h>
 #include <cooperative_groups/reduce.h>
@@ -565,10 +561,9 @@ void sdpa_vector_2pass_fallback(
   array sums(intermediate_shape, float32, nullptr, {});
   array maxs(std::move(intermediate_shape), float32, nullptr, {});
 
-  intermediate.set_data(
-      cu::malloc_async(intermediate.nbytes(), encoder.stream()));
-  sums.set_data(cu::malloc_async(sums.nbytes(), encoder.stream()));
-  maxs.set_data(cu::malloc_async(maxs.nbytes(), encoder.stream()));
+  intermediate.set_data(cu::malloc_async(intermediate.nbytes(), encoder));
+  sums.set_data(cu::malloc_async(sums.nbytes(), encoder));
+  maxs.set_data(cu::malloc_async(maxs.nbytes(), encoder));
 
   encoder.add_temporary(intermediate);
   encoder.add_temporary(sums);
@@ -663,21 +658,16 @@ void sdpa_vector_fallback(
 
 } // namespace
 
-namespace fast {
-
-bool ScaledDotProductAttention::use_fallback(
+bool supports_sdpa_vector(
     const array& q,
     const array& k,
     const array& v,
     bool has_mask,
     bool has_arr_mask,
     bool do_causal,
-    Stream s) {
-  if (detail::in_grad_tracing()) {
-    return true;
-  }
-  if (s.device == Device::cpu) {
-    return true;
+    bool output_logsumexp) {
+  if (output_logsumexp) {
+    return false;
   }
 
   const int value_head_dim = v.shape(-1);
@@ -691,29 +681,24 @@ bool ScaledDotProductAttention::use_fallback(
   const bool supported_vector_config =
       sdpa_supported_head_dim && query_sequence_length < 4;
 
-  const bool supported_config = supported_vector_config;
-
-  return has_arr_mask || !supported_config;
+  return supported_vector_config && !has_arr_mask;
 }
 
-void ScaledDotProductAttention::eval_gpu(
-    const std::vector<array>& inputs,
-    array& out) {
-  nvtx3::scoped_range r("ScaledDotProductAttention::eval_gpu");
-
-  auto& s = stream();
+void sdpa_vector(
+    const array& q_pre,
+    const array& k_pre,
+    const array& v_pre,
+    float scale,
+    array& o,
+    bool do_causal,
+    const std::optional<array>& sinks_pre,
+    Stream s) {
   auto& encoder = cu::get_command_encoder(s);
-
-  auto& q_pre = inputs[0];
-  auto& k_pre = inputs[1];
-  auto& v_pre = inputs[2];
-  auto& o = out;
-
   std::vector<array> copies;
 
   // Define some copy functions to ensure the layout of the inputs is as
   // expected.
-  copies.reserve(inputs.size());
+  copies.reserve(4);
   auto copy_unless = [&copies, &s](
                          auto predicate, const array& arr) -> const array& {
     if (!predicate(arr)) {
@@ -731,8 +716,8 @@ void ScaledDotProductAttention::eval_gpu(
   };
 
   std::optional<array> sinks = std::nullopt;
-  if (has_sinks_) {
-    sinks = copy_unless(is_matrix_contiguous, inputs.back());
+  if (sinks_pre) {
+    sinks = copy_unless(is_matrix_contiguous, sinks_pre.value());
   }
 
   // We are in vector mode ie single query
@@ -788,7 +773,7 @@ void ScaledDotProductAttention::eval_gpu(
       };
 
       o.set_data(
-          cu::malloc_async(o.nbytes(), encoder.stream()),
+          cu::malloc_async(o.nbytes(), encoder),
           o.size(),
           {str_oB, str_oH, str_oL, str_oD},
           flags);
@@ -798,8 +783,7 @@ void ScaledDotProductAttention::eval_gpu(
       encoder.add_temporary(cp);
     }
 
-    return sdpa_vector_fallback(
-        s, encoder, q, k, v, scale_, o, do_causal_, sinks);
+    sdpa_vector_fallback(s, encoder, q, k, v, scale, o, do_causal, sinks);
   }
 
   // Full attention mode should never reach here
@@ -808,6 +792,4 @@ void ScaledDotProductAttention::eval_gpu(
   }
 }
 
-} // namespace fast
-
 } // namespace mlx::core

mlx/backend/cuda/scan.cu

@@ -374,7 +374,7 @@ void Scan::eval_gpu(const std::vector<array>& inputs, array& out) {
       out.copy_shared_buffer(in);
     } else {
       out.set_data(
-          cu::malloc_async(in.data_size() * out.itemsize(), encoder.stream()),
+          cu::malloc_async(in.data_size() * out.itemsize(), encoder),
           in.data_size(),
           in.strides(),
           in.flags());

mlx/backend/cuda/slicing.cpp

@@ -24,7 +24,7 @@ void concatenate_gpu(
   std::partial_sum(sizes.cbegin(), sizes.cend(), sizes.begin());
 
   auto& encoder = cu::get_command_encoder(s);
-  out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
+  out.set_data(cu::malloc_async(out.nbytes(), encoder));
 
   auto strides = out.strides();
   auto flags = out.flags();
@@ -89,7 +89,7 @@ array compute_dynamic_offset(
   if (donate) {
     offset.copy_shared_buffer(indices);
   } else {
-    offset.set_data(cu::malloc_async(offset.itemsize(), encoder.stream()));
+    offset.set_data(cu::malloc_async(offset.itemsize(), encoder));
   }
 
   encoder.add_temporary(offset);

mlx/backend/cuda/softmax.cu

@@ -118,7 +118,7 @@ void Softmax::eval_gpu(const std::vector<array>& inputs, array& out) {
         out.copy_shared_buffer(x);
       } else {
         out.set_data(
-            cu::malloc_async(x.data_size() * x.itemsize(), encoder.stream()),
+            cu::malloc_async(x.data_size() * x.itemsize(), encoder),
             x.data_size(),
             x.strides(),
             x.flags());

mlx/backend/cuda/sort.cu

@@ -49,14 +49,12 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
     array trans = swapaxes_in_eval(in, axis, last_dim);
     in = contiguous_copy_gpu(trans, s);
     encoder.add_temporary(in);
-    out = array(
-        cu::malloc_async(out.nbytes(), encoder.stream()),
-        in.shape(),
-        out.dtype());
+    out =
+        array(cu::malloc_async(out.nbytes(), encoder), in.shape(), out.dtype());
     encoder.add_temporary(out);
   } else {
     out.set_data(
-        cu::malloc_async(in.data_size() * out.itemsize(), encoder.stream()),
+        cu::malloc_async(in.data_size() * out.itemsize(), encoder),
         in.data_size(),
         in.strides(),
         in.flags());
@@ -74,17 +72,13 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
       if (argsort) {
         // Indices in the sorted dimension.
         array indices(
-            cu::malloc_async(out.nbytes(), encoder.stream()),
-            in.shape(),
-            out.dtype());
+            cu::malloc_async(out.nbytes(), encoder), in.shape(), out.dtype());
         encoder.add_temporary(indices);
 
         // In argsort though we don't need the result of sorted values, the
         // API requires us to provide an array to store it.
         array discard(
-            cu::malloc_async(in.nbytes(), encoder.stream()),
-            in.shape(),
-            in.dtype());
+            cu::malloc_async(in.nbytes(), encoder), in.shape(), in.dtype());
         encoder.add_temporary(discard);
 
         size_t size;
@@ -104,9 +98,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             stream));
 
         array temp(
-            cu::malloc_async(size, encoder.stream()),
-            {static_cast<int>(size)},
-            uint8);
+            cu::malloc_async(size, encoder), {static_cast<int>(size)}, uint8);
         encoder.add_temporary(temp);
 
         // Start capturing after allocations
@@ -148,9 +140,7 @@ void gpu_sort(const Stream& s, array in, array& out_, int axis, bool argsort) {
             stream));
 
         array temp(
-            cu::malloc_async(size, encoder.stream()),
-            {static_cast<int>(size)},
-            uint8);
+            cu::malloc_async(size, encoder), {static_cast<int>(size)}, uint8);
         encoder.add_temporary(temp);
 
         // Start capturing after allocations

mlx/backend/cuda/ternary.cu

@@ -257,9 +257,8 @@ void ternary_op_gpu(
   auto& c = inputs[2];
   auto topt = get_ternary_op_type(a, b, c);
   auto& encoder = cu::get_command_encoder(s);
-  set_ternary_op_output_data(a, b, c, out, topt, [&](auto n) {
-    return cu::malloc_async(n, encoder.stream());
-  });
+  set_ternary_op_output_data(
+      a, b, c, out, topt, [&](auto n) { return cu::malloc_async(n, encoder); });
   ternary_op_gpu_inplace<Op>(inputs, out, s);
 }
 

mlx/backend/cuda/unary/unary.cuh

@@ -208,9 +208,8 @@ void unary_op_gpu(
     const char* op,
     const Stream& s) {
   auto& encoder = cu::get_command_encoder(s);
-  set_unary_output_data(inputs[0], out, [&](auto n) {
-    return cu::malloc_async(n, encoder.stream());
-  });
+  set_unary_output_data(
+      inputs[0], out, [&](auto n) { return cu::malloc_async(n, encoder); });
   unary_op_gpu_inplace<Op>(inputs, out, op, s);
 }
 

mlx/backend/cuda/utils.cpp

@@ -5,6 +5,7 @@
 #include "mlx/dtype_utils.h"
 
 #include <fmt/format.h>
+#include <vector>
 
 namespace mlx::core {
 
@@ -60,7 +61,7 @@ const char* dtype_to_cuda_type(const Dtype& dtype) {
     case float64:
       return "double";
     case complex64:
-      return "complex64_t";
+      return "mlx::core::cu::complex64_t";
     default:
       return "unknown";
   }

mlx/backend/cuda/worker.cpp

@@ -44,7 +44,7 @@ void Worker::commit(cudaStream_t stream) {
   }
   signal_event_.record(stream);
   signal_event_.wait(signal_stream_);
-  cudaLaunchHostFunc(signal_stream_, signal, this);
+  CHECK_CUDA_ERROR(cudaLaunchHostFunc(signal_stream_, signal, this));
 }
 
 void Worker::thread_fn() {

mlx/backend/gpu/slicing.cpp

@@ -11,7 +11,7 @@ void slice_gpu(
     array& out,
     const Shape& start_indices,
     const Shape& strides,
-    const Stream& s) {
+    const Stream&) {
   slice(in, out, start_indices, strides);
 }
 

mlx/backend/metal/CMakeLists.txt

@@ -21,19 +21,14 @@ function(make_jit_source SRC_FILE)
   target_sources(mlx PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/jit/${SRC_NAME}.cpp)
 endfunction(make_jit_source)
 
-make_jit_source(
-  utils
-  kernels/jit/bf16.h
-  kernels/metal_3_0/bf16.h
-  kernels/metal_3_1/bf16.h
-  kernels/bf16_math.h
-  kernels/complex.h
-  kernels/defines.h)
+make_jit_source(utils kernels/bf16.h kernels/bf16_math.h kernels/complex.h
+                kernels/defines.h)
 make_jit_source(unary_ops kernels/erf.h kernels/expm1f.h kernels/fp8.h)
 make_jit_source(binary_ops)
 make_jit_source(ternary_ops)
 make_jit_source(reduce_utils kernels/atomic.h kernels/reduction/ops.h)
 make_jit_source(indexing/scatter kernels/indexing/indexing.h)
+make_jit_source(indexing/masked_scatter)
 make_jit_source(indexing/gather kernels/indexing/indexing.h)
 make_jit_source(indexing/gather_front kernels/indexing/indexing.h)
 make_jit_source(indexing/gather_axis)

mlx/backend/metal/custom_kernel.cpp

@@ -32,7 +32,7 @@ std::string write_signature(
     const std::vector<Dtype>& output_dtypes,
     const std::vector<std::pair<std::string, TemplateArg>>& template_args,
     const std::vector<std::string>& attributes,
-    const std::vector<CustomKernelShapeInfo>& shape_infos,
+    const std::vector<std::tuple<bool, bool, bool>>& shape_infos,
     bool atomic_outputs) {
   std::string kernel_source;
   kernel_source.reserve(header.size() + source.size() + 16384);
@@ -88,19 +88,19 @@ std::string write_signature(
     index++;
     // 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 constant int* " + name + "_shape [[buffer(" +
              std::to_string(index) + ")]],\n");
         index++;
       }
-      if (shape_infos[i].strides) {
+      if (std::get<1>(shape_infos[i])) {
         kernel_source +=
             ("  const constant int64_t* " + name + "_strides [[buffer(" +
              std::to_string(index) + ")]],\n");
         index++;
       }
-      if (shape_infos[i].ndim) {
+      if (std::get<2>(shape_infos[i])) {
         kernel_source +=
             ("  const constant int& " + name + "_ndim [[buffer(" +
              std::to_string(index) + ")]],\n");
@@ -184,12 +184,12 @@ CustomKernelFunction metal_kernel(
     throw std::invalid_argument(
         "[metal_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);
   }
   const std::vector<std::pair<std::string, std::string>> metal_attributes = {
@@ -388,15 +388,15 @@ void CustomKernel::eval_gpu(
     index++;
     if (in.ndim() > 0) {
       int ndim = in.ndim();
-      if (shape_info.shape) {
+      if (std::get<0>(shape_info)) {
         compute_encoder.set_vector_bytes(in.shape(), ndim, index);
         index++;
       }
-      if (shape_info.strides) {
+      if (std::get<1>(shape_info)) {
         compute_encoder.set_vector_bytes(in.strides(), ndim, index);
         index++;
       }
-      if (shape_info.ndim) {
+      if (std::get<2>(shape_info)) {
         compute_encoder.set_bytes(ndim, index);
         index++;
       }

mlx/backend/metal/device.cpp

@@ -21,12 +21,12 @@ constexpr const char* default_mtllib_path = METAL_PATH;
 
 auto get_metal_version() {
   auto get_metal_version_ = []() {
-    if (__builtin_available(macOS 15, iOS 18, tvOS 18, visionOS 2, *)) {
+    if (__builtin_available(macOS 26, iOS 26, tvOS 26, visionOS 26, *)) {
+      return MTL::LanguageVersion4_0;
+    } else if (__builtin_available(macOS 15, iOS 18, tvOS 18, visionOS 2, *)) {
       return MTL::LanguageVersion3_2;
-    } else if (__builtin_available(macOS 14, iOS 17, tvOS 17, visionOS 1, *)) {
-      return MTL::LanguageVersion3_1;
     } else {
-      return MTL::LanguageVersion3_0;
+      return MTL::LanguageVersion3_1;
     }
   };
   static auto metal_version_ = get_metal_version_();
@@ -119,8 +119,10 @@ std::pair<MTL::Library*, NS::Error*> load_swiftpm_library(
   // if SWIFTPM_BUNDLE is a framework identifier, try loading from that
   auto frameworks = NS::Bundle::allFrameworks();
   for (int i = 0, c = (int)frameworks->count(); i < c; i++) {
-    auto bundle = reinterpret_cast<NS::Bundle*>(frameworks->object(i));
-    if (!strcmp(bundle->bundleIdentifier()->utf8String(), SWIFTPM_BUNDLE)) {
+    const auto bundle = reinterpret_cast<NS::Bundle*>(frameworks->object(i));
+    const auto identifier = bundle->bundleIdentifier();
+    if (identifier != nullptr &&
+        !strcmp(identifier->utf8String(), SWIFTPM_BUNDLE)) {
       library = try_load_framework(device, bundle->resourceURL(), lib_name);
       if (library != nullptr) {
         return {library, nullptr};
@@ -380,11 +382,8 @@ MTL::CommandQueue* Device::get_queue(Stream stream) {
 
 bool Device::command_buffer_needs_commit(int index) {
   auto& stream = get_stream_(index);
-  if (stream.buffer_ops > max_ops_per_buffer_ ||
-      (stream.buffer_sizes >> 20) > max_mb_per_buffer_) {
-    return true;
-  }
-  return false;
+  return (stream.buffer_ops > max_ops_per_buffer_) ||
+      ((stream.buffer_sizes >> 20) > max_mb_per_buffer_);
 }
 
 MTL::CommandBuffer* Device::get_command_buffer(int index) {

mlx/backend/metal/device.h

@@ -265,4 +265,19 @@ Device& device(mlx::core::Device);
 
 std::unique_ptr<void, std::function<void(void*)>> new_scoped_memory_pool();
 
+inline bool is_nax_available() {
+  auto _check_nax = []() {
+    bool can_use_nax = false;
+    if (__builtin_available(
+            macOS 26.2, iOS 26.2, tvOS 26.2, visionOS 26.2, *)) {
+      can_use_nax = true;
+    }
+    can_use_nax &=
+        metal::device(mlx::core::Device::gpu).get_architecture_gen() >= 17;
+    return can_use_nax;
+  };
+  static bool is_nax_available_ = _check_nax();
+  return is_nax_available_;
+}
+
 } // namespace mlx::core::metal

mlx/backend/metal/distributed.cpp

@@ -30,4 +30,9 @@ void Recv::eval_gpu(const std::vector<array>&, std::vector<array>&) {
   throw std::runtime_error("[Recv::eval_gpu] has no GPU implementation.");
 }
 
+void ReduceScatter::eval_gpu(const std::vector<array>&, std::vector<array>&) {
+  throw std::runtime_error(
+      "[ReduceScatter::eval_gpu] has no GPU implementation.");
+}
+
 } // namespace mlx::core::distributed

mlx/backend/metal/indexing.cpp

@@ -1,4 +1,5 @@
 // Copyright © 2023-2024 Apple Inc.
+
 #include <fmt/format.h>
 
 #include "mlx/backend/common/compiled.h"
@@ -8,7 +9,9 @@
 #include "mlx/backend/metal/jit/includes.h"
 #include "mlx/backend/metal/jit/indexing.h"
 #include "mlx/backend/metal/kernels.h"
+#include "mlx/backend/metal/scan.h"
 #include "mlx/backend/metal/utils.h"
+#include "mlx/dtype.h"
 #include "mlx/primitives.h"
 #include "mlx/utils.h"
 
@@ -641,4 +644,84 @@ void ScatterAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
   compute_encoder.dispatch_threads(grid_dims, group_dims);
 }
 
+void MaskedScatter::eval_gpu(const std::vector<array>& inputs, array& out) {
+  const array& dst = inputs[0];
+  const array& mask = inputs[1];
+  const array& src = inputs[2];
+
+  auto& s = stream();
+  auto& d = metal::device(s.device);
+
+  const size_t total = mask.size();
+  const CopyType ct = (total == 1)
+      ? CopyType::Scalar
+      : (dst.flags().row_contiguous ? CopyType::Vector : CopyType::General);
+  copy_gpu(dst, out, ct, s);
+  if (total == 0) {
+    return;
+  }
+
+  array mask_flat = flatten_in_eval(mask, 1, -1, s);
+  if (mask_flat.data<void>() != mask.data<void>()) {
+    d.add_temporary(mask_flat, s.index);
+  }
+
+  if (!mask_flat.flags().row_contiguous) {
+    mask_flat = contiguous_copy_gpu(mask_flat, s);
+    d.add_temporary(mask_flat, s.index);
+  }
+
+  // Prefix (exclusive) of mask → scatter_offsets
+  array scatter_offsets(mask_flat.shape(), uint32, nullptr, {});
+  scatter_offsets.set_data(allocator::malloc(scatter_offsets.nbytes()));
+  d.add_temporary(scatter_offsets, s.index);
+
+  scan_gpu_inplace(
+      mask_flat,
+      scatter_offsets,
+      Scan::Sum,
+      /*axis=*/1,
+      /*reverse=*/false,
+      /*inclusive=*/false,
+      s);
+
+  // Kernel selection/build
+  static constexpr std::string_view kBaseName = "masked_assign";
+  const std::string dtype_tag = type_to_name(out.dtype());
+  const std::string value_type = get_type_string(out.dtype());
+  const std::string contiguous =
+      (src.flags().row_contiguous) ? "true" : "false";
+  const std::string kernel_name =
+      fmt::format("{}_{}_{}", kBaseName, dtype_tag, contiguous);
+
+  auto lib = d.get_library(kernel_name, [&]() {
+    std::string source = metal::utils();
+    source += metal::masked_scatter();
+    source += fmt::format(
+        std::string(masked_assign_kernel), kernel_name, value_type, contiguous);
+    return source;
+  });
+  auto kernel = d.get_kernel(kernel_name, lib);
+
+  // Binding
+  int bind_idx = 0;
+  const int ndim = static_cast<int>(src.ndim());
+  auto& compute_encoder = d.get_command_encoder(s.index);
+  compute_encoder.set_compute_pipeline_state(kernel);
+  compute_encoder.set_input_array(mask_flat, bind_idx++);
+  compute_encoder.set_input_array(scatter_offsets, bind_idx++);
+  compute_encoder.set_input_array(src, bind_idx++);
+  compute_encoder.set_output_array(out, bind_idx++);
+  compute_encoder.set_vector_bytes(src.shape(), bind_idx++);
+  compute_encoder.set_vector_bytes(src.strides(), bind_idx++);
+  compute_encoder.set_bytes(ndim, bind_idx++);
+  compute_encoder.set_bytes(src.size() / src.shape(0), bind_idx++);
+  compute_encoder.set_bytes(mask_flat.size() / mask.shape(0), bind_idx++);
+
+  // Dispatch
+  auto group_dims = get_block_dims(total, 1, 1);
+  MTL::Size grid_dims(total, 1, 1);
+  compute_encoder.dispatch_threads(grid_dims, group_dims);
+}
+
 } // namespace mlx::core

mlx/backend/metal/jit/includes.h

@@ -11,6 +11,7 @@ const char* ternary_ops();
 const char* reduce_utils();
 const char* gather();
 const char* scatter();
+const char* masked_scatter();
 
 const char* arange();
 const char* unary();

mlx/backend/metal/jit/indexing.h

@@ -70,3 +70,7 @@ constexpr std::string_view scatter_kernels = R"(
       gid);
 }}
 )";
+
+constexpr std::string_view masked_assign_kernel = R"(
+template [[host_name("{0}")]] [[kernel]] decltype(masked_assign_impl<{1}, {2}>) masked_assign_impl<{1}, {2}>;
+)";

mlx/backend/metal/kernels/CMakeLists.txt

@@ -1,6 +1,5 @@
 set(BASE_HEADERS
-    metal_3_1/bf16.h
-    metal_3_0/bf16.h
+    bf16.h
     bf16_math.h
     complex.h
     defines.h
@@ -10,7 +9,14 @@ set(BASE_HEADERS
     utils.h)
 
 function(build_kernel_base TARGET SRCFILE DEPS)
-  set(METAL_FLAGS -Wall -Wextra -fno-fast-math -Wno-c++17-extensions)
+  set(METAL_FLAGS
+      -x
+      metal
+      -Wall
+      -Wextra
+      -fno-fast-math
+      -Wno-c++17-extensions
+      -Wno-c++20-extensions)
   if(MLX_METAL_DEBUG)
     set(METAL_FLAGS ${METAL_FLAGS} -gline-tables-only -frecord-sources)
   endif()
@@ -18,16 +24,9 @@ function(build_kernel_base TARGET SRCFILE DEPS)
     set(METAL_FLAGS ${METAL_FLAGS}
                     "-mmacosx-version-min=${CMAKE_OSX_DEPLOYMENT_TARGET}")
   endif()
-  if(MLX_METAL_VERSION GREATER_EQUAL 310)
-    set(VERSION_INCLUDES
-        ${PROJECT_SOURCE_DIR}/mlx/backend/metal/kernels/metal_3_1)
-  else()
-    set(VERSION_INCLUDES
-        ${PROJECT_SOURCE_DIR}/mlx/backend/metal/kernels/metal_3_0)
-  endif()
   add_custom_command(
     COMMAND xcrun -sdk macosx metal ${METAL_FLAGS} -c ${SRCFILE}
-            -I${PROJECT_SOURCE_DIR} -I${VERSION_INCLUDES} -o ${TARGET}.air
+            -I${PROJECT_SOURCE_DIR} -o ${TARGET}.air
     DEPENDS ${SRCFILE} ${DEPS} ${BASE_HEADERS}
     OUTPUT ${TARGET}.air
     COMMENT "Building ${TARGET}.air"
@@ -128,6 +127,30 @@ if(NOT MLX_METAL_JIT)
   build_kernel(gemv_masked steel/utils.h)
 endif()
 
+if((MLX_METAL_VERSION GREATER_EQUAL 400) AND (MACOS_SDK_VERSION GREATER_EQUAL
+                                              26.2))
+  set(STEEL_NAX_HEADERS
+      steel/defines.h
+      steel/utils.h
+      steel/gemm/transforms.h
+      steel/gemm/nax.h
+      steel/gemm/gemm_nax.h
+      steel/utils/type_traits.h
+      steel/utils/integral_constant.h)
+
+  build_kernel(steel/gemm/kernels/steel_gemm_fused_nax ${STEEL_NAX_HEADERS})
+  build_kernel(steel/gemm/kernels/steel_gemm_gather_nax ${STEEL_NAX_HEADERS})
+
+  build_kernel(quantized_nax quantized_nax.h ${STEEL_NAX_HEADERS})
+  build_kernel(fp_quantized_nax fp_quantized_nax.h ${STEEL_NAX_HEADERS})
+
+  set(STEEL_NAX_ATTN_HEADERS
+      steel/defines.h steel/utils.h steel/attn/nax.h steel/utils/type_traits.h
+      steel/utils/integral_constant.h)
+
+  build_kernel(steel/attn/kernels/steel_attention_nax ${STEEL_NAX_ATTN_HEADERS})
+endif()
+
 add_custom_command(
   OUTPUT ${MLX_METAL_PATH}/mlx.metallib
   COMMAND xcrun -sdk macosx metallib ${KERNEL_AIR} -o

mlx/backend/metal/kernels/fp4.h

@@ -49,7 +49,10 @@ struct fp4_e2m1 {
   }
 
   operator float() {
-    return FP4_LUT[bits];
+    half converted = as_type<half>(ushort((bits & 7) << 9));
+    converted *= 16384.0;
+    converted = bits & 8 ? -converted : converted;
+    return converted;
   }
 
   uint8_t bits;

mlx/backend/metal/kernels/fp8.h

@@ -1,12 +1,5 @@
 #pragma once
 
-inline float fp32_from_bits(uint32_t bits) {
-  return *(reinterpret_cast<thread float*>(&bits));
-}
-inline float fp32_to_bits(float x) {
-  return *(reinterpret_cast<thread uint32_t*>(&x));
-}
-
 struct fp8_e4m3 {
   template <typename T>
   fp8_e4m3(T f) {
@@ -14,7 +7,7 @@ struct fp8_e4m3 {
     // https://github.com/pytorch/pytorch/blob/e3643e1e0e923f0fc063dfab6f45c956d568919d/c10/util/Float8_e4m3fn.h#L148
     uint32_t fp8_max = 543 << 21;
     uint32_t denorm_mask = 141 << 23;
-    uint32_t f_bits = fp32_to_bits(static_cast<float>(f));
+    uint32_t f_bits = as_type<uint32_t>(static_cast<float>(f));
     uint32_t sign = f_bits & 0x80000000;
     f_bits ^= sign;
     if (f_bits >= fp8_max) {
@@ -22,8 +15,8 @@ struct fp8_e4m3 {
       bits = 0x7E;
     } else {
       if (f_bits < (121 << 23)) {
-        f_bits =
-            fp32_to_bits(fp32_from_bits(f_bits) + fp32_from_bits(denorm_mask));
+        f_bits = as_type<uint32_t>(
+            as_type<float>(f_bits) + as_type<float>(denorm_mask));
         bits = static_cast<uint8_t>(f_bits - denorm_mask);
       } else {
         // resulting mantissa is odd
@@ -53,7 +46,7 @@ struct fp8_e4m3 {
         ((((nonsign << renorm_shift >> 4) + ((0x78 - renorm_shift) << 23)) |
           inf_nan_mask) &
          ~zero_mask);
-    return fp32_from_bits(result);
+    return as_type<float>(result);
   }
 
   uint8_t bits;
@@ -77,11 +70,12 @@ struct fp8_e8m0 {
     bits = static_cast<uint8_t>(n + 127);
   }
 
+  operator bfloat16_t() {
+    uint16_t out = (bits == 0 ? 0x40 : (static_cast<uint16_t>(bits) << 7));
+    return as_type<bfloat16_t>(out);
+  }
   operator float() {
-    if (bits == 0xFF) {
-      return metal::numeric_limits<float>::quiet_NaN();
-    }
-    return metal::ldexp(1.0f, static_cast<int>(bits) - 127);
+    return static_cast<float>(this->operator bfloat16_t());
   }
 
   uint8_t bits;

mlx/backend/metal/kernels/fp_quantized.h

@@ -29,15 +29,31 @@ inline constexpr short get_bytes_per_pack() {
 
 template <typename T>
 static inline T dequantize_scale(uint8_t s) {
-  using FOrI = union {
-    bfloat16_t f;
-    uint16_t i;
-  };
-  FOrI out;
-  out.i = (s == 0 ? 0x40 : (static_cast<uint16_t>(s) << 7));
-  return static_cast<T>(out.f);
+  return T(*(thread fp8_e8m0*)(&s));
 }
 
+template <int bits>
+struct Quantize {
+  uint8_t operator()(float x) {
+    if (bits == 8) {
+      return fp8_e4m3(x).bits;
+    } else {
+      return fp4_e2m1(x).bits;
+    }
+  }
+};
+
+template <int bits>
+struct Dequantize {
+  float operator()(uint8_t x) {
+    if (bits == 8) {
+      return float(*(thread fp8_e4m3*)(&x));
+    } else {
+      return float(*(thread fp4_e2m1*)(&x));
+    }
+  }
+};
+
 template <typename T, typename U, int values_per_thread>
 inline void load_vector(const device T* x, thread U* x_thread) {
   for (int i = 0; i < values_per_thread; i += 4) {
@@ -62,62 +78,41 @@ inline void load_vector_safe(const device T* x, thread U* x_thread, int N) {
   }
 }
 
-template <typename T>
-void load_fp4_lut(threadgroup T* lut, uint simd_gid, uint simd_lid) {
-  if (simd_gid == 0 && simd_lid < 16) {
-    lut[simd_lid] = static_cast<T>(FP4_LUT[simd_lid]);
-  }
-  threadgroup_barrier(mem_flags::mem_threadgroup);
-}
-
 template <typename U, int values_per_thread>
-inline U qdot(
-    const device uint8_t* w,
-    const thread U* x_thread,
-    U scale,
-    const threadgroup U* lut) {
+inline U qdot(const device uint8_t* w, const thread U* x_thread, U scale) {
   U accum = 0;
   const device uint16_t* ws = (const device uint16_t*)w;
   for (int i = 0; i < (values_per_thread / 4); i++) {
     accum +=
-        (x_thread[4 * i] * lut[ws[i] & 0xf] +
-         x_thread[4 * i + 1] * lut[(ws[i] >> 4) & 0xf] +
-         x_thread[4 * i + 2] * lut[(ws[i] >> 8) & 0xf] +
-         x_thread[4 * i + 3] * lut[(ws[i] >> 12) & 0xf]);
+        (x_thread[4 * i] * Dequantize<4>{}(ws[i]) +
+         x_thread[4 * i + 1] * Dequantize<4>{}(ws[i] >> 4) +
+         x_thread[4 * i + 2] * Dequantize<4>{}(ws[i] >> 8) +
+         x_thread[4 * i + 3] * Dequantize<4>{}(ws[i] >> 12));
   }
   return scale * accum;
 }
 
 template <typename U, int values_per_thread>
-inline U qdot_safe(
-    const device uint8_t* w,
-    const thread U* x_thread,
-    U scale,
-    const threadgroup U* lut,
-    int N) {
+inline U
+qdot_safe(const device uint8_t* w, const thread U* x_thread, U scale, int N) {
   U accum = 0;
 
   const device uint16_t* ws = (const device uint16_t*)w;
   for (int i = 0; i < (N / 4); i++) {
     accum +=
-        (x_thread[4 * i] * lut[ws[i] & 0xf] +
-         x_thread[4 * i + 1] * lut[(ws[i] >> 4) & 0xf] +
-         x_thread[4 * i + 2] * lut[(ws[i] >> 8) & 0xf] +
-         x_thread[4 * i + 3] * lut[(ws[i] >> 12) & 0xf]);
+        (x_thread[4 * i] * Dequantize<4>{}(ws[i]) +
+         x_thread[4 * i + 1] * Dequantize<4>{}(ws[i] >> 4) +
+         x_thread[4 * i + 2] * Dequantize<4>{}(ws[i] >> 8) +
+         x_thread[4 * i + 3] * Dequantize<4>{}(ws[i] >> 12));
   }
   return scale * accum;
 }
 
 template <typename U, int values_per_thread>
-inline void qouter(
-    const thread uint8_t* w,
-    U x,
-    U scale,
-    thread U* result,
-    const threadgroup U* lut) {
+inline void qouter(const thread uint8_t* w, U x, U scale, thread U* result) {
   for (int i = 0; i < (values_per_thread / 2); i++) {
-    result[2 * i] += x * scale * lut[w[i] & 0xf];
-    result[2 * i + 1] += x * scale * lut[(w[i] >> 4) & 0xf];
+    result[2 * i] += x * scale * Dequantize<4>{}(w[i]);
+    result[2 * i + 1] += x * scale * Dequantize<4>{}(w[i] >> 4);
   }
 }
 
@@ -192,7 +187,10 @@ struct QuantizedBlockLoader {
             bj * bytes_per_pack),
         scales(scales_ + bi * src_ld / group_size),
         lut(lut_) {
-    load_fp4_lut(lut, simd_group_id, simd_lane_id);
+    if (simd_group_id == 0 && simd_lane_id < 16) {
+      lut[simd_lane_id] = static_cast<T>(FP4_LUT[simd_lane_id]);
+    }
+    threadgroup_barrier(mem_flags::mem_threadgroup);
   }
 
   void load_unsafe() const {
@@ -264,10 +262,7 @@ METAL_FUNC void fp_qmv_quad_impl(
     const constant int& out_vec_size,
     uint3 tid [[threadgroup_position_in_grid]],
     uint quad_gid [[quadgroup_index_in_threadgroup]],
-    uint quad_lid [[thread_index_in_quadgroup]],
-    uint simd_gid [[simdgroup_index_in_threadgroup]],
-    uint simd_lid [[thread_index_in_simdgroup]],
-    threadgroup float* lut) {
+    uint quad_lid [[thread_index_in_quadgroup]]) {
   constexpr int quads_per_simd = SIMD_SIZE / QUAD_SIZE;
   constexpr int pack_factor = 8;
   constexpr int values_per_thread = D / QUAD_SIZE;
@@ -279,7 +274,6 @@ METAL_FUNC void fp_qmv_quad_impl(
 
   thread U x_thread[values_per_thread];
   thread U result[results_per_quadgroup] = {0};
-  load_fp4_lut(lut, simd_gid, simd_lid);
 
   // Adjust positions
   const int in_vec_size_w = in_vec_size / pack_factor;
@@ -299,7 +293,7 @@ METAL_FUNC void fp_qmv_quad_impl(
 
     U s = dequantize_scale<U>(sl[0]);
     if (row * quads_per_simd + out_row < out_vec_size) {
-      result[row] += qdot<U, values_per_thread>(wl, x_thread, s, lut);
+      result[row] += qdot<U, values_per_thread>(wl, x_thread, s);
     }
   }
 
@@ -321,8 +315,7 @@ METAL_FUNC void fp_qmv_fast_impl(
     const constant int& out_vec_size,
     uint3 tid [[threadgroup_position_in_grid]],
     uint simd_gid [[simdgroup_index_in_threadgroup]],
-    uint simd_lid [[thread_index_in_simdgroup]],
-    threadgroup float* lut) {
+    uint simd_lid [[thread_index_in_simdgroup]]) {
   constexpr int packs_per_thread = 2;
   constexpr int num_simdgroups = 2;
   constexpr int results_per_simdgroup = 4;
@@ -337,7 +330,6 @@ METAL_FUNC void fp_qmv_fast_impl(
   typedef float U;
   thread U x_thread[values_per_thread];
   thread U result[results_per_simdgroup] = {0};
-  load_fp4_lut(lut, simd_gid, simd_lid);
 
   // Adjust positions
   const int in_vec_size_w = in_vec_size * bytes_per_pack / pack_factor;
@@ -358,7 +350,7 @@ METAL_FUNC void fp_qmv_fast_impl(
       const device auto* sl = scales + row * in_vec_size_g;
 
       U s = dequantize_scale<U>(sl[0]);
-      result[row] += qdot<U, values_per_thread>(wl, x_thread, s, lut);
+      result[row] += qdot<U, values_per_thread>(wl, x_thread, s);
     }
 
     ws += block_size * bytes_per_pack / pack_factor;
@@ -384,8 +376,7 @@ METAL_FUNC void fp_qmv_impl(
     const constant int& out_vec_size,
     uint3 tid [[threadgroup_position_in_grid]],
     uint simd_gid [[simdgroup_index_in_threadgroup]],
-    uint simd_lid [[thread_index_in_simdgroup]],
-    threadgroup float* lut) {
+    uint simd_lid [[thread_index_in_simdgroup]]) {
   constexpr int num_simdgroups = 2;
   constexpr int results_per_simdgroup = 4;
   constexpr int packs_per_thread = 1;
@@ -402,7 +393,6 @@ METAL_FUNC void fp_qmv_impl(
 
   thread U x_thread[values_per_thread];
   thread U result[results_per_simdgroup] = {0};
-  load_fp4_lut(lut, simd_gid, simd_lid);
 
   // Adjust positions
   const int in_vec_size_w = in_vec_size * bytes_per_pack / pack_factor;
@@ -433,7 +423,7 @@ METAL_FUNC void fp_qmv_impl(
         const device auto* sl = scales + row * in_vec_size_g;
 
         uint8_t s = sl[0];
-        result[row] += qdot<U, values_per_thread>(wl, x_thread, s, lut);
+        result[row] += qdot<U, values_per_thread>(wl, x_thread, s);
       }
 
       ws += block_size * bytes_per_pack / pack_factor;
@@ -452,7 +442,7 @@ METAL_FUNC void fp_qmv_impl(
         const device auto* sl = scales + row * in_vec_size_g;
 
         U s = dequantize_scale<U>(sl[0]);
-        result[row] += qdot<U, values_per_thread>(wl, x_thread, s, lut);
+        result[row] += qdot<U, values_per_thread>(wl, x_thread, s);
       }
     }
 
@@ -481,7 +471,7 @@ METAL_FUNC void fp_qmv_impl(
         const device auto* sl = scales + row * in_vec_size_g;
 
         U s = dequantize_scale<U>(sl[0]);
-        result[row] += qdot<U, values_per_thread>(wl, x_thread, s, lut);
+        result[row] += qdot<U, values_per_thread>(wl, x_thread, s);
       }
 
       ws += block_size * bytes_per_pack / pack_factor;
@@ -501,7 +491,7 @@ METAL_FUNC void fp_qmv_impl(
 
         U s = dequantize_scale<U>(sl[0]);
         result[row] +=
-            qdot_safe<U, values_per_thread>(wl, x_thread, s, lut, remaining);
+            qdot_safe<U, values_per_thread>(wl, x_thread, s, remaining);
       }
     }
     for (int row = 0; row < results_per_simdgroup; row++) {
@@ -523,8 +513,7 @@ METAL_FUNC void fp_qvm_impl(
     const int out_vec_size,
     uint3 tid [[threadgroup_position_in_grid]],
     uint simd_gid [[simdgroup_index_in_threadgroup]],
-    uint simd_lid [[thread_index_in_simdgroup]],
-    threadgroup float* lut) {
+    uint simd_lid [[thread_index_in_simdgroup]]) {
   constexpr int num_simdgroups = 2;
   constexpr int pack_factor = get_pack_factor<32>();
   constexpr int bytes_per_pack = get_bytes_per_pack();
@@ -545,8 +534,6 @@ METAL_FUNC void fp_qvm_impl(
   thread U scale = 0;
   thread U x_local = 0;
 
-  load_fp4_lut(lut, simd_gid, simd_lid);
-
   // Adjust positions
   const int out_vec_size_w = out_vec_size * bytes_per_pack / pack_factor;
   const int out_vec_size_g = out_vec_size / group_size;
@@ -568,7 +555,7 @@ METAL_FUNC void fp_qvm_impl(
       scale = dequantize_scale<U>(*scales);
       w_local = *((device vec_w*)ws);
       qouter<U, tn * pack_factor>(
-          (thread uint8_t*)&w_local, x_local, scale, result, lut);
+          (thread uint8_t*)&w_local, x_local, scale, result);
 
       x += block_size;
       scales += block_size * out_vec_size_g;
@@ -581,7 +568,7 @@ METAL_FUNC void fp_qvm_impl(
       w_local = *((device vec_w*)ws);
 
       qouter<U, tn * pack_factor>(
-          (thread uint8_t*)&w_local, x_local, scale, result, lut);
+          (thread uint8_t*)&w_local, x_local, scale, result);
 
       x += block_size;
       scales += block_size * out_vec_size_g;
@@ -596,7 +583,7 @@ METAL_FUNC void fp_qvm_impl(
       scale = 0;
     }
     qouter<U, tn * pack_factor>(
-        (thread uint8_t*)&w_local, x_local, scale, result, lut);
+        (thread uint8_t*)&w_local, x_local, scale, result);
   }
 
 // Accumulate in the simdgroup
@@ -975,9 +962,7 @@ template <typename T, int group_size, int bits, int D, bool batched>
     const constant int64_t* s_strides,
     uint3 tid [[threadgroup_position_in_grid]],
     uint quad_gid [[quadgroup_index_in_threadgroup]],
-    uint quad_lid [[thread_index_in_quadgroup]],
-    uint simd_gid [[simdgroup_index_in_threadgroup]],
-    uint simd_lid [[thread_index_in_simdgroup]]) {
+    uint quad_lid [[thread_index_in_quadgroup]]) {
   if (batched) {
     int M = x_shape[x_batch_ndims];
     adjust_matrix_offsets(
@@ -995,20 +980,8 @@ template <typename T, int group_size, int bits, int D, bool batched>
         s_strides,
         tid);
   }
-  threadgroup float lut[16];
   fp_qmv_quad_impl<T, group_size, bits, D>(
-      w,
-      scales,
-      x,
-      y,
-      in_vec_size,
-      out_vec_size,
-      tid,
-      quad_gid,
-      quad_lid,
-      simd_gid,
-      simd_lid,
-      lut);
+      w, scales, x, y, in_vec_size, out_vec_size, tid, quad_gid, quad_lid);
 }
 
 template <typename T, int group_size, int bits, bool batched>
@@ -1046,9 +1019,8 @@ template <typename T, int group_size, int bits, bool batched>
         s_strides,
         tid);
   }
-  threadgroup float lut[16];
   fp_qmv_fast_impl<T, group_size, bits>(
-      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid, lut);
+      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid);
 }
 
 template <typename T, const int group_size, int bits, bool batched>
@@ -1086,9 +1058,8 @@ template <typename T, const int group_size, int bits, bool batched>
         s_strides,
         tid);
   }
-  threadgroup float lut[16];
   fp_qmv_impl<T, group_size, bits>(
-      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid, lut);
+      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid);
 }
 
 template <typename T, const int group_size, int bits, bool batched>
@@ -1126,9 +1097,8 @@ template <typename T, const int group_size, int bits, bool batched>
         s_strides,
         tid);
   }
-  threadgroup float lut[16];
   fp_qvm_impl<T, group_size, bits>(
-      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid, lut);
+      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid);
 }
 
 template <typename T, const int group_size, int bits, int split_k = 32>
@@ -1170,18 +1140,8 @@ template <typename T, const int group_size, int bits, int split_k = 32>
   int in_vec_size_adj =
       tid.z % split_k == split_k - 1 ? final_block_size : in_vec_size;
 
-  threadgroup float lut[16];
   fp_qvm_impl<T, group_size, bits>(
-      w,
-      scales,
-      x,
-      y,
-      in_vec_size_adj,
-      out_vec_size,
-      tid,
-      simd_gid,
-      simd_lid,
-      lut);
+      w, scales, x, y, in_vec_size_adj, out_vec_size, tid, simd_gid, simd_lid);
 }
 
 template <
@@ -1342,9 +1302,8 @@ template <typename T, int group_size, int bits>
       w_strides,
       s_strides,
       tid);
-  threadgroup float lut[16];
   fp_qmv_fast_impl<T, group_size, bits>(
-      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid, lut);
+      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid);
 }
 
 template <typename T, int group_size, int bits>
@@ -1392,9 +1351,8 @@ template <typename T, int group_size, int bits>
       w_strides,
       s_strides,
       tid);
-  threadgroup float lut[16];
   fp_qmv_impl<T, group_size, bits>(
-      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid, lut);
+      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid);
 }
 
 template <typename T, int group_size, int bits>
@@ -1442,9 +1400,8 @@ template <typename T, int group_size, int bits>
       w_strides,
       s_strides,
       tid);
-  threadgroup float lut[16];
   fp_qvm_impl<T, group_size, bits>(
-      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid, lut);
+      w, scales, x, y, in_vec_size, out_vec_size, tid, simd_gid, simd_lid);
 }
 
 template <
@@ -1771,28 +1728,6 @@ template <
   }
 }
 
-template <int bits>
-struct Quantize {
-  uint8_t operator()(float x) {
-    if (bits == 8) {
-      return fp8_e4m3(x).bits;
-    } else {
-      return fp4_e2m1(x).bits;
-    }
-  }
-};
-
-template <int bits>
-struct Dequantize {
-  float operator()(uint8_t x) {
-    if (bits == 8) {
-      return float(*(thread fp8_e4m3*)(&x));
-    } else {
-      return float(*(thread fp4_e2m1*)(&x));
-    }
-  }
-};
-
 template <typename T, const int group_size, const int bits>
 [[kernel]] void fp_quantize(
     const device T* w [[buffer(0)]],